Richard Sutton agrees: AI progress is overestimated — but that leaves vast room to improve
RichardSSutton · x · 2026-10-13
Philosopher Dan Williams argues AI capability progress is far less than assumed: frontier models still struggle with few-shot learning, continual learning, and physical interaction, with highly uneven competence — superhuman math/coding doesn't transfer to messy white-collar work. But that's no comfort: it implies vast room to improve existing systems by changing training methods. Turing Award winner Richard Sutton replied, 'Agree on all three points.'
More from AGI Musings
- Beijing issues identity cards to AI digital humans, raising AI accountability questions — tallmetommy · 2026-10-13
- NotebookLM-generated fake podcast with 13 episodes and ads passes as real show — omooretweets · 2026-10-13
- AI is fast where verification is easy — and safe automation requires measurement first — WhatTheLJW · 2026-10-13
- Norbert Wiener on identity: 'patterns that perpetuate themselves,' visualized with Rule 30 — pickover · 2026-10-13
- Graph algorithm researchers shrug at OpenAI's general matching math result — thegautamkamath · 2026-10-13
- First Real Test of Meta's Muse Agent: A Birthday Party Email Drafted in 30 Seconds — RachelVT42 · 2026-10-13