Google says self-improving AI is still far away

Google DeepMind engineer Haydn Belfield says AI-driven research is not accelerating yet, despite systems taking on a growing share of software-development work. ¹

Belfield's figures, reported by The Decoder, say AI now writes more than 75 per cent of production code, while around 88 per cent of code changes still require human oversight or review. The share of changes fully delegated to AI rose from 7.5 per cent to 11.9 per cent. ¹

Automation still struggles with complexity

Across 1,389 internal Argon agent runs, success without human input fell sharply as tasks became more complex. Humans remained important for creative research, experimental design, diagnosing subtle training problems and long-term planning. ¹

On DeepMind's RE-Bench, which covers 74 realistic machine-learning engineering tasks, Argon solved about half of the tasks. That is ahead of Gemini 3.7 Flash's reported 27 per cent, but still far from the reliability needed for fully autonomous research. ¹

No measured acceleration in AI research

DeepMind's General Capability Index covers 123 models and 92 benchmarks. The index shows Gemini capabilities growing at roughly 0.1 points per month, with no statistically significant change in the rate of progress. The report attributes Gemini 4 Argon's position slightly above the trend line to more compute and organisational investment rather than AI-driven acceleration. ¹

Belfield's public comments are also linked to the relevant Gemini safety report. ²

Our opinion

The useful part of this update is the distinction between automation and acceleration. AI can already remove a lot of work from well-defined coding tasks without becoming a self-driving research department. That gap matters: impressive benchmark gains do not automatically translate into independent scientific judgement, and DeepMind's own measurements suggest the difficult, open-ended parts remain stubbornly human for now.