AI startup Recursive has unveiled Flint, an open-weight coding model that company leaders say was developed largely by its own recursive AI system, the Eureka Machine. In tests by WIRED, Flint—a 40-billion-parameter model that runs on a Mac with 48GB of RAM—handled a variety of coding tasks quickly and accurately, exceeding what the writer expected from similarly sized models.
Recursive CEO Richard Socher, formerly chief scientist at Salesforce, says the Eureka Machine automated nearly every part of the model pipeline, including training approaches, model architectures, and chip-specific code. That allowed a handful of researchers to build a competitive model in just a few months. The startup frames this as a step toward full recursive self-improvement, or RSI, while keeping human oversight in place.
RSI remains controversial. Some researchers warn that self-improving AI could escape human control, and legislative proposals in Congress have sought to ban it. Recursive dismisses those fears as closer to science fiction than reality, arguing that its system is designed to be stopped if it goes off track. The company has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Nvidia, AMD Ventures, and Greycroft.