This article discusses a community METR study on AI R&D acceleration at Anthropic. I did not activity on this project or see the unredacted report.1 All opinions are my own and do not portray METR or the study authors. Thanks to Ryan Greenblatt for discussion, feedback, and assistance alongside modeling, and Daniel Kokotajlo, Eli Lifland, and Hjalmar Wjik for comments.
Earlier this week, Anthropic stated in their S1 that they “may imminently meet” the threshold of “dramatically accelerated advancement because of AI” as defined in their Responsible Scaling Policy. In their latest RSP (effective July 2026), say this threshold volition be met when:
They detect or anticipate general capabilities advancement to be going twice as accelerated as the fastest sustained charge before AI was contributing much to AI R&D.2
This is plausibly attributable to AI automating AI R&D, fairly than acceleration in compute or another inputs.3
For the final few months, several of my colleagues at METR have been operating on a deep-access inquiry to comprehend how much Anthropic is being accelerated by its AI systems. From my study of their report and Anthropic’s RSP, it looks akin the “dramatic acceleration” threshold has already been crossed.4
The study authors established that:
Between October 2025 and October 2026, capabilities advancement at Anthropic has been going concerning twice as accelerated as it had been complete the former dozen months. The measure they study is an inner type of the Epoch Capabilities Index called AECI; a broad range of another supporting evidence for acceleration has been redacted. They stated that this redacted evidence was "a bulk of the evidence [they] established most informative."
While the gait may have initially started picking up for a mix of reasons, AI automation is likely the ample bulk of the logic why capabilities are still improving at this faster gait today. Their reasons for believing this have been redacted.
This is a much bigger agreement than it power audio at first. I think this study is the most crucial update we have gotten concerning AI hazard in the final brace of years — much bigger than the Hugging Face attack. Naively, it points toward much shorter timelines than most external researchers (including the AI 2027 team) have projected.
Importantly, if AI is approximately doubling capabilities advancement overall, it must be speeding up individual researchers by much more than that.
Researchers activity on improving algorithms and architectures; Anthropic has sometimes referred to these innovations as “compute multipliers.”5 Overall AI capabilities (“effective compute”) is the merchandise of compute multipliers and training compute. Because the compute multipliers are lone one input into capabilities, this method that if AI capabilities are advancing twice as accelerated general since of application progress, the charge of growth in compute multipliers would need to be more than twice as fast: perchance ~2.5-3x faster according to sensible assumptions.6 Since there’s doubt concerning the size of speedup and how much of it comes from AI, I’ll go alongside ~2-3x.
Similarly, investigator labor is lone one input to discovering these compute multipliers: they need to test their ideas, which requires compute for experiments. If we assume a extremely uncomplicated Cobb-Douglas example anywhere test compute and labor are likewise important, that method the uplift to labor approximately needs to be the square of the acceleration in compute multipliers.7 This would connote that we would need a (~2-3x)^2 = ~4-9x speed up in labor to create general AI capabilities go twice as fast.8
However, formerly capabilities start speeding up significantly, this additionally method that low-hanging create is plucked at a correspondingly faster rate. Each new compute multiplier takes additional and additional activity to detect and execute (i.e., “ideas are getting harder to find”). Before there was much acceleration from AI, labor approximately grew at several approximately fixed exponential charge from hiring additional humans. Combined alongside growth in compute, this compensated for diminishing returns adequate to keep the baseline charge of capabilities progress. But to keep a new faster charge of capabilities advancement without scaling compute faster, we would need a new faster charge of growth in general labor.
Since the individual labor power is stuck expanding at additional or small the old rate, the uplift from AI would need to get larger and larger complete period fair to prolong capabilities advancement at its new rate. That method the uplift today, after multiple months of accelerated advancement from AI,9 needs to be equal higher than ~4-9x, the evaluation for what it would obtain to double advancement at the start: maybe the necessary labor uplift is ~10x. This is extremely uncertain, and substantially additional or small is plausible. But it lines up amazingly fine alongside informal self-reports from Anthropic staff, who say they are now getting a ~12x uplift from the use of inner models.
When you’re picturing a 10x uplift, you should not be imagining that all investigator is now a squad of ten researchers — as anyone who has always done item knows, a squad of ten does not get ten times as much precious activity done per hr as one person. Instead, it’s as if all investigator at Anthropic is in a period warp bubble anywhere they cognition ten days for all day that passes in the exterior world. As if AI lets them think, read, speak, type, and do everything else ten times faster.10 Anthropic likely wouldn’t have been capable to speed up its investigation this much by hiring any figure of people.11
A sensible default anticipation is that as AI capabilities continue to improve, AI researchers volition be sped up additional and additional from automation: if today it takes them a duration to get as much activity done as they used to in a year, shortly it volition obtain a week, and afterward perchance a day. If this expanding labor uplift is powerful adequate to conquer diminishing returns, afterward in the deficiency of another bottlenecks it would translate into equal larger speed ups in the charge of AI progress: advancement volition commencement going 4x faster, and afterward 8x, and afterward 16x, and so on.
The study authors were extremely uncertain concerning this question, and say that their image changed multiple times alongside new information. But ultimately, they did not discover evidence of bottlenecks or steep diminishing returns that would get in the way of compounding acceleration.12 They additionally study seeing several (confusing) redacted evidence indicating that acceleration could already be complete 2x and expanding complete time.
If we obtain it at visage value that AI is now nearly doubling the charge of AI capabilities advancement and assume acceleration from AI is smooth, this naively implies that we would attain extreme superintelligence inside a year. It seems that about a year’s value of AI advancement got us from humble acceleration to the item anywhere AIs now create AI advancement approximately twice as fast. If acceleration from AI is smoothly increasing, afterward at this new rate, we should anticipate that we volition twice the charge of advancement again inside six months. At that point, it would obtain three months to twice it again, afterward six weeks.13
In theory, these doubling times would add up to infinite capabilities inside one year. In reality, at several item we would reach the productive ceiling on application efficiency. But I anticipate these limits to be extremely far from anywhere we are now.
If we go through this feedback loop, I anticipate that the AI systems that arrive out the another flank volition be capable to extremely quickly create novel weapons and extremely effortlessly overpower individual civilization in a conflict. If these systems were earnestly misaligned, I think they would apt disempower humanity, and this could guide to individual extinction.
The “dramatically accelerated advancement because of AI” is specified an crucial and extreme milestone since it is a powerful tangible indicator of extreme capabilities in the extremely near term. Indeed, in an before type of Anthropic’s RSP (May 2025), the highest named hazard threshold for AI investigation automation (AI R&D-5) corresponded to a slightly weaker milestone.14 According to the old RSP, Anthropic would be required to have, at a minimum, safety fine adequate to “protect against model-weight theft by state-level adversaries”15 by the period it reached AI R&D-5.
I concur that extremely powerful security, the two for example weights and algorithmic secrets themselves, is a minimum prerequisite for undergoing an intellect detonation next twelvemonth without operating extremely ample risks of slaying billions of people. The AI industry is nowhere near this bar: the two OpenAI and Anthropic lately had important safety breaches by amateur groups. Beyond powerful example importance security, I additionally think we would need sturdy logging and monitoring alongside worldwide coverage, dependable measurements of monitorability, productive countermeasures against monitor collusion, ethical alignment training techniques alongside extremely convincing model organisms or interpretability to comprehend how fine they’re working, governance infrastructure for pacing progress if we notice these alignment and authority interventions commencement to interrupt down…and many another measures if we’re going to method an intellect detonation in item akin a liable way.
Unfortunately, most of the crucial evidence that fed into these conclusions has been redacted from this report. Analytical choices volition inevitably be fraught, and the study authors say that the evidence was frequently confusing.
I rely that my colleagues did a careful and rigorous job navigating making awareness of what they saw, and I made a big update according to this report. But I desire much additional of the underlying evidence was transparent so readers don’t have to depend on trust. We may be operating out of period to get on the identical leaf concerning timelines and takeoff speeds, and it is much harder to do that whenever external scientists cannot analyze the data and test out their own models.
