if you treat it like stochastic software it is, how is 95% chance of one-shotting a problem not useful? I feel like the biggest problem most people have with LLMs is the "lack of accountability", and then the same people say "don’t anthropomorphize them"