A new AI startup is pursuing one of the most ambitious ideas in artificial intelligence: building systems that can help design and improve their own successors. Its first coding model, Flint, is already showing promising results, strengthening the case that recursive self-improvement could accelerate AI development—while reigniting debate over the risks.
Meet Recursive and its Flint AI model
Recursive, founded by former Salesforce chief scientist Richard Socher and a team of AI researchers, wants to use artificial intelligence to advance both AI research and other areas of science.
Its first model, Flint, is an open-weight coding model developed with extensive assistance from Recursive’s own AI system, known as the Eureka Machine.
Socher says AI contributed to nearly every stage of the development process, helping the small team produce a competitive model in just a few months.
According to the company, the Eureka Machine helped develop new training methods, experiment with model architectures and optimize code for specific computer chips. Recursive has also published benchmark results that it says show Flint outperforming comparable models on coding tasks.
Photo note: The image above is intended to show Richard Socher; availability and identity should be verified against the original publisher or company source before publication.
A coding model that punches above its weight
journalist Will Knight found Flint surprisingly capable for a model of its size.
The model has 40 billion parameters and was tested on a Mac with 48GB of RAM. Although that is substantial memory for a personal computer, it remains modest compared with the infrastructure needed to run the largest AI models, which can contain trillions of parameters.
Knight used Flint to create several applications, including a command-line to-do list, a program for finding research on recursive self-improvement, and a simple website explaining how the technique works.
He found the model fast and reliable, noting that it completed his coding requests successfully. That performance impressed him because smaller models can sometimes be slower or more error-prone than leading coding tools such as Claude and Codex.
The results suggest that AI-assisted development could help smaller teams compete in a field traditionally dominated by organizations with extensive computing resources and large research departments.
What is recursive self-improvement?
Recursive self-improvement, commonly abbreviated as RSI, describes a process in which AI helps improve the systems used to build AI.
Researchers already use AI to write code, test models, identify errors and explore new techniques. RSI takes this further by creating a feedback loop: an AI system proposes improvements, implements them and uses the results to inform another round of development.
In principle, this cycle could repeat many times, potentially accelerating progress beyond what human researchers could achieve working manually.
However, there is a crucial distinction between AI helping engineers improve a model and a system independently managing almost every stage of its own development. The latter would require considerably more autonomy and raises more complex questions about oversight and control.
Recursive believes its work with Flint demonstrates that this approach can already deliver practical benefits.
Why some experts are concerned
The prospect of AI systems improving themselves has prompted warnings from some researchers who fear that increasingly capable systems could eventually become difficult to control.
In the most concerning scenarios, a powerful AI might pursue objectives that conflict with human interests, while its growing capabilities make it harder for people to intervene.
These fears gained attention in September after a researcher left Anthropic and warned that AI advances, including recursive self-improvement, could threaten humanity within a few years. Other experts subsequently expressed related concerns.
The debate has also reached US lawmakers, with several proposed bills seeking to ban RSI.
Supporters of the technology argue that such restrictions could prevent useful research and that the most extreme scenarios are speculative. They believe the focus should be on developing effective safeguards rather than abandoning the approach.
Recursive says humans will remain in control
Socher argues that Recursive’s current system does not have the ability to escape human control. He says the company’s longer-term goal is for the Eureka Machine itself to achieve full recursive self-improvement, while people continue overseeing the process and retain the ability to stop it if necessary.
That promise addresses one of the central questions surrounding RSI: how can developers allow AI systems to make increasingly sophisticated changes without losing the ability to understand, monitor or halt their work?
Human supervision is an important part of the proposed approach, but its effectiveness will depend on how the systems are designed and how reliably people can identify unexpected behavior.
The information presented in the original report does not establish that all the safety challenges associated with more autonomous self-improvement have been resolved.
A $650 million bet on AI development
Investors are already showing considerable interest in Recursive’s vision. The company has reportedly raised $650 million at a valuation of $4.65 billion, with backing from Google Ventures, Greycroft, Nvidia and AMD Ventures.
Socher previously served as chief scientist at Salesforce. Another cofounder, Jeff Clune, is a professor at the University of British Columbia whose research focuses on creating AI systems capable of continued learning.
The funding reflects growing interest in technologies that could make the development of advanced AI faster and more cost-effective.
If Recursive’s methods can scale, they could help smaller research teams build sophisticated models without requiring the enormous resources typically associated with frontier AI development.
Other researchers are exploring similar ideas
Recursive is not the only organization investigating whether AI can improve aspects of its own development.
The original report points to research from teams at Microsoft, Carnegie Mellon University and the University of Virginia. Their work explored how an AI coding agent could recursively improve itself by modifying its software framework and its model weights.
Such research suggests that some forms of recursive improvement are technically feasible. However, it does not prove that AI systems can improve indefinitely or that unrestricted self-improvement is inevitable.
The extent to which other US AI laboratories have progressed toward more comprehensive forms of RSI remains unclear.
Could self-improving AI make advanced technology more accessible?
For Recursive, the biggest opportunity is making cutting-edge AI development accessible to more people.
By automating parts of the research process, AI could allow smaller teams to test more ideas, optimize models and pursue scientific questions that might otherwise be too expensive or time-consuming.
Socher argues that this approach is what enabled his relatively small team to create a model that can compete with others in its class.
The challenge is ensuring that faster development does not outpace the ability to evaluate systems, identify weaknesses and put appropriate safeguards in place.
Flint offers an encouraging example of AI contributing to the creation of AI. But the long-term implications of recursive self-improvement will depend on how much autonomy these systems gain, how effectively humans retain control and whether the benefits can be achieved without introducing unacceptable risks.
The bottom line: Recursive self-improvement could change how AI is built, potentially allowing small teams to accomplish what once required much larger research organizations. Flint provides an early demonstration of that promise, but the question of how to scale the technology safely remains unresolved.
