From I. J. Good to Flint: The Long, Strange Quest for Self-Improving AI

Conversations about recursive self-improvement (RSI) inspired us to start this series. But it’s a tricky subject. On one side, machine learning has a long history of automating parts of its own development. Researchers were using neural networks to design other neural networks well before the current GenAI boom. And of course, Jürgen Schmidhuber was exploring these ideas much earlier, with his 1987 work on “learning to learn”, proposing ways for systems to improve the learning process itself. So even the bigger ambition – a system that improves its own ability to improve – has been around for decades.

And then there’s open-endedness, which we wrote about back in 2024. How does a system keep discovering useful new things? Can it find new problems to explore as well as better solutions to existing ones? And how do those discoveries make it better at the research process itself? And why have open-ended systems remained so hard to build?

So, instead of talking about “AI builds AI” in a vacuum, we want to start building a map of the systems and ideas working toward open-ended self-improvement (RSI map). To place something on that map, we need to understand what it improves, how it chooses its next experiment, where humans guide the process, and how the results feed back into the next round.

There is a lot of fear-mongering around RSI, so this map will help us navigate those fears as well.

Today, we will review Flint and explain why it fits the map perfectly. Flint is Recursive’s open-weight coding model, designed to run locally. Allegedly, Flint was developed with help from the Eureka Machine, an automated research system that proposes ideas, implements them, and tests the results. That sounds interesting. So let’s investigate: how much of this is familiar automated research, where does open-endedness enter the picture, and what would make the process recursive?

In today’s episode:

Your email improved by a model is indeed improved, but it doesn’t make the model better. So it’s all about the loop: the recursive part is the improvement coming back to the system doing the improving.

A better learning method becomes part of how the system learns. Better research code becomes part of how it conducts experiments. A better model becomes part of the machinery developing the next model. We need to follow what happens to the result after it has been produced.

That core idea goes back decades. In his 1965 essay Speculations Concerning the First Ultraintelligent Machine, mathematician I. J. Good noted that designing machines is itself an intellectual activity. A sufficiently intelligent machine could design an even better one, which would in turn possess greater design ability – setting off what he called an “intelligence explosion.” Good was thinking about machines that surpassed human intelligence, but the loop he described should sound familiar by now.

That feedback loop is also what drove Jürgen Schmidhuber’s work. His 1987 thesis explored using genetic programming to improve genetic programming itself. Genetic programming searches for useful programs by generating variations and selecting among them. Schmidhuber extended that idea to the mechanisms modifying those programs: could the system discover better ways to make its own changes?

His Gödel machine proposal in 2003 made self-modification explicit. The system would search for a mathematical proof that rewriting some part of its software would be useful, then make the change once it found that proof. Its own proof-searching machinery was also available for modification. This was a mathematical design in which even the method for finding improvements could be improved.

Meanwhile, researchers were turning particular parts of AI development into learning problems.

In their 2016 paper Neural Architecture Search with Reinforcement Learning, Barret Zoph and Quoc Le trained a neural network to propose the architectures of other neural networks. The proposed networks were trained and evaluated. Their scores then supplied the feedback used to train the network proposing them. Researchers had automated part of model design, including learning which designs were worth trying.

That same year, Marcin Andrychowicz and colleagues published Learning to learn by gradient descent by gradient descent. The repetition in the title describes the idea. They used gradient descent to train an optimizer, implemented as an LSTM, that learned how to update another model’s weights. A part of the learning procedure that researchers usually designed by hand had become something a neural network could learn.

So machine learning had already automated architectural choices and learned ways to train models years before today’s research agents. And the ambition of improving the improvement process goes back further still.

For our map, the useful question is which parts of that process a system can change, and how those changes affect its subsequent work. This lets us examine Flint and Eureka more precisely: how Eureka’s discoveries helped build Flint, and how a model produced through that process could strengthen Eureka’s research abilities.

There is still another dimension to explore. A system can improve its methods while pursuing the same task and objective. Open-endedness asks how it keeps discovering useful new possibilities – including problems its designers did not specify at the start.

Become a Premium member to continue reading this episode of AI Builds AI.

Join Premium members from top companies like Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.

Follow us on

What is recursive self-improvement in AI?

It is a process in which an AI system improves its own ability to make further improvements – for example, by improving the models, tools, or research methods it uses.

What is open-endedness in AI?

Open-endedness describes the continued generation of results that are novel and learnable to an observer. In AI research, that can mean exploring new problems and using discoveries as stepping stones for later work.

What are Recursive Flint and the Eureka Machine?

Flint is an open-weight coding model designed to run locally. Eureka is Recursive’s automated research system, which the company says helped develop Flint.

Has Flint achieved recursive self-improvement?

Recursive says Eureka helped build Flint. The company describes stronger models making Eureka a better researcher – and helping build the next generation – as its goal for what comes next.