AI is dramatically accelerating AI research

An analysis of Anthropic’s internal AI-research progress argues that the company may already have crossed its threshold for “dramatically accelerated” AI development. A METR investigation reportedly found that capabilities advanced about twice as fast from October 2025 to October 2026 as in the preceding year, and that AI automation now explains most of the acceleration. Much of the supporting evidence is redacted, so the conclusion rests partly on trust in the investigators. The analysis reasons that doubling overall progress would require a far larger boost to research work: because software improvements and experimental compute both contribute to advances, it estimates AI may be multiplying researchers’ effective productivity by roughly ten, consistent with staff estimates of around twelvefold. The author stresses that this means researchers can accomplish far more per unit of time, not that adding people would produce the same gain.

The central concern is a feedback loop: more capable AI helps researchers make further advances, which could increase AI’s contribution and accelerate progress again. The investigation found no clear evidence that bottlenecks or diminishing returns would prevent this compounding, though the authors were uncertain and some evidence was confusing. If acceleration continued smoothly, a simple extrapolation suggests another doubling in six months, then progressively shorter intervals - potentially leading to extreme superintelligence within a year. The author treats this as a warning, not a reliable forecast, and argues that current safeguards are inadequate for such a trajectory. Needed protections include strong security, comprehensive monitoring, reliable alignment methods, and governance capable of slowing development. Redacted evidence also makes it difficult for outside researchers to independently assess timelines and risks.