Dwarkesh: ASI median arrival 2028, as AI R&D is largely verifiable like math
daniel_mac8 · x · 2026-11-16
Dwarkesh Patel now believes ASI arrives within a few years — median 2028, possibly 2027 or as late as 2030. His core argument is verifiability: AI went from failing elementary math to tackling Millennium Prize problems in 2 years because math is a verifiable domain, and AI research itself is nearly as verifiable — pretraining loss, RL environment scores, and in-context sample efficiency are all measurable. Sample efficiency / continual learning, long ML's deepest open problem, has strong verifiable proxies; solving them could trigger a compounding domino effect across AI R&D. He notes the vibe among well-informed researchers has shifted to "it's really going to happen."
More from AGI Musings
- eigenrobot cites cleaner fish stations to argue markets emerge naturally without humans — eigenrobot · 2026-11-16
- Founder retired the day his kid was born: raising children is what AI can't do — PTrubey · 2026-11-16
- AI safety event for women sparks misogyny debate over "simple terms" framing — amplifiedamp · 2026-11-16
- Doctors who ask AI to show its work are earning clinical trust the right way — MaziyarPanahi · 2026-11-16
- Asimov Press founder: academic journal innovation is strangled by indexers like Clarivate — NikoMcCarty · 2026-11-16
- Gary Marcus: I respect honest doomers, not those who make things up about me — GaryMarcus · 2026-11-16