Mathematician claims LLMs only interpolate, can't discover new math like Riemann
burny_tech · x · 2026-10-19
A user pushed back on mathematician Alex Kontorovich's view, arguing that LLMs can only interpolate between training data, not extrapolate. This means models can connect A to B via chains of known intermediate techniques—chains that would have taken humans longer to find—but they cannot make genuinely novel mathematical discoveries, such as progress on the Riemann hypothesis or P vs NP. The poster set a /remindme 1year reminder, implying the claim may be proven right or wrong within a year.
More from AGI Musings
- 'It's only intelligence if it radically expands accessible truth — otherwise it's sparkling automation' — fkasummer · 2026-10-19
- AI solves the n=1 software problem: tailored apps will replace configurable ones — sloppenheimer · 2026-10-19
- e/acc meme: "They don't think AGI could produce double-digit real GDP growth by 2030" — ChrisGPT · 2026-10-19
- Debate: Don't Automate People Out of Jobs Without Universal Shelter and Food First — broodsugar · 2026-10-19
- Hassabis: I Build AI to Solve the Puzzles of Consciousness, Time and the Universe — dioscuri · 2026-10-19
- Stripe exec: witnessing AI backlash firsthand finally makes historical tech resistance make sense — gabriel1 · 2026-10-19