Steven M. Muskal, Ph.D.

Seven companies now carry roughly a third of the S&P 500, with the Magnificent Seven peaking at 35 percent in June 2026 against prior highs near 26 percent in the 1980 energy boom and at the dot-com top. The piece's frame comes from Carlota Perez: every bubble runs two engines, a technological one and a financial one, and a study of 51 major innovations from 1825 to 2000 found equity bubbles in 73 percent of them. Which engine dominates tells you what breaks. When finance dominates, as in 2008, the leftovers are loan books that reprice; when technology dominates, as in 2000, the leftovers are physical capacity someone else inherits cheap. Only twice have both engines run at full throttle: 1873, when Jay Cooke financed the Northern Pacific on bond sales years ahead of the freight, and now. Cooke was not a fraud; he was a believer on a schedule the traffic could not meet, his firm's failure closed the New York Stock Exchange for ten days, and the line was completed in 1883 by other people after his investors were wiped out. Swap the Northern Pacific for gigawatt data centers and the bond syndicate for off-balance-sheet vehicles and circular chip-and-cloud agreements, and the structural bet is identical.

The asset at the center is the weakness, and Steve speaks to it from forty years of building models: a large language model is a very large array of numbers, billions of floating-point weights fitted by gradient descent, with no patent-grade secret inside the matrix. Chinese labs, DeepSeek most visibly, have distilled frontier models at a small fraction of the training cost and published the weights; four-bit quantization fits a 70 billion parameter model in roughly 40 GB on a workstation under a desk; Qwen derivatives were over 40 percent of new language model variants on Hugging Face by August 2025. If a half-trillion-dollar asset can be functionally reproduced by people who could not buy your chips, and then given away, it is not a moat. At best it is a puddle. The steel parallel makes the mechanism explicit, except the marginal cost of the next copy is zero: a government that wants to compress American model margins does not need an export terminal, it needs one upload. And capability is no longer the constraint. The last real jump in usefulness came from inference-time compute rather than a larger base model, a quiet admission that pretraining scaling has hit diminishing returns, while a well-engineered mid-sized open model now beats what the frontier could do eighteen months ago on the tasks a business actually cares about.

The harder ceiling gets less attention than compute: Epoch AI puts the effective stock of quality public human text near 300 trillion tokens and projects training consumes it between 2026 and 2032. The corpus has been read, access is tightening, and new writing cannot be manufactured by the people who need it. So the durable moat sits where nobody is pricing it: proprietary content, the corpus nobody else holds and that keeps growing, already fetching reported terms like News Corp's $250 million-plus OpenAI agreement and Reddit's $203 million in licensing, with the move from subscription to perpetual license a 10 to 100 fold step up for the contributor, and the asset regenerating every training cycle. A chemistry aside carries the theme, ATP for energy and GTP for information, one oxygen apart and never confused by the cell. The close is two lessons pointing the same way: hold more dry powder than is comfortable, because the thermodynamics of the correction are settled even if the kinetics are not, and keep writing, because original content is the one part of the stack nobody can distill out of someone else's model.

If you wear a Dexcom Stelo, G7, or ONE+ and your glucose chart in a third-party app runs hours behind the clock on the wall, nothing is broken. The sensor samples interstitial glucose about every five minutes and delivers it to the Dexcom app exactly as advertised. The delay sits one layer downstream, where the Dexcom app exports to Apple Health on iPhone and Health Connect on Android, and it is held back on purpose: Dexcom's own support pages state the three-hour delay in one sentence each, and none of them says why. Steve's reading is that it is a regulatory line rather than a technical one. The Dexcom app is the cleared medical display, and a three-hour hold guarantees that no app downstream of the general health stores can be mistaken for a real-time glucose monitor. The cost lands on anyone trying to connect a meal to the glucose response it caused: lunch at noon arrives in Apple Health smeared across a window wide enough to blame the wrong meal entirely. After 35-plus years of building software at the intersection of biology and computation, Steve calls it what it is, a timestamp problem, which makes it a solvable one.

The fix is a beta called Direct Sensor Reading in Food Health TxD, the diabetes companion to Food Health AI: Scan & Score, and the piece is specific about whose shoulders it stands on. On iPhone it is built on the engine of xDrip4iOS, the community CGM app, which joins the sensor's existing Bluetooth session in Coexistence mode: the Dexcom app keeps owning the sensor, no pairing code, nothing changes, and TxD writes each reading to Apple Health the moment it arrives, with backfill. On Android it is built on xDrip+ from the Nightscout Foundation, reading each value off the Dexcom app's notification and writing it to Health Connect, with an optional direct Bluetooth bridge for anyone who wants the listener to own the sensor outright. Sensor data you already paid for reaches your chart in minutes instead of hours, and the caveats are stated as plainly as the feature: the beta is off by default, TxD never pairs with or takes over the sensor, the Dexcom app keeps the medical alarms, nothing in the app is a glucose estimate or prediction, and it is not a medical device.

Minutes matter because of what they unlock. No public repository anywhere pairs raw voice or face captures with glucose values, so the only labeled data will be first-party, collected by the person whose glucose it is. With readings arriving in near real time, TxD labels each voice or face check-in with the nearest sensor reading within 15 minutes, reduces every clip to nine acoustic features and every photo to ten, and runs the statistics on the device: a day-block permutation test, leave-one-day-out ridge regression, and a three-band random forest, with the verdict shown plainly even when it is no signal yet. Two weeks of Steve's own data close the piece. The food scans carried the clearest signal: every meal under 30 grams of carbohydrate kept his rise under 30 mg/dL, a model on the scan's nutrients predicted rise sizes 28 percent better than his average, and the biggest spiker was a chocolate nut bar rather than the biggest meal. Cheek redness tracked glucose at a correlation of 0.51, and voice pitch showed a small positive tendency that cannot yet be separated from whatever else changed between days, which is exactly why the app keeps collecting.

Science is supposed to be hypothesis testing and validation, and the field quietly stopped applying that to its own back catalog. You publish, the claim enters the record, and it sits there untested against anything larger than the data you had at the time. Steve spent a few months going back through twenty five years of his own work, some of it from 1998, and rebuilding it to find out whether it still holds. Five preprints later the answer is that most of it does, and the reason a twenty five year old hypothesis can be retested at all is that the public data is now one to two orders of magnitude larger. The loop he runs today inverts the old one: resurrect the code, revalidate on the bigger data, stand up a free working website, then post the preprint and let the formal record catch up.

The revisits run through four decades of problems. The 1998 human intestinal absorption model turned out to have ridden a fortunate train and test split, yet its six descriptors still carry the signal on a set nine times larger, and the deployed model at predicthia.ai calls the right side of the 30 percent threshold 92.64 percent of the time. The 2003 toxicity method, where a compound's structural neighbors vote, lost its proprietary reference set and has been rebuilt on 11,989 public rat oral LD50 values, then extended untouched to 28 more endpoints. A new signal came with it: a compound whose structural matches reach no protein family inhibits CYP3A4 2.5 percent of the time, and one reaching sixteen or more families does so 51.2 percent of the time, a pattern that then held on 2,335 compounds measured later by the OpenADMET consortium. PharmCast removes conformer generation, which was 2.82 of the 2.86 seconds a pharmacophore fingerprint used to cost. ReverseScreen.ai puts retrieval in front of docking and hands it 35 protein sites instead of 28,579. The comparator models ask which of two, and need no structure, no pocket and no pose.

Three things tie them together, none of them planned. Ask which, not how much, because decisions are comparisons and the field spent decades predicting absolute quantities nobody uses. The expensive step is usually optional, because precision a decision never spends is cost rather than rigor. And an old question is usually underpowered rather than closed, since what expired was the tooling and the sample size, not the idea. The same turned out to be true of the people: every revisit put Steve back in touch with a co-author he had not worked with in years, and everyone said yes. The piece closes on the quieter threads, sound and the microbiome, a vitamin D question he wrote down ten years ago and answered in about a third of a second, and why both of his food apps listen.

Reading is a 5,000-year-old hack. Speaking is a 200,000-year-old operating system. Before any human could read a word, every human could tell a story, and anyone within earshot could receive it. Spoken language is old enough to be load-bearing in our neurology; writing is a tool bolted on afterward that repurposes visual circuits built for recognizing objects, which is why children pick up speech on their own and reading only with instruction. The printing press, the paperback, the inbox and the feed moved information faster than a voice ever could, but none of them changed what kind of animal we are. Steve, who has built software for three decades and is not anti-text, argues we have confused the accelerant for the engine.

From there the essay works through the consequences. Literacy has always been a gate; voice is the leveler. Reading before bed demands tight saccadic eye control at the moment the body is trying to let go, while listening asks nothing of the eyes. Typing is a very small movement for a very big brain, where oral storytelling recruits the whole body and walking measurably improves creative output. A tour of ten oral traditions, from West African griots to shadow puppetry, shows how much information has survived for centuries without a page. The toolkit for going back is practical: local dictation with Rambleproof, cloud dictation with Wispr Flow, and a cloned voice from ElevenLabs that reads back to him.

The last section treats his voice as a lab sample. Jitter, shimmer, harmonics-to-noise ratio and cepstral peak prominence track physiology, and recent studies in Diabetes Care, the Journal of Voice and IEEE JBHI link them to hypoglycemia, gestational diabetes and type 2 diabetes. Steve now records every dictation session as a longitudinal dataset and plans to correlate the voice features with continuous glucose readings. The close is a return trip, not a retreat: keep everything literacy earned, and route it back through the channel we were built for.

Steve met Tim Lewis in an observatory, where people gradually headed out and the two of them kept talking. Tim takes remarkable photographs of the night sky, plays bass to Steve's drums, and has a habit that runs through the whole conversation: once he understands something, he wants to explain it to somebody else. His path to astronomy started at eleven with library books on telescopes and a first scope so unstable it needed homemade tripod braces (and cost him a drill bit through a toe), then a long detour into music before he came back around 2000 with better tools. Looking at his images of the Pac Man and Orion nebulae, Steve's reaction was how small we are; Tim's correction was that zooming out can make our problems insignificant without making us insignificant. From there the talk moves through consciousness, Steve's comfort with not knowing, and the few seconds when a band locks in and the whole room feels different.

The middle of the conversation is about working with AI. Steve has written a lot of code and does not want to write another line; he wants to describe what he wants built, inspect what comes back, test it, and improve it, and the result has been more projects and more old questions reopened. The concrete example is a human intestinal absorption model he built at Affymax a quarter century ago from 86 hand-curated compounds. Revisited with approximately 780 data points, the original result reproduced, but trying many different divisions of training and test data showed the original division had been one of the favorable ones, and that broader picture goes into the new report. The same approach applies to consortium voting for toxicity prediction: build the site, make the model available (cHIA at predicthia.ai and ToxPred), then write the paper around work people can already use, while pressing the pharmaceutical industry to share far more of the data it already holds.

The close is about who benefits. Tim describes personal growth as an expanding sphere of care, and Steve connects it to a change since his father passed away: more of his time and money directed toward other people, with no sense of deprivation. Tim keeps returning to incentives, from an automated billing system with no human to reach to the idea of measuring a company partly by how much it improves people's lives. Neither claims to have the mechanism. What they agree on is the practical part: share what you understand, put the tools out there, and see who can use them.

Irina Ankoudinova's own technology kept making her obsolete. Now she wants to bring Ukraine's land back, and she needs about five people. The first half of the conversation is a career with a shape: every time Steve asked what happened next, the answer was a new technology arriving and making her current one unnecessary. Zinc fingers and TALENs, then CRISPR. Sanger sequencing, then massively parallel sequencing. Structural genomics, which quietly lost its funding wave. Her path runs from plant biology into Sung-Hou Kim's lab at UC Berkeley, where she was handed crystallization, the least deterministic step in all of structural biology, then on through genome editing at Sangamo, next-generation sequencing, and DNA-encoded library screening. She got hit early every time because she worked at the edge, where the hits land first, and the lesson she pulled out of it travels well past biology: do not get attached to the technology, get attached to the questions you are trying to answer.

The second half is Ukraine, where her mother's family is from, and it has the same shape at a scale that makes the first half look small. Farmland shelled and mined out of being farmland, a language turned into a border, families scattered across a continent. What she wants next is a vocation rather than a job: recovering war-damaged soil the biological way, matching plant species and microbiomes to a specific contamination and giving a plot one season to come back, rather than solving it with heavy industry. Maybe not wheat. Maybe beans. Israel turned desert into farmland and countries wrecked in the Second World War brought their land back, so the precedent exists and the experience can be shared.

The last hour is method: triage damaged ground from satellite and drone imagery, find an analog landscape you are actually allowed to walk and run controlled experiments there, and generate training labels by computation or metadata rather than by hand, the same move that turned a minute of pharmacophore calculation per molecule into a fraction of a microsecond. The idea the title comes from is two-sided: a person who left and a person who stayed each posting where they are today, a mutual check-in that happens to produce a landscape data set. Nothing here is a plan. It is a scientist with a good question and no collaborators yet, and the next step is not having the answer, it is getting five people interested in the same question.

We Are Optimizing the Wrong Objective in Biology opens with a personal anomaly: Steve's Amazon cart registered a respiratory infection about four days before his wearables did. He uses it to ask a bigger question. Medicine has spent a generation predicting protein structure and other intermediate abstractions, but nature never optimized structure. Evolution optimizes fitness through a DNA to organism to outcome loop, and that, not structure, is the objective medicine should share.

The turn is that everyday behavioral data, our purchases, wearables, and transactions, already works as a population-scale physiological sensor network no one ever pointed at health. Five peer-reviewed studies anchor it: 47,000 Fitbits tracking CDC flu curves at correlations up to 0.97, smartwatches flagging COVID before symptoms, the DETECT study across 30,000 people, and credit-card sequences sorting a city into lifestyle groups. Google Flu Trends is included as the cautionary counterexample, because ambient data is only as good as the rigor applied to it.

The right question, then, is not what structure a protein adopts, but given everything we know about a person, which intervention improves their trajectory. It is an N-of-1 argument scaled to populations, and the population evidence is starting to catch up to one man's shopping cart.

The piece opens with a dated sequence and asks the reader to take it in order. April 3, 2026: Anthropic acquires Coefficient Bio, a stealth biotech of fewer than ten people, nearly all from Genentech's Prescient Design, for roughly $400 million in stock. April 16: a collaboration with Eli Lilly, presented by Dario Amodei himself, following an earlier AbbVie partnership. May 20: a strategic agreement with Bristol Myers Squibb reported to reach more than 30,000 employees. June 9: Fable 5 ships with a routing policy that diverts, in the company's own words, the majority of biology, chemistry and life sciences queries away from the most capable model. June 30: Claude Science and in-house preclinical drug programs. August 2: reported AstraZeneca and Bristol Myers Squibb merger talks that would make Anthropic's May customer the largest drugmaker on earth. The therapeutics commitment came first, which retires the defense that the drug business is an afterthought bolted onto a fixed safety posture. In roughly fourteen months the company's own description of its safeguards moved from a very narrow set of topics to the majority of a scientific discipline, and its own published diagram labels the band it deliberately widened benign but blocked.

Steve's objection is not his own access, though the canary is personal: the flags land in his terminal several times a week on ordinary target biology and molecular mechanism work, from someone with a Berkeley Ph.D. in biophysical chemistry, four decades in computational biology and a paid enterprise relationship, and the fallback tier now carries the same intentionally broad safeguards. The real objection is the bet. More than 200 recognized cancers, an Alzheimer's label covering substantial biological heterogeneity, and over 10,000 rare diseases mostly without an approved treatment are not solved by a few large pharmaceutical companies and one AI lab's internal programs. They get solved, if they get solved, by thousands of groups attacking thousands of targets in parallel, resting on a pre-competitive commons of deposited structures, released sequences and published dead ends. Concentrating the most capable scientific reasoning behind a handful of incumbent relationships is a bet that the binding constraint in medicine is capability rather than parallelism, and Steve argues that bet is wrong in a way that costs lives rather than money. Nobody outside the company can tell whether Lilly or Bristol Myers Squibb work under the same constraints as a graduate laboratory, and that opacity, he writes, is not a side issue in the story; it is the story.

Set the ethics aside and it is still a vendor-qualification problem: availability, after two frontier models went offline worldwide for roughly eighteen days under a government directive; behavior that changes without publication, which is an audit finding whose remediation is revalidation; fallback; and the new fourth question, who else are you. Steve built the obvious architecture in July, an abstraction layer with two providers and an open-weight model in reserve, and notes the open-weight gap to the closed frontier has compressed from roughly 150 Elo points to about 30, so a filtered request now has somewhere to go. He does not want Anthropic to lose; he wants no one to be the gatekeeper, because plurality is the only appeals process the field actually has. The close is three things to publish, none of which removes a guardrail: turn the announced biology trusted-access program into a real front door with published criteria and answers in about two business days, let verified identity reach the classifier, since a system that bills a customer by name and judges them as a stranger is an unfinished architecture, and disclose the firewall, stating whether the internal therapeutic programs run against the same threshold that stops everyone else.

AI for Everyone opens in a lecture room in 1990, where a young Steven Muskal presents neural-network work and hears the sentence that closes doors: that is no big deal, we knew that decades ago. Sung-Hou Kim widens the room instead of piling on, and that choice becomes the essay's spine - expertise can humiliate the next generation, or it can help it keep going. Decades later, as AI takes off, Steve catches the same dismissive line forming in his own mouth and resolves to hold the door open instead.

What makes that invitation real is that the tools now build themselves in plain English. Steve describes shipping real products - Toast Our Friend, AI-Dad, Image Explorer, a patient CT viewer, Food Health - by saying what he wanted and letting the system build it. He turns the fear of being replaced into a better question, which parts of the work should no longer consume a life, and argues the durable edge stays human: fresh, honest, original content is the scarce resource now that models have nearly drained the web's supply, so content, not compute, makes kings.

From there the piece widens to where the race is actually won. Frontier models spread too fast through open weights to stay ahead, so deployment wins, and nations win it with millions of ordinary people turning what they know into working software - the way a medical student's four years of flashcards became Study with Hannah in a single day. If you can measure it, you can monetize it. Steve closes with four moves anyone can make now that the cost of building has collapsed: use it, share it, preserve it, and build it - choosing to be the voice that says come join the party.

That Go There starts with a phone call about an old tape: Steve's older daughter at two and a half, on the floor with a wooden puzzle of the United States, setting every state in its exact spot and naming the capitals in the small serious voice kids use when they already know they are right. She could not do it today. Neither could he. The tape still can - thirty-some years later, press play and it all comes back.

That gap, between what a mind lets go of and what a piece of media holds onto, becomes a practical argument about memory preservation. The tape survived five dying formats only because someone kept moving it forward. The Library of Alexandria in your garage does not announce when it starts to fade: find the tape, the drive, the format nobody has opened in a decade, and move it somewhere that survives the next format change - then celebrate the people on it while they can still watch.

Study with Hannah starts at a kitchen table, where Steve learns his daughter has quietly built a personal knowledge base of more than 62,000 medical study cards, backed by tens of thousands of her own annotated histology slides, radiographs, and diagrams. It is the residue of two associate degrees earned before she could vote, a psychology bachelor's, four of six years at Semmelweis University in Budapest, and a USMLE Step 1 pass well ahead of her class. Her classmates had been using all of it for free.

In about a day that conversation became StudyWithHannah.com: keyword search across her notes, semantic image search over the slide archive, and a conversational AI tutor that quizzes from her own cards and cites them instead of generic textbook filler. The essay widens from there. A stock-photo site still licenses a newspaper photo of Hannah at four years old, taken at a library event two decades ago, without her consent or a cent. Education is already a market, and students are perpetually the customer and never the seller. If schools can monetize the process of educating students, why can't the student who did the distilling be the one at the counter?

Universities Built the Bomb. Now Build This asks whether elite research universities will treat AI as a threat to scarce degrees or as the next infrastructure for public knowledge. Starting from a conversation with Sung-Hou Kim, the essay argues that Berkeley and its peers can build institution-specific AI knowledge systems from their own research, courses, libraries, and faculty expertise.

The piece is both proposal and historical challenge: use RAG, citations, privacy boundaries, licensing, and faculty incentives to make university knowledge interactive and accountable. Berkeley's cyclotron, plutonium chemistry, and Oppenheimer network become the frame for a new question: who will build the civic infrastructure for intelligence?

Toast Our Friend begins with a thought from Julie's celebration of life: the best stories often arrive too late for the person to hear them. The essay follows Scott Miller, Mike Szwajkowski, Steve Cooper, Tom Shearer, Sung-Hou Kim, and AI-Steve to show how a grief-born tool became a celebration engine for people who are still here.

The piece turns that grief-born insight into a practical invitation: create a circle, collect video toasts, photos, humor, gratitude, and memory, and let the person feel the force of the relationships around them now. The urgent question is simple: who shaped you, and have you told them?

Steve joins Lloyed Lobo on Traction for a wide-ranging conversation that starts with early neural networks for protein structure prediction and moves into AI drug discovery, AI-Steve, AI-Dad, Food Health, and the practical limits of any model when human judgment still matters.

The episode also turns personal: wearables, HRV, CGMs, sleep, recovery, cold exposure, stress, and N-of-1 experimentation. The core point is that data becomes useful only when it is personal, longitudinal, and interpreted with enough calm to avoid turning optimization into another source of cortisol.