Google DeepMind engineer says AI is not yet speeding up AI research, even as it writes most production code

Google DeepMind has seen no measurable speedup in AI research driven by AI itself, according to Google engineer Haydn Belfield, as reported by The Decoder. Belfield said self-improving AI remains far off, though a lot of automation is already happening. AI now writes more than 75 percent of production code. About 88 percent of code changes still need human oversight or review, and a human reviews every change before it is submitted. The share of code changes fully delegated to AI rose from 7.5 to 11.9 percent. Across 1,389 internal agent runs with the Argon model, the success rate without human input fell sharply as tasks became more complex. Belfield said humans remain essential for creative research, designing new experiments, diagnosing subtle training instabilities and long-term planning. A survey of twelve senior DeepMind researchers found the same pattern: automation was highest for well-defined tasks whose results can be checked and lowest for open-ended scientific ideation. The figures give a rare inside look at one frontier lab. They are relevant to the debate over whether AI systems could soon speed up their own development, a concern behind recent safety proposals.