Where's the "Intelligence Explosion"?

Ramez Naam presents a skeptical, data-driven case against the common idea that recursive self-improvement (RSI) will trigger a near-term “intelligence explosion.” He offers a five-level taxonomy of RSI, argues that meaningful gains so far are in productivity improvements for humans and models improving smaller models (Types 1-2), and notes narrow superintelligence already exists in highly verifiable domains (chess, Go, parts of formal math and coding). He cites striking examples such as claimed AI proofs in mathematics and strong internal uses of AI at companies, but stresses that these do not amount to broad, robust autonomous research or general judgment.

The central empirical claim is that the current self-improvement feedback loop is far too weak to sustain runaway growth: Naam estimates it would need to be roughly 5-10× stronger. Benchmarks and forecasts (METR, AI 2027) dramatically overstate autonomous capability compared with internal OpenAI data, which shows 80% success on research tasks only up to about 15 minutes of human-equivalent work - 16-44× shorter than some benchmark horizons. He highlights diminishing returns, exponentially rising resource costs, and the absence of observed acceleration as reasons runaway ASI is unlikely without a major conceptual breakthrough, and calls for better, more granular data to track real progress.