True RSI Isn't Here — Stop Using the Word Where Outsiders Can Hear It

True RSI Isn't Here: Stop Using the Word Loosely · Jason · 2026-09-20

Over the past few weeks, RSI has been bouncing hard through English-language tech Twitter and the trade press. Some people say the labs already have it. Some talk like the singularity is parked outside. Others read Recursive Self-Improvement in a Google paper title and hear: the model is upgrading itself in a locked room. To someone outside the industry, that sounds like artificial intelligence already making children on its own. My bar for true RSI is simple. The model iterates without people in the decision loop. Humans, at most, keep it in a sandbox for a while and can still pull the plug. By that bar, anyone saying "we already have RSI" is early. Especially when they say it where outsiders can hear. Outsiders do not have time to parse the footnote that still says "under human supervision." They hear the hard version of the word—and then they panic. What the labs actually wrote The careful documents are less dramatic than the chatter. Anthropic has published metrics on R&D automation. One figure that traveled widely: Claude "leads" roughly a quarter of model-development work end-to-end from a high-level prompt, still with people in the loop. The rung for fully autonomous successor design sits at zero in their public framing.[^1][^2] OpenAI has talked about an automated "research intern" that can carry defined tasks for days under direction, with a longer-range target for a fuller automated researcher, and has also said it does not yet know how to get all the way to aligned, full RSI safely.[^3][^4] Read those sentences all the way through and you get: AI already writes a lot of code, runs experiments, watches logs. People still set direction, approve launches, and allocate compute. The mess starts when RSI gets used as a casual shorthand in interviews, quotes, and headlines. Inside the building, speakers may mean "AI helping us build AI." Outside, the default meaning is "nobody is in control anymore." The dictionaries are not aligned. Panic scales to the heavier reading. I am not saying the labs are inventing productivity. Collaboration is faster. Headcount leverage is real. What I push back on is using RSI—knowing the public hears sandbox autonomy—as the label for today's human-in-the-loop R&D line. Then the internet adds its own chapters In mid-September, X lit up with wordplay and screenshots implying Google or DeepMind had "cracked RSI." No model card. No paper. No benchmark. Even the sharper tech blogs covering that wave treated it as unverified rumor.[^5] Almost the same week, arXiv got Dream-RSI: take past discovery trees, treat them as replay worlds, try many exploration policies offline, redeploy the winner. The underlying coding agent, the evaluator, and the weights can stay fixed.[^6] It is a clean engineering move. It improves how the search is run , not "the model grew a stronger brain by itself." Put Recursive Self-Improvement in the title and the reposts will still hear the second story. Nathan Lambert's recent Interconnects essay lands close to this. He is not buying closed-loop amplifying RSI as the near-term path. His older coinage was lossy self-improvement: models are inside the development loop, the losses are large, automatable research is narrower than the myth, parallel agents hit diminishing returns, and capital plus org politics still throttle the system.[^7][^8] He quotes Richard Ngo's line that short-timeline people may be directionally right relative to outsiders and still factually wrong about superintelligence on a few-year clock—things will move fast enough that the short-timeline camp feels vindicated anyway.[^7] That does not collide with my point. Fast is one claim. Unmanned iteration in a sandbox is another. Why the word still gets used One reason is local weather. The San Francisco lab scene was already tense; agents got more useful in early 2026; when thousands of them are running productively inside the company, anxiety climbs on its own. Jumping from "our tools suddenly work harder" to extinction talk skips stairs. From outside it can look almost religious. Lambert wrote about that temperature too.[^7] Feeling it inside does not license dropping definitions when you talk outside. Another reason is narrative competition. Sounding first on the "RSI track" helps fundraising stories, regulatory stories, slowdown stories. If the slowdown case rests on an inflated clock, and the main forecasted risks do not show up on schedule, you spend the credibility you will need the next time a real warning is due. The 2023–2024 safety fights already burned some of that store. The narrow argument Say that AI is speeding the labs' own R&D. Say the automation share is rising and people are still in the loop. Say work like Dream-RSI is cutting the cost of exploration. Do not tell the public "we already have RSI" while the definitions are still crossed. If true RSI arrives on my ruler, it should look like this: humans mostly constrain and watch from outside a sandbox; inside, the system decides how the next round changes, how it is checked, and how it hands off. Shout the word then if you want. Calling human–AI collaboration RSI today is convenient. Aimed at people outside the room, the cost is making them think that door already opened. The door is still shut. Do not take the headline's word for it. Check whether a person is still in the decision loop. References [^1]: Anthropic Institute. “When AI builds itself” / recursive self-improvement progress notes. https://www.anthropic.com/institute/recursive-self-improvement (also mirrored coverage summarizing the ~26% “leads” figure and the absence of full autonomy). [^2]: Fortune / AP. Barbara Ortutay, “What is recursive self-improvement?…” Sep 19, 2026. https://fortune.com/2026/09/19/what-is-self-improvement-rsi-full-autonomy-openai-anthropic-xai/ [^3]: OpenAI. Materials on automated research assistance and the path toward an automated researcher; public statements that aligned full RSI is not yet a solved, safe destination. See also Fortune/AP synthesis above. [^4]: OpenAI. Jakub Pachocki, “An Alien Mind,” and related policy notes on preparing for recursive self-improvement. https://openai.com/index/an-alien-mind/ [^5]: Trending Topics. “Three Letters Set the AI World Buzzing: Has Google Cracked RSI?” Sep 14, 2026. https://www.trendingtopics.eu/three-letters-set-the-ai-world-buzzing-has-google-cracked-rsi/ [^6]: Tong Zheng et al. “Dream-RSI: Recursive Self-Improvement through Evolving Worlds.” arXiv:2609.14858, Sep 14, 2026. https://arxiv.org/abs/2609.14858 · https://www.dream-rsi.com/ [^7]: Nathan Lambert. “Why I still haven’t bought into true RSI.” Interconnects, Sep 19, 2026. https://www.interconnects.ai/p/where-i-stand-on-rsi [^8]: Nathan Lambert. “Lossy self-improvement.” Interconnects, Mar 22, 2026. https://www.interconnects.ai/p/lossy-self-improvement