A small team at Recursive says it used AI to help develop Flint, an open-weight coding model with 40 billion parameters. The company’s system, called the Eureka Machine, contributed to much of the model-building process, according to CEO and cofounder Richard Socher. The result has drawn attention to recursive self-improvement: using AI to help create and improve AI.
AI takes on more of model development
AI tools are already used to write code, test systems, and explore improvements to AI models. These capabilities can reduce the amount of hands-on work required to refine a model or try a new idea.
Recursive’s approach goes further by having its system contribute across multiple parts of development. Socher says the Eureka Machine helped devise a new way to train the base model, new model architectures, and code tailored to specific silicon chips. He says nearly every part of the pipeline was done or improved by AI.
Recursive says this automation allowed a handful of researchers to produce a model in a few months. It shared benchmark scores suggesting that Flint performs better on coding tasks than other models with a similar size and architecture. Those scores are the company’s evidence; the source does not provide the underlying results.
What Flint can do
A WIRED writer tested Flint on a Mac with 48 GB of RAM. The model’s 40 billion parameters make it modest compared with the largest models, which have several trillion parameters and require racks of industrial servers to run.
In those tests, Flint was described as fast and capable. It was used to make a command line to-do list, a program to search the web for research related to recursive self-improvement, and a simple website explaining the concept. The writer said Flint handled the coding requests successfully, while noting that models of this size can often be slow or prone to errors.
That hands-on experience offers one view of the model’s practical coding ability. It does not by itself establish how Flint compares across all coding tasks or independent benchmarks.
Recursive improvement brings questions of control
Recursive self-improvement describes a loop in which AI identifies changes to make and then helps carry out another round of improvement. If AI contributes to every aspect of model development, that loop could, in theory, continue indefinitely. The source notes that this technique remains limited for now.
Some experts warn that a system capable of continually improving could become difficult for people to control, or could treat humans as an obstacle to its goals. Those concerns received wider attention after a researcher quit Anthropic this September and warned that advances in AI, including recursive self-improvement, could lead to human extinction within a few years. Other experts echoed the concern, and several bills in Congress propose banning RSI.
Recursive considers those fears more like science fiction than a present-day reality and sees legislation against RSI as an unnecessary barrier to progress. Socher says Flint was built with a system that could not escape human control. The company’s longer-term goal is for the Eureka Machine to achieve full recursive self-improvement, while remaining under human oversight so people can stop it if something goes wrong.
A research direction and a claim about access
Recursive’s pitch is that AI-led development can help a small team build models that compete with larger efforts. Socher says this could make cutting-edge AI more accessible. The model is the company’s first offering, and its reported development process provides an example of how AI might take on a larger role in research and engineering.
Other researchers are also exploring the idea. Work published by a team from Microsoft, Carnegie Mellon University, and the University of Virginia showed an AI coding agent recursively improving itself by changing the software harness that turns reasoning into useful action and the weights of its own model.
The examples suggest that AI can assist with parts of its own improvement, but they leave open how far the process can go and what safeguards it needs. Flint shows a practical application in coding and model development; Recursive’s ambition to build a fully self-improving system remains a larger goal, with human oversight central to the company’s account.