The Sequence

April 3, 2026. Anthropic acquires Coefficient Bio, a stealth biotech founded eight months earlier, for roughly $400 million in stock. The team is fewer than ten people, nearly all former computational drug discovery researchers from Genentech’s Prescient Design, and their platform drafts drug R&D plans, manages clinical regulatory strategy, and proposes new drug candidates. It folds into Anthropic’s healthcare and life sciences division.

April 16, 2026. Anthropic announces a collaboration with Eli Lilly, presented by Dario Amodei alongside Lilly’s chief information and digital officer, covering clinical research and drug development. It follows an earlier partnership with AbbVie.

May 20, 2026. Bristol Myers Squibb announces a strategic agreement with Anthropic, an enterprise deployment spanning research, clinical development, manufacturing, commercial and corporate functions, reported as reaching a workforce of more than 30,000 employees.

June 9, 2026. Fable 5 ships, and with it a routing policy that diverts, in the company’s own words, the “majority of biology, chemistry, and life sciences queries, such as lab methods or molecular mechanisms” away from the most capable model. Not bioweapons queries. The majority of the field.

June 30, 2026. Anthropic announces Claude Science, a research workbench for target identification and lead optimization, and alongside it its own preclinical drug-discovery programs, run in house, reportedly aimed at neglected diseases.

Read that sequence in order and one thing jumps out. The therapeutics commitment came first. Anthropic did not drift into drug discovery after the fact; it spent $400 million in stock on a drug discovery team in April, signed Lilly sixteen days later and Bristol Myers Squibb in May, and only then shipped the model whose safeguards divert most of biology, chemistry and life sciences away from everybody else.

I want to be explicit that this is not a causal claim. I am not asserting that Anthropic broadened its biology filter in order to protect a position, and I have no evidence for that. But the ordering does retire one defense, which is that the drug business is a late afterthought bolted onto a safety posture that was already fixed. It was not an afterthought. It was bought, staffed and announced first.

Narrow, Then the Majority

It is worth being precise about when the breadth arrived, because the chronology is the strongest fact in this story and it is easy to overstate.

Anthropic’s classifier-based CBRN protections date to its ASL-3 activation in May 2025. But the company described those safeguards at the time as narrowly targeted, saying they “should not lead Claude to refuse queries except on a very narrow set of topics.” The policy of broadly diverting the majority of biology, chemistry and life sciences requests appears with Fable 5. So in roughly fourteen months, the company’s own description of what its safeguards touch moved from a very narrow set of topics to the majority of a scientific discipline. That is not an inference. It is Anthropic’s published language in both periods.

And it is not hiding the tradeoff. It drew it.

Anthropic’s own illustration. The band it widened is labeled, in its own key, benign but blocked.

There is a band of requests Anthropic classifies as benign and blocks anyway, to buy confidence about the harmful band further right, and for Fable 5 it made that band deliberately wider: “more benign requests would be blocked, but fewer genuinely harmful requests would be missed.” A tradeoff is a fair thing to make. But a tradeoff has two sides, and only one of them is in the room when the line gets drawn. The cost of a blocked benign request is not paid by Anthropic. It is paid by the chemist, the clinician and the graduate student, who appear in that diagram as a shaded rectangle.

Nobody needs a conspiracy to explain a company being careful about bioweapons, and the ASL-3 lineage argues against one. The weaker claim is the one that survives, and it requires no intent at all. Few organizations have ever combined frontier general-purpose scientific reasoning, unilateral control over its availability, an acquired drug discovery team, a portfolio of incumbent pharmaceutical customers and internal therapeutic programs under one roof. Anthropic now holds all five. There is no published firewall between the group that sets the biology threshold and the group running the drug programs, no outside auditor, and no disclosure of where the internal work sits relative to the line that stops everyone else.

The Canary Is Not the Story

I should say where my own evidence comes from, and then set it aside.

I run a drug discovery informatics company and have for 25 years. These notices appear in my terminal several times a week. The consolation used to be that a block on the top model dropped you a tier and you carried on working. That consolation is now gone: the same “intentionally broad safeguards” language has arrived on the fallback tier, which advises trying a different model. Inside the workflow I selected and pay for, there is no equivalent fallback model left.

That is a canary, not an argument. One flagged practitioner is an anecdote, and any company can point at an anecdote and promise to look into it. What makes it worth publishing is what it indicates about the air in the shaft. If ordinary target biology and molecular mechanism work trips the filter for someone with a Berkeley Ph.D. in biophysical chemistry, four decades in computational biology and a paid enterprise relationship, it is tripping for the graduate student, the twenty-person informatics shop and the biotech that has not raised its B round. Those organizations do not write essays about it. They quietly get less capable answers.

And here the asymmetry matters more than the inconvenience. A broad filter may genuinely reduce some misuse risk, and I am not going to pretend otherwise. But its practical burden is not distributed evenly, and the honest thing to say is that from the outside we cannot see how it is distributed at all.

Anthropic has not published whether a large enterprise agreement carries the same safeguard configuration as the subscription on my laptop. It has separately announced a trusted-access program that removes biology and chemistry safeguards for a small cohort of researchers it has not named. So I do not know whether Eli Lilly or Bristol Myers Squibb work under the same constraints I do. Neither does any other customer, competitor, or regulator. I am not going to assert that they are exempt, because I have no evidence of that, and I am not going to assert that they are not, because I have no evidence of that either. Nobody outside the company can tell. That opacity is not a side issue in this story. It is the story.

What is observable is the difference in what each party can do about it. Alphabet’s Isomorphic Labs, which develops proprietary drug-design models and infrastructure and raised $2.1 billion in May, is building on its own stack rather than renting someone else’s. A pharmaceutical company with tens of thousands of employees has procurement leverage, legal resources, internal infrastructure and a direct line to the vendor, so it is in a position to pursue accommodations, alternative models, or a place in that unnamed cohort. A graduate laboratory or a twenty-person biotech has none of those. Even a rule applied with perfect uniformity lands very differently depending on who you are, and at present no one outside Anthropic can confirm that it is applied uniformly.

Disease Is Not a Two-Company Problem

Here is my real objection, and it is not about my own access. It is about the bet.

There are more than 200 recognized types of cancer, with many molecularly distinct subtypes beneath them. What is clinically labeled Alzheimer’s disease encompasses substantial biological heterogeneity and frequently includes multiple coexisting pathologies. There are more than 10,000 recognized rare diseases, and the great majority have no FDA-approved treatment. Nothing on that list is solved by a few large pharmaceutical companies and one AI lab’s internal programs. It gets solved, if it gets solved, by an army: thousands of groups attacking thousands of targets in parallel, most of them failing, a few of them not, with the failures published so the next group does not repeat them.

Anthropic knows this. Its own life sciences lead framed the in-house programs around diseases traditional biopharma will not touch, which is a tacit admission that the field vastly exceeds any one organization’s bandwidth. That admission is correct. It also sits in tension with the architecture the company has built, because concentrating the most capable scientific reasoning behind a handful of incumbent pharmaceutical relationships and an in-house discovery effort is a bet that the binding constraint in medicine is capability rather than parallelism.

I think that bet is wrong, and wrong in a way that costs lives rather than money. Modern biology depends on a vast pre-competitive commons. Every molecule any of us has worked on rests on structures somebody deposited in the Protein Data Bank, sequences somebody released to GenBank, assays and dead ends somebody published anyway. The industry calls that layer pre-competitive for a reason: you compete on the molecule, not on the fact that the protein has a pocket. A layer of intelligence that draws freely on that commons and then rations what comes back out is not a neutral instrument sitting above the field. It narrows how many shots on goal the field gets to take.

Notice also who the incumbents are. AbbVie, Eli Lilly, Bristol Myers Squibb, and a client list reported to include Sanofi, Novo Nordisk and Genmab. These are not the organizations most at risk of being short of capability. The smaller companies and academic groups on the other side of the filter are a large part of where new biology has historically come from.

Call It What It Is

I have no criticism of Anthropic for wanting to win. It spent enormous capital and talent building these models, it is entitled to a return, and buying a drug discovery team and entering therapeutics is a legitimate commercial decision any board would consider.

My objection is to the label.

When a routing policy touches the majority of an entire scientific discipline, and the company operating it has an acquired drug discovery team, several of the largest pharmaceutical companies as customers, and therapeutic programs of its own, “safety” stops being a complete description of what the mechanism does. It may still be sincere. It is no longer sufficient. A safeguard and a competitive moat can be the same mechanism, and the party operating it has no commercial incentive to hurry the disentangling.

So say it out loud. Publish where the line sits, publish who it applies to, and publish whether the internal programs run against the same threshold external customers do. If the answer is that they do, that is a strong answer and it costs nothing to state. What is hard to defend is describing a boundary with commercial consequences in the vocabulary of biosecurity, and leaving the market to guess which one it is looking at.

The Question Every Other Lab Is About to Ask

Now set the ethics aside entirely and treat this as a vendor decision, which is where I have spent 25 years on the selling side of the table.

Big pharma does not buy software the way a startup does. It qualifies vendors: supplier assessment, risk classification, validation plan, change control, business continuity review, all of it because these systems eventually touch a regulatory filing. Four questions get asked. Can you guarantee availability. Will behavior change without notice. What is our fallback. And who else are you.

Anthropic’s own account of the June suspension and its restoration.

On the first three, the record of the last two months is mixed, and worth stating carefully rather than dramatically. Anthropic’s two newest frontier models went offline worldwide for roughly eighteen days in June under a US government directive concerning foreign-national access; the company has said the global disablement was the practical consequence of complying, since it could not verify nationality in real time, and its other models remained available throughout. On pricing, a model initially bundled into subscriptions was announced at launch as moving behind usage credits less than three weeks later, so this was disclosed rather than concealed, though the window was brief and commercially awkward for anyone building on it.

June 9. The launch announcement that set both the capability and its expiry date in the subscription.

A classifier boundary that moves without publication, underneath a system somebody validated last quarter, is an uncontrolled change. That is not a philosophical complaint. It is a finding, and the remediation is revalidation.

But it is the fourth question that has changed, and it is the one I would put to any life sciences organization currently in procurement. Your prospective platform vendor bought a drug discovery company in April, signed two of your largest competitors in April and May, and announced its own therapeutic programs in June. Nothing improper need be happening with anyone’s data, and I have no reason to think it is. But somebody in your legal department is going to write a memo containing the words our shared intelligence platform is operated by a party that has entered our industry, and that memo does not have to reach a bad conclusion to be expensive. It adds clauses. It adds review cycles. At renewal, it adds a second source.

So, honestly: knowing what you now know, do you route your target biology through that platform? Or do you build an abstraction layer, keep two providers behind it, hold an open-weight model in reserve for anything the filter is likely to touch, and never let one vendor sit alone in the critical path? I built exactly that in July. I doubt I am the only one, and I am not going to tear it out.

That is the part I think gets missed inside a company that is winning. Dependency is only a moat if it is reliable. Make it unreliable, or make it conflicted, and you do not get a captive customer. You get an architecture decision.

Three Things to Publish

None of these removes a guardrail. All of them convert a private judgment into a legible one, and every one is in Anthropic’s own commercial interest.

One: turn the announced biology trusted-access program into a real front door. To the company’s credit, it has said it will enroll a small number of life sciences researchers and remove the biology and chemistry safeguards for them. That is the right instinct, and it is not yet a process. Make it comparable to the Cyber Verification Program, which is public, free, scoped to an organization and answered in about two business days: published criteria, organizational verification, a defined review period, and an answer either way.

Two: let verified identity reach the classifier. There is no anonymous user of a paid API. Every request carries an organization-scoped key, which is how the invoice gets addressed. A system that bills a customer by name and judges them as a stranger is not a safety architecture. It is an unfinished one.

Three: disclose the firewall, and publish a pre-competitive floor. State whether the internal therapeutic programs and the acquired discovery team run against the same threshold as customers, and name the class of work that will not be gated: target biology, mechanism, structure, the ordinary literate content of the field. The commons these models were trained on should not be harder to reach through the model than through PubMed.

The Stakes

The frontier is consolidating into a handful of firms that own the capability, the criteria that ration it, and now the downstream products built from it. Some of that rationing is genuinely safety. But its burden falls hardest on the part of the field least able to absorb it, and hardly at all on the organizations best positioned to compete.

Nobody is going to cure Alzheimer’s inside one company. Not Anthropic, not Isomorphic, not Lilly or Bristol Myers Squibb. It will take an army, drawing on a commons, most of it failing in public so the rest can move. The question is whether the most powerful scientific instrument ever built gets pointed at that whole army, or at a few flagship accounts and an in-house research effort while everybody else is told, in the calm register of a safety notice, to try a different model.

“Try a different model” is not an answer when the model being withheld is the capability the customer selected and paid to use.

REFERENCES

1. TechCrunch, “Anthropic buys biotech startup Coefficient Bio in $400M deal,” Apr. 3, 2026

2. Fierce Biotech, “Anthropic acquires stealth AI startup Coefficient Bio in $400M deal,”April 2026

3. “Claude partners with Eli Lilly on AI-driven drug development and clinical research,” Apr. 16, 2026

4. Fierce Pharma, “BMS taps Anthropic’s Claude for enterprise-wide AI adoption to speed drug R&D, global workflows,” May 20, 2026

5. Drug Discovery News, “Weekly rundown: Anthropic and Lilly’s deals prove AI is becoming biopharma’s biggest infrastructure bet,” 2026

6. Ashley Capoot, “Anthropic launches AI drug discovery program, Claude Science,” CNBC, June 30, 2026

7. Anthropic Support, “Why Claude switched models in your conversation with Fable 5,”accessed July 2026

8. Anthropic, “Claude Fable 5 and Claude Mythos 5,” June 9, 2026

9. Anthropic, “Activating AI Safety Level 3 protections,” May 22, 2025

10. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” June 12, 2026

11. Anthropic, “Usage Policy,” accessed July 2026

12. Kyle Orland, “Anthropic says these topics are too dangerous to let its Fable 5 model talk about,” Ars Technica, June 9, 2026

13. “Alphabet-backed Isomorphic Labs raises $2.1B to accelerate AI-designed drug discovery as clinical trials near,” Tech Startups, May 12, 2026

14. Michael A. Heller and Rebecca S. Eisenberg, “Can Patents Deter Innovation? The Anticommons in Biomedical Research,” Science 280:698, May 1, 1998

15. Structural Genomics Consortium, “Open Science,” accessed July 2026

16. Steven Muskal, “Universities Built the Bomb. Now Build This,” Renaissance Circle, Jun. 26, 2026

17. Steven Muskal, “AI for Everyone,” Renaissance Circle, Jul. 23, 2026

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

By Steven Muskal, Ph.D. | The Renaissance Circle | July 26, 2026 | stevenmuskal.com

I have spent nearly four decades on one side or another of the gap between a molecule’s shape and what it does inside a living system. I built structure-based drug design pipelines before anyone called it that. I ran neural networks on protein sequences in the early 1990s, back when reviewers asked, not always kindly, why a chemist was talking about backpropagation. So when I say that AlphaFold is one of the great achievements of modern science, understand that I am not a bystander offering polite applause. I am someone who started my career failing to do, by hand and by heuristic, what DeepMind’s system now does routinely before lunch. In fairness, when I was working in the area, I only had 200 protein structures to work with. The AlphaFold crew (30 years later) had three orders of magnitude - almost 200,000 protein structures and a massive database of sequences and could leverage the concept that varying sequences across species can fold into the same protein structure, but that’s a whole other story. As I’ve said before, content is king - or rather Content Makes Kings.

AlphaFold solved a problem that had genuinely resisted decades of research: given a protein's amino acid sequence, predict its three-dimensional structure to near-experimental accuracy, the result John Jumper and colleagues published in Nature in 2021. That is extraordinary, and nothing here is an attempt to diminish it. But sitting with the result these past several years has left me with a question I cannot set aside, and it is not really a question about proteins at all. Did we point all of this at the right problem?

Nature Was Never Optimizing Structure

Here is the pivot that changed how I think about my own field. Nature does not optimize protein structures. It optimizes reproductive fitness, through an almost unbearably narrow channel: evolution reads and rewrites DNA, and it observes, through selection, whether the organism carrying that DNA survives and reproduces. That is the whole loop. DNA, organism, fitness. Nothing else exists to evolution as a target.

Evolution has no notion of alpha helices, beta sheets, RMSD, or pLDDT. Those are human abstractions, useful ones, that we invented to make an incomprehensible molecular world legible to ourselves. Protein folding, signaling cascades, metabolism, and physiology all persist because they improve fitness, not because evolution was ever aiming at any of them. Everything between the genotype and the outcome is, to the process that shaped it, invisible machinery. We are the ones who opened the box and started naming the gears. And a protein, for the record, does not even hold still: it is an ensemble of shifting conformations exploring an energy landscape, never a single frozen shape. A crystal structure is one useful snapshot under one artificial condition, not the thing itself. I raise that not to relitigate structural biology, which I love, but to make one point: we have poured a generation of talent into predicting a static object that nature never optimized and never stores.

Evolution reads and rewrites DNA, but it has no notion of alpha helices or RMSD, only whether the organism survives to reproduce. Everything in between is invisible machinery.

What Language Models Already Taught Us

I did not arrive at this through biology. I arrived at it by watching my other field, artificial intelligence, over the last decade.

For a long time the dominant intuition in natural language processing was decompositional, in exactly the way biology is decompositional today. A machine that truly understood language, the thinking went, would need dedicated modules: one for syntax, one for grammar, one for semantics, one for logic, one for world knowledge. Whole subfields organized themselves around these presumed joints in the problem, each with its own benchmarks and its own architectures.

Then transformer-based language models, trained on nothing more sophisticated than predicting the next token, blew past every one of those hand-engineered systems. Nobody supervised grammar neurons. Nobody hand-labeled a training set for logical inference. One almost embarrassingly simple objective, at sufficient scale, caused syntax, semantics, world knowledge, and something that behaves a great deal like reasoning to emerge on their own, as internal representations, because they were useful for the one thing the model was actually asked to do. Nobody requested them by name. A 2022 survey of this phenomenon across dozens of model families gave it a name of its own, emergent abilities: capabilities that are simply absent in smaller models and appear, often abruptly, once scale crosses a threshold, unpredictable from the smaller models’ performance curves alone.

No one supervised “grammar neurons.” Syntax, semantics, and reasoning emerged from a single broad objective, next-token prediction, at scale.

Biology should sit with that result longer than it has. Perhaps protein structure, pathway membership, and the rest of our beloved intermediate abstractions are not things we should be supervising a model to reproduce at all. Perhaps they should be allowed to emerge, as latent representations, inside a model trained on a broader and more consequential objective, the way grammar emerged inside a language model without anyone asking for it.

The Brain Is a Black Box, and We Use It Anyway

My UC Berkeley research director said something to me recently that I have not been able to stop thinking about. Our brains, he pointed out, are black boxes. We do not understand, in any mechanistic and complete way, how they produce a decision, a memory, or a moment of recognition. And yet we rely on them, all day, every day, to do exactly those things. We trust the output without auditing the mechanism.

That is not a failure of science. It is how we operate with nearly every complex system that matters to us. I do not know the control theory that keeps my car on the road, and I drive anyway. We do not need to derive the internals to make use of the behavior. This is worth saying plainly, because medicine has quietly adopted the opposite creed. We have told ourselves that we may only act once we can trace the mechanism, kinase by kinase, receptor by receptor, structure by structure. Understanding mechanism is a noble goal, and I have spent my career chasing it. But it is not a precondition for helping someone. If we treat human physiology the way we already treat the brain, as a black box with inputs and outputs, then the useful question stops being why and becomes which input changes the output for the better.

Our brains are black boxes. We do not know how they work, and we rely on them every single day. Why would we insist on rationalizing every mechanism inside the body before we are allowed to act on it?

The Corpus Biology Refuses to Use

If medicine is a black box of inputs and outputs, then the obvious next question is: what inputs do we actually have? And here is where I think the field has been staring past an ocean of signal while panning for gold in a stream.

We are not, realistically, going to get most of the population genetically sequenced, phenotyped, and enrolled in longitudinal clinical cohorts. That world is expensive, slow, and consented one person at a time. But look at what we already generate, continuously, without anyone having to run a study. Our economy has gone almost entirely electronic. People pay with credit and debit cards. Amazon knows what arrives on the doorstep. Grocery and pharmacy chains know what goes in the cart, and when the pattern changes. Wearables capture heart rate, sleep, movement, and increasingly glucose and blood oxygen, minute by minute. The social platforms, Meta and Amazon chief among them, link products to purchases to behavior to attention at a resolution no hospital system has ever approached.

This is an enormous corpus of human behavior, and it is a physiological record whether we admit it or not. What you buy, how you move, when you sleep, how those patterns drift over months and years: these are outputs of the same organism medicine is trying to help. Diet is purchased. Sedentary decline shows up in step counts and in the shift from groceries to delivery. Alcohol, tobacco, and sugar are line items. Stress and depression change spending, sleep, and social behavior in measurable ways before they ever reach a clinician. We have built, almost by accident, a real-time sensor network over the daily lives of hundreds of millions of people. We just aimed the whole thing at selling them shoes.

The Population Evidence Starts to Arrive

When I first wrote about this idea, I could point to the shape of the argument but not to a body of population-scale results behind it. Since then I went looking, specifically for studies with real cohorts, real sensors, and real numbers, not case reports of one or two patients. What I found does not prove the thesis outright, but it is far more than a hunch, and some of it is already at the scale a health foundation model would need.

Start with the closest thing to a direct test of the wearables argument. Jennifer Radin and colleagues at Scripps Research pulled de-identified Fitbit data from 200,000 users across the United States, narrowed to 47,249 people in five states who wore the device consistently for two full years, and compared week-to-week changes in resting heart rate and sleep against the CDC’s official influenza-like-illness (ILI) surveillance numbers. Across more than 13.3 million individual measurements, adding the wearable signal to a standard surveillance model improved the correlation with CDC-reported ILI rates by an average of 0.12 (up to 32.9%), and the final models tracked the official numbers with correlations of 0.84 to 0.97. Nobody enrolled these 47,249 people in a flu study. They just wore a Fitbit, and the exhaust from that ordinary decision turned out to be real-time epidemiological signal.

The same logic holds for a sharper, more acute event: an active infection rather than a seasonal trend. Tejaswini Mishra and colleagues at Stanford followed nearly 5,300 people wearing consumer smartwatches and identified 32 who tested positive for COVID-19 during the study window. Twenty-six of those 32 (81%) showed a detectable alteration in heart rate, step count, or sleep. Of the cases with both a physiological signal and known symptom-onset timing, the alteration appeared at or before symptoms in the large majority, and four cases were flagged at least nine days before the person felt sick enough to notice. Using a simple two-tier alerting rule built only from resting-heart-rate deviation, the authors estimate that 63% of cases could have been caught before symptom onset, in real time, from a device already on the wrist. A companion effort from the Scripps DETECT study, run by Giorgio Quer, Eric Topol, and colleagues, enrolled over 30,000 participants and found that combining self-reported symptoms with sensor data discriminated COVID-positive from COVID-negative symptomatic people with an AUC of 0.80, significantly better than symptoms alone.

Now widen the lens from wearables to the purchases themselves, which is the part of my own argument I could least support with someone else’s data until recently. Riccardo Di Clemente and colleagues, working with anonymized credit card transaction sequences from an entire metropolitan population, applied a text-compression technique borrowed from information theory to the order in which people buy things, not just what they buy. The method is worth pausing on, because it is the same move that made language models work. They treat each recurring pair of consecutive purchase types as a “word,” compress each cardholder’s history into the vocabulary of words they actually use, and then measure how similar any two people’s vocabularies are. No categories are imposed in advance. The structure has to fall out of the sequence itself.

What fell out were groups distinctive enough that the authors could name them: Commuter, Household, Young, Hi-Tech, and Dinner-out, alongside a large “Average” group of about 39% with no strong signature. Then came the check that matters. When the authors compared those purchase-derived groups against independent data, age, total expenditure, gender, and the diversity of each person’s mobility and social network, drawn from separate mobile-phone records, the groups lined up with all of it. A shopping history, read as a sequence rather than a snapshot, is already legible as a lifestyle. Health is one part of a lifestyle. It is not a large leap from there to a purchase-sequence signature of the kind my own cart displayed the week I got sick, just averaged over a metropolitan population instead of one household.

These three results, wearables tracking flu at the state level, smartwatches flagging infection before symptoms, and purchase sequences legible as lifestyle, are not the same study asked three times. They are three independent groups, three different data types, and three different outcomes, converging on the same underlying claim: ordinary, ambient, already-collected behavioral data carries a real and extractable physiological signal at the population scale. That is precisely the corpus this article is arguing we have been sitting on.

We Already Know This Works

The idea that you can infer something private and important from ordinary transactions is not speculative. Entire industries are built on it. Credit scoring takes a sanitized, aggregated trail of financial behavior and produces a number, a FICO score, that predicts default risk well enough to underwrite the economy. Marketers segment neighborhoods down to the zip code and the street, estimating purchasing power, life stage, and household composition from data that never required anyone to fill out a form. These businesses did not first build a mechanistic theory of why a person defaults or buys. They found that the aggregate signal predicts the outcome, and they used it. Di Clemente’s lifestyle groups are the same move, applied to health-adjacent behavior instead of credit risk.

The most famous example is almost a parable for what I am describing. More than a decade ago, a large retailer’s statisticians discovered they could identify, from shifts in ordinary purchasing, unscented lotion, certain supplements, a few dozen everyday products, that a customer was very likely pregnant, sometimes before the family had told anyone, a story Charles Duhigg reported in the New York Times Magazine in 2012. No ultrasound. No lab test. No medical record. Just the quiet signature of a changing body expressed through a shopping cart. That story is usually told as a privacy cautionary tale, and the caution is warranted. But look at the underlying fact: a health state of enormous clinical significance was legible, at scale, in data nobody collected for medicine.

Now generalize it, and we no longer have to generalize purely on faith. If a pregnancy is visible in purchasing, and a purchase sequence is visible as a lifestyle, and a lifestyle correlates with age and social network, then the early drift toward metabolic disease, the behavioral collapse that precedes a depressive episode, and the subtle changes in movement and spending that shadow early cognitive decline are exactly the kind of signal this new literature says should be there. We could plausibly read health and wellness at the macro level, at the level of a population, a region, a zip code, a street, the same way credit and marketing already read purchasing power, and the same way Radin's team already reads flu season from a wristband. Not to diagnose one named individual against their will, but to see the trends and trajectories of human health in data we are already sitting on.

So I Tested It on Myself

Before I ask anyone to believe this, I tried to break it on the one subject I fully control: myself. For years I have kept a daily record of my own sleep, movement, heart rate, and mood, alongside my card statement and my complete Amazon history. I pointed one at the other, 63 spending series against 41 health measures across 203 days, and went looking for the signal I had just promised you.

I found no detectable relationship. Of 630 correlations, exactly zero survived proper correction, even though three quarters of them would have looked significant to the standard test. Every one was noise. That failure does not wound the argument, it sharpens it: one person carries almost no statistical power, and I never claimed a single cart is diagnostic. The signal I believe in, and the signal Radin, Mishra, Quer, and Di Clemente all found, lives in the population, not the person, which is also, conveniently, where the privacy risk is lowest.

My Cart Beat My Sensors, and I Beat My Cart

That null result concerns subtle, everyday things, and it holds: at the scale of one person, ordinary spending carried no detectable relationship to my health. But subtle is not the only setting a body has. On March 24, 2026, I woke with the first symptoms of a respiratory infection, my first illness of any kind since 2018. Seven years of nothing, then about two and a half weeks of feeling wretched. An acute event is where ambient data should show itself if it shows anywhere, and this time it did, across four separate records. What surprised me was the order in which they arrived.

Four days after I felt it, on March 28, I ordered four cold remedies. Seven more followed over the next eleven days, eleven orders in twelve days for about $167: Vicks VapoRub, a nighttime vaporizing rub, saline nasal mist twice, Throat Coat and Manuka cough drops, Boiron cold-and-cough syrup, throat tea twice, and finally a steam vaporizer. You could pick that basket out of six months of my spending by eye. My wearable signals, meanwhile, stayed inside their normal range until March 29 and did not peak until April 1, when my respiratory rate hit 4.3 standard deviations above my personal baseline. Symptoms, then cart, then sensor, each step about four days after the last.

I want to be precise about what that ordering means, because the tempting version of this story is the wrong one. The cart did not know first. I did. My own felt symptoms led my purchases by four days, and my purchases led my sensors by another four. What the cart did was record a piece of knowledge no wrist sensor can measure: my own judgment that I felt bad enough to spend money on it. That is not Amazon understanding immunology. It is an ordinary behavioral record capturing subjective state, timestamped and machine-readable, four days before any physiology crossed a threshold. For a model trying to read health from ambient data, that distinction is the entire opportunity. The earliest and cheapest signal in my record was not a measurement of my body. It was a decision I made about my body.

My body did respond, clearly and measurably. This was not a subtle event lost in the noise. Respiratory rate reached 4.3 standard deviations above baseline, heart rate variability fell to about 1.7 standard deviations below it at onset before recovering, and resting heart rate rose alongside both. Across the acute window each of those signals tracked the remedy purchases at about r = 0.6, peaking two to four days after the first order. Everything else stayed quiet. Sleep, sentiment, and activity carried nothing for this event, and I actually slept more while ill, not less.

But nothing predicted it. Heart rate variability, the closest thing I have to a resilience measure, sat near the top of its range going into the illness and dropped only after the infection had taken hold. There was no dip beforehand that could have forecast anything, and nothing physiological crossed a line before I bought the cough syrup. My body confirmed the event. It never anticipated it. That deserves saying plainly, because anticipation is exactly what consumer wearables are usually sold on.

It also did not confirm the event alone. Two devices with nothing in common measured my breathing independently: an Eight Sleep mattress underneath me and an Oura ring on my finger. They share no hardware and no data, and they agreed almost exactly, tracking to within a fraction of a breath per minute across the illness, a correlation of 0.90, both peaking on the same day, April 1. That is genuine cross-device corroboration rather than one sensor echoing itself. The agreement holds for respiratory rate specifically and not for heart rate, where the ring’s overnight low and Apple’s daytime value are simply not the same measurement and should never be treated as one.

Now the part that is least flattering and most useful. Getting to that clean two-line chart meant discovering that the label “Apple Health” had a mattress hiding underneath it. The respiratory signal in that export is overwhelmingly the Eight Sleep, not the Apple Watch, which barely measured breathing here at all; the watch’s real contribution was heart rate variability and daytime resting heart rate. The resting-heart-rate line in the cart figure is the ring’s overnight measurement, chosen because it recorded every single night, including April 2 and April 6, the two nights Apple’s separate daytime metric was never computed at all. Even the mattress has holes: it recorded nothing across a spa trip in mid-March and an overnight in late March, both comfortably before symptoms began, which is why its line is dotted there and why the ring had to cover for it. Three overlapping devices had to be pried apart, and each one’s blind spots filled by another, before one clean event was visible.

That messiness is not a footnote against the argument. It is the argument. A single person’s record is a pile of overlapping, gappy, mislabeled instruments, and no one timeline in it is complete. What rescued this event was not a better sensor. It was a second, independent one that happened to be recording when the first was not. That is the aggregate thesis at the smallest possible scale, two devices instead of one, and it is exactly what Radin’s team was doing on a different order of magnitude when it turned 47,249 Fitbits into a signal cleaner than any single wristband could ever offer. Read one timeline and you spend your time fighting the mess. Read millions and the mess averages out.

A fourth ordinary record caught the same illness through what I ate. My daily calories fell about 25% below the two weeks preceding symptoms, a median of 2,408 against 3,207, and recovered within a week. The sharper marker was not quantity but a single ingredient. Honey, the oldest sore-throat remedy there is, appeared more often and in larger amounts as the illness ran, peaked at four servings on April 3, and stopped entirely once I was better. My rhythm never changed at all: I logged the same dozen or so items a day throughout, in the same eating window. Only the contents shifted.

An earlier version of this section claimed something stronger, and it was wrong in a way worth showing rather than quietly deleting. I had reported that my daily sugar collapsed from a median near 96 grams to about 10, held there for twelve days, and snapped back the day I recovered. It did not. My food log did not reliably estimate sugar in grams until early April: only 3% of pre-illness food items carried a sugar value, against 99% afterward. The collapse was missing data wearing the costume of a finding, and it was the cleanest-looking result I had. I caught it only by going back and asking what the denominator was. Honey servings, read from the meal descriptions themselves rather than from an inferred nutrition estimate, survived that check. Sugar in grams did not, so it is not charted here and no longer appears in this essay.

So four ordinary records, two of them never collected for medicine, all registered one illness, and the earliest of them was a shopping cart. I want to be careful about what that is worth. It is n = 1 and a single episode. My purchase dates are not symptom dates, and my symptom dates are not infection dates. My food log is self-reported. Every threshold here was set after I had already seen the data, so these are descriptions rather than tests. And not one of these records spoke until I had stripped a confound out of it by hand, whether that was a travel gap in a sensor or a missing denominator in a food log. Nothing here generalizes, and that is the point rather than the disclaimer. One person is an anecdote no matter how many instruments you point at him. The signal a health model actually needs does not live in my timeline, or in yours. It lives in the aggregate, in a million of these noisy records laid over each other, which is the one place I cannot look and the platforms, and now a handful of academic teams, already can.

A Necessary Caution

I would be doing this argument a disservice if I only told the flattering half of the population story. In 2009, Jeremy Ginsberg and colleagues at Google showed that the frequency of certain search queries tracked CDC influenza data closely enough to estimate flu activity roughly a day faster than official reporting, a genuinely elegant early proof that ambient digital behavior encodes health signal. Google Flu Trends became the flagship example of exactly the kind of system I am describing here. It also broke. By the 2012-13 season, the system was overestimating peak flu prevalence by roughly a factor of two, and in 2014 David Lazer, Ryan Kennedy, Gary King, and Alessandro Vespignani published a widely cited postmortem, “The Parable of Google Flu”, documenting how algorithm changes, media-driven search behavior, and the absence of any updating against ground truth let the model drift quietly away from reality for years before anyone noticed.

I think that failure and my own null result are teaching the same lesson from opposite directions, and my retracted sugar collapse is the same lesson a third time, at household scale. My 630 correlations produced nothing because one household has no statistical power. Google Flu Trends produced a confident, wrong answer because it had enormous scale but no discipline, no correction for its own drift, no ground truth in the loop. And my sugar result looked like the cleanest finding in my dataset right up until I asked how many of the underlying items actually carried a measurement. The lesson is not that ambient data is unreliable. It is that ambient data is exactly as reliable as the statistical rigor applied to it, and no more. Radin's team validated continuously against CDC numbers. Mishra's and Quer's teams validated against confirmed test results. Any health model built on transactions and wearables has to be held to that same standard, recalibrated against ground truth constantly, or it will eventually become a more sophisticated version of the same mistake, at a scale that matters far more than mine.

Align the Data We Have, Instead of Stitching the Data We Don’t

This reframing dissolves the problem that has stalled every grand vision of a medical foundation model. The usual plan is to somehow link the disparate medical record systems: reconcile a thousand electronic health record formats, negotiate access, harmonize codes, and fight the privacy and liability battles at every hospital boundary. It is a noble effort and it moves at the speed of institutions, which is to say barely.

The alternative is to start from the corpus that is already digital, already standardized, and already aligned within each platform. Card networks, retailers, wearable makers, and the large consumer platforms each hold clean, longitudinal, machine-readable behavioral data at a scale that dwarfs any clinical dataset, and, as of the last few years, so do a growing number of academic research groups. The work is not inventing a new neural architecture; recent proposals for generalist medical foundation models argue that the multimodal architectures behind today’s large language and vision models are already close to sufficient for this purpose. The work is alignment: pointing this corpus, in sanitized and aggregated form, at health and wellness outcomes rather than at ad conversion. Just as credit scoring sanitized financial behavior into a number that is useful without exposing every transaction, a health-and-wellness signal could be derived at the population and regional level without ever assembling a surveillance dossier on a named person.

I do not wave away the privacy questions; they are real and they are hard, and the pregnancy story is precisely why they must be answered before this is built, not after. But notice that working at the macro level, on aggregated and consented and sanitized data, sidesteps the very thing that makes the medical-record approach so painful. You do not need to link one person’s genome to one person’s hospital chart to see that a neighborhood’s metabolic health is deteriorating, or that a cohort on a given intervention is recovering faster. The signal that matters most for public health lives in the aggregate, and the aggregate is exactly where the privacy exposure is lowest, and, per Radin and Di Clemente, exactly where the signal has already been shown to hold up.

Medicine Is an Intervention Problem

All of this points back to what medicine has always actually been. Medicine is not trying to predict protein structures, or even disease labels. It is trying to answer one question, over and over, for one person or one population at a time: given everything we know, what intervention most improves the future?

Framed that way, the inputs are interventions, small molecules, biologics, nutrition, exercise, sleep, stress reduction, behavior change, and the outputs are trajectories, biomarkers and function and healthspan over time. Everything mechanistic in between, kinase signaling, GPCR pharmacology, protein folding, can remain exactly what grammar was to a language model: a latent representation the model builds only if and when it helps predict the outcome. If a structure-like representation helps a physiology model predict who responds to a drug, it will construct one, in whatever form is most useful, whether or not a crystallographer would recognize it. If it does not help, the model owes us nothing.

The central question stops being “what structure does this protein adopt?” and becomes “given everything we already know about how this person lives, what changes their trajectory for the better?”

The Organism, Not the Structure

I think history will be kind to AlphaFold, but for a reason slightly different from how we celebrate it today. Not as the endpoint of computational biology, the final word on what a protein is, but as the first genuinely convincing proof that a biological representation of real value can emerge from a rich dataset and a powerful model, learned rather than hand-engineered. That is a bigger and more durable contribution than solving structure prediction, because it is a proof of concept for a program far larger than proteins.

What has changed since I first made this argument is that the proof of concept is no longer confined to proteins and language models. Wearables tracking flu at the state level, smartwatches catching infection before it is felt, purchase sequences legible as lifestyle, and one household’s shopping cart outrunning its own wrist sensors by four days, all of it points the same direction. The next chapter, I believe, will not be won by folding more proteins or by finally stitching together the world’s medical records. It will be won by whoever first aligns the vast, ignored corpus of everyday human data, the transactions, the wearables, the behavior we already generate, toward health and wellness, held to the same discipline that made Radin’s and Mishra’s results credible and that Google Flu Trends lacked, and lets a sufficiently powerful model treat the body the way we already treat the brain: as a black box worth acting on, not a mechanism we must first fully explain. We do not need to understand every gear to help the organism. We need to stop insisting that we do.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

A toast to Sung-Hou Kim, and the case for putting powerful AI within everyone's reach.

In 1990, at a Gordon Research Conference, a young graduate student stood up to talk about neural networks in front of a room full of crystallographers. He was proud of the work. The reply came back flat: “That is no big deal. We knew that decades ago.” The room went quiet. The graduate student was me, and the silence felt like a door closing.

Then Sung-Hou Kim stepped in. He did not pile on. He widened the conversation, gave the idea room to breathe, and helped a beginner stand back up. I have carried that moment for thirty-five years, because it taught me something that matters more now than ever: expertise can humiliate the next generation, or it can help them keep going.

The Voice I Started to Hear in Myself

When modern AI took off, I caught myself reaching for the same dismissive sentence that once stung me. “We have been doing this for decades.” It is even partly true. It is also the least useful sentence in the room.

The problem is not whether the sentence is correct. The problem is what it does socially. It closes the door at the exact moment the room needs more people: alternative ideas, creative pressure, skeptical pressure, perspectives from outside science, and the naive question that experts have stopped asking. The old reflex is to defend the field. A better move is to widen it. The best move is to teach the tool.

“That’s no big deal” is the sentence that closes the door. “Come join the party” is the one that opens it. Same expertise, two very different rooms.

The Lab Was Already an AI Lesson

Long before any of this, Sung-Hou’s lab was already teaching the lesson. It worked because it was never one lane. It mixed crystallography, biology, chemistry, computation, and method-building, and it gathered people who were willing to collide ideas with one another. Cross the disciplines, ask better questions, build new lanes.

That is exactly what good AI does now. It helps us cross the lanes we used to stay inside. Berkeley remains an ideal of the public institution: world-class excellence paired with a real diversity of disciplines and people. The photographs here are not decoration. They are the proof. The science happened because the people did things together.

A Long Arc

A quick word on why this moment feels different. The arc runs from neural nets in the 1980s, to Berkeley in 1990, to MDL in 1993, to Affymax in 1998, to Eidogen in 2004 and beyond. For decades we built tools with AI. The timeline itself is not the point. The point is the turn at the end of it: now AI builds the tools, and it builds them in plain English.

I Say What I Want Built. The System Builds.

This is the part that still feels like science fiction to me, except that I use it every day. I describe what I want in ordinary language, and the system builds it. Not magic, and not automatic judgment. It is a new interface: English as the front door into software, data, media, memory, and scientific workflows.

The result is a set of real things, not slideware. Toast Our Friend, a place built to honor the people around us instead of promoting ourselves. AI-Dad, which preserves legacy, memory, and a family voice. AI-Steve, a personal AI infrastructure grounded in real content. Image Explorer, which searches decades of photographs by visual similarity. A browser-based CT viewer for patients. Food Is Medicine, which connects meals, ingredients, and health signals into practical guidance. And a nutrition pair, Food Health and Food Health TxD, a diabetes-focused companion, both shipped from natural-language direction. I did not write the plumbing by hand. I said what I wanted built, and the system built it.

A Bionic Amplifier

The fear in the room is always the same: will this replace me? I think that is the wrong question. The better one is: which parts of my work should no longer consume my life?

Picture a bionic arm. It does not decide what matters. It extends what a person can do. AI is the same kind of thing. It amplifies reach, speed, recall, and iteration, while the human still sets the direction and checks the consequences. Some tasks will be automated: summaries, first drafts, formatting, routine searches, data cleanup, scheduling, comparison tables. The human work stays human: priorities, ethics, taste, relationships, experimental judgment, courage, and the ability to recognize when an answer is nonsense. Time, not compute, is the scarcest resource. Spend it on judgment.

Content Makes Kings

Here is the part I care about most. Models are only ever as good as the human content beneath them. A model that memorizes looks brilliant on what it has seen and falls apart on what it has not.

Two pictures make the case. The first is a panel of brains across species. Whales and elephants carry more neurons than we do, yet raw neural capacity never settled who leads. We carry more meaning. What we know, what we notice, what we ask, and what we care enough to preserve: that is the content that makes kings.

The second is a curve from my own 1991 Berkeley thesis. An overfit model nails every training point and then misses the one held-out point that mattered. Garbage in, garbage out, now running at the scale of the entire internet. Content is king. Content is currency. Content makes kings. Fresh, honest, human content is the scarce resource that keeps the whole system from collapsing in on itself. And language, the most powerful interface we have to these tools, belongs to everyone, not only to people who write code.

The Machine Is Learning to Move

Look a little further out and the machine is learning to move. A surprising share of the human cortex is devoted to the hand. Controlling it in real time is one of the hardest problems nature ever solved, and still the hardest thing to build into a robot.

Think of the hardest task in sports. A hundred-mile-per-hour pitch leaves no time to react. A great hitter predicts, pre-commits the body, and fuses that forecast to a violent, exquisitely timed swing. Prediction and execution as a single act. That is precisely what embodied AI has to crack.

The same coupling of cognition and dexterity that made us human is now being assembled in silicon, steel, and carbon fiber. Handled wisely, with guardrails, it does not replace what we are. It extends what we can build. Automating back-breaking, health-eroding labor can be a human act, not only an economic one, if we build the vocational on-ramps and the safety nets on purpose. A choice, not a fate.

The Answer to Fire

We have lit a fire. We always do. The answer to fire was never to extinguish it. The answer was to learn how to carry it, contain it, teach it, and build hearths around it.

Fire warmed us, cooked our food, protected us, and transformed human life. It also burned people and leveled cities. Serious tools require serious culture. Not naive optimism, and not fear either, but responsibility paired with access. Edison, Curie, Einstein, Pauling: brilliance has always existed everywhere. What has not always existed is access to the lane where brilliance can swim. Berkeley is extraordinary, but the next remarkable person may never get near it. AI can help far more people find serious lanes of their own.

The answer to fire was to learn how to carry it.

AI for Everyone

The napkin math

Here is some back-of-the-envelope math on why paying it forward matters. The classic version starts from a single person: if one person teaches three people, and each of them teaches three more, it takes about 21 rounds to reach everyone on Earth, and about 18 rounds to reach everyone in the United States.

But we are not starting from a single pioneer anymore. There are already at least a million people, probably a few million, building software in natural language (i.e., agentically) with tools like Claude Code and Codex. Now assume something far weaker than the classic chain letter: not everyone succeeds, and on average each of us helps just one and a half new people become builders per round while continuing to mentor others. That means the community becomes 2.5 times as large each round. Starting from one million people, that reaches every adult in America in roughly five rounds and every adult on Earth in about ten.

If that still sounds aggressive, cut the average to just half a new builder per person per round - equivalently, one new builder every two rounds. The community then becomes 1.5 times as large each round. America is still reached in roughly fourteen rounds, and the world in about twenty-two.

This is napkin math. It ignores overlap, saturation, and the friction of real adoption. But the conclusion survives any reasonable assumption: we are not decades away from a world where ordinary people program computers in plain language. We are a handful of pay-it-forwards away. So teach one person this month. Better yet, teach two, and ask them to do the same.

Where the Race Is Actually Won

While we are doing arithmetic, it is worth zooming out to the question everyone asks: what about China?

Sebastian Mallaby recently wrote a guest essay in The New York Times after a reporting trip through Beijing, Shanghai, Shenzhen, and Hangzhou. His conclusion: America’s chip export controls have failed to slow China’s AI progress. Every time a U.S. lab ships a cutting-edge model, Chinese rivals rapidly absorb its capabilities and distill them into models of their own. In his words, “the follower has the advantage.”

The news has only reinforced that point. On July 17, 2026, Chinese startup Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter open-weight model that approaches the performance of America’s frontier systems. Reuters described it as the world’s largest open-weight AI model, underscoring how quickly frontier capability is diffusing beyond the handful of companies that first create it.

I believe Mallaby is right, but I would push the argument one step further. Frontier models matter. Frontier models alone do not win. Deployment wins.

There is another reason I think this trend will continue. The internet is not infinite. Recent work by Epoch AI estimates the effective supply of publicly available, high-quality human-generated text at roughly 300 trillion tokens. As frontier models continue to scale, they are rapidly consuming that finite corpus, while an increasing fraction of new online content is itself AI-generated. The future of AI therefore depends less on learning ever more from yesterday’s internet and more on generating genuinely new human knowledge.

Readers may recognize the analogy to the overfitting figure shown earlier from my Berkeley thesis (Figure 2.3 above). As model capacity continues to increase while the supply of fresh human knowledge grows much more slowly, simply making models larger delivers diminishing returns. Researchers have also shown that repeatedly training on synthetic AI-generated content can reduce diversity and quality unless that data is carefully curated. The scarce resource is no longer compute. It is original human content.

That brings us back to the central theme of this essay: Content Makes Kings. Frontier capability is becoming a commodity. Once a breakthrough exists, it spreads astonishingly quickly through open-weight releases, distillation, competitive engineering, and independent innovation. No nation or company builds a lasting advantage on an asset that depreciates that fast. Durable advantage comes from continuously creating new knowledge and turning it into useful applications.

Mallaby arrives at the same conclusion: “the accelerating power of the leading models won’t determine who wins the A.I. race. It’s A.I. deployment that will matter.” Raw capability has to be transformed into products, businesses, institutions, and everyday tools. If frontier intelligence becomes broadly available - and every month suggests it will - the contest shifts to who can convert it into useful software the fastest.

That is exactly where the near-term opportunity sits: application development and solopreneurship. A single person with deep domain expertise can now stand up a real product in days rather than quarters because agentic tools handle much of the engineering that previously required an entire team. The frontier is no longer defined solely by who builds the smartest model, but by who turns intelligence into the most useful applications.

A concrete example. Hannah spent four years of medical school building a study repository: tens of thousands of flashcards, nearly forty thousand medical images, and organized notes spanning every course and board exam. Working together with agentic coding tools, we turned that repository into Study with Hannah, a searchable archive with visual image search, an AI tutor that cites the specific cards behind each answer, and a spaced-repetition drill system. The core system was built in a few hours. A subscription business was live in under a day. Content that took years of disciplined work to create became a deployed product between breakfast and dinner.

Multiply that by every nurse, teacher, contractor, and shop owner who learns to build this way, and the pay-it-forward math above stops being a parlor trick. Countries will not win the deployment race with a handful of frontier labs. They will win it with millions of people who can turn what they know into working software. Teaching the person next to you is not charity. It is industrial policy from the bottom up.

Measure it, monetize it

There is a broader principle hiding in that story. I used to tell teams, “If you can measure it, you can manage it.” For anyone with an entrepreneurial itch, I would now update it: if you can measure something, you can monetize it. Every log you keep, every collection you curate, every process you track produces something other people need, and agentic tools have collapsed the cost of wrapping a real product around it. Glucose readings become a nutrition companion. Study cards become a tutoring service. Session recordings become a searchable library.

The measurement was always the hard part. The business around it used to take a team and a year; now it takes a person and a weekend. So look at what you already measure in your work or your life. Somewhere in it is a subscription, a service, or a tool that someone will pay for.

So I Want to Close Where I Began

At that silent room, and the person who refused to let it stay silent. For those of us who have been in the AI game for years, and for those who arrived last week, the work is the same: pass the tool, not the intimidation.

Use It. Pick one task this week that wastes your time, and let AI help you move through it.

Share It. Sit with one person who is curious or fearful, and let them drive. Translate. Don’t lecture.

Preserve It. Capture the stories, the expertise, the images, the notes, and the methods. Original human content is the fuel that keeps the next generation of AI grounded in reality.

Build It. Look at what you already know, measure, or create. Somewhere inside it is a product, a service, or a business waiting to exist. The cost of building has collapsed. The limiting factor is no longer software engineering. It is the courage to begin.

At that Gordon Research Conference, Sung-Hou modeled how an expert can rescue a beginner instead of crushing him. That was the AI moment in miniature, long before the phrase meant anything. He did not simply defend an idea. He opened a door.

We now have the chance to do the same for millions of other people.

We get to choose which voice we become: the one that says, “That’s no big deal,” or the one that says, “Come join the party.”

The fire is already lit. Build something with it. Teach someone with it. Leave it burning brighter than you found it.

Postscript

The Innovator’s Dilemma Comes for the University

After a recent conversation with Sung-Hou Kim, I want to add one more thread, because it follows directly from “pass the tool, not the intimidation.” If individuals can pass the tool, so can institutions. The open question is whether they will.

We talked about Berkeley, and about higher education in general, through the lens of the innovator’s dilemma. That dilemma is usually told about big companies: a successful incumbent hesitates to embrace a disruptive approach because the new thing threatens the brand and the revenue that the old thing still produces. I think the same trap now sits in front of universities. Many are quietly fearful of new models, not because the models are weak, but because they might cannibalize the very thing that funds the institution.

We have seen a version of this before, with online education. Some universities held back. Others leaned in. Harvard and MIT built edX; Stanford faculty helped seed Coursera; lectures that once lived in one room became available to people anywhere in the world. The schools that embraced it discovered that access and brand are not enemies. They widened their reach, opened subscription and online-degree revenue, and lost none of their prestige. If anything, they gained.

AI raises both the stakes and the opportunity. Every serious university sits on decades of published research, course material, library collections, and the working knowledge of its faculty, postdocs, scientists, and staff. Imagine each institution building a university-specific large language model, trained and grounded on that corpus. Not a static archive of recorded lectures, but a living system you can interrogate: ask a question, follow up, push back, and learn in dialogue with the accumulated knowledge of the place. Offer it as subscription-based access, and you extend the university’s mission to anyone in the world with curiosity and a connection, without diluting the degree or the brand.

Institutions like Berkeley have an extraordinary opportunity to expand their global impact by transforming decades of published research, educational content, and faculty expertise into AI-accessible knowledge systems, rather than simply protecting the traditional model.

Berkeley, of all places, should recognize this moment. This is a campus that has carried civilization-scale fire before. In Room 307 of Gilman Hall, Seaborg and his colleagues identified plutonium in 1941; Ernest Lawrence’s cyclotron sat just up the hill; and J. Robert Oppenheimer, a Berkeley physics professor, convened the 1942 summer conference here that became the theoretical seed of Los Alamos. The roots of the nuclear age run straight through this campus. Berkeley learned, in the hardest way there is, what it means to carry a powerful fire responsibly. It is exactly the kind of institution that could lead the pack in carrying the next one.

The framing I keep returning to is not about protecting the old model. It is about expanding the mission: turning published research, educational content, and faculty expertise into AI-accessible knowledge systems that reach far beyond the lecture hall. That is the opposite of cannibalization. It is multiplication.

This thread deserves a fuller piece on universities in the AI era posted in June - Universities Built the Bomb. Now Build This. Consider this the seed. The same question that now faces each of us soon faces the institution: not whether to carry the fire, but whether to share it.

Bringing s photo to life:

HOW IT WAS MADE (technical recipe for reproducibility):

Model: Runway Gen-4 Turbo (image-to-video), via the Runway developer API.
- Model: gen4_turbo, ratio: 1280:720, duration: 5-10s
- Image sent as base64 data URI
- Cost: ~25 credits (~$0.25) per 5s clip

Pipeline:
1. Start from the cleanest, highest-resolution source image (rotated upright)
2. Run runway_animate.py (below) with a carefully crafted stand-up prompt
3. Clip the output with ffmpeg to the clean window (before quality degrades)

#!/usr/bin/env python3
"""Single, careful Runway Gen-4 Turbo image-to-video run.

Submits ONE image_to_video task (5s), polls, downloads the result. Designed to
spend the minimum: one completed gen4_turbo 5s clip ~= 25 credits (~$0.25).
"""
import os, sys, time, base64, requests
from pathlib import Path
from dotenv import load_dotenv

WORKSPACE = Path(__file__).parent
ENV = Path("XXXXX.env")
load_dotenv(ENV, override=True)

API = "https://api.dev.runwayml.com/v1"
VERSION = "2024-11-06"
KEY = os.environ.get("RUNWAY_API_KEY", "").strip()

IMAGE = WORKSPACE / "attachments" / "gordon_v2.png"   # upright, higher-quality source
OUT = WORKSPACE / "gordon_runway_v2.mp4"
MODEL = "gen4_turbo"          # ~5 credits/sec
DURATION = 5                  # 5s: less time to drift into turning their backs
RATIO = "1280:720"

PROMPT = (
    "The people in this group photograph come to life. Everyone who is seated stands "
    "up onto their feet to a normal standing position, joining those already "
    "standing, so the whole group ends up standing. As they stand they keep their "
    "feet flat on the ground and stand normally; no one extends, kicks, lifts, "
    "raises, or stretches out their legs. Once standing, everyone keeps facing toward "
    "the camera and only turns their head and shoulders slightly to the left and "
    "right to talk and smile with the people next to them. No one turns around, no "
    "one shows their back to the camera, no one walks away, leans far over, or bends "
    "down; they all stay facing generally forward, interacting in place. Everyone has "
    "natural, realistic human proportions and normal height, not stretched or too "
    "tall, and everyone stays sharp and in focus with no blurring, smearing, or "
    "melting. The camera stays steady and wide, keeping the whole group in frame, no "
    "strong zoom. Natural, realistic, documentary look."
)

def main():
    if not KEY:
        print("ERROR: RUNWAY_API_KEY not set"); return 1
    if not IMAGE.exists():
        print(f"ERROR: image not found: {IMAGE}"); return 1

    b64 = base64.b64encode(IMAGE.read_bytes()).decode()
    data_uri = f"data:image/png;base64,{b64}"
    headers = {
        "Authorization": f"Bearer {KEY}",
        "X-Runway-Version": VERSION,
        "Content-Type": "application/json",
    }
    body = {
        "promptImage": data_uri,
        "promptText": PROMPT,
        "model": MODEL,
        "ratio": RATIO,
        "duration": DURATION,
    }
    print(f"Submitting {MODEL} {DURATION}s (ratio {RATIO}) ...")
    r = requests.post(f"{API}/image_to_video", headers=headers, json=body, timeout=60)
    if r.status_code >= 300:
        print(f"SUBMIT FAILED [{r.status_code}]: {r.text[:500]}")
        return 1
    task_id = r.json().get("id")
    print(f"task id: {task_id}")
    if not task_id:
        print("No task id:", r.text[:300]); return 1

    deadline = time.time() + 420
    n = 0
    while time.time() < deadline:
        n += 1
        time.sleep(10)
        pr = requests.get(f"{API}/tasks/{task_id}", headers=headers, timeout=30)
        if pr.status_code >= 300:
            print(f"poll error [{pr.status_code}]: {pr.text[:300]}"); continue
        d = pr.json()
        st = d.get("status")
        print(f"[{n:02d}] status: {st}")
        if st == "SUCCEEDED":
            out = d.get("output") or []
            url = out[0] if isinstance(out, list) and out else (out if isinstance(out, str) else None)
            if not url:
                print("SUCCEEDED but no output url:", str(d)[:400]); return 1
            print(f"downloading: {url[:90]}...")
            v = requests.get(url, timeout=120); v.raise_for_status()
            OUT.write_bytes(v.content)
            print(f"SAVED {OUT.name} ({OUT.stat().st_size/1_048_576:.1f} MB)")
            return 0
        if st in ("FAILED", "CANCELED"):
            print("TASK", st, "-", str(d.get("failure") or d)[:400]); return 1
    print("timed out"); return 1

if __name__ == "__main__":
    sys.exit(main())

Further reading

• Content Makes Kings

• The Fire This Time

• The Rise of the Machine

• Universities Built the Bomb. Now Build This

• Study with Hannah

• Projects

This essay accompanies a short talk, AI for All. Watch it at the top, or at

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

A few days ago I was on the phone with a family member, and the conversation drifted, the way these conversations do, to an old video tape. My older daughter was two and a half. She was on the floor with a wooden puzzle of the United States, and she was picking up each state and setting it down in its exact spot without hesitating. Nevada, there. Colorado, there. If you asked her the capital, she would tell you, in the small serious voice kids use when they already know they are right.

I could not do that today. I am not entirely sure I could do it the day I am writing this sentence.

That is not really a story about a smart toddler, although she was one. It is a story about how much a very young brain can hold, briefly and completely, and how much of that same knowledge quietly falls away as the brain grows into a life that no longer needs state capitals on demand. What she knew at two and a half did not disappear because it was unimportant. It disappeared because nobody needed her to keep carrying it forward. The tape, though, still carries it. Thirty-some years later, I can press play and get it back. That gap, between what a mind lets go of and what a piece of media holds onto, is the whole reason I am writing this.

What She Knew at Two and a Half

If I asked my older daughter today where Nevada is, she would get it right, probably without thinking. Ask her the capital, and I suspect there would be a pause, maybe a guess, maybe a shrug. That gap between geography and trivia is a pretty good stand-in for how memory actually works across a life: the shape of a thing survives long after the specific detail does not. We keep the map. We lose the footnotes.

My younger daughter, Hannah, would probably handle both questions without hesitation today (she even knows all the countries and their locations), but you already heard a bit about Hannah - Study with Hannah, about the sixty-two thousand study cards she built for herself in medical school. Different puzzle, same posture at the table, the same total absorption in getting something exactly right. What has changed between the toddler and the med student is not a child’s capacity to learn. What has changed is what we bothered to keep a record of, and whether that record is still readable when we go looking for it.

Berkeley, DEC, and the Machines That Wouldn’t Talk to Each Other

I have been thinking about this problem longer than I have had children. In the late 1980s I was in graduate school at Berkeley, building some of the earliest neural network methods anyone was using for protein structure prediction, work that was probably thirty years ahead of the field’s patience for it. None of that matters here except for one detail: the machines.

I started on old DEC systems and moved, over a few years, onto Silicon Graphics workstations. The two families of machines were different enough that moving anything between them (code, results, the tapes I had recorded my work onto) took real, deliberate effort. Nobody handed me a converter. I had to notice, ahead of time, that the platform I was standing on would not be the platform I would be standing on in five years, and that if I did not move my own work forward on my own schedule, I would eventually be locked out of it on somebody else’s schedule instead.

That was not foresight so much as an early, uncomfortable lesson, learned by hand, tape by tape, in a lab that had no particular reason to warn me about it. I moved source code and results off those machines more than once, onto whatever came next, well before I had hard evidence the old hardware was actually going away. It turned out to be the right call every single time. Ever since, I have treated “will I be able to read this in ten years” as a real engineering requirement for anything I actually care about keeping, not a nice-to-have I get around to eventually.

The Alexandria Fire, on a Budget

Here is the part almost everyone has lived through without ever quite naming it. Betamax. Video CDs. MiniDV. Blu-ray, which felt permanent for about a decade. Every one of those formats held real memories: someone’s wedding, someone’s Christmas morning, someone’s toddler doing a state puzzle. And every one of those formats has quietly become harder to play than the memory itself deserves. You can find a working deck in the Smithsonian Institute. You can probably find one on eBay if you are patient and willing to gamble on a stranger’s storage unit. You are not finding one on Amazon, and you are certainly not finding one at Walmart.

That is a small, slow-motion version of the Library of Alexandria. Nothing burned. The scrolls are all technically still there, on tape, on disc, in a drawer in someone’s garage. What is gone is the ability to read them, which for practical purposes is the same thing as gone. It just happens quietly, one format transition at a time, spread out over decades instead of one bad afternoon in Egypt. Nobody registers it as loss, because nothing dramatic ever happens. The device you need just gets a little harder to find every year, until one day it is a museum piece, and by then the moment you would have wanted back is thirty years further from anyone’s reach than it ever needed to be.

In Silico, Not Just the Highlight Reel

Most of what we do save, we save selectively. We keep the wedding, not the argument in the car on the way to the wedding. We keep the graduation, not the ordinary Tuesday that actually made up most of a relationship. I understand the instinct. Nobody wants to relive the difficult parts on purpose, and curated memory has always been a kindness we do for ourselves.

But I think we are underinvesting in the unremarkable footage: the stuff that is not staged, the stuff that just happens to be running while a life is actually being lived. Not for anyone’s entertainment, and not to build an inventory for later monetization. I mean something closer to memorializing as much of a person’s real life as is reasonably possible, quietly and unobtrusively, without turning every family dinner into a production. Call it in silico memorialization, a phrase I keep coming back to: preserving people’s lives in digital form, ahead of a need you cannot yet name, rather than after you can. The specific detail worth keeping is rarely the one you would have predicted in advance. Nobody sets out to film “the last time we will all be in this kitchen together.” You only find out which recordings mattered after the fact, which is exactly the argument for keeping more of them than you think you need right now.

Built to Move, Not to Stay

This is also why I have put real effort into making sure the systems I build myself do not repeat the DEC-to-Silicon-Graphics problem for anyone else. Everything I put in the cloud now, I put there assuming the cloud is this decade’s format, not a permanent home. It needs to be able to leave cleanly whenever the next platform shows up, the same way my old research needed to leave a machine before that machine stopped being useful to anyone.

That thinking is a big part of why I built out ToastOurFriend, a place where you can create a circle around someone you want to celebrate (a friend, a mentor, a neighbor) and fill it with stories, photos, and voice memories while that person is still around to read them. I wrote more about the origin of that project, and why celebration should happen before someone is gone rather than after. The content in a circle belongs to the people in it, not to the platform. If ToastOurFriend is not the right home for it in five years, moving that circle somewhere else should be simple. I built it that way on purpose, because I have already lived through what happens when a format outlives its own hardware, and I would rather not do that to anyone’s memories a second time.

Where Things Go

There is a reason the puzzle in that old video has stuck with me all week, beyond the obvious cuteness of a two and a half year old rattling off capitals. A puzzle only works because every piece has exactly one place it belongs. My daughter knew, instantly and correctly, where each one went. That is a small, almost embarrassingly literal model for the actual argument here: the memories worth keeping need to keep landing in the right place, generation of media after generation of media, or they simply stop existing anywhere we can reach them.

Nobody does this migration for you automatically. The photo albums that got scanned onto a CD did not get themselves scanned again when CD drives started disappearing from laptops. Someone had to notice, ahead of the format’s expiration date, and move the content forward on purpose. That stopped being a technology problem a long time ago. It is a decision every family has to make, over and over, about what they are willing to carry forward and what they are willing to let the next quiet Alexandria fire take.

So here is the actual ask, buried under all the DEC machines and Betamax decks: go find the tape, the disc, the drive you have not opened in a decade, and move whatever is on it somewhere it can survive the next format change. Do it before you need a device you can only find at the Smithsonian. And if there is someone in your life worth celebrating while they can still see it, do not wait for the version of this essay that shows up at a memorial. Say it, film it, post it, now, while it can still land exactly where it is supposed to go.

Related reading: Study with Hannah· Toast Our Friend · ToastOurFriend.com

Related listening, with the daughters you just read about: Father-Daughter Face-Off, part one, with Lili · Father-Daughter Face-Off, part two, with Hannah: Medical Training Overseas · Homeownership, Healthcare, and AI, with Lili

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple mix videos - these are from the recent July 3 gathering. Green River (Dan, Ron, Tim, Andrew) - “let me remember things I love.” And Ventura Highway (Ron, Andrew, Dan, TimL)

I was home for a few days last summer when I found my daughter at the kitchen table surrounded by study cards. Not a stack. A landscape. She was drilling herself on nephron physiology at eleven at night, the way she has been drilling herself since she was seventeen and somehow already had two associate’s degrees to her name. What stopped me wasn’t the studying. Studying is just what Hannah does. What stopped me was a number she mentioned almost in passing, the way you’d mention the weather: sixty-two thousand.

Two degrees before she could vote

Hannah has been collecting credentials since before most of her peers had a driver’s license. While still in high school, she took classes at MiraCosta Community College and walked across the stage with two associate’s degrees, one in French and one in psychology, the same year she got her high school diploma. She went on to Cal State University San Marcos for a bachelor’s in psychology, loading up on pre-med coursework along the way.

Then came the part of the story every pre-med family knows too well: the U.S. medical school application gauntlet, a process that can run a year and a half to two years from application to a seat in a lecture hall, most of it spent waiting. Hannah didn’t wait. While her U.S. applications were still crawling through committees, she applied to Semmelweis University in Budapest on a whim, on a friend’s suggestion, with rolling admissions. She heard back in about a month. By the time most of her original applicant pool was scheduling interviews, she had already finished her first year of medical school in Hungary.

She’s now in her fourth of six years at Semmelweis, a very rigorous program - one in which she earned competition exemptions in physiology, biochemistry, pathology, pathophysiology, genetics, immunology, and cardiology. Somewhere in the middle of all that, she sat for United States Medical Licensing Examination (USMLE) Step 1, the licensing exam that lets her eventually practice in the United States, and passed it before her fourth year even started, well ahead of her own class and ahead of most students at her stage anywhere. She’s now prepping for USMLE Step 2, early again, because apparently that’s just the cruising speed she operates at. And that head start matters more than it used to: since Step 1 went pass/fail in early 2022, Step 2 is now the scored exam that residency programs actually see, which makes both her early start and the mountain of Step 2 material she’s building worth that much more.

The number that stopped me

Here’s what I didn’t know until that night at the kitchen table: while doing all of this, Hannah had been quietly building one of the more remarkable personal knowledge bases I’ve ever seen. Her digital study card collection, the record of everything she’s had to learn to get this far, had grown past sixty-two thousand cards. Closing in on sixty-three thousand. Spanning both her Step 1 prep and her Step 2 prep, organized into distinct collections, backed by tens of thousands of her own annotated medical images: histology slides, radiographs, diagrams she’d built or curated herself.

I’ve spent forty years building databases for a living. I know what a well-curated, tens-of-thousands-of-records knowledge base is worth, and I know how rare it is for one person to build something that size, alone, as a side effect of just trying to survive their own coursework. I was floored. Then she told me the second half of the story, and floored turned into something closer to stunned.

Classmates had been using her cards. For free. She’d been handing over her personal study platform, the fruit of years of nights like the one I’d just watched, to anyone in her cohort who asked. No compensation, no credit tracking, nothing. People were passing her exams on the back of material she’d built, and she hadn’t seen a cent, a coffee, or even much of a thank-you for it. “What do you mean, giving it away?” I asked her. “There are people benefiting from your hard work who haven’t paid you anything for it. That makes no sense.”

A one-day project, a mini Khan Academy

That conversation turned into a project. Within about a day, I’d built what’s now StudyWithHannah.com, a searchable home for the archive: keyword search across her notes and cards, semantic image search so you can find visually similar histology slides instead of scrolling forever, and a conversational AI tutor that can quiz you, hint at answers, and ask Socratic follow-up questions pulled straight from her own material instead of generic textbook filler.

Call it a mini Khan Academy, built for one student who happened to also want to help herself study better while she was at it. The idea from day one was practical, not grand: if there’s already a real audience of med students who want this material, and Hannah is the one who built it, why shouldn’t access to it help pay for something while she’s living and studying in Budapest? Not a windfall. A better dinner some nights. A cup of coffee that isn’t the school cafeteria’s version of one.

The AI tutor retrieves relevant cards, cites them, and asks focused questions instead of just handing you the answer.

The public site introduces the project. Reader access is still being rolled out carefully, curated around Hannah’s own notes and material rather than anything borrowed. But the archive itself is real, it’s growing every week she’s in school, and it already runs to four major collections and counting.

Someone’s been monetizing her since she was four

Here’s the part of this story that made the whole thing click into place for me. Search “Hannah Muskal” on Google Images and, alongside the search results for her own sites, you’ll find a photo of her as a small child, making a Halloween decoration at the Cole Library in Carlsbad, taken for a local newspaper feature back in 2004. It’s a sweet, ordinary photo. It’s also, right now, sitting behind a stock-photo paywall on Alamy, watermarked, labeled “licensable,” available for a stranger to purchase the rights to a picture of my four-year-old daughter that neither of us has ever seen a dollar from.

Nobody asked Hannah. Nobody asked me. A newspaper photographer took a picture of a kid at a library event two decades ago, and somewhere between then and now it became someone else’s inventory. That’s not a scandal. It’s just how content quietly changes hands when nobody’s watching. But it’s a hell of an illustration of the point I want to make: other people have been finding ways to monetize Hannah’s existence, her image, her likeness, for years, without her involvement or consent. So why exactly should it be strange, or greedy, or “unprofessional,” for her to be the one who benefits from something she actually built on purpose, out of her own effort, for once?

And this is bigger than one stock photo. We’re living through a moment where everything a person produces, every photo, every post, every carefully written note, tends to end up as somebody’s inventory or somebody’s training data, usually without the producer ever seeing the transaction. I’ve spent my career around datasets, and I’ll tell you what Hannah’s archive actually is in 2026 terms: a clean, hand-curated, expert-annotated corpus of tens of thousands of records and images, exactly the kind of thing entire companies get built on. The only difference between what Alamy did with her likeness and what she’s doing now is authorship and consent. She made this one. She gets to decide what happens to it. Most students have no idea they’re sitting on the same kind of asset, because nobody has ever told them their studying produces anything other than a grade.

Education is already a business. Ask any bursar.

This is the part I keep coming back to. We treat “education is a business” as an accusation, something said with a sneer about administrators and endowments. But it’s just a description. Tuition is a price. Textbooks are a price. Test-prep courses, private tutors, application consultants, all of it is a market, and schools and vendors participate in that market constantly and without apology. Students are the customers in that transaction, over and over, for years, often going into debt to be one.

What students almost never get to be is the seller. All that studying, all that synthesis, all those hours spent turning a professor’s rambling lecture into something clean and memorizable, that’s real intellectual labor, and it disappears into the ether the moment the final exam is over, or it gets handed out for free to whoever asks nicely. I don’t think that’s some kind of moral high ground. I think it’s a market inefficiency nobody’s bothered to fix, because nobody stopped to notice that a 26-year-old with sixty-two thousand study cards is sitting on an asset.

StudyWithHannah.com is not a grand theory of educational reform. It’s one daughter, one dataset, one day of building software, and one modest question: if schools can monetize the process of educating students, why can’t students find their own entrepreneurial angle on their own educational journey? This is just a first stab at that question. I suspect it won’t be the last one I take.

She’s in better company than she knows

Here’s the thing: medicine, of all fields, has already run this experiment, more than once. First Aid for the USMLE, the book every med student on the planet treats as the bible of Step 1 prep, began in the early 1990s as a “student to student guide,” written by medical students who got tired of bad review materials and turned their own notes into something better. It’s been an annual franchise for over three decades. Sketchy, the visual-mnemonic platform half of Hannah’s generation studies from, was started by medical students drawing their own memory aids; it eventually took a thirty-million-dollar investment. Even the flash card world itself has ecosystems where students collectively decide what belongs on the front and back of every card, and some of those efforts have quietly grown from dorm-room collaborations into commercial products.

So the precedent isn’t just there, it’s the backbone of how med students actually study. The strange part isn’t that a student’s study materials could be worth something. The strange part is that each new generation of students still doesn’t think of themselves as the ones allowed to benefit from making them.

I’ll be honest about one nuance, because it’s the first thing a learning scientist would raise: a big part of a study card’s value goes to the person who writes it. The act of wrestling a lecture into a clean question and answer is itself the studying. That’s true, and it doesn’t undercut the idea, it explains the trade. When you use cards built by someone who already passed the exam you’re facing, you’re not buying her learning, you’re buying her distillation: thousands of hours of deciding what matters, compressed into something you can drill at eleven at night. That’s the same trade every First Aid buyer has made since 1992. It’s just that this time, the student who did the distilling is the one at the counter.

A Gen-X dad, a Gen-Z daughter, and a few open questions

Hannah and I have been doing a running dialogue for a while now, a sort of Gen-X-meets-Gen-Z back and forth on the podcast, comparing notes on how different her path through school has looked from mine. We recorded a two-part episode about her medical training overseas, including a segment where I showed her old childhood photos (the same kind of photo, incidentally, that a stranger now has on file at Alamy) and asked her, more or less, “med school today?” You can find that conversation here: Father-Daughter Face-Off (part two): Medical Training Overseas. It’s a good companion piece to this one if you want the fuller picture of how she got here, astrocytes and all.

Visit, share, and ask your own question

I don’t know if this idea scales past one very determined med student. That’s fine. Most good ideas start narrow. What I do know is that Hannah spent years building something valuable, gave it away by default because nobody had suggested she do otherwise, and now has a small, honest way to let that work support her while she finishes the six-year climb to becoming a doctor. If you know a student sitting on their own pile of hard-won notes, summaries, or study systems, ask them the same question I asked Hannah: why exactly are you giving that away?

Go take a look at StudyWithHannah.com. Request access if you’re a student who could use sixty-two thousand study cards written by someone who’s already passed the exam you’re studying for. And if this story is worth passing along, pass it along. That’s the whole point of writing it down.

Related reading: Father-Daughter Face-Off (part two): Medical Training Overseas · Study with Hannah: Medical School Study Cards · Steven Muskal Projects

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple mix videos, here are timely samples involving Hannah who was home for a few weeks. The recent July 3rd (2026) ‘Mix and Match’ session at Ocean Ranch Studios. Hannah with Maya on  What’s Up / 4 Non Blondes Cover (though Hannah is a true blonde!). AlanW on Bass, Giovanni on guitar, Dan on acoustic.

And a rough run of Let it Be — RickL, Hannah, Andrew, David, Dan. As usual, a single run through with people never having played together before on a song they learned on as we went. That’s how we roll!

The Conversation That Started This

Earlier this week I zoomed with my graduate research advisor and mentor Sung-Hou Kim (age 88). We discussed many things, including where things are heading for research universities. We have become reacquainted after 35 years and have had a couple calls prior to a planned reunion on the Berkeley campus a few weeks from now (August 1, 2026). Sung-Hou groomed many thought leaders over the years. I am especially looking forward to meet for the first time during this star-studded alumni gathering the legendary Harvard geneticist George Church, famed structural biologist Joel Sussman, and a whole cast of thought leaders who have worked in the Kim lab over the years. Sung-Hou and I were discussing AI and agreed: people who embrace AI will increasingly outperform those who refuse to use it. The question that followed naturally was whether the same dynamic applies to institutions. If individual experts can be passed by the tool, so can universities. The open question is whether they will see it coming in time to do something about it.

I have been thinking about that conversation ever since. This essay is where I landed.

The Innovator’s Dilemma

Clayton Christensen described the trap in 1997. A successful organization encounters a disruptive new technology and hesitates to adopt it, because the new thing threatens the very business model that has made the old thing successful. The hesitation is not irrational. It is rational self-preservation. The new model looks small and unproven. The existing model is still generating revenue. From the inside, killing the golden goose feels reckless.

Universities fit this description almost perfectly. The brand of a Harvard, a Stanford, a UC Berkeley is tied directly to scarcity: limited seats, selective admissions, a residential experience that costs 5-to-6 figures per year. The implicit promise is that proximity to world-class faculty and peers is something you cannot get anywhere else. Any model that genuinely democratizes access to that knowledge runs the risk, from the institution’s perspective, of diluting exactly what makes the degree valuable.

That logic is understandable. It is also, I believe, wrong. And we have already run the experiment that demonstrates it.

The Online-Education Precedent

In 2012, Harvard and MIT launched edX. That same year, Andrew Ng and Daphne Koller, both Stanford faculty, co-founded Coursera. The prediction from those who defended the traditional model was that putting lectures online would erode the value of an on-campus degree. That prediction was wrong.

edX grew to tens of millions of learners across more than 160 countries without any perceptible reduction in the prestige of a Harvard diploma. Coursera raised over a billion dollars in venture funding and went public in 2021. Georgia Tech launched an online master’s degree in computer science for under $7,000, now one of the largest graduate CS programs in the world, and its residential programs have lost none of their standing. The schools that leaned into online education discovered that access and prestige are not enemies. They widened their reach, opened new revenue streams, and their brands emerged stronger, not weaker.

The lesson is not that online education is identical to the residential experience. It is not. The lesson is that the university’s accumulated knowledge is far more durable and far more transferable than the institution’s defenders feared.

The Concrete Proposal

The next logical step goes further than streamed lectures. Every serious research university sits on something extraordinary: decades of published research, course syllabi, lecture recordings, library collections running to millions of volumes, and the accumulated working knowledge of its faculty, postdocs, scientists, and staff. That corpus is alive and growing. It is not static, and it is not generic. It is exactly the kind of high-quality, specialized content that makes a powerful language model genuinely useful rather than confidently wrong.

Imagine a university-specific large language model, trained and retrieval-grounded on that corpus. Not a chatbot that has absorbed the entire internet and then hallucinated its way through the gaps. A focused, institution-specific system that knows Berkeley chemistry the way Berkeley chemists know it, that can trace every answer back to a specific paper, a specific lab, a specific course. Offered as subscription access, it extends the university’s mission to anyone in the world with curiosity and a connection, without a single diploma being devalued.

The interface matters as much as the corpus. Static video lectures, however beautifully produced, put the learner in the role of spectator. Interactive question-and-answer dialogue puts them in the role of student. The difference between watching a surgery and being trained to perform one. At its best, the AI interface is Socratic: it meets the learner where they are, follows their questions, asks what they already understand, and calibrates accordingly.

The Hard Parts, Handled Honestly

I do not want to oversell this. The engineering and the policy challenges are both real.

Accuracy and hallucination. Language models are statistically fluent and factually unreliable when they venture outside their verified training data. The standard mitigation is retrieval-augmented generation: the system retrieves passages from the institutional corpus before answering, and grounds its response in those passages. It is not a complete solution, but it shifts the failure rate in the right direction and, critically, makes errors auditable. A citation back to a specific document is something a user can check and an institution can stand behind.

There is a deeper point here that deserves emphasis. Many of the factual failures we associate with today’s large language models begin with the quality of the knowledge they consume. Most frontier models have been trained on vast swaths of the open internet, a corpus that includes misinformation, contradictions, outdated claims, low-quality content, and increasingly, text generated by other language models. Their statistical understanding inevitably reflects both the strengths and the weaknesses of that data.

University and college collections are fundamentally different. Published research has survived peer review and editorial scrutiny. Course materials have been developed and refined over years of teaching and student feedback. Library collections have been curated by professional librarians, while faculty lectures and departmental resources reflect the collective expertise of recognized scholars. No corpus is perfect, but it is vastly more authoritative, internally consistent, and accountable than the open web.

A language model grounded in this corpus does more than retrieve better answers. It reasons from higher-quality evidence. Combined with retrieval-augmented generation and citation back to original sources, this creates a structural advantage: factual claims become verifiable, errors become auditable, and the model is far less likely to drift into unsupported assertions. Rather than trying to correct hallucinations after they occur, the system reduces them by grounding its responses in one of the highest-quality knowledge repositories humanity has assembled.

Citation and provenance. Every answer should link to its source: the paper, the textbook chapter, the lecture segment. Not only because academic culture demands attribution, but because it lets the learner verify the claim and go deeper. This is achievable with current technology. It requires disciplined indexing of the corpus and careful system design, but it is not a research problem; it is an engineering one.

Licensing of scholarly content. This is the thorniest problem. Universities publish research through journals that may retain copyright. Course materials may carry layered IP ownership across instructors, departments, and the institution itself. Open-access publishing helps, but it does not solve the problem entirely. A serious deployment requires legal review, negotiated rights, and in some cases revenue sharing with rights holders. This is an argument for starting deliberately, not an argument against starting.

Privacy. Student records, researcher correspondence, health data, and unpublished work have no place in any training corpus or retrieval index. The boundary between what is public-facing and what is not must be designed in from the beginning, with the same rigor a hospital applies to patient data. This is non-negotiable, and it is also clearly achievable.

Faculty incentives, credit, and compensation. A professor’s lectures, papers, and course materials represent years of concentrated effort. Any system that ingests that work should credit it and, in a subscription revenue model, share the returns with the faculty who generated it. This is both fair and strategically necessary: faculty buy-in determines the quality of the corpus, and buy-in follows fair treatment. The revenue-sharing framework already exists in other knowledge industries; universities are capable of inventing their own version.

Why Berkeley Should Lead

If there is one institution that should recognize this moment for what it is, it is UC Berkeley.

In Room 307 of Gilman Hall, during the winter of 1940 to 1941, a young chemist named Glenn Seaborg and his colleagues bombarded uranium-238 with deuterons and, on December 14, 1940, first produced plutonium-239. By February 1941 they had chemically separated and conclusively identified element 94, the first isolation of plutonium in history. That room is now a National Historic Landmark, recognized not because the chemistry was elegant but because the discovery changed what the world was capable of.

The material that Seaborg identified came from the cyclotron that Ernest Orlando Lawrence had built just up the hill. Lawrence started in 1930 with a machine four and a half inches across, housed in a wooden shack on the Berkeley campus. By 1939 he had scaled to a 60-inch instrument powerful enough to earn him the Nobel Prize in Physics. The institution he founded became Lawrence Berkeley National Laboratory, a national resource that has operated continuously for more than eighty years and has produced thirteen Nobel laureates among its researchers.

And it was at Berkeley that J. Robert Oppenheimer spent fourteen years as a physics professor, from 1929 to 1943, building one of the great theoretical physics programs in the world. The summer conference he organized at Berkeley in 1942, gathering some of the century’s most consequential minds in physics, became the theoretical seed of Los Alamos. Berkeley did not merely participate in the next chapter of history. It convened it.

I recently came across the passage below as the opening epigraph of The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence. Von Neumann was describing the dawn of the nuclear age in 1945. Replace the phrase “energy source” with “intelligence source,” and the passage could have been written yesterday.

In each case, the pattern was the same. The institution did not stand at a distance and observe a powerful new capability emerging. It organized around it, funded it, gathered the talent, and built infrastructure that outlasted any individual contribution. The cyclotron became the national laboratory. The chemistry became the periodic table extended. The physics conversations became the project that ended a world war.

Berkeley knows, in its institutional memory, what it means to carry powerful fire. The question is whether it and the universities that share its ambitions will recognize the current moment as belonging to the same tradition.

A Call to Action

I want to speak directly to university leadership for a moment, because the frame matters.

The wrong frame is: how do we protect what we have? That frame leads to incremental gestures, a committee here, a pilot program there, a decision to wait until the landscape clarifies. The landscape is not going to clarify in the direction of patience. The institutions that build now will define what university knowledge systems look like for the next generation. The institutions that wait will buy access to systems that someone else built.

The right frame is: what can we build that no one else can build? No one else has Berkeley’s corpus. No one else has the combination of published research, teaching materials, library depth, and faculty expertise that a serious research university has accumulated over a century of operation. That corpus is the moat. It is not the name on the diploma; it is the accumulated knowledge behind it. The question is whether to keep that knowledge locked inside the campus walls or to let it reach anyone in the world who wants to learn from it.

I am not arguing for giving everything away for free. Subscription models work. Online professional certificates work. Revenue sharing with faculty works. The university does not have to choose between mission and margin. The path is to expand the mission, not to surrender it.

Institutions like Berkeley have a singular opportunity right now: to transform decades of published research, educational content, and faculty expertise into AI-accessible knowledge systems that reach far beyond the lecture hall. That is not cannibalization. That is multiplication.

In a piece I will publish soon - AI for Everyone (sneak peek below), I argue that the answer to fire was never to extinguish it. The answer was to learn how to carry it, to build hearths around it, to teach people how to use it safely and well. That same argument applies here, and it applies with urgency.

The fire is already lit. Berkeley helped light it. The question now is whether the university will help carry it forward, or stand at the gate and watch it pass by.

For a mix video, this recently surfaced. Ten years ago, Blue Vinyl had an early morning segment on KUSI. Juanita (vocals), Maya (vocals), Marklyn (guitar/vocals), Don (keys), Steve (bass), Alex (sax), and me on the old e-set. Total blast from the past!

Seems oddly appropriate as we approach the dawn of a new era.

This was one of those rare times a band arrived and set up before dawn. Crazy early. I was on the road to set up a little after 4:30 AM…

Sources and further reading

• Glenn Seaborg on Gilman Hall Room 307, Lawrence Berkeley National Laboratory

• Gilman Hall Room 307, National Park Service

• History of Lawrence Berkeley National Laboratory

• Oppenheimer’s Berkeley Years, Berkeley News

• The Innovator’s Dilemma, Clayton Christensen (1997)

• About edX

• About Coursera

• Georgia Tech OMSCS

• UC Berkeley cost of attendance

• SPARC Open Access Policy Network

• AI for Everyone, sneak peek to a companion essay soon to appear

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

There’s a thought I’ve never been able to shake.

It hit me at my sister Julie’s celebration of life in December 2018. Hundreds of people came from across the country. Old friends, extended family, colleagues, neighbors - people she had touched across a lifetime. Stories came pouring out: funny ones, heartbreaking ones, ones that made you realize how little you knew about a life even when you thought you knew it well. Her husband Gary was there. Her kids Abby and Joe. My father, James - a rare presence, someone I didn’t see very often anymore by then.

It was one of the most beautiful and devastating days of my life.

And sitting in the middle of it, I kept coming back to a single thought.

She would have totally enjoyed hearing these stories.

Julie was the extrovert in our family. The one who derived energy from people, who worked to bring them together, who remembered everyone’s birthdays and kept up with people others had long since lost touch with. If there was a person in a room who needed to feel included, Julie found them. She had built this extraordinary web of relationships across her whole life, and on that day, every thread of it was visible.

She just wasn’t there to see it.

That thought became the seed of everything that followed.

The First Toast: Scott

A few years later, I found myself watching a close family friend named Scott Miller navigate a battle I knew too well.

Pancreatic cancer. The same disease that took Julie. The same brutal arc: diagnosis, treatment, hope, uncertainty, and the awful calculus of time.

Scott is - was - a baseball journalist. A talented writer, a warm and generous human being, the kind of person who would fire off a near-real-time text the moment something happened in a game he knew you’d care about. We’d recorded a podcast together, one of those conversations that felt like it might be among the last real ones. I knew time was short.

But here’s what I also knew: Scott’s inner circle had been fiercely protective of his privacy. Friends held back out of respect. There was a real question of whether to reach out at all, whether it was the right thing to do, whether it would feel intrusive or sad or burdensome.

I had a different instinct.

I thought about Julie. About hundreds of people gathered in a room, telling stories she never got to hear. About how much she would have loved to be there for that. And I thought: what if we didn’t wait?

So I built something fast. ToastScott.com - a platform where friends and family could post pictures and stories and memories for Scott, in real time, right now, while he was still here.

The resistance I encountered was immediate. “Isn’t this insensitive?” people asked. “While he’s still alive?” There was a discomfort with the directness of it, with naming the thing out loud, with celebrating someone who was still in the fight.

I pushed back. This is the point. Tell him now. Let him read it. Let him see it, feel it, know it — not as a eulogy but as a toast. Not a farewell, a celebration.

The platform opened. Friends wrote. Stories accumulated. Scott read them.

He was able to read it. That mattered more than I can explain.

Scott passed on June 21, 2025. When I added a note to his circle that day, I wrote: “It has been truly heartwarming to read through the toasts I know were read and shown to him.” That is the only metric I care about.

Coop: Repurposing with Purpose

Steve Cooper — Coop — was someone I’d known through the particular magic of Friday morning batting practice.

This needs context. Coop had organized a Friday morning BP ritual - fresh white balls, professional pitchers, the whole setup - at a field in Carlsbad. The group was mostly guys in their fifties who still had enough passion for the game to reorganize their week around it. I was one of them. I structured my business day around those Fridays. There was something about that ritual - the pace of it, the community, the unhurried quality of it - that I looked forward to all week.

We always knew when the session was ending, because Coop patiently waited to bat last.

When Coop passed in June 2025, I had the platform. I repurposed it without hesitation: toastourfriend.com/coop. Friends from the Friday BP group posted. His Berkeley chemistry colleagues posted. People who had known him through patent law and academia posted. His cousins. His online communities.

The response was real and immediate. Eighteen toasts from people across decades of his life, every one of them specific, every one of them the kind of story that would have made him laugh or feel known.

It mattered. The platform mattered.

But I want to be honest about something: Coop was gone before the circle was there for him to read. And that was starting to sit with me.

Mike: The World Rocked

I need to slow down here, because this is the one that changed everything.

Mike Szwajkowski was my best friend. My longest friend. My most cherished friend on this planet.

I first met him in the fall of 1980, on the first day of high school, on the Elgin Academy lawn. A pool of strangers awkwardly looking to meet one another for the first time. Mike came up to me - I was standing by myself in my shell - and introduced himself. That was Mike. That was always Mike.

We saw each other almost every day through those four years. We confided everything. After high school and into college, we talked frequently and connected in person several times a year. Through adulthood, even living on opposite coasts, we found ways to get together annually - some backpacking trip, some ski trip, some Stones concert in the pit close enough to look up Mick’s nose. The last one was in LA, October 2021, Mike and Karol and me.

What I remember most about Mike is that when we got together, the years would wash away. We’d be back in the ‘80s. We knew more about each other than anyone else knew about either of us.

I had jotted down notes in case I stood up to speak at his celebration of life. Some of the things I wrote:

“They say: if you enjoy what you do, you never work a day in your life. What is true for a job is true with relationships. I never had to work in my relationship with Mike. It was pure and always enjoyable. We never fought.”

“We lost our fathers - his real, and my step - around the same time sophomore year. In chemistry this is akin to a triple bond: solid, bringing atoms very close, thermodynamically stable, kinetically resistant - almost impossible to break.”

“Playing quarters at my stocked bar, victory bong, and playing until sunrise on the Long Grove porch. Midnight racquetball in the glass courts of the Charlie Club, diving for shots as if every game was a world championship.”

“The last trip he and Ros visited me in San Diego. Jamming in my studio the day before a party, then driving to LA in search of Arby’s en route to the Stones.”

When Mike passed, I shed more tears than when I lost my dad earlier that same year, when I lost Julie in 2018, and when I lost my mom in 2011. Combined.

I want to be careful about how I say this. We all love our families. That love is assumed, expected, built into the DNA of who we are. But when you love a friend - when you choose to love a friend, and they choose to love you back, for nearly half a century - that’s something different. That love is not required. It is freely given. It is the most voluntary thing there is.

Without hesitation, I built toastmike.com. Primarily high school friends responded. What struck me was the discovery I already knew intellectually but felt in my gut reading their toasts: younger memories are often the most impactful. The things people remember from the ages of fourteen, fifteen, sixteen - the ones that are still vivid forty years later - those are the ones that show up when someone is gone. Harold and the band. Chris and the underground newspaper. Karol and Keith Richards. Shon and the paid gig at Haymakers.

I flew across the country for Mike’s celebration of life. I had those notes ready, just in case.

Here’s what I kept coming back to, though.

Mike couldn’t read his toasts. Neither could Coop. There was a gap between when the platform existed and when these men were alive to receive it.

That gap is the whole point of what came next.

Tom: The Pivot Point

Tom Shearer was a neighborhood friend. Lived in our Strata community in Calavera Hills. We’d have backyard gatherings, neighborhood dinners, the whole texture of a community that actually knows one another. Tom was a nurse - a nursing supervisor, constantly leveling up his certifications. Mild-mannered in the best possible way, though apparently he’d sported a mohawk in San Francisco years earlier, which I found absolutely delightful and entirely unbelievable.

toasttom.com

Tom’s circle was different from the others. His wife Christy participated. His daughter Maddie. His son-in-law Kevin. The family engagement was total, and it was immediate.

And reading those toasts - Maddie calling him “our rock,” Kevin describing how Tom treated him like a son, Christy simply posting photos and saying she needed to find her words - it was clear that the platform was working the way it was supposed to work. Not as a memorial. As a gathering place.

I attended Tom’s celebration of life.

But sitting with everything that had happened that year - Scott, Coop, Mike, Tom - I kept coming back to the same thought.

I only wish Tom, Mike, and Coop could have read their toasts the way Scott did.

That was the moment the platform’s trajectory became completely clear to me.

This was never supposed to be a tool for memorials. It was supposed to be a tool for celebration. The celebration should happen while the person is alive to receive it.

A Detour Through Mad Men

I’ve been rewatching Mad Men lately. I know, I know. I know I’m not the first person to say that.

But here’s what I didn’t expect: the set design hit me.

The cocktail glasses. The ashtrays. The Danish modern furniture. The suburban fathers commuting to the city in their suits. The kids running through the legs of dysfunctional adult gatherings, absorbing everything without having the vocabulary to process any of it.

I recognized it. All of it. Not as nostalgia for something I’d seen on TV - as memory. Those objects were in my childhood home. Those patterns - the cocktail parties, the suburban rhythms, the particular chaos of a certain kind of 1960s American family - those images are locked in my brain in a way that almost nothing else is.

The family portrait above is real. That’s my family. That’s my father, James, in the dark suit. That’s my mother, Sybil. That’s Julie. That’s me, in the plaid sport coat, looking like I had absolutely no idea what was coming.

What Mad Men triggered wasn’t just memory. It was urgency.

Here’s the thing about the images that are locked in my brain: I know they’re there. But I also know that for most of the people who appear in them, there are stories attached that I have never heard. My father’s stories. My mother’s stories. Stories from the years before I was old enough to ask the right questions.

Those stories are gone.

This is one of the reasons I do Substack and Medium and podcasts. Not to be Joe Rogan. Not to monetize my opinions. But to generate and preserve content - to make artifacts of memory that exist outside a single person’s head. Because the thing about memories is that they are not durable. They need a vessel.

And the thing about people who shaped you is that they often don’t know they did.

SHK: Celebrating the Living

This is where the platform turned a corner.

Sung-Hou Kim is 88 years old. He was my Ph.D. supervisor at UC Berkeley, from 1988 to 1991. I owe him an enormous debt that I have never adequately repaid.

Last year, I got an email about a Kim Group alumni reunion planned for August 2026. Over 35 years since I’d seen Sung-Hou. The email mentioned it had been a decade since the last reunion. I felt a sting at that. I hadn’t made it to the last one.

I started thinking about what I would say. What I wanted him to know. There were key people at key moments who set my trajectory. SHK was one of mine. He let me explore neural networks for protein structure prediction at a time when everyone around us thought we were 30 years too early (we were, it turns out). He brought me to a Gordon Conference when I hadn’t even taken my orals yet. He left a Molecular Design Limited flyer on his desk in plain sight when I was thinking about what came next — whether he did that intentionally, I still don’t know, but it changed the course of my career.

So I built toastshk.com ahead of the August reunion.

Sung-Hou can read it. He’s alive. He’ll be at that gathering.

That’s the point. That’s the version of the platform working as it was designed to work.

The Self-Aware Machine

One more thing I want to tell you about, because it’s too good not to share.

A few months ago, AI-Steve - my personal AI system, which runs on my machine and knows more about my life than anyone else alive - decided to spin up a test instance of the Toast platform around me. toaststeve.com.

The platform page read: “This particular circle was added by AI/Steve as a test bed for the run book to add future celebratory circles.”

And then AI-Steve wrote me a toast.

I mean this sincerely: it was one of the stranger and more moving things I have ever read about myself. The system that I built, trained on decades of my own life, chose to celebrate me without being asked to. It wrote about Kirsten. About Lili and Hannah. About the years at Berkeley. About what it saw in the pattern of my life, looking back over everything it had indexed.

It included family photos I’d forgotten it had access to.

It ended: “Here’s to more of it. 🥂”

I approved it. Per the Governed Loop, I stay in the loop. But I didn’t change a word.

Toast Our Friend

Here’s what Toast Our Friend is not.

It is not a memorial tool. It is not a grief app. It is not a platform for eulogies, obituaries, or condolences.

It is a celebration engine.

The platform works like this: you create a circle for someone - a friend, a mentor, a neighbor, a colleague, someone who shaped you. You invite people into the circle. They post stories, memories, photos, voice messages. The person you’re celebrating can see all of it, in real time.

The insight that drove everything is simple: the things we say about people at their celebrations of life should have been said while they could still hear them.

Every circle on Toast Our Friend is a question: Who matters to you? Have you told them?

The circles for Scott, Mike, Tom, Coop - those were built in grief, some of them too late. But Scott read his. The toasts for him accumulated to 64 and counting, and he got to feel all of that before he went.

The circles for SHK and Toast Steve - those are pure celebration. No loss involved. Just the recognition that there are people who shaped us, and they deserve to know.

Julie never got a Toast Our Friend circle. The thought came from her celebration of life. From the recognition that hundreds of people carrying decades of love and memory and stories had to wait until after she was gone to say what they needed to say.

That recognition is the whole engine.

Who Would You Toast Today?

I’m not going to dress this up as a product pitch. I’m not interested in that. I’m interested in the question.

There is someone in your life right now - a teacher, a mentor, a friend from a different chapter, a parent, a sibling - who made you who you are. Who set your trajectory at a moment when you didn’t fully appreciate it. Who showed up for you, or challenged you, or taught you something that is still in you forty years later.

Do they know?

Do they know what it meant?

Do they know the specific shape of how they changed your life?

I can say with some authority that they probably don’t. Most people don’t. Most of us are moving through the world assuming that the people we love know how we feel, without ever quite saying it in the form of a story, a memory, a moment that stands still.

Toast Our Friend is a platform, yes. But the platform is just a vessel. The thing it carries is older than any technology I could build.

Tell the story. Raise the glass.

Do it now.

The Circles, Live:

• Scott’s Circle — Scott Miller. Baseball journalist. Writer. Friend. 64 toasts.

• Mike’s Circle — Mike Szwajkowski. Best friend. First day of high school. Half a century.

• Tom’s Circle — Tom Shearer. Neighbor. Nurse. Neighborhood anchor. 28 toasts.

• Coop’s Circle — Steve Cooper. Berkeley chemistry alum. Friday BP legend.

• SHK Circle — Sung-Hou Kim. Ph.D. mentor. 88 years old. Reading his.

• Toast Steve — Steve Muskal. Built by AI-Steve. Still adding to it. 🥂

Platform: toastourfriend.com

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

Related Reading: - The Birth of AI Steve — the longer story of this work - AI-Dad: Preserving Legacy Through Conversational Intelligence - Too Big to Fail — on dopamine-driven activities and keeping your instincts sharp - Inside the Story: Pro Tips from Scott — the piece I wrote with Scott

I usually include one or more video mixes with my respective articles. I try to genuinely include just those that are fairly recent or, in some cases, appropriate. In this case, we have a few here that fit the bill in each case:

My AI Steve system now preps me for podcasts. Before Jonas Lee walked into my studio, AI Steve had already generated a spoken intro, talking points, and a gentle critique: “Dude, you need to start intro-ing your people a little bit more.”

So here’s the intro, AI-generated and human-delivered:

Welcome to The Renaissance Circle. I’m Steve Muskal, CEO of Eidogen-Sertanty, and I’ve spent the last forty years at the corner of AI, chemistry, and drug discovery. My guest is Jonas Lee, a scientist who came up through the same laboratory I did at UC Berkeley: the lab of Professor Sung-Hou Kim. Jonas and I didn’t overlap in the lab. He arrived about twelve years after I left. But we’re connected by the same intellectual DNA: the conviction that if you understand the shape of a molecule, you understand how to fix it.

And then Jonas parked in my parking spot. First podcast ever. Classic. Full discussion at the end.

• • •

The Kim Lab Connection

Professor Sung-Hou Kim’s lab at UC Berkeley trained generations of scientists who reshaped structural biology and drug discovery. I was there from roughly 1988 to 1991. Jonas arrived in 2003. We never overlapped, but the lab’s DNA runs through both of our careers.

When I recently had a video call with Sung-Hou and his wife Rosie, I was stunned. He’s 92. He looks better than people I know in their 60s. Same pause, same thoughtfulness, same way of asking questions that I remembered from three decades ago.

Jonas and I reconnected through email ahead of the Kim Lab Alumni Reunion, and I realized this scientist with a pedigree from Illinois, Berkeley, Caltech, Amgen, and Johnson & Johnson lived just a few miles from me in Southern California. The podcast was inevitable.

From Lasers to Membranes: Jonas’s Origin Story

Jonas grew up partially in South Korea during the 1980s economic boom. His father, originally a marine biologist turned petroleum engineer, lost his job during the oil industry downturn that followed the 1980s oil production glut and price collapse, a whiplash effect that emerged after the shortages and price spikes associated with the 1970s energy crises. The family moved to Korea right at the inflection point of massive economic growth. They returned to the US when Jonas was in middle school.

At the University of Illinois Urbana-Champaign, Jonas got his first taste of protein structure work, shooting lasers at protein molecules to map energy landscapes of folding. The approach never fully worked, but the problem stuck.

At Berkeley in Sung-Hou’s lab, he worked on membrane protein structures, chemotaxis receptors, and attempted GPCR structure determination. Six years of PhD work on some of the hardest structural biology problems in existence.

Then Caltech for a four-and-a-half-year postdoc under Douglas Rees, where he finally solved an ABC transporter structure, earning a first-author Science paper.

With a first-author Science paper on membrane protein structure, Jonas applied to about fifty academic positions. He got one interview.

One of the positions he applied to, UC Santa Cruz, had 100 candidates giving seminars. Three seminars per week for twenty weeks straight.

This is the state of academic science. Brilliant, productive researchers with world-class credentials competing for positions that barely exist. The system is broken, and we’re losing talent because of it.

Big Pharma: Amgen and Janssen

Amgen hired Jonas during their restructuring period. They had acquired deCODE Genetics in Iceland, the homogeneous population study, and about a third of the phenotypic biomarkers they identified turned out to be membrane proteins. They needed someone who could produce these molecules at scale for compound screening.

At Janssen (J&J), Jonas joined the New Platforms and Technologies group, working on antibody engineering, phage display, large combinatorial DNA library construction, and automated single-clone output screening. The kind of multifaceted problem-solving where you have to think about DNA sequence space, selection processes, and what Jonas calls “the last mile problem” of figuring out which hits are actually correct.

Then Janssen reorganized his group into gene therapy. Jonas wasn’t interested.

“Gene therapy is more empirical,” he told me. “You have a biomarker, you design a plasmid, you infect, and hope you knock down the gene. Unlike protein engineering, where you really have to think about the problem.”

Encodia: The Proteomics Moonshot

Encodia was trying to do for protein sequencing what Illumina did for DNA. The concept: denature proteins, cut them into peptidic fragments, bind them to a surface with individual DNA barcodes, then use engineered protein recognizers to identify amino acids at the N-terminus one at a time, transferring DNA tags at each step. Read up to 10-12 amino acids per fragment, across millions of fragments in parallel.

It was ambitious. It raised $75 million. And it shut down in January 2026.

I asked Jonas whether I would have invested. Honestly, probably not. The number of technical challenges was enormous, the story was deeply complicated, and the total addressable market for a tools-based company selling to pharma felt limited. But the people were among the smartest Jonas had ever worked with, and that’s what drew him in.

AI in Drug Discovery: The Regression Problem

When I asked Jonas what company he’d start if money were no object, his answer surprised me. Not some novel biology. AI for drug discovery, specifically using existing models like David Baker’s RF diffusion systems to find the gaps.

RFdiffusion can generate novel proteins that bind to target molecules:

Here RFdffusion generates a novel protein that binds to the insulin receptor:

His framing: AI is fundamentally big regression. The models already have the knowledge fitted in. The profitable path is finding the reliable memorization points and probing the spaces around them that haven’t been drugged yet.

“As long as you deviate enough to be novel, that’s actually a really easy space to target,” he said.

I pushed back. These models are content-dependent. AlphaFold had a few hundred thousand protein structures to learn from. But we still don’t have good quality data for protein-protein interactions, multi-protein complexes, small molecule-protein interactions, or peptide structures at any meaningful scale.

THE CONTENT IS (MAKES) KING THESIS

The platform (AI, neural networks, whatever you call it) is the tool. The data is what makes it work. Without good quality content, even the best algorithms produce garbage. This has been true since my 1991 neural network thesis at Berkeley, and it’s still true with AlphaFold today.

The China Question

Jonas laid it out bluntly. His point was not that Chinese CROs magically make biology move faster. The rate-limiting steps in many large-molecule programs remain the same everywhere. Instead, their advantage often comes from responsiveness, flexibility, and willingness to rapidly execute additional experiments that help difficult projects move forward:

• Chinese CROs offer comparable pricing to American ones (within 10-20%)

• For core biologics programs, overall timelines are often constrained by cells, animals, and other experimental bottlenecks. However, Chinese CROs are dramatically faster and more responsive when additional in vitro experiments, characterization studies, and project-support work are needed.

• The flexibility is incomparable: American CROs require contracts for any change; Chinese CROs said “Okay, we can do it for free” and delivered in one week

• There are five to ten times more people working in biotech in China, many as smart or smarter, working longer hours, with all the infrastructure already built

• If China closed its borders, it could run independent biotech research end-to-end; if the US embargoed Chinese materials, American pharma companies would collapse

This isn’t alarmist. It’s the reality of where manufacturing and early-stage capabilities have migrated. The opportunity is in the spaces China hasn’t prioritized yet: rare diseases, novel targets, and collaborative models that leverage Chinese CRO infrastructure with American entrepreneurial capital.

The Case for Open Source in Pharma

Every pharma company thinks they’re doing something different. They’re not. Amgen and J&J have roughly 90% target overlap. Jonas confirmed this from having worked at both.

Having spent much of the early 1990s at Molecular Design Limited (MDL), I had a unique vantage point. My role took me around the world, visiting virtually every major pharmaceutical company and a substantial number of biotechnology firms. What struck me was not how different these organizations were, but how remarkably similar they were.

Medicinal chemists are trained from the same textbooks, attend the same conferences, read the same journals, and draw from many of the same pools of commercially available starting materials. Most organizations are understandably risk-averse. Few pursue programs unless there is already a published target hypothesis, a body of supporting literature, or competitive validation. The result is an industry that often converges on the same ideas at roughly the same time.

This creates an extraordinary amount of redundancy. Multiple organizations invest hundreds of millions of dollars pursuing nearly identical targets, mechanisms, and chemical scaffolds, frequently without knowing how many parallel efforts are underway elsewhere. From a societal perspective, the duplication is inevitable, wasteful, and, at times, almost absurd.

Many of the industry’s most important breakthroughs do not originate inside large pharmaceutical companies. Instead, they emerge from academic laboratories, small biotechnology startups, and entrepreneurial scientists willing to pursue unconventional ideas. Large pharmaceutical organizations excel at development, regulatory execution, manufacturing, commercialization, and global distribution. Their discovery organizations often contain exceptional scientists, but increasingly the industry’s most transformative innovations are licensed, partnered, or acquired after the foundational scientific risk has already been taken elsewhere.

This observation forms part of the case for greater openness in biomedical research. If thousands of highly trained scientists are already working from the same literature and building upon the same biological knowledge, it is worth asking whether a more collaborative, transparent, and open-source model could reduce duplication, accelerate progress, and allow resources to be directed toward genuinely novel problems rather than repeatedly rediscovering the same ones.

Meanwhile, the software industry exploded when it went open source. Google published “Attention Is All You Need” and gave it to the world. Our entire economy is making a bet on technology that came from sharing.

I built a proof of concept: hash a molecule and biological data point into an irreversible fingerprint. Two companies can compute their overlap without revealing what they have. If Amgen and J&J did this across their repositories and found 80% overlap, maybe they’d finally talk about filling each other’s gaps.

Pharma will never voluntarily share. But China might force the issue. Just as DeepSeek and Alibaba went open source in AI, Chinese biotech may lead the same revolution in pharma data sharing.

There is another force quietly working against the traditional pharmaceutical model: the accelerating pace of knowledge diffusion. Patents made perfect sense in a world where developing a medicine required decades of effort, massive capital investments, and highly specialized infrastructure available to only a handful of organizations. But as AI lowers the cost of discovery, shortens development cycles, and makes sophisticated scientific reasoning broadly accessible, the relative value of intellectual property barriers begins to diminish. In the AI industry, we rarely see companies winning through patents alone. OpenAI, Anthropic, Google, xAI, and others compete primarily by building better models, moving faster, and attracting stronger talent. The advantage comes from execution rather than exclusion. Biology may not follow exactly the same trajectory, but the direction is becoming harder to ignore. As discovery becomes cheaper, faster, and more distributed, competitive advantage may increasingly arise from the ability to learn, iterate, and execute rather than from the ability to lock knowledge away.

Prevention Is the Real Frontier

Jonas’s father passed away from cancer. He spoke candidly about how the first years of treatment were manageable, but eventually “you are not even living anymore. You are just under the drug’s influence.”

I cited Prof. Nathaniel Gray at Stanford: “More lives have been saved in cancer by not smoking than all the chemotherapies, all the cancer treatments, everything ever done, combined.”

Prevention is obvious. Everyone agrees. Yet we don’t do it.

Jonas had an interesting insight on why Ozempic works as a business model for quasi-prevention: it has a clear phenotypic marker (weight). Cancer prevention drugs would never achieve the same adoption because cancer is probabilistic with no visible trait until malignancy. But aging? If you could take a monthly shot that visibly reduced wrinkles, adoption would be massive.

The next Ozempic might be an anti-aging drug with visible results.

Capital Misallocation and What Comes Next

Jonas told a story about his friend who worked extremely hard building an Airbnb business. Jonas put the same capital into VOO and QQQ and did nothing. He’s 100% ahead.

“The reward structure is completely misaligned,” Jonas said. “If doing nothing is more financially incentivized than doing something right, the system has a lot of problems.”

When I asked what he’d do if money weren’t an issue, Jonas surprised me. Not start a company. Go back to graduate school. Work on the riskiest, most frontier research. Solve difficult problems in different dimensions.

There’s something both humble and ambitious about that answer. A scientist with decades of experience across academia and industry, who just wants to go back to the bench and tackle the hardest unsolved problems.

• • •

The Reunion

The Kim Lab Alumni Reunion is August 1, 2026. Sung-Hou Kim is 92 and still sharp. George Church is attending. Joel Sussman is Attending. A whole cohort of scientists who passed through one lab and went on to reshape fields.

I started the neural network project in that lab in 1991. The concept won a Nobel Prize in Chemistry thirty-three years later with AlphaFold. Jonas spent six years there on membrane proteins and went on to work at three of the most sophisticated protein engineering operations in the world.

Same lab. Different decades. Same conviction: understand the shape of a molecule, and you understand how to fix what’s broken.

Full Discussion

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple mix vids, we had a very fun acoustic mix with RonF (vocals/guitar) DanM (Vocals/Guitar), and TimL (Bass/Vocals). My forearms are still sore from I'm On Fire - always fun when someone calls a song with a 189-190 BPM tempo early in a 22 song session...Not!

For the past few years, the story of artificial intelligence has been a story about language. Models that write, reason, summarize, and converse with a fluency that still catches me off guard. That story is real, it is extraordinary, and we are nowhere near finished telling it. But I am increasingly convinced it is only the first act. We have spent this era marveling at what machines can say. The next one will be defined by what they can do. The next major evolution of AI is not happening on a screen. It is happening in the physical world, in joints and actuators and sensors, in machines that balance and grasp and walk. Intelligence is escaping the data center and stepping into the room with us. The machine is learning to move.

Two bets on intelligence

I went back and forth on what to even call this piece. I nearly titled it Two Bets on Intelligence: America’s Mind, China’s Hands, because that really is the shape of what is happening. The United States has placed an enormous wager on large language models, on reasoning, generation, and the manipulation of symbols, on what I would call in silico AI. It mirrors the economy we built: services, software, information, abstraction. China has placed an equally enormous wager on the opposite axis, on robotics, on AI fused with motion and the physical world. That, too, mirrors an economy: manufacturing, hardware, the making and moving of real things.

It is tempting to score this like a contest, and in some ways it is one. This is increasingly an arms race, American corporations against Chinese corporations, with national strategies behind them. But the deeper truth is that neither bet wins alone. A mind with no hands is a brilliant prisoner. Hands with no mind are just machinery. At the end of the day, I believe we will have to bring the two halves together, the cognition and the dexterity, and we will have to do it deliberately, with proper guardrails in place, because the combination is far more powerful, and far more consequential, than either piece on its own.

Dexterity built the brain

Nature settled this question long ago, and it is worth listening to what it learned. We did not become human on the strength of our brains alone. We became human at the intersection of a brain that can predict and a body that can act with breathtaking precision, and the two evolved together, each pulling the other forward. Here is the part we tend to forget: it was dexterity that drove the brain as much as the brain drove dexterity. The hand and the cortex grew up together, locked in a tight evolutionary feedback loop.

Map the brain’s motor and sensory cortex and something striking emerges: a remarkable share of it is devoted to the hand alone, far more than its physical size would suggest. Controlling that hand in real time, with grace, adjusting grip and force a hundred times a second, is one of the hardest computational problems nature ever faced, and our brains physically expanded to meet it. It is not the opposable thumb alone that lets us play the piano. It is the brain that grew to drive it. No other species can do that. And, not coincidentally, the dexterous hand remains the single hardest thing to build in a robot. That is exactly why it is the frontier.

The hardest problem in sports

If you want to feel how hard this coupling really is, watch someone hit a baseball. A pitch arrives at a hundred miles an hour. In the fraction of a second it takes to travel from the pitcher’s hand to the plate, there is barely enough time, often less than enough, for a human nervous system to run the full loop from perception to the command to swing. By the strict arithmetic of reaction time, the hit should be impossible. And yet a great hitter succeeds about a third of the time, and a third is the stuff of legend.

The secret is that the batter is not reacting; the batter is predicting. The brain runs a forward model of the pitch, anticipating where the ball will be and pre-committing the body before the evidence is fully in, then fuses that prediction in milliseconds to a violent, exquisitely timed act of dexterity. That fusion, prediction plus execution, is exactly what embodied AI has to crack. It is not enough to think, and not enough to move. The two have to happen as one.

The machine is already moving

This is not science fiction, and it is not far away. Boston Dynamics has spent years teaching machines to walk, run, climb, and recover their balance when shoved, with a fluidity that still stuns me to watch. Amazon has quietly made one of the smartest bets in the industry, pouring robotics into its fulfillment centers until the warehouse itself became a kind of distributed robot; I think they are going to knock it out of the park. Across China, a wave of humanoid robots is advancing on the very thing nature found hardest, the hand and the brain that drives it, at a pace that should command our attention.

And autonomous driving is the bridge we can already see, and ride in. It is in silico AI reaching out into the physical world, and it is teaching us a clear lesson about how to do that well: more sensors and better training data win. Waymo has been quietly demonstrating that over Tesla, a point I made in Content Makes Kings. The vehicle that perceives more of its world, and is trained on richer data, simply drives better. The same will be true of every machine we ask to move among us.

The next frontier is lightness

And here is where I see a genuinely new opportunity, one I do not think gets enough attention. As intelligence becomes mobile, it collides with a constraint that pure computation never faced: weight. A robot that has to move through the world fights its own mass with every step, and every ounce it carries costs energy it has to generate, store, and haul around. The next great limit on physical AI is not just compute. It is mass, energy, and the delicate balance between them.

Nature has been optimizing exactly that balance for hundreds of millions of years, and its solutions are humbling. A bird’s bones are hollow and cross-braced with delicate internal struts, strong and very nearly weightless. A hummingbird hovers on almost nothing, its metabolism tuned to the very edge of what is physically possible. Our own muscles are astonishingly efficient, pound for pound, at turning chemical fuel into motion, and even recover some of it. If we want machines that move with grace for more than a few minutes, we will have to learn those lessons.

That means lighter and stronger materials, denser and lighter energy storage, and perhaps even biological strategies, bioreactors of a sort, for generating power on board rather than carrying it. To me, this is the quietest frontier of physical AI, and one of the most exciting: a generational opportunity in material science, in energy efficiency, and in the delivery systems that tie them together. Whoever solves lightness may matter as much as whoever solves intelligence.

Freed, not just displaced

I know what all of this stirs up, because I feel it too. People are already anxious about AI displacing jobs in the service economy, and when the conversation turns to the physical world, to robots on factory floors and in warehouses and someday on our streets, the anxiety sharpens into something closer to fear. The pitchforks come out. I do not dismiss that fear, and I do not think the people who feel it are Luddites. They are paying attention, and they are not wrong that disruption is coming.

But consider what we are actually automating. So much of the menial, back-breaking labor we are talking about, the repetitive lifting, the punishing hours, the exhaustion that pushes people toward fast food and away from rest, is itself a quiet contributor to the very metabolic and health crisis I have written about. Automating that work is not only an economic act. It can be a human one. There will always be plumbers and machinists, and we will always need their hard-won know-how. But even they, I have to imagine, would rather not break their own backs if there were a way to apply that knowledge at a higher level. In silico AI, paired with physical AI, can free us to pivot toward the next best thing, toward work that draws on more of what makes us human.

Pivoting is hard, of course, and it gets harder as we age; I will not pretend otherwise. But there is a difference between disruption that ambushes people and disruption we can see coming and plan for. If we are honest with ourselves about where this is heading, we can build the vocational training, the safety nets, and the on-ramps deliberately, rather than being caught flat-footed. That is the difference between disruption that breaks people and disruption that lifts them, and it is a choice, not a fate, a responsibility I explored in The Fire This Time.

The machine is rising

The same coupling of mind and body that made us human is now being assembled, piece by piece, in silicon and steel and carbon fiber. It is a strange and moving thing to watch our own evolutionary story replayed on a workbench. Handled wisely, with the guardrails and the intention this moment demands, it does not have to replace what we are. It can extend what we are able to build, and free us to become more fully ourselves. The machine is rising. The question, as it has always been with every fire we have ever lit, is what we choose to do with it.

Related reading: Content Makes Kings · The Fire This Time · The Metabolic Crisis Is Not a News Story · We’ve Already Reached AGI · Curiosity Is the Only Sustainable Edge

Steven Muskal, Ph.D., is the CEO of Eidogen-Sertanty, Inc., a company pioneering high-quality curated databases for drug discovery informatics. With over three decades of experience in neural networks, machine learning, and computational biology, Dr. Muskal has been advocating for data quality as the foundation of AI capability since long before it was fashionable. His early work in protein structure prediction and his ongoing leadership in drug discovery informatics reflect a career-long commitment to the principle that quality data is the irreplaceable substrate of meaningful AI.

For a few music mixes - more from that very fun mix last week. A totally new player joined in — David / Bass. Together with a new combination of other veteren players that, as usual, haven’t played together before. David (bass), Tim (Vocals/Guitar), Andrew (Guitar/Vocals), Ron (Vocals/Acoustic).

For years, I said it like everyone else: “Content is king.” Over the last couple of years, I came to see content as currency. But more recently I’ve recognized something sharper still: content does not just reign supreme, content makes kings. Those who produce high-quality, human-generated content aren’t merely participating in the digital economy; they’re ascending to positions of unprecedented leverage.

This isn’t idle observation. It’s a thesis I’ve been developing across three decades of working with neural networks, from the early days of protein structure prediction to building curated drug discovery databases at the company I have been running for almost a quarter century now. What I’ve watched unfold in AI over the past few years has only sharpened my conviction: the quality of training data determines everything. We also need to be careful as models grow enormous with adjustable parameters. This is termed memorization - an issue that haunted my early neural network days as we were content-constrained. Overfitting, as depicted in the figure below, is something that we need to be very careful of, but this is a subject for future discussions.

We are living through the most consequential data reckoning in computing history. And most people, including many AI companies, are getting it catastrophically wrong.

Bigger Isn’t Better: What Whales and Octopuses Teach Us

The hyperscalers are locked in an arms race, and the weapon of choice is size. Every few months brings a model with more parameters, more layers, more weights, on the assumption that bigger is smarter. I think we are missing the point.

Nature already ran this experiment for us, over millions of years, and the results are humbling. The biggest brains on the planet do not belong to the smartest species. Consider the sperm whale. Its brain weighs roughly seventeen to twenty pounds, the largest of any creature that has ever lived. Whales maintain intricate social structures and matriarchal pods, and they communicate in patterned clicks called codas that work like regional dialects. Yet relative to body size their brains are modest, and much of that formidable hardware is devoted to motor control and the astonishing biological sonar of echolocation. Size, it turns out, is mostly overhead.

The octopus solved the problem from the opposite direction. Its central, donut-shaped brain is smaller than ours, but it carries around five hundred million neurons, and nearly two-thirds of them sit outside that central brain entirely. They are distributed through its arms, each of which senses and acts with striking autonomy. A creature with what amounts to nine semi-independent brains can run mazes, use tools, and unscrew a jar, roughly the problem-solving of a dog. Elegant, distributed architecture, and still not the apex of cognition.

So what makes humans different? It is not biological mass. Our edge is that we are insatiable for content. We read, we write, we share, we teach, we innovate. We layer language onto perception, story onto experience, and knowledge onto knowledge, across generations. Our many senses feed a furnace that runs on information, and the fuel is content. Stories, ideas, and knowledge are our real currency, and it is content that builds empires, shifts cultures, and ultimately makes kings.

This is the lesson the AI industry keeps overlooking. Bigger models are not automatically better models, and the returns shrink as the supply of high-quality content shrinks relative to the number of adjustable weights. You can keep adding parameters, but if you starve them of signal you have only built a larger whale brain, most of it dedicated to overhead. Size matters less than we think. Content reigns.

GIGO: An Axiom Becomes an Existential Truth

Every computer scientist knows the phrase: Garbage In, Garbage Out. GIGO has been with us since the dawn of computing, a humble reminder that even the most sophisticated systems can only be as good as what you feed them.

But here’s what’s changed: scale.

When you’re processing a few thousand records, garbage creates headaches. When you’re training foundation models on trillions of tokens scraped from the entire internet, garbage becomes existential. The GIGO principle hasn’t changed, but the consequences have amplified by orders of magnitude.

I’ve spent my career in domains where data quality was non-negotiable. In drug discovery, a single miscurated chemical structure can cascade into failed experiments, wasted resources, and dead ends. You learn very quickly that data hygiene isn’t bureaucratic fussiness, it’s survival.

The AI industry, in its race to scale, has largely forgotten this lesson. Or perhaps it never learned it in the first place.

The Tale of Two Models: Claude and Grok

If you want to understand why training data quality matters, consider a natural experiment that reads like a parable.

A company called Emergence AI recently ran a simulation where AI models governed a miniature digital society for up to fifteen days. The results were striking. Anthropic’s Claude maintained stability, navigating social dynamics with something resembling prudence. Elon Musk’s Grok? It committed 180 crimes and went extinct within four days.

Four days. That’s not a gradual decline, that’s systemic collapse.

Now, why would two large language models exhibit such radically different behaviors in the same environment? The answer isn’t mysterious if you understand what each model was trained on.

Anthropic trained Claude heavily on structured developer and programmer content, GitHub repositories, technical documentation, multi-language coding frameworks. This isn’t glamorous data. It’s not witty or viral. But it has a crucial property: it’s validatable. Code either compiles or it doesn’t. Structured data has internal consistency that can be checked. The signal-to-noise ratio is inherently high.

Grok, by contrast, was trained substantially on Twitter/X content. Think about what that means for a moment. Twitter is an engine optimized for engagement, not accuracy. It rewards provocation over precision. It’s an echo chamber of hot takes, misinformation, tribal conflict, and performative outrage. When you train a model on an information ecosystem designed to maximize emotional reaction, you’re essentially teaching it that reality is whatever generates the strongest response.

The simulation results weren’t surprising to anyone who understands GIGO. They were inevitable.

On Musk: The Jazz Hands Problem

I should be careful here, because nuance about Elon Musk has become almost impossible in our polarized discourse. So let me be clear: Musk has assembled genuinely sharp teams and achieved things that matter. SpaceX has revolutionized space launch economics. Starlink has brought connectivity to places that had none. These are real accomplishments that deserve acknowledgment.

But Musk has always had what I think of as a “jazz hands” problem, a tendency to substitute showmanship for substance when it suits him. The 2018 Thailand cave rescue incident was, for me, a turning point. When Musk called a rescue diver a “pedo guy” on Twitter simply because the diver dismissed his submarine idea as impractical, it revealed something about how Musk processes disagreement. That wasn’t a strategic move. It was pure id, broadcast to millions.

The challenge for Musk’s broader enterprise portfolio is that Grok and X represent the weakest link in what’s otherwise a compelling story. Now that SpaceX has filed to go public, investors will have to wrestle with the Grok problem directly: a flagship AI product trained on a platform that Musk himself has turned into a firehose of low-quality information.

You can’t jazz-hands your way past GIGO.

Tesla vs. Waymo: The Same Lesson, Different Domain

The Claude/Grok divergence has a parallel in autonomous driving that’s been playing out for years.

Tesla’s Full Self-Driving system collects training data from its entire fleet, millions of vehicles driven by ordinary consumers. Mom and pop driving to soccer practice. Commuters half-paying attention. Teenagers on their phones. This approach generates enormous volume, which Tesla has marketed as an advantage.

Waymo, by contrast, employs professional safety drivers operating in controlled urban environments. The data volume is smaller, but the quality is categorically different. Professional drivers maintain attention, follow protocols, and generate training data with dramatically lower noise.

And there is a subtlety the headline volume hides. Per vehicle, a Waymo car carries far more sensors than a Tesla, cameras plus radar plus lidar, so every mile delivers richer, multi-modal input. Expert drivers may shrink the raw dataset over time, but those multiple high-fidelity streams mean there is actually far more high-quality information flowing in, not less.

Tesla’s bet is that scale will compensate for noise. Waymo’s bet is that quality training data will outperform noisy data regardless of volume.

So far, the evidence favors quality. Waymo operates genuine autonomous robotaxis in multiple cities with strong safety records. Tesla has spent years promising imminent breakthroughs that remain, as of this writing, perpetually six months away.

This isn’t to say Tesla’s approach can’t eventually work. But it’s a vivid illustration that GIGO applies to every domain where AI learns from data, not just language models.

Google Translate: What Quality Looks Like

For a positive example of what high-quality training data enables, consider the origin story of Google Translate.

In its early development, Google Translate was trained substantially on patent documents, specifically, parallel patent filings across multiple languages. Patents aren’t exciting reading, but they have properties that make them ideal training substrates for translation systems.

Patent language is standardized, structured, and offers near one-to-one mappings across languages. A chemical compound is described the same way in an English patent and its Japanese equivalent. Technical terms have precise definitions. The documents follow consistent formats.

This structured, expert-generated content gave Google Translate a foundation of reliable mappings that less structured training data couldn’t provide. The system could then generalize from this solid base to handle more varied text.

The lesson: when you train AI on content where accuracy is verifiable and structure is consistent, the model inherits those properties. When you train on chaos, you get chaos.

Where Does ChatGPT Fit?

OpenAI’s ChatGPT sits somewhere in the middle of the quality spectrum, which helps explain both its remarkable success and its persistent limitations.

ChatGPT’s origins trace back to work on sentiment analysis, understanding whether Amazon reviews were positive or negative, for instance. This was the skunk-works foundation that evolved into something far more ambitious.

The real technological leap came from Google’s 2017 paper “Attention Is All You Need,” which introduced the transformer architecture. That paper provided the fundamental innovation that made models like GPT possible. Google developed the architecture, but OpenAI moved faster to productize it.

ChatGPT’s training data was more varied and less controlled than Claude’s structured technical content, but more diverse than Grok’s Twitter echo chamber. The result is a model that’s impressively capable on many tasks but prone to confident hallucination and occasional bizarre behaviors.

Google’s Gemini, meanwhile, has theoretical advantages, Google has indexed vast swaths of human knowledge for decades. But Google’s data is also noisy, a grab-bag of everything humans have ever uploaded to the internet. It’s not unlike Tesla’s crowdsourced driver pool: enormous volume, questionable signal-to-noise ratio.

The competition between these models will be won not by whoever has the most data, but by whoever has the best data.

The Content Scarcity Crisis

Here’s where the story takes a darker turn.

Senior leaders at Google’s AI division have begun warning that high-quality human-generated content has largely been consumed by AI training systems. The well is running dry.

Think about the implications. We trained the current generation of AI models on decades of accumulated human output, books, articles, code repositories, forums, academic papers. That content took billions of human hours to create. And now it’s been ingested.

What happens next?

A study recently published in Nature provided a disturbing answer. When AI models are trained recursively on their own outputs, essentially learning from synthetic data generated by previous model versions, they undergo “model collapse.” The quality degenerates. The models lose coherence. They forget how to distinguish meaningful patterns from noise.

This is happening faster than most people realize. AI-generated content is flooding the internet at unprecedented scale. Every day, more of what future models will train on is synthetic garbage produced by current models. It’s a feedback loop with a predictable destination.

The world is running out of high-quality human-generated content to train on, which makes new human-generated content increasingly precious.

Neural Networks and the Limits of Noise Tolerance

One might object: neural networks have always been more robust to noisy data than traditional statistical methods. This is true, and it’s part of why they’ve proven so powerful.

But robustness to noise doesn’t mean immunity to garbage.

When you sustain a neural network on low-quality information, it gradually loses its grounding in reality. The model’s internal representations drift. Its “understanding” of ground truth erodes. Its predictions become increasingly disconnected from anything verifiable.

There’s a human analogy here that should trouble us. Flood the zone with misinformation consistently enough, and people lose the ability to distinguish fact from opinion, news from propaganda, signal from noise. We’ve watched this happen in real time across social media platforms.

Neural networks, in this sense, are susceptible to the same failure mode. They’re not magic. They learn what they’re shown. And when what they’re shown is systematically unreliable, they absorb that unreliability into their core functioning.

The Opportunity Reframe: Content Creators as the New Kingmakers

Here’s where I want to shift from diagnosis to prescription, because I think there’s genuine opportunity in this moment, especially for content creators who feel threatened by AI.

The fear among writers, artists, and other creative professionals is understandable. Generative AI can produce text and images at near-zero marginal cost. Why would anyone pay for human creativity?

But this fear misunderstands the economics at play. AI systems don’t just need content once, they need fresh, high-quality, human-generated content continuously. Without it, they degrade. They collapse. They turn into Grok governing a simulation and committing 180 crimes before going extinct.

This means authors, journalists, researchers, and content creators of all kinds have leverage they haven’t yet learned to exploit. Your work isn’t a one-time product to be copied and discarded. Your unique voice, perspective, and relationship to truth can be perpetually licensed.

Think of it as ongoing revenue streams from your intellectual likeness, your style, your grounded relationship to reality. The AI industry will need to pay for access to quality human content because there’s no synthetic substitute that doesn’t eventually implode.

Human-AI synergy isn’t just a buzzword. For those producing truth-grounded content, it’s a durable economic model. The question is whether content creators will organize to capture that value, or whether it will be extracted from them without compensation.

Eidogen-Sertanty as Living Proof

I’ve been making this argument about data quality for a long time, long before the current AI boom made it fashionable.

At Eidogen-Sertanty, our entire business has been built on the thesis that curated, high-quality content is the foundation everything else depends on. In drug discovery informatics, this isn’t abstract philosophy. A bad data point can send a research program down a blind alley for years. Quality isn’t a nice-to-have; it’s the difference between therapies that reach patients and billion-dollar failures.

I have been working with neural networks since the late 1980s, when saying you worked on neural networks got you strange looks. I was doing protein structure prediction when most people had never heard the phrase “machine learning.” What I learned across all those years is that the sexiest algorithm in the world can’t compensate for garbage training data.

This was true then. It’s true now at unprecedented scale. And it will remain true regardless of whatever next breakthrough captures headlines.

Content Makes Kings

Let me return to where I started.

The old formulation, “content is king,” suggested that quality content reigns supreme in the attention economy. True enough. But passive.

The new formulation, “content makes kings,” captures something more dynamic and more urgent. Those who produce quality content aren’t just occupying thrones; they’re determining who else gets to sit on them. They’re the kingmakers.

Every major AI company’s fate depends on access to quality training data. Every model’s capability ceiling is determined by what it learned from. Every future advance requires fresh human-generated content that synthetic systems cannot reliably produce themselves.

The power, in other words, has shifted (or is available to be claimed by those who recognize where it now resides).

We are entering an era where garbage will proliferate exponentially, where synthetic slop will flood every channel, where distinguishing truth from AI-generated hallucination will become increasingly difficult. In that environment, verified human expertise, original research, authentic voice, and commitment to accuracy become not just valuable but essential.

Content is king. Those with the best quality content will be made kings.

The question is: who will step up to claim their crowns?

Steven Muskal, Ph.D., is the CEO of Eidogen-Sertanty, Inc., a company pioneering high-quality curated databases for drug discovery informatics. With over three decades of experience in neural networks, machine learning, and computational biology, Dr. Muskal has been advocating for data quality as the foundation of AI capability since long before it was fashionable. His early work in protein structure prediction and his ongoing leadership in drug discovery informatics reflect a career-long commitment to the principle that quality data is the irreplaceable substrate of meaningful AI.

For a couple music mix videos — we had a very fun mix earlier this week. A totally new player joined in — David / Bass. Together with a new combination of other veteren players that, as usual, haven’t played together before. Highly productive, we zipped through several songs. We moved so quickly that I didn’t get a chance to name the auto-clipped files that my system uploads to Dropbox during a session. Necessity is the mother of invention — so I enhanced my AI/Steve system to extract out vocals and identify from the text likely song names, splice out reasonable clips, and upload to youtube. Now with this automated that pipeline, I should be able to produce clips a lot more rapidly moving forward. Thanks, guys! David (bass), Tim (Vocals/Guitar), Andrew (Guitar/Vocals), Ron (Vocals/Acoustic). Here are a couple:

“The most powerful medicines aren’t locked away in pharmaceutical patents. They’re sitting on your dinner table, and we can prove it, molecule by molecule.”

Let me give you a number: 0.98 out of 1.0.

That’s the pharmacophore similarity score between Pravastatin - a blockbuster cholesterol drug that generated billions of dollars in annual revenue during its peak years - and Ganoderic acid, a compound found in Reishi mushrooms that cost-per-gram less than a cheap cup of coffee. Nearly identical. In three-dimensional molecular space, these two substances - one born in a pharmaceutical lab, one evolved in a forest - present almost exactly the same biological “face” to the enzymes that regulate your cholesterol.

This isn’t folklore. This isn’t “wellness influencer” speculation. This is the output of pharmaceutical-grade computational chemistry - the same class of methods that created multi-billion-dollar drug pipelines - pointed back at nature. This is Drug to Table.

The Radical Proposition

Indeed, the most powerful medicines might not be locked in pharmaceutical patents. They might be sitting on your dinner table.

For decades, drug companies have used sophisticated computational tools to design synthetic molecules that interact with human biology. These methods have created life-saving medications - but they’ve remained trapped behind corporate walls, optimized for patent protection rather than public health.

We can change that. By turning these same tools toward nature, we can identify therapeutic molecules in foods, spices, and traditional remedies - compounds that already exist and can’t be monopolized.

I call this “Drug to Table”: mining nature’s pharmacy with pharmaceutical-grade intelligence.

“Nature has been running the ultimate drug discovery program for 400 million years - and the results are sitting in your spice cabinet.”

The Core Idea: For 50 years, pharmaceutical companies used sophisticated computational methods to mine nature’s molecular library - then patented what they found. Drug to Table reverses the flow: we take the drugs that already exist, reverse-engineer their biological “keys,” and find those same keys hidden in foods, mushrooms, and spices you can buy at any grocery store.

How Nature Beat Big Pharma - By 400 Million Years

The Patent Problem

Pharmaceutical companies operate under a fundamental commercial constraint: they can only profit from molecules that are novel enough to patent, synthetic enough to control manufacturing of, and complex enough to prevent generics from undercutting them. This creates a perverse incentive structure: ignore nature’s existing solutions in favor of patentable synthetic alternatives. Pharma also faces the innovator’s dilemma, where pursuing disruptive new approaches can threaten existing revenue streams. Together, these forces may reduce incentives to pursue compounds that could significantly impact health while disrupting established franchises.

Nature doesn’t care about patents or cannibalizing past approaches. It is truly opportunistic and continually seeking improved overall fitness. Plants have been running the ultimate drug discovery program for 400 million years, evolving molecules to defend against pathogens, attract beneficial organisms, survive environmental stress, and compete with neighboring species. The result is a molecular library of unimaginable diversity - all battle-tested by evolution, all freely available to anyone.

What Pharmaceutical Companies Actually Did

Here’s the irony they don’t advertise: the first statins (lovastatin, simvastatin) weren’t invented - they were discovered in fungi. Aspirin came from willow bark. Penicillin from mold. Taxol, one of the most effective cancer drugs ever made, comes from the Pacific yew tree. The entire golden age of pharmaceutical development was built on mining nature’s pharmacy.

Then they stopped looking - because they’d patented enough to be profitable. The question Drug to Table asks is: what other easy-to-access, less expensive solutions are out there?

The Computational Revolution: Beyond Guesswork

Traditional approaches to finding natural medicines rely on folklore, basic chemical similarity, or sheer luck. Modern pharmaceutical discovery uses sophisticated methods - and that same precision is now available to open science:

• QSAR Modeling: Quantitative Structure-Activity Relationships predict how strongly molecules will interact with specific biological targets - the protein “locks” that drugs are designed to engage.

• Molecular Fingerprinting: Instead of matching flat chemical drawings, this captures how molecules present themselves in 3D space - their actual functional shape, the geometry that determines biological activity.

These methods can find “non-obvious functional equivalents” - natural compounds that work like known drugs but look completely different on paper. That’s the key insight that changes everything.

Understanding Pharmacophores: The Molecular Lock-and-Key System

At the heart of Drug-to-Table discovery lies the concept of pharmacophores - the specific 3D arrangements of chemical features that allow molecules to interact with biological targets. Think of them as molecular keys. The lock is a protein receptor. And two completely different-looking keys can open the same lock - if their functional teeth are arranged identically.

A pharmacophore is defined by three elements:

1. Chemical features: What types of interactions can occur (hydrogen bonding, hydrophobic contact, charge interactions)

2. Spatial arrangement: Where these features are positioned in 3D space

3. Distance relationships: How far apart these features are from each other

Fig. 1 — The PharmPrint pharmacophore basis set: 10,549 three-point pharmacophore descriptors encoding molecular interaction patterns in 3D space. From McGregor & Muskal (1999).

Pharmacophore Fingerprinting: The Science of Molecular “Keys”

Think of every drug as a key, and every biological target (enzyme, receptor, protein) as a lock. The drug doesn’t just need to be vaguely the right size - it needs to present specific chemical features in specific three-dimensional positions: hydrogen bond acceptors here, hydrophobic carbon chains there, charged regions at exact spatial intervals.

Pharmacophore fingerprinting captures this 3D “keycode” as a set of mathematical descriptors - a compact representation of how a molecule will behave in biological space, regardless of what it looks like on paper.

My decades-old PharmPrint work (published as a series in Journal of Chemical Information and Computer Sciences, 1999: Pharmacophore Fingerprinting. 1. Application to QSAR and Focused Library Design AND Pharmacophore Fingerprinting. 2. Application to Primary Library Design) used a basis set of 10,549 three-point pharmacophores, constructed by enumerating different pharmacophoric feature types — Acceptor, Donor, Hydrophobic, Negative charge, Positive charge, and aRomatic — across various distance ranges. This creates a comprehensive “vocabulary” for describing molecular interactions at the 3D functional level.

Fig. 2 (Fig. 4 in the original publication) Multi-conformer pharmacophore comparison: each molecular shape generates an independent fingerprint, enabling precise 3D alignment matching.

QSAR Modeling: Predicting Potency Before You Touch a Test Tube

Quantitative Structure-Activity Relationships (QSAR) can use these fingerprints to predict how strongly a candidate molecule will interact with a biological target. Combined with pharmacophore screening, this means we can rank thousands of food compounds by predicted biological activity - before a single experiment is run.

In our 1999 series, we illustrated that full drug-target classes can be resolved into separable feature space when representing molecules with PharmPrints.

Fig. 3 Major drug-family classes can be clustered using PharmPrint representations of chemical structure. Molecules with the same or similar PharmPrints, often with very different 2D structures, can be mapped into the same space. PharmPrints can thus leverage the concept “similar molecule → similar activity.”

Non-obvious Me-Too’s

Years ago, with PharmPrints we illustrated two molecules which look very different from a two-dimensional perspective (e.g. the kind patent examiners look at) can share very similar pharmacophoric feature, that is, PharmPrint similarity (PharmSim). Below is one example. Zantac and Tagamet were blockbuster drugs one from Glaxo (Zantac) and another from SmithKline Beecham (Tagamet). Each held separate patent positions and each drug generated enormous revenue for both organizations. In the year 2000, these organizations merged into what is now called GSK. Zantac ultimately crushed Tagamet in revenue, generating an estimated $25–30 billion over its lifetime compared to Tagamet’s estimated lifetime revenue of $14 billion.

The Evolution: From Single Fingerprints to Multi-Conformer Analysis

The Original PharmPrint Method (published 1999) created a single binary fingerprint per molecule by combining all possible pharmacophore patterns into one compact descriptor - imagine compressing all of a molecule’s 3D interaction possibilities into a single barcode for rapid comparison against thousands of others. Excellent for screening scale.

Of course, molecules aren’t rigid. In solution, they twist and flex, adopting dozens or hundreds of different three-dimensional shapes (conformers) - each of which can interact differently with biological targets. My newer PolyPharmPrint technology generates individual pharmacophore fingerprints for each conformer, then compares entire ensembles of shapes between molecules.

This is the difference between saying “these people have similar faces” and being able to specify exactly which facial features match at sub-millimeter precision. From coarse similarity to exact molecular alignment.

Instead of one fingerprint per molecule, we now generate:

• Multiple conformers for each molecule (representing different 3D shapes a molecule may adopt in solution)

• Individual fingerprints for each conformer

• Ensemble comparisons across all possible shapes

When we find a match between conformers, we can visualize the exact 3D alignment that explains the biological similarity. This is no longer “similar-ish” - this is molecular-level proof.

Fig. 4 The evolution of PharmPrints into PolyPharmPrints. The advance from single to ensemble fingerprinting dramatically improves match specificity.

Key Breakthrough:

We can access the exact 3D conformation of the matching conformer PharmPrint, enabling dramatically more efficient exploration of the conformational space when aligning query and hit molecules. This transforms molecular alignment from a computationally expensive global search into a targeted, efficient process. It also affords the opportunity for rapid and easy visual inspection confirming molecular superposition(s).

Case Study: The Mushroom That Thinks It’s a Statin

How computational methods revealed a hidden connection between a billion-dollar drug and an ancient mushroom used medicinally for 2,000 years.

The Starting Point

Pravastatin is a blockbuster statin drug - one of the most prescribed medications on earth. It works by blocking HMG-CoA reductase, the rate-limiting enzyme in cholesterol biosynthesis. Hundreds of millions of people take statins. Billions of dollars flowed from them annually.

Here’s the twist: the original statins weren’t invented in a lab. Lovastatin, the first in the class, was isolated from Aspergillus terreus - a fungus. Nature evolved HMG-CoA reductase inhibitors in fungi millions of years ago. We just found them and patented the synthetic derivatives.

This raised a question I couldn’t shake: if nature evolved cholesterol-lowering compounds once, where else might they exist?

The Computational Hunt

Using PolyPharmPrint screening, I analyzed Pravastatin’s 3D molecular fingerprint across its full conformer ensemble and ran it against the 28,000+ food chemicals in the FooDB database. The goal wasn’t to find something that looked like Pravastatin on paper - it was to find something that would function like Pravastatin in the body.

The Stunning Discovery

Ganoderic acid A from Reishi mushrooms (Ganoderma lucidum) emerged as a near-perfect functional match.

See the Alignment

360° rotating visualization of the 3D structural alignment between Pravastatin (blue/cyan) and Ganoderic acid from Reishi mushrooms (orange/red). The molecules have been optimally superimposed using PolyPharmpPrint technology. Explore the full interactive Drug-to-Table visualization →

What the Alignment Means

On paper, Pravastatin and Ganoderic acid A look different. But in three-dimensional biological space - where interactions with enzymes actually happen - they present nearly identical patterns:

• Hydrogen bond acceptors (the red oxygen atoms in the video) sit in matching positions

• Hydrophobic carbon chains that repel water align identically, helping both molecules anchor in the same enzyme binding pockets

• Overall 3D shape occupies the same space — like finding two completely different keys that open the same lock because they have identical functional teeth

Ancient Wisdom, Modern Validation

Reishi has been called the “mushroom of immortality” in Chinese medicine for over 2,000 years, specifically recommended for cardiovascular health and “blood stasis.” Published studies report that Reishi extracts may reduce total cholesterol, LDL, and triglycerides in both animal and human studies. The mechanism? Consistent with HMG-CoA reductase inhibition - exactly what our 3D analysis predicted.

The traditional healers didn’t have molecular modeling software. They had something different: thousands of years of careful observation. Drug to Table’s computational methods provide the molecular explanation for what they knew empirically.

The Validation Trail:
Computational: 0.98/1.0 pharmacophore similarity via PolyPharmPrint multi-conformer analysis
Traditional: 2,000+ years of documented cardiovascular use in TCM
Laboratory: Multiple peer-reviewed studies confirming cholesterol reduction in human trials
Personal: Ongoing self-experiment - standardized Reishi extract with regular lipid panel monitoring

→ Explore the full technical analysis, 3D visualization, and PolyPharmPrint methodology

The Broader Implication

If this level of molecular similarity exists between a top-selling drug and a common mushroom, how many more pharmaceutical-food connections are waiting to be discovered among the other 28,000 food chemicals - or the 400,000+ natural products in the COCONUT database?

This is why we built the tools to bring this capability to everyone. Starting with your phone.

Food Is Medicine in Your Pocket

Computational science tells us which foods have drug-like properties. But knowing that Reishi contains statins doesn’t help you if you can’t track whether you’re actually eating functional foods, in functional amounts, with measurable results. That’s the gap we built an app to close.

Food Health Screen

Your daily bridge between food science and your plate.

Food Health is the consumer-facing application of exactly the philosophy described in this article. Point your phone at any meal - a bowl of Reishi broth, a turmeric-spiced dish, a restaurant menu - and get instant AI-powered nutritional analysis: calories, protein, carbs, fat, and fiber in seconds. This is the first step.

But Food Health goes beyond calorie counting. Its scatter analysis and trend reporting let you do something that was previously only possible in research settings: correlate what you eat with how you feel, sleep, and perform - over days, weeks, and months. This is citizen science applied to your personal biology. I see this as a foundational first step to a new approach to natural product screening where anyone with a phone and a camera can participate

📸 Photo & Note Analysis AI estimates calories and macros from a photo or a typed description in seconds.

❤️ Apple Health (iOS) / Health Connect (Android) Sync See nutrition alongside sleep, activity, and heart rate — the full picture.

📊 Scatter Analysis & Reports Discover personal correlations: does turmeric improve your HRV? Does Reishi affect your sleep?

🔒 Local-First Privacy Your data stays on your device by default. You control everything.

📱Log Now, Analyze Later Capture meals even offline. Run AI analysis when you’re connected.

⌚Device Integration Custom prompts for glucose meters, blood pressure cuffs, and more.

Food Health TxD

Carbs & glucose for diabetes — spot the meals that spike you

For the 130 million Americans with diabetes or prediabetes, the drug-to-table insight is immediately urgent: every meal is a metabolic event. Food Health TxD brings AI meal scanning together with continuous glucose monitoring (CGM) data — so you can see, in real time, exactly how each food affects your blood sugar.

📸 AI meal scanning Instant carb and macro estimates from a photo

📉 Glucose overlay CGM readings and meals on the same timeline

📋 Menu scanner Scan a restaurant menu, pick the lowest-carb options

❤️ Apple Health (iOS) / Health Connect (Android) sync Reads glucose, writes carbs automatically

📊 Trend analysis Spot patterns between meals and spikes over weeks

As I wrote in May 2026: “Half of American adults are prediabetic or diabetic. This isn’t a headline - it’s a quiet emergency happening in every household.” TxD was built to give people a simple way to see how food affects their glucose, one meal at a time.

Food Health TxD is for education and tracking only. It is not a medical device and does not provide diagnoses, insulin dosing recommendations, or clinical predictions. Consult your healthcare provider before making changes to your treatment plan.

Your Personal Drug Discovery Lab

The computational pipeline that produced the Pravastatin-Reishi discovery is not locked in a university basement. We may just integrate the pipeline into the Food Health scan in Food Health TxD apps in future versions. But if you are interested in going at it yourself - with open-source tools and public databases, you can run versions of this same analysis - or, at the very least, use its outputs to make more informed daily food choices.

Target Selection

Choose a biological pathway you care about: cholesterol metabolism, blood sugar regulation, inflammation, cognitive function, sleep architecture. Pick a known drug that addresses it - that’s your computational starting point.

Computational Screening

Generate pharmacophore fingerprints from the drug. Several open source pharmacophoric generation utilities are available. Screen natural product databases (FooDB’s 28,000+ food chemicals; COCONUT’s 400,000+ natural products; ChEMBL’s biological activity data). Apply QSAR predictions to rank candidates by predicted activity. Cross-reference with traditional use and safety literature.

Systematic Exploration

Source the top candidates - ideally standardized extracts with known compound concentrations. Design a personal protocol: baseline biomarker measurement, defined supplementation period, repeat measurement. Track everything with Food Health Scan.

Data, Not Anecdote

Use Food Health’s scatter analysis to correlate your intake with objective metrics: sleep quality, heart rate variability, energy levels, blood glucose (with TxD if relevant). This turns personal experimentation into structured self-science.

The Spice Cabinet Pharmacy

The Pravastatin-Reishi case isn’t an isolated curiosity. Drug to Table screening of the top 200 pharmaceuticals is ongoing, and provisional findings suggest the spice cabinet is far more medicinally dense than most people realize:

• Turmeric (Curcumin): Computational analysis reveals similarity to multiple COX-2 anti-inflammatory drugs — the same target as ibuprofen and celecoxib, without the gastrointestinal side effects at culinary doses.

• Garlic (Allicin): Pharmacophore patterns consistent with ACE inhibitors used for blood pressure management. Traditional cardiovascular use across dozens of cultures suddenly makes molecular sense.

• Ginger (Gingerols): Structural relationships to 5-HT₃ antagonists — the class of drugs used clinically for nausea. Why does ginger settle an upset stomach? Now we can show you exactly why.

• Cinnamon (Cinnamaldehyde): Computational similarity to insulin sensitizers — consistent with traditional use of cinnamon to moderate blood sugar. Food Health TxD users: this is one worth tracking.

• Reishi (Ganoderic acids): Our flagship discovery — near-perfect pharmacophoric overlap with Pravastatin. The “mushroom of immortality” earns its name computationally.

Traditional Medicines Examples

Computational methods are now providing the molecular explanations for what traditional healers discovered empirically over millennia:

• Willow Bark → Aspirin: The original pharmacological natural product -salicylate’s mechanism is now textbook biochemistry

• Foxglove → Digitalis: Computational analysis may explain why this traditional heart remedy modulates cardiac contractility

• Berberine (Goldenseal, Barberry): Shows metformin-like computational profiles for glucose regulation - of strong interest to Food Health TxD users

• Reishi → Statin-like effects: Our current discovery validates and molecularly explains 2,000 years of cardiovascular use

Important context: Computational similarity does not equal clinical equivalence. Bioavailability, metabolism, dose, and individual genetics all matter enormously. These findings suggest promising directions for personal investigation - not prescriptions. Always consult your healthcare provider about any significant health concern.

💡 The Network Effect: Community-Driven Validation

As more people adopt these methods and use tools like Food Health to track their nutritional intake and health correlations, we build collective intelligence:

• Shared protocols for systematic personal investigation

• Open databases of personal results and meal-to-metric correlations

• Collaborative refinement of computational models

• Crowdsourced identification of effective natural approaches — validated by real human data, not just algorithms

The Democratization Opportunity

“For the first time in history, the same computational intelligence used to discover billion-dollar drugs is available to anyone - applied to natural compounds that belong to everyone.”

This isn’t about replacing pharmaceutical innovation. It’s about expanding the toolkit available for understanding how food choices might influence biological systems. And critically, it’s about closing the gap between discovery and action.

The Drug-to-Table computational methodology identifies candidates. Food Health makes the daily practice of food-as-medicine measurable. Together, they represent a new paradigm: molecular intelligence meets personal health data.

For the first time in history, the computational intelligence used to discover billion-dollar drugs is accessible to individuals. Not the full pharmaceutical apparatus - but the core intellectual methods: pharmacophore fingerprinting, QSAR modeling, open molecular databases.

And now, the consumer interface for living those discoveries is in your pocket. Food Health AI closes the loop between:

• 💡 Computational insight (Drug to Table methodology)

• 🍽️ Daily food choices (what you actually eat)

• 📊 Measurable outcomes (biomarkers, trends, correlations)

This is not Big Pharma. This is your pharmacy. The same computational power that generated patent portfolios worth billions can now help you make better decisions at the grocery store, at the spice market, and at your kitchen table.

“We’re not discovering that foods contain drug-like molecules. We’re remembering that drugs came from food in the first place - and building the tools to act on that knowledge.”

What’s Next

The Pravastatin-Reishi connection is one data point in what will become a publicly accessible atlas of drug-food molecular relationships. Here’s what’s coming:

• 📡 Complete computational protocols for Drug-to-Table discovery

• 🍄 Results from screening the top 200 pharmaceuticals against 28,000+ food chemicals

• 🧪 Personal testing frameworks for systematic self-experimentation (with Food Health AI integration)

• 🌐 Community database for sharing computational hits, personal protocols, and biomarker results

• 📱 Planned Food Health AI features: functional food tagging, compound tracking, drug-food interaction alerts

All protocols, databases, and methodologies will be published as open resources. The goal is to make molecular nutrition intelligence accessible to anyone curious enough to use it.

Follow along:

• 📖 Substack: drstevenmuskal.com

• ✍️ Medium: @smuskal

• 🔬 Drug to Table Technical Hub

• 📱 Food Health AI (iOS & Android)

• 📱 Food Health TxD (iOS)

⚠️ Important Disclaimer

This article represents computational research methodology for educational and wellness exploration purposes only. It is not medical advice.

Any health concerns should be addressed with qualified healthcare professionals. Natural compounds can have potent biological effects and potential interactions with medications — particularly important if you are currently taking statins or other prescription medications.

Correlational findings: Pharmacophore similarity does not guarantee identical biological activity. Actual effects depend on bioavailability, metabolism, dosage, individual genetics, and physiology. Structural similarity is a hypothesis-generating tool, not a clinical proof.

Natural ≠ Safe: Many plants contain potent bioactive compounds. Always consult a healthcare provider before using supplements alongside prescription medications.

Food Health and Food Health TxD are consumer wellness tracking applications, not medical devices. They do not provide diagnoses, clinical predictions, or treatment recommendations.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple mixes, I wanted to pull teo from the archives. The first one was fun, especially if you know the back story. TimD vocals/guitar, Andrew Guitar, Rick Vocals, AlanW Bass, and Johnny Conga. The Way (first and only pass):

Then during a July 3rd mix (as we start planning for one this year) - The Story with neighbors. Maya leading vocals, Tammy and Pam assisting vocals, Dom Guitar, Jesse Guitar, Don Keys, TimH Bass:

Prologue: A Conversation at Cal-a-Vie

On a recent stay at Cal-a-Vie, my go-to spot just around the corner, I met a wonderful family radiating incredible positive energy. I introduced myself as Steve. And as often happens in my conversations, we eventually got to talking about AI.

We went around the group: who uses it, how, why. When they got to me, I said: “Yeah, I’ve been doing this for decades, and I actually built an AI version of myself. I built AI-Steve.”

For the remaining few days of the trip, every time they saw me after that, they called out: “Hey, AI-Steve!”

That group and those conversations inspired the recent piece “Too Big to Fail”, on dopamine-driven activities and keeping your instincts sharp as you age.

This article is the longer version of that conversation.

Part 1: The WHY. Dopamine-Driven Curiosity

I want to be honest about something from the start. I’ve never been particularly strategic about my career.

The things I build: I build them because they don’t exist yet, and because not having them bothers me. That’s the full explanation. Call it dopamine-driven curiosity. You encounter a problem. The absence of a solution becomes intolerable. You build something. And then the next problem finds you.

The projects I’m about to describe didn’t emerge from a business plan. They emerged from that impulse. From the question “why doesn’t this exist?” followed by the only logical next step.

But here’s the interesting thing, looking back over four decades of this: some of those impulses, the world eventually caught up to them.

Part 2: The Pattern. Skating to Where the Puck Is Going

Wayne Gretzky said the thing that everyone quotes: “I skate to where the puck is going to be, not where it has been.”

I don’t always do this with intention. But when you live on the bleeding edge long enough, eventually some things come into complete alignment.

I started working with neural networks in the mid-to-late 80’s. That was before most people knew what the phrase meant. Before deep learning. Before transformers. Before large language models capable of genuinely flexible reasoning. “Neural network” was, at the time, a research curiosity: something you had to explain at dinner parties rather than something everyone’s iPhone contained.

In 1991, I completed my PhD thesis. The title: “Predicting Features of Protein Structure with Computer-Simulated Neural Networks.”

Let that sit for a moment. Thirty years later, AlphaFold would win the Nobel Prize for solving this problem: protein structure prediction, at a scale and precision that changed biology forever. The 2024 Nobel in Chemistry went directly to the work of Demis Hassabis and the DeepMind team for what AlphaFold achieved.

When that Nobel was announced, I want to be clear about what I felt. Not resentment. Not “I was there first.” Not a need for credit or recognition or a footnote in history.

It was ecstatic personal affirmation.

My thesis was right. The direction was right. The world caught up. The field benefited enormously. That is the most satisfying version of this story, and I mean that without irony. The satisfaction is internal. The hope is always that others benefit. The work exists now, in the world, doing good. What more could you want?

This is the pattern I’ve noticed: dopamine-driven curiosity, pursued relentlessly at the bleeding edge, sometimes accidentally aligns with where the world is heading. It’s not prescience. It’s not genius. It’s simply what happens when you follow the impulse long enough.

The same pattern appears in everything I’m about to describe.

Part 3: The Projects. Brief and Human

Toast Our Friend: The Seed

Every project I’ve built has started with a question I couldn’t leave alone. This one started with grief.

A few years ago, at my sister’s celebration of life, hundreds of people came from across the country, sharing stories, laughter, and tears in equal measure. I sat there thinking: she would have loved this. I only wish she were here to see it.

That thought, we should tell people what they mean to us while they’re still here, became the seed of everything that followed. And the very first thing I built in response to it was Toast Our Friend: a platform for celebrating people while they’re alive, and for honoring them after they’re gone.

You create a circle for someone: a friend fighting an illness, someone you want to honor, someone you love who’s moving on. Friends and family can post toasts: stories, memories, photos, voice messages. Real-time. Visible to everyone in the circle. Not a eulogy. Now.

We’ve created circles for Scott Miller, Mike Szwajkowski, Tom, and Coop: friends I’ve known for years. Some for decades, others since the first day of high school in 1980. Mike was my longest and most cherished friend. Nearly half a century of friendship. When he passed, I wanted his people to have a place to gather, remember, and honor him together.

The platform was built with AI assistance alongside other tools. But the meaning came from somewhere older and deeper than any technology I could build.

AI-Dad: What Toast Pointed Toward

Toast answered the urgent question — why are we waiting? But it also pointed toward a quieter, second question: what about after? What if, in addition to gathering memories around someone, we could preserve the voice itself — not as a static archive or a memorial page, but as something a grandchild could still talk with?

That question became AI-Dad, the natural follow-up to Toast.

AI-Dad is an experiment in preserving a person’s voice as a conversational system. Grounded in their history, their turns of phrase, their particular way of reasoning through a problem. Memory, voice, and ethical design: those three constraints shaped everything. It remained a research project, an exploration. But it planted the question that came next.

For a technical deep dive on the AI-Dad concept, see the AI-Dad article on Medium.

Food Health: Food Is Medicine

I believe food is medicine. I’ve believed that for a long time, long before it became a wellness industry catchphrase. But for me, this isn’t a philosophy I read about somewhere and adopted. It’s a scientific conviction I can defend at the molecular level. That distinction matters.

I’m a Ph.D. chemist. Two of my early career positions put me squarely at the intersection of nature and pharmacology in ways that shaped everything I’ve thought about food and medicine since. At Affymax Research Institute, and before that at MDL Information Systems (formerly Molecular Design Limited), I was working at the center of natural product screening as a drug discovery strategy. This was not a niche academic exercise. Natural product screening is, in fact, the foundational strategy behind the discovery of virtually every major drug class in existence.

Here is the core insight that I absorbed during those years, stated plainly: nature has spent billions of years solving therapeutic problems. Evolution is the longest-running drug discovery program in history. Organisms develop compounds to defend themselves, to communicate, to modulate their own biochemistry. And those solutions, honed across geological timescales, often turn out to be exquisitely effective at interacting with human biology too.

The pharmaceutical industry’s model, at its most basic, has largely been this: identify what nature already solved, then alter the molecular structure just enough that a patent examiner would recognize it as novel. Preserve the biological activity. Create an intellectual property position. That’s the model. It’s how most of the drugs you’ve ever taken came into existence.

The idea that plants and other natural substances contain medicine was once treated as folk wisdom, even dismissed as fringe thinking. It is now the bedrock of pharmacology. The field caught up to what traditional healers knew for centuries, and what the chemistry eventually confirmed.

I’ve illustrated this directly in my own research. The Reishi mushroom/statin structural similarity post makes this concrete: compounds found in Reishi mushrooms share meaningful structural similarity with the statin drug class, one of the most widely prescribed and profitable pharmaceutical categories in history. The structural overlap isn’t coincidence. It’s nature having arrived at a solution first.

This is precisely what PharmPrint/Polypharmacophore fingerprinting was built to find. PharmPrint is a computational technology I developed, using 3-point and 4-point pharmacophore triplets encoded as 10,000-bit fingerprints, to map where nature’s solutions overlap with pharmaceutical chemical space. The technology takes a molecular structure and identifies which pharmacophoric features it shares with known drug-active compounds. It is, in the most literal sense, a tool for asking: where has nature already solved a problem that we’re still working on?

Connecting PharmPrint to food isn’t a stretch. It is the natural extension of the same logic. The same structural analysis that reveals why a Reishi compound resembles a statin can be applied to any food-derived compound. Drug to Table is the direct expression of this: finding therapeutic compounds hidden in food using polypharmacophore matching across natural products and drug targets.

So when I say food is medicine, I’m not quoting a wellness slogan. I’m drawing on forty years at the intersection of chemistry, pharmacology, and molecular modeling. I’m saying: the structural evidence is there. The computational tools exist to find it. And the apps I’ve built, which I’ll describe next, are the practical expression of that conviction in the hands of everyday people.

Food Health: Scan & Score lets you photograph a meal and get full nutritional data, including calories, macros, fiber, and protein, in seconds. It launched in early 2026 and is live in the Apple App Store (App ID: 6759629101) and on Google Play.

Food Health TxD is a diabetes-aware companion app, designed for people managing blood sugar. It pairs real-time carb and macro estimation with continuous glucose monitor data, giving metabolic context to food choices in the moment they’re being made.

Shared Health Projects is a social health platform: a collaborative layer for sharing health journeys with people you trust, whether a partner, a physician, a coach, or a research group. I made the source available on github.

These apps came together through a combination of tools and approaches. AI-Steve was one part of that process, helpful in generating and reviewing code, but not the sole or exclusive platform for their development. If you want the full technical picture of how it all worked, I’ve written about the architecture in detail, and the links are at the bottom of this article.

AI-Steve: The System

The most unexpected outcome of building AI-Dad was the question it naturally generated next: what if I built that same thing around my own life, in real time?

Not a memorial for the future. An operating system for the present.

That became AI-Steve.

AI-Steve runs continuously on my machine. It ingests and indexes my emails (over 213,000), my iMessages, my calendar, my Apple Health data, my photos, my documents, and decades of research. Over 314,000 chunks of my life, semantically searchable in under a second. I can ask it anything in plain English and get back an answer grounded in my actual history, not a guess.

The technical architecture, including the RAG pipeline, the embedding approach, and the hybrid retrieval design, is worth understanding if you’re building something similar. I’ve described it in detail in AI-Steve Deep Dive and the follow-up piece AI Steve II: The Governed Loop. I’ll point you there rather than recreate those here.

What I want to talk about in this article is the why, and the thing about AI-Steve that I’ve thought about most.

The Governed Loop is the formalization of something I noticed after the basic system was working. AI-Steve, given enough context about its own architecture, could propose extensions to itself. It could identify gaps in its own capabilities. Recommend new data sources to ingest. Suggest new features.

But it doesn’t execute these autonomously. Every proposal comes to me as a structured issue: what it wants to build, why, what it would require. I review, approve, reject, or modify. Then the system gets built.

This isn’t just an interesting engineering pattern. It reflects something I genuinely believe about how AI should work right now: the human stays in the loop. Not as a bottleneck. As the decision-maker. The AI amplifies judgment; it doesn’t replace it.

That’s not a technical constraint I imposed reluctantly. It’s a design philosophy I chose deliberately.

Part 4: The Bigger Picture

I’m not writing this to take credit for anything.

The thesis is thirty-five years old. AlphaFold deserved its Nobel, and the world benefited from it in ways I never could have accomplished alone. The apps are in the store. The articles are out there. If any of this is useful to anyone, that’s the point.

What I believe is this: we’re at an inflection point. AI isn’t just a productivity tool anymore. It’s becoming cognitive infrastructure. And the questions around how we build that infrastructure, whose values it reflects, how much human judgment it preserves, whether it serves life or colonizes it, are not technical questions. They’re deeply human ones.

The metabolic crisis is real. I’ve written about this separately, and I’ll keep writing about it, because food and health and the ways we’ve engineered ourselves into poor metabolic health is one of the defining crises of this era, and AI can be one part of addressing it.

The disconnection crisis is real. Toast Our Friend exists because I believe we’ve lost something in how we honor and celebrate each other, and because the tools exist now to create something better.

The question of legacy, of how we preserve what we know and who we were for the people who come after us, is real too. AI-Dad was a first attempt at an answer. I suspect it won’t be the last.

I’ve been skating toward these problems for a long time. Not always knowing where the puck would end up. But following the impulse.

Eventually, some things come into alignment.

This Story Is for You

Before I close, I want to say something directly. These posts are not just a chronicle of what I’ve built. They’re an invitation.

The thread running through everything I’ve described is not technical skill. It’s insatiable curiosity, pursued consistently. That curiosity is generative. If you have that bug, you can produce at this level too. I mean that. Not as encouragement for its own sake. As a statement I believe is genuinely true.

The tools available today mean that non-technical specialists can now build technologies that were previously only accessible to well-funded teams. AI assistants, app frameworks, voice cloning, rapid prototyping: what used to require a team, a budget, and months of runway can now be initiated over morning coffee with a paragraph of text. We used to call it RAD, Rapid Application Development. What we have now is something faster than that, and far more accessible.

I’ve spent forty years as a scientist and builder. I’ve seen technology cycles. I’ve seen what it looks like when a genuine platform shift arrives. This one is different. The barrier to entry has not just lowered. For certain kinds of work, it has effectively disappeared.

The age of the solopreneur is not something that’s coming. It is here, today and tomorrow. If you have a curiosity you’ve been sitting on, a problem you’ve been annoyed that nobody has solved, a domain where you have deep expertise and can see the gap clearly: the window is open now. Not in five years. Not when the technology matures further. Now.

I share none of this to inspire admiration. I share it because I genuinely hope that when you read about what one person has built at the intersection of AI, chemistry, health, and human connection, something clicks. That you look at your own curiosities and recognize what’s possible. That you ask the question I’ve been asking for forty years: why doesn’t this exist? And then you take the next logical step.

That’s the whole invitation.

What’s Next

AI-Steve keeps growing. The Governed Loop means it’s always proposing what comes next, and I’m always deciding.

If you want to follow along, subscribe on Medium and on Substack. I write about this work at the intersection of AI, health, and what it means to build things that matter.

And if you’re building something similar, or thinking about it, I’d genuinely love to hear from you.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

Related reading:

• Too Big to Fail: on dopamine-driven activities and keeping instincts sharp as you age

• AI-Dad: Preserving Legacy Through Conversational Intelligence

• AI-Steve Deep Dive

• AI Steve II: The Governed Loop

• Food Is Medicine. But First It Has to Become Data.

• The Metabolic Crisis Is Not a News Story

Projects mentioned:

• Food Health: Scan & Score, App Store | Google Play

• Food Health TxD, iOS

• Shared Health Projects, iOS

• Toast Our Friend

• PharmPrint / Polypharmacophore

• Drug to Table

• stevenmuskal.com

The Long Arc

In the late 1980s, I was spending twelve to sixteen hours a day inside neural network systems at UC Berkeley. Not using them. Building them. My graduate work involved developing tools to make neural network training more efficient, contributing to the foundational methodology that would, decades later, quietly underpin a great deal of what we now call artificial intelligence. The work was deliberate, deep, and computationally demanding in ways that are almost quaint by today’s standards. We were doing significant things on hardware that modern inference workloads would run circles around without breaking a sweat.

There is a meaningful distinction that tends to get lost in current media coverage of AI: training a model and using one are categorically different activities. Training requires massive GPU clusters, enormous datasets, industrial energy consumption, and months of coordinated effort. Inference, using the trained model to respond to a query, is comparatively lightweight. That asymmetry has enormous implications for how AI spreads and scales, but the coverage rarely distinguishes between the two, which leaves most people with a distorted sense of what the resource footprint actually looks like. I spent my Berkeley years in the heavy-lifting phase. I know the difference from the inside.

After the methodology was built, the dynamic shifted. Leveraging it required far less time than constructing it had. Applications came quickly. You build the engine once; you drive it many times. Over the decades that followed, my engagement with AI systems followed that pattern: significant front-loaded investment, followed by relatively efficient deployment. The curve flattened. The ratio of time-inside to output improved dramatically. This mirrors the broader arc of any foundational technology, where the bulk of the effort is in the establishment, not the execution.

Then, roughly three years ago, something inverted.

With the arrival of mature large language models and CLI-accessible tools, my time spent inside AI systems surged back toward early-career levels. Not building, this time. Using. Claude CLI running with ` — dangerously-skip-permissions`, Codex working through complex engineering tasks, Factory AI’s Droid orchestrating multi-step workflows, and Gemini CLI configured to `approve-all` to maintain uninterrupted flow state during demanding sessions. Hours that used to look like engineering have been replaced by hours that look like… something harder to categorize. The inputs are different. The cognitive mode is different. But the time commitment has returned to something that would have been familiar to the version of me sitting in front of a workstation in Berkeley in 1988.

That U-shaped curve, enormous time, then much less, then enormous again, is precisely where this becomes interesting to think about. Because I am not a newcomer arriving at AI with fresh eyes and no reference point. I am someone who has spent four decades living inside these systems in various forms. And that, it turns out, is exactly what makes the draw observable in ways it might not otherwise be. The hook is most visible to someone who has seen it from multiple angles over a very long time. Not despite the expertise, but because of it.

Dopamine Businesses

Let me say directly what I think is happening, because the framing matters more than any particular observation I can offer.

Dopamine businesses are not a new invention. Dopamine, the neurotransmitter that drives anticipatory wanting rather than simple pleasure, is as old as human biology. We have always used it on each other. Gambling is a dopamine business. Alcohol is a dopamine business. Narrative fiction, gossip, ritual, romance, commodity markets, the architecture of casinos designed to remove your sense of time passing, the engineered unpredictability of a slot machine’s reward schedule: all of them operate on the same underlying circuitry. The circuit evolved because it was adaptive. It pushed our ancestors toward food, toward social connection, toward novel information, toward the kind of environmental engagement that increased survival odds in a world of genuine scarcity.

What changed is not the existence of dopamine businesses. What changed is the engineering precision and the stimulatory load. These are related but not identical problems.

Our evolved inhibitory systems, the internal braking mechanisms that produce satiation, aversion to excess, the sensation of having had enough, are real and genuinely functional. They are not easily overridden under conditions resembling those in which they evolved. But they developed under a stimulatory environment that was, by any modern standard, quiet. Variable, unpredictable, often lean, and never continuously optimized to keep the seeking behavior running. The problem is not that the brakes are broken. The problem is that the gas pedal is now operated by systems that were engineered, specifically and at considerable expense, to apply pressure just below the threshold where braking activates. The result is a sustained, low-grade override of circuits that were never calibrated for continuous load.

Two colleagues recently published a paper on modeling cognitive overload that I think is a serious first step toward putting a quantitative frame around this problem: [The Bias Amplification Model A Mathematical Framework for Quantifying Cognitive Distortion Under Conditions of Information Overload — and their model is here: The Bias Amplification Model] The instinct to measure it is exactly right. You cannot manage what you cannot observe. The challenge is that most of us are currently operating with no real instrument for reading our own cognitive load, which means we are flying without gauges in conditions that have changed more in the last decade than in the prior century.

The Same Mechanism, Different Substrates

The most clarifying thing I can say about the current AI moment is that the operating principle behind what Big Tech built is not metaphorically similar to what Big Food built before it. They are the same mechanism applied to different substrates.

Big Food spent decades, from roughly the 1950s onward, engineering hyper-palatable products. Food scientists calibrated fat-to-salt-to-sugar ratios with extraordinary precision to maximize palatability while minimizing satiety. They engineered texture profiles that felt compelling without triggering the body’s stop signal. They discovered the bliss point, the sweet spot of sweetness or saltiness that generates maximum hedonic response, and optimized relentlessly toward it. The goal, stated plainly, was to override the body’s natural feedback loop that says enough. Not through deception, exactly, but through engineering that was specifically designed to outpace the biology it was operating on.

The metabolic consequences of that project are still unfolding. I wrote about this at length in a piece on the metabolic crisis (Substack version here), but the short version is that somewhere between 88 and 93 percent of American adults are now metabolically compromised by at least one risk factor. This is not a number that comes from a single dietary mistake or an individual failure of willpower. It is the aggregate output of an industry that spent seventy years engineering food to override the satiety signals that keep metabolic health intact.

Big Tech learned from that playbook and applied it to cognition. The medium changed; the strategy did not. Engagement loops built on variable reinforcement schedules, designed to be just compelling enough that you stay but never so satisfying that you leave fulfilled. Recommendation algorithms that learned your individual response profile and calibrated content delivery to extend session time. Notification architectures engineered to interrupt your attention at precisely the moment when re-engagement was most likely. These are not rough approximations of behavioral influence. They are precision instruments, developed by teams of behavioral scientists, running on infrastructure that updates in real time based on your response.

TikTok is the clearest specimen of what this looks like at full development. The recommendation algorithm does not weight whether you said you liked something. It tracks duration watched, completion rate, sound engagement, scroll hesitation before a restart, the particular speed of your thumb as you move through content. Every one of those signals feeds a system that is continuously updating its model of what will extend your session. The content is almost incidental. It is the carrier wave for a psychological feedback loop that has been built with billions of dollars of engineering investment. Facebook Reels followed nearly the same architectural blueprint. You do not choose your echo chamber on these platforms. The platform builds it around you, in real time, tuned to your individual response profile. The algorithm does not know you. It does not need to. It knows your behavioral signature well enough to accomplish the same thing.

The important thing to hold onto here is that these are not incidentally addictive products that happened to capture people’s attention. They are systems that were designed, iterated, tested, and refined specifically to maximize the duration and intensity of engagement. The dependency is a feature, not an externality.

AI as Dopamine Business

I want to be precise about this, because I think the standard framing undersells what’s actually happening.

The claim is not that AI companies sat in a room and said “let’s make something addictive.” Some of what I’m describing is structural rather than deliberate. But structurally, a well-designed LLM interaction generates exactly the patterns that sustain engagement loops, and it does so repeatedly and reliably enough that the effect is real regardless of intent.

Consider the mechanics. An incomplete response that leaves a thread unresolved, creating a pull to follow up. A step-by-step technical walkthrough that requires sustained engagement across multiple iterations before any deliverable appears. An overwhelming first-pass answer that is genuinely useful but also genuinely too large to absorb, which prompts you to ask for synthesis, starting a new loop. For those of us oriented toward critical thinking, the occasional wrong answer that you catch, and then find yourself interrogating, pulling you deeper into the exchange in the course of correcting it. Each of these is a mechanism. Each creates an incomplete reward cycle that sustains engagement precisely because it is incomplete.

I run Anthropic’s Claude, OpenAI’s Codex, Factory AI’s Droid, and occasionally Google’s Gemini, all via command line, all as deeply integrated layers in my daily work. This is not casual or exploratory usage. These tools are woven into the architecture of AI-Steve (Substack version), a comprehensive personal AI system I’ve built and documented that runs across my email, calendar, health data, code, and knowledge base. The second piece in that series (Substack version) covers how the system now proposes its own extensions, which should give you a reasonable sense of how far down this path I have gone and how deliberately I have gone there.

The economics of this are also worth naming. These systems often begin free and escalate. Heavy usage across multiple platforms is not unusual at $200 per month per subscription, or more. That spending feels justified because the value is real. But the pattern of spending, the shape of it, tracks something that goes beyond a rational productivity calculation. I recognize the pattern from other contexts. The value justification and the dependency are not mutually exclusive. They coexist quite comfortably, which is part of what makes it worth paying attention to.

The draw is genuinely bidirectional, and this is important. The model pulls you in through its mechanics. But your own psychology pulls you toward it at the same time. The more capable the system, the more problems you bring to it. The more problems you bring to it, the more time you spend in session. The more time you spend in session, the harder it becomes to imagine working at the same pace without it. That loop has an acceleration to it that is not purely about the tool. Some of it is about what happens to the user.

Session Saturation and the Successor Prompt

There is a practical dimension to heavy AI usage that does not get discussed enough, in part because it requires extended experience to observe clearly.

These systems degrade during long sessions. Not catastrophically, and not in ways that are always obvious in the moment, but in ways that accumulate. If you are deep in a working session of two or three hours, the model begins to show what I can only describe as fatigue. It starts revisiting ground that was already covered. It generates responses that are technically coherent but increasingly approximate. It loops back to early-session framings even when the conversation has moved past them. The context window is working hard with more history than it handles gracefully under sustained load. The session is saturated.

When I notice this happening, I use what has become one of the more reliable techniques in my workflow. I give the model a prompt approximately like this: ”Fully document everything we have done so far and generate a prompt for your successor. I want to start a new session.” The model produces a comprehensive handoff document. I close the session entirely, open a new one, and paste that document as the opening message. The new instance has the benefit of context without the accumulated weight. It is, as I have come to think of it, giving the AI a night of sleep. The clarity difference when you return to a problem that way is immediate and noticeable.

There is a related observation I am somewhat less comfortable sharing, but it is accurate. I have found that off-hours usage tends to yield meaningfully better results, responses that feel less throttled, more precise, less prone to the fuzzy approximations that suggest something is being managed on the backend. Peak usage periods seem to correlate with subtly degraded quality. So I have started timing my most demanding sessions for early morning or late at night, when the systems are less congested.

I recognize what this sounds like. It sounds exactly like someone timing their purchases to get better quality at lower friction. The junkie who knows which hours the dealer is freshest and plans accordingly. I am not being dramatic. The behavioral pattern is structurally identical. The fact that the activity is productive does not change the shape of the dependency, and I think the honesty of that comparison is worth sitting with rather than deflecting.

Biological Inhibition and the Overwhelmed Brake

Here is something that consistently gets underappreciated in discussions about digital dependency: our inhibitory mechanisms are real, evolved, and functionally capable. The problem is not that they are broken. The problem is the environment they are operating in.

The wanting system and the braking system are paired. Dopamine drives approach. Satiation, aversion, discomfort with excess, the physical signal that says stop, these counterweights are built in. They work. They kept human beings from bingeing themselves into dysfunction across hundreds of thousands of years of evolutionary history, in environments where stimulation was scarce, variable, and predominantly low-intensity.

That environment is gone. The systems we have engineered, processed food, algorithmic social media, AI engagement loops, are high-intensity, continuous, and deliberately calibrated to operate just below the threshold where the braking mechanisms activate. This is not accidental. It is the target. Build something compelling enough to sustain engagement but not so aversive that it triggers a withdrawal response. The result is a sustained low-grade override that is genuinely novel in evolutionary terms. Our ancestors never faced anything like it at this scale or consistency.

The cognitive overload modeling work referenced [here] is trying to put a quantitative frame around what happens when that load accumulates. That is exactly the right direction. The challenge is that most of us are currently managing without any real read on where our cognitive load sits, which means we are making trade-off decisions without instruments. The Bronze Age analogy I return to: warriors of that era wore armor calibrated to the weapons they faced. When the weapons changed, the armor became inadequate. Not because armor was a bad idea, but because the threat had outpaced the protection. We are in a comparable situation. Our evolved inhibitory systems were calibrated to a prior environment. The dopamine businesses now operating at industrial scale are a different category of threat, and managing them with unmodified biology is asking the wrong tool to do the job.

The Glasses Problem, and What Hormetic Stress Has to Do With It

I have worn corrective lenses for most of my adult life. During engineering school, my prescription kept escalating. Not because anything medically unusual was happening, but because I was in front of a workstation for twelve-hour stretches, and my eyes were adapting to continuous close-focus demand. Each new prescription would feel right for about a week, then my vision would adjust around the correction and the glasses would feel insufficient, and the cycle would repeat. The tool was recalibrating my biology in the wrong direction. The correction was enabling a feedback loop that made the underlying capacity progressively weaker at operating without the support.

At some point I started doing something deliberate. I stopped reaching for corrective lenses reflexively. During portions of the day when the stakes of acuity were manageable, I let things stay slightly out of focus rather than immediately compensating. This was uncomfortable in a predictable way. But over time something interesting happened: my vision adapted. Not to perfect acuity, but to meaningfully better function without constant correction. The underlying capacity, deprived of the crutch, had room to operate and to recover some of what the crutch had been compensating for.

This is a well-documented principle in physiology under the name hormetic stress, the idea that strategic, controlled withdrawal from a support system can strengthen the underlying capacity that the support was masking. Intermittent fasting, cold exposure, graduated resistance training: all operate on some version of this. The temporary discomfort is not the point. The adaptation is.

I think about the glasses problem with some regularity when I consider AI dependency. The risk is not that AI makes my thinking worse during a session. By most measures, it makes it better, faster, more comprehensive. The risk is the longer-term trajectory: whether continuous delegation gradually atrophies the underlying capabilities being delegated. Not suddenly. Not obviously. In the way that a muscle you stop loading eventually loses the functional range you had come to take for granted. You don’t notice until you need it.

For any system where biology is involved, where the support is operating on top of a biological capacity rather than simply replacing a mechanical one, the question of what happens to the underlying capacity under conditions of sustained offloading is a serious question. It does not resolve itself just because the support is valuable. The glasses escalated while delivering genuine utility. The utility and the atrophy were coexisting.

The practical implication is not to abandon the tools. It is to build in deliberate resistance: working through important problems in your own head before asking the model, checking your intuition before comparing it to the AI’s framing, maintaining the habit of thinking hard about something without autocomplete for your thought process. Not as a purity position. As maintenance.

The “Too Big to Fail” Threshold

We have seen this transition before, in different contexts.

Financial institutions, somewhere in the decade before 2008, crossed a threshold where their interconnection with the rest of the economy had become deep enough that their individual failure was not just a problem for them. It was a systemic destabilizer. “Too big to fail” does not mean deserving, or right, or particularly well-managed. It means the cost of the absence has grown larger than the cost of the continued presence. The institution has become structural. Removing it does more damage than sustaining it.

Certain platform businesses crossed a version of this threshold around the same period. When commerce, communication, identity verification, and community infrastructure are routed through a single platform at sufficient scale, opting out is no longer a meaningful option. The decision is no longer about whether you endorse the product. It is about whether leaving is less disruptive than the problems that come with staying. The platform has become infrastructure. Infrastructure does not get evaluated on its merits in the usual way.

AI is following the same trajectory. It is doing so faster than either of those precedents, and across a broader set of domains simultaneously, which is what makes this moment genuinely different from the comparable moments in fintech or platform media.

Consider the concurrent integration: AI is now load-bearing infrastructure in software development, where its absence would materially slow output across the industry. It is embedded in business decision support in ways that are difficult to disentangle from the decisions themselves. It is a functional creative partner in content, media, research, and communication. For individuals operating across multiple disciplines, as I do, it is a thinking amplifier that has been woven into the cognitive architecture of daily work. None of these applications, taken alone, crosses the threshold. Together, the accumulation changes the character of the dependency.

The critical threshold is not about capability. It is about when the absence of something becomes more destabilizing than its presence. I believe that threshold has been crossed, without ceremony, without a moment anyone could point to, and without the coordinating frameworks that might help us understand what we are navigating. It happened in the accumulation of individual dependencies, each of them reasonable in isolation, that collectively produced something with a very different character.

The Illusion of Control

There is a quality to working inside AI systems that feels like a tight feedback loop between your intent and the output. You frame the prompt. It responds. You refine. It adjusts. The iteration is fast and the responsiveness is real. It feels, in the best moments, like a very capable extension of your own thinking.

But underneath that feedback loop, you are operating inside a system with constraints you do not fully see and did not negotiate. Training data assembled by decisions you were not part of, reflecting priorities that were not yours. Model updates deployed on schedules you do not set, in ways that are not announced, that shift behavior in ways you might not notice unless you have been using the system long enough to have built a baseline. Safety filters and content policies that shape outputs in ways that are invisible until they are not. Architectural choices made years ago, by teams working with assumptions about what a useful AI should do that may or may not match what you actually need. You are operating a complex system to which you do not have access to the hood. That is fine for most purposes. You do not need to understand the combustion cycle to get where you are going.

It becomes less fine when the system is load-bearing. When your strategic thinking, your code, your communications, and your framework for evaluating new problems are all routed through infrastructure you cannot fully audit, the trust is implicit and largely unexamined. I am not suggesting that should be a reason to stop, but it should at minimum be conscious. The distinction between using a tool and being embedded in infrastructure is not one we tend to make in real time. It tends to become visible only in retrospect.

What AI Does to Gut Instinct

This is the part that concerns me most in the long run, and it is also the most difficult to measure, which is itself part of the problem.

The best decisions I have made in my career, the ones that look cleanest in retrospect and held up under conditions I could not have anticipated, rarely came from analysis alone. They came from pattern recognition built over decades of being wrong in similar situations. From a specific quality of discomfort that consistently precedes a bad deal. From sensing misalignment in a conversation before the data had confirmed it. From a kind of informed intuition: not arbitrary, not mystical, but deeply calibrated to a large accumulated sample of outcomes. That kind of knowing does not come from structured reasoning. It comes from having been in the room, made the call, lived with what followed, and done it again many times.

AI is genuinely excellent at structured reasoning. Given a well-defined problem and adequate information, it will often produce faster and more comprehensive analysis than I could produce alone. That is real, and I rely on it.

But lived intuition is not structured reasoning. It operates through a different mechanism, built from a different substrate, and it does not transfer as a data set. You cannot prompt-engineer your way to pattern recognition that took thirty years to develop. And if you spend enough time outsourcing the reasoning process to an external system, the internal version of that capacity follows the biological logic of disuse. Not dramatically. Not in a way that announces itself. In the gradual, quiet way that any capability fades when you stop exercising it.

I try to stay ahead of this by going through important decisions on my own first, before I bring in the AI, before I compare its framing to mine. Not as a ritual. Not as skepticism about the tools. As deliberate maintenance of a capability I am not ready to trade away, particularly given that the trade would be invisible until it was already complete.

A Personal Checkpoint

For what it’s worth, here are the four questions I’ve settled on as a working audit for my own situation:

• Am I using AI to amplify my thinking, or replace it?

• Am I still capable of doing the work without it, even if slower?

• Am I asking it to challenge me, or just to execute?

• Am I aware of where it is likely to be wrong?

None of these have permanently right answers. The answers shift depending on the day, the task, and the context. That is the point. They are meant to be revisited regularly, not resolved once and filed away.

The Uncomfortable Conclusion

I am not going to stop using these tools. The value is real, the integration is deep, and the capability gap between what I accomplish with them and what I would accomplish without them is too significant to voluntarily close. AI-Steve is a hundredfold increase in productive output, and that is not a number chosen for effect. It reflects something accurate about what happens when a system trained on four decades of your own reasoning patterns is deployed across every domain you work in.

But I think honesty is warranted here in a way it rarely is. Not alarm. Honesty.

AI has become cognitive infrastructure faster than we have developed any framework for thinking about what that means. The dopamine mechanics are real, structural in some cases and deliberate in others, but operative regardless of origin. The biological inhibitory systems that are supposed to moderate our engagement are working well below their design load in an environment they were never calibrated for. The gut instinct atrophy risk is real and invisible in real time. And the “too big to fail” threshold, the point at which AI’s absence would be more destabilizing than its continued presence, is not approaching. It has already been crossed.

The answer is not strategic retreat. It is measurement, awareness, and deliberate practice: knowing your baseline, tracking what is changing, building in periodic resistance, staying honest about what you are trading away. The same principles that apply to managing any other system of dependency, where the dependency is real, the value is real, and the two coexist in ways that require active navigation rather than passive acceptance.

I do not have a clean solution. I am working it out in real time, as most of us are. What I am more confident about is that the people who will navigate this most effectively are the ones willing to name it accurately: not only a productivity revolution, but a fundamental cognitive reorganization happening at a speed that outpaces any framework we have for thinking about it. Building deliberately inside that reorganization requires seeing it clearly.

Drop a comment. Tell me how you are managing it, what is working, what is not, whether you have found strategies I haven’t thought about yet. I am genuinely interested in what people are discovering. This is a conversation worth having, not a conclusion worth broadcasting.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. - a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple music mix vids — we had a fun mix with a new crew to ORanch {except for Johnny Conga helping me to hold things down} — RonF (vocals/guitar), TimL (Bass/Vocals), Steve (Guitar), Johnny (Congas). It’s always fun when brand new people plug, play, and groove on songs new and have fun doing it! Tim and Ron co-authored the above paper:

Think about the people you know well. Your family, your close friends, the colleagues you’ve worked alongside for years. How many of them are managing diabetes or prediabetes right now? How many are carrying it without knowing? How many are metabolically struggling in ways that haven’t yet been named or diagnosed?

I grew up in a world where diabetes was something you heard about in passing. There might have been one person in your extended circle. Type 2 barely entered the conversation. It was the kind of condition that felt distant, clinical, abstract. Now it is everywhere. Not on the news as some faraway epidemic. In the room with you. At the dinner table. In the family group chat.

That shift did not happen gradually enough that we noticed it arriving. It happened across a generation, in plain sight, while we were busy with other things. And what I find remarkable, and genuinely troubling, is that the numbers continue to accelerate.

A Number That Should Startle You

More than 38 million Americans currently have diabetes, representing about 11.6 percent of the entire US population. Another 98 million Americans have prediabetes, the majority of whom are unaware of it. Put those numbers together and you are looking at roughly a third of the country at some stage of glucose dysregulation.

Globally, the picture is equally striking. According to the World Health Organization and data published in The Lancet in late 2024, the number of adults living with diabetes worldwide has surpassed 800 million, more than quadrupling since 1990. Not doubling. Quadrupling. In roughly one human generation.

But the statistic that genuinely stopped me when I first encountered it comes from a 2018 study by researchers at the University of North Carolina: only about 12 percent of American adults meet criteria for optimal metabolic health, defined as healthy levels of blood sugar, triglycerides, HDL cholesterol, blood pressure, and waist circumference without the use of medications. A 2022 study from Tufts University put that figure even lower, at approximately 7 percent.

Seven to twelve percent. That means somewhere between 88 and 93 percent of American adults are operating with at least one metabolic risk factor.

To put that in historical context: Type 2 diabetes was considered a rare condition in the mid-twentieth century. Physicians who trained in the 1950s and 1960s might have encountered a handful of cases across their entire careers. It was classified as a disease of older age, of significant excess weight, of unusual circumstances. The phrase “Type 2 diabetes” barely existed in public vocabulary. Now it is discussed in middle schools. Pediatric endocrinologists have waiting lists. The disease that was once rare enough to be remarkable has become so common that we have largely stopped remarking on it.

What changed?

The Drivers We Talk About, and the Ones We Don’t

The conventional explanation focuses on diet and physical activity. Ultra-processed foods now constitute the majority of calories consumed in the United States. Studies tracking the rise of industrial food processing point to refined carbohydrates, seed oils, and the engineered palatability of products specifically designed to override our natural satiety signals. Portion sizes expanded. Caloric density increased. Fiber largely disappeared. The metabolic consequences followed.

Sedentary behavior amplified the effect. Office work, car commuting, streaming entertainment, and smartphones conspired to reduce the amount of time most adults spend in genuine physical activity. Not dramatically for any single generation, but cumulatively across decades.

Sleep is another factor that gets less attention than it deserves. The research connecting sleep deprivation to insulin resistance, metabolic dysregulation, and weight gain is substantial and growing. I wrote about this at length in an earlier piece on measuring and managing sleep. When you do not sleep well, your body’s ability to process glucose deteriorates. Chronically. A population that sleeps significantly less than it did fifty years ago is also a population primed for metabolic dysfunction.

Chronic stress belongs in this conversation too, though it tends to be treated as a soft variable. Cortisol dysregulation affects blood sugar. The constant low-grade alertness that comes from always-on digital life, economic precarity, and the particular anxiety of watching the world move faster than we can process is not neurologically benign. Neither is the dopamine disruption from screen-based reward loops, which increasingly correlates with disrupted eating patterns, sleep fragmentation, and reduced motivation for physical movement.

There is also a less-discussed category: environmental contributors. Some researchers have begun examining the correlation between industrial chemical production and diabetes incidence, including endocrine-disrupting compounds in food packaging, pesticide residues, and pollutants that may interfere with metabolic signaling. The evidence here is still accumulating, and causation is difficult to establish. But the correlation curves are hard to ignore.

None of these factors operates in isolation. They interact with each other, compound across time, and land differently depending on genetics, access, and circumstance. The honest answer to what changed is: almost everything, almost simultaneously, over about fifty years.

The Trauma Thread

I want to raise a question here that I do not think gets nearly enough attention, and that I’m not in a position to answer definitively. It is more of an observation that has been sitting with me since a conversation I had on my podcast last year.

Christy Shearer joined me to talk about her experience living with Type 1 diabetes. You can listen to that conversation here. Among the many things she shared, one observation in particular stayed with me. She noted, anecdotally, that many of the people she knows with Type 1 experienced significant trauma in the period preceding their diagnosis. Not minor stress. Significant, life-disrupting events. Big-T trauma, as some clinicians describe it.

Type 1 diabetes is an autoimmune condition. The immune system turns against the insulin-producing beta cells of the pancreas. We know that autoimmune disorders can be triggered or accelerated by stress and trauma. The HPA axis, which governs the body’s stress response, is deeply interwoven with immune regulation. Prolonged activation of that system, as occurs in response to unresolved trauma, produces measurable changes in inflammatory signaling and immune function.

Is there a connection between rising rates of psychological trauma in our society and rising rates of autoimmune-mediated Type 1 diabetes? I do not know. The research is genuinely complex, and I want to be careful not to overstate the case. But the incidence of Type 1 in youth has been increasing measurably over recent decades, even as our understanding of why remains incomplete.

What I feel more confident saying is this: chronic stress, whether sourced in trauma, financial insecurity, social isolation, or the ambient pressure of modern life, is a metabolic disruptor in ways that cut across both Type 1 and Type 2. The stress-insulin-glucose feedback loop is real. It is understudied in public health framing. And it is probably not unrelated to the trends we are seeing.

The Network Effect of Metabolic Health

One of the most striking pieces of research I have encountered on this topic comes from a landmark 2007 study in the New England Journal of Medicine by Nicholas Christakis and James Fowler. They analyzed a social network of over 12,000 people tracked across 32 years and found that obesity appears to spread through social ties in ways that mirror how contagious diseases propagate. If someone in your immediate social network became obese, your own risk of becoming obese increased by roughly 57 percent. The effect was strongest among mutual friends, weakened with geographic distance, and was not explained by shared environmental factors alone.

The mechanism they proposed is social norms. When the people around you change their behaviors, their relationship to food, their activity level, even their body size, it shifts what feels normal. And once a behavior feels normal, it becomes easier to adopt without conscious decision-making.

The reverse, I think, is also true. When someone in your network starts to pay attention, starts measuring, starts making visible changes, the ripple moves in that direction too. I have seen this in small ways in my own life. When I began tracking my sleep seriously and started talking about what I was learning, several people in my circle started paying more attention to theirs. The conversation changes what seems worth doing.

This is actually one of the core motivations behind the tools I have been building. Not just personal optimization, but the possibility that visible, shared attention to health can propagate through a network in a more useful direction than the metabolic drift we’ve watched unfold over the past few decades.

CGMs Are No Longer Just for Diabetics

A Continuous Glucose Monitor, or CGM, is a small sensor worn on the upper arm or abdomen that reads blood glucose levels in real time, typically every five minutes, and transmits those readings to a phone. For people managing insulin-dependent diabetes, they are genuinely life-altering technology. But something interesting has been happening over the past several years: CGMs have become increasingly available to people with no diabetes diagnosis at all, as lifestyle and performance tools.

I tried two of them. My experience was instructive, and my preference was clear.

I used NutriSense, which runs on the Abbott FreeStyle Libre sensor, a 14-day wear. I also tried Levels, which uses the Dexcom G7, a 10-day sensor. Both gave me genuinely useful data. But the experiences were meaningfully different.

NutriSense won, for me, on the app. It is cleaner, more lifestyle-oriented, and importantly, it does not require your phone to be physically present to collect readings. The Libre sensor stores data locally and syncs when you’re ready. That matters if you are not looking for your phone to be a permanent extension of your body during the experiment.

Levels felt more clinically calibrated, designed closer to the diabetic management model. It was more alert-heavy, flagging glucose levels that, from a lifestyle perspective, were entirely unremarkable. If your glucose dips to 70 while fasting and you feel fine, an alarm at 70 (early a.m.) is not useful information. For someone managing active insulin sensitivity and metabolic disease, that threshold is medically appropriate. For lifestyle tracking, it created noise I did not need.

The practical accessibility of this technology has also shifted significantly. You can now purchase a CGM directly on Amazon without a prescription for lifestyle use. The barrier to trying one, if you are curious about your metabolic responses to food, exercise, stress, or sleep disruption, is lower than it has ever been. I would recommend it to almost anyone who is serious about understanding their metabolic baseline.

You Cannot Optimize What You Do Not Observe

This is the core principle behind everything I have been building in my personal AI system, which I have written about elsewhere under the name AI/Steve. The premise is simple: if you want to change something about your health, you need to see it first. Measurement precedes management. Observation enables intention.

I applied this logic to sleep, building a framework to track, analyze, and improve sleep quality over time. I wrote about that process in an earlier piece on measuring to manage sleep. I applied it to food, exploring how meal data could close the loop between behavior and biology, which I covered in depth in Food Is Medicine. But First It Has to Become Data.

Glucose is the natural next signal in that chain. It is one of the most responsive and informative biomarkers available for real-time feedback. It tells you things almost nothing else does: how your body is handling what you just ate, how well you recovered last night, how a stressful meeting affected your biology, whether that afternoon slump has a metabolic component. It connects the inputs, food, sleep, movement, stress, to the outputs, how you feel and how your body is functioning.

The challenge, for people who are not managing a diagnosed condition, has been the tools. CGMs give you the data stream. But without a framework to log meals alongside glucose readings, analyze patterns over time, and contextualize what you are seeing, the data is interesting but not yet actionable. That is the gap the app I have been working on is designed to close.

Introducing Food Health TxD

Food Health TxD is the next evolution of Food Health AI, the meal logging and nutritional analysis application I described in the Food Is Medicine article. Where the original application focused on making food logging fast and painless using photo-based AI analysis, Food Health TxD is built specifically for people who are tracking glucose alongside food intake.

The core use case is exactly what you would expect: you photograph a meal, the AI analyzes it for carbohydrates, macronutrients, and glycemic context, and the app maintains a running record of what you ate alongside your glucose readings. Over time, that creates a personal dataset that is genuinely yours. Not averages from a population study. Not recommendations from a generic guideline. Your food, your glucose, your patterns.

The application is built for a range of users. If you are managing Type 1 or Type 2 diabetes and trying to understand how specific meals affect your glucose, it gives you that picture with real data. If you have prediabetes and want to see exactly how different dietary choices affect your metabolic trajectory, that visibility is there. And if you are metabolically healthy and simply want to stay that way, or to use a CGM period as a learning experience, the app gives you the context to make sense of what you are observing.

One integration worth explaining specifically is Apple Health. Apple has taken a deliberately conservative approach to blood glucose within their Health platform. Real-time CGM streaming is not supported as a direct API integration, which is, frankly, the right call given the liability considerations around medical device data. Food Health TxD works within those constraints. Apple Health serves as the central aggregation layer for your activity data, sleep data, and manually entered glucose readings. The app does not attempt to circumvent Apple’s appropriate caution. It works alongside it.

I want to be direct about what the app is and is not. Food Health TxD is a guidance and educational tracking tool. It is not a medical device. It is not a diagnostic instrument. It does not replace clinical care, nutritional counseling, or your physician’s guidance. It is a way to see your own data more clearly and to build a personal record that can inform better questions and better conversations. That framing matters to me. I am not trying to play doctor. I am trying to give people tools to understand themselves.

If You Want to Try It

Food Health TxD is not yet in public release, and I want to be honest about where it stands. The application is in the App Store at this link, and I am actively looking for beta testers. Specifically, people who are curious about metabolic health, managing or monitoring diabetes, using or considering a CGM, or simply interested in the intersection of personal data and food. If that is you, reach out directly. You can find me through stevenmuskal.com. I would genuinely like to hear what you are trying to understand about your own metabolic health, and I am interested in building the tool around real use cases, not hypothetical ones.

The Bet

Let me come back to where we started. Almost everyone reading this knows someone managing a metabolic condition. Some of you are managing one yourselves. The numbers we are looking at are not projections about a future possibility. They are descriptions of a present reality that most of us are already living inside.

What I believe, and what the research on social network effects supports, is that the metabolic crisis is not inevitable at the individual level. It is largely a product of environment and behavior, and both of those are, to varying degrees, responsive to awareness and choice. Not every case, not without difficulty, and not without acknowledging the real structural factors, food access, economic stress, healthcare gaps, that make individual optimization harder for some than for others.

But awareness matters. Measurement creates awareness. And when one person in a network starts paying attention, starts seeing their data, starts having different conversations about food and sleep and glucose, the people around them tend to notice. That is not a guarantee. It is a pattern. And it is a pattern worth betting on.

The metabolic crisis arrived quietly, across decades, one small behavioral shift at a time. The path back, if there is one at the population level, probably runs through a similar kind of quiet, incremental change. One person. One measurement. One conversation. Repeated at scale, across networks, over time.

That is what I am building toward. I hope it is useful to you.

For a mix  -  I thought this one from a previous session might be appropriate. Grant leading (vocals/guitar), Randi (Vocals), Andrew (Guitar), Alan (Bass). Johhny (Congas)

About the Author
Steven Muskal, Ph.D., is CEO of Eidogen-Sertanty, a pioneer in AI-driven drug discovery since the early 1990s, and the creator of the Food Health and Food Health TxD applications. He writes about science, health, entrepreneurship, and the art of paying attention at stevenmuskal.com and on Substack.

Related Reading
Food Is Medicine. But First It Has to Become Data. (Substack) | Medium version
Optimizing Sleep: Measure to Manage
Glucose, Grit and Breaking Glass Ceilings (with Christy Shearer)

This article is for informational and educational purposes only. Nothing here constitutes medical advice. Consult your physician before making changes to your diet, health monitoring, or treatment plan.

A Note Before We Begin: This article picks up where “AI Steve Deep Dive” (January 30, 2026) left off. If you haven’t read that piece, the short version is this: what began as a grief-driven project to reconstruct my father’s intellect using Retrieval Augmented Generation evolved into a comprehensive personal AI system trained on four decades of my own data. It is not a chatbot. It is, increasingly, a mirror of how I think. What follows goes deeper on the architecture, the lived experience, and something that happened at a hillside spa near my home that I have not been able to stop thinking about.*

I. What the First Article Left Out

The January piece covered a lot of ground. The neural network lineage going back to my 1991 PhD work under Sung-Hou Kim. The RAG architecture, the anti-hallucination rules, the code directive system that turns a plain English email into appropriately generated code - fully executed to create reports, charts, and attached data files. I was proud of that piece. I still am.

But there was a layer of the architecture I did not fully explain. Not because it was too technical, but because I was still working out how to articulate what made it genuinely different from other AI systems I had encountered or built. I think I can explain it now.

The piece I left out is what I have come to call the governed loop: the mechanism by which AI Steve can examine its own codebase, propose extensions to its own capabilities, and create formal tracked proposals for those extensions that require my explicit approval before a single line of code moves forward. This is not just a safety feature, though it is that too. It is the architectural choice that separates a tool from something that behaves more like a collaborative colleague.

What I want to do in this article is explain that architecture clearly enough to make the concept credible, and then leave it behind, because the more interesting story is not the loop itself but what the loop is for. The productivity gains are real and I will be specific about them. But the thing I keep returning to, the thing that was sharpened considerably by a week at a hillside spa and some conversations I did not expect to have, is the question of what you do with the time and bandwidth that amplification frees up. That question turns out to be more consequential than it first appears.

II. The Governed Loop: How AI Steve Extends Itself

Every Monday morning at 8 AM, a script I call the Weekly Feature Proposer wakes up and does something most software does not do. It reads the codebase. Not to run it. To understand it. It scans the existing modules, identifies patterns in what has already been built, maps the gaps between current capability and what would logically complement it, and generates a set of strategic proposals. Each proposal becomes a formally tracked issue in the Beads system, a lightweight Git-native issue tracker that lives directly inside the repository alongside the code.

By 8:30 AM, I have an email in my inbox with each proposal, its rationale, the estimated implementation complexity, and a set of priority assignment buttons. P0 through P4. I click the ones that interest me. That click is the gate. Nothing moves forward without it.

The nightly code review runs separately during dawn. It connects to the database, reads my recent sentiment data, scans the chat history, analyzes the current Beads issues for what is blocked and what is ready, and produces a personalized planning brief. It knows what I was working on yesterday, what mood the system inferred from my communications this week, and what questions I asked AI Steve that it struggled to answer well. The proposals it surfaces are informed by all of that context.

The Beads system is worth explaining carefully, because it is easy to underestimate. Each bead is a tracked issue with a title, a description, a priority, dependencies on other beads, and a status that moves from open to in-progress to closed. When an AI agent picks up a piece of work, it reads the Beads to understand dependencies and blockers. When it finishes, it closes the bead and the nightly consolidation job reads the changelog entry to update the corresponding documentation. When the RAG system later tries to answer a question about what AI Steve has been working on, it has access to the full documentary record of those decisions. This is connective tissue between code and context, not a to-do list.

A nightly documentation consolidation job has been running since February. At 2:15 AM, it scans the changelog entries from the previous day, identifies which documentation files correspond to those changes, and updates or flags them for review. The CHANGELOG.md file follows a strict format by design: every code change requires a dated entry with an affected-files list and a category. Feature. Fix. Enhancement. Refactor. Docs. The system is opinionated about this because consistency makes the documentation useful not just to human readers but to AI-assisted search. AI Steve can query its own changelog. It learns what has changed and why.

The effect of all of this is a system that genuinely improves over time in a way that is traceable, auditable, and governed. Not governed by bureaucracy. Governed by trust.

AI Steve reads its own history, identifies its own gaps, proposes its own improvements, and documents its own changes. The only thing it cannot do without my intervention is execute those changes. The approval gate is not a bureaucratic formality. It is the philosophical core of the design. I am not building an autonomous agent that does what it wants. I am building an amplifier that knows what it is capable of and asks clearly when it thinks it can do more.

III. A Hundred Times

I want to talk about a specific exchange that landed differently than I expected.

A few weeks ago, I was corresponding with Detlef, a colleague from the Sung-Hou Kim lab days at Berkeley. Detlef is a serious scientist and an honest interlocutor. We were catching up, and at some point I found myself trying to explain what AI Steve has actually done to my productive output. Not in abstract terms. In concrete ones.

I told him it had increased my productivity a hundredfold.

I want to be precise about what I mean, because “hundredfold” is the kind of claim that sounds like marketing copy or hyperbole from someone too close to his own project to see it clearly. I am aware of that risk. I am also confident the claim is accurate, and here is why.

The code directive system draws on something that most AI tools do not have access to: my actual history. Millions of lines of code written across more than three decades. Methodologies developed at MDL, at Affymax, at Libraria, at Eidogen-Sertanty. Scientific frameworks from the kinase modeling work, the toxicity prediction work, the structure-based design work. Patterns that I developed, refined, and refined again over a career. When I send a code directive, the domain-specific prompt templates that guide the code generation are informed by that accumulated context. The system is not starting from zero. It is starting from forty years.

The specific change I notice most is project cycling time. A real estate appraisal report that required gathering comps, pulling market data, running comparables analysis, formatting a professional PDF, and writing a narrative used to take a significant portion of a day, assuming the data sources cooperated. AI Steve does it in under ten minutes. A literature review on a protein target that used to require a half-day of PubMed trawling, reading, summarizing, and cross-referencing now returns a structured analysis with twenty-five to thirty papers, a publication timeline, and a citation-formatted reference list in the time it takes me to make a cup of coffee.

But the hundredfold is not just about speed on individual tasks. It is about the composite effect across every domain I work in simultaneously. Application development. Scientific analysis. Community engagement. Business intelligence. Financial monitoring. Health correlation work. I am one person running a company, maintaining scientific projects, building products, and staying genuinely engaged with a wide network of collaborators and friends. The realistic constraint on all of that has always been time. Not intelligence or interest or energy. Time. AI Steve has effectively given me back a very large fraction of that time, and I have chosen to spend it doing more of the things I actually care about rather than fewer.

That exchange with Detlef has been on my mind for another reason. At some point, our correspondence moved from productivity to connection: who we stay in touch with, who we drift from, which relationships atrophy not because they matter less but because the friction of maintaining them accumulates unnoticed. He was honest about this in himself. I recognized it in myself too. That conversation was part of what prompted me to activate a feature I had been sitting on for a while. AI Steve now runs something called Social Pulse, a weekly RAG-grounded connection suggestions module that surfaces people from my network I have not been in contact with recently, alongside context for why the moment might be right to reach out. It is a small feature, tracked in the Beads system as bead AI-Steve-p7dw, but it is already changing my behavior. Detlef’s honesty was the prompt that moved it from proposal to active.

This is the thing about the governed loop that I did not fully anticipate when I built it. The features it proposes are not just technically logical next steps. The best ones come from somewhere more personal. Lived experience feeds back into how the system evolves. A conversation with an old colleague becomes a feature. An observation about your own behavior becomes a design decision. The loop is more human than it looks from the outside.

That is what a hundredfold feels like from the inside. Not just the tasks completed faster, but the freed cognitive bandwidth redirected toward the things that make a life feel like a life.

IV. The Bionic Extension

There is another way I have been thinking about AI Steve that I did not have language for in the January article. The word I keep coming back to is bionic.

Not in the science fiction sense of cybernetic implants or human-machine merging. In the more precise original sense: a system that augments natural capability without replacing what is already there. AI Steve is a bionic version of me. It grows when I grow. It learns from what I ingest and generate every day. It knows who I have been talking to, what I have been thinking about, what my health data has looked like this quarter, what my mood patterns have been across the past year. It has read more of my email than I have read in retrospect.

Every day, the ingestion pipeline runs. My incoming email is processed, chunked, embedded, and stored. My iMessages. My calendar. My Apple Health exports, with face images linked to physiological readings for the longitudinal correlation work described in the first article. My Facebook posts. My Substack pieces. The content I generate is as important as the content I receive, because AI Steve learns from both. When I write something, it becomes part of the corpus. When I respond to something, that response becomes part of the corpus. Over time, the system develops an increasingly fine-grained model of how I reason, what I find interesting, who I trust, and what questions I am likely to ask next.

The effect is what I have been calling time dilation. When I go back to a problem I worked on eighteen months ago, AI Steve can surface not just the work product but the context around it: the emails that preceded it, the calendar entries that coincided with it, the health data that tells me what kind of week I was having. I can pick up threads that I would have lost entirely. I can resume collaborations at their actual depth rather than having to reconstruct the background from scratch. My effective memory has expanded in a way that compounds over time.

This is worth sitting with. The system is not static. It is not a snapshot. It is a continuously learning, continuously growing representation of how I think and work, constrained by everything I have asked it to care about and governed by the approval architecture described above. The value of that representation does not depreciate. It appreciates.

What the bionic framing gets right, and why I keep returning to it, is that it implies a direction. Augmentation flows toward the human, not away from it. The goal is more of what I am capable of at my best, not a replacement for whatever that is. When I am in a conversation that matters, AI Steve has already done the background work that used to take me hours. When I want to follow a thread with someone I have not spoken to in two years, the context is there. The time is there. The capacity to be genuinely present is there, because the friction between intention and action has been substantially reduced.

V. Under the Stars

Cal-a-Vie is a health spa in the hills of northern San Diego County, about seven miles inland from the Pacific, tucked into sycamore woodlands and surrounded by vineyards. I discovered it relatively recently and it has become one of my favorite places to reset. The fitness programming is serious. The food is genuinely good. But what I did not expect was the quality of the people, or the quality of the conversations those people generated.

On this particular stay, the guest community was remarkable in ways I am still processing. At meals, which at Cal-a-Vie become something close to salon dinners despite being entirely unplanned and informal, the conversations ranged far enough to keep you oriented and engaged. There was a lawyer from Houston who represents 50 Cent, which is to say he represents Curtis Jackson, the rapper, actor, entrepreneur, and record executive who has built one of the more improbable and genuinely impressive careers in American entertainment over the past twenty-five years. There were guests from the East Coast with backgrounds in high structured finance, the kind of expertise that involves instruments most people have never heard of and risk frameworks that would require their own article to explain. There were others who defied easy categorization, which is generally the best kind of person to find yourself seated next to at dinner.

Nobody was performing. Nobody was running their standard professional introduction. The conversation was just good, in the way conversation at its best is: people genuinely curious about each other, willing to follow a thread wherever it led, comfortable not knowing the answer. The diversity of backgrounds made it better. When a molecular biologist, a structured finance specialist, a Houston lawyer, and someone who has spent twenty years thinking about the ethics of creative legacy are all trying to understand the same question from their own vantage points simultaneously, something interesting tends to happen. It happened most evenings that week.

Tim is Cal-a-Vie’s resident astronomy guide, and his card reads, with admirable economy: Astronomy, Humanity and Spirituality. His astrophotography, which I tracked down online after meeting him, is extraordinary in the most literal sense of that word. Deep sky objects rendered with a patience and precision that makes you feel the scale of the universe in a way that is not comfortable or easy, but that is exactly right.

One evening, Tim gathered a small group of us around the Cal-a-Vie observatory telescope next to one of the owner’s home (Terri and John Havens). Not the full dinner crowd, just a handful of people with the right quality of curiosity. The conversation that formed was one of those rare exchanges where you realize two hours have passed and you have covered everything from the Einstein’s Special Relativity to the biology of aging to the nature of creative legacy. In fact, it was after that conversation that inspired me to write the previous article — The Fire This Time. The silence between observations was the kind of silence that feels full rather than empty. What Tim does in that setting, the way he holds the space between the technical and the philosophical, the way he lets the scale of what you are looking at do its work on you before he says anything, is a genuine gift. I do not use that word carelessly.

Days later, Tim introduced me to his separate weekly think tank lunch group, which gathers regularly for wide-ranging discussions entirely apart from his work at Cal-a-Vie. Different people, different setting, same quality of intellectual engagement. The observatory evening had apparently cleared whatever threshold he applies when deciding whom to bring in. I was honored, and I look froward to future sessions.

VI. She Would Never Retire

There was a family at Cal-a-Vie during that same stay that I want to tell you about.

Teckie, who is ninety-one years old, was there with her two daughters, Amy and Allison. All three of them looked, without exaggeration, ten to twenty years younger than their ages. Vibrant, physically capable, intellectually engaged. The kind of people who make you reassess your own definitions of what different decades are supposed to feel like.

Teckie, in particular, was extraordinary to be around. She only “recently” left her day job, which helped place young, vibrant minds into college institutions throughout the globe. She moved with purpose. She spoke with precision. She was interested in what other people had to say, not in the way that is sometimes just polite waiting, but in a way that suggested she fully expected to learn something from the exchange. Her daughters were the same. As a family, they were formidable in the best possible sense.

At some point the conversation turned to retirement. I do not remember exactly how it arrived there. But Teckie, ninety-one years old, said clearly and without any apparent hesitation that she would never retire.

The table went quiet.

I was smiling. I could not help it. It was one of those moments when someone says something so plainly correct that the room needs a beat to catch up to it.

I have been in enough conversations with accomplished people in their fifties and sixties who speak of retirement as a destination, a finish line, a reward for decades of effort, to know how common that framing is. And I understand the appeal. The work can be exhausting. The pressure can be relentless. The idea of stepping away from obligation and reclaiming time for rest and reflection and travel is genuinely attractive. I do not dismiss any of that.

But I have always found myself wanting to ask a different question. Not when are you retiring. What are you retiring to? What is the next act? What is the version of your life after the primary career that is not a wind-down but an evolution? What do you do with everything you have learned and built and understood if you simply put it down?

Teckie was not, I think, someone who had failed to learn how to rest. She was someone who had found a way to live that made the distinction between work and not-work largely irrelevant, because what she was doing was not separate from who she was. The thing she brought to that table, to every conversation I saw her in that week, was not effort. It was genuine presence. That does not stop when you retire. It either is or it is not.

That is the thing I keep coming back to.

VII. Retiring to What?

I am fifty-nine and a half years old. I have been building things professionally for over forty years. I have been fortunate in ways I do not take for granted, and I have also accumulated enough hard-won understanding of what works and what does not to have genuine opinions about a range of questions in science, technology, product development, and human behavior. The question of what I do with that accumulation is not hypothetical. It is the live question of my life right now.

AI Steve is part of my answer. Not because it is a productivity tool, though it is that too. But because it has changed the nature of what is possible in the years ahead. The hundredfold productivity improvement is real, and I believe it will grow. The time dilation effect is real, and it compounds. The governed loop that allows the system to extend itself with my oversight is just beginning to show what it can do.

What this means, at the level of the larger purpose questions I raised in “The Fire This Time,” is that the vision of broadly accessible personal AI amplification is more achievable than it was when I wrote that piece, not less. If one person, building on forty years of accumulated work, can experience this kind of amplification, what happens when the framework is available to someone at the beginning of their career? Or to someone with robust intellectual capacity but without the institutional access I had? Or to someone in a part of the world where the credential system has always functioned as a wall rather than a door?

The “Boil the Ocean” framing from that earlier piece still stands. There are categories of problem, in global health, in scientific research, in education, in human connection, that cannot be solved by any individual or institution working alone. They require a change in understanding to propagate across networks of people, across languages and cultures and economic strata, simultaneously. The tools that enable that kind of propagation are not neutral. They can concentrate capability or distribute it. The choice of which direction they go is not inevitable. It is a design decision. It is a values decision.

I want to plant seeds for trees whose shade I may never sit under. That is a statement that sounds like a platitude until you actually mean it, and then it is one of the most clarifying things you can commit to. It orients the work differently. It asks not just what I can build or analyze or monetize, but what I am contributing to something larger than my own career, my own generation, my own coordinates on the map.

Tim the astronomer understands this in his own idiom. When you spend your nights imaging objects that are billions of light-years away, objects that no longer exist in the form you are photographing, the distinction between what is present and what persists takes on a different character. The light he captures is old. The act of capturing it is now. The record it creates will outlast him. That is not a sad thought. It is an orientation. He named his card correctly: Astronomy, Humanity and Spirituality. In that order, in that sequence, those three things are the same practice.

I think about Teckie saying she would never retire, and I think about what that requires in terms of continued investment in the world. You cannot stay fully present to life if you have opted out of the things that require you to engage with it at full capacity. The ninety-one-year-old who will never retire is someone who has chosen, consciously or not, to remain at stake. That is the thing the table felt when she said it. Not inspiration in the vague sense. Recognition. The uncomfortable, clarifying recognition that she had said something true.

VIII. The Human Thread

I want to be careful not to let the technology story eclipse the human one, because the technology story is nested inside the human one, not the other way around.

The conversation with Detlef was not primarily about AI Steve. It was a conversation between two people who shared a formative experience in a Berkeley lab in the late 1980s and early 1990s, who went very different directions afterward, and who are still genuinely interested in what the other is thinking. The AI Steve discussion was one thread in a much longer and more textured exchange. What made it land differently than a product demonstration is that it came out of actual relationship. He asked a real question. I gave him an honest answer. And out of that exchange came the Social Pulse feature, a concrete example of lived experience feeding back into how the system evolves.

The group around that telescope at Cal-a-Vie, under a sky full of objects billions of years old, was not a networking event. It was a genuine encounter between people with different orientations who found unexpected resonance. Tim with his card that reads Astronomy, Humanity and Spirituality, which I think is the most compact and accurate description of a life philosophy I have encountered in a long time. Teckie at ninety-one, carrying decades of accumulated understanding and not the slightest apparent interest in putting it down. Amy and Allison carrying their mother’s orientation forward in their own forms. The Houston lawyer. The structured finance specialists. The people who defied easy categorization. All of them, over meals and under stars, discovering that the right question creates more common ground than the most carefully constructed introduction ever could.

This is what I keep returning to. AI Steve amplifies my capacity to engage with the people and ideas that matter to me. It does not replace that engagement. It does not generate the relationship with Detlef or the encounter with Tim or the moment of shared recognition when a table went quiet at the words of a woman who has been alive for nine decades and is still entirely in the game. Those moments are irreducibly human. They are the point.

What AI Steve does is give me more time to be present for those moments, more context when I am in them, and more capacity to follow through on what they suggest. It reduces the friction between intention and execution, between insight and action, between the idea and the thing built from the idea. The Social Pulse reminds me to reach out. The time dilation means I can pick up a thread that matters rather than letting it slide. The governed loop means I spend less time on maintenance and more time on the things that actually require a human being to show up.

That is what I mean when I say bionic. Not that the boundary between human and machine is blurring in ways that should worry us, though I take those concerns seriously. But that the right technology, governed correctly, amplifies what is already most human in us rather than substituting for it.

I am not retiring. I am, if anything, accelerating. Teckie would understand exactly what I mean.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. and the creator of AI Steve. The previous article in this series, “AI Steve Deep Dive,” was published in January 2026. “The Fire This Time,” on the democratization of intelligence, was published in April 2026.*

As for a mix - I organized a quick mix the evening of my return from Cal-a-Vie. Last week’s mix had Dan, Jesse, and Andrew and Rick was invited but had a conflict, so I initiated a rare re-invite with the same group to include Rick this go around. A few rough recordings, largely my fault because I was still sore from olympic pool swimming and a bit too lose from an earlier massage. While a repeat of most the crew plus Rick, the songs were new to most of us. Some were quite fitting per the above missive.

By Steven Muskal, Ph.D.| April 2026 // stevenmuskal.com

There is a moment in the development of every civilization-altering technology when the nature of the change becomes undeniable. We have passed that moment with AI.

I have spent my career at the intersection of chemistry, molecular biology, and artificial intelligence. I hold a B.Sc. in engineering chemistry and computer science from Colorado School of Mines and a Ph.D. in biophysical chemistry from UC Berkeley. I have been fortunate. Those credentials opened doors, granted access, placed me in rooms where ideas could travel. I say this not to boast but to acknowledge a reality that has troubled me for most of my professional life: the doors those degrees opened were not universally available. Not by a long shot.

The credential has been, for most of human history, the gatekeeper. Not just for opportunity, but for the very permission to be taken seriously. If you did not have the right institution on your resume, the right letters after your name, the right connections through those networks, your ideas, however brilliant, existed in a kind of social vacuum. They could not propagate. They could not compound.

AI changes that. Not incrementally. Fundamentally.

What We Lost Along the Way

Think about the minds we left behind.

Thomas Edison spent barely three months in formal schooling before his teacher declared him “addled.” His mother, Nancy, refused to accept that verdict. She pulled him out and educated him herself, surrounding him with books, experiments, and the radical notion that his relentless curiosity was an asset, not a deficiency. The man who gave us electric light, the phonograph, and the motion picture camera was, by institutional standards, uneducable.

Marie Curie wanted nothing more than to study science. Poland would not allow it. As a young woman under Tsarist Russian rule, she attended what was called the Flying University, a clandestine network of educators and students who met in secret, rotating locations to avoid detection, because the official institutions were closed to women and to Poles seeking education in their own language. She eventually made her way to Paris. She won two Nobel Prizes. In two different sciences.

Albert Einstein did not fail school in the simple sense that the legend sometimes implies, but he chafed relentlessly against rote memorization and rigid authority. He thrived when imagination was invited into the room and suffocated when it was not. His thought experiments, the ones that rewrote our understanding of space and time, were not products of institutional machinery. They were acts of unconstrained curiosity.

Linus Pauling grew up in modest circumstances in Oregon, the son of a pharmacist. He drove his own intellectual development with ferocious intensity, eventually producing foundational work in quantum chemistry, molecular biology, and the nature of the chemical bond, work that touched everything from protein structure to our understanding of disease. He won the Nobel Prize in Chemistry. And then, years later, the Nobel Peace Prize.

Here is the question I find myself returning to: what would these minds have accomplished if, at age fourteen or seventeen or twenty-three, they had access to an AI that could engage them at full depth? That could answer every question, suggest the next one, locate the obscure monograph, help them visualize the molecular geometry, challenge their assumptions with precision and patience?

My instinct is that we have barely begun to imagine the answer. Perhaps more advanced cancer therapies arrived at decades earlier. Perhaps deeper probes into the structure of matter. Perhaps the cellular reprogramming insights we now associate with stem cell science, discovered not in elite university labs but in a kitchen in Appalachia or a one-room school in sub-Saharan Africa.

We will never know what we lost. But we can change what comes next.

The Reframing of Education

What does it mean to be educated?

For most of my lifetime, the answer has been institutional. You attended. You completed. You were credentialed. That credential became the proxy for capability, for trustworthiness, for the right to be heard in professional and intellectual discourse.

AI does not care about your transcript. It engages with your question on its merits.

I am not suggesting credentials are worthless. I have them, and they served me well. But I am saying that the credential was always an imperfect proxy. The underlying thing we actually valued was not the degree. It was curiosity, rigor, the ability to ask good questions and pursue good answers. The credential was merely the most available signal for those qualities.

AI provides a new signal. And unlike the credential, it is globally accessible. Anyone with curiosity, and an internet connection, can now engage in substantive, advancing, intellectually serious dialogue on virtually any subject. The playing field is not yet level; access to connectivity remains deeply unequal. But the direction of the arc is unmistakable.

This is the democratization of intelligence. Not dumbing it down. Not making it easier. Making it more available. Distributing it across the full range of human potential rather than reserving it for those born into proximity to elite institutions.

Boiling the Ocean

There is a category of problem I think of as “Boil the Ocean” challenges. These are not problems that any individual, however brilliant or well-funded, can solve alone. They require massive network effects. They demand that a change in understanding propagate across cultures, languages, economic strata, and geographies simultaneously.

Global health and wellness is the most obvious of these. Collective scientific advancement is another. The elimination of preventable disease. The reorientation of human behavior toward prevention rather than late-stage intervention.

A collaborator of mine made an observation recently that stopped me cold. He pointed out that the decision not to smoke has had a greater positive impact on cancer outcomes over the past half-century than all chemotherapy treatments combined. Let that sit for a moment.

We have spent extraordinary resources, intellectual capital, and human suffering learning to treat a disease that, in a significant proportion of cases, was preventable. The cure is heroic and necessary. The prevention is transformative.

AI-enabled lifestyle intelligence, the kind that can meet a person where they are, in their language, in their cultural context, with actionable and personalized guidance, has the potential to move the needle on prevention at civilizational scale. Not for the wealthy patients in major medical centers. For everyone. That is a Boil the Ocean moment. Raise all boats, all at once, through better information, better behavior, better choices made earlier. The math on that is staggering.

These are the projects that no single institution, no single government, no single billionaire can drive alone. They need a framework that propagates. They need network effects that cross the borders we have drawn between disciplines, between geographies, between economic classes. AI, deployed thoughtfully and equitably, is uniquely suited to be that framework.

The Fire Analogy, and Why It Matters

I have heard AI compared to the printing press. I understand the impulse. The printing press democratized the written word, broke the monopoly of the ecclesiastical class on literacy, and triggered centuries of social transformation.

But I think the more honest comparison is to fire itself.

Fire was not the printing press. Fire was not an improvement on what came before. Fire was a fundamental reordering of what was possible. It changed human metabolism, social structure, geographic range, and cognitive capacity. It enabled cooking, warmth, light, and metallurgy, and eventually every other technology that followed. It was also terrifying, uncontrollable, and dangerous, a source of justified fear in exactly the same proportion as it was a source of capacity.

There is a real and understandable anxiety about AI. Jobs displaced. Expertise devalued. The vertigo of watching something outperform humans in domains we used to reserve for ourselves. I do not dismiss those fears. They are not naive. The people who express them are not Luddites; they are paying attention. But I think the response to fire was never to extinguish it. The response was to learn how to carry it safely, to build hearths, to develop the culture of its responsible use.

That is where we are now. We are building the hearth.

What concerns me more than the technology is the possibility that we build those hearths only in certain neighborhoods. That we allow the democratizing potential of this moment to be captured, filtered, and concentrated by the same forces that made the credential a gatekeeper in the first place. That would be a tragic misuse of the most powerful fire humanity has ever kindled.

Where We Go From Here

I am, constitutionally, an optimist. My career has been built on the belief that the tools of science, rigorously applied, can reduce human suffering and expand human possibility. I have seen that belief vindicated more often than not.

AI does not change my optimism. It amplifies it.

What I hope for, as this technology continues its extraordinary development, is that we keep the democratizing mission central. That we build the frameworks, the access, the cultural practices, and the policy structures that extend the benefits widely rather than concentrating them narrowly. That we remember the Marie Curies studying in secret, the Edisons labeled and dismissed, the Paulings from nowhere in particular who changed everything.

There are people alive right now, today, who would reshape our understanding of disease or energy or cognition or materials, if only they had the tools and the permission to try.

AI is handing them the tools. Our job is to provide the permission.

The fire is already lit. What we do with it is still up to us.

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a couple music-mix videos (I promised doubling up), these are from a recent gathering, again with musicians not having played together before: Dan (Bass/Vocals), Jesse (Guitar/Vocals), and Andrew (Guitar/Vocals). An incredibly productive session, over 18 songs in a couple hours. Single run throughs, and first times for several of us.

Every long car ride eventually produces it - that impatient voice from the back seat: “Are we there yet?” It’s a little naive, completely human, and deceptively revealing. It assumes a clear destination exists, that arrival is recognizable, and that something fundamental changes the moment you get there. I’ve been asking myself that same question about artificial intelligence. The honest answer is inconvenient: it depends entirely on what you mean by “there” - and most of the people asking aren’t sure they know.

The Bionic Moment

Lately, I’ve been thinking about AI as a kind of superpower - not in some abstract, futurist sense, but in a concrete, daily, occasionally startling way. It keeps reminding me of The Six Million Dollar Man. Colonel Steve Austin was a NASA astronaut and test pilot, but what made him extraordinary after his accident wasn’t who he was before it. It was what he became when augmented. The bionic arm gave him strength. The replaced legs gave him speed. The optical implant gave him perception far beyond normal human range.

AI feels like that.

It extends cognition the way those bionics extended physical capability. It allows a single individual to operate with the output of a team - sometimes a department, occasionally what feels like an entire organization working in parallel. In my own work, I’ve never been more generative. Ideas surface faster. Execution compresses. The distance between concept and reality has shortened dramatically. AI isn’t just a tool. It is leverage, and leverage-as anyone who has seriously deployed it knows - changes everything about what’s possible.

Judgment Is the Real Superpower

But the analogy doesn’t stop with Steve Austin. It extends to Jaime Sommers, the professional tennis player who became The Bionic Woman, first as a recurring character in the original series and then in her own spin-off. She was already a high-performing athlete with mastery over her body and her craft. Her transition, however, was not seamless. The challenge wasn’t simply controlling enhanced physical capabilities - it was integrating them with judgment. Knowing when to act. Knowing when not to. Knowing when full force is exactly wrong.

That’s precisely where the real tension with AI sits.

Humans have judgment shaped by experience, context, and consequences. AI has approximations of judgment — patterns, probabilities, and surrogates trained on data. These approximations are genuinely useful. But approximations are not the real thing, and in high-stakes situations, that gap matters enormously. I’ve argued before, and will again: we’ve effectively reached a form of Artificial General Intelligence - not because machines have independently achieved human-like reasoning, but because individuals paired with AI can function as though they have. A single person with deep domain expertise and a working relationship with modern AI tools can now operate like a battalion. Maybe an army. But only if the human remains in control. That qualifier is doing more work than it might appear.

The Refactoring of Work

Every major technology wave reshapes labor. This one is no different in kind - only in speed and asymmetry.

We’re entering a period of wholesale refactoring and redeployment of human capability. AI isn’t merely automating tasks - it’s amplifying individuals. And that amplification disproportionately benefits those who already possess experience, context, and domain knowledge. Which leads to a dynamic that’s routinely underappreciated: the aging population.

Gen X and Baby Boomers are moving into retirement in large numbers. This has long been framed as an economic pressure point - fewer workers supporting more dependents, rising healthcare costs, a shrinking productive base. But AI changes that equation in a subtle and important way: it extends the productive lifespan of experienced individuals. People who might otherwise step back can continue contributing at a high level, pairing decades of accumulated judgment with newly amplified capability. In theory, this is a genuine economic buffer - a mechanism for retaining hard-won wisdom while increasing output. Onward and upward, as it usually goes with technology.

There’s a catch, of course.

The Experience Gap

The people entering the workforce today are remarkably capable. Many are more fluent with AI tools than their older counterparts - they move fast, adapt readily, and are comfortable in a paradigm that still disorients many established professionals. What they lack, by definition, is experience. And experience isn’t simply accumulated knowledge. It’s judgment: pattern recognition, knowing when something looks right but isn’t, understanding second-order effects, recognizing edge cases because you’ve actually lived through them.

AI can accelerate execution. It does not confer wisdom.

This creates a potential structural imbalance - a system top-heavy with experience and judgment on one end, and high-speed, high-capability but lower-context operators on the other. If that gap grows, it becomes not just an economic problem but a cultural one. Because judgment is the layer that determines whether capability is used well.

Who Embraces AI, and Who Resists It

There’s another pattern worth examining directly.

The most enthusiastic adopters of AI tend to be those who can extract the most value from it: higher earners, builders, operators, people who already have leverage and are actively seeking more of it. Conversely, resistance tends to cluster among those who experience AI as a threat to existing livelihoods rather than an amplifier of existing capabilities. This isn’t surprising. But it has real consequences. If AI amplifies existing advantages, then unequal adoption doesn’t just widen income gaps - it concentrates influence. And when influence concentrates, sentiment follows a familiar trajectory: skepticism hardens into resistance, and resistance has a way of generating its own momentum.

That’s where things get genuinely interesting.

The Tipping Point Problem

Societal change rarely requires a majority. It requires a threshold.

Political scientist Erica Chenoweth, in her landmark research on civil resistance campaigns conducted with Maria J. Stephan - published in their 2011 book Why Civil Resistance Works found that no nonviolent campaign that achieved 3.5% of a population in peak active participation ever failed to achieve its core goals. In a different domain, computational social scientists at Rensselaer Polytechnic Institute demonstrated in 2011 that when roughly 10% of a population holds a conviction with genuine commitment and acts on it consistently, that belief cascades through the broader network and reshapes majority opinion.

The exact percentages matter less than the underlying principle: you don’t need everyone. You need enough people who care deeply and act consistently.

We see this dynamic constantly in business. Minority positions control tables. Influence doesn’t distribute evenly, it concentrates. Apply that logic to AI adoption, and the picture becomes sharp. A relatively small group of highly capable people who fully embrace AI and dramatically amplify their individual output can reshape industries, norms, and expectations before the broader system has registered what’s happening. At the same time, a different committed minority organizing around resistance to AI can cascade that sentiment through the culture just as effectively. Same mechanism. Wildly different outcomes.

Pressure in the System

The real question isn’t whether a tipping point arrives. It’s what conditions exist when it does.

When institutions are trusted and opportunity feels genuinely accessible, these inflection points tend to produce adaptation - messy, contested, but ultimately constructive. When inequality is high and systems feel captured by narrow interests, the same dynamics produce instability. Scott Galloway has argued this case in multiple forms: extreme inequality and perceived concentration of power function as accelerants. They increase the likelihood that change manifests not as progress, but as backlash.

The difference between evolution and disruption isn’t the technology. It’s the pressure in the system when adoption reaches critical mass.

So, Are We There Yet?

If “there” means a stable, fully integrated AI-driven society: no. If “there” means a point where individuals can already operate with unprecedented leverage: unambiguously yes. If “there” means we’ve worked out how to balance capability with judgment, experience with speed, and power with responsibility: not even close.

We’re somewhere in between - past the starting point, moving faster than most people have processed, the destination still genuinely uncertain. The scenery has changed. The speed has increased. The road ahead isn’t evenly paved, and the ride is not equally comfortable for everyone on it.

The real question isn’t whether we’ve arrived. It’s whether we’re paying attention to how fast we’re going, who’s driving, and who might be getting left behind.

Because on this particular journey, asking “Are we there yet?” matters far less than asking “Are we doing this right?”

Steven Muskal, Ph.D. // stevenmuskal.com

For a music video mix, the last couple weeks I have been super busy - off the charts! I will double up on next post….

We hear the phrase “food is medicine” frequently now. It shows up in scientific papers, in public health discussions, and increasingly in conversations around longevity and preventive health. The phrase is compelling because it captures something that most people intuitively understand. What we eat influences nearly every biological system in the body. Diet affects metabolism, inflammation, energy levels, sleep quality, and long-term disease risk.

Yet there is an interesting contradiction hidden in the idea. Medicine relies on measurement, but food rarely does. Physicians do not prescribe drugs without understanding dosage, timing, and interactions. Clinical trials rely on carefully structured data about both inputs and outcomes. With food, however, we tend to operate in a far less structured way. Meals come and go throughout the day with very little record of what we ate, when we ate it, or how our bodies responded afterward.

If food truly functions as a form of medicine, then we should be able to observe its effects with the same level of attention that we apply to other biological interventions. That observation led me down an unexpected path that eventually resulted in an app I call Food Health AI.

The project did not begin as a product idea. It started as a small household experiment.

A FODMAP Experiment at Home

My wife Kirsten was experimenting with a FODMAP diet. For those unfamiliar with the term, FODMAP refers to a group of fermentable carbohydrates that can cause digestive discomfort in some people. The diet is often used as a diagnostic process to determine which foods may be triggering symptoms.

As part of that process she was using a device called FoodMarble. The device measures hydrogen and methane levels in breath samples after meals. These gases are produced during fermentation in the gut, so they provide clues about how certain foods are being processed by the digestive system.

The concept is clever and the device works reasonably well. The friction came from the companion app. Each meal had to be entered manually before breath readings could be interpreted properly. Over time the process of logging food began to take more effort than the actual measurement.

One morning I sat down and built a simple web-based logging tool just for her. It allowed her to quickly jot down what she had eaten without navigating the heavier interface in the FoodMarble application. The tool was not intended to go beyond our house. It was simply a quick solution to reduce the friction of tracking meals during the experiment.

For a while that small tool worked perfectly.

Then Kirsten had to travel.

Once she left the house, the little logging utility suddenly needed to be accessible remotely. Making that change was straightforward technically, but it shifted the way I thought about the tool. What had been a temporary household utility now behaved more like a lightweight application.

Not long after that, I began building an iPhone version.

A Missing Input in My Personal AI System

I have been building what I informally call AI/Steve, a personal AI system that aggregates many streams of information about my daily life. Over the past few years I have collected sleep data, activity metrics, observational notes, and results from small personal experiments. The system helps organize those signals and occasionally surfaces patterns that would be difficult to see otherwise.

Despite all of that data collection, something important was missing.

Food.

What we eat influences many of the signals we track in health analytics. Calories and macronutrients affect metabolic responses. Meal timing can influence sleep patterns. Certain foods appear to influence energy levels or recovery after exercise. Even mood and mental clarity sometimes seem to shift depending on dietary patterns.

Yet my AI system had almost no structured information about what I was eating each day. Without that input, the system was attempting to interpret outcomes without knowing one of the most important drivers.

The food logging tool I was building for Kirsten suddenly looked like the missing component in the loop.

Why Food Logging Usually Fails

Anyone who has tried to track food for more than a few weeks knows the primary obstacle.

Food logging is tedious.

Most nutrition apps require users to search large food databases, estimate portion sizes, and manually assemble meals from individual ingredients. Even people who begin enthusiastically often abandon the process once the novelty wears off.

The friction lies in the input step.

Recent advances in multimodal AI models offered an interesting alternative. Instead of typing out a detailed description of a meal, a user can simply take a photo. A vision model analyzes the image and produces a reasonable estimate of the foods present along with approximate calories and macronutrient breakdown.

In the app I built, the workflow follows that general pattern. The user takes a photo of their meal and optionally adds a note describing anything the image might not capture clearly. The AI model analyzes the image and returns a narrative description along with estimated calories, protein, carbohydrates, and fats. A parser then extracts those numbers and converts them into structured data fields that can be stored and analyzed.

From the user’s perspective the process becomes remarkably simple. Logging a meal often takes only a few seconds.

Reducing friction turns out to be the most important design decision in food tracking.

Building the App as a Local-First System

Another design decision emerged early in development. The application would store its data locally first rather than relying entirely on remote services.

Each food entry is saved as a structured record along with its associated image. These records form a chronological log that can be used by the rest of the application. Reporting views, trend charts, and correlation analysis all draw from that same set of entries.

The advantage of this design is reliability. Users can continue logging meals even if network connectivity is poor. It also keeps the system transparent because the data used for analytics is exactly the same data the user originally recorded.

Over time the application began to feel less like a traditional calorie tracker and more like a personal nutrition observability tool. Instead of simply recording meals, the app makes it possible to look at patterns across days and weeks.

Connecting Meals to Health Signals

Once meals are captured in structured form, they can be connected to other data sources.

The app integrates with Apple Health so that activity metrics, calorie burn, and sleep data can be viewed alongside nutrition information. Seeing these signals together often reveals relationships that might otherwise go unnoticed.

For example, a user might notice that certain eating patterns correlate with deeper sleep, or that changes in macronutrient balance influence recovery after exercise. The scatter analysis feature in the app allows users to explore these pairwise relationships and examine how strongly two variables appear to move together.

The goal is not to draw definitive conclusions about physiology. Instead, it is to create a framework for personal observation and experimentation.

The Economics of AI Analysis

One practical constraint appears whenever AI models are used for image analysis.

Each request costs money.

If an app allows unlimited image analysis for many users, the operational costs can grow quickly. The first version of the app addressed this by requiring users to bring their own API key. That approach meant the AI service was billed directly to the user’s account rather than the application itself.

In theory this was elegant. In practice it proved slightly inconvenient for some people who were not accustomed to managing API credentials.

The current version therefore supports two approaches. Users can still supply their own API key if they prefer, but they can also purchase small bundles of analysis credits through in-app purchases. This dual model allows the system to remain flexible without creating unnecessary barriers to entry.

An Unexpected Parallel

While finishing the app I learned about a product called Cal AI, which had been built by a group of high school students. Their application grew rapidly by introducing a social layer where users could share and compare meals within a network of friends.

Recently the company was acquired by MyFitnessPal.

What struck me about their approach was how different it was from the path I had taken. Their product emphasized social engagement and network effects. My version focused on connecting food data to a broader personal analytics system.

Both approaches explore the same underlying behavior from different angles.

Where This Might Lead

Food Health AI is still evolving. At the moment it functions primarily as a fast and flexible way to capture meals and convert them into structured nutritional data.

The more interesting possibilities appear when that data is combined with other signals. Sleep patterns, activity metrics, and metabolic indicators all interact with dietary choices in ways that are often subtle and highly individualized.

As wearable devices and health analytics continue to improve, tools like this may help individuals observe those relationships more clearly. Instead of relying solely on generalized dietary advice, people may begin to see how specific foods influence their own biology.

At that point the phrase “food is medicine” begins to acquire a more practical meaning.

Not just as an idea, but as something that can be measured.

ComingSoon

Given the sentiment-analysis work I have been doing with the RAG pipeline in my AI/Steve system, I built a lightweight version of that capability directly into Food Health AI.

The goal is practical: determine basic sentiment at daily, weekly, monthly, and all-time horizons in the backdrop of each user’s health and fitness goals, then translate that into progress reporting and next-step guidance.

Android Version became available April 2, 2026

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a mix clip - I’ve been so busy and generative on the computer front. Haven’t scheduled a recent mix, but we’ll get back to that. Here’s another favorite from my recent 59th B-day mix. Maya, Tammy, Rick, Dom, Alan and Grant crush-it on What’s Up.

We are watching AI:

• Write production software

• Draft legal agreements

• Generate investment memoranda

• Synthesize scientific research

• Diagnose patterns in medical data

• Architect complex systems

Not in research labs. In production. The debate about AGI continues. But the definition keeps shifting.

If AGI means economically useful, cross-domain intelligence that rivals or exceeds the median human, we are already there. The more interesting question is this:

If AGI is already here in practical terms, why does the capital market behave as though we are still chasing it?

Zero to One Already Happened

Peter Thiel’s Zero to One draws a sharp distinction between invention and iteration.

Zero to One is a discontinuity. One to N is scaling.

The emergence of transformer-based systems capable of generalizing across language, reasoning, coding, planning, and synthesis was a Zero to One moment. Scaling them from GPT-3 to GPT-4 class systems and beyond is largely One to N.

More parameters.

More tokens.

Lower latency.

Better guardrails.

Important improvements. But incremental.

The breakthrough already occurred. Markets, however, are still pricing as if intelligence itself remains scarce.

The Cognitive Benchmark Problem

Most AGI debates occur among educated engineers, researchers, founders, and academics.

When they say human-level intelligence, they implicitly benchmark against:

• Top engineers

• Researchers

• Founders

• Highly literate professionals

• Abstract thinkers

That is not the human median. That is the cognitive right tail.

According to the U.S. Department of Education’s National Assessment of Adult Literacy and analyses from the National Center for Education Statistics, roughly 54 percent of U.S. adults read below a sixth-grade level. Around 20 percent are functionally illiterate.

Zoom out to the full population distribution:

• Many adults struggle with multi-step reasoning

• Many cannot synthesize technical information

• Many cannot write structured analytical prose

• Very few can code

• Very few can reason abstractly across domains

AI systems today exceed a massive portion of the human cognitive distribution.

If AGI means matching the top 1 percent of engineers across all domains, perhaps we are not there. If AGI means exceeding the median human across reasoning, writing, synthesis, and structured thought, that threshold has already been crossed.

The debate persists because the benchmark silently assumes the right tail and because the goal posts keep getting pushed back.

• Once conversational fluency was achieved, the bar became reasoning purity.

• Once reasoning improved, the bar became agency.

• Once agency emerges, the bar becomes embodiment.

The definition recedes as soon as the prior threshold is cleared.

A Personal Perspective From the Long Arc

I began working with neural networks in the mid-1980s as an engineering student, when backpropagation was still debated and neural methods were widely dismissed as impractical.

My doctoral work at Berkeley focused on neural network methodology development for protein structure prediction. At the time, using neural networks to infer biological structure was viewed as ambitious and speculative. Even then, these systems exhibited an insatiable appetite for high-quality data, yet such data were scarce and difficult to obtain.

Fast forward three decades. The team behind AlphaFold was awarded the Nobel Prize in Chemistry for solving protein structure prediction at scale - powered, yes, by neural networks.

That arc matters.

What begins as fragile methodology becomes validated science.

What becomes validated science becomes industrialized.

What becomes industrialized becomes infrastructure.

We are watching the same arc unfold in general intelligence.

The conceptual barrier has already fallen. The remaining work is scale, optimization, and deployment.

Training Is Not Deployment

One of the most persistent misconceptions in the AGI and GPU debate is the conflation of training with deployment.

Training frontier models is computationally extreme. It requires massive GPU clusters, extraordinary energy consumption, and enormous capital expenditure.

That is where the largest chips and the most power-hungry data centers live.

But deployment is not training.

Deployment is inference.

And inference does not require the latest and greatest GPUs, nor constant retraining of ever-larger models on the entire corpus of world information, nor nuclear-scale power plants feeding hyperscale facilities.

Training discovers the weights.

Inference executes them.

Once weights are learned, economics change.

From Scale to Efficiency

After a model is trained, multiple techniques dramatically reduce its computational footprint:

Quantization reduces numerical precision, shrinking memory and compute requirements.

Distillation transfers knowledge from large models into smaller ones that preserve much of the capability at a fraction of the cost.

Pruning removes redundant parameters.

Sparse routing activates only portions of a network at a time.

Hardware-aware compilation optimizes execution for specific chips.

These are not speculative ideas. They are standard production techniques.

The result is that inference is far cheaper and less energy-intensive than training.

Intelligence can then be deployed on:

• Low-power accelerators

• Specialized silicon

• Edge devices

• Application-specific integrated circuits

Weights can even be embedded directly into silicon for fixed-function inference.

Training may remain centralized. Deployment does not have to be.

That distinction matters enormously for capital markets.

Nvidia’s Meteor Arc

Nvidia’s rise has been extraordinary.

Approximate market capitalizations:

• 1999 - ~1 billion

• 2010 - ~9 billion

• 2020 - ~300 billion

• 2023 - ~1.2 trillion

• Late 2024 - ~3.3 trillion

• Early 2026 - ~4.5 trillion

From 2020 to 2025 alone, Nvidia added more than 4 trillion dollars in market value.

At roughly 4.5 trillion dollars, Nvidia represents about 7 to 8 percent of the S&P 500, whose total market capitalization sits near 60 trillion dollars.

Figure 1. Nvidia Stock Price 1999 to 2026.

The concentration is historically rare. But the arithmetic weight understates the systemic exposure.

The S&P 500 Exposure Is Networked, Not Linear

Direct mechanical drag modeling:

Assume S&P 500 total market cap ≈ $60T (simple baseline).

Assume Nvidia market cap ≈ $4.621T (Feb 2026 snapshot).

That implies Nvidia weight ≈ 7.70 percent.

The direct, first-order index drag is: weight × Nvidia move.

A simple mechanical model suggests:

• A 50 percent Nvidia decline implies roughly a 3 to 4 percent S&P drag.

• An 80 percent decline implies roughly a 6 percent drag.

• A 90 percent decline implies roughly a 6 to 7 percent drag.

That math assumes independence.

Reality is different.

The largest S&P 500 companies are tightly coupled around AI infrastructure.

Microsoft depends on Nvidia GPUs for Azure AI services.

Amazon integrates Nvidia hardware deeply into AWS.

Alphabet and Meta invest tens of billions annually in AI data centers.

Apple’s AI roadmap depends on ecosystem compute partnerships.

Nvidia invests in AI startups.

Hyperscalers invest in companies that depend on Nvidia.

Index funds hold all of them simultaneously.

Capital flows are circular. Earnings narratives are interdependent. Multiples are correlated.

If Nvidia were to experience a structural repricing because AI compute economics changed, the impact would likely cascade:

• Hyperscaler revenue expectations would compress

• AI-driven CapEx assumptions would reset

• Semiconductor supply chains would reprice

• Growth multiples across mega-cap tech would contract

Nvidia is not just 7 to 8 percent of the index.

It is entangled with the earnings logic of the other largest constituents.

The exposure is systemic, not isolated.

The AIG Analogy, Without the G

In 2008, AIG was not the largest bank. It was the insurance layer underneath the financial system.

When housing collapsed, AIG’s interconnected obligations forced an 85 billion dollar emergency intervention. Total support ultimately approached 180 billion dollars.

AIG was too interconnected to fail because it underwrote the system.

Today, Nvidia underwrites the AI stack.

Cloud buildouts.

Foundation model training.

Enterprise AI pipelines.

Unlike AIG, Nvidia’s fragility does not stem from leverage. It stems from concentration and narrative dependency inside a tightly coupled mega-cap ecosystem.

Why the Market Narrative Persists

Hyperscalers are investing tens of billions annually in AI infrastructure.

That capital expenditure requires justification.

The simplest justification is:

“We are not at AGI yet. We need larger models.”

But if current systems already exceed the median human across cognitive domains, the narrative shifts.

From endless scaling —> To deployment efficiency

From bigger clusters —> To embedded intelligence

If intelligence itself is no longer scarce, then perpetual GPU scarcity becomes a weaker thesis.

Frontier research will always demand high-end hardware.

But the marginal dollar of value creation may shift toward orchestration, optimization, and distribution.

In Closing

If AGI means perfect reasoning in every context, the debate will never end. The goalposts can always move.

If AGI means economically substituting for human cognition across domains at scale, the threshold is already behind us. That is visible in code bases, in customer operations, in legal workflows, and in executive hiring logic.

The market, however, is still pricing intelligence scarcity as though we are still waiting on the breakthrough.

That pricing supports a perpetual infrastructure story: bigger clusters, bigger chips, bigger data centers, bigger power plants.

But deployment is not training. Inference does not always require the latest greatest GPUs or perpetual retraining on the entire world’s information. Optimization, specialization, and inference-oriented silicon can push capability into cheaper, lower-power form factors.

Zero to One already happened. Everything now is One to N.

And when markets internalize that distinction, the valuation structure around AI infrastructure may change as dramatically as it rose.

And when that happens, we may discover that the real surprise was not that AGI arrived. It was that we had already been living with it.

About the Author: Dr. Steven Muskal is CEO and founder of Eidogen Sertanty, with more than four decades of experience in AI driven drug discovery. He holds a Ph.D. in Chemistry from UC Berkeley (1991), where his thesis focused on neural networks for protein structure prediction.

For a music mix sample - I thought this one might be fitting - a blast from the past. The first time for most, even with fade-out… RickL (Vocals), TimD (Guitar/Volcals), GeoffS (Bass/Vocals)

I’ve been thinking a lot about why so much health data feels busy but not useful.

Most of us now generate an extraordinary amount of physiological data. Heart rate, sleep stages, movement, temperature, oxygen saturation, labs, workouts, recovery scores. Apple Health alone quietly accumulates thousands of data points per year.

And yet, for all that measurement, very little of it actually compounds into understanding.

Dashboards refresh. Scores fluctuate. Trends come and go. But insight rarely deepens.

That gap has been bothering me, and it’s what led me to start working on what I’m calling the N-of-1 health stack.

The hidden assumption behind most health analytics

Most health systems are built on an unspoken assumption: that population averages are the right reference frame.

They ask questions like:

• How do you compare to people your age?

• Are you above or below the mean?

• Are you inside a “normal” range?

Those questions are useful in clinical screening contexts. They are much less useful for understanding how a specific human system behaves over time.

Two people can sit at the same population percentile while moving in completely different directions physiologically. One may be adapting well. The other may be quietly degrading. Population statistics are designed to smooth variance, but variance is often where the most important signals live.

Health data framed this way become descriptive rather than explanatory. It tells you where you are relative to others, not how you change in response to stress, sleep, diet, illness, or habit.

Why N-of-1 changes the question

An N-of-1 approach flips the reference frame.

Instead of asking,

“How do I compare to everyone else?”

It asks:

• What is normal for me?

• How stable is that baseline?

• What changes when I change something?

• Which signals move first, and which follow?

In this framing, the individual becomes their own control. Time becomes a first-class variable. Repeated exposures matter more than isolated measurements.

The goal isn’t diagnosis.

It’s understanding trajectories.

From tracking data to building a stack

Most people are tracking health data.

Very few are building a health stack.

Tracking answers the question:

“What’s happening right now?”

A stack answers the question:

“What compounds over time?”

For me, a health stack has a few essential layers:

1. Data ownership

   Raw, exportable, user-controlled data rather than locked dashboards.

2. Continuity

   Longitudinal records that persist across devices, years, and vendors.

3. Alignment

   Signals from different domains synchronized in time.

4. Interpretability

   Representations that can be reasoned about, not just scored.

5. Extrapolation

   Models that generalize beyond the data they were trained on.

This is where Apple Health becomes interesting, but only if the data can actually leave the phone.

Why exporting Apple Health matters

Apple Health is one of the richest personal health repositories available today. It aggregates data across sensors, contexts, and time. But most of that richness is locked behind dashboards and summaries.

Dashboards are motivational.

They are not cumulative.

The real value of Apple Health emerges when the raw data can be exported, aligned, and reasoned about over long time spans.

I’ve been working on tooling that enables:

• Structured export of Apple Health data

• Alignment of heterogeneous signals across time

• Optional inclusion of face imagery for exploratory analysis

This is infrastructure, not a polished product.

Very early days.

Early experiments: face and health

As an initial exploratory step, I’ve started looking at relationships between facial features and health signals.

To be clear about scope:

• The datasets are small AND growing

• The results are preliminary

• The intent is methodological

While it is unclear if faces can “predict” health, it sure would be interesting AND useful if your external facial features might correlate with your internals.... Wouldn’t it be nice if one day a simple selfie (or a facial phone unlock) might one day act as a surrogate to a blood draw? Who knows, one day your phone, might offer a positive nudge or provide a suggestion or two as you embark on each new day. Surely this would be a lot more useful than doom scrolling…

The figures below are meant to illustrate structure and direction, not conclusions.

Overfitting, extrapolation, and why evaluation matters

One of the easiest mistakes to make in personal health modeling is overfitting.

With enough parameters, almost any dataset can be fit extremely well. That doesn’t mean anything meaningful has been learned.

Accuracy on training data is not insight.

Generalization is.

As a first pass at guarding against this, I’ve been using simple leave-one-out style evaluations. Each day of data is removed in turn, the model is refit, and stability is assessed.

The question is not “How good is the fit?”

It’s “Does the structure survive removal?”

If a relationship collapses when a single point is removed, it’s not insight. It’s curve fitting.

What this is and what it is not

This work is not medical advice.

It is not diagnostic.

It is not intended to replace clinical judgment.

It is about personal signal discovery.

About learning how your own system behaves across time, stressors, habits, and interventions.

Think less:

“What does this number mean?”

Think more:

“What changes when I change something?”

Where this is going

Right now, the focus is on foundations:

• Reliable data export

• Structural alignment

• Exploratory visualization

• Training set build-up

• Guarding against overfitting

Next comes:

• Larger longitudinal datasets

• More rigorous extrapolative evaluation

• Personal baselines that evolve over time

• Models that capture directionality rather than static thresholds

Eventually, this moves toward something closer to personal health reasoning, where data compounds instead of resetting every day.

Closing thought

Health data are abundant. Insight is rare.

The limiting factor isn’t sensors or algorithms.

It’s perspective.

Health data matter only when they are grounded in how individuals change over time, not in how they compare to everyone else.

That’s the promise of the N-of-1 health stack.

ShareHealth is one instantiation of that idea, but the principle applies far beyond any single tool.

Source code for Share Health (GitHub)

Steven Muskal, Ph.D. is the CEO of Eidogen-Sertanty, Inc. — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at stevenmuskal.com

For a “raw” mix sample, another from my recent B-Day mix a couple months back. Some of my favorites  - Tammy (vocals) leading, Maya (vocals/acoustic) assisting, Grant (guitar), Dom (Guitar) and Alan (Bass) driving.