Large language model AIs might seem smart on a surface level but they struggle to actually understand the real world and model it accurately, a new study finds.
As such, it raises concerns that AI systems deployed in a real-world situation, say in a driverless car, could malfunction when presented with dynamic environments or tasks.
This is currently happening with driverless cars that use machine learning - so this goes beyond LLMs and is a general machine learning issue. Last time I checked, Waymo cars needed human intervention every six miles. These cars often times block each other, are confused by the simplest of obstacles, can’t reliably detect pedestrians, etc.
An important characteristic of a model is “stability.” Stability means that small changes in input produce small changes in output.
Stability is important for predictability. For instance, suppose you want to make a customer support portal. You add a bot hoping that it will guide the user to the desired workflow. You test the bot by asking it a bunch of variations of questions, probably with some RLHF. But then when it goes to production, people will start asking it variations of questions that you didn’t test (guaranteed). What you want ideally, is that it will map the variants to the best workflow that matches what the customer wants. Second best would be to say “I don’t know.” But what we have are bots who will just generate some crazy off-the-wall crap, and no way to prevent it.
We can’t even teach humans a coherent world model.
I would argue humans often have a world model that is too coherent. If you ask a flat earther about their beliefs they will always argue that the earth is flat and evidence to the contrary is manufactured or interpreted wrongly. That is a completely absurd world model, but perfectly coherent.
I think they mean WE struggle to understand these things have no understanding, probably because they are struggling with it also.
it guesses the next word, that is literally all it does, it’s not trying to build a model of reality to more accurately guess. It has no fidelity and anyone taking it seriously has themselves failed the turing test.