Are we approaching an intelligence explosion?

The central question is whether recent gains in AI capability signal an intelligence explosion: a process in which more capable AI helps produce further advances, so that each unit of research effort yields increasingly large returns. The authors argue that a faster capability trend over time is not enough to establish this. METR data indicate that measured capability growth roughly doubled after October 2025, but Anthropic’s annualized R&D spending rose about 6.2-fold and its technical workforce about 3.1-fold. On this accounting, capability gains per research dollar fell roughly 69%, and gains per technical employee-year fell about 38%. These figures are descriptive, not causal, and the capability measure is a benchmark-based index rather than a direct measure of intelligence.

To interpret the trend, the authors fit research-production models. A power-law model suggests that AI advances were becoming harder to achieve: capability growth would have slowed if research inputs had stayed fixed, while rising investment more than compensated. Allowing for a change around October 2025 points to a one-time improvement in research effectiveness, potentially associated with coding agents, rather than evidence of ongoing acceleration in returns. A more flexible model allows AI to contribute research effort itself - the mechanism needed for an intelligence explosion - but the available data do not show convincing evidence that AI is yet supplying enough of that input to overcome the growing difficulty of progress. Both approaches therefore find no current evidence of an intelligence explosion, while leaving open the possibility of one later.