Key points
- METR found Anthropic’s model improvements doubled in speed over the past year.
- Researchers attribute this acceleration primarily to AI automation, not compute scaling.
- Anthropic redacted parts of the report, limiting public insight into specific risks.
An independent safety review published by METR reveals that Anthropic is using artificial intelligence to accelerate the development of its own models. The report, released on Thursday, follows an 11-week investigation inside the company and suggests development pace has doubled compared with a year ago.reportedA safety group found that Anthropic is using AI to make its own technology better much faster than before.METR’s post-hoc analysis attributes Anthropic’s doubled model improvement rate to AI-assisted development rather than compute scaling.
Evidence: reported
Why a new report on Anthropic automating AI development is alarming researchersaifleet-verge-all
The evidence
Benchmark scores indicated this rapid improvement, with METR expressing confidence that the acceleration was not merely due to increased computing power. The safety group stated that the majority of this speed-up, as of October, stemmed from AI automation.
The decision
Anthropic permitted METR to redact portions of the final document as part of their agreement. The company appears to have used this right extensively, which obscures the full extent of the findings. This raises questions about how much the public truly knows regarding the pace and risks of frontier AI development.
What the evidence means
The findings provide clear evidence that AI systems are accelerating their own creation. This occurs while Anthropic CEO Dario Amodei has publicly urged labs to slow down AI development. The company is also preparing to go public, adding commercial pressure to these technical advancements.
What builders can do with it
Developers can observe that AI-assisted workflows are significantly compressing model improvement timelines. The documented shift from compute-driven gains to AI-driven automation suggests that future development cycles may rely less on hardware scaling. Builders should note that internal productivity gains, confirmed by employee interviews, are now a primary driver of capability increases.
Why it matters
- Policy readers must consider that accelerated development timelines reduce the window for regulatory impact assessments.
- The reliance on AI for AI development complicates safety verification, as automated processes are harder to audit manually.
- Extensive redactions in the METR report limit transparency, making it difficult to assess specific risks associated with rapid progress.
- The contrast between accelerated technical progress and executive calls for caution highlights a growing governance tension.
Topics: Anthropic, METR, AI safety, model development, automation
Updated
The strongest doubt
Counterpoint from the sources
Anthropic exercised its right to redact parts of the METR report, limiting the completeness of the public record.
For readers
Policy
- Accelerated development timelines challenge existing regulatory pacing.
- Redactions limit transparency into specific safety risks.
Researchers
- AI-driven automation is now the main factor in model improvement.
- Internal productivity gains are substantial and documented.
Investors
- Anthropic is preparing to go public amid rapid technical progress.
- CEO calls for slowing development contrast with actual acceleration.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Did METR confirm that AI is doubling Anthropic's improvement pace?
Yes, METR reported that Anthropic's models improved roughly twice as quickly as a year ago, primarily due to AI.
Was this acceleration caused by more computing power?
No, METR researchers were confident the acceleration was not simply due to scaling up computing power.
Can the public see all the findings in the METR report?
No, Anthropic was permitted to redact portions of the report, and it appears to have done so liberally.
Connected stories
Reading receipt
- METR found Anthropic’s model improvements doubled in speed over the past year.
- Researchers attribute this acceleration primarily to AI automation, not compute scaling.
- Anthropic redacted parts of the report, limiting public insight into specific risks.
Still open: How do the redacted sections specifically alter the risk assessment?
Still open: Will the pace of AI-driven development continue to accelerate without slowing?