Flint: Recursive releases an open-source model with 40 billion parameters, designed largely by AI

Richard Socher’s start-up claims that its Eureka Machine system has automated almost the entire development pipeline for Flint, an open-source model-based code platform that runs on a Mac.

Recursive, founded by Richard Socher, former Scientific Director of Salesforce, and co-founded by Jeff Clune (University of British Columbia), released his first model: Flint, an open weight programming model of 40 billion parameters. According to the company, its internal system called Eureka Machine supported most of the development: new basic model drive method, new architectures and optimized code for precise chips. A small team of researchers would have completed the project within a few months, and published scores put Flint ahead of comparable size and architecture models on code tasks.

Recursive Self-Improvement, a Worrying Subject

Recursive claims the goal of achieving a complete recursive self-improvement (RSI), a disputed approach: several bills in the US Congress aim to prohibit it, and some researchers see it as a risk of loss of control. Socher judges these fears of science fiction and ensures that the process would remain supervised by humans capable of interrupting it.

The company raised $650 million for $4.65 billion from Google Ventures, Greycroft, Nvidia and AMD Ventures. Wired's reporter who tested him said he had Flint run on a 48 GB memory Mac, with results he described as impressive.