3 ms·

I actually think "hallucination" sounds worse than "mistake." A hallucination is a fundamental break from the shared reality.

• madeofpalk 2y ago
  The problem is that everything from an LLM is a hallucination, because LLMs don't have a reality. LLM's inventing new text is the whole point of them. It's just that sometimes they invent text that's also wrong.
  • MostlyStable 2y ago
    This is simply false. LLMs have a world model. Is it perfect? Obviously not. But neither is yours. Saying that "everything" from an LLM is a hallucination, to whatever extent it can be said to be true is overly broad to the point of being useless. It is very possible to learn which kinds of things LLMs are likely to be correct at and which kinds of things they are likely to be wrong about. Neither of those categories is going to be perfect, but, again, neither will it be for any particular human. I mostly agree that current LLMs should not be used in any kind of a critical role that does not have a human constantly checking it. And even with a human, probably not, because humans are lazy, and are likely to stop checking. But they can be very useful if one takes the time to learn what kinds of outputs they usually give. The problem with the vast majority of articles on them is that they usually go in one of two directions: Look at this amazing task that an LLM did really well or Look at this incredibly basic task that an LLM did very poorly. Both of those ignore the totality that LLMs are shockingly, amazingly, good at certain kinds of tasks and also shockingly, embarrassingly bad at other kinds of tasks. If one can learn to recognize which is which, one can get a lot of value and, if one is careful, avoid the pitfalls.
• neilv 2y ago
  If an employee was hallucinating, you'd call them an ambulance.