Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All
The AI Futures Project updates its timelines after METR estimated Anthropic's AI progress is running about twice its historical pace, with roughly 80% of that speedup attributed to AI itself. They argue this pulls their forecast for a Superhuman AI Researcher into 2027 and makes pausing or slowing frontier development urgent.
- METR estimates Anthropic's AI R&D progress is about 2x its pre-October 2025 trend, with ~80% of the acceleration due to AI, implying roughly a 1.7x AI-driven speedup.
- The authors' models move their Superhuman AI Researcher median to around early-to-late 2027, several months ahead of their earlier AI 2027 scenario, which had put it later.
- Much of METR's evidence was redacted, so the 2x and 80% figures are hard to verify, and the authors say their own analysis is preliminary.
METR recently estimated that Anthropic's AI R&D progress is running about twice its historical rate, with a plausible range of 1.3x to 3x. They attributed "perhaps ≈80%" of that speedup to AI itself, which the authors of this post translate into roughly a 1.7x AI-attributable acceleration. If that holds, the AI 2027 scenario is about five months behind schedule in the wrong direction, meaning things are moving faster than that forecast expected. The authors say this surprised them: as recently as three months ago, even the most aggressive model in their suite gave less than 4% odds that progress would be sped up 1.7x by November 2026. They also stress that 1.7x AI-attributable acceleration is a bigger deal than it sounds, since it is far more than making researchers 1.7x more productive when you account for compounding effects.
The forecasting centers on two milestones. The first is the Superhuman AI Researcher (SAR), roughly the point where a company would rather fire its human researchers and rely on automated ones than keep the best human team with no AI help. The second is TED-AI, systems better than top human experts at nearly all tasks, including automating white-collar work. Across several simple extrapolations (research taste trends, self-reported labor uplift, Anthropic's own automation-level charts, and Tom Davidson's toy model), SAR lands somewhere between mid-2027 and mid-2028. Their filtered version of the AI Futures Model, conditioned on the METR numbers, pulls the SAR median forward to February 2027. They also flag a second data point, a roughly 1.2x acceleration in May 2026, which slows the forecast back down, so the picture is not one-directional. The post is candid that the modeling is preliminary and that the conclusions are not firm.
The policy conclusion is the sharpest part. The authors argue that if SAR arrives in 2027, the situation becomes urgent: companies are not on track to align such a system, recursive self-improvement with a misaligned SAR would be dangerous, and training one could expose the training recipe and possibly model weights to theft by China. Their preferred response is an industry-wide pause or substantial pacing of frontier development, ideally backed by an international agreement. The authors also invite corrections, noting that errors in the modeling are possible but, in their view, unlikely to change the high-level picture. The biggest uncertainty they identify is whether near-AC coding skills are driving more R&D uplift than their model assumed, which would make a steeper, faster takeoff more likely.