Where's the "Intelligence Explosion"?
Ramez Naam argues AI's recursive self-improvement loop is 5-10x too weak to sustain a runaway intelligence explosion despite rapid progress.
Noah Smith hosts a long analysis by futurist Ramez Naam arguing that a fast-takeoff 'FOOM' to superintelligence is unlikely because the current AI self-improvement feedback loop would need to be roughly 5-10x stronger to sustain itself. The piece distinguishes five types of self-improvement, notes evidence exists only for productivity gains (Type 1) and partial autonomy (Type 2), and cites METR task horizons, ECI scores, and OpenAI data showing diminishing returns. The debate matters because OpenAI, per Noam Brown, ranks AI doing AI research as its top priority.
32