OpenAI's unreleased model claims more than 370 math results at about three hours apiece

Take OpenAI's average at its word and the batch used roughly 1,110 hours of computing time, at about three hours for each of 370 results ^[4][12]. On the Australian Financial Review's count of 377, the total is about 1,131 hours ^[2][13]. OpenAI did not say what hardware ran those hours or what they cost, and neither report connects the results to the company's spending or revenue.

Suppose outside mathematicians confirm most of the results. OpenAI would then have produced solutions to open problems at a few hours of computing each ^[4]. Four of the seven Millennium Prize problems, each with a $1 million Clay Mathematics Institute award attached, would carry an OpenAI solution or partial result ^[5][6][14]. In a second outcome, a meaningful share prove partial, or turn out to rest on the kind of steering Tristan Buckmaster of New York University raised, and the count falls ^[16]. In the third, the results hold but the skill stays inside mathematics. The case for heavier AI capital spending then gains little from them.

Fortune reports that AI researchers think training on hard math may teach logical reasoning and persistence, and may help in math-heavy fields such as physics or economics ^[17]. It also reports that it is unclear how those skills carry into law or business strategy, fields that involve reasoning but have no objectively verifiable correct answer ^[18]. I think the third outcome is the most likely on this evidence. The batch shows capability where a proof can be checked and says little so far about work where an answer cannot be checked. The counter-thesis is the researchers' own transfer argument. It would be borne out if the same internal model produced results in a field without a checker and outside experts accepted them.

Those hours went into a model OpenAI has not released, or rather into a showcase ^[3]. Fortune notes that AI companies have been targeting math problems to display their models' capabilities ^[7]. For now the compute has bought a published demonstration from a system OpenAI keeps internal ^[1][3].

Mathematicians split on the method. Dan Litt, a University of Toronto mathematician, said several of the published solutions touched problems he is interested in ^[9]. "My view is that this is great for mathematics," he said ^[8]. Fortune reported that other mathematicians called the approach OpenAI and other AI companies have taken an assault on mathematics as a human academic discipline ^[11].