Three Letters Set the AI World Buzzing: Has Google Cracked RSI?

A few strange capital letters were enough to send the AI crowd into overdrive. Lyra, a well-connected leaker, posted “huge congRatulationS Indeed! @GoogleDeepMind,” and the highlighted letters spelled RSI, short for recursive self-improvement. Within hours the post had thousands of likes, X had pushed it into trends, and plenty of people were acting as if Google had already done the big thing. It hasn’t been proven. But the rumor landed on top of real movement inside Google, which is why people paid attention.

RSI, in the AI sense, means a system that improves its own code, training methods and architecture with little meaningful human help. The theory is simple enough to scare and excite people at the same time: one improvement makes the next one easier, and the pace starts to feed on itself. For many researchers, that is the threshold that matters on the road to AGI. But AGI and RSI are not the same thing, and a model that tweaks parts of a human-built system is still a long way from one that redesigns its own research agenda.

Google has been giving the rumor oxygen for months. Reuters reported in late summer that Sergey Brin is pushing the company’s AI resources toward self-improving systems and urging key staff to close the gap with frontier labs. Brin is also said to be steering Gemini from a Google microkitchen. Around a thousand researchers are reportedly working on related efforts. Then there’s the management reshuffle: Demis Hassabis is handing over operational leadership of DeepMind to focus on AGI, and chief scientist Jeff Dean is leaving. That is not exactly the backdrop of a company standing still.

Google has also talked about automation in plain sight. Fortune reported that the company credits the rapid cadence of its smaller Gemini Flash models partly to long-running AI-agent loops that recursively evaluate and refine the underlying models. Four Flash models in a little over a hundred days is a brisk pace, and it suggests a lot of machine help somewhere in the process. Even so, Google’s flagship Gemini 3.5 Pro still hasn’t shipped, despite Sundar Pichai having promised it for the summer, and internal prototypes reportedly didn’t show enough progress over Flash to justify launch.

The evidence for RSI, though, stays thin. There is no model, no paper, no benchmark, no date and no clear definition of what “reached RSI” even means. The one solid technical example people point to is AlphaEvolve, which improved a matrix multiplication kernel by around 23 percent and cut Gemini’s training time by about one percent. Useful. Impressive. Not self-bootstrapping intelligence. Until Google shows consecutive improvement cycles without human intervention, this is still a very loud rumor wrapped around a real trend.