From I. J. Good to Flint: the long, strange quest for self-improving AI

By Turing Post · October 28, 2026.

This episode of Turing Post's "AI Builds AI" series opens with a quote from Richard Socher (interview with TechCrunch, edited for length): his approach is to use open-endedness to achieve recursive self-improvement, something he says no one has achieved yet. Socher argues that many people assume it simply happens through "auto-research": you ask an AI to improve something else, whether a machine learning system or a letter, but that, he says, is not recursive self-improvement, it is just improvement.

Turing Post explains that these conversations about recursive self-improvement (RSI) inspired the series. It notes that machine learning has long automated parts of its own development: neural networks were used to design other neural networks well before the current generative AI boom, and Jürgen Schmidhuber explored these ideas with his 1987 work on "learning to learn." The ambition of a system that improves its own ability to improve has therefore existed for decades.

The newsletter also introduces the notion of open-endedness: how a system keeps discovering useful things, whether it can find new problems as well as better solutions to existing ones, and why open-ended systems have proven so hard to build. The goal is to build a "map of RSI": a map of systems and ideas working toward open-ended self-improvement. To place something on it, one needs to understand what it improves, how it chooses its next experiment, where humans guide the process and how results feed back into the next round. According to the text, the map will also help navigate the fears surrounding RSI.

Today's case study is Flint, Recursive's open-weight coding model, designed to run locally. Flint was reportedly developed with the help of the Eureka Machine, an automated research system that proposes ideas, implements them and tests the results. The episode's question is how much of this is already-familiar automated research, where open-endedness comes in, and what would make the process truly recursive.

What makes self-improvement recursive: the text notes that an email improved by a model stays improved, but that does not make the model better. What is recursive is when the improvement flows back into the system doing the improving: a better learning method becomes part of how it learns, better research code becomes part of how it runs experiments, and a better model becomes part of the machinery that develops the next one. One has to follow what happens to the output after it is produced.

Historical background cited: in his 1965 essay "Speculations Concerning the First Ultraintelligent Machine," mathematician I. J. Good noted that designing machines is itself an intellectual activity, so a sufficiently intelligent machine could design a better one, which would in turn have greater design capability, triggering what he called an "intelligence explosion." Schmidhuber's 1987 thesis explored using genetic programming to improve genetic programming itself, that is, for the system to discover better ways of making its own changes. His Gödel machine proposal (2003) made self-modification explicit: the system would search for a mathematical proof that rewriting part of its software would be useful and apply the change upon finding one, and could also modify its own proof-search machinery.

In parallel, researchers turned parts of AI development into learning problems. In 2016, Barret Zoph and Quoc Le published "Neural Architecture Search with Reinforcement Learning": a neural network proposed architectures for other networks, which were trained and evaluated, and their scores were used to train the proposing network. That same year, Marcin Andrychowicz and colleagues published "Learning to learn by gradient descent by gradient descent," in which they used gradient descent to train an optimizer implemented as an LSTM that learned to update another model's weights.

The takeaway from this section is that machine learning had already automated architecture decisions and learned training methods long before today's research agents. For the map, the useful question is which parts of the process a system can change and how those changes affect its subsequent work; this will allow a more precise examination of how Eureka's discoveries helped build Flint and how such a model could strengthen Eureka's research capabilities. Another dimension remains to be explored: a system can improve its methods while always pursuing the same task and objective, whereas open-endedness raises the question of how it keeps discovering new possibilities, including problems its designers did not specify at the outset.

The rest of the episode (novelty search and POET, what open-endedness brings to RSI, what Eureka contributed to Flint, whether it is just sophisticated "autoresearch," where the loop closes, and FAQs) is behind the paywall for Premium subscribers, so this summary does not cover those sections.