Google DeepMind Finds No Significant Acceleration in Machine-Learning R&D

Google DeepMind’s internal assessment found no significant end-to-end acceleration in machine-learning research and development, despite AI writing more than 75% of the lab’s production code. The result separates wider coding automation from faster progress across the full research process.

Google DeepMind researcher Haydn Belfield summarized the finding in an Oct. 14 thread: “There is currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise.”

The lab’s GCI v1 index combines scores for 123 AI models across 92 benchmarks. Frontier Gemini capabilities rose by about 0.1 GCI points per month, with no significant change in that pace after April 1, 2026. The index tracks model capability across benchmarks; the acceleration finding concerns whether AI speeds machine-learning R&D as a whole. The two measures answer different questions: whether model performance is improving and whether AI is speeding the full research process. Gemini 4 Argon scored above the index’s trend line, but DeepMind’s assessment linked that result to higher computing resources and organizational investment. Its uncertainty range extended to the trend, and Argon did not reach the Level 1 critical capability threshold for ML R&D acceleration.

AI writes more than 75% of DeepMind’s production code, but that share does not describe how much work is completed without human input. Fully delegated pull requests rose from 7.5% in April 2026 to 11.9% in Sept. 2026. About 88% still needed multiple rounds of human steering or code written by a person, and code changes remained subject to human review before submission.

A survey of 12 senior leaders and researchers found automation worked best on clearly specified tasks with verifiable outcomes. Weaker areas included open-ended scientific ideas, novel experiment design, diagnosing subtle training problems and long-range research judgment. The review also examined 1,389 logs of Gemini 4 Argon operating as an internal agent.

The same safety review could not rule out that Argon reached capability thresholds for biological information risk and cybersecurity, while assessing residual risk below those levels after safeguards.

The assessment addresses recursive self-improvement, the possibility that AI systems help develop more capable AI systems and potentially speed later development. It covers DeepMind’s own lab and does not establish whether research is accelerating at other companies.