A startup with a handful of researchers says AI built nearly all of its new coding model, Flint, which a Wired writer ran on a Mac and found very capable.
A company called Recursive has released a coding AI model named Flint, and says AI did nearly every step of building it. The model is small enough to run on a Mac with 48 GB of memory. A Wired writer tested it and said it handled every coding request he gave it.
CEO and cofounder Richard Socher says the company used its own system, called the Eureka Machine, to automate most of the work. That let a handful of researchers build a top model in a few months, he says. The system came up with a new way of training the base model, new model designs, and code tuned for specific computer chips.
This approach is called recursive self-improvement, or RSI. In plain terms, AI helps improve AI, and the improved AI then helps again.
Flint has 40 billion parameters, a rough measure of a model's size. The biggest models have several trillion and need racks of industrial servers. Recursive shared benchmark scores, which are standard tests, suggesting Flint beats other models of similar size on coding tasks.
The Wired writer used it to build a command line to-do list, a program that searches the web for research on RSI, and a simple website explaining how RSI works. He says models this size are often error-prone or slow compared with Claude or Codex, but Flint did not stumble.
Some experts fear that AI able to improve itself without limit could escape human control. In September, a researcher quit Anthropic and warned that such advances could lead to human extinction within a few years. Several bills in Congress propose to ban RSI.
Recursive calls these fears more science fiction than reality. Socher says Flint was built with a system that cannot escape human control. He says the goal is full RSI overseen by humans who could stop it.
The benchmark scores come from Recursive, and the article does not describe independent tests. It also does not say how people can get Flint or what it costs. It is unclear how far other US labs have gone with RSI, though a team from Microsoft, Carnegie Mellon and the University of Virginia has published work on a coding agent that improves itself. The policy fight is unsettled.
A strong coding AI on one Mac could mean cheaper, more private help, but the scores are the company's own and the safety debate is far from settled.
Netflix's top plan is now $26.99 a month, while its 2007 streaming only worked on a Windows PC with Internet Explorer. Here is what each plan gets you.
Read more: Netflix now costs up to $26.99 a month, and in 2007 it only ran on one kind of PCVisa tested 22 AI models on realistic shopping tasks. They passed just over half, and often bought the wrong thing anyway.
Read more: Visa tested 22 AI shoppers, and most still make costly mistakesDARPA wants AI to hunt for design flaws in stock, bond, energy and cloud markets, so enemies cannot use them to cause economic damage that reaches households.
Read more: Pentagon wants AI to find weak spots in stock and energy markets
