3 ms·

> So? That's how you learned everything you ever read in a book. Proof please.

IceMetalPunk 4y ago
Imagine a person is born blind, and from birth is kept in a room and never allowed outside. They are provided with food and a bed, along with a library of audiobooks to listen to whenever they want. From those audiobooks, they learn all about the world around them: about trees, and grass, and wind, and water, and buildings, and birthdays (they've never celebrated), and fireworks, and puppies, etc. Would you say that, because this person has never experienced these things other than via language, they don't -- and can't possibly -- understand anything about the world? I would argue they certainly do understand, and an LLM is doing the same thing. In a less extreme, hopefully more relatable, example: think back to when you were in school. Everything you ever read in a textbook or on a whiteboard/blackboard or in a homework packet. Are you telling me that either (a) you didn't learn any of it, or (b) you personally experienced everything you read about in order to learn it? Because if there's something you understood by reading about it, despite having no other sensory experience of it, then you learned by finding patterns in the language -- you must have, for it was the only data available to you. Yes, multi-modal learning has the potential to enhance that learning greatly (hence why multi-modal models are all the rage this year), but humans do in fact learn through language alone sometimes, and we still call that understanding.
- righttoolforjob 4y ago
  Thanks for the expanded thoughts. I see two problems: 1) You seem to include any kind of processing in "finding patterns". I wanted to both get your definition for "finding patterns" as well as the proof that this is how humans learn. 2) In the example of reading a book in class you fail to include the past history of experiences/learnings that takes place outside of class and which helps you to put the words in context. This might very well be a major differentiating factor compared to the first example. To give a silly example: If I kick a wall and get a pain response through my toes which teaches me to not do that again, did I learn this by observing a pattern? Is that really a useful model for how the learning took place? Surely there was a pattern, perhaps even several ones, no denying that, but I highly question the accuracy and thus general usefulness of this model for learning. Depending on definition, finding patterns could mean just about anything, thus meaning nothing.
  - IceMetalPunk 4y ago
    I guess I'd pose the question to you, then: for that first example of the blind person who has experienced basically nothing other than language, what more could they learn from beyond finding patterns in the sounds of language? If they hear "Leaves are green", but they've never seen colors and have never been around anything like a tree or leaves, what other methods could they possibly use to learn that leaves are green beyond just recognizing that the sounds of the word "green" come after the sounds of the words "leaves are"? As for the second example, while I agree that external context improves understanding, do you really think it's required for understanding? For instance, I took AP chemistry in high school, and much of that was learning about quantum mechanics, electronic configurations, etc. I have zero external experience for any of that -- hell, at the time, I hadn't even ever seen any microscopy of an electron cloud. So there's no context for those concepts outside of knowing how the words fit together and deriving meaning from that. Yet I'd still claim I learned the material (and my grades in that course would agree).

Imagine a person is born blind, and from birth is kept in a room and never allowed outside. They are provided with food and a bed, along with a library of audiobooks to listen to whenever they want. From those audiobooks, they learn all about the world around them: about trees, and grass, and wind, and water, and buildings, and birthdays (they've never celebrated), and fireworks, and puppies, etc. Would you say that, because this person has never experienced these things other than via language, they don't -- and can't possibly -- understand anything about the world? I would argue they certainly do understand, and an LLM is doing the same thing. In a less extreme, hopefully more relatable, example: think back to when you were in school. Everything you ever read in a textbook or on a whiteboard/blackboard or in a homework packet. Are you telling me that either (a) you didn't learn any of it, or (b) you personally experienced everything you read about in order to learn it? Because if there's something you understood by reading about it, despite having no other sensory experience of it, then you learned by finding patterns in the language -- you must have, for it was the only data available to you. Yes, multi-modal learning has the potential to enhance that learning greatly (hence why multi-modal models are all the rage this year), but humans do in fact learn through language alone sometimes, and we still call that understanding.

righttoolforjob 4y ago
Thanks for the expanded thoughts. I see two problems: 1) You seem to include any kind of processing in "finding patterns". I wanted to both get your definition for "finding patterns" as well as the proof that this is how humans learn. 2) In the example of reading a book in class you fail to include the past history of experiences/learnings that takes place outside of class and which helps you to put the words in context. This might very well be a major differentiating factor compared to the first example. To give a silly example: If I kick a wall and get a pain response through my toes which teaches me to not do that again, did I learn this by observing a pattern? Is that really a useful model for how the learning took place? Surely there was a pattern, perhaps even several ones, no denying that, but I highly question the accuracy and thus general usefulness of this model for learning. Depending on definition, finding patterns could mean just about anything, thus meaning nothing.
- IceMetalPunk 4y ago
  I guess I'd pose the question to you, then: for that first example of the blind person who has experienced basically nothing other than language, what more could they learn from beyond finding patterns in the sounds of language? If they hear "Leaves are green", but they've never seen colors and have never been around anything like a tree or leaves, what other methods could they possibly use to learn that leaves are green beyond just recognizing that the sounds of the word "green" come after the sounds of the words "leaves are"? As for the second example, while I agree that external context improves understanding, do you really think it's required for understanding? For instance, I took AP chemistry in high school, and much of that was learning about quantum mechanics, electronic configurations, etc. I have zero external experience for any of that -- hell, at the time, I hadn't even ever seen any microscopy of an electron cloud. So there's no context for those concepts outside of knowing how the words fit together and deriving meaning from that. Yet I'd still claim I learned the material (and my grades in that course would agree).