(“…a computer doesn’t, like, know what an apple is maaan…”)
I think you’re misunderstanding and/or deliberately misrepresenting the point. The point isn’t some asinine assertion, it’s a very real fundamental problem with using LLMs for any actually useful task.
If you ask a person what an apple is, they think back to their previous experiences. They know what an apple looks like, what it tastes like, what it can be used for, how it feels to hold it. They have a wide variety of experiences that form a complete understanding of what an apple is. If they have never heard of an apple, they’ll tell you they’ve never heard of it.
If you ask an LLM what an apple is, they don’t pull from any kind of database of information, they don’t pull from experiences, they don’t pull from any kind of logic. Rather, they generate an answer that sounds like what a person would say in response to the question, “What is an apple?” They generate this based on nothing more than language itself. To an LLM, the only difference between an apple and a strawberry and a banana and a gibbon is that these things tend to be mentioned in different types of sentences. It is, granted, unlikely to tell you that an apple is a type of ape, but if it did it would say it confidently and with absolutely no doubt in its mind, because it doesn’t have a mind and doesn’t have doubt and doesn’t have an actual way to compare an apple and a gibbon that doesn’t involve analyzing the sentences in which the words appear.
The problem is that most of the language-related tasks which would be useful to automate require not just text which sounds grammatically correct but text which makes sense. Text which is written with an understanding of the context and the meanings of the words being used.
An LLM is a very convincing Chinese room. And a Chinese room is not useful.
As an example of that, try asking a LLM questions about precise details about the lore of a fictional universe you know well, and you know that what you’re asking about hasn’t ever been detailed.
Not Tolkien because this has been too much discussed on the internet. Pick a universe much more niche.
It will completely invent stuff that kinda makes sense. Because it’s predicting the next words that seem likely in the context.
A human would be much less likely to do this because they’d just be able to think and tell you “huh… I don’t think the authors ever thought about that”.
About the only useful task an LLM could have is generating random NPC dialog for a video game. Even then, it’s close to the least efficient way to do it.
There’s a lot of stuff it can do that’s useful, just all malicious. Anything which requires confidently lying to someone about stuff where the fine details don’t matter. So it’s a perfect tool for scammers.
As a counter: you only think you know what an apple is. You have had experiences interacting N instances of objects which share a sufficient set of “apple” characteristics. I have had similar experiences, but not identical. You and I merely agree that there are some imprecise bounds of observable traits that make something “apple-ish”.
Imagine someone who has never even heard of an apple. We put them in a class for a year and train them on all possible, quantifiable traits of an apple. We expose them, in isolation, to:
all textures an apple can have, from watery and crisp to mealy
all shades of apple colors appearing in nature and at different stages of their existence
a 1:1 holographic projection of various shapes of apple
sample weights, similar to the volume and density of an apple
distilled/artificial flavors and smells for all apple variations
extend this training on all parts of the apple (stem, skin, core, seeds)…
You can go as far as you like, giving this person a PhD in botanical sciences, just as long as nothing they experience is a combination of traits that would normally be described as an apple.
Now take this person out of the classroom and give them some fruit. Do they know it’s an apple? At what point did they gain the knowledge; could we have pulled them out earlier? What if we only covered Granny Smith green apples, is their tangential expertise useless in a conversation about Gala apples?
This isn’t even so far fetched. We have many expert paleontologists and nobody has ever seen a dinosaur. Hell, they generally don’t even have real, organic pieces of animals. Just rocks in the shape of bones, footprints, and other tangential evidence we can find in the strata. But just from their narrow study, they can make useful contributions to other fields like climatology or evolutionary theory.
We could, with our current tech and enough resources, make something that matches the complexity of the human brain. You just need a shit ton of processing power and lots of well groomed data. With even more dedication we might match the dynamic behavior, mirroring the growth and development of the brain (though that’s much harder). Would it be as efficient and robust as a normal brain? Probably not. But it could be indistinguishable in function; just as fallible as any human working from the same sensory input.
At a higher complexity it ceases being a toy Chinese Room and turns into a Philosophical Zombie. But if it can replicate the reactions of a human… does intentionality, personhood or “having a mind” matter? Is it any less useful than, say, an average employee who might fuck up an email or occasionally fail to grasp a problem or be sometimes confidently incorrect?
I think you’re misunderstanding and/or deliberately misrepresenting the point. The point isn’t some asinine assertion, it’s a very real fundamental problem with using LLMs for any actually useful task.
If you ask a person what an apple is, they think back to their previous experiences. They know what an apple looks like, what it tastes like, what it can be used for, how it feels to hold it. They have a wide variety of experiences that form a complete understanding of what an apple is. If they have never heard of an apple, they’ll tell you they’ve never heard of it.
If you ask an LLM what an apple is, they don’t pull from any kind of database of information, they don’t pull from experiences, they don’t pull from any kind of logic. Rather, they generate an answer that sounds like what a person would say in response to the question, “What is an apple?” They generate this based on nothing more than language itself. To an LLM, the only difference between an apple and a strawberry and a banana and a gibbon is that these things tend to be mentioned in different types of sentences. It is, granted, unlikely to tell you that an apple is a type of ape, but if it did it would say it confidently and with absolutely no doubt in its mind, because it doesn’t have a mind and doesn’t have doubt and doesn’t have an actual way to compare an apple and a gibbon that doesn’t involve analyzing the sentences in which the words appear.
The problem is that most of the language-related tasks which would be useful to automate require not just text which sounds grammatically correct but text which makes sense. Text which is written with an understanding of the context and the meanings of the words being used.
An LLM is a very convincing Chinese room. And a Chinese room is not useful.
As an example of that, try asking a LLM questions about precise details about the lore of a fictional universe you know well, and you know that what you’re asking about hasn’t ever been detailed.
Not Tolkien because this has been too much discussed on the internet. Pick a universe much more niche.
It will completely invent stuff that kinda makes sense. Because it’s predicting the next words that seem likely in the context.
A human would be much less likely to do this because they’d just be able to think and tell you “huh… I don’t think the authors ever thought about that”.
About the only useful task an LLM could have is generating random NPC dialog for a video game. Even then, it’s close to the least efficient way to do it.
There’s a lot of stuff it can do that’s useful, just all malicious. Anything which requires confidently lying to someone about stuff where the fine details don’t matter. So it’s a perfect tool for scammers.
As a counter: you only think you know what an apple is. You have had experiences interacting N instances of objects which share a sufficient set of “apple” characteristics. I have had similar experiences, but not identical. You and I merely agree that there are some imprecise bounds of observable traits that make something “apple-ish”.
Imagine someone who has never even heard of an apple. We put them in a class for a year and train them on all possible, quantifiable traits of an apple. We expose them, in isolation, to:
You can go as far as you like, giving this person a PhD in botanical sciences, just as long as nothing they experience is a combination of traits that would normally be described as an apple.
Now take this person out of the classroom and give them some fruit. Do they know it’s an apple? At what point did they gain the knowledge; could we have pulled them out earlier? What if we only covered Granny Smith green apples, is their tangential expertise useless in a conversation about Gala apples?
This isn’t even so far fetched. We have many expert paleontologists and nobody has ever seen a dinosaur. Hell, they generally don’t even have real, organic pieces of animals. Just rocks in the shape of bones, footprints, and other tangential evidence we can find in the strata. But just from their narrow study, they can make useful contributions to other fields like climatology or evolutionary theory.
An LLM only happens to be trained on text because it’s cheap and plentiful, but the framework of a neural network could be applied to any data. The human brain consumes about 125MB/s in sensory data, conscious thought grinds at about 10 bits/s, and each synapse could store about 4.7 bits of information for a total memory capacity in the range of ~1 petabyte. That system is certainly several orders of magnitude more powerful than any random LLM we have running in a datacenter, but not out of the realm of possibility.
We could, with our current tech and enough resources, make something that matches the complexity of the human brain. You just need a shit ton of processing power and lots of well groomed data. With even more dedication we might match the dynamic behavior, mirroring the growth and development of the brain (though that’s much harder). Would it be as efficient and robust as a normal brain? Probably not. But it could be indistinguishable in function; just as fallible as any human working from the same sensory input.
At a higher complexity it ceases being a toy Chinese Room and turns into a Philosophical Zombie. But if it can replicate the reactions of a human… does intentionality, personhood or “having a mind” matter? Is it any less useful than, say, an average employee who might fuck up an email or occasionally fail to grasp a problem or be sometimes confidently incorrect?
4.7 bits per neuron, 10 bits in thinking, huh? This is good information, thank you :3