Ex-Anthropic Safety Lead: RSI Hype Ignores Amdahl's Law Human Bottlenecks
joshua_saxe · x · 2026-09-13
Joshua Saxe, former head of Anthropic's safety team, pushes back on the sensationalized RSI narrative circulating in policy circles:
- Amdahl's law math: removing humans might speed LLM architecture search 1000x, but if that's a small slice of the pipeline and the rest only improves 2x, total speedup is roughly 4x — comparable to earlier unlocks like RL/synthetic data replacing manual SFT.
- Some parts are decelerating: parameter count growth is slowing, making an idealized "foom" scenario doubtful.
- Epistemic asymmetry: only a handful of people are positioned to judge RSI, and they work in secrecy; meanwhile, those without inside knowledge are memeifying RSI as the central policy construct of Q3 2026.
He concedes plenty is being automated, but calls claims of breaking through Amdahl-bounded productivity gains an unearned claim.
More from AGI Musings
- Guardian podcast examines "AI psychosis" among chatbot true believers — nordicinst · 2026-09-13
- tszzl: models aren't autonomously producing research ideas — RSI is not here yet — i_dg23 · 2026-09-13
- Why pure commercial incentives will force frontier AI labs to slow down — scottleibrand · 2026-09-13
- Reddit essay warns RSI race without interpretability is AI's biggest existential risk — Glittering-Neck-2505 · 2026-09-13
- repligate Reposts 'Target Fixation' Wikipedia Entry in AI Context — repligate · 2026-09-13
- AI discourse wrongly assumes the field vanishes in five years, argues repligate — repligate · 2026-09-13