AI Futures Project pulls its superhuman AI researcher forecast into 2027

The AI Futures Project has shortened its AI timelines, now saying a Superhuman AI Researcher (SAR) is likely to arrive in 2027. In a post dated November 10 (flagged by Techmeme), the group cites a METR report on research speed at Anthropic.

What METR measured at Anthropic

According to the AI Futures Project’s account, METR’s report appeared on November 5. METR judged that Anthropic’s AI R&D progress over the past year runs at about 2x its pace before October 2025, and said it would not be surprised by anything from 1.3x to 3x. It guessed that most of that acceleration comes from AI, which the authors translate into roughly a 1.7x AI-driven speedup.

The new forecast and how far it moved

The authors, led by Eli Lifland, Brendan Halstead and Daniel Kokotajlo, say they now expect a SAR in 2027 assuming no substantial deliberate slowdown. They give a 35-65% chance that Top-Expert-Dominating AI (TED-AI) arrives before 2028. Lifland’s own medians moved to August 2027 for a SAR and September 2028 for TED-AI, while Kokotajlo’s previous position was a 50% chance of a SAR before 2029.

The shift is large against the group’s own earlier work. They made their previous forecasts less than three months ago, and say even their most aggressive model-based distribution gave under 4% to a pace of 1.7x or more by this month. Compared with the AI 2027 scenario, which predicted a 2x R&D uplift in December 2026, they now see reality running about five months ahead.

A call built on evidence its authors cannot fully see

The post is openly critical of the evidence trail. It quotes METR as saying that Anthropic redacted a majority of the evidence METR found most informative, and the authors urge fuller disclosure. They also read the system card for Anthropic’s newest model as ambiguous about whether the company accepts METR’s estimate. The card says the model “does not double” the internal rate of overall progress, according to the post. Anthropic recently released Mythos 5.2 and Fable 5.2, priced at $8 per million input tokens and $40 per million output tokens, according to Techmeme.

The authors list their own caveats. The speedup may be confined to well-specified, benchmark-friendly tasks, which would lengthen timelines. Non-AI factors may explain more of the acceleration than assumed. A separate METR team’s preliminary numbers show much less speedup on a different measure, and the authors say such measures are sensitive to method. They also say the post was written quickly and may contain modeling errors, though they consider the high-level conclusions robust.

The group does not hide its stance. It argues that if a SAR arrives in 2027, it is urgent to pause or substantially pace frontier AI development, preferably through an industry-wide pause followed by an international agreement. It says companies are not on track to align such a system. Those are the authors’ policy views, not findings from METR.

For readers, the practical point is narrow: a well-known forecasting group has moved its estimate of AI automating AI research toward next year, based on one outside assessment of one lab. The estimate is not independently confirmed, and METR’s full report was not reviewed for this article; its figures here come from the AI Futures Project’s summary.