Fastest matrix multiplication team: AI can't understand its own proofs
luismbat · x · 2026-11-17
Researchers behind the world-record fastest matrix multiplication algorithm shared a counterintuitive observation from using AI in mathematical research: the obfuscated proofs AI generates are hard not only for humans but also for the AI itself.
- Models consistently failed to build on ideas they had generated themselves
- When humans presented a clean simplification of those ideas, the models (and the humans) could suddenly take further steps
- The authors credit the algebraic decomposition mainly to the AI and the algorithmic framework using it mainly to themselves
The methodological takeaway: AI outputs need human simplification before re-feeding, forming an effective human-AI iteration loop.
More from AGI Musings
- Harvard Class That Bans Laptops Draws 180 Applicants for 90 Seats — ValerioCapraro · 2026-11-17
- With algorithms and AI-fueled output, the economy increasingly resembles a lottery — StewartalsopIII · 2026-11-17
- Why hasn't OpenAI answered Opus 5.5 and Fable 5.2? Redditor speculates nascent RSI at Anthropic — Robot_Apocalypse · 2026-11-17
- The hidden cost of AI speed: premature closure ends thinking too early — DrKavner · 2026-11-17
- After 40 years of searching, he gave up on note-taking systems entirely — bfrench · 2026-11-17
- SF law firm drops automatic offers for summer associates, tells new lawyers to self-train in AI era — jkubicki · 2026-11-17