Report: No 1M-GPU Vera Rubin run in 2027, but a 500K run next year; 15-25% compute earmarked for alignment
teortaxesTex · x · 2026-09-14
A resurfaced claim suggests the rumored 1M-GPU Vera Rubin training run may not happen in 2027, though a 500K-GPU run is expected next year. Additionally, 15%-25% of compute would be dedicated to model monitoring and alignment research. The commentary ties into Sam Altman's safety statement that "pacing" is not "stopping" — progress continues but should be slower than it otherwise could be — alongside OpenAI's welcome of a consistent federal frontier-AI safety framework with ideas like independent auditors.
More from AGI Musings
- Scott Alexander's pandemic prediction piece is being reread as an AI-risk manual — davidmanheim · 2026-09-14
- Keynes quote on post-scarcity purpose resurfaces amid AI economic transformation — paulnovosad · 2026-09-14
- As AI makes execution easier, taste becomes the scarce skill — alexmacgregor__ · 2026-09-14
- 99% of people talking about 'agentic workflows' have no idea what it means — prajdabre · 2026-09-14
- AI 2027 forecast: superhuman AI to outpace the Industrial Revolution, with two endings — Chris_Armstrong · 2026-09-14
- Anthropic signs up to $517 billion in compute deals in 11 months, Exponential View reports — Exponential View (Azeem Azhar) · 2026-09-14