Shing-Tung Yau launches a math-and-AI lab alliance; its headline result, a preprint claiming positive curvature on all 28 smooth 7-spheres, credits AI only with exploring strategies and calculations

Why it matters

【Key points】

On October 11, Fields medalist Shing-Tung Yau launched the Yau Mathematics and AI Laboratory Alliance at Tsinghua University. Eight institutes, including Tsinghua's Yau Mathematical Sciences Center, BIMSA and SIMIS, joined, with over 1,000 researchers. Its four priorities are math-specific large models, AI-assisted formal proof and conjecture discovery, interpretable AI and cross-disciplinary computing. It also released more than ten AI-and-math results and directions, including the degenerate Arnold conjecture and 7-dimensional homotopy spheres.

【Verifiable】

Beijing Daily reported the member list, the priorities and the division of labor: in the 7-sphere work, AI helped explore proof strategies and simplify calculations, computer algebra checked identities and interval bounds and produced reproducible certificates, and mathematicians did the reasoning, proofs and writing. That work is arXiv preprint 2609.29426 by Yang-Hui He, Ziran Liu and Yau, which claims a metric of strictly positive sectional curvature on all 28 smooth homotopy 7-spheres, by building positively curved metrics on two 7-disks and gluing them with the Reiser–Wraith theorem. Its acknowledgments thank "ChatGPT 6 Astra and Claude Pro" for help exploring proof strategies and calculations; code, exact-arithmetic certificates and an independent checker are public. Earlier attempts covered only the Gromoll–Meyer sphere; a September 11 preprint (2609.12882) says its metric and proof were found by an AI research agent called Odin.

【Our read】

The template matters more than the roster: models suggest and compute, computer algebra certifies, humans argue and take responsibility. If 2609.29426 survives review it is among the weightiest pure-math results AI has touched, but it is a preprint, and certificates check the numerics, not the construction. Human-led and AI-led claims on the same problem will now be tested. No model timeline or compute plan yet.