I wrote an explainer of METR’s report on Anthropic’s AI acceleration, largely confirming their findings. But with time to reflect and new evidence, I’ve become less and less comfortable with the position I endorsed. This is especially salient in light of the AI Futures Project’s Q3.5 update, which uses the METR findings to forecast that AI systems will be superhuman at AI research in 2027. They put a 35-65% chance on top-expert-dominating AI (TED-AI) – AI that matches or exceeds top human experts at virtually all cognitive tasks – by the end of 2027. In short, they predict superintelligence in 2027.
I don’t think this is likely.
1. Anthropic’s acceleration is only evidence about AI R&D automation, not superintelligence
A key assumption of the AI Futures Model – and in turn, of the AI safety community – is that the jump from superhuman AI researcher (SAR) to TED-AI is just another milestone of AI research. Under this assumption, any evidence of accelerating AI R&D automation is also evidence for superintelligence soon.
But people who have long timelines usually dispute this linkage! They argue that the transition from SAR to TED-AI is bottlenecked by external factors, like the ability to collect data in domains outside AI research. But the METR report only concerns progress towards AI automation of AI research – not whether that automation will lead to superintelligence. The link from SAR to TED-AI is the part that’s most cruxy for disagreements about superintelligence, and METR doesn’t have anything to say about that.
I find a fuzzy-task bottleneck generally plausible, but I haven’t thought about it nearly enough to wade into the debate. I might have a take later on that. But the important point is that the METR report is not evidence in favor of the SAR to TED-AI transition being as simple as it is in the AIFM. If you believed AI progress would hit a ceiling before generalizing to superintelligence, the METR report has no evidence that should change your mind about that.
2. Experimental compute is a likely driver of AI acceleration
METR was clear that faster scaling of training compute wasn’t a big part of the explanation for the acceleration. But they conspicuously did not rule out experimental compute as an important part of the explained AI progress. This is a major problem, because my calculations and others (e.g. Ajeya Cotra) calculated acceleration from AI uplift, assuming experimental compute stayed on trend.
But imagine experimental compute grew at twice its trend level. Ryan Greenblatt’s model predicts a 1.24x overall capability progress, holding everything else fixed, including AI uplift.1 Thus, experimental compute could make up substantially more of the explanation than METR acknowledges, and require much less AI labor uplift to rationalize their acceleration.
In fact, it’s extremely likely that experimental compute has grown above trend – because it’s optimal to scale up experimental compute if you’re seeing any AI uplift on research labor! When your researchers become more effective, you want to give them more experimental compute to make progress. It would be crazy if Anthropic saw researchers get drastically more productive due to AI, and decided not to scale its experimental compute above trend. Yet this is exactly what you need to assume if you’re going to attribute the vast majority of the speedup to AI uplift rather than experimental compute.
Thus, I expect experimental compute to be roughly as important as AI uplift in explaining Anthropic’s acceleration – which diminishes AI uplift as the majority explanation for the acceleration.
3. Anthropic’s AI labor uplift is implausibly high
METR describes informal Slack polls where staff used emoji reactions to report how much the latest model sped up their own work. Reported uplift reached 12x in October 2026 – i.e. researchers were able to do their work 12x faster because of AI. METR treats these polls as very loose evidence in their report, but both my explainer and Ajeya Cotra’s explainer used this figure to show how the estimated AI acceleration could be driven entirely by AI uplift.
In my earlier explainer, I showed plugging a 7x serial labor acceleration into Greenblatt’s model was sufficient to get the AI-attributable speedup to about 1.6x. But I’ve since thought it through, and realized a 7x serial labor acceleration has insane implications. If AI really made each researcher 7x faster, then Anthropic would rather fire 98% of its researchers than give up frontier AI models; alternatively, it would rather lose 91% of its experimental compute than give up frontier AI models.2 These are truly staggering numbers, and I suspect that if you asked Anthropic researchers whether they agreed with these statements, they would disagree.
Luckily, OpenAI actually asked researchers this question, and they estimate that the median researcher would give up 30% of team compute or 40% of headcount to use frontier AI rather than GPT 5.2. Under the same assumptions, those answers imply a serial labor acceleration of about 1.3x, dramatically lower than 7x.3
I see the intuitive appeal of a speedup like 7x. I wrote my review of the METR report in eight hours, and without AI I probably would have taken a full workweek – that’s a 5x speedup! But I also know that that was a specific burst of work, not a generalizable claim about all of my work. Still, I might be tempted to report a 5x uplift if I was asked about the productivity effect of AI.
There’s also a potential priming effect from asking people to report uplift in multiples. People might say they were 12x as productive, but I’d bet they wouldn’t say they were 1,100% more productive if asked to express their uplift as a percentage.
For more realistic uplifts like the ones implied by the OpenAI survey, the model predicts about 1.10-1.12x overall acceleration from AI labor uplift at the same level of accumulated research progress. I expect the true uplift was higher than that, but not that much higher.
4. DeepMind is still a frontier lab
SemiAnalysis famously argued that DeepMind was no longer a frontier lab, pointing to departures from its reinforcement learning teams and poor compute allocation. I also personally counted DeepMind out and considered AI to be a two-horse race, almost entirely based on the belief that better internal models would keep OpenAI and Anthropic ahead.
But now Gemini Carbon and Gemini Neon are slaying benchmarks, and Neon in particular is capturing the vibe on Twitter. I have yet to see a great explanation for how DeepMind caught up if its internal models were weaker than OpenAI’s and Anthropic’s! Notably, its research automation results look closer to OpenAI’s than to METR’s depiction of Anthropic’s acceleration. I see DeepMind’s return as evidence against AI uplift being a critical determinant of frontier AI progress today.
Maybe DeepMind is still behind on internal models? Of course, there’s probably still a gap. But six months ago, the public reference points were Mythos Preview, GPT-5.5 Pro, and Gemini 3.5 Flash. That’s a real hydrogen bomb versus coughing baby comparison. Now, at least it’s somewhat even. The fact that the public gap has closed is evidence against an RSI-driven lock-in story, even if the internal gap hasn’t closed.
5. People’s updates are suspiciously asymmetric
When the METR report came out, people posted about feeling physically ill about these results. Every METR researcher on my timeline was advocating for people to take 2027 superintelligence seriously. AIFP made their emergency update, and 80,000 Hours made an explainer of the METR report.
But I have not seen any of those people update their views in response to OpenAI’s much more pessimistic report about research acceleration. This sets off alarm bells to me! Why are people only taking the information on acceleration seriously?
You can make good arguments for why the OpenAI report is compatible with short timelines. But I’ve heard almost nothing about it from the people I know who updated on the METR report. This is really bad vibes, and suggests that people aren’t using all the information out there to make timelines forecasts.
Why does this matter?
I’m hardly the first to point this out, but it feels like we’re seeing a preference cascade towards believing in extremely short timelines. This ties the case for AI risk to extremely short timelines. AI will still be risky if it shows up two years later, and there’s a real trade-off to making your bets on it showing up in 2027 as opposed to 2029.
I should be clear that I am not a “long-timelines” person. My reasoning about AI futures is exactly as sophisticated as this: relative to 2022, we’re probably about 60% of the way towards superintelligence (for which I accept AIFP’s notion of top-expert-dominating-AI). Therefore, I expect superintelligence in 2029 by default.
Superintelligence in 2029 versus 2027 seems like a tiny difference in the grand scheme of things. But it’s absolutely enormous in terms of what we should prioritize. For example, should we prioritize work on the 2028 election? Should we drop everything and push for a pause right now? Should we prioritize building AI safety in China, which would only pay off on (slightly) longer timelines? I’m seeing a push towards deprioritizing “long-timelines” work (i.e. anything that pays off in more than 6 months) and I don’t think that’s a good idea.
In Greenblatt’s model, software R&D scales as R = L^0.55 * C^0.45, where L is research labor and C is experimental compute (both can be normalized relative to a baseline trend). With L unchanged and C doubled, R grows by 2^0.45 = 1.37x. Greenblatt attributes 2/3 of capability progress to software and 1/3 to training compute scaling, so this further attenuates overall progress to (2/3) * 1.37 + (1/3) * 1 = 1.24x. Thus, a doubling of experimental compute scaling above trend would predict a 1.24x acceleration in AI even if there was no AI uplift.
Both tradeoffs hold software R&D output fixed to see how Anthropic would trade off resources. For compute, Greenblatt’s model implies 7^0.55 * C^0.45 = 1. Solving gives C = 7^(-0.55/0.45) = 0.09, so the team could give up 91% of experimental compute. For headcount, I assume effective labor scales with the square root of headcount H due to parallelization frictions. Then at fixed compute, 7 * H^0.5 = 1, so H = 7^(-2) = 0.0204 and the team could give up 98% of headcount and still maintain the same output with a 7x serial acceleration.
OpenAI surveyed teams on what they would trade to restore access to coding agents better than GPT-5.2 for three months. Treating team compute as a proxy for experiment compute and a 30% reduction in compute meaning C = 0.7, how would L need to change in order to keep R fixed? Greenblatt’s model implies L^0.55 * 0.70^0.45 = 1, so L = 1.34. For the 40% headcount reduction, using the same square-root assumption as before gives L * 0.60^0.5 = 1, so L = 1.29.




The point on the OpenAI research acceleration report is particularly apt. Like it shows time horizons for 80% accuracy of ~15 minutes, that's way too low for such short timelines to make sense.
Re: why this matters.
I think waking people up to urgently working on this now rather than later is robustly good. it’s not as if we have a finite pool of resources allocating to short and long timelines. The short timeline people are most likely growing that pool the most.