https://chatgpt.com/g/g-2fkFE8rbu-dall-e/c/6772fc7f-2754-800d-ab66-c2271a4fbde9
I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines.
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.
It's only when the dreams go into deemed factually incorrect territory that we label it a "hallucination". It looks like a bug, but it's just the LLM doing what it always does.
I hate the term "hallucination" when applied to Large Language Models (LLMs). Hallucination implies that there is some sort of entity that is mis-perceiving reality, like a teenager in a bad 60s LSD movie. In one sense LLMs can't hallucinate, there is no entity to mis-perceive anything. When an LLM provides information at odds with reality, for example a reference to an article that doesn't exist, it's not misperceiving reality, it's simply doing what it does. LLMs build strings of words that make some sort of sense to humans. Those word strings may or may not conform to our human perception of reality
LLMs don't have a model of reality. They have a model of language as trained from the internet. In whatever sense internet language reflects reality, that's as close to real world knowledge as an LLM will get.
In order to make LLMs more accurate, they need to be connected to other modules that more closely model reality. For example, to do arithmetic without errors, the LLM needs a calculator agent.
No comments:
Post a Comment