I have many conversations with people about Large Language Models like ChatGPT and Copilot. The idea that “it makes convincing sentences, but it doesn’t know what it’s talking about” is a difficult concept to convey or wrap your head around. Because the sentences are so convincing.
Any good examples on how to explain this in simple terms?
Edit:some good answers already! I find especially that the emotional barrier is difficult to break. If an AI says something malicious, our brain immediatly jumps to “it has intent”. How can we explain this away?
So there’s two different things to what you are asking.
(1) They don’t know what (i.e. semantically) they are talking about.
This is probably not the case, and there’s very good evidence over the past year in research papers and replicated projects that transformer models do pick up world models from the training data such that they are aware and integrating things at a more conceptual level.
For example, even a small toy GPT model trained only on chess moves builds an internal structure of the whole board and tracks “my pieces” and “opponent pieces.”
(2) Why do they say dumb shit that’s clearly wrong and don’t know.
They aren’t knowledge memorizers. They are very advanced pattern extenders.
Where the answer to a question is part of the pattern they can successfully extend, they get the answer correct. But if it isn’t, they confabulate an answer in a similar way to stroke patients who don’t know that they don’t know the answer to something and make it up as they go along. Similar to stroke patients, you can even detect when this is happening with a similar approach (ask 10x and see how consistent the answer is or if it changes each time).
They aren’t memorizing the information like a database. They are building ways to extend input into output in ways that match as much information as they can be fed. In this, they are beyond exceptional. But they’ve been kind of shoehorned into the initial tech demo usecase of “knowledgeable chatbot” which is a less than ideal use. The fact they were even good at information recall was a surprise to most researchers.
Thanks for your thorough answer.
I’ll see if I can find that article/paper about the chess moves. That sounds interesting!
Could it be that we ascribe an LLM with conceptual knowledge while in fact it is by chance? We as humans are masters at seeing patterns that aren’t there. But then again, like another commenter said, maybe the question is more about conscience itself, and what that actually means. What it means to “understand” something.
So the paper that found that particular bit in Othello was this one: https://arxiv.org/abs/2310.07582
Which was building off this earlier paper: https://arxiv.org/abs/2210.13382
And then this was the work replicating it in Chess: https://www.lesswrong.com/posts/yzGDwpRBx6TEcdeA5/a-chess-gpt-linear-emergent-world-representation
It’s not by chance - there’s literally interventions where flipping a weight or vector results in the opposite behavior (like acting like a piece is in a different place, or playing well he badly no matter the previous moves).
But it’s more that it seems unlikely that there’s any actual ‘feeling’ or ‘conscious’ sentience/consciousness to understand beyond the model knowing what the abstracted pattern means in relation to the inputs and outputs. It probably is simulating some form of ego and self, but not actively experiencing it if it makes sense.