Google Co-founder Sergey Brin: The company has begun entering the recursive self-improvement phase, models are co-developing the next version with researchers

Google co-founder Sergey Brin confirmed during the Gemini 4 Argon roundtable discussion that the company has begun entering the recursive self-improvement phase, with models co-developing the next version alongside researchers.

Brin pointed out that the team embarked on this path starting from Flash 3.6, 3.7, and 3.8, and now researchers are collaborating hand-in-hand with the models to iterate on the next version, with the pace continuing to accelerate.

He made it clear that by the time the outside world sees Gemini 4 Argon, the internal team will already be using models that are several generations ahead, and this pace is expected to only increase.

Google's security report shows that AI agents have written the majority of production code and bug fixes, although code changes still require manual review before submission.

Logan Kilpatrick previously stated that Gemini 3.5 to 3.8 were released in continuous intervals of three to four weeks, showing early signs of recursive self-improvement, and Gemini 4 is expected to further deliver on this.

In market mechanisms, the buyers are Google's internal research teams and computational resources, while the sellers are the current model capabilities and the progress of external competitors; the events are driven by model-assisted coding and iteration, with funds and computational power flowing towards internal advanced model training, benefiting laboratories that master the recursive flywheel, while putting pressure on competitors with slower release rhythms.

In internal testing, employees have been using a stronger version called Carbon, with some comparing its coding capabilities to Anthropic's Opus 5.5, and the training needs have shifted from requesting chips to requesting Argon quotas.

Source: Public information