Google’s Pichai Says Fully Self-Improving AI Remains Years Away as Models Help Build Successors

Google CEO Sundar Pichai has said the company remains far from achieving fully autonomous artificial intelligence capable of recursively improving itself, even as its researchers increasingly use AI agents to help develop more advanced models.

Speaking on the Big Technology Podcast this week, Pichai described an early stage of recursive self-improvement (RSI), in which AI systems assist human researchers with different aspects of model development. Rather than operating independently, the systems currently work alongside researchers, automating selected tasks and helping coordinate the process of building successive generations of models.

"We are pretty far from the end state of RSI," Pichai said. "The right way to describe it is, you have RSI loops along with AI researchers in various aspects of model development. I would say you almost have agents helping AI researchers, in some automated way, work together to build these models."

Pichai expects the process to advance further into next year, suggesting that AI's role in developing future systems will continue to expand. His comments offer a more measured assessment than the language used by some other figures at Google, who have described the company as beginning to enter a cycle in which AI helps build its own successors.

AI-assisted research and fully recursive self-improvement represent different levels of automation. Current systems can help write code, analyze results, and support researchers, but a genuinely self-improving system would need to do substantially more: improve its own capabilities, evaluate those improvements reliably, and use the results to develop subsequent versions with progressively less human intervention.

Google's progress suggests that parts of this feedback loop are already emerging. It does not establish that the company has reached a stage where AI can independently drive its own development.

Brin Sees Google Moving Towards an RSI Feedback Loop

Earlier this month, Google co-founder Sergey Brin said the company's Flash series of models and Gemini 4 were demonstrating how AI could help develop subsequent AI systems. He said Google was "starting to get onto that RSI treadmill."

"And now our researchers and the model work hand in hand to make the next version of the model," Brin said.

Brin has become increasingly involved in Google's Gemini programme and has pushed for resources to be directed towards projects he considers especially effective, according to Business Insider. His emphasis on recursive self-improvement reflects the growing importance of using AI to accelerate the research process itself, rather than relying exclusively on human engineers and scientists to build each new generation of models.

The potential benefits are substantial. Developing advanced AI systems requires extensive software engineering, experimentation, evaluation, and computational resources. If AI tools can automate more of these activities, researchers may be able to test more ideas, identify weaknesses faster, and improve models more rapidly.

That could also intensify competition among leading AI developers. Companies that use their existing models to accelerate research could gain an advantage if the resulting improvements translate into better systems, more efficient development, and faster product releases.

However, the process is not automatically self-sustaining. AI-generated code and research recommendations must still be tested, and improvements in one part of a model's development may introduce problems elsewhere. Researchers also need reliable methods to determine whether a new version represents a genuine improvement rather than a system that performs better on selected benchmarks while becoming less dependable in other settings.

Pichai's description indicates that Google continues to rely on human researchers within these loops. The company is automating elements of model development, but the human role remains central to coordinating and advancing the work.

A report published earlier this month by Gladstone AI offered a more immediate timeline for the next stage. Most researchers interviewed from companies including Anthropic and Google DeepMind reportedly believed recursive self-improvement could begin within the next three to 12 months, potentially sooner without government intervention.

That estimate represents researchers' expectations, not confirmation that fully autonomous self-improvement is imminent. The timeline also depends on how RSI is defined. AI systems already assist in developing other AI systems, but the point at which that assistance becomes sufficiently autonomous and capable of producing sustained improvements remains unsettled.

The difference between the forecasts highlights a broader uncertainty in the industry, where researchers now expect AI to accelerate AI development, but there is no established timetable for systems to take over the process independently.

AI-Generated Code is Changing Google's Engineering Workflow

Pichai also discussed the extent to which AI has become embedded in Google's software development process. The company said in April that AI generated 75% of its new code, but the CEO suggested that the figure no longer captures how extensively AI influences engineering work.

"That number in our world today is like meaningless in the sense that every part of code we now touch is influenced by AI deeply," he said.

His comments indicate that measuring AI adoption solely by the proportion of code generated by a model may understate its influence. Engineers can use AI to suggest changes, debug existing software, review implementations, and improve code written primarily by humans. In those cases, AI may influence the final product without generating the entire block of code.

Pichai said Google's Gemini 4 Argon would enable employees to tackle harder problems. He also reported a one-third decrease in code rollbacks, in which software is reverted to an earlier version following a bug or other system problem.

"And so I think people are writing code more productively, and better code with it," he said.

The reported decline in rollbacks offers a potentially important measure of engineering quality. If sustained across comparable workloads, fewer rollbacks could indicate that AI-assisted development is helping engineers produce more reliable software. However, the figure alone does not establish how much of the improvement resulted from AI, what period or projects it covers, or whether other changes to Google's engineering processes contributed.

The contrast between productivity and quality is becoming crucial as technology companies increase their use of coding assistants. Generating code faster does not necessarily reduce the total cost of software development if the output requires extensive review, testing, or correction. The commercial value lies in translating productivity gains into faster delivery, fewer defects, and lower overall engineering costs.

Google's comments indicate that the company is evaluating AI not simply as a tool for producing code, but as part of a broader development process in which researchers and engineers use models throughout their work.

Pichai also told the Sources podcast this week that he uses AI to track Google's model development and internal discussions, illustrating how the technology is being incorporated into management and information-gathering tasks as well as technical work.

Google's efforts point towards a potential shift in the economics of AI development. If models become increasingly capable of assisting with research, coding and evaluation, the cost and time required to build subsequent systems could decline, allowing developers to experiment more frequently.

But that possibility should not be confused with evidence that AI progress will accelerate without limit. More automated development could also make errors harder to detect, particularly if systems generate code, evaluate their own output, and use those evaluations to guide further changes. Reliable independent testing and human oversight would remain important as the process becomes more automated.

Google's immediate objective appears to be expanding the contribution AI makes to research while retaining human involvement in the development cycle. Brin's description of researchers and models working together and Pichai's more cautious account are consistent on the central point - AI is already helping build future AI systems, but the degree of autonomy remains limited.