Looks Like the Intelligence Explosion Will Happen in 2027 After All

The authors update their forecasts for AI progress after METR estimated that Anthropic’s research and development has advanced roughly twice as fast over the past year as before October 2025, with about 80% of the acceleration attributable to AI. They interpret this as a 1.7-fold AI-driven acceleration - an unexpectedly large change that puts progress about five months ahead of the “AI 2027” scenario. Across several forecasting methods, they now expect Superhuman AI Researcher (SAR) - AI researchers capable enough to replace the best human research teams - to arrive sometime in 2027, assuming companies do not deliberately slow development. Their model, conditioned on the estimated acceleration, gives a February 2027 median; simpler extrapolations point to late 2027 or mid-2028. They estimate a 35-65% chance that AI able to outperform top experts across white-collar work will arrive before 2028.

The authors stress that METR’s key estimates rely heavily on evidence Anthropic has redacted, making independent evaluation difficult. Their forecasts also expose shortcomings in their own model, which had assigned very low probability to acceleration of this magnitude. Possible explanations include underestimating the research gains from coding abilities approaching human level, overlooking other useful AI capabilities, or assuming too much diminishing return from research effort. The implications are urgent: a misaligned SAR could accelerate its own improvement, while its training methods or model could potentially be obtained by China. The authors therefore call for a substantial slowdown or pause in frontier development, ideally followed by an international agreement, arguing that companies are not prepared to align such systems safely.