Recursive says AI handled most development of its first coding model, Flint

Recursive used AI to automate most of the development of its first coding model, Flint, CEO Richard Socher told WIRED in an October 21 feature. The open-weight model has 40 billion parameters, and reporter Will Knight found it fast and capable in hands-on coding tests—a promising result for a company pursuing AI that improves AI.

The development system behind Flint is called the Eureka Machine. Socher says it helped a handful of researchers build the model in just a few months. His account puts AI inside the process of designing and training a model, rather than describing it only as a tool for writing the software around one.

Socher credits the Eureka Machine with devising a new way to train the base model, new model architectures and code optimized for particular chips. These are different parts of the development pipeline: how a model learns, how it is structured and how its software runs on hardware. He says AI contributed to nearly every part of that pipeline.

Recursive also supplied benchmark results that it says show Flint beating coding models of similar size and architecture. That is a comparison with a defined peer group, not a claim that Flint beats every coding model. The scores are company-supplied; Knight’s separate hands-on work offers a narrower view of how the model performs on actual requests.

Nearly every aspect of the entire model pipeline was actually done and improved upon by AI

Richard Socher, Recursive CEO and cofounder, speaking to WIRED

Knight ran Flint on a Mac with 48 GB of RAM. He reported that it was fast and successfully handled every coding request he tried. That stood out against his experience with models of this size, which he described as sometimes slow or error-prone compared with Claude or Codex.

The tasks Knight tried

• A command-line to-do list.
• A program to search the web for research about recursive self-improvement.
• A simple website explaining how recursive self-improvement works.

Those examples span a small utility, a research-search program and a website. They support Knight’s positive assessment of the requests he tested, rather than a general guarantee of coding reliability. They also test Flint’s output, not the Eureka Machine’s ability to keep producing better models without human direction.

Recursive self-improvement means more than using AI to help with a development task. In the loop described by WIRED, an AI system decides which improvements to make, carries them out and begins another round. If AI could contribute across the whole development process, that cycle could theoretically continue indefinitely. The technique remains limited today.

Socher says full recursive self-improvement is the goal for the Eureka Machine, not a milestone he says Flint has already established. He told WIRED that humans would oversee that system and could stop it if it went wrong. He also described the system used to build Flint as incapable of escaping human control.

The dispute concerns what happens as that loop becomes more capable. WIRED describes fears that indefinitely improving AI could escape human control; Recursive rejects those fears as more science fiction than reality. Socher’s case is that the approach can make advanced AI development more accessible, allowing a small research team to produce a competitive model.