Recursive's Flint built by Eureka Machine in four months

Recursive's founders include Richard Socher and former DeepMind researcher Tim Rocktäschel, and the company came out of stealth in May 2026 with $650 million raised at a $4.65 billion valuation — before it had any concrete technical results to show. Flint is the first public evidence that the Eureka Machine can carry an idea from architecture search through quantization and kernel optimization to a shipped model. The specifics are notable: starting from 35 billion total and 3 billion active parameters, the system tested variants and landed on 40 billion total and 2.5 billion active; for quantization it went against common practice by cutting the output layer to four bits, adding a distillation step that Recursive says needed roughly half as many training steps as the previous method, leaving the finished NVFP4 model only 0.7 points behind the 16-bit version. The company's longer-term goal is a superintelligence that keeps improving itself, with each model making the Eureka Machine a better researcher. That leap is not yet demonstrated, and the debate is live: former Google DeepMind research lead Oriol Vinyals considers AI-assisted self-improvement realistic but thinks an intelligence explosion is unlikely, and a study of more than 700 task logs found AI agents handle large portions of the work while humans still make most of the critical research calls. Other labs are moving in the same direction — Anthropic says Claude now handles a large share of its own development work, and OpenAI deploys AI agents across its research pipeline — so Flint is best read as one concrete data point in that broader push rather than proof that the loop closes.