Post

LLM fine tunning to reduce token use in writing code tasks is bad, because it generates a more complex code by trying to "compress"/optimize the use of sequences that are costly for LLMs, but helpful for humans.