Recursive focuses on using AI to develop AI

Summary

Recursive, led by former Salesforce Chief Scientist Richard Socher and valued at $4.65 billion, released Flint. This 40-billion-parameter open-weight coding model was largely developed by the company's own AI system, Eureka Machine. This marks the transition of Recursive Self-Improvement (RSI) from theory to practice. WIRED reporter Will Knight tested Flint on a 48 GB Mac, finding it fast and reliable for coding tasks in its size class.

Why it matters

RSI is one of the most debated areas in AI safety. A small team using AI for model training, architecture, and hardware optimization without human intervention could dramatically accelerate the rate of capability gains. This news emerged amid real political tensions, with multiple Congressional bills proposing to ban RSI and a former Anthropic researcher issuing an existential risk warning. Under these conditions, Recursive’s open-weight and public demonstration serves as both a technical milestone and a direct challenge to the regulatory narrative.

Highlights

• Flint (40 billion parameters) was trained using an 'almost entire process' automated by Eureka Machine. This process includes new foundation model training methods, architectures, and chip-specific code optimization.
• Recursive raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Co-founder Jeff Clune researches unsupervised learning AI at UBC.
• In Will Knight's field test on a 48 GB Mac, Flint 'nailed every one of my coding requests.' The model outperformed typical models of its size. However, Knight notes that frontier models operate with trillions of parameters and industrial server racks.

Critical take

The article reads more like a promotional profile than independent reporting. Recursive’s own benchmark claims are presented without third-party verification. The 'scarily good' framing is quoted from the company’s positioning. Dismissing RSI risks as 'science fiction' is attributed to the CEO without meaningful counter-arguments from safety researchers. While the single positive field test on a consumer Mac is useful, the piece fails to address whether the claimed autonomy of Eureka Machine over the entire process is reproducible or merely a curated narrative supporting the $4.65 billion valuation.