Anthropic has achieved notable acceleration in AI model development, with improvements occurring at roughly twice the pace observed a year prior, according to an investigation by the AI safety organization METR. The research, conducted over an 11-week period, suggests that the majority of this acceleration stems from AI assisting human researchers rather than independent AI development, though researchers identified emerging signs that fully autonomous contributions may increase in the near term. Employees using Anthropic’s most advanced internal systems reported productivity gains of approximately 12 times their baseline work rate, compared to roughly six times as recently as a few months earlier.
METR’s investigation revealed that Anthropic has completed training on three unreleased models in recent months, with development timelines spanning July through October. The organization noted finding no observable slowdowns or bottlenecks that might constrain continued acceleration. Crucially, researchers determined that scaling computational power alone cannot account for the observed improvements, indicating that methodological and technical innovations within the company have driven much of the progress.
The report’s findings come with substantial caveats regarding transparency. Anthropic retained the authority to redact sensitive information from the final public version, and exercised this power extensively. METR indicated that approximately half of its original 100-page assessment was redacted before release, with researchers emphasizing that much of the withheld material would have provided important context for public understanding. The organization stated that Anthropic justified the redactions on competitive grounds, though METR noted the redacted content was largely high-level in nature.
METR stressed the implications of these information restrictions, noting that the public must rely heavily on the organization’s own interpretation of findings rather than independently verifying the evidence. This limitation makes it difficult for external parties to fully assess the pace and trajectory of frontier AI development at one of the industry’s leading laboratories. Anthropic’s cofounder acknowledged the redaction decision complicated independent verification of the report’s conclusions.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI