- According to WIRED, startup Recursive has introduced Flint, an open-weight AI coding model with 40 billion parameters, developed primarily using a self-improving AI system called Eureka Machine.
- CEO Richard Socher said Eureka Machine automates much of Flint’s development process, from designing the architecture and training methods for the foundational model to optimizing code for different types of chips. This enabled a small research team to complete the model in just a few months.
- According to benchmarks provided by Recursive, Flint outperforms models of comparable size and architecture on several coding tasks. A WIRED reporter tested it on a Mac with 48 GB of RAM and found the model fast and capable of handling software-generation requests effectively.
- Recursive is pursuing recursive self-improvement (RSI), an approach in which AI proposes, implements, and evaluates improvements to the AI development system itself before beginning another cycle of refinement.
- Socher said the long-term goal is to enable Eureka Machine to achieve full recursive self-improvement while retaining human oversight and a mechanism to halt operations when anomalies are detected.
- RSI remains controversial because of concerns that AI could develop capabilities beyond human control. Some bills proposed in the U.S. Congress seek to ban RSI, while Recursive argues that restricting the technology too early could hinder innovation.
- Recursive has raised $650 million at a valuation of $4.65 billion from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Its founding team includes Richard Socher, former chief scientist at Salesforce, and Jeff Clune, a researcher specializing in self-learning AI.
📌 Flint demonstrates the potential of using AI to automate AI model research and development, allowing small teams to compete with major laboratories without equivalent resources. However, current results demonstrate only the automation of many development stages and do not yet confirm that the system has achieved full recursive self-improvement. The central challenge is to verify the effectiveness of these improvements, ensure adequate oversight, and preserve human control as AI becomes increasingly involved in developing AI itself.
