Anthropic is using AI to build better AI: METR report sparks concern

Anthropic is successfully using artificial intelligence to build even more powerful AI systems, according to a new report published Thursday. Following an 11-week investigation by the AI safety research organization METR, findings show that the company's models are improving roughly twice as quickly as a year ago with no signs of slowing down.

• METR published a report Thursday after an 11-week investigation into Anthropic's operations.
• Anthropic's models are improving about twice as fast compared to the previous year.
• Employees utilizing internal models reported roughly a 12-fold increase in productivity.

Accelerated AI Progress and Unreleased Models METR's findings indicate that Anthropic achieved approximately two years of historical progress over the past year alone. Crucially, researchers are confident this acceleration was not driven solely by scaling up computing power, but rather by AI itself contributing significantly to the research process. Internal surveys showed that employees using Anthropic's top internal model were about 12 times as productive as they would have been without assistance. Furthermore, Anthropic completed training on three unreleased models—codenamed Model 2, Model 3, and Model 4—over the preceding six months. This rapid pace of development underscores a future where AI systems could soon improve themselves with minimal human oversight.

Heavy Redactions and Industry Backlash The report's publication faced notable transparency hurdles because Anthropic exercised its contractual right to redact large portions of METR's findings. The company argued that releasing these specifics would compromise commercial sensitivities and hand competitors a roadmap for faster advancement. Anthropic co-founder Sam McCandlish defended the redactions despite acknowledging they make independent verification difficult. The widespread redactions and the paradox of Anthropic publicly calling for slowed AI development while accelerating its own capabilities have drawn severe backlash and raised critical questions among AI safety researchers globally.