Dwarkesh argues AIs in the intelligence explosion will master 'fuzzy' work too
trevposts · x · 2026-11-14
Responding to skepticism that AIs only excel at measurable tasks, Dwarkesh offers two arguments: generalization is already emerging at less than 1/1000th of the RL scale expected during the intelligence explosion, and evolution produced highly general human brains from an extremely crude verifiable signal (gene survival). trevposts agrees this implies AIs will soon handle the hardest parts of AI research better than humans, making recursive self-improvement plausibly less than 6-24 months away.
More from AGI Musings
- T-shaped vs H-shaped: debate over what talent thrives in the LLM era — dotey · 2026-11-14
- Mike Frank: maxed-out compute won't make AI magic—it's still just software — MikePFrank · 2026-11-14
- Sakana AI Keynote in Seoul: The Path to Recursive Self-Improving AI — Stefania_druga · 2026-11-14
- Blogger: AI cognition will soon surpass all of humanity — answer is symbiosis, not bans — danfaggella · 2026-11-14
- Dwarkesh's footnote on compute bottleneck to RSI strikes a chord in AI circles — zetalyrae · 2026-11-14
- Robotics researcher pushes back on latency generalizations in AI debates — KyleMorgenstein · 2026-11-14