After AGI, a new term is dominating AI conversations: RSI, or Recursive Self-Improvement. It means an AI that makes itself smarter without human help. Several major labs are working on it right now, and it may be the most important concept in AI you have not heard of yet.
The Gist
- RSI stands for Recursive Self-Improvement: an AI system that continuously upgrades itself without human intervention
- Anthropic (via Andrej Karpathy’s Auto-Research project), Richard Socher’s Recursive Superintelligence, and Adaption’s AutoScientist are all actively working on it
- Current AI can make small self-improvements, but still struggles with self-direction (experts disagree on timelines)
What RSI Actually Means (In Plain English)
AI usually gets smarter because humans improve it. A team of researchers studies how the model behaves, identifies its weaknesses, adjusts the training process, and releases a better version. That loop requires human effort at every step.
RSI removes the human from that loop. Recursive Self-Improvement means an AI system that can identify its own weaknesses, figure out how to fix them, and make itself better without waiting for a team of researchers to do it. The “recursive” part is important: once it gets better, the improved version is now smarter at improving itself, and so on, in a self-reinforcing cycle.
Why does this matter? Because if a system can improve itself faster than humans can study and guide it, the rate of progress stops being controlled by how fast human researchers work. It becomes limited only by available computing power. That is a very different kind of AI development than what we have seen so far.
This is why RSI is making experts nervous and excited at the same time. It is the mechanism behind what some people call “the intelligence explosion”: a scenario where AI progress accelerates so fast that it becomes difficult to predict or control. The term AGI (Artificial General Intelligence) described the destination. RSI describes the engine that could get us there.
Who Is Working on It Right Now
RSI is not a theoretical concept being discussed in papers. Multiple teams are actively building toward it. At Anthropic, Andrej Karpathy, who recently joined the company after years at OpenAI and Tesla, is leading a project called Auto-Research. The goal: build AI systems that can conduct their own research and development cycles.
Richard Socher, a well-known AI researcher, has launched a company called Recursive Superintelligence specifically focused on this problem. Adaption, another startup, has a product called AutoScientist with a similar aim. These are not university research projects. They are funded companies with teams building practical systems.
Current progress is real but limited. AI systems today can make incremental improvements to smaller models, and some can already write their own code. But they still struggle with self-direction: understanding what to prioritize, how to verify their own work, and how to set their own goals. These are the hard problems that remain unsolved.
Expert opinion on timelines is genuinely divided. Some researchers expect rapid progress that could lead to significant breakthroughs within a few years. Others expect slower, plateau-prone development where the hardest challenges take much longer. What almost nobody argues is that RSI is irrelevant. The disagreement is about when, not whether.
Why You Should Care About This Now
RSI might sound abstract, but it changes something concrete about how you should think about AI tools. Every AI assistant you use today was improved by a human team over months or years. A system with RSI capabilities could improve itself over hours or days. The tools available to you in two years could be dramatically more capable than anything you can try today.
This also explains why so much money and talent is flowing into the companies working on frontier AI. The potential upside of reaching genuine RSI is enormous, but so is the importance of making sure it happens safely. That tension is why Anthropic, founded specifically to make AI development safer, is now one of the most valuable companies in the world at nearly a trillion dollars.
For you as a beginner trying to understand AI: the key idea is that the pace of AI development is not fixed. RSI is the reason researchers and investors talk about AI progress in years rather than decades. Whether it happens gradually or in a leap, self-improving AI would fundamentally change how quickly the tools around us evolve. Following that development, even from a distance, is worth your time.
Follow the story on AI Noobies.
About the Author
