MIT Professor: Models Cross a Discovery Threshold Around 0.76 Log-Perplexity
ProfBuehlerMIT · x · 2026-11-23
MIT professor Markus Buehler, drawing on OpenAI's "Perplexity and Intelligence" data and @merettm's analysis, argues AI scaling is entering a regime of qualitatively new open-ended discovery capability.
Key points
- Early scaling laws showed next-token prediction improving along clean power laws, unlocking fluent conversation, reasoning, coding, and math in sequence.
- Broader capability categories also follow a scaling law: around 0.76 log-perplexity on OpenAI's benchmark, models began producing results on extremely hard open math problems.
- The accompanying chart maps milestones (fluent conversation → thinking → useful coder → cybersecurity expert → Millennium Prize math problem) to perplexity tiers, with the next transition labeled as unknown future capabilities.
- Mechanism hypothesis: lower tiers reflect deeper representations capturing shared structure across distinct problems, enabling cross-domain transfer and recombination toward genuine scientific novelty.
Buehler expects the next expansion into lower tiers to be even more consequential for scientific discovery.
More from AGI Musings
- Noam Brown's 2023 bet lands: 10K agents solve Navier–Stokes, Riemann proof claimed — haider1 · 2026-11-23
- 2004 Letter Predicted Mathematicians Would Be the First Profession AI Makes Obsolete — ericwdolan · 2026-11-23
- If AI proves Riemann, formal verification becomes the infrastructure of all software — inductionheads · 2026-11-23
- Can We Engineer Serendipity? Rethinking social networks with agents — Borthwick · 2026-11-23
- Should lab mathematicians get better models than the public? — inductionheads · 2026-11-23
- LeCun: we supposedly have superintelligence but still no level-5 autonomous driving — ylecun · 2026-11-23