DeepMind Finds No Sign AI Is Speeding Up Its Own Progress

One of the louder fears in AI is that the technology will start improving itself, each generation building the next a little faster until the curve goes vertical. Google DeepMind has just published a safety report that, on its own evidence, finds nothing of the sort happening yet. The findings, released around 14 October alongside the lab's Gemini 4 Argon work, were summarised on X by DeepMind researcher Haydn Belfield and covered by The Decoder and others.

"There is currently no clear end-to-end acceleration in machine learning research and development, whether or not driven by AI," Belfield wrote. The team put the risk of AI-driven research acceleration at currently low. Its internal yardstick, an index it calls GCI, shows frontier capability climbing at a steady pace of roughly 0.1 points a month, with no statistically significant change in that rate since 1 April 2026.

Plenty of automation, not much lift-off

What makes the report interesting is that it is not arguing AI does little. The opposite. DeepMind says AI now writes more than 75 percent of its production code. The catch is in the next number: about 88 percent of code changes still pass through human review, and the share of changes handed off to AI with no human in the loop rose only from 7.5 percent to 11.9 percent between April and September. A lot of code, still a lot of supervision.

The autonomy drops off fast as the work gets harder. Across 1,389 internal agent runs, the success rate without human help fell sharply as task complexity rose. On RE-Bench, DeepMind's own test of research engineering, the Argon model solves about half of the 74 tasks. Useful, clearly, but a long way from a system that can run the lab by itself.

A pointed counterpoint

The report lands as a direct answer to the people warning that recursive self-improvement is already under way, a group that includes some of the field's most senior figures. It fits a pattern we have seen this month, with researchers pushing back on the idea that models can yet do original research on their own, and it cuts against the more dramatic timelines for runaway AI.

A caveat is in order. Almost everything here comes from DeepMind's own measurements and from Belfield's posts rather than an independently audited paper, and a company reporting that its own technology is not dangerously self-accelerating has an obvious interest in that conclusion. Still, the specific numbers, the 75 percent, the 88 percent, the steady monthly climb, show up consistently across the coverage. Taken at face value, they describe a field moving quickly and in a straight line, not one bending upward on its own.