For decades, AI researchers have chased a loop. Use AI to build a better AI, let that better AI build an even better one faster, and keep going until nobody can predict what comes next. The idea is called recursive self-improvement, and it stayed a theory because AI able to do research and write its own code has existed for only a couple of years.
METR, a nonprofit that studies whether AI developers could lose control of their most advanced systems, spent eleven weeks inside Anthropic to find out whether the loop has started. Its assessment concludes that Anthropic is likely improving its models on clearly defined tasks at around twice the rate it managed between March 2024 and October 2025, roughly two years of progress in one. It puts perhaps 80% of that acceleration down to the company’s use of AI.
The company’s own figures point the same way. As of May 2026, Claude, its AI model, wrote more than 80% of the code merged into Anthropic’s codebase, according to a company essay. Before Claude Code, its coding tool, launched in February 2025, the share was in low single digits. Staff polls, which the assessment treats as loose evidence, put employees’ speedup from the newest internal model at 12 times.
Still, the loop is not closed. Researchers there told METR that AI speeds up their work mainly as an assistant, and cannot yet make fully autonomous contributions to the most important parts of training a model. They did describe “signs of life” suggesting such contributions could come soon.
A paper co-written by Geoffrey Hinton, often called the godfather of AI, models what happens once AI can do all the research. At today’s costs, one frontier developer’s computing power could run an AI workforce equal to at least millions of top human researchers. If returns on research held at historical levels and nothing else got in the way, progress would speed up tenfold within about 1.5 years of full automation, at which point a year of today’s progress would take about five weeks.
That speed could dramatically bring forward AI’s benefits, the authors write, but capabilities could also grow far beyond what society can keep up with, and humanity could lose control over superhuman AI systems. How soon either arrives depends on the step METR found missing, AI doing the most important research on its own.




