Google CEO says fully self-improving AI is still years away

The limits of self-improvement

Google CEO Sundar Pichai recently addressed the concept of recursive self-improvement, often referred to as RSI. This is a process where an artificial intelligence system is capable of redesigning and improving its own underlying architecture without human intervention. Pichai noted that while the technology is moving forward, the company is still very far from reaching this final state.

The current reality involves a hybrid approach where AI agents work alongside human researchers. Instead of a closed loop where AI builds AI indefinitely, the process remains a collaborative effort. Pichai explained that these agents assist in various aspects of model development to make the process more efficient.

This distinction is important for understanding the current trajectory of the industry. While the goal of autonomous evolution is a major milestone, the current phase is defined by AI as a sophisticated tool for human experts. Pichai expects this continuum of progress to continue moving forward into the next year.

Internal coding and productivity

While full RSI is distant, Google is already heavily integrating AI into its internal software engineering workflows. The company previously disclosed that 75% of all new code written inside Google is now generated by AI. Pichai suggested that this specific percentage is almost meaningless because nearly every piece of code the company touches is now deeply influenced by these tools.

The impact of these tools is measurable in the quality and reliability of the software being produced. Pichai reported that Google has seen a one-third decrease in code rollbacks. A rollback occurs when software must be reverted to a previous version because of a bug or a system issue, indicating that AI is helping engineers write more stable code.

By using AI to handle repetitive tasks and suggest improvements, developers can focus on more complex problems. Pichai believes that these tools allow the team to be more productive while simultaneously improving the overall quality of the output. This represents a shift in how software is built rather than a shift in who is building it.

The RSI treadmill

Not all perspectives within the company are as cautious as Pichai's regarding the timeline for self-improvement. Google cofounder Sergey Brin has taken a more active role in the company's Gemini efforts and has pushed for a focus on RSI. Brin suggested that the company is already starting to get onto what he calls the RSI treadmill.

Brin pointed to the Flash series of models and Gemini 4 as evidence of AI helping to build itself. He described a scenario where researchers and models work hand in hand to create the next version of the technology. This suggests a push to accelerate the transition toward more automated development cycles within the company's most important projects.

Brin's influence has been significant in shifting resources toward projects he deems most effective. His involvement highlights a potential internal tension or diverse set of goals regarding how quickly Google should pursue autonomous model evolution. While the CEO emphasizes the distance to the end state, the cofounder highlights the progress already being made on the path toward it.

Industry expectations vs. reality

The internal debate at Google mirrors a broader conversation happening across the wider AI research community. A report from the firm Gladstone AI recently surveyed researchers from major labs including DeepMind and Anthropic. These experts provided a much more aggressive timeline for the arrival of recursive self-improvement.

The survey indicated that many researchers believe RSI could begin within the next three to 12 months. Some of these experts even suggested that this timeline could be shortened if there is no government intervention. This creates a stark contrast with Pichai's more conservative view that the technology is still in its early days.

The gap between these two views may depend on how one defines the 'start' of recursive self-improvement. If it is defined as any instance of AI assisting in its own development, the industry may be closer than Pichai suggests. If it requires a fully autonomous loop, the company's more cautious stance may be the more accurate reflection of current capabilities.

Monitoring the development

To manage these complex developments, Google is using AI to oversee its own internal progress. Pichai revealed that he uses AI tools to track the company's model development and monitor internal discussions. This adds another layer of automation to the company's operations, where AI helps manage the very process of creating new AI.

This oversight allows the company to maintain a bird's-eye view of how different models are performing and how researchers are interacting with them. It also helps in identifying bottlenecks or trends in the development cycle. By using AI to track these metrics, Google can better allocate resources to the projects that show the most promise.

As the company moves toward the next generation of models like Gemini 4 Argon, these management tools will become even more critical. The goal is to tackle harder problems by creating a more streamlined and data-driven development environment. While the end state of RSI is far off, the infrastructure to support it is being built every day.