so we have this program that takes some abstract schema and generates a bunch of different representations of it. when i started working on it, there was a bunch of spaghetti procedural string munging code in every renderer. So i wrote this templating system to separate the different concerns: the procedural translation between domains from the declarative representation in the target domain - "how to make" from "what is made". This makes a nice programming surface that can be read and written, the last human being to do any real work with it before "AI" was very pleased with it, saying "you really cooked with this" after having worked with the more spaghettilike code.
we have reached the point of LLMs where their utility is dramatic brute force, fable is good because it just keeps generating until it gets it right. since this templating system is not oversampled in the training data, though, and the documentation is in a slightly different location than the code, fable just brute forces through it and not only undoes the work of the templating system, but uses the templating system to introduce two novel forms of spaghetti and code divergence.
so anyway. asbestos et al. actively undoing any design work. a pulp mill that pressure-casts every problem into statistical boilerplate broth. fractal regression as a service
https://github.com/linkml/linkml/pull/3818#pullrequestreview-4990129294