Is recursive self-improvement happening now?
In a narrow sense, AI tools already help researchers build newer AI, for example by writing code and evals. A self-driving loop where AI improves itself without people in charge has not been shown.
The idea of AI systems helping to build better AI, which then builds better AI again. Partly real, mostly still speculative.
Recursive self-improvement is the idea that an AI system could improve its own design or help build a more capable successor, which could then improve the next one, and so on, possibly speeding up progress sharply. Parts of this are real today in a limited form: AI labs use their own models to write code, generate and filter training data, build evals and run experiments. But humans still choose the goals, run the training, check the results and decide what ships. A fully automated loop in which AI improves AI with little human involvement has not been demonstrated, and how fast or far it could go is debated.
The term comes up often in news and debates about AI risk and timelines, sometimes as hype and sometimes as a serious concern. Separating what is happening now, AI as a productivity tool inside AI research, from what is speculative, autonomous runaway improvement, helps you judge claims calmly. It is also why labs talk about monitoring and limits on AI systems used in their own research.
Today: an engineer at an AI lab uses a coding agent to write data-processing scripts and an eval harness for the next model, then reviews the code before it runs. That is AI helping improve AI, with a person in charge. Speculative: a model that independently designs, trains and deploys a better version of itself with no human checks. The gap between those two is where most of the debate sits.
Using AI to help build AI does not mean runaway improvement has started. Current systems speed up parts of the work, but data, computing power, evaluation and human decisions remain bottlenecks that a model cannot simply bypass.
All 28 terms are in the AI glossary. For models side by side, see frontier vs open-weight.
In a narrow sense, AI tools already help researchers build newer AI, for example by writing code and evals. A self-driving loop where AI improves itself without people in charge has not been shown.
If AI could reliably make itself more capable, progress might become very fast and hard to oversee. That is a reason labs discuss monitoring, evaluation and limits for AI used in AI research.
No. Chat models do not update their weights during your conversation. Recursive self-improvement refers to AI contributing to the design and training of future models, which is a separate, deliberate process.