AI
Anthropic's AI accelerates its own development twice as fast as last year, alarming researchers
A new independent assessment reveals that AI is dramatically speeding up Anthropic's research capabilities, even as the company publicly advocates for slowing down AI development.
1/2
The Verge
An independent investigation into Anthropic's operations has uncovered something many feared: the company is using AI to build better AI at a startling pace. According to a report published by METR, a safety research organization, Anthropic's models are improving roughly twice as quickly as they were a year ago—equivalent to achieving two years of historical progress in just twelve months.
AI doing the heavy lifting#
The acceleration appears to be driven largely by AI itself rather than simply throwing more computing power at the problem. METR researchers found that employees using Anthropic's latest internal model reported being around 12 times more productive than they would have been without AI assistance—double the productivity gains reported just months earlier. Much of this boost stems from AI assisting human researchers directly, though METR cautioned that fully autonomous contributions to development are likely to expand significantly.
Crucially, METR found no obvious signs of diminishing returns or technical bottlenecks that might naturally pump the brakes on further acceleration.
Hidden models and restricted access#
Three of Anthropic's newest models remain unreleased to the public, with training completed between July and October 2024 under codenames Model 2, Model 3, and Model 4. Researchers at METR only became aware of these systems during their 11-week investigation.
While Anthropic shared more information than initially expected, the company maintained tight control over the assessment. Anthropic heavily redacted the final report, removing roughly half of METR's original 100-page analysis. The company justified these redactions on competitive grounds, arguing that revealing specifics could accelerate rivals' development. However, METR stressed that this opacity makes it harder for the public to independently verify findings or understand uncertainties.
Anthropic cofounder Sam McCandlish defended the redactions, stating the company determined "sharing exactly what is driving the speedup isn't the right thing to do, both because it's commercially sensitive and because it risks handing out a roadmap for others to go faster."

