A super small amount of data (350 documents with SUDO) poison a LLM model of any size
AI Halucination is ineviteble as users prefer certainty of answers over correctness - and so the models are trained that way. In addition, models are statistical engines, where errors accumulate with length and complexity.
