Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All

Open original

The AI Futures Project argues that a METR estimate of Anthropic's AI research pace makes superhuman AI researchers likely by 2027, which they say calls for urgent pausing or pacing of frontier development. Much of the evidence behind the estimate was redacted, so the forecasts rest on partly unverifiable inputs.

  • METR estimates Anthropic's AI progress is about 2x its historical rate, with roughly 80% of that acceleration attributed to AI, implying about 1.7x AI-driven speedup.
  • The authors say this is far faster than their models predicted, and that the 1.7x figure matters more than it sounds because it is not the same as making staff 1.7x more productive.
  • Their forecasts put a Superhuman AI Researcher around 2027, and they argue this makes a pause or international agreement urgent.
  • Anthropic redacted most of the evidence METR found most informative, so the 2x and 80% numbers are hard to check independently.

The core claim is that METR's recent assessment of Anthropic implies AI is speeding up Anthropic's own AI research by about 1.7x, and that this pushes the timeline for a "Superhuman AI Researcher" (SAR) forward. METR estimated Anthropic's progress is roughly twice its historical pace, with perhaps 80% of that acceleration attributable to AI. The authors (the AI Futures Project team, including Daniel Kokotajlo) say this is well ahead of their prior model: Daniel, the most bullish of them, had put only about a 50% chance on SAR before 2029, and the new numbers would put SAR in 2027. They say the 1.7x figure surprised their model, which gave under 4% odds to that level of speedup in November 2026 even under its most aggressive settings. They also stress that METR redacted much of the evidence it found most informative, which makes the estimate harder to check from the outside. On forecasts, the team now expects SAR sometime in 2027 absent a deliberate slowdown, with a 35 to 65 percent chance that "Top-Expert-Dominating AI" (AIs better than top human experts at nearly all tasks, including white-collar work) arrives before 2028. These forecasts come from several simple extrapolations (research-taste trends, self-reported labor uplift, Anthropic's own automation-level charts, and Tom Davidson's toy model) plus a filtered version of their detailed model. Those methods land between mid-2027 and mid-2028. Conditioning on the 1.7x figure pulls the SAR median to about February 2027, though a lower 1.2x reading from May 2026 pushes it back somewhat. The authors are candid that they can't yet explain why their model underpredicted, and they list candidate causes: near-coding-automation skills giving more uplift than expected, non-coding skills like data generation contributing more, or diminishing returns being modeled too harshly. Each would point toward a faster takeoff. Because of that timeline, the authors argue an industry-wide pause or substantial pacing of frontier development, followed by an international agreement, is urgent. Their reasoning is that companies are not on track to align a SAR, that using a misaligned SAR for recursive self-improvement would be dangerous, and that training one soon could let China steal the recipe or weights. They openly invite corrections and say errors in the modeling are possible but unlikely to overturn the high-level conclusions.

Ask about this article