Google cofounder Sergey Brin says the company's AI development has reached the point where its models are actively helping to build their own successors — a dynamic known as recursive self-improvement, or RSI. Speaking in a roundtable discussion about the creation of Google's new Gemini 4 model, Argon, Brin said AI development was accelerating internally and that the company 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 during the roundtable, which Google shared on Wednesday. His comments were first reported by Business Insider. The remarks put Google's cofounder — who has taken an increasingly hands-on role in the Gemini program — squarely inside the industry's most consequential debate: whether AI systems will soon improve themselves faster than human engineers can. For continued coverage of that debate, follow our AI research news.
What Brin Pointed To as Evidence
Brin pointed to the cadence of Google's Flash model releases leading up to Gemini 4 as evidence that the feedback loop is already turning. The implication is that each generation of models is shortening the time required to produce the next one, as automated coding, evaluation, and research tooling compound.
Google's own safety documentation for Argon appears to support parts of that picture. According to the model's safety report, "AI agents now operate as high-leverage assistants that author a majority of production code and bug fixes, though code changes remain subject to human review before submission." In other words, the company is describing a workflow in which models produce most of the code — while humans retain final say over what ships.
The Wider RSI Debate
Anthropic and OpenAI have also said that AI may be on the verge of RSI, and a report from the firm Gladstone AI published this week lays out the ways AI companies are moving toward it. According to Business Insider's account of the report, most researchers interviewed — working at companies including Anthropic, OpenAI, and Google DeepMind — believe RSI will begin within the next 3 to 12 months, and could arrive sooner in the absence of government intervention.
Why the Timeline Matters
Recursive self-improvement is not just another capability milestone. If models become the primary labor force for AI research itself, the pace of progress becomes harder to predict and harder to govern. That is why RSI sat at the heart of warnings from prominent AI researchers last month that humans could lose control of the technology if sufficient guardrails are not put in place.
Brin's framing was notably matter-of-fact. Rather than treating the RSI treadmill as a distant hypothesis, he described it as an operating condition of Google's model development program — researchers and models working "hand in hand" on each successive version.
The Counterpoints
Not everyone inside Google agrees on how fast the flywheel is spinning. Reporting in the same news cycle noted that Google DeepMind has found no significant acceleration in machine-learning research and development in its own evaluations, a conclusion that diverges from the more alarmed rhetoric coming from some rivals. The gap between those two readings — Google's cofounder describing a treadmill that has started moving, and Google's safety researchers measuring no dramatic speedup — is itself a snapshot of how uncertain the field remains about its own trajectory.
There are also structural caveats. Agents that author a majority of production code still operate under human review, and much of what counts as "AI automating AI research" today involves narrowing search spaces, generating candidate solutions, and running evaluations — high-leverage assistance rather than autonomous discovery.
What to Watch
Three things will tell observers whether Brin's treadmill metaphor is bearing out. First, the release cadence of Google's own models, where shorter intervals between generations would be the clearest visible signal. Second, whether other frontier labs begin quantifying how much of their research and engineering pipeline is model-executed, as Google's Argon safety report has done. Third, whether governments act on reports like Gladstone AI's, whose interviewed researchers explicitly conditioned their timeline on the absence of intervention.
The Economics Behind the Treadmill
There is also a straightforward commercial logic to Brin's comments. If models meaningfully shorten research cycles, the marginal cost of each new model generation falls, and the company that closes the loop first gains a compounding advantage in both capability and price. Google has already signaled willingness to compete aggressively on the second axis, undercutting rivals' API pricing with Argon while releasing Flash-class models for high-volume workloads. A genuine RSI flywheel would supercharge that strategy — which helps explain why Google has every incentive to describe it as already underway.
For now, the significance of the roundtable is positioning: one of Google's founders is publicly describing recursive self-improvement not as a risk scenario but as the company's working reality. That reframing — from thought experiment to roadmap — may prove to be the most newsworthy thing about the Gemini 4 launch cycle, regardless of where benchmark scores land.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
