Google CEO Pichai: We still have a long way to go before reaching the "final state" where AI can improve itself
Google CEO Sundar Pichai said that AI is still far from achieving the final form of recursive self-improvement (RSI); at present, AI agents only assist researchers in developing models with a certain degree of automation, and the technology is still in its infancy. He expects further progress next year. In April this year, Google disclosed that 75% of newly added internal code was already generated by AI, and Pichai said that this proportion can no longer accurately reflect the actual situation; in addition, the number of times Google rolled back software versions due to program errors or system failures decreased by one-third.
Body · AI translation
Google CEO Pichai: We still have a long way to go before the "final state" where AI can improve itself
2026-10-24 07:37 IT Home - Qingyuan
IT Home, October 24: According to Business Insider, Google CEO Sundar Pichai believes there is still a long way to go before AI can fully autonomously improve itself.
Pichai revealed that Google researchers have begun using AI agents to assist in model development, but the technology is still in its infancy. "We are still far from the final form of recursive self-improvement (RSI). To be precise, at every stage of model R&D, AI researchers have begun working in tandem with self-improvement loops. You could say that AI agents are assisting researchers in developing models together in a somewhat automated way."
Pichai expects further progress next year.
Google co-founder Sergey Brin said earlier this month that the company's Flash series models and Gemini 4 have already demonstrated the ability of AI to participate in its own development. Currently, researchers are working hand in hand with AI models to develop the next generation of models.
IT Home learned from the report that in April this year, Google disclosed that 75% of newly added internal code was generated by AI. But in Pichai's view, simply counting the proportion of AI-generated code can no longer accurately reflect the actual situation. "For us, that number is already meaningless, because every piece of code we process now is deeply influenced by AI."
Pichai also mentioned that Gemini 4 Argon will help people tackle more complex programming challenges. The number of times Google has had to restore software to older versions due to program errors or system failures has also dropped by one-third. "So I think AI tools have not only improved programming efficiency, but also improved code quality."