Google says self-improving AI is still far off, internal index shows no research acceleration

Google DeepMind engineer Haydn Belfield shared internal findings: AI has so far produced no measurable acceleration in AI research itself. AI now writes over 75% of production code, but about 88% of code changes still require human oversight or review, with a human gate before every commit; the share of code changes fully handed to AI rose from 7.5% to 11.9%. The more complex and time-consuming the task, the less likely AI is to complete it independently: across 1389 internal Argon agent runs, success without human intervention dropped sharply as complexity rose. A survey of twelve senior DeepMind researchers also found automation is highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation. On RE-Bench, an internal benchmark of 74 real ML engineering tasks, Argon solved only about half—a big jump over Gemini 3.7 Flash's 27%, but still far from the near-perfect performance full automation would require. The team also built a General Capability Index (GCI) covering 123 models and 92 benchmarks, showing Gemini's capabilities growing steadily at about 0.1 GCI points per month with no significant speedup; Gemini 4 Argon sitting slightly above the trend line is attributed to more compute and organizational investment, not AI-driven acceleration. For teams doing model evaluation and R&D workflow automation, this means keeping humans on creative research, new experiment design, diagnosing subtle training instabilities, and long-term planning in the short to medium term; Google currently rates the risk of AI-driven research automation as low, but is building measurement infrastructure to track whether acceleration signals emerge. This conflicts with the "intelligence explosion" view warned about by Hinton, Bengio and others.