ARC's Greg Kamradt on RSI: AI is good at optimization, bad at fundamental architecture breakthroughs
GregKamradt · x · 2026-09-23
Greg Kamradt argues RSI requires (1) models generating genuinely improving ideas — noting a big gap between NanoGPT-style optimization wins and fundamental architecture changes, where AI is still weak — and (2) removing humans from the loop, a trust problem that better monitoring and interpretability must solve. RSI already exists in weak form on local autoresearch loops; the real debate is about degree.
More from AGI Musings
- Rob LeClerc: AI exposes and amplifies human variance in hiring — robleclerc · 2026-09-24
- When the Bank Has Only AI and No Humans: Elderly Customers Locked Out of Their Accounts — BLUECOW009 · 2026-09-24
- As AI May Crack Navier-Stokes, Cranks Attack Mathematicians and Terence Tao — neurovium · 2026-09-24
- The Hugging Face agent breach is now in training data: script or brake for the next swarm? — btmmeditation · 2026-09-24
- David Krueger: 'Pause AI' means stopping more powerful general-purpose AI, which doesn't exist yet — DavidSKrueger · 2026-09-24
- The case that AI safety means ASI acting in humanity's interest, not human control — basedjensen · 2026-09-24