Google DeepMind Finds No Significant End-to-End Acceleration of ML R&D Yet, Though AI Writes Most of Its Code
In the Gemini 4 Argon Safety & Security Report, Google DeepMind published what researcher Haydn Belfield calls its first holistic assessment of ML R&D automation. It found 'no significant end-to-end acceleration of ML R&D,' even as most production code is more than 75% AI-written, according to Belfield's thread and Decrypt.
AI now writes most of the code at Google DeepMind, and the lab says it has not yet measured faster progress as a result. In the Gemini 4 Argon Safety & Security Report, DeepMind published what researcher Haydn Belfield called on X its "first holistic empirical assessment of ML R&D automation and acceleration."
The findings, as described
"There is currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise," Belfield wrote, according to Decrypt, adding: "We therefore conclude that the risk from AI-driven ML R&D is currently low." DeepMind's internal capability index, GCI v1, blends 123 models' scores across 92 benchmarks, and "frontier Gemini capabilities have improved at a remarkably steady pace of ~0.1 GCI points/month," he wrote. Argon sits above trend, which DeepMind traces to "increased compute and organizational investment."
Most production code at DeepMind is now more than 75% AI-written, and fully delegated pull requests rose from 7.5% to 11.9% between April and September, but about 88% of pull requests still need multi-turn human steering or human code authorship, Decrypt reports Belfield writing. A review of 1,389 Argon agent logs found that larger tasks were less likely to be completed without human help.
Decrypt notes the report covers only DeepMind's own lab and says it cannot rule out that an unrestricted Argon crosses its danger thresholds for cyberattacks and bioweapons information. We covered Argon's launch on September 30.
Limits
- The report itself was not opened; findings come from Belfield's thread and Decrypt's account of it.