Ex-DeepMind research head Vinyals: AI self-improvement won't spark an explosion
The Decoder · rss · 2026-09-12
Oriol Vinyals, until recently head of research at Google DeepMind, thinks a sudden intelligence explosion via recursive self-improvement is unlikely. AI could speed up research 10x, but hits two bottlenecks: 'research taste' for generating ideas, and reliably judging results; reward hacking and the speed of light add further limits. He co-founded startup Discovery Loop with Jeff Dean, Sanjay Ghemawat and Quoc Le to tackle exactly these bottlenecks.
More from AGI Musings
- Agents slash the cost of consumer vigilance, clawing back what businesses skim from inattention — signulll · 2026-09-12
- Fields Medallists warn AI's benchmark-chasing is 'severely misaligned' with mathematics — hbouammar · 2026-09-12
- Google researcher: non-transferable adoption costs make waiting the rational move, slowing AI impact — moultano · 2026-09-12
- Silent model deprecations set precedents while AI moral status debate keeps being deferred — repligate · 2026-09-12
- Researcher: doomer arguments would land better without anthropomorphic language — round · 2026-09-12
- Fortnow: Persistent Writable Memory Could Enable Recursive Self-Improvement — fortnow · 2026-09-12