Only a Constellation of Stars
We have been promising to cure all disease for decades. The latest promise comes with $2.1 billion and a Nobel prize.
TL;DR — Demis Hassabis says his startup Isomorphic Labs will “solve all disease” — an overly optimistic claim from a Nobel laureate who raised $2.1 billion based in part on the strength of that promise. Unfortunately, the claim is the latest in a 55-year pattern of cure-all promises from credentialed institutions, none of which has delivered, most of which produced real, albeit mundane progress. There is no public evidence that this is going to change.
“Everything is Gonna Burn / We’ll All Take Turns / I’ll Get Mine Too” - Pixies
Co(n)founding
I’ve co-founded three things worth laying claim to in my life (so far). The Houston Equal Rights Alliance (HERA), a gay rights marriage advocacy non-profit. The Regeneron Genetics Center (RGC). And, in the early 1990s with two college friends (Simon Taylor and Derek Stanford), a record label called Isomorphic Records. We were nerds in college bonding over topology puns and experimental music. The label name is a reference to an old math joke about a topologist who can’t tell the difference between his coffee cup and his donut. Because two shapes are isomorphic if one can be continuously deformed into the other, the coffee cup and the donut are topologically identical. Records and CDs are similarly isomorphic (to each other and to coffee cups and donuts) hence, Isomorphic Records.
Two weeks ago the same topology reference — in the form of Isomorphic Labs — raised $2.1 billion. The implicit claim embedded in that name is that AI drug design and biological reality are the same shape, deformed by different forces; solve one and you’ve solved the other. “AlphaFold solves protein structure” is a problem in chemistry and geometry. “Isomorphic Labs solves all disease” is a problem in physiology, immunology, pharmacology, trial design, and at least four other disciplines that do not yield to abstract mathematics. The function from the first to the second is not continuous. It is not even well-defined.
Back to Isomorphic (Labs, not Records) in a minute — but first, the RGC. We announced the RGC to the world on January 13, 2014, the day before Illumina launched the HiSeq X. There was even a New York Times article about the launch1, and fun fact: George had to borrow Len’s shirt for the photo shoot. The RGC was set up to do roughly what Demis Hassabis is now saying his startup Isomorphic Labs will do: use large-scale data and computational biology to inform drug discovery. We started with a hundred-thousand-person sequencing partnership with the Geisinger Health System.2 We hit a million exomes sequenced in early 2020, the first organization in the world to reach that number; we crossed two million in early 2023 and are approaching three million as of this writing.
It has been, in my estimation, an unqualified success. The RGC has delivered more than forty novel drug targets and therapeutic programs into Regeneron’s pipeline, across chronic liver disease, obesity, cancer, and neurodegenerative conditions.3 Among the most prominent: ANGPTL3, which inspired a drug approved by the FDA in 2021 for an inherited form of high cholesterol;4 GPR75 for obesity; HSD17B13 and CIDEB for chronic liver disease; INHBE for favorable fat distribution and protection from diabetes. We published the genetic and clinical work in peer-reviewed journals over the course of a decade, and we contributed a number of open-source tools to the field including REGENIE, a whole-genome regression method widely used for biobank-scale GWAS.5 I’m retiring next month, extremely proud of what we’ve accomplished by bringing together a collaborative team of scientists and staying focused on the data.
We never said we would cure or solve all disease. Not in the founding press release. Not in any subsequent press release. Not in any of the numerous peer-reviewed papers the RGC has published over the past decade in Science, Nature, Nature Genetics, NEJM, and Nature Communications. The 2014 announcement said the new center’s objective was “to expand the use of human genetics for defining disease targets and improving the drug development process.”6 Verbs you can audit. Define. Improve. This essay is about why some ventures in this space speak the way the RGC spoke at its founding and others speak the way Isomorphic speaks. The answer, from a decade inside one of them, is not primarily a question of character.
I think for the most part everyone is honestly trying to do their best to help humanity. The question isn’t character. It’s incentive architecture — and I’m not letting myself off the hook for my own failures (and they are legion) inside that architecture. The latest person (certainly not the only and unfortunately not the last) making the loudest version of the promise is perhaps the most credentialed person who has ever made it, and the gap between his credentials and his claim is significant.
The Achievement That Earned the Right
Demis Hassabis is the co-founder and CEO of Google DeepMind, the founder and CEO of Isomorphic Labs, a knighted British scientist, an AI advisor to the UK government, and a 2024 Nobel laureate in Chemistry. He shared the Nobel with John Jumper for their work on AlphaFold, an artificial-intelligence system that predicts the three-dimensional structure of proteins from their amino-acid sequences. (The other half of the prize went to David Baker, at the University of Washington, for computational protein design.) The achievement is not contested. Protein structure prediction was one of the longest-standing open problems in molecular biology. AlphaFold solved a version of it and released structures for two hundred million proteins to the world. More than three million researchers have used it. The Nobel committee made a reasonable, defensible call.
AlphaFold is the real thing. It belongs in the same paragraph as Sanger sequencing and the polymerase chain reaction. Nothing in what follows is meant to diminish that. The reason to be careful with what Hassabis is now claiming is that the AlphaFold credential is so big it does work — rhetorically, financially, politically — far outside the bounds of what was actually achieved scientifically. The credential writes checks the achievement can’t cash.
On 60 Minutes
On April 20, 2025, Hassabis sat down with Scott Pelley on CBS’s 60 Minutes. The segment did the things 60 Minutes segments do — establishment lighting, awed cutaways, a robot playing with colored blocks (and I love good B-roll robot footage) — and somewhere in the middle, Hassabis said this:
“Proteins are the basic building blocks of life. So, everything in biology, everything in your body depends on proteins. You know, your neurons firing, your muscle fibers twitching, it’s all mediated by proteins. But 3D protein structures like this are so complex, less than 1% were known. Mapping each one used to take years. DeepMind’s AI model did 200 million in one year.”
Then Pelley pivoted, and Hassabis said this:
“So on average, it takes, you know, ten years and billions of dollars to design just one drug. We can maybe reduce that down from years to maybe months or maybe even weeks. Which sounds incredible today but that’s also what people used to think about protein structures.”
And then this:
“I think that’s within reach. Maybe within the next decade or so, I don’t see why not.”7
The “that” in the last sentence is curing all disease. Pelley had asked. Hassabis answered. Eleven days earlier, on April 9, on Reid Hoffman’s Possible podcast, he had given a slightly different version of the same line: “I think maybe in the next 10, 15 years we can actually have a real crack at solving all disease.” On X, two days after the 60 Minutes segment aired, Hassabis posted: “perhaps one day even solve all disease.”8 By the time the Isomorphic Labs Series B closed thirteen months later, “solve all disease” was Isomorphic’s formal mission statement, repeated verbatim in the press release, in the Thrive Capital quote, and in Hassabis’s own statement on the round.9
In the 60 Minutes quote he establishes that proteins matter. (They do.) He establishes that mapping them was an unsolved problem. (It was.) He establishes that AlphaFold solved it. (More or less.) And then — in a single transition so smooth most viewers will not catch it — he implies that AlphaFold can now collapse the timeline for designing a drug from ten years to weeks. This isn't a lie and it isn't a confabulation. It's a smooth movement from protein structure prediction (solved) to drug discovery timelines (not), whose connective tissue does not survive scrutiny. That implication is central to the story behind a $2.1 billion fundraise. It is also incomplete in several important ways, but they all require more biology than a 60 Minutes segment will ever air.
Solve, Not Cure
On April 9, 2025, on Reid Hoffman’s podcast, Hassabis said: “solving all disease.” Eleven days later, on 60 Minutes, he used “cure” — saying we could cure all disease within the next decade with the help of AI. On April 22, two days after the segment aired, his X post used “solve.” Across two weeks in April 2025, on the most-visible interviews of his career, he used both words interchangeably.
By 2026, only one word remains. At Davos in January 2026, Fortune’s editor-in-chief Alyson Shontell interviewed him for the magazine’s 500: Titans and Disruptors of Industry series; the episode was published in February: “You’re doing two huge jobs at once… And you have some lofty goals, like you say, we’re going to solve all disease. You don’t say cure, you say solve.”10 He did not correct her. He explained the choice. The cure-to-solve migration is the kind of careful rhetorical shift that happens between an interview and a $2.1 billion Series B.
Fortune’s Allie Garfinkle, who spent three days inside Isomorphic Labs for a January 2026 profile, summarized Hassabis’s position: “The idea of ‘solving disease’ is broader and more practical than eliminating illness once and for all. There’s a reason that he doesn’t say ‘cure.’ While you can’t promise no one will ever get sick again, he says, you can develop a systematic, repeatable, and scalable process — powered by advanced AI and technology platforms — for discovering, designing, and optimizing drugs or treatments as needs arise.”11
The power of “solve all disease” is that it lets the speaker move between two meanings: a literal public meaning that sounds like the end of illness, and a technical fallback meaning that means a better, repeatable drug-discovery process. Those are not the same claim. The first is a transformation; the second is a platform. The first raises $2.1 billion from sovereign wealth funds. The second describes — fairly accurately — what every major pharmaceutical company has been trying to build for forty years, and what dozens of well-funded biotechs are currently trying to build with AI tools. A competitive bet inside a crowded industry is not a redefinition of medicine. The scientifically defensible version is what he tells Fortune. The CBS version is the one most useful for fundraising.
Isomorphic’s then-president, Colin Murdoch, told CNBC in April 2025: “When AlphaFold came along, I think Demis and I looked at this and we thought, ‘Well, what could we do with AlphaFold?’”12 The goal of solving all disease, in Murdoch’s telling, came after the tool. It was not a problem statement that found the right tool. It was a tool that found the most ambitious problem statement available. The phrase is doing rhetorical work that the platform is not yet doing scientifically.
Seven days after the Series B closed, on the main stage of Google I/O 2026, Hassabis used the keynote to repeat the mission line and raise it. AI will, he said, “reimagine the drug discovery process with the goal of one day solving all disease,” and we are “standing in the foothills of the singularity.” The phrase did not retreat after the round closed. It escalated.
The Mona Lisa Depends on the Paint
Derek Lowe is a medicinal chemist. He has worked in drug discovery for major pharmaceutical companies since the late 1980s. He is currently a director in Chemical Biology and Therapeutics at Novartis. For more than twenty years he has written the blog In the Pipeline at Science, which is the closest thing the drug industry has to a public conscience. Lowe is not an AI skeptic. He has called himself a “short-term pessimist but a long-term optimist” about AI in drug discovery, and he uses ML tools in his own work.
On April 21, 2025 — the day after the 60 Minutes segment — Lowe published a piece called “The End of Disease.”13 It is one of the most patient, careful, exasperated pieces of writing about AI hype out there. Lowe is tired of having to be the guy who explains this, and he says so. He cites a phrase he coined in 2007: the Andy Grove Fallacy, named for Intel’s late CEO, who once asked why drug discovery couldn’t just “get with the modern age” and move as fast as software development. The fallacy describes a recurring belief among smart people from outside biology: that drug discovery is slow because the people doing it are slow.
Lowe’s position is that this belief is wrong, always was, and will be wrong even after the next generation of AI tools is built. The rate-limiting step in drug discovery is not the speed at which we generate hypotheses. It is that biology is more complicated than our models of biology, and a faster model of inadequate biology gets you a faster wrong answer.
He goes after the protein-centricity directly. Hassabis’s claim that “everything in your body depends on proteins,” Lowe writes, “is such a reductionist take that it really doesn’t get you anywhere useful. It’s like saying that everything in the Mona Lisa depends on the paint.”14
The deeper layer is the analogy itself. The Mona Lisa is paint, in the trivial sense. But knowing the chemical composition of every pigment Da Vinci used will not tell you why people line up at the Louvre. Lowe’s point is that biology is the same. You can know the structure of every protein and not know what makes a person sick. You can predict the binding pocket and not know whether modulating the protein in that pocket will help, hurt, or do nothing. This is not a theoretical objection. It is the dominant empirical finding of half a century of pharmacology.
Drug discovery is a stack of five problems: pick a target whose modulation will treat the disease in humans; design a molecule that binds it specifically; get the molecule to the right tissue at the right concentration; do not kill the patient; prove it works in a clinical trial. AlphaFold meaningfully helps with one of these — molecule design. Genetic target validation, what the RGC and its peers do, helps with another half. The other three are the ones that kill 90% of drugs, and they kill them for reasons that have little to do with protein structure or human genetics. Industry-standard numbers: about 90% of drugs that enter Phase I do not make it to approval; roughly 70% of programs fail in Phase II; drugs that survive take 10 to 15 years and cost on the order of $1 to $2 billion per approved molecule.15
Data, Not Methods
In October 2018, I published a short piece in the Regeneron Stories series titled “When the Student is a Computer: Teaching Machines to do New Tricks.” The thesis:
“Clear away the hype, and most AI/ML methods are narrowly focused — just fancy names for accurate, lightning-fast computational pattern recognition. In fact, some of these methods are not even new or innovative, just old-fashioned data analysis with improved branding. Unfortunately, the fancy name evoking the alluring idea of computers thinking like humans obscures the fact that the methods themselves aren’t nearly as important as the data from which they learn.”16
Five months earlier, Daphne Koller had launched insitro — a portmanteau of in silico and in vitro that names their thesis: wet-lab data generation and ML have to be done together. Koller’s credentials run parallel to Hassabis’s — MacArthur fellow, an early machine-learning faculty hire at Stanford for eighteen years, and the Rajeev Motwani Professor of Computer Science there, Coursera co-founder, elected to the National Academy of Engineering and later the National Academy of Sciences, and the former Chief Computing Officer of Calico. Her launch post is the road not taken:
“There is a lot of hype today around machine learning, with hyperbolic promises that it will magically solve all of humankind’s problems… We at insitro don’t expect ML to be the solution to all of the problems in drug development, nor to be the magic bullet that helps find a treatment for every disease.”17
Eight years on, insitro has raised approximately $643 million in equity across three rounds plus roughly $150 million in collaboration revenue from Gilead (NASH), Bristol Myers Squibb (ALS), and Eli Lilly, had three targets nominated by BMS in its ALS program, and produced no FDA-approved drugs. Its current marketing has drifted some — phrases like “Virtual Human™” and “industrialized, repeatable process” do real work in the press releases — but none of those phrases is solve all disease. Koller’s incentive architecture, like the RGC’s, allowed her to make a smaller claim, and she made it.
We do not have enough high-quality, human, longitudinal, multi-modal data about disease to support the kind of system-level inference “solving all disease” would require. We have a lot of static data, a lot of single-tissue data, wildly uneven coverage across populations. The places where we have rich human data — a handful of biobanks, a handful of integrated health systems — produced a generation of genetic discoveries I am proud to have been part of, and that work has so far translated into a small number of drugs and a much larger number of insights still waiting on translation. The translation step is the hard part. Hassabis knows this. He works in this field. The fact that he says “solve all disease” anyway is not a failure of his understanding. It is a feature of his function in the discourse.
The Receipts
Isomorphic Labs closed a $2.1 billion Series B on May 12, 2026. Thrive Capital led. Alphabet and GV continued. New money came from MGX (Abu Dhabi sovereign wealth), Temasek (Singapore sovereign wealth), CapitalG, and the UK Sovereign AI Fund. The combined geography of the round — American venture, Alphabet, Gulf state, Singaporean state, British state — is a story in itself: sovereign capital is now part of the AI drug discovery pitch, and the British government is investing directly in the company led by the man who is also the British government’s AI advisor.9
Three things are missing from the announcement, and the things that are missing are the things that matter.
No compound names. No disease targets. No specific Phase I timeline beyond “by the end of 2026.” Isomorphic has been operating for five years. It announced partnerships with Novartis and Eli Lilly in January 2024 and added Johnson & Johnson subsequently.18 None of the partnerships has produced a publicly named candidate.
Zero patients dosed. As of this writing, no Isomorphic-designed molecule has entered a human clinical trial. The original public commitment was to have a drug in clinical trials by the end of 2025. That date passed. At Davos in January 2026, Hassabis revised the target to end of 2026 and told Bloomberg he had misspoken — he had been referring to pre-clinical trials, which the company has begun.19 Pre-clinical and clinical are categorically different things. Pre-clinical work is animal models. Clinical work is humans.
No public data. AlphaFold 3, the model underneath IsoDDE, was published in Nature in May 2024, and the publication itself drew significant scientific criticism for being released without open-source code or weights — unusual for the DeepMind team that made AlphaFold 2 a model of open science. The code was eventually released in November 2024 after pressure from the research community. The reasons given for the initial restrictions were commercial. Isomorphic was the commercial reason.
The slipped deadline is not damning. Drug development deadlines slip all the time; it is the modal outcome. What is worth noticing is the combination: a public promise that did not happen, a quiet revision reclassifying the original commitment as a misstatement, an opaque pipeline, and a $2.1 billion raise made on the basis of a mission statement that cannot be falsified within the lifetime of the people making the investment. None of those properties are unusual. They are the established pattern. We have been here before. Many times.
Fifty-Five Years of the Same Promise
The first “we will end disease” claim I can find with full presidential backing is from December 23, 1971. Richard Nixon, beleaguered, looking for a legacy issue, signed the National Cancer Act and pledged $1.6 billion over three years. At the signing ceremony, politicians compared the effort to the moon landing; the bill’s framers had set a public target of curing the most common cancers within five years, by the nation’s bicentennial on July 4, 1976.20 Five years. Cancer would be solved by July 4, 1976.
It was not. By the bicentennial Nixon had bigger personal issues than cancer to deal with, and cancer was still killing approximately 370,000 Americans a year. By 1975, network news was already calling the effort a failure. “There is little tangible to show for the money that has been spent,” NBC reported in August 1975, four years in. By the war’s fiftieth anniversary, oncologist Otis Brawley reflected: “I wish that we had tried to convince people this was an investment in research which was long-term. I wish some people had not assumed we would make tremendous insights very quickly. Unfortunately, human nature is such that if people had realized that this was a 40, 50, 60 year commitment they wouldn’t have gone for it.”21 The honest sentence is the one that does not raise money.
It is easy to hate on Nixon, but the cure-all pitch did not start with him. It is the oldest pitch we have. Gilgamesh sought a plant at the bottom of the sea that promised rejuvenation four thousand years ago; a serpent stole it before he could use it. Chinese emperors died chasing the elixir of immortality. Newton spent more of his adult life on alchemy than on physics, looking in part for the philosopher’s stone that, among its other properties, would cure all disease. By 1843 the first patent-medicine almanacs were circulating in America, and by the 1893 Chicago World’s Fair the Rattlesnake King was selling bottled snake oil from the main stage.
Then came antibiotics, the polio vaccine, smallpox eradication — the framework worked, dramatically, for a specific class of disease, and the realistic memory of that success is what every subsequent overclaim has been parasitic on. The figure who carried the pitch into modern American political language was Mary Lasker, the medical philanthropist who, with her husband Albert, founded the Lasker Awards. Her 1969 full-page newspaper ads — “Mr. Nixon: You Can Cure Cancer” — were the proximate cause of the Cancer Act. Everything since is operating inside the same four-thousand-year groove, like the Scala Sancta in Rome.
What actually happened over the next fifty years: cancer mortality in the United States peaked in 1991 and has fallen about 33% since then. The decline is real and represents millions of lives saved. It is also overwhelmingly attributable to smoking reduction (which gave us most of the lung-cancer numbers), to screening and early detection (mammography, colonoscopy, Pap smears), and to targeted therapies in specific cancers (imatinib for CML, trastuzumab for HER2-positive breast cancer, immune checkpoint inhibitors in melanoma and a few other indications). The decline did not come from a cure. It came from a lot of slow, expensive, incremental, mostly successful work that does not fit on a magazine cover.
This sequence of subsequent promises reminds me of the ‘Lucy and the football’ running gag in Peanuts:
July 17, 1990. George H.W. Bush signs Presidential Proclamation 6158 declaring the 1990s the Decade of the Brain. The proclamation lists Alzheimer’s, Parkinson’s, stroke, schizophrenia, autism, spinal cord injuries, depressive disorders, and epilepsy. Real scientific advances follow — brain imaging, computational neuroscience, the genetics of Huntington’s and Alzheimer’s.22 Thirty-five years on, in 2026, the number of these diseases for which we have a cure remains zero.
June 26, 2000. Bill Clinton, Tony Blair, Francis Collins, and Craig Venter announce the completion of the Human Genome Project’s first draft in the East Room of the White House. The line I want to make sure you hear, from Clinton: “It is now conceivable that our children’s children will know the term ‘cancer’ only as a constellation of stars.” Venter, at the same event, projected the potential for zero cancer deaths within our lifetimes.23 Twenty-five years on: cancer is the second-leading cause of death in the United States. The term is not only a constellation of stars.
February 16, 2011. IBM’s Watson beats Ken Jennings on Jeopardy! Memorial Sloan Kettering, MD Anderson, and others line up. In April 2015, IBM CEO Ginni Rometty goes on Charlie Rose and calls healthcare Watson’s “moonshot.” By February 2017, MD Anderson cancels its $62 million partnership after roughly four years and zero patients.24
September 2013. Larry Page announces Calico, a Google-funded life sciences company; Time’s cover asks, in huge letters: “Can Google Solve Death?” Twelve years and more than $3 billion in combined Alphabet and AbbVie commitments later, Calico’s most prominent candidate failed in an ALS trial in January 2025.25 In November 2025, AbbVie ended its 11-year collaboration with Calico, laying off roughly 100 scientists. Death has not been solved.
January 2014. Regeneron launches the Regeneron Genetics Center as a wholly-owned subsidiary. I am one of the four people named in the founding press release with a direct RGC role. The stated objective is the one I gave you in the opening: define disease targets, improve the drug development process. Twelve years later: approaching three million exomes sequenced, with a goal of ten million on the horizon; peer-reviewed work in the major genetics journals; tools and data used across the field; more than forty novel drug targets and therapeutic programs delivered into Regeneron’s pipeline; one FDA-approved drug and multiple clinical-stage medicines in development. Depending on the comparison, that is either a respectable yield or a sobering one. It is not “all disease.”
December 1, 2015. Mark Zuckerberg and Priscilla Chan announce the Chan Zuckerberg Initiative, with a backdrop the next year asking, “Can we cure all diseases in our children’s lifetime?” $3 billion is pledged.26 By 2023, Chan described CZI as a “teeny-tiny” player compared to NIH. In November 2025, CZI relaunched with renewed AI-driven ambition and — explicitly — “meeting the biomedical goal even earlier” thanks to AI.27 The goalpost moves with the technology. The phrase stays.
January 2016. Vice President Joe Biden launches the original Cancer Moonshot, $1.8 billion authorized through the 21st Century Cures Act. February 2022. President Biden relaunches the Moonshot with the first numerically falsifiable cancer goal ever set by a U.S. administration: cut the age-adjusted cancer death rate by 50% over 25 years, by 2047. April 2023. NIH publishes an analysis: if current trends continue, the death rate will fall by 44%, not 50%, by 2047. To hit the target, the annual rate of decline would need to accelerate from 2.3% to 2.7%.28 The Moonshot is the most honest of these efforts because it set a quantified, falsifiable target. We are, as of this writing, on track to miss it.
2018 through 2024. The AI drug discovery wave. BenevolentAI, Recursion, Exscientia, Insilico. Recursion’s 2021 IPO raised $436 million on the AI-speeds-discovery promise. By August 2024, Recursion and Exscientia had lost roughly 80% of their peak market value and merged for cash runway.29 BenevolentAI delisted from Euronext Amsterdam in March 2025 and was absorbed by Osaka Holdings. FDA approvals of drugs designed by AI-first companies, as of this writing: zero. Two recent qualifiers bound the pattern without changing it — Takeda’s December 2025 Phase 3 win for zasocitinib30 and Insilico’s June 2025 Phase 2a publication in Nature Medicine for rentosertib, the first peer-reviewed Phase 2a where both target and molecule originated in generative AI.31 Both are real. Neither is “all disease.”
November 2021. Demis Hassabis and Colin Murdoch launch Isomorphic Labs out of DeepMind. The mission: solve all disease. Series A in March 2025: $600 million, led by Thrive Capital. Series B in May 2026: $2.1 billion. Patients dosed: zero. Target Phase I entry: end of 2026, slipped from end of 2025.
May 27, 2026. This essay is published at Tears in Rain, and there is still no tangible evidence of an AI-enabled transformation of medicine that puts us anywhere near a trajectory that could plausibly ‘solve all disease’ or ‘cure cancer’.
The Pattern
Fifty-five years. Built on an ancient concept. Four presidential initiatives, two billion-dollar philanthropies, and a handful of corporate research programs (including the RGC). Tens of billions of dollars in philanthropy. Hundreds of billions of dollars in public investment. Trillions in industry R&D. A staggering amount of real science. Some genuinely better treatments for some specific diseases. No “solve all disease.” Not even close.
The pattern does not show that the people making these claims were lying, that the work they funded was useless, or that the next promise will fail the same way the others did. Each effort produced real science, and the cumulative effect is the medicine I and everyone I love can access today. Childhood leukemia is now usually survivable. HIV is a manageable chronic condition. There are gene therapies for sickle cell disease and spinal muscular atrophy, and checkpoint inhibitors for cancers that were uniformly fatal twenty years ago. The work was real.
What the pattern shows is that the phrase “solve all disease” — or whatever its cognate is in any particular decade (“cure cancer,” “end disease,” “solve death,” “revolutionize the diagnosis, prevention, and treatment of most, if not all, human diseases”) — is doing specific work in our discourse, and that work has very little to do with predicting what will actually happen. The phrase is a fundraising primitive, a political mobilization tool, a magazine cover, a competitive moat for an institution that needs to look like the entity capable of reaching the most ambitious thing. It has never been a description of a tractable engineering project with a defined timeline or falsifiable deliverable.
Every decade or so, a new technology arrives that is powerful enough to make the old phrase look plausible again, and the phrase gets dusted off and put back into circulation. Recombinant DNA in the 1970s. Polymerase chain reaction in the 1980s. The Human Genome Project in the 1990s. RNAi and microarrays in the 2000s. CRISPR and big-data EHRs in the 2010s. AI in the 2020s. Each technology is real. Each technology contributes something. None of them, in isolation or in combination, has been the technology that ends disease, because disease is not a thing that ends. Disease is what bodies do.
The RGC did not promise to cure all disease. It also did not exist outside the rhetorical ecosystem I have just described. We benefited from the rising tide of capital and attention flowing into “AI for drug discovery” as a category. The 2018 piece I quoted earlier was published in Regeneron Stories, the company’s corporate communications channel, and uses phrases like “scientific revolution in genomics” and “turbo-charge AI/ML strategy.” Those are marketing phrases. They are softer than “solve all disease” — considerably so — but they are not the dry technical language of a peer-reviewed paper. I wrote them. I would write them less breathlessly today; I would also concede that at the time, writing them less breathlessly would have been institutionally costly. Anyone inside any of the ventures named in this essay is operating inside the same incentive system. The question is not who is virtuous. The question is what the incentive system rewards.
The RGC’s framing was modest because the institutional setup did not require, and would have been undermined by, the larger version of the phrase. We were a wholly-owned subsidiary of a profitable biotech. Our funding came from existing drug revenue, not from venture rounds whose next valuation depended on the story getting bigger. Our success metrics were countable: papers, validated targets, programs informed, drugs filed. Specific metrics produce specific rhetoric. Moving metrics produce moving rhetoric. This is structural, not moral. Had the RGC been venture-backed in 2014 — had each round depended on a story bigger than the last — I do not know we would have stayed inside the circumspect, technical version. I would like to think we would have, but sometimes you do what you must to raise the money when you need it.
What I Got Wrong
The structural pull I’ve described doesn’t stop at institutions with more modest goals. Three places where, with hindsight, I think that even our temperate framing at the RGC stretched what the science could deliver.
Genetic validation moves the probability, not the outcome. The ratio of genetic discoveries published to drugs approved at the RGC is on the order of a hundred to one, depending on how you attribute. That is comparable to industry baselines. Genetic validation moves the marginal probability of approval from extremely low to low; the base rate is still grim. I was, at various points in this work, overly optimistic about what genetic data alone could do. The data does what the data does, though we’ve seen that layering modalities beyond genetics and phenotypes provides meaningful additional discovery leverage.
We sequenced where the data and samples were, not necessarily where it was most needed or valuable. Geisinger MyCode is rural Pennsylvania, overwhelmingly of European ancestry. UK Biobank is overwhelmingly British white. The work we did with these cohorts was foundational and the publications carry; the population coverage is also a real limitation. I’m proud of the diversification work that did follow. The RGC has worked hard to build diversification into our data in later years — the Mexico City Prospective Study, the DiscoverMe collaboration in Durban, Together for CHANGE with Meharry Medical College — but I would argue for moving those efforts earlier in the program if I were starting over. The genetic data we have is not yet the genetic data we need.
We measured what we could control. The success metrics we chose — validated targets, programs informed, papers, drugs filed — were the metrics we could deliver against. A research center inside a publicly-traded biotech needs to deliver against measurable outcomes or it does not survive twelve years. Twelve years is roughly the duration required to do anything useful at all in this field, and a non-sustaining RGC contributes zero drugs. I stand by the calls we made. Institutional metrics and ‘solve all disease’ rhetoric do different work in the same ecosystem, and I worry that the work institutional metrics do is closer to that rhetoric than I want to admit.
Credit Where It Is Due
When asked in an interview with Varun Mayya at the Indian Institute of Science in February 2026 what would actually count as AGI, Hassabis proposed the Einstein Test: train an AI system on all human knowledge up through 1911, then see if it can independently derive general relativity. He added, plainly, that current systems — including his own — cannot come close.32
At least the Einstein Test is a falsifiable definition of AGI. It rules out benchmark hacking, viral apps, and vague statements like “outperforms humans at most economically valuable work.” It demands a leap into territory the training data has not been to. It demands the thing that distinguishes science from recombination. That said, if we had a model that could do everything any human can do except intuit quantum from the published science of 1911, it would be so miraculous that it would almost certainly be hailed as god-like, if not AGI. If human capability was reducible to a soundbite-able test, it wouldn’t be worth building an artificial version. Which is why AGI is the ideal MacGuffin for fundraising.
Days after proposing the Einstein Test, at the main stage of the India AI Impact Summit, Hassabis predicted AGI within five years. The same person says we can solve all disease in a decade. The same person raises $2.1 billion against the disease claim while admitting AGI is one or two breakthroughs away. Holding all of that simultaneously is the part that requires a pedant’s perspective.
The careful technical Hassabis is the credential. The expansive public Hassabis is the fundraise. They are both real. The gap between them deserves special attention.
Being Boring
I wrote last week about what the safety briefing binder says in a nuclear physics lab. Gravity is the weakest force in the universe, and the leading killer of graduate students. The strong nuclear force is the most powerful and the one almost no one is harmed by. The four forces, ranked by what they actually do to people, are the exact inverse of the four forces ranked by their textbook strength.33
Drug discovery has its own boring reality, and it works the same way. The cinematic version of the field — the version that funds the magazine covers and startups — is “cure cancer with AI.” The actual version, the one written down in pharmacovigilance databases and Phase II clinical trial reports and post-mortem case studies of failed programs, is much less impressive. Drugs fail because the target was wrong, or because the model species did not predict the human response, or because the liver did something the cell line did not, or because the patient population was more heterogeneous than the trial design assumed, or because the placebo arm did unexpectedly well, or because the regulator wanted endpoints the program had not been powered to detect, or because a competitor got there first and the standard of care changed under the program’s feet, or most likely for some reason that we can’t for the life of us figure out. The reasons drugs fail are mundane. The mundanity is the binder.
This matters because the cinematic version of AI drug discovery is now setting expectations across the entire policy and capital ecosystem. Sovereign wealth funds are investing in a $2.1 billion round against “solve all disease.” The UK government is structuring its biotech policy around the prospect that the Hassabis claim is approximately correct. Capital is allocating into the glossy magazine cover version. But the rate-limiting step in drug discovery has not visibly moved enough to support the advertised timeline, and not only because protein structure was never the gate.
The honest version of the next decade in AI-assisted drug discovery looks like this: a small number of well-validated targets in well-characterized disease areas will produce candidate molecules faster and somewhat better than they otherwise would have. Some of those molecules will become drugs, after the usual 10-to-15-year clinical timeline. The drugs that succeed will join a portfolio that has been expanding at roughly the same rate for sixty years — about fifty new molecular entities approved annually by the FDA, give or take a noisy decade. The total number of diseases for which we have effective treatments will continue to slowly grow. The burden of untreated disease will not go to zero, because biology is gravity, and gravity wins.
Twenty-five years ago Bill Clinton stood in the East Room and said our children’s children would know the word “cancer” as only a constellation of stars. Some of the children born after that statement have children of their own now, and cancer is the second-leading cause of death. The stars are still up there. The work is still here. We have made enormous progress, in mundane ways, against specific diseases, by patient effort sustained over decades. We will continue to make progress in the same way. Nothing in the current public evidence makes “solve all disease” plausible over the next several decades. We will not solve it because solving is the wrong verb. Bodies do not get solved, and curing a disease simply means everyone dies of something else.
Everything is gonna burn, we’ll all take turns, I’ll get mine too.
1. Andrew Pollack, “Aiming to Push Genomics Forward in New Study,” The New York Times, January 13, 2014.
2. The Regeneron Genetics Center–Geisinger DiscovEHR collaboration, announced January 2014. Initial commitment: 100,000 exomes linked to electronic health records. First report on 50,726 participants: Dewey et al., Science, 2016. The collaboration has since expanded to roughly 250,000 participants.
3. On pipeline outputs (more than forty novel drug targets and therapeutic programs across chronic liver disease, obesity, cancer, and neurodegenerative conditions): Aris Baras, Paul & Daisy Soros Fellowships profile, February 2026; Regeneron press release, January 13, 2025.
4. The ANGPTL3 loss-of-function genetic association was reported by RGC and collaborators in Dewey, Gusarova, Dunbar et al., NEJM 377 (2017): 211–221. FDA approval for homozygous familial hypercholesterolemia occurred on February 11, 2021; pediatric expansions followed in 2023 and 2025.
5. REGENIE: Mbatchou et al., “Computationally efficient whole-genome regression for quantitative and binary traits,” Nature Genetics 53 (2021): 1097–1103.
6. “Regeneron Launches New Human Genetics Initiative,” Regeneron press release, January 13, 2014.
7. Demis Hassabis, interview with Scott Pelley, 60 Minutes, CBS, aired April 20, 2025. Full transcript at cbsnews.com.
8. Demis Hassabis (@demishassabis), X, April 22, 2025: “perhaps one day even solve all disease.” Earlier remarks on the Possible podcast, April 9, 2025: “real crack at solving all disease.”
9. Isomorphic Labs Series B, $2.1 billion, May 12, 2026. Led by Thrive Capital with Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund. Hassabis is founder and CEO of Isomorphic Labs and serves as an AI advisor to the UK government.
10. Alyson Shontell, interview with Demis Hassabis, Fortune 500: Titans and Disruptors of Industry, recorded at Davos January 2026, published February 2026.
11. Allie Garfinkle, “The pioneer behind Google Gemini is tackling an even bigger challenge—using AI to ‘solve’ disease,” Fortune, January 22, 2026.
12. Colin Murdoch, interview with CNBC, April 9, 2025.
13. Derek Lowe, “The End of Disease,” In the Pipeline, April 21, 2025. The “Andy Grove Fallacy” was coined by Lowe in his 2007 piece “Andy Grove: Rich, Famous, Smart and Wrong,” In the Pipeline.
14. Lowe, “The End of Disease,” as above.
15. Phase-transition probabilities: Likelihood of Approval from Phase I is 6.7% (Citeline, 2014–2023); Phase II survival is roughly 28–33% across recent industry analyses (Norstella 2025; Wong, Siah & Lo, Biostatistics 20, 2019). Cost per approved molecule of $1–2 billion: DiMasi, Grabowski, Hansen, Journal of Health Economics 47 (2016): 20–33; Wouters, McKee, Luyten, JAMA 323 (2020): 844–853.
16. Jeffrey Reid, “When the Student is a Computer,” Regeneron Stories, October 29, 2018.
17. Daphne Koller, “insitro: Rethinking drug discovery using machine learning,” Medium, May 1, 2018. insitro has subsequently raised ~$643M in equity across three rounds, with collaboration revenue from partnerships including Gilead (NASH, 2020), Bristol Myers Squibb (ALS, 2020), and Eli Lilly.
18. Isomorphic Labs partnership announcements: Novartis and Eli Lilly, January 7, 2024; Johnson & Johnson added subsequently.
19. Hassabis interview at Davos, January 2026; “misspoke” clarification reported by Bloomberg, May 12, 2026.
20. National Cancer Act of 1971, signed December 23, 1971. Otis Brawley, “The 50 years’ war,” Cancer 127 (2021): 4534–4540.
21. Otis Brawley, quoted in Retro Report, “Nixon and the Long, Somewhat Successful, War on Cancer,” 2021.
22. Proclamation 6158, July 17, 1990. See also Goldstein, “Decade of the brain,” Western Journal of Medicine 161 (1994): 239–241.
23. Remarks by the President on the Completion of the First Survey of the Entire Human Genome, East Room, the White House, June 26, 2000. Clinton White House archive transcript includes Venter’s and Collins’s remarks.
24. Schmidt, “M.D. Anderson Breaks With IBM Watson,” JNCI 109 (2017): djx113. Strickland, “How IBM Watson Overpromised,” IEEE Spectrum, April 2, 2019. Watson Health sold to Francisco Partners for ~$1 billion in January 2022.
25. Calico Life Sciences and AbbVie press release, January 6, 2025: fosigotifator (ABBV-CLS-7262) failed its primary endpoint in the HEALEY ALS Platform Trial. AbbVie ended the 11-year collaboration with Calico in November 2025; combined Alphabet and AbbVie commitments to the collaboration reached approximately $3.5 billion by the 2021 extension. Nature, “Calico, Alphabet’s anti-ageing play, goes it alone,” December 12, 2025.
26. Chan Zuckerberg Initiative announced December 1, 2015. Chan, at UCSF September 21, 2016: “Can we cure all diseases in our children’s lifetime?”
27. Priscilla Chan, Fortune Impact Initiative interview, September 13, 2023. Jon Cohen, “AI drives dramatic expansion of Chan Zuckerberg Initiative’s funding to end all diseases,” Science, November 2025.
28. Biden White House, “Fact Sheet: President Biden Reignites Cancer Moonshot,” February 2, 2022. NIH news release April 17, 2023; Shiels et al., “Opportunities for Achieving the Cancer Moonshot Goal of a 50% Reduction in Cancer Mortality by 2047,” Cancer Discovery 13 (2023): 1084–1099.
29. Recursion–Exscientia merger announcement, August 8, 2024; closed November 20, 2024. BenevolentAI delisted from Euronext Amsterdam, March 12, 2025; merged with Osaka Holdings. As of May 2026, no AI-first drug discovery company has received FDA approval for an internally discovered molecule.
30. Takeda press release, December 2025: positive Phase 3 results for zasocitinib (TAK-279) in plaque psoriasis (LATITUDE trials). Takeda acquired the program from Nimbus Therapeutics in December 2022 for $4 billion upfront. The JH2-allosteric TYK2 mechanism originated in Bristol-Myers Squibb medicinal chemistry work (2016); optimized using Schrödinger’s physics-based FEP+.
31. Xu et al., “A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial,” Nature Medicine, June 3, 2025. 71 IPF patients; mean FVC change of +98.4 mL on rentosertib vs. –20.3 mL on placebo. Both target and molecule originated in Insilico Medicine’s generative AI platform.
32. Hassabis, interview with Varun Mayya, Indian Institute of Science, Bangalore, February 17, 2026. Hassabis predicted AGI “perhaps within the next five years” at the main stage of the same India AI Impact Summit, February 19, 2026 (Business Today).
33. See Danger and the Four Forces, Tears in Rain, May 19, 2026.
Jeff Reid writes Tears in Rain, and Claude helps.
Disclosure: I am a co-founder of the Regeneron Genetics Center. I’m retiring next month with associated equity. I have no consulting, advisory, or financial relationship with Isomorphic Labs, Alphabet, DeepMind, insitro, or any of the other companies discussed here. Opinions and ideas expressed here are mine.