OpenAI has published 722 mathematical manuscripts produced by an unreleased internal frontier model, organised into 372 families of related results.[2] The collection covers open questions across mathematics and theoretical computer science, with some results accompanied by machine-checkable Lean formalisation.[2]
The scale is the startling part.[2] OpenAI says it tested the model on roughly 4,000 problems and that each reported result used, on average, compute equivalent to about three hours of ChatGPT Pro thinking.[1] Rather than one celebrated breakthrough requiring years of attention, the release presents hundreds of claimed advances generated through a repeatable evaluation process.[2]
That is the basis for The Rundown’s striking conclusion that “progress in math now scales with chips instead of humans”.[5] It is an interpretation rather than a settled scientific law.[5] The underlying change is that more compute can now be directed at more open problems through the repeatable process OpenAI describes.[1]
The release is not the same as 722 independently confirmed breakthroughs.[3] OpenAI says the manuscripts are at different stages of verification, not all have Lean formalisation, and some unformalised results could contain errors.[2] Scientific American reported that mathematicians may need months to determine which proofs are correct, novel and important, while an OpenAI spokesperson said many results were not yet understood by the company’s own mathematicians.[3]
Formal verification also makes mathematics unusually suited to this kind of acceleration.[1] Lean can check whether a proof follows logically from a formal statement.[2] Biology and medicine require empirical validation: useful hypotheses must survive laboratory experiments and, for treatments, further preclinical and clinical assessment.[4]
There are already signs that compute can accelerate the earlier stages of those fields.[4] A peer-reviewed Nature paper reported that Google’s Co-Scientist generated biomedical hypotheses that researchers tested in acute myeloid leukaemia cell lines, human liver organoids and antimicrobial-resistance research.[4] The authors described those experiments as preliminary validation, not a replacement for the preclinical and clinical work required before a treatment can be trusted.[4]
So the evidence does not support claiming that AI will instantly produce medical cures.[4] A more defensible projection is that systems could generate and prioritise more candidate explanations, targets and treatments for expert review and testing.[4] The breakthrough may be a new abundance of plausible answers, followed by a harder problem: deciding which ones deserve scarce human attention and real-world testing.[4]
[1] OpenAI: Sharing AI progress in mathematics
[2] OpenAI mathematics repository
[3] Scientific American: OpenAI unleashes hundreds more math results
[4] Nature: Accelerating scientific discovery with Co-Scientist
[5] The Rundown AI: OpenAI’s math avalanche continues
Source: OpenAI