OpenAI's Millennium Prize breakthrough: compute scaling has ~4 OOMs left
sytelus · x · 2026-09-09
sytelus offers cold takes on AI's Millennium Prize Problem solution:
- Timeline: Buckmaster et al's headline results came out just a month ago, and these models take 2+ months to train — it's extremely unlikely OpenAI's model saw that recent work. Córdoba and Martínez-Zoroa's results were fully public on arXiv anyway; any agent swarm would explore them given compute.
- Scaling: The biggest takeaway is compute scaling keeps working for at least 4 more orders of magnitude, so the scaling party should last 2+ more years.
- Against cope: "They just threw compute" ignores that humans doing 100 years of math are also just throwing compute; "the proof is slop" ignores that math has no obligation to be human-readable — bounded human minds just need time to digest.
- Scope: A massive milestone worth celebrating unreservedly, but likely not parallelable in physics/bio, where wins came from verifiability + heavy RL on math rather than physical-world experimentation.
More from AGI Musings
- Terence Tao: identifying promising problems is now the scarce resource in the AI era — anshulkundaje · 2026-09-09
- Coders push back on 'superhuman AI soon': don't take digital-physical transducers for granted — jwt0625 · 2026-09-09
- Kimi paper said to 16x global compute sparks debate on US-China AI race — pstAsiatech · 2026-09-09
- Frontier labs quietly building recursive self-improvement, thread claims — 0xsachi · 2026-09-09
- Prediction: Big AI will start buying Big Pharma — not for the drugs, for the data — MannyKayy · 2026-09-09
- A magic lamp thought experiment on when AI solving math helps and hurts — sandersted · 2026-09-09