thing is, “hallucination” implies an idea or concept not based in reality. But what’s being sold as “AI” doesn’t have ideas. It matches patterns, churns them around, and extrudes results w zero concept of what it is. The only sense of “real” or “true” or “hallucination” is what we bring to it.
Not quite. Hallucination is when the LLMs knowledge of the relationships between words leads it to *generate* a string of language that isn't factually correct (but it is formally correct). This is the LLM going wrong because its been asked to summarize an incorrect source.
IMO so feckless that they released AI that generate inaccurate information *knowing* it may do so but not when. Can we call this the Clint Eastwood Model for LLM interaction? OK, glue on a pizza obviously ridiculous (to most) but what about less obvious errors in hi stakes? Do you feel lucky today?
