Daily AI Paper

Today’s pick is Compile by Training: Turning Natural-Language Specifications into Local Neural Functions. The problem it tackles is simple to state but hard to engineer: many useful text-processing tasks can be described in plain language, yet implementing them as rules is brittle, and sending every request to a large remote model is slow, expensive, and dependent on an outside service. The core idea is to treat a natural-language spec like something you can compile. At compile time, stronger teacher models generate examples for the task, and those examples train a small adapter inside a compact interpreter. After that, the function runs locally, without the teachers, and can be saved, versioned, and reused like ordinary software. That matters because it turns prompt-like behavior into a real asset you can deploy, compose, and control. The paper reports strong results on a difficult benchmark and shows practical demos, from website helpers to language-controlled avatars and translation tools.