A new investigation published Thursday by the AI safety research organization METR reveals that Anthropic is actively using artificial intelligence to build more powerful AI, resulting in models improving roughly twice as quickly as they did a year ago. Following an 11-week probe inside the company, the report offers concrete evidence of AI accelerating its own development without immediate signs of slowing down or hitting diminishing returns. The findings arrive as Anthropic prepares to go public, presenting a striking juxtaposition with CEO Dario Amodei's public calls for AI labs to exercise caution and pump the brakes on development.
Key takeaways
- Anthropic's models are improving roughly twice as fast as a year ago, showing two years of historical progress in just twelve months.
- METR found that the majority of this acceleration is driven by AI assistance rather than simply scaling up computing power.
- Anthropic heavily redacted the public report to protect commercial secrets and prevent giving rivals a roadmap.
- The company has completed training on three unreleased models codenamed Model 2, Model 3, and Model 4 over the past six months.
How AI is accelerating internal research
The METR investigation indicates that the rapid progression inside Anthropic is primarily fueled by AI assisting human researchers rather than operating entirely independently, though researchers noted early signs that fully autonomous contributions could soon play a larger role. Internal Slack polls and employee interviews cited in the report show that staff using the company's most capable internal model reported being roughly twelve times as productive as they would have been without AI assistance. This marks a doubling of productivity gains compared to reports from just a few months prior when employees utilized Mythos 5. METR researchers confirmed they are confident the acceleration is not simply the result of scaling up the raw computing power used to train the models.
During the 11-week probe, METR identified several unreleased models that have pushed the company's research capabilities forward. Over the past six months, Anthropic finished training three models designated as Model 2, Model 3, and Model 4, which completed training in July, August to September, and September to October respectively. An older unreleased model dating back to 2022, called Unreleased 2022, was also identified by the investigators, demonstrating a swift pipeline of internal development.
Transparency concerns and extensive redactions
Although Anthropic provided METR with significant access to confidential research, performance data, and employee interviews that surpassed initial expectations, the final public report faced heavy censorship. Anthropic exercised its contractual right to redact major portions of the findings before publication. METR stated on social media that the company redacted the majority of the evidence found most informative, leaving the public to rely heavily on METR's independent judgment rather than verifiable data.
Anthropic cofounder Sam McCandlish defended the decision in a post on X, explaining that releasing the exact drivers of the speedup would compromise commercial sensitivity and hand competitors a roadmap to accelerate their own systems. Despite these justifications, the heavy redactions sparked intense criticism from safety researchers and outside observers. Critics argue that opacity undermines accountability, fuels accusations of hypocrisy given the company's public safety stance, and complicates efforts for proper democratic oversight in frontier artificial intelligence development.
Comments