Post
LLMs are token-processing engines. *all they do* is take in tokenized input and spit out output. I say "input and output" because it's not even *language,* the *language* can only happen through human interpretation of those inputs and outputs, externally to the model.
5 reposts47 likes