Flint Is Only the Start: Recursive's Eureka Machine Learns to Improve Itself

Key points

Key takeaways from MTS interview with Recursive co-founder and CEO Richard Socher (October 2026):

Flint, the launch. Recursive released Flint, which Socher describes as the best and fastest local coding model available to download — fully open source and better than comparable models in the category, "pretty much all of which are Chinese" — making it one of the first American-made open-source models at the frontier.

The Eureka Machine. The model was built largely by AI: an automated research system in which a community of AI scientists and models ideate, implement and validate ideas in an open-ended fashion. Socher says nearly every aspect of the model pipeline was done and improved upon by AI, an automation level no other company has reached.

Open-endedness over narrow goals. Recursive argues that chasing a defined goal directly misses innovations — the microwave analogy, radio-wave researchers discovering melted chocolate, illustrates combining interestingly different ideas across branches, inspired by biological, cultural and technological evolution.

Rainbow teaming. Co-founder Tim Rocktäschel's idea of co-evolving attacker and defender language models, inoculating against each newly discovered attack, is cited as the open-ended answer to safety problems such as prompt injection — preferred over pure RL chasing a single benchmark.

Against outlawing self-improvement. Socher pushes back on calls to ban recursive self-improvement, arguing that enforcement would require a totalitarian surveillance state; "regulate applications of intelligence, not intelligence itself" — as with self-driving cars and drugs, not smart people.

Ideation is the frontier. Using AI to implement ideas is the consensus piece of recursive self-improvement; generating genuinely novel ideas is the least consensus capability. The combinatorial space of existing concepts is underexplored, and deleting an idea can itself be novelty — the transformer paper's core insight being that no recurrent layer was needed.

A DecaNLP flashback. Socher's early paper unifying every NLP task into one prompted neural network was rejected by reviewers, then picked up by Alec Radford and Ilya Sutskever and cited in the early GPT papers — an idea he says is now "like breathing air."

Read more: MTS