Chapotraphouse

I'd just like to interject for a moment. What you're referring to as hallucinations is, in fact, a missunderstanding of LLMs/LLM Assistants, or, as I've recently taken to calling it, the "creativity vs. accuracy problem." Hallucinations are not an issue unto themselves but rather the expected behavior of a fully functioning LLM made useful by an LLM Assistant, a system often more complex than just the LLM itself, comprised of a GUI frontend and systems and logic backend that simulates the experience of conversational interaction utilizing an LLM.

Many AI users are consuming the hallucinations of an LLM system every day without realizing it. Through a peculiar turn of events, the term hallucinations, which is widely used today, is often used to identify when the output goes into deemed factually incorrect territory, and many of its users are not aware that LLMs are basically dream machines, where we direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful.

There really are hallucinations, and these people are consuming them, but it is just a natural byproduct of the system, not a "bug," but just the LLM doing what it always does. LLMs do not have a "hallucination problem" because, in some sense, hallucination is all LLMs do. The LLM is an essential part of the whole system but not as useful by itself; it can only become more useful in the context of a LLM Assistant. There are many ways to mitigate hallucinations in these systems - using Retrieval Augmented Generation (RAG) to more strongly anchor the dreams in real data through in-context learning is maybe the most common one. Disagreements between multiple samples, reflection, verification chains. Decoding uncertainty from activations. Tool use. All an active and very interesting areas of research.

All that said, I realize that what people actually mean is they don't want an LLM Assistant (a product like ChatGPT etc.) to hallucinate. I know I'm being super pedantic but the LLM has no "hallucination problem." Hallucination is not a bug; it is LLM's greatest feature. The LLM Assistant has a hallucination problem, and we should fix it.