A Tsunami Hit Mathematics. Academia Just Built a Surfboard.

On October 11, 2026, the Yau Mathematical and AI Laboratory Alliance was established at Tsinghua University. Shing-Tung Yau leads it. It brings together eight major mathematical institutions: Qiuzhen College, the Yau Mathematical Sciences Center, the Beijing Yanqi Lake Institute of Applied Mathematics, the Morningside Center of Mathematics, the Shanghai Institute for Mathematics and Interdisciplinary Sciences, the Southeast University Yau Center, the Hetao Institute for Mathematics and Interdisciplinary Sciences in Shenzhen, and the Chinese University of Hong Kong’s cross-disciplinary mathematical science academy. The alliance has more than 1,000 scientists. Its stated directions include mathematics-specific large models, AI-assisted formal proof and conjecture discovery, interpretable AI, and intelligent computing in cross-disciplinary fields.

To understand why this alliance exists, start four days earlier.

On October 7, OpenAI posted 722 AI-generated mathematics papers on GitHub. They covered number theory, algebraic geometry, and other fields. Some pushed open problems forward. Some falsified existing claims. Within a day, OpenAI retracted three manuscripts and revised fourteen. Terence Tao shared a joint statement from a human mathematics association. It called on mathematicians to stop cooperating with OpenAI and argued for a return to a scientific vision centered on human understanding.

Yau called it a tsunami. “Like a tsunami,” he said, “how should we respond?” The metaphor captures a structural passivity. A tsunami does not change direction because you plant your feet more firmly on the shore. You either ride it or you are submerged.

Yau chose to ride. He did not choose to ride by asking mathematicians to reject AI. He chose to organize them. The alliance is a defensive piece of academic infrastructure. It also funds technical work. Its larger purpose is to help the academic community reclaim standard-setting power and agenda-setting power after AI has begun to reshape mathematical research.

The integral offensive of industry and the differential dilemma of academia

Industry has already moved far.

Three overseas labs have taken three different routes. All three share one feature: their scale and iteration speed exceed what any single academic institution can carry.

Google DeepMind has taken a neuro-symbolic plus reinforcement learning route. In May 2026, it released AlphaProof Nexus. The agent combines large language model proof generation with Lean formal verification. It autonomously solved nine of 353 open Erdős problems. One had been open for 56 years. The cost was as low as $7.50 per research-level problem. The marginal cost of attacking frontier mathematics with AI is falling faster than intuition suggests.

Anthropic has taken a single-model long-horizon autonomous formalization route. On September 4, its Claude model completed the first end-to-end Lean formalization of Fermat’s Last Theorem in 11 days. Mathematicians had generally expected years. Claude wrote about 13 million lines of Lean code and completed 30,300 machine-verifiable theorems. It consumed about 6 billion output tokens. Kevin Buzzard of Imperial College London led a community project to formalize Fermat’s Last Theorem. Its first-stage technical blueprint alone ran 86 pages. Buzzard said that before, he was 99.9 percent sure the proof was correct. After Claude’s work, he was 100 percent convinced.

OpenAI has taken an agent parallel search route. In its Navier-Stokes project, about 10,000 concurrent agents ran for 88 hours. They exchanged 2.7 million messages, consumed 130 billion output tokens, and cost millions of dollars. The logic is compute covering as much proof search space as possible. Depth from a single model matters less.

Yau has a useful summary. From ancient times to now, progress in mathematics and basic science has mostly relied on the thinking of a few people. It is a “differential” progress. Artificial intelligence is an “integral” effect. It stacks continuously and produces exponential speed. The observation exposes a structural trap. Academia is used to producing differential-level innovation. It is now forced to compete in an integral-level field. A mathematician’s life work may be formalized, verified, or surpassed in a single model iteration. Yau himself admitted that a year earlier he would not have thought AI would bring such a large challenge to mathematics.

The compute gap and the anxiety of subjectivity

The alliance has left room for industry cooperation. Its operating mechanism includes shared computing platforms and large-model training resources. But the compute reality is harsher than the intention.

AlphaProof’s training pipeline consumed about 100,000 TPU-days. That is equivalent to 2,000 TPUs running for 50 days. No single academic alliance can absorb that. In China, the compute situation is tighter. At WAIC 2026, industry participants described the year as “one card hard to find.” Supply fell short of demand. High-end GPU cards sold out as soon as they arrived. At universities, compute long remained concentrated in top research groups. When several groups train models at the same time, “compute queuing becomes normal, and research progress is paused.”

This is the deepest structural contradiction. Academia has the authority to define problems and judge truth. It lacks the compute and engineering resources to turn that authority into frontier model capability. Industry has resources. Its output mode conflicts with academic norms.

The OpenAI release of 722 manuscripts shows both the energy and the chaos. Xi Nanhua, president of the Chinese Mathematical Society, pointed out that AI companies batch-produce results in a way that conflicts with peer review. Companies can publish material that has not been rigorously checked. Tao’s shared resistance statement is a protest against a mode of knowledge production that bypasses the academic community. If academia relies entirely on industry APIs, research agendas can be pulled by commercial logic. Academic subjectivity disappears. That is the existential question the alliance must face.

What academia still owns: data standards and specialized tooling

Academia is not empty-handed. It owns something industry cannot produce on its own.

The alliance has already published results that show the division of labor. In work on seven-dimensional homotopy spheres and positive sectional curvature, AI helped explore proof strategies and simplify calculations. Computer algebra systems verified exact identities, interval bounds, and reproducible certificates. Mathematicians conceived the proof, checked the result, and wrote the argument. In work reconstructing three-dimensional black hole entropy from a discrete spectrum, AI helped construct counterexamples, generalize results, and write Lean-checked formal proof code. Researchers posed the questions, revised conjectures, analyzed the physics, and checked the formal statements.

Industry and academia are dividing the work. Industry has compute and engineering. But high-quality mathematical data and standardization depend on academic judgment. Building formal mathematics libraries, checking rigor, deciding which conjectures deserve compute: these cannot be done by a model alone. Yau said that making machines think and reason requires learned researchers. He also said that many scholars need to work together to make original mathematical AI models.

The other key variable is specialized tooling. Fermat’s Last Theorem was formalized in 11 days, with 6 billion output tokens and 13 million lines of code. The model was not smarter in some general sense. The project used engineered project management. The grand goal was decomposed into a manageable dependency graph. A Prove2Me-style system scheduled model calls, managed proof states, and controlled context size. Peng Tianyi, an Anthropic researcher and Tsinghua Yao Class alumnus, initially wanted to test whether Claude could push the project forward. It quickly exceeded expectations. Dozens of Claude agents worked at the same time. They defined mathematical concepts, proved intermediate propositions, and combined results into higher-level conclusions. Humans occasionally gave directional prompts: “The Jacobian variety as a scheme has high priority,” or “Finish Mazur’s theorem soon.”

Systems engineering around the model shapes its capability as much as parameter count does. The alliance lists AI-assisted formal proof and conjecture discovery as a key direction. Its main output may not be a single model. It may be an operating system for mathematical research: data formats, verification toolchains, tooling frameworks, and evaluation protocols. That soft infrastructure may last longer than any one model.

Tao’s mirror: a parallel path

The alliance is not alone. A parallel exploration is underway on the other side of the Pacific.

On September 18, 2026, Terence Tao announced on his blog that the Science and AI Research Foundation, which he co-founded, had launched the Open Math Model Initiative. Its aim is to build open-weight models for science and open-source tools. It invites the mathematical community and its supporters to build open models together.

The keywords are open weights and community control. It explicitly joins academia and industry. It follows open development, community-owned data, shared intellectual property, and open community governance. It focuses on daily research scenarios such as understanding arguments and checking literature. Tao has chosen a path different from OpenAI and Anthropic. He is not chasing a single problem beyond human ability. He is building open, reusable academic infrastructure so that mathematicians benefit in their ordinary workflows.

The Yau alliance and the Tao initiative share the same logic. Both recognize that academic research dependent purely on industry APIs is unsustainable. Both need academic community control over data and tools. The difference is organizational. Yau’s alliance is more institutionalized and rests on a large national network. Tao’s initiative gives more control to the community and works in the open. Their parallel appearance is itself a signal. Academia’s response has moved from shock to construction. AI will change mathematics. The fight is over who defines how.

What success depends on

The alliance has a chance to change the balance of power in AI plus mathematics. Several assets help. It has organized academic credibility: eight top mathematical institutions and more than 1,000 scientists. That gives it natural authority over data standards and evaluation protocols. It has early demonstrations: the seven-dimensional homotopy sphere work and the black hole entropy work show a feasible model of AI assistance under mathematician leadership. It has institutional interfaces for industry cooperation: shared compute and model resources are built into its design.

The challenges are just as serious. The structural compute shortage will not disappear because an alliance is founded. Industry iteration will not slow because academia organizes. The tension between academic norms and commercial logic will not resolve because both sides express willingness to cooperate.

Success depends on three concrete tasks.

First, it must build a set of mathematical data and evaluation standards recognized by industry and become the court for AI mathematical capability. That requires operable rules for formal mathematics library standardization, benchmark design, and rigor verification. If it cannot, the alliance remains an internal academic club.

Second, it must form a feasible academic-industry compute cooperation mechanism rather than a statement of intent. A layered model is possible. Academia handles data standards, benchmarks, specialized mathematical tooling, and formal verification pipelines. Industry provides base models and compute. The feasibility depends on intellectual property, attribution, and agenda control.

Third, it must produce reusable specialized tooling and formal toolchains that reduce dependence on giant models. The Fermat case shows that engineered tooling can push the boundary of model capability outward. If the alliance can build a general mathematical research framework that lets ordinary mathematicians use AI for complex formalization, its impact will exceed any single model release.

Yau said at the founding ceremony: “If handled well, scholars can, with AI’s help, make epoch-making impact and achieve great results. If handled poorly, it may bring a larger development gap.” The sentence applies to the alliance itself and to the fate of Chinese mathematics in the AI age.

The tsunami has arrived. The alliance is an organized attempt to surf. Whether the surfers ride the wave depends on whether they build a board strong enough before the wave breaks. That is the thing academia is least good at.

In a university server room, a queue sheet is taped to the door. Next to the estimated available time, someone has written, in pencil, “heh.”