OpenAI's mathematics claims, a Fields Medalist's safety role and a student Olympiad show how AI is being tested in mathematics and who is joining that process.
Artificial Intelligence··Morning
From claim to verification
OpenAI says Astra solved or made major progress on 10 hard problems in mathematics and theoretical computer science. According to Donga Science, AI generated the arguments in the company's 250-page manuscript, human researchers organized the text, and the proofs were formalized with Lean software. That division of labor makes the distance between a model proposing a result and the mathematics community accepting it easier to see. Experts interviewed by Donga Science credited formal verification and the possibility of opening new research directions. They also identified a substantial omission in the presentation of successful cases: failed attempts, an overall success rate, and the precise contribution of the people who prepared the work were not disclosed. Hyunwoo Kwon of Brown University said mathematics papers normally take at least 6 months to review and recalled that errors were found in some results OpenAI had previously presented as solved. The immediate focus therefore shifts from an impressive solution count to the process by which independent specialists can test the proofs.[1]
A mathematician joins safety work
A new institutional move puts a mathematician inside this debate. The Decoder reports that newly awarded Fields Medalist Jacob Tsimerman is joining OpenAI to work on AI safety. Tsimerman, a number theorist at the University of Toronto, describes AI as a highly transformative technology and argues that much more effort should go into safety. He connects the contribution mathematicians can make to a concrete problem: present AI development is largely empirical, while strong guarantees about how the systems work remain scarce. A paper he co-authored last year examined scenarios in which AI could contribute to human extinction; The Decoder also notes that experts dispute this position. Tsimerman expects AI to surpass human performance in mathematics research soon. His move into a laboratory's safety work expands mathematics' role beyond supplying difficult questions for models. Formal guarantees, definitions of risk, and boundaries on system behavior also become places where mathematical expertise can shape how a model is built and evaluated.[3]
From classroom to research lab
At the other end of the expertise pipeline, students at the International Olympiad in Artificial Intelligence combined data analysis, programming, model training, optimization, and testing in one problem-solving process. VnExpress International reports that Vietnam's team won two gold, one silver, and four bronze medals. The individual programming contest drew 471 students in 108 teams representing countries and territories; Vietnam selected its eight competitors in two rounds from 1,188 students across 34 provinces and cities. The competition does not resolve the independent-verification problem described by Donga Science, and its participants are not doing the same work as Tsimerman in a safety role. The three developments nevertheless reveal a common institutional path. Students learn to build and test models, research laboratories use them to generate claims about difficult problems, and established mathematicians debate the guarantees and risk measures that should govern those systems. AI's place in mathematics therefore extends beyond the performance of one model. Educational selection, laboratory divisions of labor, formal tools, and lengthy specialist review now form different stages of the same developing ecosystem.[2], [1], [3]
