AI Check · Updated October 18, 2026 · A recursive self improvement story with two versions. On October 14, 2026, Google cofounder Sergey Brin dropped into a developer roundtable and said the company was getting onto "that RSI treadmill" — models helping build their successors. DeepMind's own safety report, published around the same time, tells a more measured story. This check puts both side by side.
The quote heard around the AI world. Google had just introduced its new flagship model, Gemini 4 Argon, in a "Google for Developers" roundtable hosted by Logan Kilpatrick and posted to YouTube on October 14, 2026. About three and a half minutes in, cofounder Sergey Brin — who returned from retirement and is credited as a core Gemini contributor — dropped in unannounced. REPORTED · BI / BADIYAA / NEOTEO
Brin spoke for roughly 45 seconds, pointing to the run of smaller Flash models Google shipped before Gemini 4 as his evidence:
"And if you look at even before Gemini 4, when we had Flash 3.6, 3.7, 3.8, we're starting to get onto that RSI treadmill. And now the, you know, our researchers and the model work hand in hand to make the next version of the model."
He added that he had "popped in here randomly" because he and Melvin Johnson had been discussing the next model, that "probably by the time folks are viewing this, internally we're already using a couple models ahead," and closed with: "I expect it'll only continue to accelerate." The Flash models had iterated through versions 3.6, 3.7 and 3.8 in three-to-four-week cycles — pace that, Brin suggested, reflected early signs of models aiding their own evolution. REPORTED
Recursive self-improvement (RSI) is an AI system that helps design a more capable successor, which then helps design the next one. In theory, each turn of that loop runs faster than the last — and that is why the phrase makes AI safety researchers nervous. RSI is the threshold where humans stop being the fastest improvers in the loop, and the moment a capability jump could arrive faster than oversight can adapt. It is one of the oldest and most debated ideas in AI safety, and Brin just put Google publicly near its doorstep.
DeepMind's Gemini 4 Argon safety report describes a process with extensive AI help — and continuing human control. The measured numbers: MEASURED · DEEPMIND SAFETY REPORT
Fully delegated code changes — proposed, written and submitted by AI with no human authoring — rose from 7.5% in April 2026 to 11.9% in September 2026. That means roughly 88% of proposed changes still required human involvement: human-written code, repeated human steering, or both.
Argon completed about half the tasks in an internal AI R&D evaluation — below the level that would constitute full automation of research work. AI agents now author a majority of production code and bug fixes, but every code change is reviewed by people before it is integrated.
The analysis found no statistically significant acceleration in overall model progress attributable to AI assistance. Greater investment and compute were considered the likelier explanation for recent gains. Humans still select research directions, design experiments, and diagnose training failures.
Read the wording again: Brin did not say Gemini is improving itself without people. He described researchers and the model working "hand in hand" — a claim about speed and division of labor, not about autonomy. IMPORTANT CAVEAT
That distinction is the whole story. "Onto the treadmill" is not "running it alone." A 7.5% → 11.9% delegated share is a trendline, not a takeover. The honest read of both sources: AI assistance in AI research is real and growing, full automation is not here, and the measured evidence does not yet show the flywheel spinning on its own. A Gladstone AI report this week found most researchers across the major labs expect RSI to begin taking hold within three to twelve months — an expectation, not a measurement. REPORTED · GLADSTONE AI
Here is why the safety question survives the measured numbers. The delegated share climbed from 7.5% to 11.9% in five months. If that curve keeps climbing — and Brin says "it'll only continue to accelerate" — the question is not whether AI helps, but who reviews, who gates, and who can stop the loop when the delegated share reaches 30%, 50%, or more. Right now the answer is reassuring: human review gates every submission, humans pick the experiments, humans diagnose the failures. The thing to watch is whether the oversight scales at the same speed as the delegation. OUR TAKE
7.5% → 11.9% is the number to track in the next safety report. A continued climb is expected; a sudden jump would be the real news.
Writing code is one thing; completing whole R&D tasks is another. Argon is at ~half. Watch for a model that completes the loop end to end.
The critical safety metric is not how much the model does — it is whether human review, experiment selection, and failure diagnosis stay in human hands as the delegated share grows. If review ever becomes a rubber stamp, the treadmill is running on its own.
On October 14, 2026, in a Google for Developers roundtable introducing Gemini 4 Argon, Brin said Google was "starting to get onto that RSI treadmill," with researchers and the model working "hand in hand" to build the next version. He pointed to the rapid Flash 3.6/3.7/3.8 release cycle as early evidence and said internally Google was already "a couple models ahead." He did not claim the model improves itself without people.
Recursive self-improvement (RSI) is when an AI system helps design a more capable successor, which then helps design the next one. In theory each turn of the loop runs faster than the last, which is why the phrase makes AI safety researchers nervous: it could compress years of progress into months, with less time for oversight at each step.
Fully delegated code changes rose from 7.5% in April 2026 to 11.9% in September 2026 — meaning roughly 88% still required human involvement. Argon completed about half the tasks in an internal AI R&D evaluation, below the level of full automation. All code changes are still reviewed by people before integration, and humans still select experiments and diagnose training failures.
No. Analysis across 92 benchmarks found no statistically significant acceleration in overall model progress attributable to AI assistance; greater investment and compute were considered the likelier explanation for recent gains. This is the key gap between Brin's "treadmill" framing and the measured evidence.
Watch three signals: (1) the fully-delegated share of R&D work — 7.5% to 11.9% is the trendline to track; (2) whether models start completing whole R&D tasks rather than half of them; and (3) whether oversight keeps pace — who reviews, who gates, and who can stop the loop as the delegated share climbs.
Sources: Business Insider (Oct 15, 2026), Badiyaa, NeoTeo, webpronews, Times of India — reporting on the October 14, 2026 Google for Developers roundtable and DeepMind's Gemini 4 Argon safety report. Brin's remarks are quoted verbatim (46 words); all other figures are paraphrased from the cited reporting. "Recursive self improvement" is not yet volume-researched — no search-volume claims are made on this page.