Google cofounder Sergey Brin says the company has started using Gemini to help build Gemini, a feedback loop the AI field calls recursive self-improvement, or RSI.
Brin said Google is "starting to get onto that RSI treadmill" during a roundtable on Gemini 4 Argon. Google shared the video itself. He pointed to the Flash models Google released before Gemini 4 as evidence. The quote in the excerpt we reviewed breaks off mid-sentence: "And now our researchers and the model work hand in hand to make [...]". We did not review the full remarks, so the excerpt includes no figures on how much of the work the model handles.
RSI matters because of what it implies. In AI safety debates, it describes systems that improve their own successors, with each round arriving faster than the last. A lab saying it has stepped onto that path is a notable claim, though this one comes from a company executive in a company-published video.
Brin's own wording is more modest than the concept. A treadmill suggests steady effort rather than takeoff, and "hand in hand" suggests researchers remain in the loop. A model that assists engineers is a different thing from one that directs its own upgrades, and the fragment we have points toward the first. Which of the two Google is describing, and how it would show the difference, is the question the full remarks would need to answer.