Brin talks about Gemini 4: Google AI development accelerates and models embark on the road of self-improvement

Abstract:

According to "Business Insider",

Google may be approaching a critical moment in the field of AI: model self-improvement.

Brin

This concept is called recursive self-improvement (RSI), where AI helps develop more advanced AI systems, triggering a flywheel effect and forming an ever-accelerating cycle that drives the development of new models faster and faster.

Brin pointed to the Google Flash series of models launched before the release of Gemini 4 as evidence that this trend is taking place. He said the company was "beginning to embark on a cyclical path with RSI."

“Now, our researchers and models have begun working together to create the next version of the model.” He said during the roundtable discussion. Google shared the content of the discussion on Wednesday.

Others within Google appear to agree with Brin. Vihan Jain, the person in charge of the Gemini recursive self-improvement work at Google DeepMind, published an article on the X platform and talked about how deeply the Argon model plays a role in its own iterative evolution process.

He wrote: "When your post-training needs shift from applying for original chips to applying for Argon usage quotas, you know the RSI flywheel has shifted into a new gear."

Brin is becoming increasingly involved in the Google Gemini project and is using his influence as a co-founder to direct resources to projects he believes are most effective, including pushing the company to pay more attention to RSI.

In the roundtable discussion, Brin said that he was excited about the pace currently displayed by Google's teams, and said that by the time people saw the video about Argon's release, Google had actually advanced to several later models.

Anthropic and OpenAI also stated that AI may be about to enter the RSI stage. A new report released this week by AI company Gladstone AI explains how various AI companies are moving towards RSI. Most of the researchers interviewed for the report, from companies such as Anthropic, OpenAI and DeepMind, believe recursive self-improvement will begin within the next three to 12 months. Without government intervention, they said, the process might have started earlier.

In a security report released for the new Argon model, Google stated that some aspects of the AI ​​development cycle are being automated. The report states: "AI agents have now become assistants that can greatly improve efficiency, responsible for writing most of the code in the real operating environment and fixing vulnerabilities, but code changes still must be reviewed manually before submission."

However, RSI is at the heart of warnings issued by AI researchers last month. They believe that without adequate safeguards, humans could lose control of the technology.