The AI Futures Project, the group behind the AI 2027 scenario, has shortened its timelines. In a Q3.5 update dated 10 November, all six authors now forecast that a “Superhuman AI Researcher,” or SAR, will likely arrive at some point in 2027, assuming no substantial deliberate slowdown. They also put a 35–65% chance on “Top-Expert-Dominating AI” — systems that can automate all white-collar work — arriving before 2028.
“Now, all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown).”
— Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden, in Q3.5 Timelines Update
A SAR, as they define it, is an AI that could fully automate coding at a frontier lab and match the research taste of that lab’s best human researcher — the ability to pick directions, choose experiments and interpret the results. In plainer terms, they say it is roughly the point at which a company would rather rely solely on automated researchers than keep a team of the best humans and no AIs. They open the post by warning that the modelling may contain errors, while arguing that those errors are very unlikely to undermine the high-level conclusions, and they ask to be corrected if they have made mistakes.
The shift is large by their own recent standards. In a mid-August update, Kokotajlo, whom the new post calls the most bullish of the group, had a median of December 2028 for SAR and had earlier assigned a 50% chance of it arriving before 2029. His all-things-considered median is now May 2027. Eli Lifland’s is August 2027, with a September 2028 median for top-expert-dominating systems. Those personal medians are conditional on companies not slowing development any more than they have in the recent past.
The authors say the single biggest source for the change is a 5 November assessment by METR of AI research and development at Anthropic. From 12 August to 29 October, Tom Cunningham, Nate Rush, Parker Whitfill and Hjalmar Wijk had high access to internal data, including benchmark series, speedup analyses and staff interviews. They conclude that, as of October, Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025. They attach high uncertainty to that figure, and say they would not be surprised if the true acceleration were closer to 1.3 times or 3 times the historical rate.
METR also guesses that a majority — “perhaps ≈80%” — of the current, October, acceleration is due to AI being used in the company’s own R&D, with non-AI factors likely playing a large role in late 2025. On the public numbers they can share, scores on AECI, Anthropic’s internal version of the Epoch Capabilities Index, have grown since October 2025 at an estimated 1.9 times the earlier rate, with a 95% confidence interval of 1.5 to 2.8. That, they say, implies about two years of historical progress in the past year, and they are confident the speed-up is not from a faster rise in training compute. Interviews and surveys are consistent with AI assistance driving the current acceleration, though staff say the most important contributions are not yet fully automated.
“Anthropic redacted a majority of the evidence we found most informative for these conclusions.”
— Tom Cunningham, Nate Rush, Parker Whitfill and Hjalmar Wijk, in METR’s Independent Assessment of AI R&D Acceleration at Anthropic
The AI Futures authors say those estimates, if right, imply a 1.7-times acceleration attributable to AI, relative to historical levels, and that this is about five months ahead of the original AI 2027 scenario. They stress that a 1.7-times lift in the overall pace of progress is a larger claim than researchers becoming 1.7 times more productive, because overall progress also depends on compute and on diminishing returns to extra research effort. Even Kokotajlo’s most aggressive earlier model-based distribution, they write, had assigned less than 4% probability to a speed-up of at least 1.7 times by November 2026. They also urge Anthropic to release enough information for outsiders to assess the redacted evidence; METR notes that the 80% attribution, unlike the AECI trend, has no comparable public chart behind it.
One of the simpler methods in the update extrapolates Anthropic’s R&D Automation Index, which rates work from AL0 (no AI involvement) to AL5 (the AI “operates fully autonomously, with no human in the loop”). As of August, Anthropic said Claude was not operating fully autonomously for any measured subset of AI R&D work, while it “led” 26% of that work — completing most of a task end-to-end from a high-level prompt, with a human supervising. The AI Futures authors’ best guess is that a SAR arrives when nearly all tasks are at AL5. Fitting a logistic curve, and assuming each level begins registering about seven months after the previous one, they get 95% of tasks at AL5 in November 2027. They call the method conservative because it does not model further speed-up from automation, and they flag high uncertainty about how long AL5 will lag behind AL4. The published chart notes that the AL5 curve is assumption-based: no task was rated AL5 in any month of the data.
In a post linking the update, Kokotajlo wrote that if METR is right, “that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario,” and added, “I am scared, sad, and angry.” The authors argue that a SAR in 2027 would make it urgent to pause or substantially “pace” frontier development, with an industry-wide pause followed by an international agreement as their ideal. They also say companies are not on track to align such a system. Those are their recommendations, not findings of the METR assessment itself, which found no clear bottlenecks that would prevent further acceleration in the next 12 months while warning that the reviewers may have missed some, and that much of the evidence they saw had more than one plausible reading.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
