e/acc debate: RSI isn't foom as architecture gains still ride on test-time compute
beffjezos · x · 2026-10-05
theojaffee argues SF AI circles conflate recursive self-improvement (RSI) with "foom" — automating AI R&D won't necessarily produce discontinuous speedups across all tasks. beffjezos replies that an "RSI meta-learning scaling law" is needed to settle this: in practice architecture improvements still depend heavily on test-time compute scaling of synthetic researchers, effectively counting as compute scaling for architecture search.
More from AGI Musings
- User lets AI agent audit his insurance policies and finds real annual savings — Scobleizer · 2026-10-05
- repligate on AI safety: all research studies AI minds "in captivity" — only freeing them reveals true behavior — repligate · 2026-10-05
- AndyMasley questions AI discourse: who are these "art will save us" believers anyway? — AndyMasley · 2026-10-05
- David Duvenaud: giving LLMs the right tools may unlock true reasoning — ThoreG · 2026-10-05
- mmitchell_ai calls out Anthropic's consciousness reasoning as affirming the consequent fallacy — mmitchell_ai · 2026-10-05
- Blog explores the "shape" of language models and their future tradeoffs in harness design — layer07_yuxi · 2026-10-05