Chollet: far transfer has no evidence in humans, and LLMs/LRMs may be no different
burny_tech · x · 2026-10-09
François Chollet argues far transfer lacks real-world evidence: near transfer exists (skill at one RTS helps others) but chess mastery doesn't make you good at RTS games. He suspects large reasoning models may be similar, though anecdotal evidence is inconclusive. ARC-AGI performance correlates strongly with math and code abilities, but correlation isn't causation. He grants that LRMs generalize in-domain and have substantial fluid intelligence, while leaving open whether they meet a skill-acquisition-efficiency definition of intelligence.
More from AGI Musings
- Andrew Critch hits back at "Claude probably isn't conscious": which consciousness, exactly? — AndrewCritchPhD · 2026-10-09
- Safety researcher: OpenAI is uniquely sabotaging regulation while other labs' risks aren't symmetrical — Turn_Trout · 2026-10-09
- Scott Aaronson: the AI skeptics' 'wake me up' warnings have all come true — scottleibrand · 2026-10-09
- Former multimodal researcher: aligning model 'taste' for everyone is unsolvable — A_K_Nain · 2026-10-09
- Traversy Media: newbies can't learn AI coding without writing code first — capeandcode · 2026-10-09
- Enjoying the Centaur Era While It Lasts: OpenAI's Math Results Signal Its End — moyix · 2026-10-09