Jay Kreps: the frontier AI race has a built-in self-stabilizing dynamic — weak models rebuild compute
olcan · x · 2026-10-28
Jay Kreps (Conflenger co-founder) argues against the common expectation that RSI will magnify small leads and let one lab run away with everything. The frontier AI race has a built-in self-stabilizing dynamic pushing the other way:
- If your model is spectacular, you sell out: you may have to cannibalize nearly all training and research compute just to serve inference and keep customers from rioting, plus divert much effort to inference efficiency — so your next model's quality may suffer a lot.
- If your model is bad, no one uses it: you're left with abundant idle inference compute, letting you scale faster and experiment more on the next run, potentially producing a much better model.
In meme form: big compute builds strong models → strong models lead to little compute → little compute leads to weak models → weak models lead back to big compute.
More from AGI Musings
- GEO: the KDD-2024-defined practice of writing content to be cited by AI answers — robertonappe · 2026-10-28
- AI skeptic's job-destruction worry overlooks countless newly solvable problems — robleclerc · 2026-10-28
- Grant agencies flooded with applications — and admins haven't connected the dots to AI — wandedob · 2026-10-28
- Frontier Models as CapEx, Open Models as OpEx: Where AI Token Spend Is Heading — curious_vii · 2026-10-28
- Chris Albon's thought experiment: how would AI history differ if another community found scaling? — chrisalbon · 2026-10-28
- AAAI/ICLR Reviewer: About Half of Reviews Look AI-Generated, Peer Review in Crisis — dsmirza · 2026-10-28