Why a new report on Anthropic automating AI development is alarming researchers

A new report from the AI safety organization METR, based on an 11-week investigation inside Anthropic, found that the company's AI models are improving roughly twice as fast as a year ago, with little sign of slowing. METR concluded that most of this acceleration is due to AI itself rather than increased computing power, and that Anthropic employees using internal models self-reported being about 12 times as productive. The report also revealed that Anthropic has completed training three unreleased models and that the company extensively redacted the public version of METR's findings, citing commercial sensitivity and the risk of helping rivals. Anthropic cofounder Sam McCandlish defended the redactions as the "right call." The findings have provoked backlash among AI safety researchers, who see a contradiction between Anthropic's public calls for caution and its role in accelerating AI progress, and who warn that recursive self-improvement may be closer than expected. Critics, including METR and outside researchers, are calling for greater transparency and independent oversight of frontier AI labs, with some policymakers urging Congress to ban recursive self-improvement.

(Source：The Verge)