Recur

About this project
Recur started with a question I've become increasingly interested in: what happens when an AI system encounters the same kind of problem again and again? Most of the time, we ask a model to solve the problem again, even when earlier interactions may already contain enough information to do something more reusable. I wanted to explore whether those interactions could become small, verifiable programs instead.
Recur collects demonstrations of recurring tasks and uses them as evidence for program synthesis. Its compiler searches a typed dataflow language made up of readers, transforms, and emitters, looking for a program that can reproduce the observed computation. Rather than accepting a candidate because it looks plausible, the system executes it against the original examples and checks whether its behaviour holds up under further verification.
That verification process includes consistency checks, leave-one-out recompilation, generated in-domain inputs, and explicit decline behaviour. Accepted programs are stored in a registry alongside their demonstrations and verification results. When a future request matches one of those programs, Recur can offer the reusable solution and, once accepted, execute it locally through a deterministic virtual machine without another model call.
The distinction between generating an answer and establishing that an answer is correct is central to the project. A language model can provide useful evidence about a task, but its wording alone does not establish the mathematical or logical truth of the result. Recur explores an alternative in which reusable computation must stand on its own evidence, with the system able to decline or fall back to a verified lookup when it cannot find a reliable solution.
This is one of my more exploratory projects, and I enjoy it precisely because it raises questions that go beyond implementing a familiar application. It has pushed me to think about program synthesis, verification, the limits of model-generated answers, and what it might mean for an AI system to become more useful through experience rather than simply repeating the same process.