Google DeepMind Engineer Says Self-Improving AI Is Far Off, Internal Index Shows No Acceleration

Google DeepMind engineer Haydn Belfield recently stated that there is no evidence that AI itself is accelerating AI research, countering claims that "self-improving AI" is imminent. This view contrasts with recent warnings from prominent researchers, including Geoffrey Hinton, about the risks of an "intelligence explosion."

According to Belfield, although AI now writes more than 75 percent of production code, approximately 88 percent of code changes still require human oversight or review, and every change is reviewed by a human before submission. The share of code changes fully delegated to AI rose only from 7.5 to 11.9 percent. In tests involving complex Argon agent runs, the success rate without human input dropped sharply as task complexity increased.

To track progress systematically, the DeepMind team built a General Capability Index (GCI) covering 123 models and 92 benchmarks. Data shows Gemini model capabilities are growing at a steady pace of about 0.1 GCI points per month, with no statistically significant acceleration. The team attributes Gemini 4 Argon sitting slightly above the trend line to more compute and organizational investment rather than AI-driven acceleration.

On DeepMind's internal RE-Bench benchmark, which includes 74 realistic ML engineering tasks, Argon solves about half of them. While this is a big jump from Gemini 3.7 Flash at 27 percent, it remains far from the near-perfect performance required for full automation. Belfield emphasized that humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities, and long-term planning.