Interview
Grant Sanderson (@3Blue1Brown) – AI disproved a famous math conjecture. Now what?
- AI progress in mathematics is projected to serve as a leading indicator for broader AI capabilities, though solving high-level problems like the Riemann Hypothesis is not expected to trigger an instant "aha moment" enabling all human tasks due to persisting narrowness in capabilities.
- The rate of solving complex mathematical problems and performing white-collar work is constrained by the same limiters, with a "100-year verification loop" potentially required for new concepts; however, AI may solve the Riemann Hypothesis by connecting distinct fields, an ability distinct from standard white-collar work.
- Over the next five years, the most useful AI progress is expected to involve filling the landscape of connections between fields, such as the Langlands program, driven by AI acting as a supercharged connector of ideas rather than just solving specific problems.
- In the next three years, headline-making results will likely involve AI erasing context to try fresh approaches, utilizing systematic entropy increases and parallelization to test diverse heuristics, potentially disrupting the "theorem economy" by automating proof while humans focus on insight generation.
- AI is anticipated to generate novel fields and solve Millennium Prize problems within 5 to 10 years, creating massive demand for human curation and explanation, with economic value materializing primarily through incremental engineering and simulation improvements rather than immediate shocks.
- A subjective "tone shift" will likely mark the next benchmark, shifting mathematician roles from problem-solving to curating infinite AI-generated ideas and deciding research directions, while the most stable mathematical career over 50 years is expected to be teaching due to its relational nature.
- Formalization tools like Lean may become crucial for AI to extend mathematical libraries endlessly, though natural language verification with meta-verification could suffice; however, error rates in natural language proofs remain intolerable without formal verification guarantees.
- AI performance in creative tasks differs from brute-force geometry, and success in drawing "unlikely connections" depends on training environments that incentivize exploration over token optimization, with digital minds leveraging parallelization to apply capabilities simultaneously.
- Challenges remain in AI's ability to understand people and emotions ("theory of mind"), leading to struggles in writing and "jujitsu-ing" a student's mental framework; effective student learning will likely rely on human-produced artifacts guiding AI to prune and refine understanding.
- Solving problems via "alien" or incoherent methods risks human verification failures, while the ideal outcome involves AI finding "compressed" elegant theories; current progress is driven by grindability and verifiability, unlike computer use hindered by real-world constraints.
- Material numbers and timeframes include a "100-year" verification period for concepts, a "3-year" horizon for context-resetting breakthroughs, and a "5 to 10-year" window for potential Millennium Prize solutions and novel field generation.
- Risks include the potential for AI to generate proofs that are incomprehensible to humans for years or incorrect, the devaluation of proof credit, and the possibility that current research areas become disconnected from physical applications due to the sheer volume of new math.