Ramez Naam: seminal takeoff models don't match empirical data, multi-agent scaling gains far flatter than predicted
ramez · x · 2026-11-11
Ramez Naam credits the seminal takeoff-modeling papers (e.g., Eth & Davidson) but argues their models fail against empirical data, especially multi-agent scaling. An accompanying chart contrasts hypothetical RSI models (Chan et al. 2026: 2x labor → 2.6x research rate; Eth & Davidson 2025: 2x → 2x papers) with real-world intelligence gains that are far flatter — logarithmic on ARC-AGI/MolExpL, and Davidson-Houlen power law of just 0.6 with available compute (0.3 with fixed compute). His point: models deserve trust only insofar as they predict real-world phenomena, and many current models need updating.
More from AGI Musings
- AI safety now draws teen savants; practitioner argues against dropping out of college — austinc3301 · 2026-11-11
- "I asked for an AI to do my job—should've specified keeping me" — kuanhoong · 2026-11-11
- A safety-talent startup tried this and died on program length — austinc3301 · 2026-11-11
- Safety community's best spend: a 2-year institution teaching LLM internals — viemccoy · 2026-11-11
- Reddit take: $200k GDP per capita is the threshold for a post-labor AI world — deferare · 2026-11-11
- Who owns scientific progress under gradual disempowerment? — najoungkim · 2026-11-11