Google DeepMind Sees No Sign of AI Research Acceleration

Google DeepMind engineer Haydn Belfield recently revealed that the company has observed no measurable acceleration in AI research driven by automated systems. While AI now generates over 75 percent of Google's production code, human oversight remains vital. Around 88 percent of code modifications still require human review, and every single submission undergoes manual inspection. Meanwhile, the proportion of code changes fully delegated to AI has grown moderately, moving from 7.5 percent to 11.9 percent.

Performance tracking across complex engineering tasks illustrates the limits of current automation. In a study evaluating 1,389 internal runs of Google's Argon agent, success rates without human guidance dropped sharply as task difficulty increased. On DeepMind's RE-Bench—an internal benchmark consisting of 74 realistic machine learning engineering tasks—Argon completed roughly 50 percent of the tests. This represents a clear jump over Gemini 3.7 Flash at 27 percent, but remains far below the accuracy required for full autonomy.

To evaluate progress systematically, the team built a General Capability Index (GCI v1) covering 123 models and 92 benchmarks. The metric shows Gemini model capabilities expanding at a steady rate of about 0.1 GCI points per month, showing no sign of exponential speedup. Although Gemini 4 Argon placed slightly above this trend line, researchers attribute the shift to extra compute and organizational resources rather than self-improving code. Additionally, a survey of twelve senior DeepMind researchers confirmed that while predictable tasks automate well, humans remain indispensable for creative research, long-term planning, and designing novel experiments.

DeepMind currently rates the risk of automated AI research acceleration as low, though it is building measurement infrastructure to track future shifts. These empirical findings directly contrast with recent warnings from over 20 prominent computer scientists, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki. Those experts argued that an intelligence explosion could fully automate AI development within a few years, compressing years of work into months.