Altman on automating AI R&D as researcher argues superfast recursive self-improvement won't be a sudden leap
SydSteyerhart · x · 2026-10-30
- Sam Altman says we know how to get models to find 5% compute efficiency wins, but nobody yet knows how to prompt a model to discover an architecture as significant as the transformer — though applying the same pipeline that made models good at math to AI research will probably work.
- @ArtemisConsort pushes back on the idea of extremely fast RSI: modern AI is an empirical science inside a formal system, more like biology than chess or math. Validating an architectural change requires implementing, retraining and deploying at great cost with fuzzy benchmarks, so recursive self-improvement is likely gradual rather than a sudden leap.
Related event: Altman on Automated AI R&D: 5% Gains Doable, Breakthroughs Unknown(2 posts)→
More from AGI Musings
- Diamandis: AI got 280x cheaper, yet the housing crisis persists — software scales, homes don't — PeterDiamandis · 2026-10-30
- François Fleuret: 'we've reached a point where I'd say... maybe' on AGI — francoisfleuret · 2026-10-30
- Zeb Anderson on AI access: as broad as possible, with different boundaries for children — sierracatalina · 2026-10-30
- VC takes: the biggest moats right now are philosophy, people skills, and defense tech — moonsandhues · 2026-10-30
- Is sycophantic AI inflating research publications? Two mechanisms worth questioning — KLdivergence · 2026-10-30
- DeepMind launches AGI economics research program with "The Economics of Intelligence" — soumitrashukla9 · 2026-10-30