Transcription

• Speaker #0

  Next time, the ones who hold the gate. That was last week's promise. This week, three labs proved something. Only one of them let the world check. On Tuesday, OpenAI announced it had solved one of the Millennium Prize problems. 88 hours, 10,000 agents running in parallel. The announcement came on a phone call, with reporters on the line. As of September 10th, nobody has seen that proof, not even the mathematician, who had been working on the problem for a year. Meanwhile, another lab was filing its own. 13 million lines, public, checkable by anyone. So which of those two results is actually a result? Welcome to the week of September 7th, 2026, Synth AI Cast. The AI that listens for you. Episode 24, the second of season two. This week is not about what AI can do. It is about what we can verify. And those are two completely different stories. First Signal. Tuesday, September 8th. OpenAI published an extraordinary claim. An internal model never released. had solved Navier-Stokes. Those are the equations that describe how fluids move. Air, water, blood, hurricanes. One of the seven Millennium Prize problems, each carrying a $1 million bounty since the year 2000. The open question was whether those equations can break, whether they can predict something physically impossible, a point in a fluid reaching infinite speed. mathematicians call that a blow-up. OpenAI says its model proved it can happen. The model ran around 10,000 sub-agents. 88 hours to reach the result. TechCrunch puts the whole week at 300 billion output tokens, roughly $22.5 million of compute. OpenAI says it will not claim the million-dollar prize. And then there is what was not done. As of September 10th, The proof is not published. It was described on a press call. Described, not filed. Two mathematicians had been working the same front for almost a year, Tristan Buckmaster at New York University and Levent Alpoge, a mathematician employed by Anthropic. They were using models from both houses. OpenAI and Anthropic. paid for by Buckmaster out of his own research funds. The night before the announcement, they published their own breakthrough on the Euler equations, with the formal version, the one a machine can check, line by line. Anyone can download it and verify it today. On September 7th, Buckmaster published a full-page statement. In it, he says OpenAI offered him sole credit. on one condition, that Alpogae's name come off the work. OpenAI denies it. The company says the effort started on September 1st, triggered by a rumor, and that it never saw any of their work. It adds that it cannot rule out that de-identified data from their use of its products may have improved its models. Buckmaster himself is careful. He repeats that he has not seen OpenAI's proof, and that he does not know what the model did, or how. Second signal. Four days earlier, another lab had done the exact opposite. On September 4th, Anthropic announced a complete formal proof of Fermat's last theorem. A claim scribbled around 1637, in the margin of a book, by a French lawyer doing maths in his spare time. It took 350 years to settle, proved by Andrew Wiles. In 1995, 129 pages and months of human checking, Claude rewrote that proof, in a language machines can verify. Eleven days of work. largely on its own. 13 million lines of code and 29,500 intermediate theorems along the way. It is the largest proof of its kind ever written. And it closes something else. There is a famous list of 100 formalization challenges, kept for about 20 years. This was the last one still standing. Nature reports the project was expected to take about 10 years for humans. One honest aside, and you may already be thinking it. Claude is the model that produces this episode, so we apply the same test as to everyone else, no exceptions. And here is what the test says. The proof was filed publicly, in the open. Kevin Buzzard, who was leading the community effort on that same theorem, compiled the code himself, and signed off on it. The first attempt had failed. It only worked with a tool built somewhere else. Prove to Me, developed by a group at Columbia University. Third Signal, Shorter, on September 8th. Google DeepMind published Alpha Genome Atlas, a one petabyte database. It predicts the effect of almost every possible change to a single letter of human DNA, taken one at a time, around 9 billion variants. Every prediction carries an impact score, validated against a clinical database, that matters in a very concrete way. When a genetic test comes back with a variant nobody recognizes, a clinician today often has nothing to say about it. A map like this is how that answer starts to exist. And here is the thing. You can download it. It is the only one of this week's three results. that an outside lab can test against its own patients without asking anyone's permission. So let us line the three of them up. 88 hours, described on a phone call. 11 days, filed on a public platform. One petabyte, downloadable tonight. The fastest of the three is the one nobody can check. and machine checking matters more when the author is a model. You cannot cross-examine a model about its own reasoning. You can only run its work through something that does not care who wrote it. This argument has been settled once before, in Paris. In 1864, two scientists were at war over where microorganisms come from. Louis Pasteur on one side, Pouchet on the other. Both of them said they had the evidence. The Academy of Sciences appointed a commission to decide. Pasteur ran his experiment again, in front of it. Pouchet withdrew before the test. History remembers the one who agreed to be watched. Last week we asked, who holds the gate? Here is this week's answer. The gate is not held by whoever announces the loudest. It is held by whoever agrees to be checked. And the number to keep? is 11 days, not 88 hours. 11 days is slower, but at the end of the 11 days, there is a file you can open, because the bottleneck in science has just moved. For centuries, the hard part was producing a proof. Today, the hard part is reading it. Three things to try this week. Five minutes. Look up the word lean. It is the language these proofs are written in. See what a machine-checkable proof actually looks like. It is far less intimidating than it sounds. 30 minutes. Take the last AI announcement that impressed you and look for one single thing, the link to something you can open, a repository, a paper, a data set. If you cannot find it, that is not a detail. One hour. At work, pick a number you publish often. And write down, next to it, how somebody else could reproduce it. That is exactly the move we have been talking about. We listen to the experts so you do not have to. And we check every fact. Then we hand you what matters. In around 10 minutes. Thank you for giving us these minutes every week. If this episode was useful, tell someone about it. Five stars, a review. It matters more than you would think. One last fact before we go. Three days after shipping its most powerful model, OpenAI's own chief scientist wrote it down in plain words, that nobody, himself included, has solved the safety of these systems. We know who agrees to be checked, but who writes the shared rule? Next time. The ones who write the rule. See you next week.