← The feed Many Minded · The brief
KNOWLEDGE · forward · impact 3/5 · 2026-11-15

AI Futures Project now expects a superhuman researcher in 2027

On November 10 the team behind AI 2027 moved its forecast earlier. All six now expect a superhuman AI researcher sometime in 2027.

The AI Futures Project, the forecasting group behind the AI 2027 scenario, published a timelines update on November 10. All six of its forecasters - Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden - now expect a Superhuman AI Researcher to arrive sometime in 2027, assuming companies do not substantially slow down on purpose. Their individual midpoints fall between May and November 2027. Kokotajlo's is May 2027. Lifland's and Dean's are in August.

They define that milestone as an AI that matches the research judgment of the best people at a frontier lab and can do the coding on its own, roughly the point at which a company would rather rely on it than on the best human researchers working without AI. They also put the chance of "Top-Expert-Dominating AI," systems better than top human experts at nearly every task, arriving before 2028 at 35 to 65 percent.

The biggest reason for the shift was a November 5 report from the evaluation group METR. METR had spent eleven weeks looking at Anthropic's internal research materials, model results and staff interviews. Most of the staff on that review concluded that, as of October 2026, Anthropic was likely improving its models on well-specified tasks about twice as fast as it had between March 2024 and October 2025. METR guessed that about 80 percent of that pickup came from Anthropic's own use of AI. The AI Futures team reads that as AI itself speeding up AI research by about 1.7 times the historical pace. Our earlier account of the METR report is at manyminded.com/a/metr-says-ai-systems-are-now-feeding-back-4zlfo.

This is a forecast, and the evidence has real gaps. METR said it would not be surprised if the true speedup were closer to 1.3 times or 2.8 times the old rate, and it cited "substantial methodological uncertainty." It also said Anthropic redacted most of the evidence METR found most informative. In August, most of the same forecasters had placed their midpoints for top-expert-dominating AI in the 2030s. Even Kokotajlo's most aggressive model at the time gave that 1.7 times pace less than a 4 percent chance by November 2026.

If the new dates are roughly right, the same loop - AI systems helping to build the next ones - could compress years of research into a much shorter span, including work on energy, medicine and materials. That is not guaranteed. The range is wide, much of the underlying evidence is still hidden, and this group has moved its dates later and earlier before. What we can say now is narrower: a careful outside review found AI already contributing to faster AI research at one leading lab, and a team that studies those trends now thinks a researcher-level system is more likely next year than they did three months ago.

Source: AI Futures Project