Of course the “understanding” of an LLM is limited. Because the entire technology is new, and it’s far from being anywhere close to being able to understand to the level of a human.
But I disagree with your understanding of how an LLM works. At its lower level, it’s a bunch on connected artifical neurons, not that different from a human brain. Now please don’t read this as me saying it’s as good as a human brain. It’s definitely not, but its inner workings are not so far. As a matter of fact, there is active effort to make artificial neurons behave as close as possible to a human neuron.
If it was just statistics, it wouldn’t be so difficult to look at the trained model and identify what does what. But just like the human brain, it is incredidbly difficult to understand that. We just have a general idea.
So it does understand, to a limited extent. Just like a human, it won’t understand what it hasn’t been exposed to. And unlike a human, it is exposed to a very limited set of data.
You’re putting the difference between a human’s “understanding” and an LLM’s “understanding” in the meaning of the word “understanding”, which is just a shortcut to say that they can’t be compared. The actual difference is in the scope of understanding.
A lot of the efforts in the AI fields gravitate around imitating a human brain. Which makes sense, as it is the only thing we know that is capable of doing what we want an AI to do. LLMs are no different, but their scope is limited.
I think comparing an LLM to a brain is a category mistake. LLMs aren’t designed to simulate how the brain works - they’re just statistical engines trained on language. Trying to mimic the human brain is a whole different tradition of AI research.
An LLM gives the kind of answers you’d expect from something that understands - but that doesn’t mean it actually does. The danger is sliding from “it acts like” to “it is.” I’m sure it has some kind of world model and is intelligent to an extent, but I think “understands” is too charitable when we’re talking about an LLM.
And about the idea that “if it’s just statistics, we should be able to see how it works” - I think that’s backwards. The reason it’s so hard to follow is because it’s nothing but raw statistics spread across billions of connections. If it were built on clean, human-readable rules, you could trace them step by step. But with this kind of system, it’s more like staring into noise that just happens to produce meaning when you ask the right question.
I also can’t help laughing a bit at myself for once being the “anti-AI” guy here. Usually I’m the one sticking up for it.
Of course the “understanding” of an LLM is limited. Because the entire technology is new, and it’s far from being anywhere close to being able to understand to the level of a human.
But I disagree with your understanding of how an LLM works. At its lower level, it’s a bunch on connected artifical neurons, not that different from a human brain. Now please don’t read this as me saying it’s as good as a human brain. It’s definitely not, but its inner workings are not so far. As a matter of fact, there is active effort to make artificial neurons behave as close as possible to a human neuron.
If it was just statistics, it wouldn’t be so difficult to look at the trained model and identify what does what. But just like the human brain, it is incredidbly difficult to understand that. We just have a general idea.
So it does understand, to a limited extent. Just like a human, it won’t understand what it hasn’t been exposed to. And unlike a human, it is exposed to a very limited set of data.
You’re putting the difference between a human’s “understanding” and an LLM’s “understanding” in the meaning of the word “understanding”, which is just a shortcut to say that they can’t be compared. The actual difference is in the scope of understanding.
A lot of the efforts in the AI fields gravitate around imitating a human brain. Which makes sense, as it is the only thing we know that is capable of doing what we want an AI to do. LLMs are no different, but their scope is limited.
I think comparing an LLM to a brain is a category mistake. LLMs aren’t designed to simulate how the brain works - they’re just statistical engines trained on language. Trying to mimic the human brain is a whole different tradition of AI research.
An LLM gives the kind of answers you’d expect from something that understands - but that doesn’t mean it actually does. The danger is sliding from “it acts like” to “it is.” I’m sure it has some kind of world model and is intelligent to an extent, but I think “understands” is too charitable when we’re talking about an LLM.
And about the idea that “if it’s just statistics, we should be able to see how it works” - I think that’s backwards. The reason it’s so hard to follow is because it’s nothing but raw statistics spread across billions of connections. If it were built on clean, human-readable rules, you could trace them step by step. But with this kind of system, it’s more like staring into noise that just happens to produce meaning when you ask the right question.
I also can’t help laughing a bit at myself for once being the “anti-AI” guy here. Usually I’m the one sticking up for it.