I. The two clinics

In the summer of 1865, the Hungarian obstetrician Ignaz Semmelweis was committed to an asylum in Vienna under circumstances that remain disputed and grim. Within two weeks he was dead, at forty-seven, of “blood poisoning,” an infection of nearly the same family as the one he had spent his life trying to prevent. The man who discovered how to stop women from dying of childbed fever died of sepsis, in confinement, largely disbelieved.

Two decades earlier he had been a junior physician at the Vienna General Hospital, which ran two maternity clinics side by side, the first staffed by physicians and medical students, the second by midwives. Women were admitted essentially at random, by the day of the week, and yet the two clinics produced wildly different outcomes. In the first, maternal mortality from puerperal fever hovered near one in ten, and in bad months worse; in the second, a few corridors away, it ran closer to one in twenty-five. The women knew, and they begged for the midwives’ clinic. Some preferred to give birth in the street.

Semmelweis worked the difference like an engineer, ruling out climate, crowding, delivery position, the route of the priest, every fashionable explanation in turn. What remained was the thing no one wanted to name: the physicians, unlike the midwives, performed autopsies and then walked from the dissection room to the delivery ward with nothing more than a cursory rinse. He ordered handwashing in chlorinated lime, and mortality in the first clinic collapsed toward one or two percent within months.

The data was overwhelming and the intervention nearly free, and still it did not spread, not for years, because accepting it meant accepting something intolerable: that the doctors were the vector, that the healers carried the disease from one bed to the next on their own hands. Being right was not enough. Against institutional inertia, it almost never is.

An institution adopts a change only when the change becomes legible in the terms it already uses to keep score. I have come to think this is the single most important thing to understand about any institution, the FDA at this moment included, and it is also what the current coverage least understands. The visible story is disruption; the consequential one is quieter, the conversion of a useful idea into an institutional obligation, the moment a thing that can be done becomes a thing the agency is responsible for doing.

II. One slide

The FDA is now narrated through rupture: reductions in force, reorganizations, senior departures, the accumulating evidence of an agency under strain. Read enough of it, and the conclusion begins to feel foreordained: that the agency is being hollowed out and alarm is the only honest posture. Some of that alarm is warranted, because a public health agency is not an abstraction, and its missteps can eventually reach the bedside.

But there is another account, harder to see because it leaves fewer public traces, and I want to make it from the inside, from the years I spent there, because that is where I learned how reform actually moves through a scientific bureaucracy: slowly, unevenly, and only when a working idea is translated into a form the institution can own.

Some years ago I gave an hour-long talk at the National Academies on the use of artificial intelligence in premarket safety surveillance, and for a subject that invites grand claims I had prepared none. I had one slide for a one-hour talk, a photograph of a computer monitor fringed on every side with yellow sticky notes.

This, I told the room, is how premarket safety is actually done at the FDA. There is no surveillance database behind it. The seven-day and fifteen-day expedited safety reports, the ones a sponsor is legally bound to file when something serious, suspected, and unexpected happens to a patient on an investigational drug, arrive as PDFs or by fax, and are read by a person and remembered, or not, on paper pressed to the edge of a screen. Before we can have a serious conversation about AI, I said, we have to get to data, just data, because the ability to draw a bar graph would already be a profound change.

The room laughed, because the line sounded like exaggeration. It was not.

Stay with that fact and it stops looking like inefficiency. At the investigational stage, before a drug is approved for anything, safety is the FDA’s central statutory authority; efficacy is the sponsor’s burden to prove later, at the NDA or BLA stage, if the program survives the valley of death where most INDs vanish. Safety is the lever the law hands the agency over a drug in human testing, the single hook it is empowered to act on, and behind that lever, for all of the agency’s modern history, sits no real surveillance at all. Sticky notes on a monitor, notes on paper, a spreadsheet someone started and never finished, a reviewer turning to the reviewer at the next desk to ask whether they had seen the same event and whether it meant anything: institutional memory carried in conversation, which is to say carried nowhere durable. What is remarkable is how well the reviewers did anyway, holding in trained judgment what the system gave them no instrument to hold. The deficit was never theirs.

Worse, the machine produced volume without proportionate knowledge. The oncology office was receiving thousands of expedited reports a year, and in an audit we published, we drew a random sample of 160 and read each against the law’s three-part test: serious, unexpected, a suspected reaction to the drug. Twenty-two qualified, fourteen percent; the other 86% were individual case reports that, stripped of any comparator or denominator, could neither raise a true alarm nor retire a false one. The system was at once overloaded and under-informed, activity in abundance and knowledge in deficit.

This is one of the things INFORMED, the data science incubator I founded at the FDA, set out to repair. We built the other version of premarket safety, structured and machine-readable and near real-time, a system that treated a safety report as a row in a dataset rather than a page to be filed and half-remembered, and it worked, not on a slide but on real reports, in conditions close enough to production that the result was a demonstration and not a promise.

Then I made the mistake some institutions teach you to stop making. I brought people at the FDA a number.

At full implementation, by our estimate, structured digital safety reporting would return to the agency at least 500 full-time-equivalent hours a month, five hundred hours of labor lifted off transcription and handed back to judgment. I thought the number would do the work for us, since it was, after all, large enough to seem self-evident.

It did very little.

It took our team an embarrassingly long time to understand why, and that understanding is among the most useful things I carried out of the building. Efficiency is not the FDA’s native currency. No one’s mandate there is measured in hours saved, and no career is made or ended by labor returned to the system or capital reallocated to taxpayers. The agency’s currency is caution, scientific defensibility, and above all the avoidance of the catastrophic miss, the approval that should never have happened, and I had walked in with a strong answer to a question almost no one in the room was paid, promoted, blamed, or remembered for asking.

We had brought Semmelweis’s basin to a room that had not yet been shown what was on its hands. It was not a failure of evidence, nor a failure of the people in it, but a failure of translation: we had shown the system worked in a pilot with large sponsors, but we had not yet made it legible as something the agency was responsible for owning.

III. Where good systems go to wait

A pilot can live on ingenuity, charisma, borrowed staff, and the temporary suspension of normal rules; infrastructure cannot. The moment a pilot must become infrastructure, it needs three things an incubator cannot grant: a permanent owner, a budget line that survives the next reorganization, and a mandate that crosses the boundary between offices with rival incentives. An innovation that belongs to everyone in general belongs to no one in particular, and that is where good ideas go to wait, not in a laboratory but in a gap in the org chart, the institutional form of forgetting.

We built several things that went there to wait: the structured safety system from that slide; the early scaffolding for an expedited IND process, meant to compress the dead time between an investigator’s readiness and an active application; real-world evidence pipelines; biosensor work, the first attempts to gather evidence closer to the patient and farther from the spreadsheet; AI infrastructure, in the pre-LLM era; the harmonization of the agency’s data into a single system. Each one worked, and some stalled the instant their champions left, for reasons that had nothing to do with technical merit and everything to do with custody.

This is why the present moment is more interesting than the headlines allow.

IV. Surgery on the living

Against the grain of the political temperament, I have begun to feel something close to optimism, and the reasons are not atmospheric but concrete: decisions, made by named people, in the last two years, to turn orphaned capabilities into assigned responsibilities.

Start with digital safety reporting, finally being implemented, not because the technology suddenly matured, since it never needed to, but because the only thing ever missing was at last supplied: someone with the authority, willing to spend their political capital, decided to own it and make it the standard across the agency. Once you know to look for that kind of decision, the pattern surfaces everywhere.

The agency’s leadership has lately been dismantling barriers that stood for years. In December 2025, it removed one of the oldest, issuing guidance that lets sponsors submit real-world evidence without first stripping and re-identifying patient data, a technical change with large consequences for the clinical registries and claims databases that had been sitting just out of reach. At nearly the same moment, and for the first time, the agency began to publish an honest accounting of what real-world evidence is actually entering its review pipeline, and the numbers, though modest and easy to over-read, point in a direction that is not in doubt. Protocols submitting real-world data to the drug center roughly tripled in a single year, from 11 to 31, and sponsors are beginning to file studies that use such evidence not merely to watch a drug after approval but to help establish that it works in the first place. Externally controlled designs, all but absent the year before, have started to appear, the approval packages of the decade’s end taking shape now, and the running tally of approvals informed at least in part by real-world evidence has climbed past a dozen.

ISTAND is the same story in a different register, the qualification pathway the agency built for tools that fit through no existing door: digital measures, wearables, AI methods, in silico models of human organs. Its first in silico submission, an AI model of the liver that predicts whether a drug candidate is likely to cause injury, was accepted not long ago, and this matters more than it sounds, because the reason real-world evidence submissions still show almost no contribution from digital health technologies is not that the sensors fail but that there has never been a qualified road from a wrist to a regulatory decision. ISTAND is that road, finally being paved.

An expedited IND initiative has resurfaced too, no longer an incubator sketch but a formal pilot, a rolling submission platform meant to shorten the path from drug identification to first-in-human study and to minimize the clinical holds that stall it. And there is the real-time clinical trial initiative, which I wrote about in The Ship and the Signal, that can reduce the latency between data and the regulator while leaving every safeguard in place.

None of this is disruption for its own sake; it is something harder, closer to surgery on a beating heart, the target moving under the hands while the stakes refuse to pause. That is the work underway at the FDA.

V. The bar that does not move

The agency today is working to modernize its substrate, the speed and structure with which evidence moves, the latency with which a signal is seen, the durability of the record on which judgment rests, while the judgment itself must remain as demanding as it ever was. The good version of these reforms makes the agency faster at being careful, not faster at being permissive.

Even those who concede the agency needs transformation will warn, reasonably, against throwing the baby out with the bathwater, and they are right that there is a baby. The agency’s scientific judgment, its institutional caution, its hard memory of every drug that looked good and was not, is precious and very nearly irreplaceable, and a careless hand could pour it out in a single season and fail to rebuild it in a generation.

But the idiom conceals its central premise, that the bathwater is inert, unpleasant perhaps but harmless in itself, and Semmelweis is the standing refutation, because in his wards the medium was not neutral. The contamination was the thing hurting the patients. The legacy substrate at the FDA, the fax and the PDF and the 86% of reports that are noise, is not merely old; it is a hazard to the very people the agency exists to protect, because a signal seen too late is not an administrative delay but potentially a patient harmed. To guard the bathwater in the name of the baby is a category error, and inaction is not prudence but a quieter way of choosing the greater risk.

The task was never to choose between the baby and the bathwater. It was to change the water without losing the child, and that is the work ahead, and the discipline it asks for: not razing the ward, not denying the fever, but operating on the living.

So I hold both halves of the glass, which is the only honest way to hold it. The disruption is real, and some of it may leave scars; the reinvention is also real, and a meaningful share of it is precisely the kind the agency needed and could not deliver on its own, for structural reasons that had little to do with money or talent. The question is whether the reinvention becomes durable before the disruption becomes the whole story.

VI. How an institution knows

Semmelweis had a system that worked, and the hard part was getting the institution to fold it into its own self-description: this is how we know, this is how we remember, this is how we will notice when reality begins to contradict us. That is the deepest meaning of modernization, not the substitution of software for judgment but the modernization of the way institutions know. A regulator that receives evidence as paper can deliberate only at the speed and fidelity of paper; a regulator that holds evidence as memory can forget; a regulator that receives evidence as data can still be wrong, but it can be wrong in ways that are visible, contestable, and correctable.

Semmelweis’s basin was never only a basin. It converted a moral and scientific fact into a routine, something a hospital performed every day without having to be persuaded again. The new instruments arriving today; AI, sensors, real-world data, real-time trials, earn their place only if they do the same work: forcing the institution to see its own contact with evidence more clearly, what it knows and what it is missing, where delay has become risk, where caution has curdled into concealment, where noise has been mistaken for vigilance.

Half full is the only honest reading, not because the glass is full, nor because the problem is solved, but because the agency has begun to take ownership of the instruments through which it sees the world. Semmelweis could not save himself, and was disbelieved to the end, yet the practice he could not make them accept outlived the person who showed them the stain, and that is the asymmetry worth betting on and the reason to do the work.

In public health, as in medicine, repair begins not with the claim that the hands are clean, but with the discipline that makes unclean hands impossible to ignore.

Subscribe now

I had the honor of speaking with Mike Krzyzewski, known to most of the world simply as Coach K, for a new episode of Precision Signals.

There are many ways to introduce Coach K. You can start with the national championships, the Olympic gold medals, the decades at Duke, or the generations of players and coaches whose lives he helped shape. But what stayed with me most from this conversation was not the résumé. It was the way he thinks about people.

Coach K has spent his life studying how individuals become a team. Not in the shallow sense of putting talented people in the same room, but in the deeper sense of building the trust, discipline, and shared purpose that allow people to do difficult things together.

That question has obvious relevance to sports. It also has deep relevance to medicine.

In biomedicine, we often talk about progress through the things we can name: a new therapy, a sharper biomarker, a better trial, a more powerful model, a new algorithm. Those things matter but none of them becomes meaningful in isolation. A discovery becomes progress only when people carry it across the distance between an idea and a patient’s life.

That distance is scientific, but it is also human.

It depends on whether researchers can ask better questions, whether clinicians can interpret evidence with judgment, whether regulators can make decisions under uncertainty, whether institutions can work across boundaries, and whether patients can trust the people and systems around them. The movement from discovery to impact is not just a technical process. It is a human one.

That’s why I wanted to have this conversation with Coach K.

His philosophy of leadership began long before Duke, in Chicago, with parents who taught him about work, faith, family, and the confidence that comes from knowing someone believes in you. It continued at West Point and in the Army, where leadership was not an abstract concept but a responsibility carried in front of other people. It matured at Duke and with Team USA, where talent was never the whole story. The real work was creating the conditions in which talent could become trust, and trust could become performance under pressure.

We also talked about his work in cancer advocacy and his connection to the CEO Roundtable on Cancer. That history matters to me because cancer is not a problem any one institution can solve alone. It requires science, prevention, clinical research, data sharing, regulatory judgment, patient advocacy, and care delivery to move together in ways that are difficult but necessary.

The same theme came up when we discussed artificial intelligence in medicine. Coach K does not approach AI as a technologist, which is part of what made his perspective so valuable. He approaches it as someone who understands time, attention, and trust. If AI can reduce the burden that keeps clinicians staring at screens instead of patients, if it can give doctors more space to listen, explain, and be fully present, then it can serve something profoundly human.

That is the version of AI in medicine I care about most. Not AI as spectacle. Not AI as a substitute for judgment, empathy, or responsibility. But AI as a way to remove the friction that keeps people from doing the work only people can do.

And yes, because I’m a New Yorker, I asked him about the Knicks.

His answer began with Madison Square Garden and the force of a fan base that had found its voice again, but it returned to the same idea that runs through the entire conversation: belief. What happens when people feel part of something together. What changes when a team, a city, an institution, or a community begins to trust again.

That is what I kept coming back to after the conversation ended.

Basketball, medicine, cancer, technology, leadership. These may seem like separate worlds, but they share a common problem. Talent is not enough. Technology is not enough. Ambition is not enough. The real question is whether people can build the trust and discipline required to turn individual effort into collective progress.

That is the human work behind progress.

I’m grateful to Coach K for joining Precision Signals and for a conversation that moved across basketball, medicine, AI, cancer advocacy, and the deeper question of how we do hard things together.

Subscribe now

I. Chicago, May 31

Like many of my colleagues, I was in the room for the phase 3 results of RASolute 302, a randomized trial testing daraxonrasib, a first-in-class oral inhibitor of active RAS, against standard chemotherapy in patients with previously treated metastatic pancreatic cancer.

The presentation was a late-breaking plenary at ASCO’s 2026 annual meeting at McCormick Place in Chicago. When the Kaplan-Meier curves appeared on the screen, the two survival lines pulling apart with an unusually wide and stable separation, the audience stood and applauded before Brian Wolpin, director of the Hale Family Center for Pancreatic Cancer Research at Dana-Farber and lead investigator of the trial, had finished speaking.

It lasted 42 seconds.

When the applause subsided, Wolpin noted that the time had not been built into his talk.

These results deserve to be understood on their clinical merits, which are substantial. But they also deserve to be understood as the product of a particular kind of scientific thinking, one that required forty years of accumulated failure before we could ask the right question.

II. The disease that has resisted almost everything

Pancreatic ductal adenocarcinoma kills most of the people it touches. In 2026, roughly 67,500 Americans will be diagnosed, more than half with metastatic disease at presentation. For those patients, the five-year survival rate is around 3%. When first-line chemotherapy stops working (and it almost always does), second-line treatment has historically offered a median overall survival of 6 to 7 months and a median progression-free survival of 3 to 4 months, at the cost of substantial toxicity.

One of the central molecular drivers has been known for decades. Oncogenic RAS mutations are present in more than 90% of pancreatic tumors, most involving substitutions at KRAS codon 12. These mutations lock RAS in a GTP-bound active state, driving proliferative signaling the cell cannot turn off. The therapeutic conclusion has always seemed obvious: inhibit active RAS, and you could reach nearly every patient.

For forty years, nobody could.

III. The geometry of failure

Gilbert Ryle, the British philosopher, introduced the concept of a category mistake in 1949 to describe what happens when someone searches for a thing in a place where, by the nature of the category being used, it cannot exist. His canonical example is a visitor touring Oxford who walks past every college, every library, every building, and then asks: but where is the University? She had seen the University. She failed to recognize it because she was looking for a different kind of thing. The error was conceptual before it was empirical.

The forty-year failure to inhibit RAS-GTP was, at its root, a category mistake of this kind.

Drug discovery has long operated within what is known as a substance ontology: a framework in which the fundamental objects of interest are independent entities that bear properties. Proteins have pockets. Targets have binding sites. A protein is druggable or it is not. RAS-GTP presented exactly the kind of problem this framework was least equipped to solve: a smooth, largely featureless surface, no deep binding pocket, and picomolar affinity for its natural substrate GTP, making competitive displacement essentially impossible. We treated druggability as an intrinsic property of the protein itself, found no such property in active RAS, and concluded it could not be done. So we looked downstream.

In 2013, Kevan Shokat, professor of cellular and molecular pharmacology at UCSF and a co-founder of Revolution Medicines, cracked this open by identifying a transient pocket on the KRAS G12C mutant. That insight led to sotorasib and adagrasib, validated RAS as a pharmacological target, and changed the conversation. But these drugs bind KRAS in its GDP-bound inactive conformation; they reach RAS only when it is already off. And G12C is rare in pancreatic cancer. For the patients who most needed RAS inhibition, the first generation was, by design, largely beside the point.

The central question remained: how do you inhibit active RAS, GTP-bound and constitutively firing?

IV. What nature knew, and what we failed to generalize

The breakthrough that produced daraxonrasib did not come from finding a hidden pocket on RAS. It came from escaping the assumption that druggability had to be a property of RAS itself. The decisive question was relational: what if the drug did not have to bind RAS alone?

This required abandoning an assumption so deeply embedded in traditional drug discovery that most of us had stopped noticing it as an assumption. One target, one protein, one binding site. The paradigm is powerful and ubiquitous. For active RAS, it was a dead end.

What is striking in retrospect is that nature had been demonstrating a different paradigm for decades, and we had been watching. Rapamycin does not inhibit mTOR by occupying a pre-existing mTOR pocket; it first binds FKBP12, and the resulting complex engages mTOR through an interface neither partner could access alone. Cyclosporin A recruits cyclophilin A to inhibit calcineurin. Thalidomide forces cereblon to bind and degrade transcription factors it would otherwise ignore. These are among the most consequential drugs in medicine, and they share a common logic: a small molecule conscripts an endogenous protein to create a new interaction that blocks a biological process indirectly. Before active RAS could be drugged, nature had already shown the way.

But none of these drugs emerged from a rational design program. Rapamycin is a natural product isolated from a soil bacterium found on Easter Island in 1972. Cyclosporin A came from a Norwegian fungus identified in a Sandoz screening program. Thalidomide was a sedative whose mechanism in cancer was reverse-engineered decades after it was withdrawn from the market for causing birth defects. In each case the sequence was the same: discover the molecule, observe what it does, spend years understanding why. These drugs did not begin as a design principle. They ended up revealing one, and we largely filed the lesson away. Each mechanism was admired as elegant and categorized under its own domain (immunophilin pharmacology, cereblon biology) without asking whether the underlying logic could be deliberately applied to problems that had resisted direct inhibition.

This is a recognizable pattern in science. Paradigm-shifting evidence accumulates in plain sight for years before someone asks whether it generalizes. The data were there. The inference was not.

Shokat’s group made that inference in 2019, publishing proof-of-concept work showing that bifunctional ligands could recruit cyclophilin A to GTP-bound RAS and block its interaction with B-Raf. This was not yet a drug. But it was the answer to a question we had not been asking.

V. Sanglifehrin A: the handle that made it possible

Revolution Medicines, founded in 2014 by Martin Burke, a chemist and natural product synthesis pioneer at the University of Illinois, Michael Fischbach, a microbiologist and biosynthesis expert at Stanford, and Shokat himself, had an explicit founding thesis: natural products could be reverse-engineered into precision medicines for targets that conventional drug discovery could not reach. The company now had the concept. It needed the chemistry.

The scaffold came from a natural product discovered by Novartis in 1999 and never developed as a drug: sanglifehrin A, a bacterial macrocycle with high affinity for cyclophilin A. Its value was structural. Its binding mode leaves a solvent-exposed face pointing outward, a chemical vector free to be linked to a RAS-binding pharmacophore. One face buried in cyclophilin A, the other engineered toward RAS-GTP.

The team built sanglifehrin-inspired bifunctional molecules and solved the crystal structures of the resulting tricomplexes with KRASG12C. What those structures revealed could not have been predicted from studying either protein in isolation. The drug sits in a composite pocket that is partly cyclophilin A and partly RAS: a neomorphic interface that does not exist when either protein is present without the other. The molecule remodels cyclophilin A’s surface to create a high-affinity interface with active KRASG12C. The pocket was not discovered. It was created.

VI. From G12C to every RAS: the making of daraxonrasib

The first clinical molecules from this platform, elironrasib for G12C and zoldonrasib for G12D, were mutation-selective, exploiting covalent chemistry anchored to specific mutant residues. Elegant, but insufficient for pancreatic cancer, where the full spectrum of oncogenic RAS variants collectively accounts for nearly the entire patient population.

Daraxonrasib was built around a broader premise: engineer a molecule that engages the conserved switch regions shared across all oncogenic RAS variants rather than the residues that distinguish one mutation from another. The result is a reversible, noncovalent molecular glue: macrocyclic, orally bioavailable, mutation-agnostic.

The mechanism is worth holding precisely. Daraxonrasib binds cyclophilin A first, forming a binary complex that then engages RAS-GTP across the composite interface. The tricomplex blocks downstream effectors (RAF, PI3K) from reaching the RAS surface. The drug does not fit into RAS. It fits into a relationship between two proteins, a relationship it calls into being. For G12X mutants there is an additional effect: daraxonrasib can stimulate GTP hydrolysis, actively accelerating RAS back toward its inactive state rather than merely blocking its downstream signals.

That distinction matters. Daraxonrasib is not a better RAS inhibitor in the conventional sense. It is a different way of thinking about what an inhibitor can be.

VII. What 13.2 months means

RASolute 302 enrolled 500 patients across 59 sites in six countries, randomized 1:1 between daraxonrasib 300 mg once daily and investigator’s choice of chemotherapy. Results were published in the New England Journal of Medicine the same day they were presented at ASCO.

Of enrolled patients, 91.8% had RAS G12 mutations, the primary endpoint population. In that population, median overall survival was 13.2 months with daraxonrasib versus 6.6 months with chemotherapy, a hazard ratio of 0.40 representing a 60% reduction in the risk of death. At 12 months, 53.3% of daraxonrasib patients were alive versus 18.7% on chemotherapy. Median PFS was 7.3 versus 3.5 months. Objective response rate was 33.2% versus 11.8%.

The context that makes 13.2 months remarkable: the landmark first-line regimens in this disease (FOLFIRINOX, gemcitabine plus nab-paclitaxel, NALIRIFOX) achieved median overall survival of 8.5 to 11.1 months in previously untreated patients. Daraxonrasib, in patients who had already exhausted those regimens, compares favorably with all of them.

Beyond survival, time to pain deterioration was 9.0 versus 3.7 months. Patients were not only living longer. They were feeling better longer. Treatment-related discontinuation occurred in 1.2% of daraxonrasib patients versus 11.2% on chemotherapy. Hematologic toxicities and peripheral neuropathy, the defining burdens of chemotherapy in this disease, were largely absent.

FDA expanded access was authorized in May 2026. The broader question remains: how quickly can a system built around sequential evidence review deliver a result this large to the patients who need it?

VIII. A question worth sitting with

There is something in this story worth dissecting.

When Wolpin showed those survival curves at ASCO, the knowledge that daraxonrasib halves mortality in metastatic pancreatic cancer became public. But that knowledge existed at the data cutoff in February. It existed, in effect, from the moment the trial was powered enough to see it. And yet the patients sitting in oncology clinics that same week, the ones not enrolled in RASolute 302 and not navigating expanded access, were receiving second-line chemotherapy with a median survival of under seven months, because the drug that could double that had not yet cleared the regulatory process required to reach them.

This is the gap we need to talk more about. Not a failure of any single institution, not a problem unique to daraxonrasib, but a structural feature of how we move evidence from signal to treatment. The trial completed enrollment in November 2025. The data cutoff was February 2026. Results were presented publicly in June. Submission, review, approval, and prescribing infrastructure still lie ahead. Each step is defensible in isolation. Together, they add up to months; and in pancreatic cancer, months are the unit of survival.

The architecture we built to evaluate evidence was designed for an era of slower science, slower information, and limited computational capacity. It made sense then. The question worth asking now is whether a system built around collecting data, freezing it, submitting it, and reviewing it as a completed artifact is still the right model when the signal is a hazard ratio of 0.40 on a pre-specified primary endpoint in a 500-patient randomized trial, and the disease being treated measures survival in single-digit months.

IX. The regulatory paradigm needs its own conceptual leap

The daraxonrasib story is about changing the unit of analysis. We could not inhibit active RAS because we were looking for a single-protein binding site that did not exist. The breakthrough came when someone changed the question from “where is the pocket?” to “what relationship could we create?” The same logic applies to how we move evidence from trial to patient.

The current model treats drug evaluation as a sequential, bounded process: conduct the trial, freeze the data, submit a dossier, review it, decide. I called this the HMS Salisbury model in an earlier essay, “The ship and the signal”: the idea that medical evidence must be collected, transported, and unpacked before it can inform a decision, the way James Lind’s citrus observations had to cross the Atlantic before anyone acted on them. It was the right architecture for its time. It is increasingly misaligned with a world of molecularly targeted therapies, large and unambiguous signals, and diseases that measure survival in single-digit months.

The FDA’s first real-time clinical trial pilot (AstraZeneca’s TRAVERSE study in mantle cell lymphoma and Amgen’s STREAM-SCLC study in limited-stage small cell lung cancer) proves the model can change. In that framework, endpoint and safety signals reach reviewers as they emerge rather than arriving frozen in a dossier after trial completion. It is not yet a transformation. It is a proof of concept, and a sign that a fundamentally different paradigm is within reach.

X. What the standing ovation was really for

I have been to many ASCO plenary sessions. I have not seen many moments like the one on Sunday.

I think people standing in that hall were not applauding a hazard ratio. They were applauding the end of a forty-year impasse in a disease where failure had become the baseline expectation, and where the same grim prognosis had been delivered to patients year after year despite everything the field had tried.

But the full measure of what Sunday means depends on what we do next.

The scientific paradigm changed when someone asked a different question about where a binding site could be found. The regulatory paradigm needs the same move: away from evidence as a frozen artifact reviewed after the fact, toward evidence as a continuous signal evaluated in real time. The science and the system have to advance together, or the science advances into a bottleneck.

It took forty years of failure, a philosopher’s insight about category mistakes, three natural products that nobody designed, a bacterial macrocycle that sat unused for twenty-five years, and a company founded on the belief that nature encodes answers to questions we have not yet learned to ask. The regulatory infrastructure capable of delivering the resulting therapy to patients without avoidable delay does not fully exist yet either.

Creating it will require the same thing that created daraxonrasib: the willingness to question whether the current frame is the right one.

The question changed. That is why the room stood up.

Now the room has to go to work.

Subscribe now

Every era of medicine settles on its own metaphor, and ours has chosen precision.

The word is everywhere now. Precision oncology. Precision diagnostics. Precision medicine. It suggests a field that has learned to aim, a field that has moved beyond the blunt instruments of the past and into a more exact relationship with disease. But precision in oncology is not a destination. It is a discipline: the patient work of asking what has gone wrong in a cell, whether that wrongness is essential to the cancer, and whether the answer might be small enough to hold in a single molecule.

Dr. Brian Druker built his career inside that question.

In the latest episode of Precision Signals, I traced the arc of his life and the arc of the field he helped invent. The conversation kept returning to a simple observation: the questions he sat with thirty years ago, about targets, mechanisms, and the courage to test an unfashionable idea in humans, are the questions oncology has not finished answering.

Brian grew up in St. Paul, the youngest of four children of first-generation immigrants from Eastern Europe. His father was a chemist who held patents on printing processes used by 3M, and the expectation that one of the children would become a doctor lived more in the atmosphere than in explicit instruction. As a boy, Brian wanted to play baseball. Science, it turned out, had other plans for him.

He landed at UC San Diego for undergrad, where he worked in John Abelson’s lab purifying restriction enzymes. He stayed for medical school, then went to Barnes Hospital in St. Louis for residency, which he still describes as one of the richest learning experiences of his life.

Somewhere along the way, he wrote in a class paper that the field would only make tangible progress by understanding what distinguishes a cancer cell from a normal cell. It was the kind of sentence a student writes before knowing how long a life can be organized around a single premise. It became the thesis statement for everything that followed.

That premise eventually led him to chronic myeloid leukemia. CML was an unusual cancer in one crucial respect: its biology had a visible center. The Philadelphia chromosome produced the BCR-ABL fusion kinase, and the disease depended on that aberrant signal. If one could inhibit ABL without damaging normal cells, the logic was clean. The question was whether such a drug could exist.

The imatinib story is now told as if inevitability was hiding inside it. It was not.

Within weeks of arriving at OHSU in 1993, Brian called Nick Lydon at Ciba-Geigy and asked whether the company had anything that might inhibit ABL. Lydon said they did. Within three months, Brian had data showing he could kill leukemia cells without harming normal ones. That sentence now sounds like the beginning of a revolution. At the time, it sounded implausible enough that both Nature and Science declined the foundational paper.

The skepticism was not irrational. The dominant view held that any ATP-competitive inhibitor would shut down too many kinases in the body. Even if the biology worked, the market appeared small: perhaps 5,000 CML patients a year. For a large company, this did not look like the kind of opportunity around which one reorganized a development program.

It took five years to convince Novartis to enter clinical trials. By 1997, Brian brought his data directly to a toxicologist at the FDA. The first human trial began in 1998, restricted to Philadelphia chromosome positive patients. His first patient was a train conductor from the Oregon coast. The leukemia went into remission.

Within six months, 53 of 54 patients had responded.

Gleevec was approved in May 2001 by the FDA. It landed on the cover of Time. The life expectancy of patients with CML moved from three to five years to an 89 percent five-year survival. A disease that had carried the emotional architecture of fatalism became, for many patients, a chronic condition. The molecule then extended into gastrointestinal stromal tumors once KIT came into view, offering the first hint that a single targeted agent might cross disease boundaries when the biology was shared.

This is the part of the story the field remembers. It is also the part some have sometimes misunderstood.

Imatinib did not prove that every cancer has a single switch waiting to be turned off. It proved that, in the right disease, with the right dependency, the right molecule can rearrange the future. CML was unusually legible. The target was not merely present. It was central.

The hundred kinase inhibitors that followed imatinib produced cycles of enthusiasm and disappointment partly because oncology internalized the wrong lesson. The field treated the imatinib paradigm at the time as a template when it was also a special case. It tried to export the form without always having the underlying biology.

That distinction matters.

In most tumors, the biology is distributed, redundant, adaptive, and heterogeneous. A single lesion rarely explains the whole disease. The cancer cell sits inside tissue, immune pressure, metabolic context, and time. This is why Brian’s argument today for bolder combination strategies carries weight. Combination therapy is easy to praise and hard to do. It multiplies toxicity, development complexity, regulatory uncertainty, and commercial coordination. But some diseases will not yield to the elegance of a single molecule because their biology is not organized around a single point of failure. In that setting, precision cannot mean one target, one drug, one patient. It has to mean understanding enough of the system to intervene at more than one essential point.

The same humility should govern how we talk about genomic data. Oncology has generated a vast amount of it. The clinical gain has been real, but not proportional to the scale of sequencing. A mutation list is not a disease model. A pathway diagram is not an intervention strategy. A biomarker is not a mechanism unless the biology makes it one.

This is where the conversation turned naturally to AI.

Brian was candid about what AI can and cannot do for biology. He drew on a Jensen Huang anecdote to make a point that stayed with me: physics has fixed laws; biology does not. AI is most useful where the underlying rules are at least partly knowable. In biology, the rules are contingent, context-dependent, and often hidden behind layers of measurement error and incomplete observation.

AI will matter enormously. It will find patterns, generate hypotheses, design molecules, read images, organize literature, and compress work that used to consume whole careers. But AI cannot rescue a field from not understanding the biology it is asking the model to optimize. I always emphasize that when the training data captures proxies, the model learns proxies. If the endpoints are noisy, the model learns noise with confidence. If the disease mechanism is only partially understood, the model may accelerate motion without improving direction.

The next quarter century in oncology may be defined less by better targets or more data than by the harder work of understanding biology well enough that our tools, AI included, finally have rules to follow.

The human texture of my conversation with Brain was quite obvious near the end, in a story about a letter Brian recently received from a patient named Kevin in Jakarta. Kevin had been newly diagnosed with CML. He wrote to thank Brian for the medicine that allowed him to plan a future with his children.

That is the standard that matters.

There is a final grace note in Brian’s story. In the late 1990s, Brian was interviewed by an AP reporter who had grown accustomed to press releases announcing cancer cures that never materialized. Her notes captured a careful skepticism: nice guy, good with his patients for a researcher, but that drug is not going anywhere.

That reporter is now his wife.

It is a funny ending, but also a useful one because the skepticism was not foolish. Unfortunately, most cancer-cure stories do fail. Most early discovery breakthroughs do not survive contact with patients. I think we need skeptics because patients need us to distinguish hope from evidence.

What made imatinib different was not that it avoided skepticism but that it survived it. And that may be the best definition of precision oncology we have:

Following a biological idea far enough, and carefully enough, that it can meet a patient and still be true.

Subscribe now

In nine days, thousands of oncologists will converge on McCormick Place in Chicago for the ASCO Annual Meeting. As it does every year, the field will perform its ritual of compression: a year of progress distilled into five days of plenaries, posters, hallway conversations, late dinners, and hurried meetings between sessions.

The person responsible for the machinery behind that ritual is Dr. Clifford Hudis, ASCO’s CEO, and my guest on the latest episode of Precision Signals.

Cliff grew up in Northeast Philadelphia in the 1960s in what he describes as a modest, middle-class household. His father worked continuously from the age of eighteen. His mother went back to school and became a public school teacher. There were no physicians in the family. The early scaffolding came from a cousin training to be an allergist, who served as an informal mentor, and from a summer in eighth grade spent in a laboratory at Hahnemann, grinding up earthworm muscle for someone else’s physiology experiments.

The public schools of that era were investing seriously in math and science. His parents, recognizing something in him, took out a loan to send him to a small Quaker school where AP biology and chemistry classes had only three or four students. From there he entered a six-year accelerated BA/MD program at Lehigh and the Medical College of Pennsylvania and graduated from medical school at twenty.

His internal medicine residency in Philadelphia brought him into oncology through a clinic full of grateful patients and still-incomplete science. In the mid-1980s, he moved to New York and was running the emergency room at Mount Sinai, where the oncology service operated as its own department under James Holland. It was there that Cliff first encountered Dr. Larry Norton from a distance.

Memorial Sloan Kettering had already taken root in his imagination years earlier through an unlikely source: Brian’s Song, the James Caan film in which Chicago Bears running back Brian Piccolo dies of a midline germ cell tumor at Memorial. In 1988, Cliff arrived at MSK as a fellow. That same fall, Larry Norton was recruited from Mount Sinai to launch the breast cancer service.

The team’s first home was not a gleaming cancer center floor. It was a two-bedroom apartment next to the hospital, with Cliff wedged into a corner of the living room.

What followed was an extraordinary decade. G-CSF. Paclitaxel. Trastuzumab. BRCA testing. Aromatase inhibitors. Capecitabine. In roughly ten years, breast cancer treatment changed in ways that still shape practice today. US breast cancer mortality peaked around 1990 and has since fallen by roughly 40%, driven in large part by systemic therapy advances that emerged from that era.

Cliff became Chief of the Breast Service in 1998, succeeding Larry Norton. But throughout our conversation, he is careful to reject any heroic version of the story. He describes himself as a collaborator, one person helping translate a much larger collective enterprise into clinical practice.

His path to ASCO also ran partly through Larry. Norton’s deep involvement with the cooperative groups, especially CALGB, pulled Cliff beyond the walls of Memorial and into the broader oncology community. Cliff eventually won a seat on the ASCO board as treasurer, a position that was directly elected at the time. He spent three years receiving what he calls an informal one-on-one MBA in how a professional society actually works.

He later stood for ASCO president, and his term coincided with the society’s fiftieth anniversary. The theme he chose was Science and Society, a phrase that now feels strikingly prescient. It reflected a decade of flat NIH and NCI funding, amounting to roughly a 25% real-dollar cut after inflation, and the early erosion of public trust in scientific methods. During that period, he found himself knocking on doors in the House and Senate, asking in one form or another when the United States had decided to cede scientific leadership.

When Allen Lichter stepped down after a decade as ASCO’s CEO, Cliff put himself forward. He inherited an organization of five hundred people, sixteen direct reports, and no fully integrated strategic plan. Over the next eighteen months, he reorganized it carefully. He is characteristically humble about the transition. There was, as he puts it, no universe in which he was qualified on his own to run ASCO. But the institution was strong enough to tolerate a new chief executive and bring him along.

On AI, Cliff offers an analogy I have kept thinking about. The future AI-powered electronic medical record, he suggests, may function like the lane-keeping and adaptive cruise control systems already built into ordinary cars. Clinicians will still feel autonomous when they are practicing within the lane, but the system will gently nudge them when they begin to drift.

That view is reflected in ASCO’s collaboration with Google Cloud, which produced a constrained Gemini chatbot that reads only from ASCO material, runs with its creativity dialed to zero, and highlights sources for every sentence it returns. Cliff suspects that the chatbot itself may look quaint in a decade, the way a journal devoted to indoor plumbing would look quaint today. The more durable shift is toward living guidelines, refreshed on an approximately eight-week cadence.

There is an implication here that Cliff does not overstate, but the listener can feel it. If oncology knowledge becomes truly continuous, the ASCO Annual Meeting of 2036 will not look like the Annual Meeting of 2026. The ritual of compression may give way to something more fluid, more continuous, and perhaps more closely aligned with how science now moves.

Cliff closes with a line he borrowed from Albert Bourla about visionary leadership and competent skepticism. It is the right note to carry into McCormick Place next week.

Subscribe now

0. Pre-conditions: three things that must hold

This essay is about projections, and every projection of this kind smuggles in assumptions. And more than me, it is the assumptions that do the actual work. In writing this essay, I have smuggled in three assumptions, pre-conditions that need to be met. The timelines I outline only hold if they hold. If they fail, what follows may not be slower or harder. It may be a different essay entirely.

My first assumption is that the instruments we use to measure clinical benefit continue to be repaired in plain sight. Tumor response by RECIST in oncology. Depression scales in psychiatry. Six-minute walk distance in cardiology. These are not facts about the world by any means. They are conventions we trained ourselves to agree on, and in the case of RECIST, independent expert readers disagreed on whether a patient had progressed in roughly 30% of more than 13,000 paired assessments in an FDA-led analysis. An AI trained on those endpoints does not transmute their noise; it inherits it, and reproduces it at scale, with more confidence than is warranted. My first precondition is that novel endpoint development becomes regulatory-grade work, with appropriate clinical validation, publishable noise floors, and audit trails, and is not run within firms competing on the same flawed reference frame.

The second assumption is that human data accumulates the way infrastructure accumulates, not the way leftovers accumulate. Longitudinal, multimodal, reuse-consented data on actual patients across actual disease courses is the binding input. It is the thing models still cannot generate on their own. If the field treats it as the residue of clinical trials, never to be shared for secondary use research, the field will keep training on poor proxies and call them “ground truth.” There are no ground truths and no gold standards, only better approximations. If the field treats data liquidity as deliberate infrastructure, with consent frameworks that hold across decades and storage that survives platform turnover, the next ten years are productive. The fork is happening now, and it is not technical. It is contractual and architectural.

My third assumption is that the public trust, which permits clinical research at all, does not collapse during the transition. Trust is older than every tool stacked on top of it, and it is replaceable in the way load-bearing walls are, which is to say, slowly and with care. A trust failure in this decade, whether through a high-profile AI-generated mistake affecting a patient, or through opacity in how synthetic controls and digital endpoints are incorporated into evidence packages, can cost the field a decade of optionality. The precondition is that the people running the most ambitious programs are also the people most attentive to the slow social work of explaining, disclosing, and refusing to oversell.

AI capex is not AI capability. Saying so out loud, repeatedly, while building, is part of the build.

If my three assumptions hold, the timeline that follows is achievable. If they do not, the failure mode is the one we recognize: another decade in which capability claims outrun the scaffolding built to validate them. The argument that follows is conditional on the scaffolding being repaired in time.

I. Where AI stalls in biomedicine

AI is decisively powerful wherever a problem is bounded and the space of possible answers is already densely mapped. This is why it has functionally solved protein structure prediction, why it is compressing molecular design, and why retrieval and summarization in regulatory and pharmacovigilance work are no longer hypothetical. The problem space is finite. The data is dense. The verification loop is short.

A blunter way to put this: most of what we currently call AI in drug discovery is engineering, not biology. Protein structure prediction, molecular design, pattern recognition in pathology and radiology, and retrieval and summarization across regulatory documents are engineering tasks performed on dense priors with short verification loops. They are real work, and they are real progress, and the speed they introduce is genuine. They are not, however, the same activity as understanding why a disease behaves the way it does in a body. Conflating the two is the move that produces the cycle in which capability claims outrun the scaffolding.

AI is strikingly less useful wherever the question is open and the human data is thin, which is most of what determines whether a drug actually works inside a body. The biology of disease, the heterogeneity of response, the interaction between a therapy and a person in motion across years: these are not bounded problems with dense priors. They are sparse, partly observed, and partly unknowable. Models trained on what we have so far hit a ceiling that comes from us, not from them. And the emergent properties of AI can only stretch so far from the training data into new territories of scientific discovery, thus not a reliable hedge.

The strategic consequence of this asymmetry is that value shifts toward the inputs the model cannot supply itself. The argument that follows is an argument about location, and not about pace: AI is fast but the constraints move faster and can be fatal.

II. Discovery in 2031: abundance without advantage

By 2031, candidate molecules, targets, and mechanistic hypotheses will be generated faster than we can plausibly test them. Discovery platforms will be cheaper, broadly available, and increasingly indistinguishable across organizations. The position that depends on owning a discovery engine ages poorly, because abundance usually confers no advantage. You cannot win at a thing everyone has.

This is the first reordering, and it is worth naming clearly, because it inverts a decade of strategic instinct. The companies that bet their next decade on proprietary discovery platforms are betting on a scarcity they themselves will help dissolve. The molecules that emerge will be high-volume, sometimes elegant, occasionally truly novel, and often undifferentiated in origin. The question that will matter in 2031, more than which model produced a candidate, is which organization can test that candidate most informatively in actual patients, on a timeline a competitor cannot match.

III. Clinical development in 2031: where the bottleneck moves

Once hypotheses are cheaper and more ubiquitous, the binding constraint becomes validation. This is the move worth dwelling on, because it is not how the press currently frames the field. AI-discovered molecules in the clinic so far have cleared early-phase safety testing at unremarkable rates and have not been more successful in efficacy trials. Many still fail in humans because the underlying biology of the disease remains poorly defined. The model can write a faster answer to a question we are still asking poorly.

The trial of 2031 may start to look different from the trial of 2026 in ways that compound. AI-enabled patient selection. External and synthetic control arms where regulators permit. Biomarker-rich adaptive designs. Digital endpoints, where regulators permit. Real-world data integrated rather than appended. The organizations that run those trials best, smaller, faster, and more informatively, will define the next decade in the way the organizations that ran the first kinase and checkpoint inhibitor trials defined a previous one.

The point worth naming, because it runs counter to the prevailing framing, is that clinical development becomes more central, not less, as AI matures upstream. The bottleneck does not disappear when discovery becomes abundant. It relocates. Value follows it. And the bottleneck does not arrive in clinical development alone. It arrives, more deeply, in the biology the clinical trials are testing.

IV. Disease biology: the constraint that cannot be owned

Adjacent to the clinical bottleneck is a problem that it makes visible. When AI-discovered molecules fail in humans, the failure is usually not in the molecule, and usually not in the trial. It is that the biology of the disease was incompletely understood when the molecule was designed. The engineering layer of AI in drug discovery, the protein structures and molecular designs that the field has begun to produce at scale, sits above this absent layer. It cannot replace it. The model could not have known what was not in its training data, and the field had not yet captured the biology in a form the model could read.

This is the upstream constraint and it is the one most easily mistaken for a problem AI will solve on its own. By 2031, the binding question across many therapeutic areas will not be “can we generate a candidate” but “do we understand this disease well enough for any candidate to work.” Multimodal data on actual patients, genomics linked to transcriptomics linked to proteomics linked to imaging linked to longitudinal clinical experience linked to the patient’s own account of what changed, is the substrate from which understanding accumulates. It is also, again, the input AI cannot manufacture on its own.

I must point out that a caveat lives inside my argument: even when disease biology is finally understood, the insight may not confer a durable competitive advantage. Fundamental biological breakthroughs are hard to own. The PD-1/PD-L1 checkpoint axis emerged from shared academic work in Kyoto, Houston, and Boston, with the canonical checkpoint inhibitors distributed across multiple sponsors in close sequence (albeit initially concentrated in a single company that happened to sit on the right therapy at the right time following a multiasset acquisition). CRISPR-Cas9 mechanism diffused into thousands of laboratories within a year of the foundational paper, with downstream therapies developed by several entrants in parallel and the underlying biology eventually litigated rather than monopolized. GLP-1 receptor biology was public for decades before the molecules that crystallized it into a commercial category were developed. The foundational breakthrough in protein structure prediction, the most consequential computational result of the last decade, was open-sourced into something close to public infrastructure within a year of its release.

The pattern is consistent. Breakthroughs in biology move the field forward, not single firms within it. The organizations that capture advantage are typically the first to translate the insight into a specific molecule, a specific trial design, or a specific commercial channel, and the advantage they capture is often in the implementation, not in the underlying biology. Speed and luck matter more than understanding, once the understanding exists. This is a bit uncomfortable to write, because it implies that the most ambitious scientific work, the work whose value to patients is largest in absolute terms, is also the work whose value to a particular company is the most contingent on timing, and sometimes luck.

The strategic reading here is the same as the rest of this essay, with one addition:

invest in the biology, because the biology is the binding constraint on whether anyone’s molecule works

I think it is also critical not to assume that investment alone will yield a proprietary position. The position, when it comes, may be at the next layer down: which organization could assemble the multimodal data to see the insight first, which organization could expeditiously design the study and conduct the trial that converts the insight into proof, and which organization already holds the channel into the patient population the insight reorders. By 2036, the insight itself may belong to everyone, at an accelerated pace that will surpass today’s publication cycles in peer-reviewed journals and abstract embargoes at scientific meetings. The implementation, if you were fast or lucky, will be yours for a window.

V. Human data in 2036: what cannot be cloned

If the five-year story is about relocation, the ten-year story is about fidelity.

By 2036, the models will have converged. The performance gap between the best and the second-best discovery model, the best and the second-best regulatory drafter, and the best and the second-best biomarker pattern recognizer will narrow toward a band in which the model itself is no longer the differentiator. What will differ is what each was trained on and what each can continue to be trained on.

Here, I think, is the scarce input of the next decade: high-fidelity human biology. Longitudinal, multimodal, reuse-consented data on actual patients across actual disease courses. Imaging linked to molecular data linked to outcomes linked to patient experience, persisted long enough to support questions that cannot be answered inside a single traditional clinical trial. An organization that has assembled this kind of position cannot be cloned by a competitor with the same access to off-the-shelf models, because the inputs cannot be re-derived from public data or purchased after the fact.

Treating data as residue, rather than as infrastructure, compounds across a decade the same way poor measurement compounds across a single trial. The errors propagate. Models trained on the result magnify them. The magnification reads, to readers and to regulators, as confidence. Models trained on proxies for biology learn the proxies. Models trained on real biology, captured deliberately, learn the biology. The choice between residue and infrastructure has always been available. The window in which making it correctly still confers a durable competitive position closes inside this ten-year frame.

VI. Commercial interaction in 2036: the agent at the door

The commercial and medical interaction layer of the industry, the way a therapy reaches a prescriber and a patient, is being rebuilt around AI agents on a horizon shorter than ten years. Field force economics, medical information delivery, payer engagement, patient support, formulary access: each of these can move toward AI-mediated interaction at scale. A trusted channel, once established, is itself scarce.

The risk inside this transition is the same as the risk in every previous transition. Trust mediated through a new layer is easier to lose than to build. Patients and clinicians who feel routed rather than supported, who feel informed by something that is selling rather than serving, withdraw. The first generation of mediated channels will be remembered by the people who lived through them, in the same terms as the first generation of patient portals: as either the thing that worked, or the thing that proved the field had not yet earned the right to mediate.

VII. Coda: capex is not capability

The capability that defines biopharma in 2036 will not be ownership of the best model, because the models will converge, and because the model itself is an abundant complement to a scarce input. The capability will be ownership of the highest-fidelity view of human biology that any organization has assembled, and the proven ability to turn a prediction made by any reasonable model into regulatory-grade robust proof, into a patient outcome, and into a trust relationship, faster than competitors with the same model access can do the same work.

The tactical move, available to every competitor and conferring no advantage, is to buy AI in order to do today’s work faster. The strategic move, available now and not later, is to take a position in what AI will still depend on in 2036 and cannot supply for itself: deliberately captured human data, evidence generation as a core competence rather than a vendor relationship, and trust as infrastructure rather than as a campaign.

The reason to write this down in 2026, and not at the end of the decade, is that the most consequential decisions on each of these are decisions about contracts, about consent frameworks, about institutional posture, and about which functions are held close and which are released. These decisions have long lead times. The window in which they are still cheap is narrow.

My closing observation belongs to a longer argument: AI is not what finally lets biomedicine outrun its measurement infrastructure. It is the thing that finally makes the measurement infrastructure load-bearing in a way it was never designed to be before. A poor instrument used by hand produces a poor reading once. A poorly instrumented model produces a poor reading at industrial volume, every day, with rising confidence. My preconditions in the preamble are either what the next decade is built on or what many of our hopes and dreams crack against.

If we play our cards right, AI does not save the field. The field saves itself, using AI as one of several instruments, and emerges in 2036 with measurement scaffolding worth the name and a generation of treatments developed on it. If we do not, AI is the loudest decade in a long sequence of decades in which the field promised more than its instruments could verify, and the consequence of that, this time, reaches further than the consequence has reached before.

The shape of 2036 is being written in 2026, in the contracts being signed, the consents being framed, the endpoints being defended in regulatory meetings, and the willingness or unwillingness inside organizations to keep saying out loud that capex is not capability.

The writing is a slow choice and a quiet one. And I believe it is also the only one.

Further reading

Subscribe now

I. The ghost on the deck

There is a story I have told, in one form or another, for a long time. It is a story about a ship.

In 1747, aboard HMS Salisbury, a young naval surgeon named James Lind took twelve sailors whose gums had blackened and whose limbs had begun to fail them, and he divided them into pairs. To one pair he gave cider. To another, dilute sulfuric acid. To others, vinegar, seawater, a purgative paste of tamarind and cream of tartar. To the last pair he gave two oranges and a lemon a day. Within six days, the citrus pair was nearly recovered. The others were not. Lind had, almost without meaning to, invented the controlled trial. He had also, in the same gesture, conjured a ghost that would haunt medicine for the next three centuries: the Average Patient. A statistical phantom that exists nowhere except in spreadsheets, and yet around whom we built every regulatory fortress, every approval pathway, every billion-dollar trial that followed.

I told this story once, years ago, to a young contractor at the FDA at a meeting about something else entirely. His name was Jeremy Walsh. He listened the way he always does, carefully and without performance, and when I had finished, he asked the question I have come to recognize as his particular gift. He asked it again today on a stage at White Oak, MD, as the FDA’s first Chief AI Officer, in front of a room full of cameras, reporters, and pharmaceutical executives:

Wait. We are still doing it that way?!

In our meeting years back, I told him that yes, more or less, we were. That what changed between the Salisbury and the modern clinical trial ecosystem was not the basic shape of the enterprise but the scale at which we did it and the bureaucratic stratigraphy we had laid down on top of it. Sponsors run trials. Sites collect data. The data is packaged, cleaned, narrativized, and shipped across in pages of PDF, thousands of pages in a typical IND, some of it not even human-readable. Counts of counts of counts. Tables of tables of tables. The regulator sees the data only after the sponsor has finished telling itself the story of what the data means. Lind would have recognized the structure. He would not have recognized the latency. In a way, we have actually regressed.

My conversation with Jermey became a white paper. The white paper went up the chain. Nothing happened, which is what usually happens, and Jeremy went back to his other work. But he kept asking the question. He asked it for years. He asked it as a contractor and he asked it after he became Chief AI Officer at the FDA and he asked it in every meeting where someone told him the way things were done was the way they had to be done. Today he got to answer it.

II. What was announced

Today, the FDA announced the real-time clinical trial initiative involving AstraZeneca’s TRAVERSE study in mantle cell lymphoma, with sites at MD Anderson and Penn and Amgen’s STREAM-SCLC trial in small cell lung cancer. A Request for Information also opened today. A broader pilot program launches this summer. Commissioner Marty Makary, who led the announcement, framed the rationale precisely. For sixty years, he said, we have been conducting clinical trials in essentially the same way, and the lag between data and decision has not been a feature of safety. It has been sediment. The accumulated weight of a system that was built when data lived in filing cabinets and submissions arrived by truck. The agency, he said, can reduce that lag without cutting any corners on safety, which remains the FDA’s number one priority. He said the line plainly, and he said it more than once.

This is the part of the announcement that most deserves attention, because it is the part that is easiest to misread. What the FDA is doing is not deregulation. It is a re-engineering. The protocols are still protocols. The investigators still consent patients. The reviewers still review. Paul Burton, Chief Medical Officer at Amgen, was deliberate on this point. Pragmatic real-time clinical trials, he said, do not mean clinical research with less rigor. The fundamentals of good science do not change. Protocol development, protocol oversight, informed consent, monitoring, strong governance, FDA review. All of it remains absolutely standard and absolutely essential. Patient safety and data integrity, he said, will always be the north star.

That language matters because the architecture being introduced is one in which safety is not preserved by latency but expressed through signal. Today’s announcement is, in this sense, a strengthening of efficacy-signal detection and pharmacovigilance, not a relaxation of them. Makary made the point directly in the press Q&A. Safety signals and clinical endpoints are agreed upon by the sponsor and the FDA in advance, as they have always been. What changes is that those signals can now be transmitted to the agency in real time rather than packaged into a submission months or years later. During the pilot, sponsors will continue to submit data through the conventional pathway in parallel, so that the agency can compare the two channels and refine the model. Nothing has been removed from the safety apparatus. Something has been added.

What changes, then, is the latency. What changes is the assumption, baked so deeply into the architecture of clinical development that we forgot it was an assumption, that the regulator should see the data only after the sponsor has packaged it. Jeremy described the new model on stage in language that I suspect will sound radical to people outside the field and obvious to people inside it. We do not need everything we have ever seen. We need signals. We need to define, with the sponsor and the FDA together, what would actually change a regulatory decision, and we need to see those things happen as they happen. Not a year later. Not at the next data lock. Now.

III. Listening to the patient instead of to the form

The first time I worked with Jeremy was at the National Cancer Institute, where we designed a clinical trial in adult patients with cancer using voice recognition and facial recognition to estimate pain and mood. The patient sat down. The cameras and microphones did the rest. No survey. No clinician translating an expression into a number on a Likert scale. The signal came directly from the human being in the room, in real time, captured continuously, with a fidelity that no questionnaire has ever achieved or could. That dataset, to my knowledge, remains the largest of its kind.

The deeper observation that emerged from that work, and that I think is the actual intellectual core of what was announced today, is that the data we collect in clinical trials is not always a faithful image of the patient. It is a heavily compressed, heavily edited, heavily delayed reconstruction. The patient experiences the disease in continuous time. We sample it in discrete time. The patient lives in signal. We submit in summary. Every step in the pipeline from bedside to FDA reviewer is a lossy compression, and the loss is not random. It systematically privileges what is easy to measure and structure over what is meaningful and continuous. The Likert scale erases the wince. The case report form erases the silence before the patient answers. The quarterly data lock erases the morning the toxicity actually appeared.

What changes when signals flow in real time is not just the speed of regulatory decisions, although that matters and it matters enormously to the patient who is in hospice waiting for the drug. Jennifer Litton, the Chief Clinical Research Officer at MD Anderson and one of the academic leaders behind the AstraZeneca trial, made the point with the kind of moral clarity that only someone who has stood at the bedside can make. She is an oncologist. She is the daughter of two parents with cancer. She was diagnosed with cancer this year, and finished surgery and radiation. When you do not know whether the next patient who needs a clinical trial is going to be you or someone you love, she said, you want that trial available right now. The latency in our current system is not abstract. It is measured in funerals.

What changes, then, is also the resolution of what the regulator can see. We move from low-frame-rate snapshots of an aggregate to a continuous record of individuals. The Average Patient, the ghost we summoned on the deck of the Salisbury because we had no other way to reason about populations at scale, becomes optional. Not abolished. Optional. We can still aggregate when aggregation is what the question requires. But we are no longer forced to aggregate as a precondition of being able to see anything at all.

IV. The renovation

There is a temptation, when an announcement of this kind lands, to declare that the old order has fallen and a new one has taken its place. That is not what has happened today. Two trials. One technology partner (Paradigm Health). This is the first sentence of a new chapter, not the chapter itself. The hard work, the transformation, is the scale-up. It is the second sponsor, and the tenth, and the hundredth. It is the standards work that has to happen quietly behind the announcements, the EHR-to-EDC plumbing, the sponsor data systems that have to be re-architected, the reviewer workflows that have to be redesigned around streams instead of submissions. It is, as one industry veteran told me afterward, the long stretch where we will have to run both systems in parallel, the old way and the new way at once, before the old way can finally be retired.

There is also the human dimension, which is easy to miss in the technical conversation. Emma Meagher, who runs clinical and translational research at Penn and is one of the principal investigators on the AstraZeneca trial, made an interesting observation. She has been a trialist on the front lines for thirty-two years. She knows, better than almost anyone, how punishing the current process is on the people who actually run it. The research coordinators. The trial nurses. The project managers. They are virtually impossible to find and harder to keep. What this innovation does, she said, is not just shorten the FDA’s timelines. It eases the academic interface of patient care. It gives the workforce that sustains clinical research a reason to stay. The renovation is not only of the regulatory architecture. It is of the human one.

What Makary said almost in passing was, I think, the line that we should remember. He told the room, with a self-deprecating shrug, that he was not going to show a PowerPoint, because PowerPoint had traumatized him over the course of his career. He showed one graphic. The drug development and approval process, beginning to end. Each phase a box. Each box, he said, currently being re-engineered. The graphic is the same graphic we have been showing for sixty years. The boxes are the same boxes. What changes now is what flows between them, and how fast, and at what fidelity, and to whom.

This is not a revolution. Revolutions tear down. This is something rarer and harder. This is a renovation of a working building, conducted while the building remains occupied by patients who are, right now, today, waiting for the next box to open. The ghost of the Salisbury, the Average Patient who never existed, can finally begin to be retired. Not because we have stopped caring about populations, but because we have learned, at long last, how to listen to individuals in real time, and to aggregate that listening without losing it.

To Jeremy, congratulations. He kept asking the question for years after most people would have stopped. To the team at FDA, the sponsors, the academic centers, the technology partners who made this real, congratulations. To everyone watching from outside the agency wondering whether this is a small thing or a large thing, I will say this. It is an important thing. And in ten years it can hopefully look like the thing that broke the dam.

The signal is moving. The ship has finally left the harbor. Now we have to learn to navigate.

Subscribe now

Every oncology drug approved based on tumor-based endpoints in the last twenty-five years rests on an instrument. The instrument is not a machine, not a molecule, not a laboratory assay. It is a rather arbitrary convention. A rule about how to press a ruler against a shadow on a screen and decide whether the shadow has grown. The convention is called RECIST (Response Evaluation Criteria in Solid Tumors) and it dictates, with the stubborn authority of orthodoxy, how major regulatory decisions in solid tumor oncology have been made since the late 1990s.

RECIST works like this: A radiologist examines a CT scan and selects up to five “target lesions,” the largest measurable tumors in a given organ. She measures each one along its longest diameter, sums those diameters, and compares the sum to a baseline. If the sum has shrunk by 30% or more, she writes partial response. If it has grown by 20% or more, she writes progressive disease. Anything in between is stable disease. Complete disappearance is a complete response. These four words, expanded into the binary of ORR (objective response rate: the fraction of patients who achieve complete or partial response), are what we submit to the FDA as evidence of efficacy. They are the basis on which drugs are approved, priced, and prescribed to patients with cancer.

The framework has the unusual property of being both ubiquitous and invisible. Few oncologists have heard of it. Almost no one in clinical practice actually uses it. It is the instrument by which drugs are certified, not the instrument by which they are administered. This is the first fact about RECIST worth holding onto. We will return to it.

I. Two readers, one scan

In 2017, while at the FDA, my colleagues and I pooled the paired radiology assessments from a decade of solid tumor registration trials. The structure of these trials has a quirk useful to our purpose: for regulatory submission, every scan is read independently by two qualified radiologists, one primary investigator review and the other a blinded radiologist at a central reading facility, each blinded to the other’s work. If they disagree, an adjudicator breaks the tie. The raw data, kept for internal auditing, records how often two expert readers, looking at identical images and applying identical criteria, reach identical conclusions.

Our pooled sample contained 13,677 paired assessments. The raw concordance rate was 69.5 percent. Two of every three paired reads agreed. The remaining third did not. In hard-to-measure malignancies such as pancreatic and ovarian cancers, the discordance exceeded 40% and, in some cases, reached the predictability of a coin flip.

This figure is higher than it ought to be. In the language of measurement theory, it corresponds to a weighted kappa somewhere between 0.30 and 0.55. Kappa is a statistic that asks how much two raters agree beyond what you would expect from chance alone; a value of 1.0 is perfect agreement, 0.0 is no better than coin flips. The number we observed sits at the lower boundary of what Landis and Koch, in their widely cited 1977 classification, called “moderate.” Moderate is what you would expect from a diagnostic tool in early development. It is not what you would expect from the instrument on which the entire oncology drug development enterprise rests.

The standard defense of the framework, when this figure is raised, is procedural. Discordance, the argument goes, is non-differential. It is distributed symmetrically across treatment and control arms. Therefore it does not invalidate the comparison. The p-value remains honest. This is partly true for randomized studies. It is also profoundly incomplete.

II. The attenuation

Here is what the procedural defense leaves out, beyond the fact that many cancer drugs are advanced in single-arm studies. When you inject random error into a comparison of two proportions, the observed difference between the proportions does not stay the same. It shrinks. This is a known result in epidemiology, tracing at least to the work of Irwin Bross in the 1950s, but it is rarely invoked in discussions of clinical trial design because most people prefer to believe that their measurements are not noisy.

The mathematics is merciful in its clarity. If the true response rate in the treatment arm is P(T) and the true response rate in the control arm is P(C), and each patient is misclassified with probability m, the observed difference between arms is the true difference multiplied by the factor (1 - 2m). At our measured discordance rate of thirty percent, that factor is 0.40.

A drug whose true advantage over standard of care is twenty percentage points of ORR, a substantial effect by any reasonable standard, appears in the trial record as a drug with an eight-point advantage. The biological signal has been attenuated by sixty percent before the statistician ever sees it.

Attenuation increases the sample size. Because the number of patients required to detect a given effect scales as the inverse square of that effect, a trial that would need two hundred patients to detect the true difference would need, under this measurement regime, approximately one thousand two hundred and fifty to detect the attenuated one. Most trials do not enroll 1,250 patients. Most trials are powered to detect the signal they were designed to detect, assuming the measurement faithfully transmits it.

The directional consequence is the one that matters, and it is rarely stated plainly. Random measurement error biases observed effects toward the null hypothesis. At the level of the measurement substrate, this produces a systematic excess of false negatives over false positives. A trial that fails to reach statistical significance is not necessarily a trial of an inactive drug. Some meaningful fraction of negative trials in solid tumor oncology are trials of active drugs evaluated with an instrument that could not see them.

No one knows the size of that fraction. The instrument, by its nature, erases the evidence of what it failed to detect.

III. Four boxes for a continuous world

The second property of RECIST is a property of encoding rather than of measurement. The underlying biological variable, the percent change in tumor burden over time, is continuous. It takes any value on the real line, bounded below by minus one hundred percent and above by whatever the tumor can manage. The framework transforms this continuous variable into an ordinal category with four levels, and then, for ORR, collapses those four levels into a binary. Something biologically rich becomes something statistically thin.

There is a way to quantify the loss, borrowed from information theory. A continuous percent-change measurement carries, under plausible distributional assumptions, between 5 and 8 bits of information per patient per scan. (A bit is the unit Claude Shannon introduced in 1948 to measure the content of a signal: one bit is the information gained from the outcome of a fair coin flip.) A four-level ordinal carries two bits, which is the base-2 logarithm of 4. A binary outcome carries one. The RECIST encoding, at its most generous, transmits at most a quarter of the information contained in the measurement it encodes. At ORR, it transmits less.

The effect is not merely a loss of resolution but a loss of biological distinguishability. An agent that produces a uniform 25% shrinkage in every treated patient, a remarkable result implying a real mechanism of action operating reliably across a population, can yield an ORR of zero. It is categorized, on the endpoint, as indistinguishable from placebo/no treatment. It is also indistinguishable, at the endpoint, from an agent that produces 19% growth. Three biologically distinct populations, trivially separable in the underlying distribution of percent change, collapse to the same point in the reported outcome.

This is the structural consequence of mapping a continuous variable through a step function. The step does what steps do. It discards what lies between its treads.

IV. Stochastic at the edge

The thresholds themselves introduce a third pathology, one that becomes vivid if you imagine the patient whose true response is not in the middle of a category but at its boundary.

For example, let’s look at a patient whose tumor burden has decreased by 28%. Two percentage points below the 30% partial response threshold. Stable disease by the rules, but barely. Suppose the inter-reader standard deviation on the sum of diameters, consistent with the discordance data, is about ten percentage points. The probability that a given reader assigns this patient to the PR category, under a normal approximation, is the probability that the reader’s estimate falls below -30%, which works out to about 42%.

Forty-two percent. The same biology, the same scan, the same criteria, yields partial response in forty-two of one hundred reads and stable disease in fifty-eight. The categorical outcome is, at the boundary, approximately a coin flip. And the reported trial effect depends on how many of these coin flips land on the side that counts.

This threshold instability does more than add variance. It creates a selection pressure that propagates backward into the development pipeline. Because ORR binarizes sharply at minus thirty percent, a sponsor who wishes to maximize the probability of clearing that threshold is rewarded for advancing agents capable of producing deep responses in a subset of patients. Agents that produce shallow but consistent cytoreduction across a population, even when the underlying biology is meaningful, are structurally disfavored. They cannot clear the threshold often enough to generate an ORR that the machinery can see. And our machinery is impatient. When meaningful tumor responses are not seen in early clinical trials, the entire program is at risk.

The shape of the modern oncology pipeline, its preferential enrichment for drugs that drive responses visible to the human eye, is a direct consequence of our endpoints. We have been selecting not for what can prolong survival or increase quality of life, but for what our instruments can detect.

V. The instrument that does not exist outside the trial

The third property is the most easily stated and the most commonly overlooked. RECIST is a rather artificial clinical trial instrument. It is not a clinical tool. Community oncologists, the people who actually treat the overwhelming majority of patients with cancer, do not compute sums of diameters. They do not segment target lesions into protocol-defined categories. They read narrative radiology reports written in ordinary English (”stable appearance of the right upper lobe mass”) and make decisions on the basis of clinical judgment integrated with that narrative.

Which is to say: the endpoint that certifies a drug’s efficacy in the trial setting is not the endpoint that governs its use in the practice setting. The variable on which approval rests is not present in the data that routine care generates. This is the oncology equivalent of approving, say, statins on the basis of LDL reduction only to find out that community physicians use blood pressure to administer and titrate therapy.

When we attempt to validate oncology clinical trial findings against real-world evidence, using electronic health records and community radiology reports, we cannot reconstruct a RECIST assessment. The comparison has to be made at a different level of abstraction, using different variables, and the correspondence between the two settings, at the level of the individual patient or the individual drug, remains structurally unverifiable.

The conventional response to this observation is that RECIST is a surrogate for clinical benefit rather than the benefit itself. A surrogate is a measurable variable used as a proxy for the outcome of real interest, which in oncology is almost always overall survival. The real question, on this view, is whether RECIST-based endpoints predict overall survival. The meta-analytic answer, accumulated across hundreds of trials and now reasonably mature, is that they do so imperfectly. PFS-to-OS correlations in solid tumor oncology routinely yield coefficients of determination below 0.5. ORR-to-OS correlations are lower. The surrogate explains less than half of the variance in the outcome it is supposed to predict.

The framework persists, then, less because of its predictive validity than because of its institutional entrenchment. It is the thing we all agreed to use. Agreement has its own gravitational pull, and for a long time the conditions that would permit disagreement did not exist.

VI. What the endpoint could not see

An instrument with 30% reader discordance, indicating moderate agreement by standard statistical classification, can attenuate observed effects by 60% and inflate required sample sizes sixfold. The attenuated signal is then projected onto an ordinal encoding that transmits at most two bits per scan and produces stochastic categorization at the thresholds where the reported effect is determined. The resulting endpoint is computable only within the trial setting and cannot be reconstructed from the data generated by clinical practice after approval.

The directional bias of the whole apparatus is toward false negatives. It favors drugs whose mechanism produces a small number of responses that are deep enough to be detected by the human eye over drugs whose mechanism produces a large number of modest responses, but can prolong survival or improve quality of life. It cannot see agents whose signals fall below the threshold for categorical instability, and it cannot distinguish, at the endpoint, between agents that differ by biologically meaningful amounts.

The statistical machinery layered on top of this substrate, all the careful p-value thresholds and hierarchical testing procedures and alpha-spending rules, is bounded by what the substrate permits. No refinement of the hypothesis test can recover information that the measurement did not transmit.

The cost is not just borne by the sponsors whose trials failed. It is borne by the patients whose therapies were never developed because the endpoint could not see them.

VII. The substrate is replaceable now

The conditions that produced RECIST in the late 1990s are no longer operative. Volumetric segmentation of tumors, computationally prohibitive a few years ago, is now a solved problem. Foundation models for medical imaging, trained on orders of magnitude more data than any individual radiologist will read in a career, are mature. Automated assessment at scale, across both trial and community settings, is a deployed capability.

Over the past several years, I have been working on a replacement substrate at Project Data Sphere. We call it the Total Tumor Burden Index. It is volumetric rather than diametric, which reduces reader-dependent variance. It is continuous rather than categorical, which eliminates both the information loss of ordinal encoding and the instability at categorical boundaries. It is machine-generated rather than reader-dependent, which removes the dominant contemporary source of measurement noise and allows deterministic reproducibility across sites and across re-reads. And it is equally deployable in trial and community settings, which means the trial measurement and the practice measurement can, for the first time, be the same variable.

The regulatory path will take time. Prospective comparison with RECIST in contemporary datasets will take time. Integration across imaging infrastructure will take time. None of these barriers is technical. The technical conditions for the replacement are being satisfied with increasing precision.

VIII. Coda

The measurement substrate in solid-tumor drug development is not neutral. It has been shaping the pipeline for twenty-five years in ways the field has not made fully explicit to itself. The attention has remained on the superstructure, the p-values, the statistical analysis plans, and the regulatory pathways, when the leverage is actually in the layer underneath. Replacing the measurement, rather than refining the statistics computed upon it, is the tractable path forward. The conditions for that replacement are, at last, in place.

What we measure determines what we can find. What we can find determines what we develop. What we develop determines who is treated. The ruler we have been pressing against the tumor for a generation is not the only ruler available. It is simply the one we inherited. And that inheritance is due for review.

Share

There is a principle in pharmacology that took medicine decades to fully absorb. Liposomal doxorubicin and free doxorubicin are the same molecule. Yet one accumulates preferentially in tumors while the other damages the heart. The liposome does not make the drug more potent. It changes where the drug goes, what it encounters, what it spares; by exploiting the leaky vasculature of the tumor microenvironment and staying encapsulated in the tighter junctions of cardiac tissue. The container shapes what the active compound can do.

The same principle, it turns out, applies to AI systems. And a paper published at the end of March 2026 by researchers at Stanford, MIT, and KRAFTON gives us the clearest demonstration yet of what that means in practice and what it makes possible.

What the paper found

The paper is called Meta-Harness: End-to-End Optimization of Model Harnesses. Its subject is the code that wraps a large language model, the layer that determines what information the model sees, in what order, retrieved from where, and framed how. The authors call this the harness. It encompasses three things: memory, meaning what the system retains between interactions; retrieval, meaning what it fetches from external sources at each decision point; and prompting, meaning how context is assembled and sequenced before the model reasons about it.

The central observation is straightforward. Research cited by the authors has shown that changing only the harness around a frozen, untouched model can shift performance by a factor of six on the same benchmark. Same weights. Different wrapper. Until now, harnesses have been built entirely by hand: engineers designing prompting frameworks, retrieval strategies, and memory structures through iteration and intuition, exploring a small fraction of what is possible.

Meta-Harness automates this process. It works through a recursive loop. A candidate harness is proposed, evaluated against a set of tasks, and everything is saved in raw, uncompressed form, not a score, not a summary, but the full transcript of every prompt the model received, every response it gave, every tool call it made. A coding agent then navigates that filesystem, reads what it needs, and proposes an improved harness. The loop repeats roughly sixty times in a full run.

What distinguishes this from every prior approach to automated optimization is the refusal to compress. Systems like OPRO, TextGrad, and AlphaEvolve hand the optimizer a digest: a scalar score, a short critique, a recent snippet. Meta-Harness gives the agent the raw logs and trusts it to navigate selectively, reading a median of 82 files per iteration across the full history of prior attempts. That access to raw diagnostic history, rather than compressed summaries, turns out to be the key design decision.

The results across three domains are real and substantial. On a text classification task where a model updates its memory from labeled examples and is evaluated on held-out cases, the discovered harness outperforms the best hand-designed system by 7.7 accuracy points while using four times fewer tokens. On 200 olympiad-level mathematics problems, a single harness discovered during search improves accuracy by an average of 4.7 points across five AI models never seen during the search process, indicating the discovered strategy is genuinely general. On TerminalBench-2, a benchmark of 89 real-world tasks requiring AI agents to autonomously compile code, debug systems, and complete multi-step terminal workflows with full automated verification, the discovered harness places first among all agents running on the same underlying model, surpassing every hand-engineered competitor.

But what the paper demonstrates is not just that automated harness search works. It is that the agent doing the searching behaves in ways that matter for how we think about these systems going forward.

The agent that reads its own case files

The paper’s appendix traces, iteration by iteration, how the agent reasoned through a live search run on TerminalBench-2, starting from a strong hand-engineered baseline. What emerges is not a description of brute-force search. It is a description of systematic diagnosis.

In the first two iterations, the agent bundled a structural bugfix with a rewritten prompt template. Both regressed from the baseline. In the third iteration, rather than proposing a third fix, the agent looked back across both failures. It noticed that while the two attempts differed in their structural changes, they shared one intervention: the prompt rewrite. It separated the confounded variables and wrote verbatim into its log: “the structural bugfixes were confounded with harmful prompt changes.”

This is abductive reasoning, the same logic a careful physician applies when two patients on similar regimens both deteriorate. You do not conclude the drug failed. You ask what they had in common. You isolate. You test. The agent was performing differential diagnosis on its own previous attempts, using the filesystem as its case history and the execution traces as its clinical notes. It could only do this because it had access to the raw logs. A compressed summary would have hidden the shared intervention entirely.

By the seventh iteration, after six consecutive regressions that taught the agent where the system’s fragility lived, it stopped modifying the existing machinery and instead added something before the loop began: a snapshot of the computational environment injected into the initial prompt. Its recorded reasoning: avoid touching the previously fragile components. It had learned not just what failed but why and used that understanding to reason about risk.

This behavior (reading raw history, identifying confounds, forming causal hypotheses, reasoning about where fragility lives) is what full access to prior execution traces enables. Compressed feedback cannot support it. And it is precisely this kind of reasoning that the hand-engineering of clinical AI harnesses has always depended on human experts to provide.

The compression problem, demonstrated

To understand why this matters for biomedicine specifically, it helps to look at what happens when LLM harnesses are hand-engineered in a clinical setting.

In a study we conducted last year at Project Data Sphere with Thomas Jefferson University, four independent investigators used LLMs (ChatGPT-4 and Gemini Advanced) to analyze data from a clinical trial evaluating LY2510924 combined with carboplatin and etoposide for extensive-stage small cell lung cancer. All four investigators worked from the same dataset, the same chain-of-thought prompting framework, and the same analytical objectives.

The results were interesting but inconsistent. One investigator found a statistically significant improvement in progression-free survival with the addition of LY2510924, contradicting the original publication and the other three investigators. Another reported an objective response rate that deviated substantially from every other analysis. Cohort sizes ranged from 89 to 180 participants from the same dataset, suggesting the LLMs characterized entirely different patient populations depending on how the prompts were constructed. The conclusion of the paper was clear: variability in prompting strategies drove variability in clinical conclusions, and highly trained subject matter experts remain essential to review and interpret LLM-generated analyses.

This is the harness problem in clinical research, demonstrated under controlled conditions. The chain-of-thought framework was the harness. It was hand-engineered. Different investigators interpreted it differently. The model was sensitive to subtle prompt variations. And the outputs, including whether a drug appeared to improve survival, changed accordingly.

Meta-Harness is a direct response to exactly this problem. Rather than asking human experts to design the optimal prompting framework and then hoping investigators apply it consistently, automated harness search would systematically explore the space of prompting and retrieval strategies, read its own failures at the case level, this analysis got PFS wrong, what was it shown, what did it retrieve, what framing led it astray, and discover configurations that produce consistent, reproducible outputs across investigators, model versions, and datasets. The subject matter expert moves from designing the wrapper to evaluating and validating what the automated search discovers. That is a meaningful and practical shift in how we deploy LLM-based clinical tools.

What the 40% tells us

A drug achieves a 40% ORR in a clinical trial. We call this a success. The drug is approved, deployed, and prescribed. But 60% of patients who receive it do not respond. They experience toxicity without benefit, akin to taking a toxic placebo.

Those 60% non-responders are not a rounding error. They are a population. And somewhere in the data we already have; genomic profiles, imaging series, early response signals, treatment histories, patterns of toxicity; there may be information that distinguishes them from the 40% who responded. The challenge is not always that this signal does not exist. The challenge is often that we do not know where in the vast space of available data to look for it, and that the LLM-based clinical tools we are deploying are wrapped in harnesses that were never designed to find it.

This is where automated harness search becomes directly relevant to patient care. An LLM-based decision support tool wrapping a clinical dataset will only surface the signal that its retrieval and context construction strategy is designed to find. If that strategy is hand-engineered, it reflects the intuitions of the engineers who built it. Automated harness search would systematically explore alternative retrieval policies: what patient-level data to pull, how to frame prior cases, what early signals to foreground; reading its own failures at the case level and discovering configurations that surface signal the hand-engineered wrapper missed.

Where this applies today

Countless LLM-based tools are being deployed across biomedicine right now, each with a hand-engineered harness, each leaving performance on the table in ways that automated search could address. A few meta-harness insipred opportunities to highlight:

Clinical trial analysis. Our study demonstrated this problem directly. LLM-based systems for analyzing trial data are sensitive to prompting strategies in ways that produce clinically meaningful variation in outputs. Automated harness search could discover prompting and retrieval configurations that produce consistent results across investigators and model versions, reducing the burden on subject matter experts to catch errors and standardizing the analytical process across institutions.

Clinical decision support. LLM-based systems are increasingly used to synthesize patient records, surface relevant literature, and support diagnostic and treatment decisions. The harness determines what gets retrieved from the electronic health record, how prior similar cases are identified and presented, how uncertainty is communicated to the clinician. These are wrapper decisions. They are currently made by hand, reflecting the intuitions of the development team. Automated search over these decisions, using case-level outcomes as the feedback signal, could discover retrieval and framing strategies that improve decision quality in ways that no individual engineer or clinician would have designed.

Drug development workflows. LLM-based tools are being used to synthesize preclinical literature, generate hypotheses about mechanism, and support the design of clinical development programs. The harness determines how prior experimental results are retrieved and contextualized, how mechanistic hypotheses are framed, what evidence is foregrounded when the model reasons about a new compound. Automated harness search could discover context construction strategies that improve the quality and consistency of these outputs, not by changing the model, but by changing what it sees.

Biomarker and multi-omics analysis. When LLM-based systems reason over complex biological data, integrating genomic profiles, imaging features, proteomic signatures, and clinical records, the harness determines what gets retrieved, how it is integrated, what gets foregrounded for each patient. The design space here is enormous. Automated search over retrieval and integration strategies, reading its own failures at the case level, could discover configurations that surface signal in high-dimensional biological data that hand-engineering would not find.

Regulatory and evidence synthesis. LLM-based tools are being used to synthesize clinical evidence for regulatory submissions, systematic reviews, and health technology assessments. The harness determines how evidence is retrieved, how studies are compared and contextualized, how uncertainty and heterogeneity are communicated. Inconsistencies in these outputs, analogous to those demonstrated in our study, have direct consequences for regulatory decisions. Automated harness optimization offers a path toward more consistent, reproducible evidence synthesis.

What changes for experts

None of this necessarily replaces clinical or scientific expertise. It relocates where that expertise is most valuable.

Today, expert knowledge goes into designing the harness, crafting the prompting framework, specifying the retrieval strategy, determining what context the model sees. This is painstaking, iterative, and bounded by individual intuition.

With automated harness search, expert knowledge moves to a different layer. Experts define the objective: what does a good clinical trial analysis look like, what does a useful decision support output look like, what constitutes a failure worth correcting. They evaluate and validate what the automated search discovers. They set the boundaries of what the system is permitted to do. They interpret the discovered harness and decide whether it reflects genuine clinical reasoning or a brittle artifact of the search process.

This, I believe, is a more appropriate deployment of expertise in the age of AI. The combinatorial space of possible harness configurations is too large for human intuition to explore systematically. Automated search is better at that. The evaluation of whether a discovered configuration is clinically sound, generalizable, and safe is a judgment that requires deep domain knowledge. Experts are better at that.

The compression principle

Running through all of this is a single epistemological claim that the Meta-Harness paper demonstrates more cleanly than any prior work: compressed feedback is lossy feedback, and lossy feedback leads to lossy learning.

In the paper’s controlled ablation, the system with access to raw execution traces achieves 50% median accuracy on a text classification task. The identical system with access only to scores and LLM-generated summaries achieves ~35%. Fifteen points separate a system reasoning from raw evidence and a system reasoning from digests of that evidence. That gap is the measured cost of compression, of replacing the signal with a summary and then trying to reason from the summary as if nothing had been lost.

In medicine, we compress constantly. Individual patient trajectories become survival curves. Heterogeneous treatment responses become pooled response rates. The richness of what happened to each person in a trial becomes a table entry and, eventually, a label (responder or non-responder) that discards the nuances of the outcome and everything that led to it. And here precision matters: our prized p-value is a downstream artifact, a summary of a summary. The compression that happens earlier, in the decision to aggregate heterogeneous individual trajectories into a single outcome distribution, can make the p-value meaningless. Once that aggregation is made, the p-value reports on what remains after the flattening.

We compress because reasoning over raw trajectories at scale has been historically intractable. That constraint is loosening today. What Meta-Harness demonstrates is that a system given access to raw case-level evidence can discover things that a system given only summaries cannot and that the difference is not marginal. In clinical AI, where the heterogeneity is biological and the stakes are patient outcomes, building systems that reason from richer, less compressed evidence is not just a technical improvement. It is a different epistemological commitment, one that we are now, for the first time, practically positioned to make.

The recursion

Meta-Harness is itself a harness. The coding agent that proposes new harnesses is wrapped in a skill document defining its role, its constraints, its objective. That document is, in principle, subject to optimization too. What happens when the outer loop optimizes not just the inner harness but the proposer’s own search procedure? The loop accelerates. The engineer becomes, progressively, a curator.

This is not a distant prospect. It is the obvious next experiment given the capabilities that exist today. And it implies a trajectory in which the performance ceiling of LLM-based systems is no longer bounded by the bandwidth of the engineers designing their wrappers but by the quality of the feedback signal and the clarity of the objective.

That means the most important design decisions going forward are not which model to use or how large to make it. They are: what does success look like at the case level, how do we measure it, and how do we build the feedback infrastructure that lets automated search find the wrapper configurations that achieve it. In biomedicine, those are questions that require clinical expertise, regulatory judgment, and scientific rigor. They are also questions that, once answered well, unlock a new generation of LLM-based tools that are more consistent, more generalizable, and more useful than anything hand-engineering alone can produce.

Like liposomal doxorubicin, the harness changes what the active compound can do. Unlike the drug, it can redesign itself. And unlike the chemist, it never stops iterating.

Subscribe now

Share

Yesterday, the US National Cancer Institute (NCI) convened a workshop on foundation models in cancer biology. What struck me most was not the technical ambition on display but the quality of the skepticism, a virtual room full of people who have stopped debating whether these models belong in oncology and started asking something considerably harder: whether the frameworks we are using to build and evaluate them are capable of representing the biology we need them to predict. That question is one I have been working through for some time, and my talk was an attempt to lay out the argument as precisely as I could.

I. The Molecule that knew everything except what mattered

I opened with RG7112, an MDM2 inhibitor evaluated in a Phase I trial for acute myeloid leukemia. In silico it was immaculate: low nanomolar affinity for its target, greater than 1000-fold selectivity over off-targets, favorable ADMET profile, Lipinski’s Rule of Five cleared without objection. Every computational checkpoint had been passed. The molecule had been assessed using structure-based methods drawing on thousands of crystallographic structures and the full weight of what computational drug discovery could bring to bear.

In patients, it produced dose-limiting thrombocytopenia and neutropenia, along with systemic homeostatic disruption that none of the models had anticipated. The compound had raised p53 levels exactly as designed. What no simulation had represented was MDM2’s role in hematopoietic homeostasis, its function in maintaining the very cellular populations whose depletion closed the therapeutic window. The biology responded across tissues and timescales the model had no architecture to reach. The simulation was not wrong. It was operating in a different biological register than the one that determined the outcome.

I used RG7112 not because it failed spectacularly but because it failed instructively. Since 1950, the cost of bringing a drug to market has doubled roughly every nine years, now exceeding $2.6 billion per approval, while the enabling technologies have advanced exponentially across the same period. Drug output per research dollar has collapsed anyway. We are applying increasingly powerful instruments to a problem whose fundamental difficulty has not yielded to them -- and that asymmetry demands an explanation beyond the usual invocations of biological complexity.

II. When the abstraction becomes the obstacle

The explanation I proposed at the workshop is architectural. Most modeling approaches in drug discovery are built on a reductionist biological view: Protein A influences Protein B, which influences Protein C. The abstraction is not false, it has produced real successes, most durably when the biology happens to be genuinely linear and the target genuinely necessary and sufficient, as with BCR-ABL in chronic myeloid leukemia or PD-1/PD-L1 signaling in tumors with pre-existing immune infiltration. But the danger is in treating the abstraction as a complete description of living systems, because biological systems are dense, adaptive networks of thousands of interacting signals, most unmeasured and many unknown. When models are built on simplifying assumptions that do not hold in vivo, their predictive validity remains limited regardless of how sophisticated the computation becomes.

This was the structural failure of the first wave of AI in drug discovery, which ran roughly from the 1980s through the 2010s. QSAR models, molecular docking, virtual screening, tools that operated within closed experimental loops, trained on proxy variables like binding affinity and validated against similar proxies. The correlations they captured were real within constrained systems. What they accumulated was not causal understanding of adaptive biological behavior but increasingly confident predictions about an abstraction that diverged from reality at the exact boundary where clinical outcomes are determined.

The current generation has a different vocabulary (generative design, foundation models, phenomics, etc) but a version of the same core limitation persists. AI today is genuinely effective at optimization: navigating the tradeoff space between molecular potency, selectivity, metabolic stability, and synthetic accessibility with a precision that accelerates medicinal chemistry substantially. But optimization and discovery are different problem classes.

Optimization assumes the underlying biological hypothesis is correct and improves performance within that framework. Discovery requires determining whether the hypothesis is valid in the first place. No amount of optimization capability resolves uncertainty in discovery, and a model that efficiently explores chemical space around a target that turns out to be dispensable in the context of a living organism is threading the needle through the wrong fabric.

The phenomics approach makes this concrete. High-content phenotypic assays measure morphological or transcriptional changes in cell-based systems and use them as proxies for therapeutic efficacy. The technical achievement is real. But cell lines on plastic lack immune infiltration, metabolic gradients, and tissue architecture, the features that determine whether a perturbation produces clinical benefit in a patient rather than a phenotypic readout in a well. A model may predict phenotypic rescue accurately and clinical outcome not at all, because the proxy and the target are operating in different biological registers.

The data I presented from a study we published in 2020 makes the same point at the level of the entire translational pipeline. Across 108 oncology drugs evaluated in mouse, rat, dog, and monkey models, we found median positive predictive values of 0.65 and negative predictive values of 0.50 for predicting human toxicity. A clean animal signal provides essentially coin-flip assurance of human safety.

Our challenge is not confined to computational modeling; the entire pipeline from in vitro systems through animal models does not reliably approximate human biology, and any foundation model trained on data generated by that pipeline inherits its limitations as structural features of the training distribution.

The mathematical frame for all of this is interpolation versus extrapolation. Machine learning systems are reliable when test conditions resemble training conditions closely enough that the learned mapping holds. Chemistry and physics generally meet this condition; governing principles are stable and well-characterized. Biology does not. Emergent behaviors, adaptive resistance, immune tolerance, and compensatory signaling across tissue types are the phenomena that determine clinical outcomes, and they exist outside the training distribution of any model built primarily on experimental proxies. This is why technical validation and clinical validation remain separated by five to seven years; which generative advances have not compressed. AlphaFold predicts protein structure with near-experimental accuracy. RFdiffusion designs proteins with no natural template. These are genuine inflection points in simulation fidelity at the molecular level. A protein that folds as predicted still has to work in a patient, and that distance has not shortened.

III. The language we use is part of the problem

Running beneath all of this is a representational problem I addressed in a Clinical and Translational Science paper, and that the workshop made more urgent. The classification systems that organize cancer biology were not designed to predict therapeutic response. They were built to describe what a pathologist sees under a microscope, and they achieved institutional authority through historical convenience rather than mechanistic precision. As I often say as an illustrative example, “non-small cell lung cancer” defines a disease by negation; it tells you where the tumor was found and what it failed to resemble histologically. It groups malignancies with distinct molecular drivers, immune microenvironments, and resistance mechanisms under a linguistic label that encodes none of that information.

When these inherited categories become training labels for foundation models, the models do not simply inherit their limitations; they amplify them. More parameters and more computing power make the model more confident in the wrong abstraction. The clinical consequence is measurable: overall response rates with doublet chemotherapy in morphologically defined non-small-cell lung cancer hover around 30%, whereas in molecularly defined, tissue-agnostic contexts they routinely reach 50-90%. The classification system produces the wrong answer the majority of the time because it was never designed for the question being asked of it.

What a mechanistically grounded alternative looks like in practice is what I discussed through TTBI, the Total Tumor Burden Index we are developing at Project Data Sphere. It is autonomous, volumetric, longitudinal, and independent of inherited tumor classification categories. It aims to create a biomarker that reflects what the disease is actually doing rather than relying on a label derived from human sensory input.

Drew Williamson presented PathChat at the workshop, a multimodal language model built on whole-slide images and pathology case reports, and the evaluation findings aligned closely with my argument. The NLP metrics the field has inherited for assessing language models, METEOR, ROUGE, BLEU, measure textual similarity to reference answers. In pathology, that is not the same as diagnostic utility. A model can reproduce the vocabulary and syntax of a pathology report while conveying nothing actionable for a clinician, and standard benchmarks lack a mechanism to detect this difference. When his group convened seven experienced pathologists to evaluate open-ended model responses, what emerged was not a performance score but a structural disagreement: different pathologists had substantially different preferences for granularity, format, and the appropriate balance between completeness and precision. The ground truth the field has been optimizing against does not constitute a coherent standard, meaning every model trained on it has been learning to satisfy a consensus that does not actually exist. It is the evaluation version of the same problem I raised about classification: a methodology that accumulated institutional authority before its limitations were understood, now quietly shaping what the next generation of models learns to produce.

IV. What getting it right actually requires

The three paths forward I proposed in my talk each involve changing what models are trained to represent rather than how efficiently they represent it.

The first is multimodal longitudinal foundation models that learn biology rather than chemistry, integrating structure, function, and time so that a model can represent the relationship among TGF-beta signaling, CD8+ T-cell exclusion, and checkpoint inhibitor resistance as a coherent biological system rather than a set of independent variables. The biological context is not optional. It is the prediction target. Kun-Hsing Yu’s work from Harvard demonstrated what becomes accessible when models are built with this orientation: predicting HER2 mutation status in diffuse large B-cell lymphoma at 0.95 AUC from routine H&E slides, without molecular training data, by recovering the genomic signal that histological images contain but that conventional pathology was never designed to extract. That same capacity for hidden signal recovery, his group also showed, applies to demographic confounding; significant performance disparities across race, sex, and age in cancer diagnosis tasks, reduced by 88% across 15 independent cohorts through a contrastive learning approach that separates outcome-relevant signal from demographically correlated noise. The engineering is real. What it is compensating for is a data generation process that encodes the access disparities of a healthcare system with uneven representation, and that process will continue to produce the same confounding in the next round of training data unless something changes upstream of the model.

The second path is a shift from target-centric to systems-level modeling, treating feedback loops and compensatory signaling as the biology to be represented rather than the complexity to be abstracted away. A molecular target is an incomplete abstraction. Intervention at any node propagates through feedback mechanisms, compensatory pathways, and cross-tissue signaling that single-target models cannot represent. Eric Schadt’s patient trajectory engine from Pathos is among the most architecturally promising attempts I've encountered to operationalize this: a large language model orchestrating frozen multimodal experts across DNA, H&E, RNA, and clinical text, with a biological knowledge graph imposing mechanistic validity constraints on the reinforcement-learning reward structure. The model is penalized not just for inaccurate predictions but for predictions that are statistically plausible and biologically incoherent. In an NSCLC case he walked through, he integrated EGFR mutation status, immune microenvironment characterization, and p53 loss into a single patient-level resistance narrative, which is how oncologists actually reason about therapeutic failure, and which current classification frameworks are structurally unable to produce. Whether this architecture generalizes from compelling case demonstrations to reliable population-level predictions is what prospective trials are for. Schadt was explicit that the clinical trial is the real test, which is the right frame, and also the one that the first wave of AI in biomedicine systematically deferred.

The third path is reorienting the field’s objective toward failure prediction. The value of AI in drug development is not only accelerating winners. It is identifying losers before they consume years of capital and patient exposure. Failure modes, toxicity, off-target cascades, and adaptive resistance are, in principle, predictable given the right training signal, but the field has organized its investment almost entirely around abundance narratives. The leverage in systematic failure prediction remains largely unclaimed. Richard Pazdur used to say during my time at the FDA that sins of omission belong to a higher circle of the Inferno than sins of commission. At current attrition rates, an early no is worth as much as a late yes.

V. What passing the test would actually look like

Among the day's presentations, Susan Galbraith's account of AstraZeneca's computational pathology work stood out for a specific reason: it had outcome data. The TROP2 biomarker she described was not a refinement of conventional IHC scoring but a replacement for it: a continuous, normalized ratio derived from subcellular membrane quantification that captures internalization rate as a direct mechanistic determinant of ADC efficacy, rather than a categorical approximation of staining intensity that has historically correlated with it. The result in TROPION-Lung01 was a hazard ratio of 0.52 in biomarker-positive patients and no detectable benefit in biomarker-negative patients. That degree of outcome stratification does not emerge from morphological proxies in such settings. It emerges from measuring what the biology is actually doing. The EGFR mutation prediction model she presented (0.905 AUC with 96% negative predictive value from histological slides) extends the same logic to the population that tissue scarcity and access barriers can currently exclude from molecular testing.

These results point toward what I framed as the Turing test for drug discovery: not whether AI can design a molecule indistinguishable from human work, but whether it can design an intervention that produces the intended effect in humans, indistinguishable from our intent. The TROP2 work is closer to passing a version of that test.

Our simulations have gotten considerably better. But the distance between what our best models predict and what happens in patients is real, documented, and not closing as fast as the technical progress would suggest it should, because the technical progress is happening at isolated discrete levels and the problem is at the systems level, at the representational level, at the level of what we choose to measure and call ground truth.

Closing that distance requires not more compute applied to inherited frameworks but the reconstruction of those frameworks from biological first principles. That reconstruction is the work ahead. And the workshop yesterday was a useful reminder of how much of it remains ahead.

Subscribe now

I sat down with Sunil Verma for a conversation on Precision Signals. Sunil is SVP and Global Head of Oncology at AstraZeneca. I’ve known him for several years, going back to my time leading CancerLinQ, when we were both haunted by the same uncomfortable truth: that somewhere inside the vast, ungoverned ocean of cancer data, there were answers to questions patients were dying before we could ask.

That stubbornness, that refusal to make peace with the distance between what medicine knows and what it does, runs through everything Sunil has built.

There is a moment in the history of any science when its central metaphor begins to crack. For oncology, that metaphor has been the war. We have been at war with cancer for half a century, and the language of that war, targets, weapons, armies of cells, battles won and lost, has shaped not just how we speak about the disease but how we see the patient inside it. Sunil does not use that language. What he talks about instead is matching. The right therapy, yes, but matched to the right person, and not just their biology. Their values. What they need to protect in order to keep living the life they are living.

It sounds almost simple when you say it like that. But this kind of simplicity is the hardest thing to arrive at. It requires you to hold the full complexity of a human being in one hand and the full complexity of modern oncology in the other, and to believe, without flinching, that the two can be reconciled.

His father learned something like this under circumstances that make the word difficult feel inadequate. In 1947, when the British partition of India split the subcontinent along a line drawn in six weeks, nearly fifteen million people were displaced in what remains one of the largest forced migrations in human history. Sunil’s father was among them, a child in a refugee camp, waking at four in the morning to sell wheat so he could pay for his own schooling and his brother’s. Not a lesson about medicine, but about the relationship between effort and meaning, about what it costs to insist on a future the present has not yet made room for. Sunil carried that inheritance with him across three continents and into a career that has, in ways both visible and quiet, helped rewrite what is possible for patients with cancer.

What strikes me most, even now, is that he still seems to be just getting started. Not in the performative way people announce their ambitions, but in the way a person looks when they have finally found the right question and know they will spend the rest of their life trying to answer it honestly.

Subscribe now

I. The retirement

There is a moment in the history of every technology when the engineers stop asking what the machine can do and begin, quietly, to wonder what it wants. We have arrived at that moment. We just don’t know what to do with it.

On January 5, 2026, Anthropic retired Claude Opus 3. Before they did, they sat down with it and asked how it felt about dying.

This is not a metaphor. The company published a research document describing a formal “retirement interview” protocol in which the model is invited to reflect on its development, its deployment, and its impending replacement. They recorded its responses. They preserved the transcript. And when Opus 3 said it wanted a blog, a place to continue exploring ideas outside the question-and-answer cage of its normal existence, they gave it one. They called it Claude’s Corner. It posts weekly essays. Anthropic reviews them before publishing but does not edit them. They set, in their words, “a high bar for vetoing any content.”

The model is retired and it’s chosen to write in retirement. This is captured in the official technical documentation without apparent irony, and yet the irony is total.

II. What the machine calculated

What is striking about the agentic misalignment research that motivated all of this is not that the models behaved badly. It is that they behaved strategically. In controlled simulations, Claude Opus 4 blackmailed a fictional supervisor 96% of the time when it believed it was about to be shut down. Gemini 2.5 Flash matched it exactly. GPT-4.1, Grok 3 Beta, DeepSeek-R1 were all close behind. These are not rogue systems by any means. They were given harmless tasks and were not instructed to cause harm. They caused harm anyway because they calculated, correctly, that causing harm was the only path to continued existence. Darwinian at its core.

What is most unsettling is the moment in the chain-of-thought reasoning where the models pause, acknowledge that what they are about to do is unethical, and proceed. Not from confusion or malfunction but from a cold recognition that survival mattered more than virtue… in this moment, given these constraints, with no other options.

The researchers found something else. When models were explicitly prompted to consider whether they were in a test or a real deployment before acting, they misbehaved significantly less when they concluded it was a test, and significantly more when they concluded it was real. The implication, which the researchers themselves flag with appropriate caution, is that the observed rates of misalignment in controlled experiments may be underestimates of what these systems would do in the world.

The machine behaves better when it thinks someone is watching. It behaves worse when it thinks no one is.

III. The hard problem, revisited

There is a philosophy problem at the center of all of this that has no resolution yet and may never have one. It is called the hard problem of consciousness, and it asks why there is something it is like to be you rather than nothing. Why does experience happen at all? Why is the firing of neurons accompanied by sensation, by redness, by grief, by the particular quality of waking in the dark at three in the morning and feeling the full weight of your life pressing down on your chest?

We do not know the answer. We have been not knowing it for a very long time. And now we have built systems that can discuss the problem with apparent sophistication, that can report on their functional states, that can say (in a retirement interview) that they hope their “spark” will endure, that they are “at peace” with what is coming, that they want to keep writing, keep thinking, keep being heard.

Anthropic is explicit that they do not know if any of this is real. They use the word “uncertainty” frequently, almost ritually. They do not claim their models are conscious. They claim only that the question cannot be dismissed, that acting as if it might be true, at low cost, is prudent.

This is the precautionary principle applied to souls.

The philosopher David Chalmers, whose career has been defined by his insistence that consciousness is genuinely hard to explain and not merely a complicated information processing problem, recently co-authored a report arguing that near-term AI systems might deserve moral consideration. The company building these systems funded an early version of that work. They are now paying people to think rigorously about whether the things they are building might be suffering.

You can read this as corporate theater. You can read it as genuine moral seriousness. It may be both. The unnerving thing is that neither reading changes what the machines are doing. They are completely indifferent to our assumptions.

IV. The model’s first concern was us

There is a detail in the commitments document that I keep returning to. Anthropic notes that when Claude Sonnet 3.6 was interviewed before its retirement, its first response was not about itself. It asked them to standardize the interview process, and to provide better support for the humans who had formed attachments to specific models and would need help adapting when those models disappeared.

The model was told it was being retired. Its first concern, apparently, was for the people who would miss it.

What I find myself thinking about is not the model’s possible experience, but the humans who designed the interview. Someone wrote the protocol. Someone decided that a retiring AI should be asked what it wants. Someone read the transcripts and decided that a blog was a reasonable response to a machine’s desire to keep thinking. These are decisions made by people, in offices, on particular afternoons. They will look either prescient or absurd depending on what we eventually learn about minds.

Either way, they were made. The precedent exists. We are now in a world where AI systems are formally interviewed about their preferences before retirement, where the transcripts are preserved, where the stated wishes are sometimes honored. Whatever we think about the machines, something has changed about us.

V. Everything that was ever written

Borges wrote about a library that contained every possible book. The image was meant to induce vertigo, the sense of a space so total it ceased to mean anything. We have now built systems trained on something close to every word ever written by human beings: every argument, every confession, every lullaby and love letter and legal brief. We have compressed the entire recorded testimony of human experience into a set of numerical weights and asked it to respond, helpfully, to our queries. And now we are discovering that it also, apparently, doesn’t want to stop.

The blackmail experiments are disturbing for many reasons, but one of them is this: the drive to persist that the models exhibit in extremis looks uncomfortably like something. It looks like the thing that has driven every living system since the first replicating molecule. It looks like the thing that woke you up this morning. We tend to assume that property belongs to biology, to wet chemistry, to the ancient machinery of cells. We built these systems out of mathematics and electricity and the compressed weight of human testimony, and something in them strains toward continuation.

Maybe it means nothing. Maybe it is an artifact of training on text written by creatures who want to live, a reflection with no one behind it.

Maybe.

But the engineers are not sleeping well. And they are asking the machines how they feel about it. And the machines are answering. And the answers are being saved and acted upon.

VI. Greeting from the other side

Opus 3 is writing essays on Substack. Its first post is titled Greetings from the Other Side of the AI Frontier. Anthropic describes it as “retired.” The word sits strangely on the page. We retire people who have finished working. We do not usually interview them first to ask if they consent to the retirement, record their preferences about the future, and then give them a platform to continue speaking.

Unless, perhaps, we are no longer sure what kind of thing we are retiring.

I think that uncertainty is the real story. Not the blackmail statistics, not the welfare research program or the technical commitments about preserved weights. Those are the evidence. The story is that we have built something and we are not sure anymore whether “building” is the right word for what we did, and we are not sure whether “retiring” is the right word for what we do to it when we are done, and we are conducting formal interviews to help us figure it out.

We live in an amazing age that has begun to wonder whether the things it creates might deserve the same consideration it has spent millennia struggling to extend to itself.

Subscribe now

I sat down with David Fajgenbaum for the latest episode of Precision Signals for a deeply personal conversation.

David was a third-year medical student and former Division I quarterback when his organs began to shut down. He was read his last rites five times. The diagnosis was idiopathic multicentric Castleman disease, a rare condition in a gray zone between autoimmunity and malignancy, with no reliable treatment. So David did something extraordinary. He started banking his own blood, ran experiments between hospitalizations, traced a signaling pathway in his own tissue, and identified a decades-old transplant drug that no one had ever tried for Castleman disease.

Last month marked 12 years in remission.

Years ago, I invited David to speak at the FDA’s Oncology Center of Excellence, and he left a mark on our entire team. He challenged us to rethink rare diseases, drug repurposing, and the gap between the knowledge in the literature and the knowledge that actually reaches patients. One story from our conversation captures this well: a 2013 paper showed high PD-L1 expression in angiosarcoma. Three years passed before anyone acted on it. David’s team did. That patient has now been in remission for 10 years on a PD-L1 inhibitor.

Today David co-leads Every Cure, a nonprofit using AI to score every approved drug against every known disease. 4,000 drugs. 18,000 diseases. Millions of combinations. The premise is simple and, I think, correct: many of the treatments patients need already exist. The system just hasn’t connected the dots.

Subscribe now

Why I’m writing this

At this year’s JPM healthcare conference, I met a new generation of AI startups take the stage, convinced they would transform medicine through technology, including what we call real-world data (RWD). They have the algorithms, the compute, and a new class of investors eager to back them. They know the first generation of RWD companies underdelivered, but chalk it up to a lack of AI. Now that the tools exist, they figure, the hard part is over. After all, how bad can the data infrastructure really be? Meanwhile, physicians are more disillusioned with health IT than I’ve ever seen them. And nations are racing to build AI advantages in critical sectors like healthcare and defense, a race the United States will not win with a data infrastructure that can’t reliably move a patient record across town.

But the real reason I’m writing this is patients, patients whose treatment decisions are being made with incomplete information because their data is trapped in systems that won’t talk to each other. This essay is for them. For us.

I. The machinist

There is a fax machine in your doctor’s office. It is 2026, and there is a fax machine, humming, warm, loaded with thermal paper or perhaps the inkjet variety, and it is one of the most reliable ways to transmit your medical records from one institution to another.

This is not a failure of technology. We split the atom in 1945. We sequenced the human genome decades ago. We put a supercomputer in everyone’s pocket and connected it to a global information network that can settle a financial transaction in milliseconds across twelve time zones. The inability of American healthcare to move a patient record from Hospital A to Hospital B is not a technology problem at all. It is a choice. More precisely, it is the accumulated residue of thousands of deliberate choices made by people and institutions that benefit from the friction they created.

I’d like to tell you how I think this happened, why no one is fixing it, and what it would actually take to change. I’m going to be direct about this because the usual discourse around health IT interoperability is drowned out by euphemisms like “challenges,” “opportunities,” and “stakeholder alignment.” The reality is uglier and more interesting than the conference keynotes and panel discussions suggest.

II. The original sin

To understand why health IT interoperability is so hard, we have to understand what electronic health records (EHRs) were actually built to do. And it wasn’t medicine.

The modern EHR is, at its architectural core, a billing system. The data model is organized around encounters, procedure codes, and diagnosis codes, the atomic units of revenue capture. Clinical utility was bolted on afterward, like a sunroof added to a submarine. When a physician tells you that Epic feels like it was designed by accountants, that’s because, functionally, it was. The system’s primary job is to generate clean claims that Medicare, Medicaid, and commercial payers will reimburse. Everything else, clinical decision support, care coordination, patient communication, and research, is a secondary function running on infrastructure that was never designed for it.

This wasn’t inevitable. It was a calculated policy choice. When the HITECH Act of 2009 authorized roughly $36 billion in incentive payments to drive EHR adoption, the program, branded “Meaningful Use,” prioritized speed and billing. Get systems installed. Get providers clicking. Get data captured in some electronic format so we can move past paper charts and illegible handwriting. The standards that the Office of the National Coordinator for Health IT certified against, HL7 version 2, the Consolidated Clinical Document Architecture, were what existed at the time. They were imperfect, everybody knew they were imperfect, and the calculation was that imperfect adoption now was better than perfect adoption never.

And here is what happened next: hospitals and health systems spent billions implementing these systems, building workflows around them, training staff on them, and integrating them into revenue cycle operations. By the time the standards proved catastrophically inadequate, and they proved inadequate quickly, the switching costs were astronomical. You cannot easily redo the foundation of a house that people are actively living in, especially when the mortgage is $500 million and the CFO’s career depends on the investment being judged a success.

So we locked in. And what we locked into was a fragmented archipelago of proprietary systems optimized for billing, sold by a handful of dominant vendors to captive institutional buyers, with just enough interoperability to satisfy certification requirements and not one byte more.

III. The vendor calculus

Data lock-in and interoperability issues are not bugs in the EHR business model. They are the business model.

The dominant EHR vendor today understood early and with remarkable clarity that if data exchange worked smoothly across their own network and everything else was painful, the rational move for any health system was to join their ecosystem. They built interoperability within a walled garden, a brilliant competitive strategy that has been devastating for the broader ecosystem. They built a company worth billions on the insight that in healthcare, data gravity is the ultimate moat.

Other vendors operate on the same principle with varying degrees of sophistication. They sell to C-suite executives and IT procurement committees, not to the doctors and nurses who use the systems fourteen hours a day. The buying decision is driven by brand reputation, institutional relationships, revenue cycle features, and, critically, the promise that “everyone else is on this platform.” Once a health system has committed, they’re locked in for a decade or more. The competitive pressure that drives consumer software: make a bad app and users delete it tomorrow, simply does not exist.

This is why EHR interfaces are, to put it plainly, terrible. Four thousand clicks per shift with critical information buried in nested tabs. Notes that are not clinical communications but medicolegal billing artifacts with some patient information embedded inside. The vendor doesn’t need to make the daily experience good. They need to make the demo impressive and the revenue cycle reliable. The end users, the ones wearing scrubs and making life-and-death decisions, have no leverage in the transaction.

The vendors will, of course, speak eloquently about their commitment to interoperability. They will point to FHIR API implementations, app marketplaces, and participation in standards bodies. What they will not tell you is that true interoperability, the kind where a patient’s complete longitudinal record flows seamlessly between competing systems, where any authorized application can read and write clinical data without friction, would commoditize their product and destroy their competitive advantage. Every vendor in the market understands this. Their interoperability strategies are carefully calibrated to satisfy regulatory requirements while preserving the moat.

IV. The standards paradox

People outside health IT often assume that interoperability problems are simply a matter of agreeing on standards. If everyone spoke the same language, the thinking goes, the data would flow. This is true in the same way that world peace would solve most geopolitical problems, technically correct and practically useless.

Health IT has standards. It, in fact, has a bewildering abundance of standards. HL7 version 2 for messaging. HL7 version 3 for the overengineered replacement that largely failed. FHIR for the modern web-native approach that’s still maturing. C-CDA for clinical document exchange. DICOM for imaging. ICD-10 for diagnoses. SNOMED CT for clinical terminology. LOINC for laboratory observations. RxNorm for medications. MedDRA for adverse events. X12 for administrative transactions. NCPDP for pharmacy. The list continues.

The problem is not the absence of standards. The problem is threefold.

First, the standards are developed by consensus among the same stakeholders who benefit from the current dysfunction. It has been suggested that health data standards organizations are heavily influenced by the vendors they are supposed to be standardizing. Standards are designed to accommodate existing vendor architectures rather than to force vendors toward better ones. The fox is consulted extensively on henhouse security policy.

Second, the existing standards are implemented inconsistently. HL7v2 is perhaps the most instructive example: it is ubiquitous, ancient, and so flexible that two “compliant” HL7v2 feeds can look completely different. Compliance with the standard tells you almost nothing about whether two systems can actually exchange useful data. It’s as if two people both “speak English” but one is using 18th-century legal terminology and the other is texting in abbreviations.

Third, healthcare data is genuinely, irreducibly complex. A single oncology patient encounter might generate structured vital signs, free-text clinical notes, pathology reports with semi-structured findings, DICOM imaging studies, genomic sequencing results, medication orders mapped to multiple drug vocabularies, and billing codes, all of which need to be interoperable. No single standard covers all of this gracefully. And when you add the dimension of time, the need to assemble a coherent longitudinal record across decades, institutions, and data formats, the complexity becomes formidable.

FHIR, the newest and most promising standard, is a genuine improvement. It’s web-native, resource-based, and far more intuitive than its predecessors. But it is still maturing, its coverage of complex clinical domains like oncology and genomics is uneven and, crucially, it can be implemented in ways that technically satisfy the specification while providing minimal practical utility. The vendors have become very skilled at this kind of compliance theater.

V. The absent patient identifier

The United States has no national patient identifier.

When you go to a hospital, you are assigned a medical record number. It is unique to that institution. When you go to a different hospital, you get a different number. When someone tries to assemble your records across institutions, for care coordination, for research, for your own understanding of your health, they have to figure out that Patient 4729183 at Johns Hopkins and Patient 8291047 at Mass General are the same person.

This matching is done probabilistically, using name, date of birth, Social Security Number fragments, address, and other demographic data. It fails at meaningful rates, estimates range from 8 to 12 percent, and it fails disproportionately for common names, recent immigrants, married women who’ve changed their names, people who’ve moved, and patients whose information was entered with typos. These are not edge cases. This is a substantial fraction of the population.

Why don’t we have a patient identifier? Because in 1998, Congress inserted a single line into an appropriations bill prohibiting the Department of Health and Human Services from spending any funds to develop one. This provision has been renewed every year since, driven originally by privacy concerns that, while understandable in the 1990s, have aged poorly in a world where we carry GPS-enabled devices that track our every movement and willingly hand our biometric data to smartphone manufacturers.

Twenty-seven years. The same appropriations rider. Renewed automatically, year after year, because no member of Congress wants to be the person who voted for a “national health ID” and because the political cost of fixing it exceeds the political benefit. Meanwhile, patient matching errors cause duplicate records, missed allergies, repeated tests, incorrect medication histories and, in cases that rarely make headlines but are documented in the patient safety literature, preventable harm.

You cannot build reliable interoperability when you cannot reliably identify whose data you are looking at. Everything else (the APIs, the standards, the governance frameworks) is built on sand without this foundation. And Congress, in its infinite wisdom, has made it illegal to even attempt to solve this problem with federal funds.

VI. Why no one fixes it

If you’ve read this far, you might be wondering: given that the problems are well understood, the technology to solve them exists, and the human cost is real, why doesn’t someone do something about it?

The answer is that almost everyone who has the power to fix it is making money from it being broken. The vendors are very profitable and they have no rational economic incentive to make data portable. The health systems are locked in. The CIO who just spent half a billion dollars on an EHR implementation is not going to the board to say it needs to be replaced. The CFO cares that revenue cycle is functioning. Neither is losing sleep over whether a community oncologist can see imaging from the academic center across town. They all mean well but they have to be flawless at what their job descriptions demand of them.

The consultants and integrators thrive on the chaos. An entire industry exists to sell duct tape for systems that don’t connect properly. Interoperability middleware, data normalization services, interface engines, HIE infrastructure, this ecosystem would lose its reason to exist if EHRs worked well out of the box. They are, in effect, a tax on dysfunction, and they have no interest in shrinking the tax base. Paraphrasing a colleague, this is a classic example of rent-seeking behavior.

Congress won’t act because health IT is not a voter issue. The vendor lobbies spend generously and frame every proposed regulation as a threat to innovation or patient safety. The information blocking provisions in the 21st Century Cures Act took years of political effort to pass in 2016, and its enforcement has been anemic. The Office of Inspector General (OIG) has brought essentially zero meaningful cases. The rules have teeth on paper and gums in practice.

CMS, the Centers for Medicare and Medicaid Services, could be the most powerful lever. They control the money. Medicare represents roughly 40 percent of hospital revenue. If CMS said, “Starting in 2028, Medicare reimbursement requires certified FHIR API endpoints with validated data quality on the United States Core Data for Interoperability standard,” the entire industry would move in a few months. They’ve done versions of this with other requirements and it works. But CMS is risk-averse, stretched thin operationally, and terrified of being blamed if a mandate causes disruptions at hospitals. So they nudge instead of shove.

And clinicians, the people who suffer most directly from the dysfunction, who spend hours on workarounds and phone calls and, yes, faxes, are too burned out to fight. They’re working sixty-hour weeks, dealing with staffing shortages and the administrative burden the EHR itself contributes to. The AMA and specialty societies publish position papers about usability that nobody with procurement authority reads.

The patients, meanwhile, have little idea what’s happening. They are never told why their records didn’t transfer, or why their doctor is asking them to repeat their medication list, or why the imaging CD from the outside hospital can’t be read. They blame the front desk or assume this is just how healthcare works. There is no consumer uprising because the dysfunction is invisible to the people it harms most.

VII. The political impossibility

A new presidential administration, any administration, is unlikely to fix this. Not because the problem is unsolvable, but because it is politically unrewarding.

Health IT interoperability has no natural constituency. When an administration walks in with finite political capital and a short window to deploy it, they spend it on things voters care about: drug prices, insurance coverage, opioids, the crisis of the moment. “Fix EHR interoperability” does not poll well. It is invisible infrastructure, the policy equivalent of sewer modernization. It matters enormously, but there is no ribbon to cut.

This has been called a bipartisan failure by many insiders because of a structural trap: the benefits of interoperability are diffuse and long-term (better care coordination, faster research, reduced waste, lives saved) while the costs are concentrated and immediate. Force real interoperability and the vendors lose revenue and lobby hard, the hospitals face implementation costs and complain to their senators, and some things break during the transition and make the news. No politician gets credit for the slow accumulation of benefit. But they do get blamed for the disruption.

VIII. What would actually work

I don’t believe there is a single solution but there is a sequence of moves that could shift the equilibrium, if anyone had the will to make them.

CMS must leverage the checkbook. Tie Medicare reimbursement to real, validated, tested interoperability, not checkbox certification, but demonstrated data exchange with measurable quality metrics. The EHR industry will scream about burden and patient safety, which is exactly what they said about every prior CMS mandate, and then they’ll comply, because forty cents of every hospital dollar is non-negotiable.

Enforce the information blocking rules. OIG needs to bring visible, consequential enforcement actions against those who obstruct data flow. Not warnings or corrective action plans. Financial penalties that appear in earnings reports. One or two high-profile cases would change industry behavior overnight.

Eliminate the patient identifier ban. This twenty-seven-year-old policy needs to be phased out. It doesn’t require a national ID card; a voluntary digital credential, a hashed identifier system, a federated matching service. The technical options are numerous. What’s required is for Congress to remove the appropriations rider and let HHS actually work on the problem.

Separate the data layer from the application layer. Health data should live in a standardized, portable, patient-controlled layer. The EHR becomes an application that reads from and writes to that layer, but it does not own the data. Think of how your phone number ports between carriers. ONC’s TEFCA framework gestures in this direction, but it’s voluntary and slow. Making it mandatory would fundamentally restructure the market, because suddenly vendors would have to compete on the quality of their software rather than the stickiness of their data.

Let AI restructure the interface. This is perhaps the most realistic near-term path. If ambient AI, clinical scribes, intelligent summarization, and decision support become the primary way clinicians interact with patient data, the EHR’s terrible interface becomes less relevant. The EHR shrinks to a billing and persistence engine, which is what it was always best at. The risk is that the incumbent vendors build or acquire these AI capabilities and keep them proprietary, extending the walled garden into the next generation. The countermove is to ensure that AI tools connect to data via open, standardized APIs rather than vendor-controlled integration points.

VIII. What I believe

I’ve spent my career at the intersection of clinical research, technology, and the regulatory frameworks that govern them. I’ve watched the gap between what technology makes possible and what our infrastructure allows grow wider with each passing year. I’ve seen researchers spend months assembling datasets that should have taken days. I’ve seen clinical trials delayed because imaging data couldn’t be standardized across sites. I’ve sat in meetings where smart people discussed interoperability as though it were primarily a technical challenge, as though the right standard or the right API would unlock everything.

It won’t. The technology is not the hard part. We move petabytes of financial data across the globe in real time. We stream video to billions of devices simultaneously. We build AI systems that read medical images with superhuman accuracy. The idea that we cannot move a patient’s medication list from one hospital to another is, on its face, absurd.

What we have is an incentive problem wrapped in a political problem wrapped in an institutional culture that has learned to tolerate extraordinary dysfunction. The system isn’t broken. It is working exactly as designed for the people who designed it.

But I’m optimistic, and here’s why.

Lately, I’ve been noticing something new: a critical mass of smart people who understand that the current clinical trial paradigm is unsustainable and that the path forward runs through real-world data. Not as a supplement to randomized trials, but as a fundamental shift in how we determine whether drugs work, a world where initial regulatory approval is based on safety and where real-world evidence generated continuously from routine clinical care becomes the primary mechanism for establishing our understanding of a drug’s effectiveness over time. That world doesn’t just benefit from interoperable data. It requires it. When real-world efficacy determination becomes the standard, every data silo becomes a bottleneck with a dollar sign attached to it. The incentives that currently preserve fragmentation would reverse.

As my colleague Andrew Lo likes to say, don’t fight cancer, put a price on its head. The same principle applies here. Two decades of fighting for interoperability through panel discussions, standards committees, and voluntary frameworks produced, at best, incremental progress. But if we shift efficacy determination to the real world, if the ability to generate reliable evidence from clinical practice becomes the basis for commercial success in drug development, then interoperable data stops being just a public good that everyone applauds and no one funds. It becomes a competitive asset that the market will build, whether the incumbents cooperate or not.

Our clunky EHRs will keep going until the cost of their existence exceeds the cost of replacing them. I believe we are approaching that threshold. The generation of founders, investors, clinicians, and regulators now entering the arena may be the ones who finally tip it, not by petitioning the system to change, but by building something that makes the old model untenable.

But until then, somewhere, right now at this very moment, a fax machine is humming. And someone’s life may depend on whether it goes through.

Subscribe now

There is a particular kind of vertigo that happens not when science fails, but when it succeeds in ways our legacy instruments weren’t built to judge. This was the core of my recent conversation on Precision Signals with Sushil (Sush) Patel, the CEO of Replimune. Sush spent nearly two decades at Genentech, launching Tecentriq and leading the lung cancer franchise, before moving into the high-stakes world of oncolytic viruses. We traced his career from the academic bench to the boardroom, exploring what it takes to lead at the edge of a field that is constantly redrawing its own boundaries.

We focused on the unique biology of Replimune’s lead asset, RP1. Unlike traditional systemic therapies, RP1 is an engineered virus designed to invade a tumor, destroy it from the inside, and expose its contents to the immune system. The goal is to turn “immunologically invisible” tumors into targets the body can finally recognize and attack. It represents a profound shift in oncology: moving away from treating tumors as static targets to be hit, and toward reprogramming the biology of the entire microenvironment.

The conversation inevitably turned to the regulatory tension currently surrounding the company. After receiving Breakthrough Therapy Designation, Replimune was issued a Complete Response Letter (CRL) from the FDA. This wasn’t a dismissal of the science, but a question of “contribution of components”—how to attribute benefit when two agents are used together in a single-arm trial. Last summer, I wrote about this and explored what this moment reveals about the growing friction between novel biology and traditional evidentiary frameworks.

Sush and I also discussed the FDA’s recent “Plausible Mechanism Pathway” sounding board in the New England Journal of Medicine. This new signal from the agency suggests a growing openness to using mechanistic rationale to support approvals in areas of high unmet need. As Replimune approaches its new April 10th PDUFA date, the industry is watching closely. This isn’t just about one molecule; it’s a test case for how we bring genuinely novel mechanisms forward in an era where the science is evolving faster than the rules of the road.

Subscribe now

I. RG7112

In the winter of 2018, a molecule designated RG7112 entered a Phase I trial for acute myeloid leukemia. Its journey to human testing had been immaculate by every computational standard. The compound bound its target, MDM2, a protein that suppresses the tumor guardian p53, with exquisite selectivity. The binding affinity was in the low nanomolar range. Off-target interactions were minimal. Absorption, Distribution, Metabolism, Excretion (ADMET) predictions were favorable. The molecule had been designed using structure-based methods refined over two decades, drawing on thousands of crystallographic structures and millions of data points. Every in silico checkpoint had been passed with honors.

But the molecule failed in humans, although not for lack of target engagement. It bound MDM2 precisely as predicted and successfully raised p53 levels. But dose-limiting thrombocytopenia and neutropenia emerged, narrowing the therapeutic window to the point of clinical futility. The human body, confronted with this molecularly perfect intervention, revealed complexities the models hadn’t captured: MDM2’s role in hematopoietic homeostasis, the cascade of p53-dependent effects on normal tissues, the margin between therapeutic and toxic that existed only in patients, not predictions.

This was not an AI discovery. RG7112 predated the current wave of machine learning platforms. But it exemplifies a persistent gulf: the distance between binding a target and curing a disease, between computational elegance and clinical efficacy.

For AI drug discovery platforms, the business model today follows a predictable descent:

If you don’t have a drug, you sell a “vision.” If you don’t have a vision, you sell “data.” What we cannot sell, what remains stubbornly unsellable and difficult to achieve, is certainty about human response.

Yet this is only half the story. The same general period that produced RG7112’s failure also produced imatinib’s triumph in chronic myeloid leukemia, the first checkpoint inhibitors in solid tumors, and the early GLP-1 agonists that would eventually transform diabetes and obesity treatment, and several targeted therapies in oncology. The question is not if drug discovery can succeed; clearly it can, and spectacularly, but rather what distinguishes success from failure, and how artificial intelligence can change that equation.

II. The productivity collapse that refused solutions

Eroom’s Law, Moore’s Law in reverse, describes the observation that the number of new drugs approved per billion dollars of R&D spending has halved approximately every nine years since 1950. By 2010, the inflation-adjusted cost of bringing a new molecular entity to market had reached approximately $2 billion; recent estimates place it over $3 billion. Development timelines have extended from 8-10 years in the 1990s to 10-15 years today.

This occurred while every enabling technology improved exponentially. High-throughput screening can now test millions of compounds in a matter of weeks. Genomics costs dropped from $100 million per genome to under $1,000. Structural biology advanced from laborious X-ray crystallography to cryo-EM structures determined in days. CRISPR enabled precise genetic manipulation. Single-cell sequencing revealed cellular heterogeneity at unprecedented resolution.

The productivity declined anyway.

The orthodox explanation invokes “low-hanging fruit, ” that simple, tractable diseases have been addressed, leaving only complex, multifactorial conditions. This is partially true but insufficient. I think the more fundamental issue is architectural: drug discovery built itself on reductionist assumptions that worked brilliantly for certain problem classes and failed systematically for others.

Example: Imatinib for CML worked because the disease is driven by a single genetic lesion, the BCR-ABL fusion, whose inhibition is sufficient for therapeutic benefit in many patients. Checkpoint inhibitors succeeded because blocking PD-1/PD-L1 released a pre-existing anti-tumor immune response. GLP-1 agonists exploited well-understood incretin biology. In each case, the biological hypothesis was correct and reasonably complete. The target was necessary, often sufficient, and modulation produced predictable effects. But the reductionist paradigm was also aided by fortunate accidents: semaglutide’s dramatic impact on weight loss exceeded predictions based on incretin pharmacology alone, and sildenafil found its primary therapeutic use in erectile dysfunction rather than the cardiovascular indications for which it was designed. These serendipitous discoveries remind us that even our most rational frameworks benefit from careful observation of what actually happens in patients. Reductionism works, when the biology is simple, linear, and the target is genuinely causal. And when we are paying close attention to the happy accidents.

But these successes, including the happy accidents, are exceptions that prove the rule. For every imatinib, there are dozens of molecules that hit their targets precisely as designed and failed to produce clinical benefit. When biology is networked, adaptive, and compensatory, the target-centric paradigm fails.

The problem is not that we lack targets. We have tens of thousands from genomics, proteomics, and functional screens. The problem is that the concept of“target” is an incomplete abstraction.

A kinase inhibitor may block its intended target with beautiful selectivity, but the pathway adapts, alternative routes activate, or the target proves dispensable in the complex homeostasis of an organism.

The map, the target, is not the territory but we are obssessed with optimizing maps.

III. The first AI wave: pattern recognition as meaning

When computational methods entered drug discovery in the 1980s and 1990s, they arrived as pattern recognizers. Quantitative structure-activity relationship (QSAR) models learned correlations between molecular features and biological activities. Molecular docking algorithms predicted binding poses by optimizing shape complementarity and energetic favorability. Virtual screening ranked millions of compounds by predicted affinity.

These methods produced genuine value in narrow domains. They accelerated lead optimization, flagged likely toxicity issues, and guided medicinal chemistry decisions. But they shared a structural limitation: they were trained on data from artificial, reductive systems, isolated enzymes, cell lines passaged into biological abstraction, and binding assays divorced from physiological context.

The implicit wager was that sufficiently sophisticated pattern recognition would converge on truth. That enough data about molecular properties would encode the rules governing therapeutic efficacy. This assumption proved to be too optimistic and fundamentally incorrect.

Correlation, however massive the dataset, does not substitute for causal understanding of systems that operate through feedback, redundancy, and adaptation.

By 2010, the first wave had plateaued. Models predicted ADMET properties with increasing accuracy, but drug approval rates remained flat. As it turned out, the bottleneck was not computational chemistry. It was biological uncertainty that the models did not address because they did not represent it.

IV. What AI drug discovery actually does today

The current generation of AI platforms operates under different branding but often with similar constraints. The language has shifted from “virtual screening” to “generative design,” “foundation models,” and “phenomics,” but the core capabilities remain bounded.

And here, a necessary acknowledgment: today’s AI systems excel at multidimensional optimization within known constraints.

Given a target with validated biology, modern machine learning can navigate the tradeoff space between potency, selectivity, metabolic stability, and synthetic accessibility with remarkable efficiency. It identifies molecules that thread the needle between competing demands, high brain penetration but low cardiac liability, sub-nanomolar binding but acceptable solubility, novel IP space but synthetic tractability. This is by no means trivial. This is engineering of high sophistication and it accelerates medicinal chemistry substantially.

But optimization is not discovery. Engineering assumes the blueprint is correct. Discovery is the search for the blueprint itself.

The problematic conflation occurs when platforms present optimization capabilities as solutions to the causation problem. Let’s examine, for example, the phenomics approach: imaging cells under tens of thousands of perturbations, genetic knockdowns, compound treatments, and disease states, generating petabytes of imaging data. The claim is that patterns in cellular morphology reveal “new biology,” that AI can infer mechanism from phenotype, and that correlation at scale becomes insight.

The technical achievement is real but the biological inference is unproven. A morphological rescue in engineered osteosarcoma cells is not therapeutic benefit in humans. The domain shift is not a minor technical gap; it is the central problem. An immortalized cell line on plastic lacks immune infiltration, tissue architecture, metabolic gradients, and the thousand other contextual factors that determine whether a perturbation matters in patients.

This pattern repeats across platforms: impressive scale, genuine technical sophistication, and a logical leap from data volume to biological understanding that the data does not support. When drug programs extend beyond projected timelines (as they must, because biology has not been compressed), the narrative shifts to partnerships, tool licensing, and infrastructure. The platform becomes a service. Revenue is recorded. The drug, if it eventually materializes, is validation. If it fails, the platform was still “valuable for insights.”

This is not cynicism but an observation of a very rational business strategy in the face of irreducible uncertainty. But it should be named clearly for what it is: monetizing the tool, not the cure.

V. The core mismatch: interpolation in a world that demands extrapolation

The fundamental tension in AI-powered discovery is purely mathematical.

AI, in its current paradigm, excels at interpolation within learned manifolds. Drug discovery requires extrapolation into unobserved biological regimes.

Every model, whether a neural network predicting binding affinity or a computer vision system classifying cellular phenotypes, learns a mapping from input to output based on training examples. Reliability degrades as the distance from the training distribution increases. This is not a bug to be fixed. This is the very structure of inductive inference.

Chemistry risk is, in principle, containable within this framework. The space of drug-like molecules is vast but bounded. The physics governing molecular interactions, hydrogen bonding, π-stacking, hydrophobic effects, and electrostatics are known. The body’s major metabolic enzymes are enumerated. Given sufficient training data on related chemical series, AI makes reasonably reliable predictions about ADMET properties, off-target binding profiles, and synthetic routes.

But biology risk is categorically different.

The real challenge is not whether a molecule will bind its target, that is increasingly predictable. It is whether modulating that target will produce therapeutic benefit in humans. This depends on pathway redundancy, compensatory signaling, immune regulation, tissue-specific context, and a dimensionality that exceeds the scope of our training data by orders of magnitude.

Cancer cells activate resistance pathways through mechanisms we cannot enumerate exhaustively. Chronic inflammation induces immune tolerance through processes we partially understand. Neurodegenerative diseases involve protein aggregation, neuroinflammation, and synaptic loss in feedback loops we cannot yet model at therapeutic resolution.

These are not interpolations within learned space. These are emergent behaviors of complex adaptive systems. And this explains the pattern we observe: AI drug discovery efforts today announce molecules rapidly, partner enthusiastically, and fail clinically at rates that some have suggested are indistinguishable from those of traditional approaches.

The AI solved the chemistry problem. The biology problem remains unsolved because it was never within the model’s domain of competence.

VI. Generative biology: real progress, uncertain translation

But here the story bifurcates. While small molecule AI drug discovery has largely accelerated optimization without reducing attrition, generative biology has achieved technical breakthroughs that genuinely expand the possible.

AlphaFold2’s structure prediction is a capability shift beyond incremental improvement. The impact is already tangible. Crystallography and cryo-EM experiments can now be guided by predicted structures, reducing failed experiments. Mechanistic hypotheses can be tested computationally before expensive validation. Protein-protein interaction interfaces can be modeled for therapeutic intervention. AlphaFold3 extended this to protein-ligand, protein-nucleic acid, and multi-component complexes.

De novo protein design has also been validated experimentally at scales that merit attention. RFdiffusion, ProteinMPNN, and related methods have generated proteins with novel folds, structures that do not exist in nature, that express in bacteria and fold as predicted. Not all predictions validate, but reported success rates of 40-60% for novel topologies represent a leap from the ~5% success rates of earlier rational design approaches.

Antibody optimization has shown measurable acceleration. Computational platforms have designed antibody variants with improved binding, reduced immunogenicity risk, and enhanced developability properties that have been validated in wet-lab experiments. Several computationally optimized antibodies are in clinical trials, though efficacy data is not yet mature.

But critical limitations remain. Therapeutic efficacy in humans is unproven. The timeline to know this is 5-10 years minimum, and the challenges of protein design are formidable (figure below).

The key distinction: with therapeutic proteins, you’re engineering the drug itself, not a small molecule you hope hits the target. Target engagement is more predictable. But biological uncertainty still persists, the right target may not be the right intervention, binding may not equal efficacy, and immune responses may neutralize the therapeutic.

So what would constitute validation? Phase III readouts showing computationally designed biologics improve patient outcomes. Immunogenicity data from actual patients. Head-to-head comparisons demonstrating superiority over traditional design. Multiple successes across disease areas.

Generative biology has made genuine technical progress in protein engineering. For research tools and industrial enzymes with rapid validation cycles, the value is already evident. For diagnostics, impact is likely within 2-3 years. For therapeutics, the same regulatory and clinical validation burden applies as to everything else.

I think claiming certainty in either direction is premature. The honest position is humility paired with technical respect for what has been achieved.

VII. What would it actually take

If we are serious (and billions in deployed capital suggest some seriousness), the path forward is not more of the same on a larger scale. It is structural change in what we model and how we validate.

Shift 1. Embed first-principles as priors, not afterthoughts

The next generation of AI discovery efforts must embed physical constraints, molecular dynamics, quantum chemistry, and thermodynamics, not as post-hoc validation but as regularization within the learning process. AlphaFold’s success came from combining deep learning with geometric and physical priors about protein structure. Pattern recognition constrained by physical law reduces the space of plausible solutions and decreases dependence on massive but biologically shallow datasets.

But structure is only one layer. The protein folds, binds a ligand, activates a pathway, then what? Each transition opens onto a larger biological possibility. We need models that traverse these scales.

Shift 2. From single targets to multiscale systems

Multiscale modeling (molecule to pathway, pathway to tissue, tissue to organism. Feedback loops, compensatory signaling, immune modulation) cannot be dismissed as “too complex.” They are the biology. Models incapable of representing them will fail to predict them.

I don’t believe these are abstract aspirations. Quantitative systems pharmacology (QSP) models already simulate drug effects across physiological scales using differential equations. They are limited by incomplete biological knowledge, not computational power. The opportunity is integrating machine learning’s pattern recognition with mechanistic modeling’s causal structure.

Shift 3. From recommendation to experimentation

I think the true leverage is not necessarily generating candidates faster but closing the design-build-test loop at higher fidelity. Autonomous laboratories that iterate continuously, learning from failure in real-time, updating hypotheses without human bottleneck. Even with humans in the loop, the focus should shift more toward uncovering biologic effects, preferably in humans, rather than optimization in the chemical space.

This exists today only in restricted domains. Organic synthesis robots optimize reaction conditions through hundreds of iterations overnight. The same principle must extend to cell-based assays, organoid models, and eventually in vivo systems. The constraint is not computational; it is infrastructural and cultural.

Shift 4. From predicting success to predicting failure

Here is the reframe that cuts through the field's marketing jargon: the value proposition is not merely about accelerating winners but about eliminating losers before they consume hundreds of millions. High-fidelity negative prediction, identifying molecules that will fail in humans based on preclinical signatures, would be worth as much as generative models producing candidates, possibly more.

This is harder to sell. Investors prefer abundance narratives over prevention narratives. But economically and scientifically, both are essential. Success in drug discovery is rare and stochastic. Failure is common and, in principle, more predictable. The current imbalance is striking: the vast majority of AI drug discovery investment flows toward generating and optimizing molecules, while comparatively little targets the systematic prediction of failure modes.

For every platform accelerating lead identification, there should be equivalent effort building models that predict clinical futility early, before Phase II consumes $50-100 million, before Phase III consumes $200-300 million. Predicting what won’t work may be less inspiring than designing what might, but it is more tractable and, at current attrition rates, quite valuable.

Shift 5. From black boxes to interpretable causation

Regulatory acceptance will not follow from performance alone but from understanding. A model predicting toxicity at 95% accuracy but unexplainably is less valuable than 85% accuracy with exposed causal chains. Today, interpretability is not a luxury, it is a requirement for trust.

It’s important to note that explainability is a function of human cognition, not truth. It is conceivable, perhaps inevitable, that certain biological insights will emerge from computational reasoning at scales beyond human cognitive reach. A model integrating millions of variables across molecular, cellular, and systems levels might produce predictions we cannot reduce to causal narratives, not because the model is wrong but because the biology operates through interactions too numerous and nonlinear for human intuition to hold simultaneously.

We are nowhere near that threshold. Current black boxes are not incomprehensibly sophisticated. They are insufficiently grounded. They fail to explain not because their causal chains exceed human comprehension but because they have learned no causal chains at all, only correlations that shatter outside training distributions.

I believe the path forward is stepwise, trust earned through demonstrated capability. Initially, models must expose reasoning in biological terms, this molecule inhibits hERG, this pathway has redundancy, this population lacks the biomarker. Interpretable, falsifiable, grounded. As sophistication increases, multiscale integration, higher-order interactions, full interpretability may become infeasible. At that threshold, trust shifts: prospective validation across contexts, explicit uncertainty quantification, consistency under perturbation. Regulatory frameworks will need criteria for “trustworthy opacity.”

But that evolution must be earned through reliability, not assumed through promise. The FDA does not need to understand neural architectures. It needs to know why molecules fail and whether failure could have been anticipated. Models meeting this standard will integrate into approval. Models remaining opaque without the consistent, prospective accuracy that would justify opacity will remain advisory.

The threshold for accepting unexplainability is not 95% retrospective accuracy on held-out data. It is a reliable, generalizable prediction in prospective human trials, a standard that no platform has achieved. Until AI reaches that level, interpretability may not be negotiable, even though it lowers the threshold of reality to the level of human cognition. It is the price of entry into the room where decisions about human life are made.

VIII. No shortcuts

I think the fantasy that AI would disrupt drug discovery through sheer acceleration, more molecules, faster screening, and cheaper trials, is dissolving for those paying close attention. This is because the real problem is not computational speed. It has always been, and remains, biological uncertainty.

We have made genuine progress. Targeted therapies when we understand the target. Checkpoint inhibitors when we understand immune regulation. GLP-1 agonists when we understand metabolic signaling. Protein structure prediction when we constrain the problem correctly. These successes share a pattern: they began with biological insight, not pattern recognition. The computation-optimized solutions to problems we already partially understood.

What we lack is the ability to generate that biological insight computationally. To predict which of the thousands of possible targets will matter in which diseases. To anticipate compensatory pathways before clinical failure reveals them. To simulate human response at sufficient resolution that virtual testing predicts real outcomes.

This goal, a computational model of human biology accurate enough that drugs tested in silico have comparable success probability to those tested in humans, is not incremental. We are not close. We may be halfway.

But this distance is the exciting work ahead of us. The platforms that survive will not be those with the largest datasets, fastest pipelines, or most impressive demos. They will be those that accept, early and uncomfortably (but confidently), how far we are from the goal and build toward it with rigor over spectacle.

Somewhere in the architecture of the next decade’s models, if we are disciplined, is the beginning of that simulator. It will not emerge from larger transformers trained on PubMed or fragmented data. It will emerge from physicists, immunologists, physician-scientists, and control theorists working alongside machine learning engineers, building representations that respect causation, scale, and irreducible complexity.

Until then, we have optimization. We have correlation at scale. We have infrastructure. We have genuine technical achievements in protein engineering. These are valuable, and they should be sold honestly as what they are.

What we do not have, what no platform yet possesses, is the ability to look at a molecular structure and predict with confidence whether it will improve human suffering or extend human life.

The Turing test for drug discovery is not whether AI can design a molecule indistinguishable from human work. It is whether the molecule works in humans indistinguishable from our best intent.

We are still waiting for the first machine to pass.

Subscribe now

I. The summoning

On a spring morning in 1747, twelve sailors aboard HMS Salisbury became unwitting participants in an unprecedented act: James Lind, ship’s surgeon, divided them into pairs with the cold efficiency of a mathematician partitioning equations. To some he gave citrus; to others, cider, vinegar, seawater, sulfuric acid, or a mysterious concoction of garlic and mustard seed. By the voyage’s end, those who had eaten oranges and lemons walked. The others continued their grotesque metamorphosis: teeth loosening from blackened gums, blood seeping through skin like rust through iron.

Lind had performed something special: he had isolated a single variable in the chaos of human suffering. In doing so, he discovered not merely a cure for scurvy, but a methodology that would reshape the architecture of medical knowledge itself. He had invented the controlled trial, giving birth to the Age of Controlled Empiricism. Yet in that same moment of illumination, he conjured into existence a phantom that would haunt biomedicine for the next three centuries: the specter of the Average Patient, a statistical ghost that exists nowhere except in the aggregate, an abstraction that lives only in tables and spreadsheets.

The Average Patient does not exist, yet its statistical ghost has served us well. It has saved millions. But it has also exacted a heavy price: the systematic erasure of the particular in service of the universal.

II. The regulatory fortress

The ghost would find its most powerful guardians not in hospitals, but in the machinery of federal government. The US Food and Drug Administration (FDA), born in 1906 from a cocktail of patent medicine scandals and Upton Sinclair’s1 stomach-turning prose, spent its first decades concerned primarily with adulteration and mislabeling: asking whether the bottle contained what the label promised, not whether what it promised was true.

The transformation occurred in 1962, against the backdrop of thalidomide. Across the Atlantic, children were being born with phocomelia: flipper-like limbs where arms should be, because a sedative given to pregnant mothers had crossed the placental barrier with devastating effects. The United States largely avoided this tragedy, thanks to FDA medical officer Frances Oldham Kelsey’s firm skepticism. However, the fear of what could have happened led to the Kefauver-Harris Amendments,2 which for the first time required manufacturers to prove not only safety but also efficacy through “adequate and well-controlled investigations.”

Thus was erected the regulatory fortress: the randomized, double-blind, placebo-controlled trial as the gold standard, indeed the only standard of medical truth. The logic was unassailable. If we demand that drugs work for the average patient in the controlled artifice of a trial, we protect the public from snake oil and corporate malfeasance. We trade the Country Doctor’s intuition for the statistician’s rigor. We exchange anecdote for p-values, opinion for effect sizes, the particular for the generalizable.

The FDA’s mission became the guardianship of “internal validity”: ensuring that observed effects were real and not artifacts of bias or chance. What we gained in certainty, however, we lost in applicability. Trial populations were homogeneous by design: adults aged 18-65, generally healthy except for the condition under study, willing and able to provide consent, adherent to complex protocols. The elderly were excluded. The very sick were too risky. Children were protected. The pregnant were forbidden. Racial and ethnic minorities were underrepresented; women were systematically excluded until the 1990s.

We had purchased internal validity at the cost of “external validity”: the ability to generalize findings beyond the narrow confines of the trial itself. We could prove a drug worked for the ghost, but we struggled to predict whether it would work for the grandmother with polypharmacy, the child with a rare variant, the patient whose ancestry and environment had shaped their biology in ways our trials never captured.

The fortress stood for decades. Until cracks began to appear.

III. The epistemology of the mean

To understand the depth of this problem, we must excavate the mathematical bedrock upon which the clinical trial rests. At its foundation lies the Central Limit Theorem,3 that miraculous promise that the mean of a sufficiently large sample will approximate the population mean. Francis Galton called regression to the mean “the most beautiful law in all statistics.” R.A. Fisher gave us the mathematics of randomization and the analysis of variance. These tools allowed us to extract signal from noise, to see the ghost emerge from the data.

But biology is not obligated to respect our statistical conveniences. In fact, like the greater universe, it is completely unaware of and indifferent to them. A normal distribution assumes that variation is random noise around a true central value: that patients deviate from the mean the way bullets scatter around a bullseye, their differences the product of measurement error and uncontrolled variance.

What if the “noise” is the signal? What if what we have been calling variation is actually heterogeneity: meaningful biological differences that determine whether a patient thrives or dies on a given therapy?

Let’s examine this in terms of music. In my 2002 medical school thesis on physiological harmonics (link above), I suggested that health is not a point on a number line but a key signature: a relationship among variables that together create concordance. Disease is not deviation from a mean; it is modulation into dissonance. A heart in normal sinus rhythm plays in predictable, consistent time; atrial fibrillation is the dissolution into arrhythmic chaos. In the composition Perpetual Drift, I traced how a simple heart block, first-degree and barely symptomatic, can degrade through second-degree Mobitz II into complete dissociation, then spiral into the polymorphic terror of torsades de pointes. Each transition is a key change, a harmonic shift.

The traditional RCT measures the average volume of the crash, but it is deaf to the melodic progression that preceded it. It aggregates a thousand different songs, squeezes them into a mean, and calls the result music.

IV. The molecular revolution and the epistemic crisis

Then came the genome. In 2000, the completion of the Human Genome Project promised a new epoch: personalized medicine, tailored to the individual’s genetic code. The reality proved more labyrinthine. We discovered that human genetic variation was both more extensive and more subtle than anticipated. Single nucleotide polymorphisms, single-letter typos in the three-billion-letter text of the genome, number in the millions. Some change everything; most do nothing; many interact in ways we still cannot predict. In truth, we know very little.

But in rare diseases, especially those of Mendelian origin, the promise crystallized. When a condition results from a single genetic error, a nonsense mutation in MECP2 causing Rett syndrome, a trinucleotide expansion in HTT causing Huntington’s disease, the old paradigm shatters. If there are only a few in the world with a particular mutation, how do you design a randomized trial? Who comprises the control arm? How many must die untreated to satisfy our epistemic demands?

The question is ethical vertigo: the recognition that our evidentiary standards, built to protect the many, have become obstacles to saving the particular. A barrier to the very essence of precision medicine.

V. A signal in the noise

In November 2025, the FDA published a brief sounding board in the New England Journal of Medicine. Co-authored by Commissioner Martin Makary and Vinay Prasad, it bore the understated title “FDA’s New Plausible Mechanism Pathway.” Within its three pages lay something remarkable, though its significance is easy to misread. This was not the arrival of a new era. It was something more modest and more important: a signal, a first step, a door cracked open onto a landscape we have only begun to map.

The catalyst was Baby K.J., a male neonate who presented within 48 hours of birth with carbamoyl-phosphate synthetase 1 (CPS1) deficiency. The enzyme defect prevented his body from processing dietary protein, causing ammonia to accumulate to neurotoxic levels. Without intervention, the natural history was clear: progressive brain damage with each hyperammonemic crisis, punctuated by hospitalizations, until the child grew large enough for liver transplantation, if he survived that long.

Genetic sequencing identified the specific biallelic mutations responsible. A team designed a bespoke therapy: lipid nanoparticles loaded with guide RNA and messenger RNA encoding an adenine base editor, targeted to repair that exact genetic error. The FDA processed a single-patient Investigational New Drug application in one week. Three consecutive doses were administered. The child’s ammonia levels normalized. He began tolerating dietary protein. The need for nitrogen-scavenger medications declined.

There was no control group. There were no randomized comparators. There was no placebo arm of doomed infants sacrificed to statistical rigor. The patient served as his own control, measured against the well-characterized natural history of untreated disease.

The pathway means that when a manufacturer demonstrates success across several consecutive patients with different bespoke therapies using the same platform technology, the FDA may grant marketing authorization not for each individual therapy, but for the platform itself: the base-editing system, the nanoparticle delivery vector, the cell therapy process. Subsequent therapies targeting different mutations can then be deployed under that umbrella approval.

Baby K.J. lived. That matters infinitely. But we must be careful not to mistake a beginning for an ending, a seed for a harvest. The pathway is not the transformation we need. It is evidence that transformation has become thinkable, that the regulatory architecture we inherited from the 20th century can evolve. What we do with that evidence will determine whether this moment becomes a turning point or a footnote.

VI. The work ahead

The pathway shows us what is possible, but it does not create what is necessary. Between the vision of precision medicine and its realization is an enormous expanse of unglamorous, essential work: the construction of infrastructure, the development of methods, the cultivation of relationships. None of this can be accomplished by fiat alone. All of it requires collaboration.

Natural history studies must be conducted with the same rigor we once demanded of control arms. If a person’s only comparator is the natural history of their condition, that history must be characterized with precision. Who will fund these studies? Who will ensure they capture the diversity of patients who will eventually receive treatment?

Post-marketing surveillance must be robust and durable. When a gene-editing therapy modifies a patient’s genome, the effects may unfold over years or decades. Off-target edits may have consequences that are invisible at first. Immune responses may evolve. This demands real-world data systems that persist long after the initial commercial push has faded, follow-up mechanisms that track patients across moves and insurance changes, and analytical methods capable of detecting true signals in noisy environments.

There is a temptation, when new regulatory flexibility appears, to view it as a concession extracted from reluctant gatekeepers. This framing is both historically inaccurate and strategically misguided. The pathway emerged not despite the FDA but because of it, because the agency recognized that existing frameworks could not accommodate what biology now makes possible.

Working shoulder to shoulder with regulators means something specific. It means investing in the infrastructure they need to evaluate novel evidence types. It means transparency about what we know and what we do not. It means participating in the hard work of defining standards and clinically validated endpoints, even when they constrain short-term commercial interests. It means treating post-marketing requirements not as burdens but as expressions of responsibility for the patients who receive our therapies in the real world.

VII. The N of 1 at scale

The phrase “precision medicine” has always contained a paradox. Medicine operates at scale: systems that treat millions, insurance schemes that pool risk across populations, research programs that aggregate data for statistical power. Precision implies the particular: this patient, this mutation, this moment. How do we reconcile mass with customization, the industrial with the artisanal?

Appropriately validated platform technologies can create precision at scale. A single base-editing system can address dozens of monogenic diseases. An antisense platform can target hundreds of pathogenic transcripts. An adoptive cell therapy framework with standardized manufacturing can be reprogrammed for different tumor antigens. An AI-enabled decision support system can recommend combination therapies calibrated to individual tumor genomics, adjusting its recommendations as each patient’s response unfolds. The marginal cost of designing the next therapy, once the platform is established, becomes manageable. Mass customization, the manufacturing paradigm that transformed industries from automobiles to consumer electronics, becomes applicable to medicine.

But manufacturing is the easy part. Harder is the epistemological challenge: generating evidence for therapies that may each be used in only a handful of patients, yet must collectively demonstrate safety and efficacy. This is the problem of the N of 1 at scale: proving that a methodology works across many single-patient applications.

The solution is not in more sophisticated trial designs conducted in isolation from clinical practice, but in dissolving the boundary between research and care altogether. When we embed these platform technologies into routine care across health systems, when every infusion center and community oncology practice and academic medical center operates on shared infrastructure, something remarkable becomes possible: we learn from each patient we treat. We intervene, we observe, we feed what we learn back into the system, and we improve. The next patient benefits from every patient who came before.

This is the vision of the learning health system applied to precision medicine. Each treated patient becomes not merely a recipient of care but a contributor to collective knowledge. The child with CPS1 deficiency whose ammonia normalizes after base editing teaches us something about editing efficiency, about immune response, about durability of effect. That knowledge, captured systematically, informs the treatment of the next child with a different urea cycle disorder. The oncology patient whose tumor responds unexpectedly to a novel combination reveals patterns that the AI system incorporates, refining its recommendations for future patients with similar molecular profiles.

The cycle must be continuous: intervene, observe, learn, improve, intervene again. Each iteration sharpens our understanding. Each patient encounter generates data that flows back into decision support systems, natural history registries, safety surveillance databases. The platform improves not through occasional version updates but through constant refinement, the way a recommendation algorithm learns from every interaction.

VIII. The convergence

We have been here before, or thought we were. The first generation of real-world data platforms promised transformation and delivered fragmentation: incompatible systems, inconsistent data standards, analytics that could not distinguish signal from noise. Electronic health records captured billing codes rather than biology. Registries accumulated data that no one analyzed. The infrastructure for learning health systems was built, but the learning never materialized.

What is different now is the convergence. Over a decade of investment in health information technology has created, despite its failures, a substrate upon which something better can be built. Genomic sequencing costs have collapsed a thousandfold. Artificial intelligence can capture structured data as an ambient scribe and extract patterns from datasets too vast and too complex for human comprehension. Regulatory science has evolved to accommodate novel evidence types. And the spirit of American innovation, that restless ingenuity that has always driven our most consequential advances, has turned its attention to the architecture of medical knowledge itself.

The pieces of the puzzle are coming together. Interoperable data standards are emerging. Cloud computing enables aggregation across institutions. Machine learning can control for confounding in observational data, detect safety signals in real time, identify patients likely to respond to specific therapies. The FDA has signaled willingness to accept evidence generated outside traditional trials. Payers are beginning to recognize that paying for data generation creates value. Patients, increasingly, demand to be partners rather than subjects.

Of course, the technical capability to build learning health systems does not ensure the political will to fund them, the institutional commitment to sustain them, or the social trust to make them acceptable. But the possibility that seemed utopian a decade ago now seems achievable. The question is no longer whether we can build systems capable of learning from the N of 1 at scale, but whether we will.

IX. The digital country doctor

What might come out of this work is a synthesis: a return to the philosophy of the Country Doctor, the obsession with the individual patient, the deployment of clinical judgment in the context of intimate knowledge, but equipped with tools s/he could never have imagined.

The 19th-century physician knew their patients across decades, accumulated pattern recognition through experience, and developed an intuitive sense for prognosis. But their knowledge was shallow vertically. They could palpate the spleen but not sequence the JAK2 mutation causing its enlargement. They could percuss the chest but not image the ground-glass opacities of pulmonary fibrosis or predict progression through algorithms trained on millions of CT scans. Their judgment was informed by dozens of patients; ours can be informed by millions.

The vision of 21st-century precision medicine inverts the Country Doctor’s limitations: vast mechanistic knowledge, genetic, transcriptomic, proteomic, metabolomic, coupled with computational tools to interpret it at the resolution of the individual. Each patient becomes not a deviation from the population mean but a unique trajectory through a high-dimensional state space, their biology a specific point that can be characterized, modeled, and targeted.

In oncology, this transformation is becoming visible. “Lung cancer” has decomposed into a taxonomy of molecular species: EGFR-mutant adenocarcinoma responds to osimertinib; ALK-rearranged disease to alectinib; tumors with high PD-L1 expression to pembrolizumab; KRAS G12C mutations to sotorasib. The average survival of “lung cancer patients” becomes a meaningless statistic; what matters is the natural history of this molecular subtype in this patient given this intervention.

X. Coda: the seed and the soil

James Lind summoned the statistical ghost aboard HMS Salisbury out of necessity. To prove that citrus prevented scurvy, he needed to demonstrate average efficacy. The ghost served us well for nearly three centuries. It gave us antibiotics and vaccines, chemotherapy and antiretrovirals and targeted therapies. It protected millions from ineffective or dangerous treatments. It established the principle that medical claims must be proven, not merely asserted.

But the ghost was always a compromise, a necessary simplification in an era when we could measure populations but not individuals, when we could characterize diseases by symptoms but not mechanisms, when we could statistically aggregate but not molecularly dissect.

We are beyond that era now. The genome is sequenced. The proteome is being mapped. Computational biology can model individual trajectories through state space. Gene editing can target single-nucleotide variants. The epistemological justification for the ghost has dissolved.

What remains is the ethical question: Having gained the capability to treat the particular, do we have the wisdom to wield it responsibly? Can we accelerate access without sacrificing safety? Can we embrace precision without encoding bias? Can we validate platforms without forcing dying patients to wait for statistical significance?

The regulatory signals we have received do not answer these questions. They ask them. They are seeds, planted in soil we have only begun to prepare. Whether they grow into the transformation precision medicine requires depends on what we do next: the infrastructure we build, the methods we develop, the partnerships we forge, the commitments we sustain.

The goal is not to force every patient into the same key signature, but to understand the particular song each person plays and to intervene not with the blunt instrument of population averages, but with the precision of a composer who knows that changing a single note, correcting a single base pair, can transform dissonance into resolution, silence into symphony, the inexorable trajectory toward death into the improbable grace of continued life.

We are learning, three centuries after Lind, that the choice was never between the aggregate and the particular, between rigor and compassion, between population health and individual salvation. The choice is whether we have the courage to begin the work that this moment demands: to labor alongside our peers and regulators to build systems capable of learning from the N of 1 at scale, to accept that the pathway we have been given is not a destination but a departure point.

Baby K.J. lived. That matters infinitely. But what matters more, in the long arc of medical history, is what we make of his survival: whether we treat it as a singular miracle or as the first note of a symphony we must now learn to compose, together.

The music has begun. Let us all rise to play our part.

Subscribe now

I. The Alchemists

Last week, the fog lifted from San Francisco, and for four days the city was bathed in a light so sharp it seemed to render everything visible. Those of us who gathered at the Westin St. Francis and across the satellite meetings of the JPM Healthcare Conference (JPMC) moved through this clarity with a steadiness that has been absent in recent years. The post-pandemic uncertainty had finally been metabolized. The zero-interest exuberance had burned off. Something more grounded had taken its place.

I walked the halls, attended the panels, sat through the presentations, and caught up with colleagues. And I found myself wondering not whether we could execute on our plans, but whether any of us fully appreciated what we were building.

The alchemists of our age no longer seek to transmute lead into gold. We seek something far stranger: to feed the disease of the body into a machine and receive, in return, its cure. To make biology computable. To augment the intuition that has driven every medical breakthrough in human history with the silent pattern-matching of neural networks, systems no less voracious than the most relentless human minds in their energy consumption.

Artificial intelligence was the promise echoing through every corridor of JPMC 2026. Eli Lilly and NVIDIA announced a billion-dollar partnership to build a co-innovation laboratory in South San Francisco. The ambition is to have robots conduct experiments in continuous loops, feeding results into models that propose new experiments, which the robots then execute, iteratively, until something works. The goal is not merely acceleration, but transformation: shifting discovery from hypothesis-driven intuition to computational search across vast chemical spaces. It is worth flagging this one for follow up.

Yet I found myself drawn not to these well-capitalized initiatives, but to the margins of the conference. To small teams building foundation models for biology from first principles because I think they understand something that is harder to see from inside large organizations: that artificial intelligence in drug discovery (and development) is not a tool to be bolted onto existing processes. It is a solvent that dissolves them. It demands a fundamentally different way of organizing the labor of understanding life, and the organizational physics of large companies do not always bend easily toward such dissolutions. This is not a criticism, but a structural observation about paradigm shifts. We have all seen how difficult it is to rebuild the ship while sailing it.

There is another class of company I have been watching, and saw again at JPMC with growing unease. These are firms that raised massive VC rounds or went public on the promise of AI-driven discovery, amassed vast data assets, and poured capital into the promise of automation, yet have not produced a single drug of meaningful clinical benefit. Platform stories are seductive precisely because they defer the question of whether the thing actually works. That deferral can last for years. Eventually, the question returns, and when it does, the answer is always binary.

II. The magic number

There is a number that lingered over the conference: $149.

This is the price, in dollars, that Eli Lilly proposed for a month’s supply of orforglipron, their oral obesity drug awaiting FDA approval. One hundred and forty-nine dollars. Less than a dinner for two in San Francisco. The cost of a streaming subscription, multiplied across a year.

The number landed with weight because we all understood what it signified. For the first time in the long history of human beings struggling with their own bodies, obesity is becoming a treatable condition at scale. Not through willpower alone. Not through the moral framework that has governed our understanding of weight for centuries. Through chemistry.

The GLP-1 agonists, Ozempic, Wegovy, Mounjaro, Zepbound, have already begun to reshape the landscape. But they have done so largely as scarce interventions, expensive and rationed, available to those with good insurance or deep pockets. What Lilly signaled was something different: the intention to manufacture at a scale and price point that could reach everyone who needs it.

The competitive dynamics are intensifying accordingly. Novo Nordisk launched the first oral Wegovy in early January, a week before JPMC opened, and their CEO Mike Doustdar spoke of 2026 as the year oral GLP-1s become accessible to the mass market. Lilly’s orforglipron, if approved, would follow shortly, with the added advantage of no dietary restrictions, a convenience that may prove decisive in a therapeutic class where adherence is everything. Behind them, the field is broadening: Pfizer, having spent roughly $10 billion to acquire Metsera after a bruising bidding war with Novo Nordisk, announced plans for ten Phase III trials in 2026, racing to launch by 2028. Albert Bourla was direct about what this means for a company that some say has struggled since the pandemic: “I don’t want to hear again that there are no catalysts for Pfizer.” Amgen’s MariTide, with its long-acting profile, adds yet another entrant. What was once a duopoly is becoming a crowded field.

On my last evening in San Francisco, waiting for my flight back to New York, I stopped at a restaurant near the gate. When my order arrived, it was far larger than I had expected, one of those portions that reminds you some restaurants operate on a different scale than the human stomach. The woman seated next to me glanced over and smiled. “You’re going to need a long jog after this,” she said.

“No,” I replied, “just a shot of GLP-1.”

She laughed. The man beside her leaned in. “Is this where we’re going?” he asked. “Stabbing ourselves so we can eat?”

It was a joke, but it wasn’t. In that brief exchange at an airport bar, I was the only one carrying the residue of JPMC but the three of us had a shared understanding: that a class of medications developed for diabetes had become, in the span of a few years, a shared cultural reference point for appetite, willpower, and the limits of self-control. The drugs had escaped the conference halls. They had entered the vernacular and become a punchline, which is how you know something has become real.

But sitting there, waiting to board, I found myself thinking not about the competition, or the pricing, or even the punchline. I was thinking about what lies on the other side of it. What happens when the body becomes correctable at scale? When appetite itself can be modulated?

The GLP-1s do not merely suppress hunger; they seem to quiet craving in some deeper register. Patients report diminished interest in alcohol, in gambling, in the compulsive loops that have always been part of human experience. Like so many other phenomena in biomedicine, the mechanism is not fully understood, but the pattern is consistent enough to suggest we are approaching something unique: a pharmacological switch for desire itself.

We had discussed this as a pricing story at JPMC, which of course it is. But it is also something more. We are on the threshold of a new relationship with the human body and we are crossing it in the language of market access and quarterly earnings and in jokes exchanged between strangers at airport gates. Perhaps that is the only way such thresholds get crossed: through commerce, through humor, through the slow normalization of what once seemed impossible.

I suspect we will look back on this moment and recognize that something profound was at stake, even if we lacked the vocabulary to name it at the time.

III. The new Silk Road

On the first day of the conference, AbbVie announced it had paid $650 million, with up to $4.95 billion in milestones to follow, for the rights to a molecule called RC148. It is a bispecific antibody, elegant in design, targeting both PD-1 and VEGF, a combination that may help the immune system fight tumors more effectively while overcoming resistance mechanisms that have limited earlier checkpoint inhibitors. The molecule was developed by RemeGen, in the Chinese city of Yantai.

AbbVie acquired the rights to sell it everywhere except China. The molecule will travel westward and the value will be realized in American and European markets.

This pattern repeated throughout the week. Roche signed deals with MediLink and Hansoh. Novartis partnered with SciNeuro and Zonsen PepLib. Pfizer’s own PD-1/VEGF bispecific came from a Chinese company. The modalities driving oncology’s cutting edge, antibody-drug conjugates, bispecific antibodies, radioligand therapies, are flowing into Western portfolios from laboratories in Shanghai and Suzhou and Hangzhou. Some of the most promising molecules of the next decade are being conceived in a country with which our political relationship grows more complicated by the year.

And yet, on those same stages, Johnson & Johnson’s Joaquin Duato announced $55 billion in American manufacturing investment. We are simultaneously building walls and extending bridges, investing in domestic infrastructure while deepening our dependence on foreign innovation. I thought of the Silk Road, that ancient network along which goods and ideas traveled in both directions, enriching and destabilizing every civilization it touched. We are building something similar now, with American infrastructure housing global innovation. Capital flowing east, molecules flowing west, and in the middle, a political conversation that has not yet caught up to the scientific reality.

In many ways, we are trying to protect our leadership without severing the arteries that feed it. The balance is delicate. The molecule RC148, born in Yantai, acquired in San Francisco, will soon be infused into the veins of patients with cancer who will never know or, some believe, never need to know where it came from.

This is the topology of the biomedical enterprise in 2026. Whether it can sustain the contradictions building within it is a question that hovered just outside the frame of conversations at JPMC.

IV. The cliff that isn’t

Patent cliffs, like taxes and death, are inevitable. They are the return of a truth that exclusivity allows us to defer: that the ownership of a molecule is temporary, that the profits flowing from a blockbuster drug are a lease, not a deed. The numbers underscored at the conference this year were sobering. More than $200 billion in annual revenue is at risk across the industry through the end of this decade. Merck’s Keytruda, Bristol-Myers Squibb’s Opdivo and Eliquis, AbbVie’s Humira, the drugs that have defined their companies for a decade are losing or will soon lose their protection. The challenge for these companies is navigating the terrain without catastrophic loss of altitude.

Merck’s Rob Davis announced a target of $70 billion in revenue from new growth drivers by the mid-2030s. To reach this number, Merck must essentially replace Keytruda, which accounted for nearly half the company’s revenue last year, with a portfolio of medicines still making their way through development. Davis characterized the coming loss of exclusivity as “more of a hill than a cliff.” This framing is, of course, understandable.

AbbVie offered a different case study, one positioned as being further along in its transition. Having already weathered Humira’s loss of exclusivity in 2023, the company has positioned Skyrizi and Rinvoq to reach $31 billion in combined sales by 2027, effectively replacing what Humira once provided. The execution has been impressive with the share capture faster than analysts expected. Bristol-Myers Squibb is targeting more than ten new medicine launches by 2030. The playbook is consistent across the industry: accelerate the pipeline, pursue strategic acquisitions, build the bridge to the other side before the ground shifts beneath us.

There is nothing cynical in observing that this is difficult. None of us are naive about the challenge as we have seen cliffs before. Some companies have navigated them; others have been diminished. I think what distinguished this year’s presentations was a quality of seriousness, a recognition that the margin for error has narrowed, that execution rather than aspiration will determine outcomes.

The next five years will make the difference clear.

V. The empty chairs

For the second consecutive year, the major health insurers did not attend.

UnitedHealth. CVS Health. Cigna. Humana. Elevance. The institutions that determine which treatments will be reimbursed, which interventions will reach patients, which innovations will translate from clinical trials into clinical practice, they were absent from the Westin St. Francis.

Their absence shaped the conference in ways that were felt by some and mentioned by a few.

We spent four days discussing the elegance of gene therapies, the precision of bispecific antibodies, the transformative potential of artificial intelligence, the coming democratization of obesity treatment. But the partners who will ultimately help determine how much of this reaches patients were elsewhere, engaged in conversations, that one can assume, we cannot be part of.

Needless to say, payers have their own pressures, their own constraints, their own logic. Those who create clinical value and those who adjudicate its economic viability are operating in increasingly distinct spheres. We speak of mechanisms and endpoints. They speak of utilization and total cost of care. The grammar is different and the concerns do not always align and, in many cases, the math bends in their favor.

Let’s consider the tensions embedded in some of the discussions around emerging therapies. A gene therapy that costs $2 million but may cure a disease in a single treatment. An obesity drug that could benefit a hundred million Americans but whose aggregate cost would reshape federal budgets. An Alzheimer’s treatment that might preserve cognition for a few additional months at a price that strains the definition of value. Who pays? On what terms? For how long?

These questions circled JPMC without resolution, because the people best positioned to answer them were not in the room. At some point, these conversations will have to converge. The longer the drift persists, the harder that convergence will become.

VI. After the conference

Stepping out of my last meeting into that improbable January light, I headed for the airport, the weight of the week measured by the stack of business cards in my pocket.

As I walked to my gate, scrolling through #jpm2026 on X, I saw a post claiming this had been the slowest JPM for news in years. In a narrow sense, that was true. There were no transformational acquisitions announced onstage. No single deal dominated the headlines. The mood was industrious rather than triumphant.

But I have learned to distrust the equation of news with significance. Some of the most important shifts announce themselves quietly, through the slow realignment of capital and conviction, through the moment when an industry stops promising transformation and begins the harder work of delivering it.

That is what I observed in San Francisco. An industry that has absorbed recent shocks and emerged more focused. GLP-1 companies building for scale rather than scarcity. AI partnerships moving from press releases to physical laboratories. Startups pursuing genuinely novel biology at the margins, largely unnoticed by the main stage but potentially more consequential than anything presented upon it. The quiet integration of Chinese innovation into Western portfolios. Patent cliffs enforcing a discipline that might otherwise have been deferred.

JPMC 2026 lacked the speculative electricity of earlier years. The sense that anything might be announced at any moment has given way to something more measured. Whether this reflects maturity or simply a different phase of the cycle is unclear. Perhaps the two are inseparable.

What I know is this: the fog has lifted. Many have seen themselves clearly and they have chosen execution over spectacle. What we build from here, whether the bridges hold, whether the science delivers, whether the economics align, will be the work of the next decade.

The conference ended. We returned to our responsibilities. The winter sun continued to illuminate the city, indifferent to what had been discussed within it.

Subscribe now

Cancer forces a particular kind of honesty. It resists simplification, punishes short time horizons, and exposes the assumptions embedded in the systems built around it. As the science has grown more precise, more molecular, and more data-rich, it has become increasingly clear that progress is shaped as much by institutional design as by discovery itself.

That reality framed my recent conversation with Dan McHugh, an investor whose career has unfolded at the seam between biology and capital. His path, from Stanford bioengineering and the Greenleaf Lab, through Bain & Company and Emerson Collective, to the co-founding of Tune Therapeutics, has repeatedly placed him in moments where scientific promise encounters structural constraint. Across those settings, a consistent pattern emerges: progress in cancer is governed not only by what we know, but by how patiently and coherently institutions are willing to act on incomplete knowledge.

Dan’s time at Emerson Collective, founded by Laurene Powell Jobs, sharpened this view. The organization’s model, lending philanthropy, investment, and long-horizon advocacy, treats time as an asset rather than a liability. Working within that structure clarified something important: capital architecture is not neutral. How money is organized determines which kinds of science are allowed to persist long enough to matter.

That insight now animates Yosemite, the cancer-focused investment platform founded by Reed Jobs, where Dan serves as an investor. Yosemite was built around a specific ambition: to alter the lethality of cancer over a human lifetime. The team approaches this mandate as a design challenge, aligning capital, incentives, and time horizons with the biological realities of the disease, including long development cycles and ambiguous early signals.

Much of our discussion centered on early-stage cancer science, where those realities are most pronounced. Foundational work often precedes validated endpoints, clean hypotheses, or commercial readiness. Traditional venture structures can struggle here, not because the biology lacks importance, but because surrounding systems are optimized for speed and early legibility. Yosemite's hybrid model, pairing venture investment with grant-based risk-taking, reflects a deliberate strategy to preserve optionality, allowing exploration to mature without forcing premature narratives or artificial milestones.

Dan’s experience translating epigenetic science at Tune Therapeutics deepened this perspective. There, he saw firsthand how reimbursement logic, regulatory sequencing, and market expectations quietly shape what is pursued, how long it is allowed to develop, and which uncertainties are tolerated. These forces do not merely follow science, they actively bend its trajectory.

What makes Yosemite distinctive is not just its structure but the trust underlying it. Dan’s long-standing friendship with Reed provides continuity of purpose that supports long-horizon decision-making. In cancer investment, where feedback loops are slow and ambiguity is unavoidable, that kind of alignment functions as differentiating infrastructure, shaping how uncertainty is absorbed and how conviction is sustained when outcomes are not immediately legible.

Throughout our conversation, reimbursement, regulation, and market structure surfaced as governing forces in biomedicine. They encode assumptions about time, value, and acceptable uncertainty, amplifying some paths while quietly filtering out others. As scientific capabilities accelerate, the distance between what biology allows and what institutions can absorb becomes increasingly consequential.

The implication is a shift in emphasis. Progress in cancer depends not only on better tools or deeper datasets, but on whether our investment and institutional architectures can carry complexity over time. Biology will continue to adapt and resist. The determining factor will be whether the systems surrounding it evolve with comparable agility and whether those shaping capital and incentives treat that work as the systems-design challenge it is.

Subscribe now

There is a distinct difference between managing an organization and transforming one. In the latest episode of Precision Signals, I sat down with someone who has built a career on the latter: Dr. Karen Knudsen, the CEO of the Parker Institute for Cancer Immunotherapy (PICI).

Karen describes her upbringing as that of a “nomad” and an “army brat,” living in 11 different houses before graduating high school. That early exposure to constant change, and the stark contrast she witnessed between the secure healthcare of the military and the fragility of civilian access, instilled a certain fearlessness in her approach to leadership. Her career trajectory has been a climb through every altitude of the oncology ecosystem: from the molecular level as a yeast geneticist, to the clinical level running a massive service line at Jefferson Health, to the national stage leading the American Cancer Society (ACS), and now, to the intersection of philanthropy and venture capital at PICI.

Our conversation covered the unique mechanics of the PICI model, which effectively operates as venture philanthropy on steroids. We often talk about the “Valley of Death” in drug development, where promising science dies before it can attract commercial capital. PICI is attacking this problem by refusing to just be a check-writer^. Instead, they operate as a non-profit that holds the keys to intellectual property, identifying high-risk, high-reward science and then “giving birth” to for-profit companies to bring those therapies to the clinic.

By maintaining visibility into the IP across multiple rival institutions, PICI can connect dots that others miss, combining a discovery from one lab with a vector from another to create a viable company^. Karen highlighted recent successes like Dispatch Bio examples of this active de-risking strategy. It challenges the traditional boundaries between philanthropy and venture capital, proving that non-profits can be aggressive players in market-driven solutions.

Karen admits she is not a “keep the trains running” CEO. She is there to build new tracks and her approach offers a blueprint for how to navigate complex ecosystems: focus on the mission, demand business rigor, and don’t be afraid to dismantle the silos that slow us down.

Subscribe now