Google has confirmed Gemini 4 Argon is now helping build the next version of Gemini — a milestone in recursive self-improvement (RSI), where AI systems assist in developing their own successors. Sergey Brin made the statement during a Google roundtable on 8 October, held in Washington, D.C., shortly after receiving the National Medal of Science. The video was published by Google on 17 October and reported by Business Insider and The Next Web. Brin said: “And now our researchers and the model work hand in hand to make the next version of the model.” He added Google is “starting to get onto that RSI treadmill”, citing Flash models released ahead of Argon as early evidence. Google DeepMind’s newly published safety and security report for Argon confirms AI agents now author a majority of production code and bug fixes at the unit, though all changes still require human review before submission. The report says 75%+ of production code at Google DeepMind is now AI-written. Fully delegated pull requests rose from 7.5% in April to 11.9% in September. But 88% still need repeated human steering or human-authored code. Humans remain essential for choosing research directions, designing experiments, and diagnosing training issues. Argon solves about half of Google’s internal test set of AI research tasks. Far short of full automation. Google tracks model progress via an internal index built from 92 benchmarks. Its monthly growth rate rose from 0.091 points before 1 April to 0.118 points after, a change the report calls not statistically significant. Google attributes Argon’s performance lift to increased investment and training compute, not AI-driven acceleration. Vihan Jain, co-lead of RSI and AI for science at Google DeepMind, wrote on X that Argon is “embedded in its own evolution”. He noted: “When your post-training demand shifts from asking for raw chips to asking for Argon quota, you know the RSI flywheel has reached a new gear.” A Gladstone AI survey of researchers at Anthropic, OpenAI and Google DeepMind found most expect RSI to begin within three to 12 months. Unless blocked by US government action. That possibility looms: last month, Congressman Ro Khanna announced plans to introduce legislation banning self-improving AI. OpenAI has urged the US to lead global standards instead. Google DeepMind’s report rates the risk from AI-driven AI R&D as “currently low”. It finds “no significant end-to-end acceleration of AI R&D, AI-driven or otherwise”. The report surveyed 12 senior leaders and researchers. Argon launched publicly this week. No pricing, regional availability, or consumer-facing features were disclosed in the sources.