I like the goal of this. As expected, I don't really understand the math/concept of this. It sounds like it caches some neural network activity and exports it to be run later. So I suppose this can't be used for things like image or video generation.

This looks cool, but I wonder how well their trained compiler generalizes to new task families. They trained on 29 specific types of tasks, with 800 sub tasks and many rephrasings of each one (the specs). They hold out some specs for validation, but don’t seem to have held out a full task family and maybe not even full sub tasks?

If the compiler can’t generalize well to unseen tasks then it’s effectively acting as a fancy router to one of 29/800 predefined LoRAs.

> PAW reframes the foundation model from a per-input problem solver into a tool builder: invoked once per function definition, it produces a small reusable artifact whose subsequent calls per function application are cheap and offline.

Umm you can just get the LLM to spit out real functions instead of fuzzy functions and just run those real functions through real interpreters, which is also "cheap" and "offline".