Recursive Launches 40 Billion-Parameter Flint Model, Using AI to Build AI
Updated
Updated · WIRED · Oct 21
Recursive Launches 40 Billion-Parameter Flint Model, Using AI to Build AI
3 articles · Updated · WIRED · Oct 21
Summary
Flint, a 40 billion-parameter open-weight coding model from Recursive, was built in just a few months with AI automating nearly every stage of development, CEO Richard Socher said.
Recursive says its Eureka Machine handled most of the model pipeline and generated innovations in training, architecture and chip-specific code, helping Flint outperform similarly sized coding models on shared benchmarks.
The launch is an early real-world test of recursive self-improvement, in which AI systems help redesign and improve themselves with less human input.
That approach is drawing scrutiny: some experts warn unchecked self-improving AI could become dangerous, and several bills in Congress propose banning recursive self-improvement.
Recursive, valued at $4.65 billion after raising $650 million, argues human oversight can contain the process and says the technique could let small teams build cutting-edge models more cheaply.