Over the last year or two, AI coding tools have become good enough that it no longer makes sense to dismiss them outright. They can explain unfamiliar code, generate boilerplate, scaffold new features, write tests, and often get you 70% of the way through tasks that would have previously required uninterrupted focus. Even when they are wrong, they are frequently wrong in useful ways. They can give you momentum. At the same time, after using them regularly, I kept running into a much less impressive limitation. They are surprisingly bad at remembering things that are already known.

While building Theta, a functional programming language that compiles to WebAssembly, I encountered an interesting challenge. In functional languages like Theta, first-class functions and closures are fundamental. They allow functions to be passed as arguments, enabling core abstractions such as map, reduce, and filter.