OpenAI's math solutions land, and research leaders inherit a funding question

OpenAI published a set of solutions to longstanding open problems in mathematics, and according to The Verge it landed like a bombshell, setting off a broad argument inside the field ^[1]. The Verge's London-based AI reporter Robert Hart, who interviewed leading mathematicians for the piece, describes a math world that is shell-shocked and in an existential crisis about what mathematicians are for going forward ^[2][3].

The timing matters more than the result. Hart dates the shift to the last six months to a year, in which models went from very terrible at this work to, in his words, seemingly genuinely quite good at a professional level ^[4]. He frames it as mathematics absorbing in that window what other fields have spent roughly five years struggling with ^[5], which is a compression of somewhere between five and ten times ^[6].

The capability profile is lopsided in a way that should make anyone cautious about extrapolating. Hart says the models remain truly terrible at arithmetic, at days of the week, and at time, citing Verge reporting by Elissa Welle that ChatGPT could not tell time and still cannot ^[7]. He checked the old benchmark and found the models can now count the R's in strawberry, though both he and Decoder's host said they suspect that particular case is hard-coded, which neither offered evidence for ^[8]. Hart's explanation for the gap is that advanced math is largely reasoning rather than calculation, and that academic papers often contain no numbers at all ^[9]. What the newer systems are good at, he says, is forging connections between areas and applying old methods in new ways ^[10].

That is the part with budget consequences. The Decoder framing puts it directly: what are academic grants and university programs for, if they exist to train new generations of humans to find and solve outstanding problems and frontier models simply answer those problems ^[11]. The awkward feature of most research funding is that it buys the search, and the search is exactly the artifact that a published solution set devalues fastest. Verifiable, hard, closed-form problems are the easiest thing for a lab to demonstrate on and the easiest thing for a funder to score. Problem selection, taste, and the unglamorous work of checking machine-produced arguments are none of those things, and they are what is left.

Two cautions before anyone restructures a doctoral program. The first is that a six-to-twelve-month capability move ^[4] is thin evidence for a five-year pipeline decision, and the podcast is an account of the reaction rather than of independent verification. The second is the possibility Decoder raises itself: that the attention on mathematics is largely a marketing exercise for frontier labs with no stake in the discipline's future ^[12]. Both readings are consistent with the same publication event.

Watch whether the labs release enough method detail for mathematicians to reproduce the results rather than only react to them, and who absorbs the verification cost if they do not. Watch also whether the claimed transfer of these skills to other domains materializes ^[13], because that is the assumption most likely to be quietly priced into hiring and grant decisions before it has been tested.