Do verifiable metrics overestimate RSI?

The author argues that metrics like Anthropic's AECI may overestimate the rate of recursive self-improvement in AI capabilities by failing to distinguish between general ML research skills and narrow overfitting to verifiable tasks. While METR recently estimated a rapid acceleration in model capabilities driven largely by AI contributions, the author suggests this early progress could saturate and lead to subsequent deceleration rather than immediate artificial superintelligence. Open article →