OpenAI says an internal model resolved 100+ open math problems, and the advisory group it named can't slow it down

OpenAI announced on September 21 an independent Advisory Group on Mathematics and AI: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, Melanie Matchett Wood. The post states that an internal model which began training on August 28 has resolved more than 100 long-standing open problems, Navier-Stokes among them. (OpenAI)

Read the governance terms, because they're specific and they're published. The group takes no payment. It holds no decision-making power. Members may publicly criticize OpenAI. And it "will not be responsible for advising us on how to pace our internal progress on mathematics." Its assigned task is coordinating the release of results. Terence Tao's blog carried the group's own announcement the same day, noting it's hosted at the Institute for Advanced Study at agmai.org, and that it formed after OpenAI approached some members about an external advisory board and they chose to build an independent body instead. (Terence Tao)

So: nine of the most credentialed mathematicians alive, assembled around a capability claim, explicitly barred from touching the one lever anyone outside the company would want them to touch. You can read that as honest scoping, and there's a case for it. An unpaid external group with veto power over a lab's research schedule isn't a thing that would survive contact with reality. You can also read it as OpenAI buying the appearance of external oversight for the price of zero dollars and zero authority. Both readings fit the published facts.

The next day, Timnit Gebru and Emily Bender published a rebuttal in MIT Technology Review taking apart three of the summer's headline claims, the math results among them. They cite mathematicians who found the results less novel than reported, and NYU's Tristan Buckmaster, who published a statement alleging OpenAI took others' work. (MIT Technology Review)

I can't referee the Navier-Stokes claim and neither can you. What I can say is that "resolved more than 100 long-standing open problems" is doing a lot of load-bearing work with no list attached, and a named researcher is publicly alleging attribution problems. Those two facts sitting next to each other is the actual state of the evidence today.

Grant Sanderson made the sharper version of this argument on Tao's blog earlier this month: once a machine can produce proofs, proof stops working as a proxy for understanding. The advisory group's real job, whatever the charter says, is deciding what "we solved it" is going to mean from here.

Hold the claim loosely. Wait for the problem list.