Google CEO says fully self-improving AI remains years away despite recent advances

In a recent episode of the Big Technology Podcast, Sundar Pichai said Google’s AI systems are still far from the ultimate goal of recursive self-improvement, despite the growing role of AI agents in model development. He described current efforts as a partnership where agents assist researchers, but emphasized that true RSI loops are not yet realized. Sergey Brin, Google co-founder, echoed this sentiment, pointing to Gemini 4 and the Flash series as early signs of AI contributing to its own evolution, and noted the company is moving onto an "RSI treadmill."

A Gladstone AI survey of industry researchers, including those from Anthropic and DeepMind, indicated expectations that RSI could emerge within three to twelve months, possibly sooner without regulation. Pichai also highlighted that AI now touches virtually every line of new code at Google, with internal reports showing a roughly 33% drop in code rollbacks, suggesting higher productivity and quality. He added that AI tools are also used to monitor model development and internal discussions.