AI's mathematical results are not a bad thing

Salvatore Sanfilippo argues that AI-generated mathematical results need not diminish human research: a difficult proof can become something people study, simplify, and use to develop new techniques. He connects the reaction to these results with the discomfort of seeing a coding agent quickly perform work that once distinguished a skilled programmer.

Key points

• Capability gains can feel like a loss of personal value: Sanfilippo suggests that when a rare skill becomes readily reproducible by AI, some practitioners experience it as a reduction in their own worth. He compares mathematical breakthroughs with a coding agent producing a program in half an hour that might have taken its user a week or a month; this is an illustration, not a measured productivity comparison 00:01–01:05(opens in a new tab).
• He presents singularities as a reminder of a model's limits: Discussing a result he attributes to OpenAI, he describes fluid equations producing infinite velocities in exceptional conditions, whereas physical systems consist of discrete particles. His point is that a simulation blowing up need not always mean a software bug or finite-precision error. The video does not establish the exact theorem, assumptions, or which simulations it applies to 01:05–03:17(opens in a new tab).
• More knowledge and a healthy research process are different questions: In Sanfilippo's account, Tao worries that externally produced AI proofs could be a net negative for mathematics by weakening incentives to develop techniques while pursuing hard problems. Sanfilippo responds that a proof still adds information, even if its effects on the trajectory of human research remain a separate question 03:17–05:22(opens in a new tab).
• A difficult proof can become a new research task: Sanfilippo proposes taking a formal proof in Lean, a language and tool for machine-checked proofs, apart and simplifying it until it becomes understandable. He argues that this work could itself yield new mathematics, with AI helping researchers interpret results and pursue further constructions. This is a proposed use of the output, not a demonstration that such simplification has succeeded for the result discussed 05:22–06:45(opens in a new tab).
• He argues that researchers can redirect their work: Sanfilippo says the mathematical community retains substantial control over which problems it studies and could organize around interpreting AI results. He also speculates that resistance partly reflects threatened status. That explanation of other people's motives is his interpretation, not evidence that their objections are unfounded 06:25–08:40(opens in a new tab).

Relevance to coding-agent work

Editorial takeaway: Sanfilippo's distinction suggests treating generated code as material to inspect and understand, rather than treating its speed of production as a verdict on the developer's value. The video does not provide a code-review workflow or show that fast generation makes software correct.

Watch the full video(opens in a new tab).