Sergey Brin has put a name on what Google DeepMind has been hinting at for weeks. At a roundtable about Gemini 4 Argon, he said Google is "starting to get onto that RSI treadmill", with RSI meaning recursive self-improvement, according to Business Insider via AOL, dated October 15, 2026. The claim is real and attributed; the evidence offered is thin, and the company's own safety report points to humans still reviewing the work.
We could not open Business Insider directly, so the quotes below are from its syndicated copy. Where a detail comes only from a secondary summary, we say so. For the background of Brin's return to hands-on AI work, see our earlier post on Brin's Gemini 4 push, and for the definition see what recursive self-improvement means.
TL;DR: claimed vs. verified
| Claim | Source | Status |
|---|---|---|
| Brin: Google is "starting to get onto that RSI treadmill" | Business Insider (via AOL), roundtable shared Oct 14 | Reported; roundtable transcript not seen by us |
| "Our researchers and the model work hand in hand to make the next version" | Same | Reported quote |
| Google will be "a couple of models ahead internally" by the time people watch the Argon video | Same, paraphrase | Reported; unverifiable |
| AI agents author a majority of production code and bug fixes | Google's Argon safety report, as quoted by BI | Quoted from Google's own report |
| Code changes remain subject to human review before submission | Same report | Quoted from Google's own report |
| Some AI R&D is automated | Same report | Quoted; scope not specified |
| Researchers expect RSI within 3 to 12 months | Gladstone AI report, as cited | Survey of expectations, not a demonstration |
| Google is near autonomous RSI | Social media posts | Unverified rumor |
What did Brin actually say?
According to the article, Brin spoke at a roundtable about Gemini 4 Argon, which Google shared on Wednesday, October 14. He said AI development is accelerating internally and that models now help build their successors. As evidence he pointed to the Flash model releases that preceded Gemini 4 and said Google was "starting to get onto that RSI treadmill."
The most quoted line is: "And now our researchers and the model work hand in hand to make the next version of the model." He added that he was excited by the cadence teams were seeing and that by the time viewers watched the Argon release video, Google would be a couple of models ahead internally.
Note the wording. "Treadmill" and "hand in hand" describe a collaboration with people in the loop, not a system that improves itself unattended. That is a meaningful gap between the headline phrase and the sentence behind it.
What does Google's own safety report add?
The article quotes Google's Argon safety report in two places that matter. First, AI agents act as "high-leverage assistants" that author a majority of production code and bug fixes. Second, "code changes remain subject to human review before submission." The report also describes some parts of AI research and development as automated.
Taken together, the reading that fits the facts is: AI writes most of the code inside Google's AI work, humans review it, and some research steps are automated. That is significant for productivity and for safety, but it is a long way from a model rewriting its own training run. Our post on the Argon safety report covers its monitoring details.
What about Carbon and the "RSI flywheel"?
Business Insider reported earlier that staff are testing a stronger internal Gemini 4 version called Carbon, which some compared to Opus 5.5 for coding. We covered that report, and its limits, in a section of our Gemini 4 and internal coding skepticism post. Carbon is a codename from anonymous sources; no release date or pricing exists.
Vihan Jain, an RSI leader for Gemini at Google DeepMind, wrote on X that "the RSI flywheel has reached a new gear" once demand for Argon compute outpaced demand for raw chips, according to the article. That is an internal usage signal: researchers wanting model quota instead of GPUs suggests they use the model as a lab assistant. It is not independent evidence about the quality of the results.
Why is everyone talking about RSI now?
The article notes that Anthropic and OpenAI have also said AI may be near RSI, and cites a Gladstone AI report in which most interviewed researchers at Anthropic, OpenAI and DeepMind expect RSI to begin within 3 to 12 months, possibly sooner without government intervention. It also notes researchers' warnings last month that humans could lose control if guardrails are inadequate.
Reuters-based reporting, relayed in a Yahoo Finance summary, says Brin has used his influence as co-founder to push resources toward AI that can improve with less human intervention, after pressure from rival progress. Together these explain why the topic is spreading: major labs are describing it, and a founder is directing resources toward it.
It also helps to separate three meanings that get blurred:
- AI as a productivity tool for AI researchers. Models write code, run experiments, and summarize papers. This is happening now, and it is what the safety report describes.
- AI-driven research loops. Agents propose ideas, run training experiments and iterate, with humans checking. We covered a version of this in Dream-RSI from Google DeepMind.
- Autonomous recursive self-improvement. A system improves itself repeatedly without human steering. Nobody has shown this.
Brin's statement fits the first two. Social posts claiming DeepMind is "near autonomous RSI" assume the third.
What are people asking?
Is "RSI treadmill" a milestone or a metaphor? On the evidence, a metaphor for pace. A treadmill implies constant effort without a finish line, and the supporting sentence involves researchers working alongside the model. A milestone claim would come with a measurement, such as the fraction of a model's training improvements that were proposed and validated by earlier models.
Why would a company say this publicly? There are several plausible reasons, and the article does not say which applies. Talent and investor signaling is one; genuine excitement about internal tooling is another; competitive pressure after rivals' statements is a third. None of these require the claim to be false, and none make it verified.
Does automated code writing equal self-improvement? Only if the code being written is the code that makes the next model better, such as training pipelines, data filters, evaluation harnesses and inference kernels. A majority of production code and bug fixes could be mostly product and infrastructure work. The safety report excerpt does not break the share down.
Why does human review matter so much? It is the control. As long as a person approves each change, the loop has a gate, and the safety case rests on that gate staying meaningful. If reviewers become a rubber stamp because the volume of AI-written code is too high, the gate weakens even though the policy still says humans review. That is the concrete risk worth watching, and it is a common concern whenever generation is cheaper than checking.
What would change the picture? A published analysis showing model-proposed changes that improved benchmark results without human design, third-party replication on open systems, or a safety framework threshold triggered by automated AI R&D. Several labs' published frameworks already treat automated AI research as a capability to track, which makes disclosures from the labs themselves the most useful signal.
How to check claims like this
- Look for numbers. "A majority of production code" is a share; ask for the share, the review rate and the defect rate.
- Look for cadence. If development truly accelerates, release intervals and capability gaps between versions should shrink in public data.
- Look for third-party evals. Independent tests of AI research agents are the only way to separate marketing from capability.
- Read the safety report. The human review line is the most concrete statement of what is and is not autonomous.
- Discount rumor. Anonymous posts that a lab "could announce RSI anytime" lack any source.
What it means for builders
Nothing changes in your API today. But expect faster model turnover: if Google really is "a couple of models ahead" internally, public releases may arrive on shorter cycles. That favors abstractions over hard-coded model names, regression suites that run on every new version, and prompts that do not depend on one model's quirks. Our comparison of Gemini 4 Argon with Opus 5.5, Grok 4.7 and GPT-6 Astra is a baseline to re-run when the next version lands.
For teams worried about agents in production, the lesson from Google's own caveat is practical: even the lab saying its models help build their successors keeps human review in the loop. Do the same.
What we still do not know
- The full roundtable transcript beyond the quotes above.
- How "majority of production code" is measured, and over what period.
- How much of AI research and development is automated, and which steps.
- When Carbon or any later model will ship.
- Whether independent evaluators will be able to test claims of accelerated cadence.
Related reading
- Sergey Brin returns to a hands-on role for Gemini 4
- What is recursive self-improvement?
- Gemini 4 Argon safety report
- Gemini 4 and the internal Carbon build
- Dream-RSI: DeepMind teaches agents to improve their search
- Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
Based on Business Insider's October 15, 2026 report as syndicated; claims about Google's internal work are the company's and its executives', not independently verified.