Researchers debate whether AI training runs will eventually last more than a year
willdepue · x · 2026-07-29
Will Depue argues we may eventually see training runs lasting more than a year, but maybe not if RSI-style self-improvement takes off.
He suggests the real tradeoff is between waiting for more base-model progress versus starting earlier, since the upside of a newer base model for self-improvement could outweigh the cost of delay. A reply adds that current training runs still rarely exceed six months and that compute ramps make long pre-commitments look increasingly silly.
Related event: Why AI Model Training Cycles Struggle to Exceed Six Months(5 posts)→
More from AGI Musings
- Over 1,100 Frontier AI Employees Sign 'Call for Help' Letter Warning of Accelerating Risks — peterwildeford · 2026-07-30
- AI alignment may be optimized for what companies think users need, not what they need — StewartalsopIII · 2026-07-30
- ACT-R, Soar and Sigma are still a better way to think about mind architecture — lauriewired · 2026-07-30
- Aphasia patients still solved logic tasks, challenging language-only theories of reasoning — lauriewired · 2026-07-30
- The AI Era Developer Paradox: Addicted to Building Tools, Refusing to Use Them — AGI Hunt · 2026-07-29
- Opinion: Labs Will Offer Persistent Memory, But Hoard True Continual Learning — VraserX · 2026-07-29