The trophy cabinet
Things we would have framed, had we thought of them first
Seeded November 2026 Written November 2026
At the end of Programming was my art, I retreated into amateur physics. Contemporary AI could hardly scratch the surface of science, could it? Surely, this time, I had climbed far enough up the food chain.
I should have checked first.
While I was busy losing an art competition to Claude, the machines had been quietly filling a trophy cabinet. Not with pictures. With proofs, algorithms, protein structures and gold medals, the kind of work that gets a human a professorship, a prize named after somebody dead, or at the very least a framed certificate in the hallway that their parents show to every visitor.
So I took inventory. The rules are simple. No art, no music, no poetry. No "it's faster at arithmetic", because computers have been faster at arithmetic since the 1940s, and nobody ever threw a parade for a pocket calculator. Every entry has to pass one test: if a human had done this, would we have been ever so proud of them?
We would have built them statues. Several of them, and one would need a very large plinth.
For most of the twentieth century, chess was our favorite proxy for intellect. If a machine could play chess, the argument went, it could think. In 1997, Deep Blue beat Garry Kasparov, and we promptly moved the goalposts. Deep Blue didn't think, we decided. It merely searched two hundred million positions a second. Chess, it turned out, had never been a test of intellect after all.
Go was the real test. You can't brute-force Go. You need intuition, and experts confidently predicted that a machine would need another decade to beat a top professional.
In March 2016, AlphaGo beat Lee Sedol four games to one. In the second game, it played move 37, a shoulder hit on the fifth line that every professional watching thought was a mistake. It won the game. In 2019, Lee Sedol retired, explaining that even if he became the best player in the world, there was now something above him that couldn't be beaten.
After that, the games fell on something like a demolition schedule.
- 2017: AlphaZero was given nothing but the rules of chess, shogi and Go, played against itself, and within 24 hours was superhuman in all three. Thousands of years of human Go theory, reinvented and then improved upon in less time than it takes to get over jet lag.
- 2017 to 2019: Libratus beat professionals at heads-up no-limit poker, then Pluribus beat them at a six-player table. Bluffing, the most human skill there is, turned out to be a matter of getting the probabilities right.
- 2019: AlphaStar reached Grandmaster in StarCraft II, and OpenAI Five beat the reigning world champions at Dota 2.
- 2022: Cicero played Diplomacy, a game won by forming alliances and then betraying them, and finished in the top 10% of human players. Hardly anyone it played against suspected it was a machine. I have opinions about what it says about us that the first game where AI blended in perfectly was the one about lying to your friends.
- 2023: Swift beat three human world champions at first-person-view drone racing, in the physical world, at speeds where a human pilot's reflexes are the bottleneck.
Here's the part I like. After AlphaGo, human professionals got better. Researchers who analyzed millions of professional Go moves found that the quality of human decisions had been flat for decades, then jumped after 2016, as players started studying the machines and playing moves nobody had dared to play before. The machines didn't just beat us at our own game. They taught us how to play it.
Still, these were games. We invented them for fun, and we could always console ourselves that none of it was real.
Then the machines turned to things that were real.
Proteins are the machinery of life, and what a protein does depends on the shape it folds into. In his Nobel lecture in 1972, Christian Anfinsen argued that the sequence of amino acids alone should determine that shape. Working out how became one of the great open problems in biology. For decades, the reliable way to find a protein's shape was to crystallize it and shoot X-rays at it, a process that could swallow a PhD student's entire doctorate for a single structure.
In 2020, AlphaFold 2 predicted structures with accuracy comparable to the experiments themselves, and the organizers of the field's own competition declared the problem largely solved. Two years later, DeepMind released predicted structures for more than 200 million proteins, essentially every protein known to science. At the old rate of one PhD student per structure, that is many times more doctorates than every university in history has ever awarded.
In 2024, Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for it, alongside David Baker, who had spent decades on the same problem with a lot more human sweat.
Notice who got the Nobel, though. The humans who built the machine. That's going to become a theme.
Some algorithms are so fundamental that every computer on Earth runs them constantly, and so well studied that nobody seriously expects to improve them. Multiplying matrices is one. Sorting is another.
In 1969, Volker Strassen showed that two 2×2 matrices can be multiplied with seven multiplications instead of the obvious eight. Applied recursively, that gives 49 multiplications for 4×4 matrices, and for more than half a century, nobody beat 49 for general 4×4 matrices of complex numbers. Generations of very clever people tried.
In 2022, AlphaTensor found 47, though only in the binary arithmetic computers use for some special purposes. Then, in 2025, AlphaEvolve found 48 for complex numbers. One multiplication. It sounds like nothing. It's the first improvement in 56 years to a result every computer science student learns about.
AlphaEvolve was pointed at more than fifty open problems in mathematics. It matched the best known answer in about three quarters of them, and beat it in about one in five. Among other things, it raised the lower bound on the kissing number in eleven dimensions to 593. That's one more than before, in a question mathematicians have been arguing about since Newton and Gregory argued about it in 1694.
And there's AlphaDev, which in 2023 found faster routines for sorting very short lists. Not in theory, in the actual machine instructions. They were merged into LLVM's standard C++ library, which means that if you've run software compiled against it since, there's a fair chance some of your data was sorted by a routine no human wrote.
Any one of those would make a career. I've written a lot of performance-critical code in my time, and the best day I ever had looked nothing like "improved Strassen".
The International Mathematical Olympiad is two days, four and a half hours each, six problems, and the most gifted teenagers on the planet. A gold medal there follows a mathematician around for life. Terence Tao won one at thirteen, and people still bring it up.
In 2024, Google DeepMind's AlphaProof and AlphaGeometry 2 scored 28 out of 42, one point short of gold. They solved problem 6, the hardest, which only five of the 609 human contestants managed. There were caveats. Humans translated the problems into the formal language Lean first, and some solutions took up to three days.
In 2025, the caveats went away. Gemini Deep Think and an experimental OpenAI model both scored 35 out of 42, gold, reading the problems in plain English, writing proofs in plain English, and staying inside the four and a half hours.
In 2026, in Shanghai, the dots model from Xiaohongshu scored 42 out of 42, graded by the IMO's own coordinators. Seven of the 666 human contestants did the same. At least one other lab has claimed the same score.
Programmers didn't fare any better. In September 2025, at the ICPC World Finals, the world championship of university programming, an OpenAI system solved all twelve problems under the same five-hour limit as the students. The best human team solved eleven. Gemini solved ten, including one that no human team solved at all.
There's one human holdout I'm fond of. In July 2025, at the AtCoder World Tour Finals in heuristic programming, a ten-hour optimization marathon, the Polish programmer Przemysław Dębiak beat OpenAI's model into second place. Afterwards he said he was exhausted, and that humanity had prevailed, for now. He used to work at OpenAI. I'm choosing to read that as the parent beating the child at arm wrestling one last time.
The obvious objection is that olympiad problems have known answers. Somebody wrote them, somebody solved them, and somebody checked them. Winning a competition is not the same as doing research.
That objection lasted about six months.
Paul Erdős posed problems the way other people breathe, and attached cash prizes to many of them. Over a thousand are collected on erdosproblems.com, each marked open or solved. They range from things a strong graduate student could knock off in a weekend to things that have resisted the best minds for eighty years.
They made a natural target, and the first attempts were embarrassing. In October 2025, OpenAI researchers announced that GPT-5 had solved ten Erdős problems. It hadn't. It had found papers where humans had already solved them, which is useful, but which is called a literature search. The announcement was deleted. On Christmas morning 2025, somebody posted an AI-generated proof of another one, and within hours somebody else pointed out that Erdős himself had solved it in 1977.
Still lying like they're running for office, I thought. Then the lies started coming true.
On January 4th, 2026, Terence Tao confirmed that Erdős problem #728 had been solved largely autonomously by GPT-5.2 Pro, with the proof checked in Lean, and nobody could find a prior solution. Tao was careful to call these the lowest-hanging fruit: problems that fall to standard techniques nobody had gotten around to trying. He estimated that only one or two percent of open problems are that easy.
One or two percent of the open problems in mathematics is a lot of problems.
By May, a DeepMind agent had autonomously resolved nine of 353 open Erdős problems, at a few hundred dollars each. A graduate student costs more than that per week, and they want weekends.
Then, on May 20th, 2026, came one that doesn't fit in the low-hanging-fruit basket. In 1946, Erdős asked how many pairs among n points in a plane can be exactly one unit apart, and conjectured that the best known constructions were essentially optimal. Mathematicians broadly believed him for eighty years. An internal OpenAI model found a construction that beats them. Not by much, but by enough to prove him wrong. Tim Gowers, a Fields medalist, called it "a milestone in AI mathematics", and said he'd accept it into a top journal without hesitation.
And then, in September, things got silly.
The Navier–Stokes equations describe how fluids flow, from the coffee in my cup to the weather over Reykjavík. Since Jean Leray asked in 1934, nobody has known whether their solutions can blow up, developing infinities out of perfectly smooth beginnings in finite time. It's one of the seven Millennium Prize Problems, with a million dollars attached. Only one of the seven had ever been solved.
On September 8th, OpenAI announced that its agents had found a blow-up in exactly the version the prize requires, with a proof formally checked in Lean. About ten thousand agents worked for 88 hours, exchanging some five million messages, at a cost of several million dollars. Twelve hours earlier, Tristan Buckmaster and Levent Alpöge had announced closely related results for the Euler equations, also with AI help. Both teams built on analytic work by Diego Córdoba and Luis Martínez-Zoroa, done the old-fashioned way, with chalk.
If it holds up, it's the biggest result in the history of AI, and possibly the biggest result in mathematics this century. Nobody has finished checking it yet. A Lean proof is only as good as the statement it proves, and humans still have to confirm that the statement is really the Millennium problem and not a lookalike. There's also an argument about who did what first, with OpenAI admitting that rumors of the Buckmaster–Alpöge progress inspired its own effort. Charles Fefferman, who wrote the official statement of the problem, called Córdoba and Martínez-Zoroa the heroes of the story. Buckmaster thinks Martínez-Zoroa deserves a Fields Medal.
Notice who's getting the medals, again.
Mathematics gets the headlines because a proof is either right or wrong, and that makes for good scoreboards. The rest of science has been quieter about it, but no less thorough.
The European Centre for Medium-Range Weather Forecasts runs what was, for decades, the best weather forecast on the planet: physics simulated on a supercomputer, refined by generations of meteorologists. In 2024, DeepMind's GenCast beat its ensemble forecast on 97% of the targets it was tested on, up to fifteen days out, and produced each forecast in eight minutes on a single chip. Living in Iceland, I have strong feelings about weather forecasts, and I'll take whichever one is right.
In early 2025, José Penadés and his team at Imperial College London asked Google's AI co-scientist a question they had spent about a decade answering: how certain genetic elements jump between bacterial species. Their answer hadn't been published. In two days, the AI proposed the same hypothesis as its top suggestion. Penadés's first reaction was to email Google and ask whether it had access to his computer. It didn't.
In 2022, a reinforcement learning system learned to control the magnetic coils of a real tokamak in Switzerland, holding plasma in shapes that physicists had needed separate, hand-tuned controllers for. Fusion has been thirty years away for seventy years, and this didn't change that. It did turn one of the hardest control problems in physics into a training run.
In March 2026, a Google team used Gemini Deep Think with tree search to solve exactly the integral behind the gravitational radiation emitted by loops of cosmic string, where earlier attempts had only managed approximate answers for limiting cases. The agent found six different analytical methods, one of them borrowed from quantum field theory. An exact closed-form answer to a messy physics problem is the kind of thing that used to make a theorist's year.
Not every entry survives inspection. In 2023, DeepMind announced that GNoME had discovered 2.2 million new crystals, 380,000 of them stable. Materials scientists were less impressed. Many of the "new" materials turned out to be minor variations on known ones, and a crystal you've predicted but nobody has made isn't a discovery yet. That one stays out of the cabinet until somebody builds something with it.
Here's the inventory, with what a human would have gotten for each.
| Year | What the machine did | What a human would have gotten |
|---|---|---|
| 2016 | Beat Lee Sedol at Go, with a move no professional would have played | A place in every Go textbook, permanently |
| 2020 | Predicted protein structures as accurately as experiments | A Nobel Prize, which was in fact awarded, to the people who built it |
| 2023 | Wrote faster sorting routines for LLVM's C++ standard library | The quiet gratitude of every compiler engineer |
| 2024 | Out-forecast the world's best weather model on 97% of targets | The undying love of everyone who has ever planned a barbecue in Iceland |
| 2025 | Beat Strassen's 4×4 matrix multiplication for the first time in 56 years | A footnote in every algorithms course, forever |
| 2025 | Gold at the IMO, and all twelve problems at the ICPC World Finals | A lifetime of being introduced at parties as "the one who…" |
| 2026 | Solved Erdős problem #728, the first one largely on its own | A paper, and Terence Tao's attention |
| 2026 | Disproved Erdős's unit distance conjecture, after eighty years | A paper in a top journal, by a Fields medalist's own account |
| 2026 | A perfect score at the IMO, officially graded | A seat at the table with the seven humans who did the same |
| 2026 | A blow-up in the Navier–Stokes equations, claimed and under review | A million dollars, if it holds, and a statue |
Lay it out like that, and the pattern from Evolution has left the body shows up again. Games took twenty years. Olympiads took three. Open research problems went from embarrassing to an eighty-year-old conjecture in about seven months.
In fairness to the humans, the machines are still the machines. In October 2026, OpenAI dumped 722 mathematical manuscripts onto GitHub in a single day, claiming progress on everything from the Riemann hypothesis to the unique games conjecture. Its own researchers hadn't read all of them. Only about half came with machine-checked proofs, and some of those proofs turned out to prove something weaker than the paper claimed. Three manuscripts were withdrawn within a day. One mathematician said a paper was impossible to read without AI help, and a Fields medalist said his field of research had been annihilated. Whether that was a compliment, nobody has yet worked out.
Buckmaster, one of the people who actually solved something that September, publicly apologized for one of his own papers and called it AI slop. The first machine-generated proof he received was, by his account, the most horrendous he had ever read. When mathematicians built a test out of genuinely unpublished research problems, called First Proof, the best AI systems got six or seven out of ten roughly right, and Scientific American graded that a C-minus. Twenty-eight Fields medalists have signed a declaration criticizing AI companies for treating unsolved problems as benchmarks.
So, yes. They still lie like they're running for office. They're just winning more elections.
Every entry in the cabinet passed my test. If a human had done any one of them, we'd have been ever so proud. Some of them, a human would have dined out on for the rest of their life.
But pride is a human thing. AlphaGo didn't celebrate move 37. AlphaFold didn't fly to Stockholm. The ten thousand agents that spent 88 hours on Navier–Stokes didn't go out for drinks afterwards. Every trophy in the cabinet was collected by a human: the ones who built the machine, the ones who posed the problem, the ones who did the hard analytic groundwork with chalk, and the ones who sat down afterwards to check whether any of it was true.
In Programming was my art, I described the job that's left for the humans: intent, judgment, taste, skepticism, and the occasional "what the fuck did you do here?" Look at the cabinet again and you'll find exactly that job, done at the very top of the field. Tao, checking the proofs and saying what they're worth. Thomas Bloom, who runs the Erdős problems site, sorting the genuine solutions from the rediscovered ones. Fefferman, naming the heroes. Gowers, saying this one is real.
Human judgment is still the term in my formula that's made of meat. It's also, for now, the term deciding what goes in the cabinet.
Back in 2022, I watched Midjourney eat the illustrators and thought: fortunately, I'm a programmer. In May 2026, I lost that one too, and retreated into physics, on the assumption that science, at least, was out of reach.
I've now read the inventory. Science was never the high ground. It was the trophy cabinet, and the machines have been filling it for a decade while the rest of us argued about whether they could draw hands.
Fortunately, I'm only an amateur physicist.
It's hard to lose a job you don't have.