i wonder if i can use the adapter finetuning stuff to feel like i am moving forward on the semibalanced trees a little
i'm imagining a model that mutates code in a way that is helpful: stabilising it, simplifying it, anything helpful for the goal if the code were small enough, it might fit right into a language model. there may even be long-context language models on huggingface, like transformers-xl, that adapter transformers could just load right up
