DeepMind paper says LLMs need abductive reasoning, not just compression or proofs
burkov · x · 2026-07-29
Core claim
Tom Zahavy of Google DeepMind argues in an ICML 2026 position paper that LLMs will not reach foundational scientific breakthroughs through data compression or deductive proof alone.
Argument
- He uses Einstein’s development of General Relativity as the example of a discovery process that depended on abductive reasoning.
- The paper claims true invention requires physically consistent world models, not just better memorization or theorem-proving.
- The post frames this as an AI tutor-readable position paper rather than a benchmark result or model release.
Related event: DeepMind Paper Outlines LLM Limits in Scientific Discovery(5 posts)→
More from AGI Musings
- Why one AI bull says the boom still has room to run: TAM, RSI, agents and HBM — johncoogan · 2026-07-29
- A borg agent essay imagines software that can survive, copy itself, and move servers — sull · 2026-07-29
- Gary Marcus says a 900-ELO chess bot is nowhere near AGI — GaryMarcus · 2026-07-29
- A blunt take says elite fear of ASI is driving today’s AI debate — AIandDesign · 2026-07-29
- Kara Swisher and Former Commerce Secretary Gina Raimondo Discuss AI Job Disaster Congress Refuses to Fix — kylemwright · 2026-07-29
- AI safety is repeating the mistake of assuming a U.S.-China race is inevitable — voooooogel · 2026-07-29