Shing-Tung Yau launches math-AI alliance of eight institutions aiming to network more than 1,000 researchers

Yau's own diagnosis is dispersal. "The U.S. and the West are gathering top-level mathematics experts to work with AI and explore advanced problems, while China's efforts are scattered," he said ^[11]. An alliance answers that directly (the South China Morning Post counts a network of more than 1,000 scientists ^[3], so the 1,000 in the other account is better read as a floor than a target ^[2]). It answers the other half of his complaint, the model gap, less directly, because the models are listed as plans ^[4] and the collaboration system is the part with a stated shape ^[2].

The yardstick he chose is a month of OpenAI output. "If you ask Chinese AI models to do what OpenAI has done over the past month, the gap is very large," he said ^[10]. The batch OpenAI released averages about 1.9 manuscripts per group of results ^[9], and the company's claim to have used AI to solve the Navier-Stokes problem in 88 hours ^[17], or about 3.7 days ^[18], is its own. Yau is measuring against figures OpenAI supplied itself ^[16].

We think the announcement is a bet on people and not yet on money, or rather, a bet on people with the model-building left as a plan. The published reports do not include a budget, a funder or a compute allocation, and neither ties the alliance to a government program; sedaily.com calls it a self-help move ^[7]. Yau directs the Yau Mathematical Sciences Center at Tsinghua ^[6], and a state-directed reading would be an inference from where he sits.

If Yau is right that the problem is dispersal, a shared system is the right first purchase and output should follow. If dispersal is only part of it, the alliance produces papers and conjecture lists while the model gap he describes stays open. A third possibility is that the label gathers work the eight institutions were already doing. A specialised model, or conjectures and formalised proofs ^[5] that mathematicians outside the alliance can check, would favour the first reading. A year of announcements without either would favour the second or third.