METR Report: Anthropic Uses AI to Build AI, Doubling Speed and Raising Safety Alarms

When AI Starts Accelerating AI Development, Why Are Researchers Nervous?

After completing an 11-week investigation, AI safety research organization METR released a new report whose core conclusion points directly at a reality that is already unfolding: Anthropic is using AI to build more powerful AI, and the effect is quite significant—the company's models are improving about twice as fast as they were a year ago, with little sign of slowing down.

What makes researchers alarmed about this finding is not "AI writing code" itself, but that the doubling of improvement speed is occurring in a stage with extremely high safety requirements: the development process of frontier models itself. When automation begins to permeate the core loop of AI R&D, humans' ability to control the direction and pace of development becomes a key variable.

Behind the Doubling: The Flywheel of Automated R&D

According to METR's report, Anthropic's model improvement speed has roughly doubled compared with a year ago. This number needs to be understood in context: it does not describe a one-off leap in inference performance, but an overall acceleration in model iteration efficiency. In other words, within the same time window, the R&D loop turns faster.

This acceleration usually comes from automation across multiple stages of the R&D process—from code generation and experiment design to evaluation feedback. When AI is used to optimize AI's own training and development process, a self-reinforcing flywheel forms: stronger models bring more efficient R&D tools, and more efficient R&D tools in turn produce stronger models. METR's investigation tracked this process over 11 weeks and concluded that there is currently no sign of deceleration.

For Anthropic, this is undoubtedly good news; but for researchers concerned about AI safety, the increased speed of the flywheel means the time window left for safety evaluation, alignment verification, and external oversight is narrowing in tandem.

Rate of model improvement in the context of AI R&D usually refers to the magnitude of progress in model capability or performance per unit of time, which can be measured by benchmark score gains, improvements in training efficiency, or the length of the cycle from idea to deployment of a new version. What METR means by "doubling of improvement speed" is not that a single training run is faster, but that the output per unit time of the entire R&D loop—hypothesis generation, experimentation, evaluation, feedback—has increased. Understanding this distinction is important: gains from compute bring point acceleration, while acceleration of the R&D loop comes from process-level efficiency gains, which are harder to obtain directly by adding resources and also harder for outsiders to observe.

Researchers' Concern: The Race Between Speed and Oversight

The core logic behind the alarm raised by the report is not complicated: if the degree of automation in AI R&D continues to increase, then opportunities for "humans to make judgments at key nodes" will correspondingly decrease. A long-standing concern in the safety research community is that when R&D speed exceeds the speed of safety evaluation, risks accumulate without being fully tested.

As a research organization focused on AI safety, METR's value lies in providing observable empirical evidence rather than remaining at the level of hypothesis. The fact that model improvement speed has doubled turns "whether automated R&D is exceeding human oversight capacity" from a theoretical discussion into a question that requires empirical answers. The report does not conclude "stop," but it puts the speed issue on the table—in AI safety discussions, pace itself is a risk variable.

AI alignment refers to the field of technical research aimed at making AI systems' goals and behavior consistent with human intentions and values, covering everything from training methods (such as reinforcement learning from human feedback) to evaluation and red-teaming. The reason alignment verification receives more attention when automated R&D accelerates is that it relies heavily on human judgment: evaluators need to define what counts as "acceptable behavior" and design test scenarios that can expose problems. When model iteration cycles shorten, if the evaluation stage still runs at the original manual pace, it will become a bottleneck—or worse, be treated as a step that can be skipped. The essence of the speed problem is whether the manpower and time required for alignment verification can keep up with the frequency of model updates.

Industry Implications of Automated R&D

Zooming out from the Anthropic case, this report touches on a common trend across the entire frontier AI industry. Using AI to assist AI R&D is already common practice among leading labs; the difference lies only in how many stages are covered by automation and how deeply it is integrated. The particularity of the Anthropic case is that its speedup has been quantified and publicly discussed, providing an external reference point.

This also creates industry-level tension: competitive pressure encourages labs to advance automation as quickly as possible to gain iteration advantages, while safety governance requires explainability, auditability, and intervenability. The two are not inherently opposed, but when speed doubles, how to maintain the effectiveness of oversight mechanisms is a question every company needs to answer. METR's report is in effect a reminder that increased automation should be accompanied by a simultaneous increase in oversight capacity, rather than remedial action afterward.

Auditability in AI governance refers to the degree to which external or internal reviewers can trace the system development process, understand the basis for decisions, and review key results. It differs from explainability: explainability focuses on why a model produces a certain output internally, while auditability focuses on whether the R&D process leaves verifiable records—who made what changes and when, how experiments were designed, and whether evaluation results were truthfully reported. The higher the degree of automation, the more intermediate steps in the process are generated by AI, and the greater the challenge to auditability, because reviewers must judge not only the final model but also the code, experimental plans, and evaluation conclusions produced by AI. This is precisely what safety governance means by "oversight capacity should improve in tandem with automation."

Conclusion: Fast Is Not Necessarily Bad, But It Must Be Seen

METR's report does not engage in doomsday narrative; what it does is more modest: measure, record, and publish. The observation that model improvement speed has roughly doubled and shows no sign of slowing is itself worthy of serious attention from the industry. For Anthropic, the efficiency advantage of automated R&D is real; for the research community, the oversight gap brought by speed is equally real.

The real question is not whether AI is helping build stronger AI—that is already an established fact—but whether humans still retain enough leverage as the R&D flywheel spins faster and faster. The value of this report lies precisely in making this leverage question impossible to ignore.