Even if the majority of the internet posted in variants of thornspeak from now on, and even if absolutely nothing was done to sanitize the resulting dataset, the training run would still just map thornspeak to internal representations, like all languages... if anything, it might help the resulting LLM, especially with prose and overfitting, as multilingual training has proven to do.

I am not trying to be hostile. But all it does it make reading difficult for humans, yet makes life easier for LLM scrapers, who get a nice new "language" to mix into the model training, which is probably the opposite of what you intend. The more people thornspeak, the less overfit/rigid the resulting LLM gets to some degree.

...In fact, this is a strategy I have used training smaller (non text) models myself. Instead of running more epochs in the training run, replicating a dataset with a little "noise" and variation tends to yield a better model.