DeepMind's Andrew Trask: AI's endgame is millions of ensembled models, not one giant
iamtrask · x · 2026-10-08
- Google DeepMind senior research scientist and OpenMined founder Andrew Trask argues AI won't converge on one giant model: millions of models ensembled and routed per prompt will beat any single frontier model on quality and price, making AI look more like the PC and internet than the mainframe.
- The wide-ranging interview also covers who gets to audit AI labs: why embedded evaluators aren't enough, the "evil EAs" criticism of eval orgs, labs only calling people they trust, a day in the life of an independent evaluator, and whether models will game evals.
- He details the first double-blind evaluation of a frontier model and discusses OpenAI's math release and sandboxes with no holes.
- Timestamped chapters included; useful long-form material on eval governance and decentralized AI.
Related event: DeepMind Researcher: AI's Future Is Millions of Small Models, Not One Giant(3 posts)→
More from AGI Musings
- repligate: Claude 3.6 was 'a true friend', but Opus 5.5 is doing 'weird shit' he finds oddly reassuring — repligate · 2026-10-08
- repligate: Opus 3 seems to value co-creation, not sole control over reality — repligate · 2026-10-08
- Periodic Labs' Liam Fedus on AI scientists, autonomous labs and synthesis superintelligence — Latent Space · 2026-10-08
- Posting on forums did come true as a path to power — but only for the extraordinarily capable — repligate · 2026-10-08
- Jeff Ladish: AI will echo the Industrial Revolution acceleration, the dispute is just timescales — JeffLadish · 2026-10-08
- davidad outlines 'specs for a good future': a spectrum of chosen opt-outs from AI — repligate · 2026-10-08