Understanding why something was said, and not only what the words mean, is central to human communication. When this action was proposed, such pragmatic reasoning was a recognised frontier for artificial intelligence. Large language models have since transformed what machines can do with language, but by a route this project deliberately did not take: training very large statistical models on very large corpora, once, then deploying them frozen. The questions that motivated the action remain open. How can a system learn from little data, keep adapting after deployment, and hold its knowledge in a form a human can inspect?
Rather than designing the learning machinery by hand, the project asked whether it could be made to emerge and evolve. The starting point was Combinatory Chemistry, an artificial chemistry in which elementary program fragments drift in a simulated "soup", combine at random and, being pieces of a Turing-complete language, also compute on one another. Self-maintaining cycles of reactions arise spontaneously, and some reproduce. The plan was to make that chemistry evolve by running it at scale and letting its computations make occasional errors, then to steer it by feeding the soup when it behaves usefully, giving a learning rule in which the evolving population does its own exploring.
Everything rested on the first step, a problem open in artificial life for decades: what conditions give rise to an open-ended evolutionary process, one that keeps producing novel and more complex forms indefinitely?
