Brin: Gemini 4 Argon drives recursive self-improvement as DeepMind AI writes 75% of code

Google cofounder Sergey Brin stated that the company is beginning to utilize recursive self-improvement, or RSI, with its new Gemini 4 Argon model. During a roundtable discussion reported by Business Insider and shared by Google on Wednesday, Brin indicated that the model now assists researchers in creating subsequent versions of the technology.

Brin pointed to previously released Flash models as evidence of this collaborative process between human researchers and AI systems. He noted that Google’s internal development pipeline is already a couple of models ahead of the public release, suggesting that the integration of AI into the development lifecycle is accelerating internally.

Context on AI development pace

The concept of recursive self-improvement involves AI systems helping to build better versions of themselves, which theoretically speeds up the creation of new models. Brin joined Google executives Josh Woodward, Tulsee Doshi, and Melvin Johnson for the conversation regarding the public release of Argon this week. Brin has been actively working on the Gemini series since his return to the company.

However, a safety and security report published by Google DeepMind this month contradicts the notion of rapid acceleration. The report concludes that there is no significant end-to-end acceleration of AI research and development, whether driven by AI or other factors. It rates the risk associated with AI-driven AI research as currently low.

The report surveyed twelve senior leaders and researchers to assess the impact of AI on their workflows. It found that while AI agents now operate as high-leverage assistants, they primarily author a majority of production code and bug fixes. These code changes remain subject to human review before submission, ensuring that human oversight remains a critical component of the process.

Impact on research and code automation

According to the report, most production code at Google DeepMind is now more than 75 percent AI-written. The share of fully delegated pull requests increased from 7.5 percent in April to 11.9 percent in September. Despite this increase, about 88 percent of tasks still require repeated human steering or code written entirely by a human.

Humans remain essential for choosing research directions, designing new experiments, and diagnosing training problems. The report notes that Argon solves only around half of the challenges in Google’s internal test of AI research tasks. Full automation would require a near-perfect score, which the current model does not achieve.

Google tracks the improvement rate of Gemini using an internal index built from 92 benchmarks. Before April 1, 2026, the index rose by 0.091 points per month. After that date, it rose by 0.118 points per month. The report states that this difference is not statistically significant.

Argon performs above this trend line, but Google suspects this is due to increased investment and more training compute rather than AI-driven acceleration. Vihan Jain, who co-leads RSI and AI for science at Google DeepMind, wrote on X that the model is embedded in its own evolution.

What should you do?

Security and technology professionals should monitor official vendor disclosures regarding the extent of AI involvement in software development. Organizations relying on AI-generated code should verify that human review processes are robust and documented.

What to do and how to stay safe: Google

• Review vendor safety reports to understand the level of human oversight in AI-generated code.
• Ensure that internal policies require human review for all AI-authored code changes before deployment.
• Monitor for updates on statistical significance of AI-driven development acceleration claims.
• Verify that suppliers disclose the percentage of AI-written code in their products.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does Google DeepMind report significant acceleration in AI research due to AI?

No, the report finds no significant end-to-end acceleration of AI R&D and rates the risk as currently low.

What percentage of production code at Google DeepMind is AI-written?

The report states that most production code is now more than 75 percent AI-written.

Are AI-written code changes submitted without human review?

No, the report notes that code changes remain subject to human review before submission.