I've tested Mistral Large 4 on our STEMBench - 20 problems for each of 6 domains - and it got a pretty nice score. It's definitely capable of doing original research across domains and is pretty even. Great win for the European AI!

What an incredible set of results! Quasi-Riemann Hypothesis and Hodge Conjecture for abelian varieties are probably my favorite ones, but there's also so many smaller gems. I really like Fontaine-Mazur modularity at prime 2 and modularity over imaginary quadratic fields: first proper AI exploration in the Langlands program. It will take mathematical community months to properly digest everything, but one thing is clear - there are no subdomains safe from AI now.

Another amazing result (towards P=NP?), and a very interesting release scheme: - Anthropic finds the proof together with a Lean formalization (so they know it's correct); - They contact the experts - the authors - offer compensation and access to Claude for a help with a write-up. Tbh, this looks like one of the good approaches to AI-academia collaboration.

We're testing our new benchmark. STEMBench is a set of 120 research-level problems from 6 domains - Biology, Chemistry, Physics, Astronomy, Earth science, Engineering - 20 problems per domain. Very cool to see that our initial calibration works well and the benchmark is far from saturated (max score is 4 per domain). Detailed testing with a dedicated website incoming. Interestingly, it seems like physics will be next after mathematics. Math is not included in this benchmark, we have ErdosBench for that. DM me if you want to see your model tested.

Big if true. No comment on the proof, as it's outside of my expertise and 140 pages long, but this looks like how things might look for the foreseeable future. Authors spent 2 months exploring 100+ approaches with various SOTA models: GPT-6 Astra, GPT-5.6 Sol, Claude Fable 5 and Fable 5.1. Monte Carlo of ideas.

I've run these new results through GPT-6 Astra + Opus 5.5 and got even more counterexamples in families that strengthen Landreman’s results. The automated paper is here: Figure by Opus 5.5 (+some calculations), final write-up by GPT-6 Astra. No Lean verification, just adversial LLM check-up.

"we ask them to stop testing advanced mathematical problems on proprietary models" That won't happen. As a former academic mathematician I can understand where this set of recommendations is coming from: it's basically trying to have this new entity "AI labs" fit into the structure of academia. But what will happen is academia will be made to fit into the AI age. One of the things missing from this text is "AI labs" are just a small part of the incoming wave. Already most of the AI generated results are coming from amateurs or former mathematicians who enjoy thinking about particular problems or just like testing LLMs on math. More useful would be a set of recommendations for human mathematicians in the age of AI.

Old Problems, New Proofs Very cool initiative described by Terence Tao on his blog. A math hackathon to find new proofs of known statements, better explanations and generally help deepen human understanding of various pieces of mathematics.

A Severe Misalignment of AI in Mathematics New blog post by Terry Tao and a new anti-AI declaration signed by 25 Fields medallists. I'm not sure what the intendent effect should be. The ban of AI use in mathematics and science in general because of "math community"? Fortunately math community is broader than just math academia, and thanks to AI, booming like never before. Also this "mass production of true/false statements" is such a wrong description. As if there's no proofs by AI that you can study, deepen your knowledge, build upon, engage. The "AI slop" is still better written than what 90% of mathematicians write. Mathematics doesn't belong to math academia. It is a tool to understand the world and help other domains with necessary tooling. AI boosts the understanding and engages wider community. Swan song of academia. Link: