GlossaryUpdated
RSI (recursive self-improvement)
RSI in AI means recursive self-improvement: an AI system that helps build a better version of itself, and that better version is in turn better at building the next one.
RSI in AI stands for recursive self-improvement: an AI that helps build a better version of itself, where that better version is in turn better at building the next one. That is the full form of the acronym in AI and machine learning. "Recursive" means the loop repeats. Each round of improvement could make the next round faster. That is what separates RSI from normal progress, where people use today's tools to build tomorrow's.
If you searched for "RSI meaning", note the other common meanings: repetitive strain injury, and the Relative Strength Index used in stock charts. Neither applies when the topic is AI.
What it means for you at work
If AI helps build the next AI, new versions can arrive more often, and each one behaves a little differently. A prompt that worked well in spring may give different answers in autumn, because the model behind your tool changed.
Three habits help:
- Keep a few test examples. Save five to ten real requests, with answers you were happy with. When your AI tool updates, run them again and compare.
- Notice when the model changes. Many AI apps show which model you are using. When it changes, check your most important prompts.
- Write clear rules, not hints. "Keep it under 150 words and follow our refund policy" holds up across model changes better than "use common sense".
The LLM evals page explains how to turn those test examples into a simple check.
Where the idea comes from
In 1965 the statistician I. J. Good wrote that a machine smarter than any person could design even better machines. He called the result an "intelligence explosion". It is the path most often suggested toward artificial superintelligence (ASI): AI that would beat the best people at almost everything. For decades this was a thought experiment.
What changed is that AI can now do some of the work of building AI. It can write code, create and grade practice examples for training, and help run experiments. So the question has moved from "could this happen?" to "how much of the next model did the current model help build?"
Weak RSI and strong RSI
Most arguments about RSI are two people meaning different things. It helps to split the term in two:
- Weak RSI: AI helps build AI. People still run the work, but AI does a growing share of it: writing code, grading answers during training, finding bugs, drafting reports. This is happening now.
- Strong RSI: AI improves itself on its own. The AI proposes changes to itself, makes them, checks the result and repeats, with each round faster than the last. This has not been shown.
When an AI company says it is working toward RSI, it usually means the weak form as a research goal. When a critic calls RSI hype, they usually mean the strong form.
Real examples so far
No AI runs the full loop by itself. But single steps of it have real examples:
- AlphaGo Zero (DeepMind, 2017). It learned the board game Go by playing millions of games against itself, and each new version trained on the games of the one before. Within three days, it beat the earlier version that had defeated world champion Lee Sedol, by 100 games to 0. But it only improved at one game, and it could not change its own design.
- AI grading AI (2022 onward). Methods such as Anthropic's Constitutional AI have an AI, not a person, judge answers during training. One step of building AI moved from people to AI.
- AlphaEvolve (DeepMind, 2025). An AI system built on Google's Gemini models searched for faster ways to run calculations. One of its findings sped up part of Gemini's own training, saving about 1% of training time. The gain was small, but the AI helped improve the system that produces it.
- AI writing code at AI companies. AI companies, Anthropic among them, have said publicly that their own models now write a large share of their code.
The pattern is the same in every case. AI does single steps. People still choose what to try, approve the changes and run the training.
What Anthropic's S-1 filing says about RSI
In November 2026 RSI moved from research papers into a stock market filing. On 2 November Anthropic, the company behind Claude, filed an updated S-1 (the document a US company files before it sells shares to the public). Its risk section uses the term directly. Using AI to help build more capable AI "could also pose a considerable safety risk", the filing says, and Anthropic's Responsible Scaling Policy sets out how it will watch for that "as we approach what has been termed 'recursive self-improvement.'" The filing came five months after Anthropic's June 2026 report on recursive self-improvement, which judged the most likely path to be AI speeding up engineering work while people keep control of research.
Three points in the filing matter for reading the headlines:
- The threshold. Anthropic wrote that its internal estimates for future models, not its current ones, "indicate that we may imminently meet" its own policy's definition of "dramatically accelerated progress attributable to AI". That definition is the RSI threshold people are searching for.
- Not crossed yet. The same paragraph says no current model has been designated as over that threshold. It also admits the company has not always noticed a big speed-up until after it happened.
- People stay in charge. Anthropic wrote that it does not expect to safely let Claude design and build new frontier models on its own "without meaningful human supervision or understanding", and that it is considering policy changes to keep people in control.
Read it with care. A filing like this is a legal list of risks and also a sales document for investors, as TechCrunch noted in its 2 November report. It is a company saying weak RSI is close by its own measure. It is not proof that strong RSI exists.
How AI companies measure it: RSI evals
RSI evals are tests of how much of the work of AI research an AI can do by itself. OpenAI's public examples include MLE-bench, which scores AI on machine learning competitions, and PaperBench, which checks whether an AI can repeat the results of published research papers. The research group METR built RE-Bench, which compares AI with human experts on research tasks under the same time limit.
These tests measure pieces of the work, not the full loop. OpenAI's Preparedness Framework, its public safety policy, lists AI self-improvement as one of the risks it tracks.
Why the loop may stall
Nobody knows how far the loop goes. Sceptics point to five limits that a smarter AI does not remove:
- Computer power. Every idea still has to be tested by training a model, which takes time on expensive hardware.
- Real-world data. Some knowledge only comes from watching the world, and that takes time.
- Testing. A loop only keeps a change if the new version scores higher. If the test is flawed, the AI can learn to pass the test without getting better at real work.
- Errors that add up. A small flaw in one version gets built into the next, like a photocopy of a photocopy.
- Harder gains. The easy improvements come first. Each later one is harder to find.
Both sides agree on the near-term effect: AI tools will keep changing. Keep the tests that tell you what changed.
A prompt for checking a model update
When your AI tool moves to a new model, run your saved examples again and use this:
I am checking whether a new AI model version works as well as the old one for my task.
<task>[paste the prompt you use]</task>
<old_answers>[paste 5 to 10 answers from before the update]</old_answers>
<new_answers>[paste answers to the same requests from the new version]</new_answers>
For each pair, say whether the new answer is better, worse or the same, using these rules: [your rules].
Then list any change in length, format, tone or refusals that I would notice.
End with a recommendation: keep using it as is, keep using it with prompt changes (say which), or switch back if my app allows it.
RSI (recursive self-improvement) in the news
Questions
What does RSI stand for in AI?
In AI, RSI stands for recursive self-improvement: an AI helping to design, train or build its own next version, where each better version is better at improving the next. Outside AI, RSI usually means repetitive strain injury or the Relative Strength Index used in stock charts. Neither applies when the topic is AI.
Is recursive self-improvement already happening?
The weak form is. AI companies use their current models to help write code, create and grade training examples, and run experiments. The strong form, an AI that improves itself without people in charge and gets faster each round, has not been shown.
What are RSI evals at OpenAI?
They are tests that measure how much of the work of AI research an AI can do by itself. OpenAI's public examples include MLE-bench, which scores AI on machine learning competitions, and PaperBench, which checks whether an AI can repeat the results of published research papers. OpenAI's Preparedness Framework, its public safety policy, lists AI self-improvement as one of the risks it tracks.
What did Anthropic's S-1 say about the RSI threshold?
In its updated S-1 filing of 2 November 2026, Anthropic wrote that its estimates for future models indicate it may soon meet its Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI, the line usually called recursive self-improvement. It also wrote that no current model has crossed that threshold, and that it does not expect to let Claude build new frontier models without meaningful human supervision.
Is RSI the same as AGI?
No. AGI, artificial general intelligence, is a level: AI that can do most thinking work a person can. RSI is a process: AI helping to build the next AI. An AI could help a lot with its own next version long before it reaches a general, human level.
Why does RSI matter if I just use ChatGPT at work?
Because it can make AI tools change more often. When the model behind your tool changes, a prompt that worked well may start to give different answers. Keep a few real examples with answers you liked, and run them again when your tool updates.
Related terms
- ASIAI that would beat the best people at almost everything. It does not exist yet.
- Intelligence explosionThe idea that self-improving AI could get smarter very fast, all at once.
- LLM evalsRepeatable tests that check whether AI answers are good enough.
- Agent memoryWhat AI saves so it remembers you and your work next time.
- Context compactionWhen a long AI chat is summarized to make room so it can keep going.