Raw LLM Responses

it's scary people trust in mathematical functions that predict multinomial distributions so much. that is literally all LLMs are, where each layer itself could be considered a function and the entire model a composite function. Very powerful mathematical functions, but you could legitimately write them out and compute them on paper if you had enough time (would be insane of course especially with LLMs but still, technically possible). Basically searching through the sea of chaos via stochastic gradient descent for a mathematical function that no one fully understands that can generate multinomial distributions such that when sampled from allow for believable text generation.