Prediction works based on the attention mechanism, and current humans don't speak like cavemen - so how could you expect a useful token chain from data that isn't trained on speech like that?
I get the concept of transformers, but this isn't doing a 1:1 transform from english to french or whatever, you're fundamentally unable to represent certain concepts effectively in caveman etc... or am I missing something?
Okay maybe not exactly caveman dialect, but text compression using LLM is definitely possible to save on tokens in deep research.