METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July

AI development cycles are shrinking, meaning breakthroughs could occur faster than anticipated.

A nonprofit research organization, METR, has found that Anthropic's AI models are improving at a significantly accelerated rate, approximately doubling in speed compared to previous periods. This acceleration appears to be driven by Anthropic's own AI systems being used to aid in research and development, with estimates suggesting AI contributes around 80% to this speedup. METR gained eleven weeks of access to Anthropic's internal data, including benchmark results and training metrics, to conduct this audit. While Anthropic has a policy to flag such acceleration as a risk indicator, METR suggests the company may have crossed its own defined warning threshold back in July. A significant portion of the most informative evidence was redacted by Anthropic, citing confidentiality concerns that could benefit competitors. METR's audit also noted that Anthropic canceled a planned review of compute allocation between capability and safety research.

The findings suggest a potential feedback loop where AI rapidly improves AI, creating a scenario where the pace of advancement becomes unpredictable and difficult to control. This acceleration, if it continues unchecked, could lead to the "sudden emergence of extreme and risky capabilities" that present novel safety challenges. The redaction of key data by Anthropic also raises transparency concerns, complicating external verification of AI development safety claims and potentially hindering independent oversight of the industry's rapid progress.

This situation highlights the growing reliance of AI developers on AI itself for research, a trend that could democratize advanced AI development or conversely, concentrate power among those with the resources to implement such systems. The debate over transparency and auditing is critical, as it directly impacts the ability of researchers and regulators to assess and mitigate potential risks associated with increasingly powerful AI systems.