LLMs are, fundamentally, compressed lookup tables that map input -> input + next token. Or, If you like, input -> input + list of possible next tokens with prob

LLMs are, fundamentally, compressed lookup tables that map input -> input + next token. Or, If you like, input -> input + list of possible next tokens with probabilities.