The Nobel Turn

“I’m flabbergasted. I had no idea this would happen.” — Geoffrey Hinton, in his Nobel telephone interview, October 8, 2024

The man who had spent the previous eighteen months telling anyone who would listen that his life’s work might end the human species was, on the morning of October 8, 2024, in a budget hotel in California with no internet, waiting to be put inside an MRI machine. Geoffrey Hinton had a scan booked. He had no plans beyond it. When the phone rang at an hour that felt like the middle of the night, a Swedish voice on the line told him that he and John Hopfield had won the Nobel Prize in Physics.

He thought, briefly, that it might be a prank. Then he understood that it was not, and that he was going to have to cancel the MRI. In the recorded telephone interview the committee posts within minutes of every announcement, he sounds genuinely disoriented, a seventy-six-year-old who cannot quite locate himself. He said he was flabbergasted. He said he had had no idea this would happen. He was, he noted, in a cheap hotel that had no connection to the outside world, which was an awkward circumstance in which to become a Nobel laureate.

The committee’s citation was for “foundational discoveries and inventions that enable machine learning with artificial neural networks.” Hopfield, at Princeton, had built a network in the early 1980s that could store a pattern and reconstruct it from a fragment, an associative memory grounded in the physics of spin glasses. Hinton had taken those ideas further, into the Boltzmann machine and then into the methods that let multi-layered networks learn for themselves. They were being honored, in other words, for the machinery underneath everything in this book. And the physicists noticed something strange about the choice. This was the Nobel Prize in Physics, the most conservative prize there is, going to a computer scientist and a man whose great work had migrated decades earlier out of physics and into a discipline physicists did not entirely recognize as their own. Some of them said so, with varying degrees of grace.

Hinton did not use his moment to bask. He used it, as he had used most of his recent moments, to warn. Asked on the call whether he had regrets, he gave the answer he always gave, that if he had not done the work someone else would have. Then he said the thing the committee had not asked about, that these systems might turn out to be more intelligent than the people who made them, and that nobody had a plan for what to do if they were. A man receiving the highest honor his field can confer, using the podium to caution the world against the field. The reporters covering the prize had to hold two facts in their heads at once, and most of them simply reported the warning and moved on, because the warning was the story they already knew how to tell.

The deeper story arrived the next morning, and almost no one outside structural biology understood at first how large it was.

On October 9, the day after the physics announcement, the Royal Swedish Academy gave the Nobel Prize in Chemistry to three men. One half went to David Baker, at the University of Washington, for computational protein design, for teaching computers to invent proteins that do not exist in nature. The other half went jointly to Demis Hassabis and John Jumper, of Google DeepMind, for protein structure prediction. Two Nobel Prizes in forty-eight hours, the physics prize for the foundations of neural networks and the chemistry prize for what neural networks had done to biology. The field had, in the space of two mornings, been handed something close to a coronation.

The chemistry prize mattered more than the headlines suggested, and to see why means going back fifty years, to a question that had defeated some of the best minds in science.

A protein is a chain. The cell builds it as a string of amino acids, one after another, like beads threaded in a sequence the genes specify. But a protein does not stay a string. The instant it is made, it folds, collapsing on itself into a dense, specific, three-dimensional knot, and that shape is the whole point. The shape is what lets a protein do its job, whether the job is carrying oxygen, cutting another molecule, holding a cell wall together, or fitting a virus to a lung cell like a key into a lock. The shape gives the function. The function gives a target, something a drug can be built to block or to fix. The shape is everything, and the shape is invisible.

For decades, the only way to see a protein’s shape was to crystallize it and fire X-rays through it, or to freeze it and image it with electron microscopes, painstaking experimental work that could take a graduate student a year or more to resolve a single structure and cost real money each time. Meanwhile the gene sequencers had gotten cheap and fast, so biology knew the bead-string of hundreds of millions of proteins and the folded shape of almost none of them. The gap between what could be read and what could be seen widened every year.

The dream was to skip the experiment entirely, to compute the shape directly from the sequence, to look at the string of beads and predict the knot. This was the protein-folding problem, and Cyrus Levinthal had pointed out as far back as 1969 why it looked impossible. A modest protein has so many physically possible ways to fold that if it tried them one at a time, sampling each in the time physics allows, it would take longer than the age of the universe to find the right one. Real proteins fold in microseconds. They were solving, instantly and reliably, a problem that brute computation could not touch. Nobody knew how, and nobody could copy them.

To keep the field honest, a structural biologist named John Moult had set up a tournament. Every two years since 1994, the Critical Assessment of Structure Prediction, CASP, ran a blind contest. Experimentalists who had just solved a protein’s structure but not yet published it would hand the sequences to the organizers. Teams around the world would submit their predicted shapes. Then the answers were revealed and the predictions scored against reality, on a scale where a hundred meant a perfect match and the high nineties meant accuracy good enough to call experimental. For twenty-five years the scores crept upward slowly, the way scores in a mature field do, a point or two a cycle. The problem was hard and getting marginally less hard, and the people working on it expected to spend the rest of their careers chipping at it.

In late 2020, at the fourteenth CASP, a team from DeepMind broke the contest.

DeepMind by then was the London lab Demis Hassabis had founded in 2010 and sold to Google in 2014, and it had spent the intervening years proving that game-playing systems could find moves no human had ever conceived. The same instinct that pointed it at Go had pointed it at folding. The lead on the folding effort was John Jumper, a physicist by training who had drifted into the messy biophysics of proteins and then into machine learning, a young researcher most of the field had never heard of. His system was called AlphaFold, and its second version did something CASP had never seen. Across the hardest targets it posted a median score of 92.4, on that scale where the high nineties means as good as the experiment. For a large share of the proteins it was handed, AlphaFold2’s prediction was indistinguishable, atom for atom, from what the crystallographers eventually found in the lab. It had not edged out the other teams. It had walked off the chart.

Moult, who had run the tournament for a quarter century and had watched the slow grind year after year, said what a scientist almost never lets himself say. In some sense, he told the assembled researchers, the problem is solved. A fifty-year grand challenge, the kind of problem people built whole careers around not solving, had fallen to a neural network, and the man who had spent his life measuring the field’s progress was telling the field that the race was over.

What made it a public event rather than a private triumph was what DeepMind did next. In July 2021 the lab published the method in Nature and, against every commercial instinct, released the code so that anyone could run it. The same week, by a coincidence the field still marvels at, David Baker’s lab at the University of Washington published its own system, RoseTTAFold, in Science, arriving at the same mountaintop by a different path. Baker had spent decades on Rosetta, software that designed proteins from scratch, and his group had absorbed the new deep-learning ideas and built a folding predictor of its own. The two halves of the eventual chemistry prize had announced themselves in the same seven days.

Then DeepMind did the thing that turned a result into infrastructure. Working with the European Bioinformatics Institute, it built the AlphaFold Protein Structure Database and gave it away. It started, in July 2021, with the human proteome and a handful of model organisms, around 350,000 structures. A year later, in July 2022, it expanded to roughly 200 million, nearly every protein cataloged in science, the folded shape of almost the entire known biological world, free to anyone with a browser. A graduate student who once would have spent a year crystallizing one protein could now look up its predicted structure in seconds, for nothing. By 2026 the database had been used by more than two million researchers in around 190 countries. It was, by a wide margin, the most useful thing artificial intelligence had yet done for people who were not in the business of building artificial intelligence.

The use cases arrived faster than the lab could count them. Groups working on malaria parasites, on the enzymes that digest plastic, on antibiotic resistance, on the proteins that knot together inside neurons in Parkinson’s disease, all reached for AlphaFold’s predictions as a starting point that had not existed the year before. The predictions were not always right, and they came with confidence scores precisely because they were not. What changed the economics of an entire discipline was that a guess good enough to design the next experiment around now arrived in seconds and for free. Work that began with a year of structure-solving could now begin with a year of structure-solving skipped.

What separated this from a single lucky strike was that it kept happening. AlphaFold was not the only thing DeepMind’s science group, run by Pushmeet Kohli, aimed the technique at, and the pattern repeated across domains that had nothing to do with biology. In November 2023 the lab published GraphCast, a weather model that produced a ten-day global forecast in under a minute, on a single machine, and beat the European Centre for Medium-Range Weather Forecasts, the gold standard, on roughly nine of every ten things it was scored on, including the tracks of cyclones, against a physics-based system that ran for hours on a supercomputer. The same month it published GNoME, which proposed 2.2 million new crystal structures and flagged some 380,000 as stable enough to possibly make, a haul that materials scientists compared, half seriously, to centuries of conventional discovery. The following July, two more DeepMind systems, AlphaProof and AlphaGeometry 2, sat the problems of the International Mathematical Olympiad, the hardest pre-college math competition in the world, and scored 28 points out of 42, solving four of the six problems, a single point short of a gold medal and squarely at silver-medal standard. AlphaProof cracked the hardest problem on the exam, one only a handful of the six hundred human competitors managed, by searching for proofs in a formal language a computer could check. The Fields Medalist Timothy Gowers verified the solutions himself and confirmed they were real.

A machine could now fold a protein, forecast a hurricane, propose a material, and prove a theorem at the level of the planet’s best teenage mathematicians. The technique that the prologue of this story watched four companies bid for in a Lake Tahoe hotel room, the discredited idea that had spent thirty years in exile, was solving the problems that human beings had filed under too hard.

It was also, at exactly the same moment, learning to do something darker, and the people who built the science tools knew it. The diffusion models that let a stranger paste a celebrity’s face into a fake nude, the systems that swallowed the catalogs of artists and newspapers without asking, the video generators that frightened Hollywood into pausing studio expansions, ran on the same family of ideas. They were siblings of the thing that mapped the proteins. The same mathematics that learned the statistical shape of a hundred million human faces well enough to forge a new one learned the statistical shape of two hundred million proteins well enough to predict the next. The transformer architecture under ChatGPT was the architecture under AlphaFold. There was no clean line between the counterfeit and the cure. There was one technology, pointed in two directions, and the direction was a human choice.

The cure did not even escape the argument that defined the counterfeit. In May 2024, DeepMind and its drug-discovery spinout published AlphaFold3 in Nature, a major leap that predicted proteins tangled with DNA, RNA, the small molecules that become drugs, and the ions that switch them on, the actual machinery of the cell in interaction rather than isolation. It was the version that mattered most for medicine. And this time, DeepMind did not release the code. AlphaFold3 came as a web server with usage limits, no downloadable model, because the commercial value now lived in a company, Isomorphic Labs, that DeepMind had spun out in 2021 to design drugs and that Hassabis also ran. The reproducibility of the science had collided with the economics of the asset.

The scientific community revolted. A letter to the editors of Nature, signed at first by ten researchers and within weeks by more than a thousand, argued that a paper without runnable code was not a reproducible result and should not have been published as one. The disclosure in the AlphaFold3 paper, the letter said, was appropriate for an announcement on a company website but fell short of the standard a scientific journal was supposed to enforce. They were, in lab coats, fighting the exact battle the artists and the open-source crowd were fighting over image and language models, the battle over whether the most powerful tools should be public goods or guarded assets. DeepMind, caught between its scientific reputation and its commercial ambition, blinked. In November 2024, six months after publication and a month after the Nobel, it released the AlphaFold3 code for academic and non-commercial use, with the weights available on request, reserving the commercial rights for Isomorphic. The open-versus-closed fight that ran through the whole of this era did not spare even the work the Nobel committee was about to crown.

Because the commercial side was real, and it was large. In January 2024, Isomorphic Labs signed research deals with two of the world’s biggest drugmakers, Eli Lilly and Novartis, worth as much as $1.7 billion and $1.2 billion respectively in upfront payments and milestones, the pharmaceutical industry betting serious money that predicting a protein’s shape translated into designing a drug that fit it. In early 2025 Isomorphic raised $600 million in its first outside round, led by Thrive Capital, to push toward its own AI-designed medicines. By that spring the company was saying publicly that it expected to put its first drug candidates into human trials before long, cancer drugs among the earliest, though as of the spring of 2026 it had not confirmed that a trial had actually begun. The promise was concrete and the timeline was honest about being a promise. Hassabis, the founder who had told the world in 2010 that he meant to solve intelligence and then use it to solve everything else, was now running the half of that sentence that came after the comma.

When the chemistry committee called him on October 9, Hassabis was at home in London. He took the news, by his own account, more calmly than Hinton had taken his, though he described the feeling as surreal. John Jumper, who had led the folding work and had turned thirty-nine that January, became the youngest chemistry laureate in seventy years, a researcher who four years earlier had been unknown outside a niche of biophysics and was now standing in the lineage of Marie Curie and Linus Pauling. David Baker, who had been designing proteins by computer since before deep learning was respectable, had lived to see his unfashionable field become the most fashionable in science.

The doubling was the whole point of the two mornings. The physics prize honored the people who had built the neural network, decades ago, when it was a discredited idea kept alive by a few stubborn believers. The chemistry prize honored the people who had aimed it at a problem nature had hidden for fifty years. The same machine had earned both, and the same machine, in the same months, was forging fake pop stars and ingesting the work of living artists and frightening newspapers into court. The committees in Stockholm had decorated one face of the technology. The other face was in the headlines every week. They were the same face, seen from two sides, and the year 2024 had insisted on showing both at once.

Hinton, for his part, did not change his mind because they gave him a medal. If anything the prize sharpened the contradiction he had been living in. He had built the thing. He believed the thing might be the most dangerous tool humanity had ever made. The Swedish Academy had now certified the first half of that sentence in gold, and he spent his prize-winner’s platform reciting the second half. The protein database that two million researchers were using and the deepfakes that flooded social media platforms both ran on his foundations, and he was proud of one and afraid of all of it, and he saw no contradiction in being both because there wasn’t one. The technology did what it was pointed at. The question that the Nobels had not answered, the question no committee could answer, was what it would be pointed at next.

Because while the world spent 2024 arguing over fakes, suing over copyrights, writing rules, and handing out prizes, the machines had quietly begun to do something they had not done before. They had started to stop and think before they answered. And a thing that could think before it answered was about to stop being a thing that answered at all, and start becoming a thing that acted.