The Brief
How AI Works 4 min read

AI Accelerating AI: What the Anthropic Report Actually Reveals

NAVION

Share

A new independent report has surfaced evidence that AI systems are now meaningfully accelerating their own development, not as a theoretical future scenario, but as something already measurable inside one of the world’s leading AI laboratories. The findings come from METR, an AI safety research organization that spent eleven weeks investigating Anthropic’s research and development processes. What the report shows, and what it deliberately cannot show, is equally important.

Two Years of Progress in One: The Acceleration Finding

The headline number from METR’s investigation is striking. Anthropic’s models are improving roughly twice as quickly as they were a year ago. METR’s researchers described this as making “roughly two years of historical progress over the last year,” based on benchmark scores.

What makes this finding significant is the explanation behind it. Historically, AI progress has been driven largely by scaling: throwing more computing power at training. METR’s researchers said they are confident that is not the primary driver here. As of October, the majority of the acceleration is attributed to AI itself, and that contribution has been growing over time.

The evidence for this comes partly from internal surveys. Anthropic employees self-reported being around 12 times as productive using the company’s most capable internal model compared to working without AI assistance. That figure was roughly double what staff had reported just a few months earlier, when using an earlier internal model called Mythos 5. These are self-reported numbers, which carry inherent limitations, but the direction and magnitude of the shift is notable.

For now, METR’s assessment is that most of this acceleration comes from AI assisting human researchers rather than AI operating independently. The researchers did not find clear signs of diminishing returns or obvious bottlenecks. They also warned that “there are some signs that fully autonomous contributions will play a larger role soon.”

The Redaction Problem: What the Public Cannot See

Anthropic gave METR access to confidential research, model performance data, and employee interviews. METR acknowledged this exceeded what was initially agreed. That transparency, however, has a significant asterisk.

Anthropic retained the right to redact portions of the final report before publication, and it exercised that right extensively. METR submitted a report of approximately 100 pages. In a post on X, METR stated that Anthropic redacted “a majority of the evidence we found most informative.” The researchers noted that much of the redacted material could have “multiple plausible interpretations,” meaning the public cannot evaluate the evidence or the uncertainties surrounding it. As the report put it, the public is left relying on METR’s judgment “much more than would be ideal, given the weight of the questions answered here.”

Anthropic’s cofounder Sam McCandlish acknowledged the redactions in a public post, framing them as a deliberate trade-off. His stated reasoning: releasing the specific drivers of the acceleration would be commercially sensitive and could hand competitors a roadmap for moving faster, which he argued would make the risks worse, not better. Anthropic has said it plans to publish its own report on AI progress in the coming weeks.

The redaction issue matters beyond this specific report. Independent assessments carry reputational weight for the companies that commission or permit them. When those assessments are substantially redacted by the subject of the investigation, the independence of the finding becomes harder to evaluate. METR itself flagged this tension directly.

What Recursive Self-Improvement Actually Means for Everyone

This is what most coverage of the report underplays. The concept at the center of METR’s findings is recursive self-improvement: the idea that AI systems could autonomously develop increasingly capable successor systems, without meaningful human intervention, potentially compressing decades of technological change into years or even months.

Until recently, this was largely a theoretical concern. The METR report provides concrete, if partial, evidence that the early stages of this dynamic are already underway inside at least one major AI lab. Anthropic is not alone: OpenAI stated in September that it is “making strong progress” toward building an AI researcher capable of advancing deep learning and alignment research by March 2028.

The broader implication is about the pace of change relative to the pace of safety work. If AI systems are accelerating their own development faster than researchers can evaluate the risks, the gap between capability and understanding widens. METR found no observable bottlenecks that would naturally slow this down.

There is also a structural tension worth naming. Anthropic has been among the most vocal advocates for responsible AI development, and its CEO Dario Amodei has publicly called for labs to slow down. The METR report describes a company that is, at the same time, using AI to move faster than ever. That is not necessarily hypocrisy: it may reflect a genuine belief that being at the frontier is safer than ceding it. But it is a tension the public deserves to understand clearly.

In Short

AI is already accelerating its own development in measurable ways, not just in theory. METR’s eleven-week investigation inside Anthropic found that the company’s models are improving roughly twice as fast as a year ago, with AI itself identified as the primary driver. The full evidence behind that finding has been substantially redacted. What remains visible is enough to raise serious questions about whether the field’s safety frameworks are keeping pace with the technology they are meant to govern.

Based on reporting from The Verge.

Written by

NAVION