Terence Tao: Math 2.0 [pdf]
Rapid advances in AI-assisted theorem proving are forcing mathematicians to rethink what counts as progress in their field. Commenters debate whether machine-generated proofs that humans struggle to understand advance mathematics or risk “sterilizing” research by shortcutting the exploratory process, using Tao’s controversial cancer-drug analogy to argue over how much mechanistic understanding society should demand from AI-derived results. Underneath are broader worries about careers, incentives, and whether future math should prioritize human insight over sheer problem‑solving throughput.
Cancer-cure thought experiment & medicine analogy
- A slide about an AI-designed cancer “cure” sparked the biggest controversy.
- Many commenters argue the example is misguided: medicine routinely uses drugs whose mechanisms are poorly understood but empirically safe and effective (e.g., anesthesia, acetaminophen).
- Several say that if they were terminal, they’d gladly take an AI-designed drug that passed a proper phase 3 trial, regardless of mechanistic understanding.
- Others defend the thought experiment as being about alignment: an AI could “game” flaws in trial design or modeling, so relying only on trial results without deeper understanding is risky.
- The wording changes to the slide (more cautious, less absolutist) are noted; some still find it confusing or misleading.
Understanding vs efficacy and alignment
- One camp prioritizes empirical success: if a treatment, algorithm, or proof works and passes robust tests, mechanism is secondary.
- Another camp argues understanding mechanisms is crucial for safety, generalization, and improving methods, and that “it seems to work” is not enough in the long run.
- Analogy: running unverifiable black-box binaries vs having source code and architecture.
AI in mathematics: process, proofs, and “Math 2.0”
- Core claim of the talk (as interpreted): pure math is uniquely focused on understanding; proofs are valuable as part of a coherent “theory codebase,” not just as yes/no answers.
- “Open problems as lighthouses”: solving them too quickly via opaque AI proofs can sterilize surrounding areas by cutting off the human exploratory process and technique development.
- Concern: AI proof dumps may “strip-mine” approachable problems without yielding new human-comprehensible methods, worsening the “junior mathematician” pipeline and incentive structures.
- Counterpoint: the OpenAI-style results already seem to challenge long-held assumptions (e.g., sub‑O(n log n) FFT-style results) and are actively seeding new human work and refinements.
Impact on careers, incentives, and identity
- Many see mathematicians going through a visible grief process similar to what software engineers, journalists, etc., faced with AI.
- There’s debate over whether this is primarily emotional/job-protection vs a legitimate concern about how to measure contributions and allocate funding when AI can solve benchmark problems.
- Suggestions include: shifting evaluation toward exposition, distillation, new concepts, dependency maps, and organizing large proof ecosystems.
Comparisons to software engineering and other domains
- Strong parallels are drawn:
- AI code vs AI proofs; “it runs” vs “it’s maintainable and understandable.”
- Medicine’s empirical pipelines vs math’s proof culture.
- Potential monoculture and loss of diverse approaches as AI centralizes problem-solving.
- Several argue that, regardless of flaws, AI tools for math and code will be used aggressively; the real work is to adapt institutions, norms, and roles around that reality.