Interview
Terence Tao – How the world’s top mathematician uses AI
- By 2030, models are expected to achieve stunning proficiency in tasks like college-level math and face recognition, while within two to five years, the cost of idea generation may drop to near zero, shifting the bottleneck to verification and potentially allowing humans to generate thousands of theories.
- Within a decade, AI is predicted to perform the bulk of tasks currently done by mathematicians, including work published in papers, leading to a hybrid scientific state where AI handles breadth and humans provide depth, a future described as unrecognizable from current methods.
- AI acceleration is anticipated to make new discoveries and breakthroughs occur more quickly, with specific expectations that a few more Erdős problems may be breached during the next major model advance, potentially moving beyond the current plateau of low-hanging fruit.
- A semi-formal framework mimicking scientific dialogue could emerge for semi-automated validation of conjectures, preventing current reinforcement learning techniques from easily hacking the process, while new professions may arise to perform ablation on giant Lean proofs and find more elegant solutions.
- The transition to full AI autonomy for solving Millennium Prize problems is expected to take longer than the current capability to dominate mathematics, as models currently lack all necessary ingredients to fully replace intellectual tasks without further breakthroughs.
- Society faces the risk of inhibited progress if the optimization of schedules and digital discovery destroys serendipity, eliminating accidental interactions and random inspiration necessary for high-temperature problem solving.
- Human reviewers are already overwhelmed by AI-generated submissions, indicating that verification and assessing what ideas move the subject forward will become the primary constraint in the research ecosystem.
- Cryptography based on primes could face a "big, big shock" and be rapidly abandoned if the Riemann hypothesis fails, as this would imply exploitable patterns and a secret patent for primes.
- Hedge funds will likely continue preferring astronomy PhDs for their expertise in extracting signals from random data and maximizing information density, while students at the high school level may gain non-traditional opportunities to make real contributions using AI tools and Lean.
- Historical parallels suggest mathematicians will shift to different types of problems once AI handles low-level tasks like generating code or log tables, leaving experts in single areas (hedgehogs) and generalists (foxes) relevant as AI takes over trivialities like reformatting.
- AI capabilities could enable the generation of millions of variants of functions like the zeta function to discover patterns that transform problems into different mathematical areas, creating new connections and types of mathematics.
- Once models reach specific competence levels, millions of copies could be allocated a million dollars of inference compute each to simulate 100 years of subjective research time across 100 different problems simultaneously.