Google is deploying artificial intelligence in its AI model development process, but it drew a line at saying it is already entering the stage of recursive self-improvement (RSI).

Business Insider reported on Oct. 23 that Google CEO Sundar Pichai (순다르 피차이) appeared on the Big Technology podcast and said Google researchers are developing models with the help of AI agents.

Pichai said the current RSI loop in AI operates as a form of collaboration with researchers. He said it is a stage in which AI agents help researchers in an automated way to build models together.

He said he expects the same trend to continue next year and make progress. He said AI use inside Google is expanding into areas that research and develop next-generation models.

Earlier this month, Google co-founder Sergey Brin (세르게이 브린) also mentioned that the Flash family of models and Gemini 4 show an approach in which AI makes itself. He has recently been more deeply involved in Gemini development and has exercised influence to focus resources on projects judged to be more effective. That included a move to raise the priority of RSI-related work.

Industry forecasts differ on when self-improving AI will begin. A Gladstone AI report released earlier this month said many researchers at major companies, including Anthropic and DeepMind, saw RSI as potentially beginning within 3 to 12 months. It also included a view that the timing could come sooner if there is no government intervention. Pichai, however, put more weight on a collaboration system between researchers and agents than on full self-improvement as a description of Google’s current state.

Google said in April that 75 percent of new code was generated by AI, but Pichai said that figure alone makes it hard to explain the current situation. He said the number is meaningless to the point of being insignificant because most code handled by developers is influenced by AI.

Pichai cited Gemini 4 Argon and said developers will focus on harder problems. He said rollbacks, which revert code to previous versions due to bugs or system issues, have fallen by one-third inside Google. He said AI tools are helping people write code more productively and produce better-quality code.

He also said on another podcast, Sources, that he is using AI to grasp the status of Google’s model development and internal discussions. That means AI use is expanding beyond research and coding to support executive decision-making. The key question ahead is how far the autonomy of agents assisting researchers will rise and how much the gap will narrow between industry expectations for RSI and the actual pace of development.