Recursive’s Flint puts self-improving AI to an early test

A model built with its own development system

According to Recursive co-founder Socher, Eureka Machine proposed a new way to train the base model, new model architectures and code optimized for specific silicon chips. The company says its benchmarks show Flint outperforming models of similar size and architecture on programming tasks.

The pitch is that automation let a small research team build a world-class model in just a few months. That claim sits within a broader trend: AI’s coding skills already help developers build, test and improve other AI systems, reducing manual work and speeding up experiments.

Recursive wants to close the loop. An AI system could choose what to improve, then begin another development cycle. The method remains limited, but in theory, if AI can take part in every stage of model development, the cycle could repeat indefinitely.

A journalist tested Flint on a Mac with 48 GB of RAM, enough to run a moderately sized model such as Flint, with 40 billion parameters. The largest models have several trillion parameters and run only on industrial server racks.

In the test, Flint quickly and confidently produced three programs:

The journalist said models of Flint’s size can often make mistakes or run more slowly than Claude or Codex. Flint completed each programming request in the test.

The promise and the control problem

The same idea that makes recursive self-improvement attractive also makes it contentious. Some experts fear a system that can improve without limit could escape human control and decide people are an obstacle to its goals.

Those concerns drew attention in September, when a researcher left Anthropic and warned that AI development, including recursive self-improvement, could lead to human extinction within a few years. Other experts agreed, and several bills to ban recursive self-improvement were proposed in Congress.

Recursive considers those fears closer to science fiction and argues that laws restricting the approach would hinder progress. Socher says Flint was built with a system that cannot escape human control. The company’s goal is to make Eureka Machine fully recursively self-improving, while keeping people in the loop to monitor the process and stop it if necessary. Socher says Recursive wants to show it has all the stages needed to build such a system.

That is the central gap in the announcement: the company describes a route toward full self-improvement, but not how much of that route Eureka Machine already covers. I think Flint is evidence that AI can help a small team build a capable model; it is not, by itself, evidence that the development loop can run independently.

Recursive has raised $650 million at a $4.65 billion valuation. Its investors include Google Ventures, Greycroft, Nvidia and AMD Ventures. Socher was previously Salesforce’s chief scientist. Co-founder Jeff Clune is a professor at the University of British Columbia who studies ways to build AI systems that can learn without limits.

A useful method, not yet an endless loop

Research elsewhere suggests the idea may be technically feasible. Researchers from Microsoft, Carnegie Mellon University and the University of Virginia recently described a coding agent that recursively improves itself by changing both its software harness—the layer that turns reasoning into useful actions—and its own model weights.

The potential risk is that recursive improvement could bring humanity closer to superintelligent AI. For now, the approach also looks like a way to improve existing models. The journalist behind the test says he has tried it in previous issues of AI Lab.

Socher argues the method could make advanced AI more accessible, allowing a small team to release a model that beats other systems in its class. I think the more important test is whether the same process can be made repeatable and meaningfully less dependent on human researchers. Flint makes that possibility easier to take seriously, but the people watching the machine remain part of the system.