Where's the "Intelligence Explosion"?

Ramez Naam argues that fully autonomous recursive self‑improvement (RSI) is unlikely to trigger a near‑term runaway “FOOM.” He lays out a five‑type taxonomy of RSI ranging from productivity gains to an accelerating, self‑sustaining loop, and points out that narrow superintelligence already exists in highly verifiable domains like chess, formal math proofs, and parts of coding. Using measured data, he claims the current AI self‑improvement feedback loop would need to be roughly 5-10× stronger to sustain itself and run away, and that commonly cited benchmarks and forecasts (METR, AI 2027, ECI extrapolations) overstate how well models perform on real, messy research tasks.

Naam backs this with internal company data showing a large gap between benchmark task horizons and real autonomous research performance: models reached about 80% success only on tasks roughly 15 minutes long (OpenAI reported ~86% success for sub‑15‑minute tasks), whereas METR and AI 2027 predict hours of human‑equivalent task length - differences of roughly 16× to 44×. Anthropic’s data shows models accelerating R&D work but no examples of fully autonomous completions. He emphasizes diminishing returns, exponentially rising resource requirements, and the increasing difficulty of new ideas, concluding that rapid AI progress is likely but not the kind of sudden, explosive takeover that would require a major conceptual breakthrough; better, more granular data is needed to reassess that conclusion.