Recursive launches Flint, a self-built AI coding model showcasing recursive self-improvement
Recursive introduced Flint, a 40-billion-parameter coding AI that was mostly designed and optimized by its own recursive system, the Eureka Machine.
Recursive unveiled its new AI coding model, Flint, which contains 40 billion parameters and was largely created by the company’s own recursive AI system, the Eureka Machine. Richard Socher, Recursive’s CEO and co-founder, explained that the system automated key steps such as novel training techniques, architecture designs and silicon-specific code, allowing a small research team to produce a world-class model in months.
Benchmark results released by the firm show Flint surpassing similarly sized models on programming tasks, and the author’s personal tests confirmed fast, accurate code generation. The startup has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia and AMD Ventures, and counts former Salesforce chief scientist Socher and UBC professor Jeff Clune among its founders. Despite Recursive’s confidence that human oversight prevents runaway behavior, critics cite recent warnings from AI researchers about the dangers of recursive self-improvement, and several U.S. bills now seek to prohibit the technique.
Why it matters
Flint demonstrates that AI can largely design its own successors, accelerating development but also raising safety and regulatory concerns.
In this story
