Sergey Brin: Google AI models are already helping create their successors as the industry enters the era of recursive self-improvement

During a public discussion dedicated to the company's newest Gemini 4 Argon model, Google co-founder Sergey Brin stated that internal AI development processes at Google are already accelerating because the models themselves are helping create their successors faster. According to him, even before the release of Gemini 4, the Flash model lineup gave the company reason to believe that structural changes in the development process were already underway, and Google 'has started pulling into the RSI lane' of recursive self-improvement. 'And right now our researchers and the model itself are working hand in hand to create the next version of the model,' Brin emphasized. Google released the recording of the discussion last Wednesday, giving the statement the character of an official public address rather than a casual comment.

What is recursive self-improvement and why is it changing industry logic

The principle Brin is discussing is designated in the professional community by the acronym RSI (Recursive Self-Improvement). While in robotics the optimal production scheme is described by the phrase 'machines making machines', in AI model development a similar principle boils down to the fact that each new version of a model is created with the participation of previous versions. This generates a compounding rather than a linear pace of development: the better the model, the faster and better it helps create the next one. Brin emphasizes that this exact dynamic allows Google to remain among the industry leaders and significantly increase the pace of AI enhancement.

Competitors confirm: The RSI threshold is approaching for the entire industry

Google's statements are not isolated. Competitors Anthropic and OpenAI have also reported that their AI systems are already on the verge of transitioning to recursive self-improvement. According to a survey conducted by analytical firm Gladstone AI among developers, the majority of respondents — including representatives from all three mentioned companies — are convinced that RSI will become routine practice within the next 3 to 12 months. At the same time, analysts acknowledge that this scenario could materialize faster if government regulation in the AI sector does not force developers to slow down. Thus, this is not a single technological breakthrough by one company, but a systemic shift affecting the entire industry.

Practical implementation: AI agents are already writing primary code

In its own report on the Argon model, Google recorded that a number of AI research and development elements are already automated. Specifically, AI agents function as high-yield assistants, acting as authors of the main part of productive program code and fixes. At the same time, all model-suggested code changes undergo human review immediately before approval, which preserves the control function for developers. This means that RSI at Google is no longer a hypothetical concept, but a working process embedded in daily development.

Brin's personal role and rumors about the next Carbon model

According to Business Insider, Sergey Brin personally oversees the direction of Google Gemini's development, determining strategic priorities and insisting on greater concentration on the RSI principle. The company's co-founder is, by his own account, very pleased with the pace of progress: 'When ordinary people are just watching videos about Argon's release, Google developers might already be a couple of models ahead.' According to some reports, Google is already testing a new model under the working title Carbon, which in programming tasks may well compare to Anthropic Opus 5.5. These details have not yet been confirmed by an official announcement, but are consistent with the acceleration dynamics described by Brin.

What this means for the market and regulators

If RSI truly becomes the norm in the coming months, the logic of AI competition will fundamentally change: the advantage will shift not to the company with the most computing resources, but to the one that builds the closed-loop 'model creates model' cycle faster. For regulators, this means that the window for forming the rules of the game is narrowing: the higher the rate of self-improvement, the harder it is to track which model changes were made autonomously versus by human decision. Brin's statement that without regulation the industry will accelerate is effectively a public signal to regulatory bodies: either they intervene now, or they record an already established reality.