Google's Haydn Belfield: no measurable AI research acceleration yet

Haydn Belfield's findings come as Google Deepmind pushes automation deeper into its own workflow. AI now writes more than 75 percent of the lab's production code, a figure that would have been hard to imagine a few years ago. Yet the same internal data shows the limits: about 88 percent of code changes still need human oversight or review, and the share of changes fully delegated to AI rose only from 7.5 to 11.9 percent. That gap between drafting and delegating helps explain why Belfield sees no measurable acceleration in AI research itself. The pattern shows up in the agent data, too. Across 1,389 internal Argon agent runs, success without human input dropped as task complexity rose. A survey of twelve senior Deepmind researchers backed that up, with automation strongest on well-defined, verifiable work and weakest on open-ended scientific ideation. On the in-house RE-Bench, which includes 74 realistic ML engineering tasks, Argon solves about half, far better than Gemini 3.7 Flash at 27 percent but still far from what full automation would require. Rather than rely on impressions, Belfield's team built the General Capability Index covering 123 models and 92 benchmarks. Gemini capabilities are growing at about 0.1 GCI points per month, with no significant change in speed; Gemini 4 Argon sits slightly above the trend line, which the team attributes to more compute and organizational investment rather than AI-driven acceleration. The team rates the risk from AI-driven research automation as low for now but is building measurement infrastructure to track signs of acceleration. Belfield's view contrasts with a recent paper by more than 20 AI researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki, who warned of a possible 'intelligence explosion' within a few years.