Press release – AI is making AI smarter, faster, with no slowdown in sight

Press release – AI is making AI smarter, faster, with no slowdown in sight

METR, an independent AI capabilities evaluator, spent 11 weeks assessing whether AI is itself accelerating frontier AI R&D, what experts call “recursive self-improvement” (RSI), at frontier developer Anthropic. On 5 November, METR released a public report, which finds that AI is likely playing a key role in the acceleration of AI research.

While markets don’t seem to have grasped the implications of this groundbreaking publication yet, Ajeya Cotra, a METR researcher who was not involved in the report, says: “This is a much bigger deal than it might sound at first. I think this report is the most important update we have gotten about AI risk in the last couple of years — much bigger than the Hugging Face attack.”

In response, the global civil society federation PauseAI is intensifying its call for an urgent global pause agreement. Similarly, some US lawmakers (Congressman Greg Casar, New York State Assemblymember Alex Bores) are demanding a ban on RSI.

Speed and implications of RSI-powered AI development

The report estimates that Anthropic is now improving its models’ capabilities about 1.9x faster than it did between March 2024 and October 2025, possibly significantly faster. Across the whole industry frontier, the figure is about 1.4x.

PauseAI’s CEO, Maxime Fournes, warned: “Recursive self-improvement, which is AI being used for improving AI, is key to the threat model which top scientists, and the global PauseAI community, have been warning about for years, and it is arriving earlier than the vast majority of experts had predicted. METR, who evaluated Anthropic’s AI R&D acceleration, warns of “a significant risk” that this speed-up gets even faster, possibly superexponentially. We could lose control of AI, which will then outsmart humanity and could end life on Earth, possibly within months, not years.”

AI R&D progress cannot be monitored closely enough to be reliably “paced”

In a recent post, Anthropic describes a future “pacing effort”, in which AI companies would slow down and monitor frontier research instead of pausing it.
Maxime Fournes said: “If no one knows how to build AI safely, why build it at any pace? As AI accelerates its own development, it will soon be powerful enough to overpower us, or take down critical infrastructure, simply because a bad actor instructs it to do so or because an AI which the public doesn’t even know exists escapes onto the Internet and wreaks havoc.

The report was heavily redacted by Anthropic, including much of the evidence METR relied on to form conclusions. This means the world is already losing grip on AI that will soon be smarter than all of us. We are already in the dark about what Silicon Valley is growing in its labs, and how fast and powerful development is.

The fact that even impartial evaluators cannot obtain direct access to these systems and share their findings with the world shows that we cannot just hope to slow down AI research and monitor it. The report confirms that we are well on our way to full, irreversible loss of control. That’s why we need to pause, not pace, AI research: zero frontier training is a simpler verification target than some fraction of it.”

About PauseAI

PauseAI is a civil society organisation with chapters and groups in 20+ countries. We are concerned citizens from around the world coordinating peacefully to form a focused, united constituency demanding a global pause on the development of advanced AI until we know how to build it safely. Several PauseAI chapters and partners will be holding protests around the world this month; contact us for more details about them.

Contact
Raluca Spătaru, Country Lead, PauseAI România – contact@pauseai.ro / 0731.904.248

Key facts

  • Anthropic is likely improving its models at a rate approximately twice as fast as it did between March 2024 and October 2025. METR considers that the actual rate could be anywhere between 1.3x and 3x.
  • METR estimates that about 80% of the current acceleration is due to using AI to build AI. METR is convinced that training compute is not the cause.
  • METR found no clear bottlenecks to further acceleration over the next 12 months. Therefore, there is “a significant risk” of even faster acceleration.
  • Researchers report signs of fully autonomous AI contributions to AI development.
  • Ajeya Cotra notes in a blog post that if acceleration continues steadily, doubling times shorten (six months, three months, six weeks, etc.). This would lead toward extreme superintelligence within a year, before reaching limits that are likely very distant.
  • METR says that censored evidence from the confidential report, to which only Anthropic has access, makes such hyperbolic (continuously accelerating) growth “more plausible.”
  • Anthropic removed “most of the evidence” that METR considered most relevant. These include forecasts for future models and METR’s analysis of potential obstacles to further acceleration.
  • The investigation was also supposed to analyze how much of its compute Anthropic allocates to safety research versus capability development. Anthropic canceled this audit halfway through. Meanwhile, the lab’s september 17 post named compute among the most easily verifiable resources and a lever for timing dangerous AI research.
  • METR did not receive direct access to systems. METR states that it avoids descriptions of the process that “would discourage companies from participating in exploratory collaborations.” In other words, evaluators depend on the goodwill of the labs they evaluate to obtain any form of access.
  • METR changed its interpretation of key data several times as Anthropic provided more context. This shows how complex and uncertain this activity is.

Leave a comment

Your email address will not be published. Required fields are marked *