Sergey Brin Says Google’s AI Models Are Now Training Themselves

How Gemini Is Learning to Improve Itself

Google cofounder Sergey Brin revealed that the company’s Gemini AI models are beginning to assist in building their own successors during a roundtable discussion on Gemini 4 Argon. Speaking at the event, which was recorded and shared by Google on Wednesday, Brin described the development as the company „starting to get onto that RSI treadmill,” referencing recursive self-improvement in AI systems. He noted that earlier Flash models released prior to Gemini 4 provided the foundation for this progress. According to Brin, researchers and the AI models themselves are now collaborating directly to develop the next generation of the technology.

Brin explained that the breakthrough came from observing how the Flash models—precursors to Gemini 4—demonstrated capabilities that allowed them to contribute meaningfully to future model design. Rather than relying solely on human engineers, Google’s AI systems are now involved in tasks such as architecture search, training optimization, and performance evaluation. This shift marks a move toward AI-assisted AI development, where the model helps identify weaknesses and suggest improvements in its own codebase. Brin emphasized that this feedback loop is still in early stages but shows promising signs of accelerating innovation cycles.

What Does This Mean for the Future of AI Development?

The implication of models training their successors is significant for the pace and direction of AI advancement. If successful, this approach could reduce reliance on manual tuning and enable faster iteration cycles, potentially leading to more rapid progress in AI capabilities. Brin cautioned that the process remains experimental and requires careful oversight to avoid unintended behaviors or stagnation in learning. Still, he expressed confidence that combining human expertise with AI-driven insights will yield more robust and adaptable systems over time. The company plans to continue refining this method across future Gemini iterations.

What is the RSI treadmill in AI development? The RSI treadmill refers to recursive self-improvement, where an AI system enhances its own design or training process, enabling successive generations to become more capable without direct human intervention at every step.

Frequently Asked Questions

How are Flash models related to Gemini 4? Flash models were released by Google before Gemini 4 and demonstrated early abilities in assisting with model development tasks, serving as a proof of concept for the self-improving approach now being scaled in Gemini.

Is Google fully automating AI development now? No, Brin clarified that human researchers remain essential in guiding the process; the AI models currently assist rather than replace human judgment in building next-generation systems.