Google DeepMind has been pulled into a new round of RSI speculation after a pair of posts on X pointed to what some users took as a sign of recursive self-improvement inside Google.
The first post came from the X account lyra (@lyraxana) at 3:21 a.m. Taipei time on Sept. 12. The message read, 「huge congRatulationS Indeed!」 and tagged Google DeepMind. The unusual capitalization of R, S, and I formed the acronym for Recursive Self-Improvement, or RSI. According to the source report, the post drew more than 1,000 likes and quickly fueled discussion online.
Lyra later said Google revoked all GDM API keys
More than an hour after the acrostic-style post, lyra replied to a post by Google AI Studio product lead Logan Kilpatrick, saying there was no need to revoke all GDM API keys out of fear and that Google could have simply reached out directly, because the company had deeper infrastructure problems.
That claim has not been independently verified. As of publication, Google has not publicly responded, and DeepMind has not issued a statement.
Second post showed a model name in a JSON-style screenshot
Roughly three hours later, another X account, Lentils (@Lentils80), posted a screenshot showing what appeared to be a JSON response. In the image, one model was listed as models/rsi-model-liverl-le, with the display name RSI Model LiveRL LE.
Lentils framed the post hypothetically, writing that it would be crazy if Google DeepMind internally had something by that name and also had 10 exclusive training slots labeled from 00 through 09 under 「rsi-model-liverl-ns-xx」. The source report says only the first part appears in the screenshot. The reference to 10 training slots was not shown in the image and was added by the poster.
Format matches Google API structure, but the numbers are not unusual
The screenshot’s field structure lines up with the return format used by Google’s generative language API when listing models. It showed a maximum input size of 1,048,576 tokens and a maximum output size of 65,536. It also indicated support for generateContent and batch generation, with thinking set to true.
Those specifications, by themselves, do not establish what the model does. The report says lyra had posted another new Google model, omni-live, more than seven hours earlier, and that model carried the same 1 million input limit and the same 65k output limit. It adds that these figures are standard across the Gemini family, so they do not prove any direct link to RSI.
Google has been talking about RSI-related loops for months
The rumor spread quickly in part because Google had already laid groundwork for that interpretation. When Gemini 3.8 Flash was introduced on Sept. 2, Google said in its official materials that the model was accelerated by long-running AI agent loops that recursively evaluate and refine the underlying model.
On the same day, DeepMind researcher Shunyu Yao wrote on X: 「one small step for the model, one giant leap for RSI.」
Recent media reports added more context around internal priorities
Reuters reported in August that Google co-founder Sergey Brin had asked DeepMind employees to refocus on Gemini in an effort to catch up with rivals. One of the directions mentioned in that report was recursive self-improvement, described as using AI to help design, code, or train the next generation of itself. The source article also said Demis Hassabis moved into a chairman role during the same period, while Koray Kavukcuoglu took on broader operational authority.
Business Insider reported this week, after speaking with eight current and former employees, that Brin had been working from a converted break room in Mountain View and weighing in on model scale, release timing, and the path toward AGI. The report also said he pushed efforts to train Gemini by monitoring how employees write code. Google declined to comment on that report.
No official explanation for the model name
As of publication, neither Google nor DeepMind had explained the appearance of rsi-model-liverl-le. In the same discussion thread, the X account Token Gremlin pointed to a simpler reading: the leaker had not tested the model, had not seen it run, and did not know what RSI in the model name actually referred to. In that view, the market reaction was built largely around a string identifier and a suggestive acronym.
What can be stated from the available information is narrow. The screenshot posted by Lentils showed a model name and a display label, plus token limits of 1,048,576 for input and 65,536 for output. Google has not confirmed that the model exists, and the screenshot itself cannot be independently authenticated.