DeepMind: No Measurable Acceleration of Research by AI

At DeepMind, automation is progressing within workflows, but the team does not observe a self-sustaining acceleration of research. Internal indicators show a steady evolution, far from the "explosion" mentioned by some researchers. Complex tasks and validation remain under human control.

Measured steady progress, low risk assessed by the team

To track the progress of its models, the DeepMind team has designed a General Capability Index that covers 123 models and 92 benchmarks. The results show that the capabilities of Gemini increase by about 0.1 GCI points per month, with no notable acceleration. The slight lead of Gemini 4 Argon over the trend is attributed to increased investment in computation and organization, rather than an acceleration driven by AI itself. The GCI v1 index indicates steady growth, without statistically significant acceleration. DeepMind currently assesses the risk associated with AI-driven research automation as low, while developing infrastructure to detect any potential signs of acceleration in the future. This position contrasts with that of more than twenty AI researchers, including Geoffrey Hinton, Yoshua Bengio, and Jakub Pachocki from OpenAI, who warn of a possible "intelligence explosion" and believe that AI systems could automate all AI research and development in the coming years, thereby significantly accelerating the pace of advancements.

Increasing code automation but under human control

At DeepMind, over 75% of production code is now generated by AI. Although this process is automated, nearly 88% of code changes still require validation or review by a person, and each modification is subject to human verification before being approved. The share of code modifications made entirely by AI has increased from 7.5% to 11.9%.

Agent capabilities: effective on structured tasks, limited on ideation

The AI agents developed by DeepMind see their performance decline as tasks become more complex. For 1,389 internal executions of the Argon agent, the proportion of success without human assistance drops significantly on difficult tasks. Haydn Belfield indicates that human presence remains essential for creative research, the development of novel experiments, the identification of subtle training instabilities, as well as for long-term planning. According to twelve senior researchers at DeepMind, automation works best for precise and verifiable tasks but is less effective for open scientific ideation. On the internal benchmark RE-Bench, which includes 74 realistic machine learning engineering tasks, the Argon agent manages to solve about half, while Gemini 3.7 Flash only succeeded in 27%. These performance levels remain below the near-perfect requirement necessary for total automation. According to Belfield, AI capable of self-improvement is not yet a reality, even though the progress in automation is notable.