They proved it on the same day

Two AI labs, working apart, say their models broke the same thirty-year-old belief about how hard computation is, and both found it on 16 October. The maths may hold. What it tells us about the labs is already clear.

On Monday evening, a little after half past nine London time, OpenAI put up a post with a title so mild you could scroll past it. "Working with academic communities to build scientific understanding," it said, which is the kind of phrase that usually sits above a grant announcement. Underneath it were three claims, any one of which would have been the biggest result of a working mathematician's life. Its models, OpenAI said, had proved Sogge's local smoothing conjecture in every dimension, and with it the Kakeya and Stein restriction conjectures. They had produced a rigorous existence proof for quantum chromodynamics, including pure Yang–Mills theory, which is part of one of the Clay Millennium Prize problems. And in the middle of the list, in a heading that reads like a newspaper splash, they had shown that "The Exponential Time Hypothesis is false."

Greg Brockman posted it with the line "AI starting to help us improve our understanding of the universe," and about an hour and a half later added a second post that did more to calm the timeline than any FAQ could have: "p still probably doesn't equal np."

Then, a few hours later, while most of Britain was asleep, Levent Alpöge, a mathematician at Anthropic, posted a single quotation and a GitHub link. "The core algorithm was found on October 16 by an internal Claude research model with 14M output tokens and no human input." It was a paper titled "A 2o(n) Time Algorithm for 3-SAT," credited simply to Claude, Anthropic, refuting the very same hypothesis. Same target, same broad shape of answer, and, according to both companies, the same day.

That coincidence is the story, and I think it is a bigger one than the proof.

What actually fell over

It is worth being plain about what the Exponential Time Hypothesis is, because half of X spent Monday night confusing it with Ethereum. (The best reply I saw was from an account calling itself crypto customer support, explaining that one ETH was going to independent peer review and the other "was refuted by the market without any review.")

The hypothesis comes from Russell Impagliazzo and Ramamohan Paturi, and it concerns 3-SAT, the textbook problem of deciding whether a long list of small logical constraints can all be satisfied at once. Brute force means trying every combination, which doubles in cost with every variable you add. Clever algorithms have shaved that down over the decades, but the hypothesis said you could never shave it down to nothing: some exponential cost, some fixed fraction of that doubling, would always remain. It was never proved. It was simply believed, firmly enough that it became load-bearing. OpenAI's post says it has been used in hundreds of papers to argue that problems in graph theory, computational biology, databases and geometry can't be solved faster than they already are. Anthropic's paper says Claude's own literature search turned up more than 800 papers resting on it or on its stronger sibling, the Strong Exponential Time Hypothesis.

Both labs now claim an algorithm that counts the satisfying assignments of a 3-SAT formula, not merely finds one, in what complexity theorists call subexponential time. Ryan Williams of MIT summarised OpenAI's version as running in roughly exp(n/(log n)1/2+o(1)) time. Sébastien Bubeck, who has spent twenty years watching this field, called it "maybe the biggest surprise I have seen" in that time.

Here is the part the headlines skipped. Neither algorithm is useful, and both papers say so with something close to pride. Anthropic's paper notes that for a formula with about five clauses per variable its bound doesn't beat plain brute force until the number of variables passes roughly ten to the ten-thousandth power, a number with no physical meaning at all, and it says outright that the result "says nothing about solving formulas in practice." Your SAT solver will not get faster this week. What changes is the floor beneath a great deal of theory. Hundreds of "this probably can't be done faster" arguments lose the assumption they were standing on, and someone now has to go back through them and see which can be rebuilt on other ground.

Two labs, one Friday

Now hold the dates side by side. OpenAI says its system found the disproof on 16 October and had it formalised in Lean, the proof-checking language mathematicians have increasingly trusted with this sort of thing, by the 17th. Anthropic says its internal Claude model found its core algorithm on 16 October too, with no human input, using fourteen million output tokens, and that the full Lean development runs to 1,050 modules, built on top of Mathlib and checked both by Lean's own kernel and by a separately written checker called nanoda. Anthropic's paper, deposited on the Hexagon preprint server by Jelani Nelson on Anthropic's behalf, says in its provenance note that "independently and concurrently, an internal model at OpenAI claims to have proven a similar result," and cites OpenAI's preprint in its reference list.

Some people online smell a stunt in that timing, and I understand the instinct. Two rival companies announcing the same breakthrough within hours of each other does not happen by accident very often. But I think the more uncomfortable reading is the more likely one, which is that it did not need to be arranged. If two independent systems, trained by two different teams on two different stacks, both reach into a thirty-year-old open problem and pull out the same kind of answer in the same week, that tells you something about where the frontier is. It is not a secret held by one lab. It is a tide, and it is coming in under everyone at roughly the same rate.

Calculus had Newton and Leibniz. Evolution had Darwin and Wallace. Simultaneous discovery is what science looks like when a field is ripe, when the tools and the questions have lined up so neatly that whoever is standing nearest picks the fruit. What is new is that the people standing nearest are not people. They are research models running inside two companies a few miles apart in San Francisco, and they arrived on the same Friday.

Noam Brown at OpenAI put the speed of it best. "Only 2.5 years ago," he wrote, "AI labs would highlight their frontier LLM performance on grade school math benchmarks." He meant it as a boast, and it is one. It is also a measure of how short the runway has become for everyone else, from journals to universities to the people whose careers are built on knowing a field better than anything else in the room.

Who checks the checkers

This is where the two announcements quietly part ways, and where I would want readers to slow down.

OpenAI did not just post a PDF. It recruited what it calls "shepherds": the original proposers of the hypothesis, Impagliazzo and Paturi, together with Williams, to organise an open examination of the ETH claim. The physics result goes to a group led by the Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, and that group has chosen to hold back OpenAI's raw output and write the result up in its own words first. The terms OpenAI published are unusual for a company valued, as Al Jazeera put it, at $852bn. No NDAs, no pay, no academic credit for the models, and freedom for the humans to say whatever they find. Williams says his group asked for all the raw data, PDFs and Lean source to be made public, and for OpenAI to fund a conference for the whole community affected by the work, and that OpenAI agreed to both. Boaz Barak, who works at OpenAI and teaches at Harvard, called it "the benefit of human understanding without gatekeeping."

Then, through Lance Fortnow, came a sentence from the three shepherds that I suspect will outlast the announcement itself. "Our involvement should not be taken as certification of the claim. A result of this significance requires careful, independent scrutiny."

Anthropic's paper is just as frank, in a different way. Its appendix says the Lean development "was written by AI models and reviewed by other instances of the same models," and that "at the time of writing no person is known to have checked it." I read that line three times. It is honest, and it is also a small landmark. We now have a major mathematical claim, formally verified by machine, written by a machine, reviewed by copies of that machine, which no human being is yet known to have read through.

Lean matters enormously here, and I don't want to undersell it. A formal proof that compiles is far stronger evidence than a confident PDF, and it is the reason serious people are taking both claims seriously rather than laughing them off. But formal verification proves that the stated theorem follows from the stated definitions. It does not prove that the definitions say what the headline says, and the gap between the two is exactly where careful humans earn their keep. That is why Williams and his colleagues insist they are examining, not endorsing, and why the arrival of a second, independent proof is, oddly, the most reassuring thing that happened on Monday. Two different machines making the same mistake about the same thirty-year-old problem would be a stranger story than either of them being right.

The part that should keep academics up

OpenAI's post is candid about its own discomfort. "Choosing initial shepherds necessarily shapes who gets involved first," it says, "and creates asymmetry of information." That is the whole problem in one sentence. When results like this arrive faster than any field can digest them, whoever a lab chooses to call first gets a head start on understanding the most important development in their discipline. OpenAI's answer is shepherds and openness. Anthropic's answer, at least this time, was to put everything on GitHub and a preprint server and let the field come to it. Neither is obviously right, and OpenAI admits as much. "We do not yet know which of these three approaches will be most useful."

Barak, in another post, made the point that matters most to the people on the receiving end: "it is up to each community to decide how best they want to address the question of AI's impact on their field." That is the right instinct, but it assumes the communities get time to decide. Physics has already chosen to slow down and write the Yang–Mills work up by hand before anything is released. Theoretical computer science has chosen to open the books immediately and argue in public. Mathematics, on the Kakeya result, has simply been handed the output and left to organise itself.

What strikes me, reading all of it, is how much of the labour has flipped. For centuries the hard part of research was finding the result, and explaining it was the reward. Now the finding happened on a Friday in mid-October, twice, for the price of some compute. The hard part, the slow and human part, is turning that raw output into something a person can understand, teach and build on. OpenAI's own phrase for its manuscripts is "raw research data." That is a remarkable thing for a company to call a disproof of the Exponential Time Hypothesis, and I think it is accurate.

Where this leaves us

The cautious version of this piece, the one I'd sign my name to, is this. Two frontier labs say their models independently refuted a central conjecture in computer science on the same day. Both proofs are formalised in Lean. Neither has yet been confirmed by independent human experts, and the people best placed to do that have said clearly that their involvement is not a stamp of approval. If the results hold, nothing practical changes tomorrow, but a great deal of theory has to be rebuilt, and part of a Millennium Prize problem is sitting with a group of physicists, waiting for its write-up.

The less cautious version is the one I keep coming back to. Monday was not really about 3-SAT. It was the day it became hard to pretend that frontier research capability belongs to any one company, and the day the question shifted from whether machines can find deep results to whether humans can keep up with understanding them. Brockman is probably right that P still doesn't equal NP. But the people who will eventually prove it, one way or the other, may well be reading a machine's working when they do.