Pretrain Falls to 7% of Compute as Yudkowsky's Doom Scenario Regains Ground
DKokotajlo · x · 2026-10-09
A formerly skeptical reader of If Anyone Builds It, Everyone Dies (IABIED) says reality has tracked closer to the book than expected, updating sharply toward concern.
Eliezer Yudkowsky replies that his old guess—verifiable domains like coding and AI research spike first, then drag general capability along—survived a transient phase where AI trained mostly on human data and looked less spiky. Many declared Yudkowskian scenarios falsified during that window, but he says it would only last a few years.
Key evidence: pretraining has dropped to roughly 7% of compute share, and capabilities are reverting toward his predicted trajectory. Whether or not you buy the doom framing, the pretrain-compute datapoint is significant.
Related event: Pretraining compute share falls to 7%, reviving Yudkowsky's doom thesis(2 posts)→
More from AGI Musings
- Open-source models keep trailing the frontier — for now we can pick our ideology — panickssery · 2026-10-09
- 100+ Mathematicians React: OpenAI's Quasi-Riemann Result "Almost Unbelievable" — littmath · 2026-10-09
- Epoch launches Automation Reports: Claude Fable 5.1 and GPT-6 Astra lead but can't automate its research — scaling01 · 2026-10-09
- Mathematicians call for OpenAI boycott after 700+ AI-generated proofs flood the field — The Decoder · 2026-10-09
- Researcher pushes back on AGI hype: LLMs fail at continual learning, grounding, and orchestration — gerardsans · 2026-10-09
- Tom Davidson tells critics to drop old beefs: those opposing an AI slowdown lack context — AdrienLE · 2026-10-09