Brin Says Google Entering Recursive Self-Improvement Phase

Google co-founder Sergey Brin said the company is beginning to enter a stage of recursive self-improvement (RSI), with the Gemini model increasingly helping develop its successor models. Brin made the remarks at a panel discussion on the development process of the new model Gemini 4 Argon, published on October 14, according to aitimes.com.

Brin said that by the time an introduction video for a new model is released, several more follow-up models are already in development internally. He cited the successive releases of the Gemini Flash models before Argon and said a cycle of using existing models for successor development has begun.

Panel participants said AI agents are used for monitoring large-scale training runs, detecting configuration errors and bugs, and checking for duplicate data. Google's published Argon safety report states that AI agents write most of the code and bug fixes used in actual services, while code changes undergo human review before being applied.

Google vice president Josh Woodward said professional writers edited the model's answers directly and this was used as reference data for training, focusing on reducing clichéd phrasing and unnecessary list-making. Participants also said token limits were increased so the model can reason sufficiently when solving difficult math or science research problems.

Earlier comments from Brin

In May of last year, at a Google I/O 2025 talk, Brin said he was confident Gemini would be the first AGI. In February of last year, he sent a message to employees urging them to win the AGI race and suggested working 60 hours a week. Last month, it was reported that Brin was using a small kitchen space at the Mountain View, California office as a work base alongside core developers.