Harald Helfgott: ‘Mathematics and the AI crisis’

Posted Oct 1, 2026 on the blog ‘The value of the variable’ (Tao’s list attributes it to Harald Helfgott). Summary from an automated reading: he considers three scenarios and examines the first, an ‘AI autumn’ in which progress slows but AI stays a transformative tool. ‘I think the best attitude here is polite agnosticism, mixed with a readiness to accept extreme outcomes.’ He argues that AI filled a real gap (‘We had no search engine for mathematics’), that ‘Whatever is done easily by a machine loses prestige’, and that ‘It is also hard not to suspect that LLMs are leaning hard on the literature’. He discusses credit, publication and doctoral training, and names models he used (Claude Sonnet 4.5 to Opus 5.5, ChatGPT 5.2 to GPT-6 Astra, Gemini, AlphaEvolve, Mistral). A postscript mentions OpenAI’s ‘October 6 mathdump’. Quotes were extracted by a summarising tool and should be re-checked against the page.