LessWrong essay: LLMs are moving from memorization to making new discoveries
gleech · x · 2026-10-10
A LessWrong essay, 'The Convex Hull Armada', revisits Dwarkesh Patel's 2023 question to Dario Amodei: why do models that memorize all human knowledge make no new discoveries? The author argues RL elicitation of new connections is bearing fruit, tracing the hypothesis through recent mathematics examples. Speculative, written with Claude Opus 5.5 as a feedback partner.
Related event: Tao's Blog and LessWrong Debate LLMs Moving from Memorization to Discovery(2 posts)→
More from AGI Musings
- Pedro Domingos: Superintelligence Will Be Commonplace — pmddomingos · 2026-10-10
- Why some AI boosters resent mathematicians: AI culture as psychiatric solvent — _onionesque · 2026-10-10
- 'Meat proxy era' banter: humans mocked as placeholders in the loop — max_paperclips · 2026-10-10
- Tao blog essay on LLMs lands amid OpenAI Navier-Stokes announcement and misconduct claims — gleech · 2026-10-10
- Musk: Digital Optimus beats Diablo halfway through with no APIs, just screen pixels — XFreeze · 2026-10-10
- Vibe coding has underdelivered: why waiting for the next model kills software commitment — sull · 2026-10-10