AI Reshapes the Rules of Mathematical Research and Education: Terence Tao (陶哲轩) Confirms Students' Closed-Book Exam Scores Are Already Trending Downward

Tao's warning: performance slipping, or the test is changing?

It has been reported that Fields Medalist Terence Tao (陶哲轩) has observed a downward trend in students' closed‑book mathematics exam scores, a signal many educators see as evidence that AI is already changing how mathematical knowledge is produced and assessed. Tao’s reported comment landed amid a broader debate: are students failing to learn, or are traditional closed‑book assessments simply the wrong metric in an era of ubiquitous computational assistance? The claim, reportedly made in public commentary, has reignited discussion about academic integrity, pedagogy and the core goals of math education.

Powerful, open models accelerate the shift

Part of the pressure comes from dramatically more capable AI tools. It has been reported that NVIDIA (英伟达) this month released Nemotron 3 Super, an open‑weight model with a 1.2‑billion‑parameter MoE backbone (120 billion in total, 12 billion active at inference) and a 1 million‑token context window — designed for agentic, multi‑step workflows and capable of loading entire codebases for end‑to‑end debugging. Open‑weight models like this — whose parameters and training recipes are publicly available — lower the bar to deploying advanced reasoning systems, making it easier for students and researchers to outsource steps of problem solving and for institutions to reconfigure workflows.

What this means for research, classrooms and policy

The practical consequences are immediate and messy. Educators face “context explosion” and new forms of shortcutting; researchers are experimenting with AI‑assisted proofs and conjecture generation; and institutions must decide whether to harden proctoring, redesign assessments toward process and understanding, or embrace AI as a collaborator. Geopolitics matters too: export controls on chips and debates over cloud access aim to shape who can run the largest models, yet open‑weight distribution and multi‑cloud availability (it has been reported that the model is accessible via major cloud vendors and Hugging Face) complicate containment strategies.

How should math be taught when an AI can show steps but not insight? That question frames the near term. Policymakers, universities and the global research community now face a choice: double down on surveillance and restriction, or retool curricula to value creativity, model‑guided reasoning and judgment over recall. Terence Tao’s reported observation is a warning shot — a prompt to redesign the rules, not merely to lament their erosion.