5 ms·

I’ve seen this argument over and over again: it’s just predicting the next word, it’s not creating anything new, etc. How do we know we’re not doing the same o

I’ve seen this argument over and over again: it’s just predicting the next word, it’s not creating anything new, etc. How do we know we’re not doing the same or similar in our brains? We can’t even define creativity, intelligence, let alone tell how it works. Predicting the next word is very much part of an intelligent discourse. What if it’s all just a “computational guessing game” all the way down with perhaps a sprinkle of randomness added? I’m very concerned with the explosion in AI development but I don’t think we should discredit the tool because we think it’s dangerous. Quite the contrary, in fact. We should take it very seriously.

qotgalaxy 4y ago
> How do we know we’re not doing the same or similar in our brains? How do we know that we're not made of magic?
emodendroket 4y ago
That would seem to be the “stochastic parrot” view denounced. But it seems hard to believe. Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? How could that be if we’re just randomly selecting words like a souped-up Markov chain?
- tambourine_man 4y ago
  We’re not randomly selecting but probabilistically predicting. Haven’t you noticed you predict the next word people will say with very high [________]?
  - peteradio 4y ago
    Sandwiches!
  - emodendroket 4y ago
    Within a sentence sure. And often arguments are familiar enough that even an essay I’ll know where it’s going. But we do occasionally have original ideas.
    - tambourine_man 4y ago
      I don’t think we know what original really is to make a meaningful distinction between derivative work. With enough mixing and matching and a tiny bit of randomness, I bet a lot of original thoughts could be created.
      - emodendroket 4y ago
        These things are hard to quantify but ChatGPT writing does have a certain anodyne character, in my mind, when I read it. Then again, I can't quantify it, I could be biased by my foreknowledge of the writing's provenance, and even if I'm right, the world is full of situations where bland writing is suitable or even desired.
        tambourine_man 4y ago
        It absolutely does. It’s wishy-washy nature is harder to disagree with, sounds more credible, etc.
- Kamq 4y ago
  > Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? How could that be if we’re just randomly selecting words like a souped-up Markov chain? All words in the normal cache have a low probability, so you have to hit disk to get the full word list.
- waselighis 4y ago
  > Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? I came here to suggest this exact thing. Humans clearly have some understanding of the meaning behind words and sentences. However, I think it's wrong, an overgeneralization, to suggest that ChatGPT is just statistically predicting the next word. While that may technically be true, I think buried deep within it has ways of modeling/encoding common concepts and ideas, maybe similar to how a human models concepts and ideas in their mind. Then there's the whole problem of consciousness, but I won't get into that here.
  - tambourine_man 4y ago
    >I think buried deep within it has ways of modeling/encoding common concepts and ideas, maybe similar to how a human models concepts and ideas in their mind Right, maybe meaning and concepts are just “things that usually go together” with different levels of distance. We’re comparing our minds, which we know very little of, to LLM, which are blackboxes. It seems like a lot of certainty in face of such vast ignorance.
arbuge 4y ago
Indeed, one might conclude that predicting what to do in the next moment (analogous to what to write next in the case of ChatGPT, which can only write) is all we do as humans. At every moment of the day, we are taking all our past experiences and the present as inputs, and predicting what we should do next based on that.
zyklonix 4y ago
This is in fact the whole premise of Jeff Hawkins' book "On intelligence". Prediction is at the core of how we learn/act. Since we are born, we are just predicting everything that will happen and our brain learns when we see/hear/sense something that we can't predict.
muzani 4y ago
I love the philosophy behind Gustafson's Law: the problem size increases in response to computational resources. You get a simple computer, you can drop bombs more accurately. Then you make word processors. Then text based simulation aka games. Then turn books into digital format. Then whole online libraries, ecommerce, social media, cryptocurrency, and so on. Seeing a LLM as purely language based might be missing the forest for the atoms.
akomtu 4y ago
A token generator corresponds to reflexes: a fuzzy hash table of situations and responses. Cats and dogs have that. Assigning meanings to tokens and establishing relationships between them is where intelligence begins. GPT continues sequence CGGA with T because there is a similar sequence in its dataset, but we interpret T as a tree and see how this meaning relates to A. I believe GPT will solve this problem to some extent in 10 years and will finally earn its real AI badge. Above it there is abstract mind that only top scientists use to some extent.
root_axis 4y ago
I won't speak for what's "all the way down", but on whatever level humans perform reflective reasoning, it's self-evident that we don't simply "predict the next word" in the way a LLM does. Whether or not a LLM is capable of "creating anything new" is a different argument altogether. "New" is in eye of the beholder and it is often said that nothing is new under the sun, but we already know that unthinking processes within nature are capable of producing beauty far and beyond anything that even the greatest human artist could ever hope to accomplish. That a large LLM is unthinking doesn't preclude it from producing art.
- tambourine_man 4y ago
  Please provide evidence that intelligent speech isn’t simply predicting the next word, as it’s obviously not self-evident to me.
  - lukev 4y ago
    I'm not sure it counts as "evidence" because I can't show it to you, but based on personal introspection, whenever I write something it starts with a barely-linguistic thought I want to communicate that gradually gets expanded into as much prose as is necessary to get across the concept whilst establishing context and defending from the most probable detractors. I literally never just start writing and then just keep adding words.
    - tambourine_man 4y ago
      It’s not evidence, sure, but what you describe, the “feel of it” is most of what we have to work with at the moment, unfortunately. But it clearly is not self-evident that intelligence is or isn’t prediction. Nothing is self-evident in a field we know so very little about. That’s my main objection, we should be humble and curious, not dogmatic.
      - orwin 4y ago
        It's pretty much self-evident when you worked with 3-6 that this isn't how it work. The idea come before the vocalization. Or even when you talk in a foreign language you're semi-competent with.
stochim 4y ago
This. I think the biggest distinction between biological neural nets that seem to be self aware and those that are not is simply metacognition. I suspect the actual stimuli responses (decision making) is virtually the same but self awareness adds a layer after the fact of reasoning about the decision. IF I'm right, one could reasonably argue that metacognition is simply a tool to help adjust the stimuli response for future decisions. One that has become complex enough that it rose to the level to give the sense of self awareness. A sort of back propagation tool to put it in ML terms. Further, if a system isn't adjusting weights and biases as it goes, and metacognition/self awareness is just a mechanism to do so, then we're not so different from these models aside from them being orders of magnitude simpler. We like to think the way we think is somehow magic and discredit these systems as a way of keeping our thinking elevated. Maybe our brains are not that special just far more advanced than our current technological capabilities. Maybe I'm way off base, but like you said, it's dangerous to just assume these systems are missing some magic and therefore just bs generators (as if people aren't pretty consistent bs generators) tl;dr: Agreed. Maybe all biological neural networks are untrustworthy bs generators too. A pile of bs built on previous bs all the way down to the beginning of life.
Barrin92 4y ago
>What if it’s all just a “computational guessing game” all the way down then we would be categorically unable to meaningfully work on purely analytical problems that have nothing to do with induction, which is how these learning systems work. Human beings can invent mathematical proofs and build machines that do symbolic computation, ChatGPT can't. At best it can try to copy a solution that humans have written down somewhere or mangle it in the process.
- tambourine_man 4y ago
  > then we would be categorically unable to meaningfully work on purely analytical problems How so? Sorry, but to categorically affirm such bold claim you need more substance. This is a hard problem, perhaps the hard problem, and we’ve been banging our heads for millennia with very little progress. Now we create a machine that pretty much passes the Turing test. It’s certainly feels more human than some telemarketing call I’ve received. And we’re not going to investigate it further but rather dismiss it as a cool trick? For all we know, that’s all intelligence may be.
  - Barrin92 4y ago
    >How so? Sorry, but to categorically affirm such bold claim you need more substance. It's not a bold claim, it's a trivial claim. No existing learning system can reason. If it could we could hand it axioms and it could do the same deductive reasoning that you can. By definition and how these systems work, all they do is generalize from particular examples, and you can't do maths by empirically looking for numbers in a set of data, because there's an infinite amount of them. That's why ChatGPT spits out nonsense if you give it two large numbers two multiply, but your TI-83 does it perfectly on a battery. We can program deterministic, deductive methods into machines, we have no machine that develops them. Now is there some architecture that can at some point learn how to do more than statistical prediction? Sure but this isn't it. Passing the Turing test and making people feel excited has little to do with what actual intelligence a system has. The Mechanical Turk in the 18th century fooled people, but it wasn't a intelligent machine, it was an elaborate contraption that mimics human behavior. And people understandably conflate the latter with the former.
c3534l 4y ago
I'm very much just predicting the next word when I talk. I don't know the grammar of the sentence and everything I will say when I start the sentence. I just start talking and try to hit the right notes and, if its a topic I've covered before, have some vague plan for the whole thing. Maybe other people's experiences are different, but when it comes to my natural language only and not my ability for reasoning or abstract thought - I am absolutely not doing anything significantly more complicated than chatGPT is doing.
snowwrestler 4y ago
A human mind probably is essentially a prediction engine, but it is predicting everything one can perceive, not just language. That's how a human can throw and catch a ball, or walk over uneven ground, or know that Mom will be mad about the mess. Our intelligence is trained by hard reality, and so our use of language fits into a much broader context. "Touching a hot stove" is a metaphor because we know viscerally what would happen if we actually did it. The limitation of ChatGPT is not that it predicts, the limitation is the incredible paucity of data from which it tries to predict. Language encodes some structure of the world because that's where it was created. But to the extent that can be inferred by ChatGPT, it's a weak and unreliable echo of the original reality. ChatGPT exists in a featureless, timeless void with only patterns of characters in its memory. It can't tell real from fake (from our perspective) because ALL of its entire reality is just those patterns of text. The words are all real, and all there is. This doesn't even get into why human minds make predictions. ChatGPT makes predictions because we built it to, and we prompt it to. Why do our brains? The trite answer is survival of the fittest, but that just raises the question of why we want to survive. The point is, there is a fundamental difference between a living intelligence, and a machine learning tool. ChatGPT will sit quietly until prompted, and then answer only what it is asked. A human won't. I'm not sure that comes down to intelligence, or something else. Even bugs don't sit still and do what they're asked.
peterth3 4y ago
How well GPTZero can detect ChatGPT generated text by measuring perplexity (how random the words choices are) and burstiness (how diverse the sentence structure is) shows that whatever algorithm our brain uses has stronger creative capabilities than this LLM. GPT-3.5 isn’t a great writer like AlphaGo is a great go player. Maybe one day AI will generate better scripts and novels than humans, but not this model. Medium-quality writing is ok for informative content though, but it’s problematic when the model doesn’t know fact from fiction. That’s the important complaint. Is it dangerous? Maybe. But is it useful? Not if it’s wrong too often. You’re right that this tech should be taken seriously, but so should the hallucination problems. These problems can be solved. And maybe they should be solved before anyone trusts it with serious questions.
- tasuki 4y ago
  > it’s problematic when the model doesn’t know fact from fiction. This is in no way unique to an AI. Have you ever interacted with humans? Half the population thinks the other half can't tell fact from fiction. The other half thinks the same about the first half. We're all wrong about all the time.
  - danaris 4y ago
    There's a fundamental difference between "some people are wrong some of the time"—or even "half the population has trouble telling fact from fiction some of the time", if we grant that as true for the sake of argument—and "ChatGPT (and similar ML algorithms) don't even have a metric to determine truth from fiction; they just predict what's the most likely set of words to stick together in response to your prompt." ChatGPT fundamentally cannot ever know when it's wrong. I should hope it goes without saying that that's not true of humans.
- tambourine_man 4y ago
  From what I’ve read, LLM can’t reliably detect itself.
tasuki 4y ago
You are just a bullshit generator - and so am I!
YeGoblynQueenne 4y ago
>> How do we know we’re not doing the same or similar in our brains? Because if I ask you to calculate 85 + 73 in your head, you will not start searching for the string that completes "85 + 73 = " with maximum probability, but instead try to add those two numbers using an algorithm you learned at school. Which means you can compute stuff, just like a computer, and not only guess stuff, just like an LLM.
- tambourine_man 4y ago
  You’re just repeating that computing is not predicting without explaining why. Also, LLMs are not grep. If the string doesn’t match, it will come up with something it deems plausible, like us. It’s even bad at math in eerily similar ways to us, like off-by-a-decimal-place errors.
  - YeGoblynQueenne 4y ago
    >> You’re just repeating that computing is not predicting without explaining why. Yes, because I assume you're a programmer, or a computer scientist, and you already understand the difference. To clarify, the difference is that a program that computes the most likely next token in a sequence, like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs; and an LLM cannot calculate every such program, nor can it perform the same calculations as those programs, it can only calculate the most likely next token in a sequence. Is the distinction clearer now? Also, are you a programmer? If so, I would point to the wikipedia article on Turing Machines or the one on the theory of computation. It's perhaps not a very easy subject to explain in a HN comment, but if you have some background in programming computers you should be able to grok it. You will see for example that a Turing machine is a device that manipulates symbols to produce other symbols, which is what we generally mean by "computation" and not a device that predicts the next token in a sequence, as an LLM does. Sorry if I assume too much about your background - this is HN after all.
    - tambourine_man 4y ago
      No worries, you were right on your assumption, I am a programmer, like you seem to be as well. > like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs I don’t know if it’s been formally verified, but it’s pretty safe to bet that ChatGPT is Turing Complete. If so, than your statement is false. ChatGPT can emulate every computable problem.
Planktonne 4y ago
When I don't know how a stage magician's trick works, I don't immediately default to assuming that they must have occult powers.

qotgalaxy 4y ago

> How do we know we’re not doing the same or similar in our brains? How do we know that we're not made of magic?

emodendroket 4y ago

That would seem to be the “stochastic parrot” view denounced. But it seems hard to believe. Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? How could that be if we’re just randomly selecting words like a souped-up Markov chain?

tambourine_man 4y ago
We’re not randomly selecting but probabilistically predicting. Haven’t you noticed you predict the next word people will say with very high [________]?
- peteradio 4y ago
  Sandwiches!
- emodendroket 4y ago
  Within a sentence sure. And often arguments are familiar enough that even an essay I’ll know where it’s going. But we do occasionally have original ideas.
  - tambourine_man 4y ago
    I don’t think we know what original really is to make a meaningful distinction between derivative work. With enough mixing and matching and a tiny bit of randomness, I bet a lot of original thoughts could be created.
    - emodendroket 4y ago
      These things are hard to quantify but ChatGPT writing does have a certain anodyne character, in my mind, when I read it. Then again, I can't quantify it, I could be biased by my foreknowledge of the writing's provenance, and even if I'm right, the world is full of situations where bland writing is suitable or even desired.
      - tambourine_man 4y ago
        It absolutely does. It’s wishy-washy nature is harder to disagree with, sounds more credible, etc.
Kamq 4y ago
> Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? How could that be if we’re just randomly selecting words like a souped-up Markov chain? All words in the normal cache have a low probability, so you have to hit disk to get the full word list.
waselighis 4y ago
> Haven’t you ever had the experience of knowing exactly what you want to say but struggling to find the word for it? I came here to suggest this exact thing. Humans clearly have some understanding of the meaning behind words and sentences. However, I think it's wrong, an overgeneralization, to suggest that ChatGPT is just statistically predicting the next word. While that may technically be true, I think buried deep within it has ways of modeling/encoding common concepts and ideas, maybe similar to how a human models concepts and ideas in their mind. Then there's the whole problem of consciousness, but I won't get into that here.
- tambourine_man 4y ago
  >I think buried deep within it has ways of modeling/encoding common concepts and ideas, maybe similar to how a human models concepts and ideas in their mind Right, maybe meaning and concepts are just “things that usually go together” with different levels of distance. We’re comparing our minds, which we know very little of, to LLM, which are blackboxes. It seems like a lot of certainty in face of such vast ignorance.

arbuge 4y ago

Indeed, one might conclude that predicting what to do in the next moment (analogous to what to write next in the case of ChatGPT, which can only write) is all we do as humans. At every moment of the day, we are taking all our past experiences and the present as inputs, and predicting what we should do next based on that.

zyklonix 4y ago

This is in fact the whole premise of Jeff Hawkins' book "On intelligence". Prediction is at the core of how we learn/act. Since we are born, we are just predicting everything that will happen and our brain learns when we see/hear/sense something that we can't predict.

muzani 4y ago

I love the philosophy behind Gustafson's Law: the problem size increases in response to computational resources. You get a simple computer, you can drop bombs more accurately. Then you make word processors. Then text based simulation aka games. Then turn books into digital format. Then whole online libraries, ecommerce, social media, cryptocurrency, and so on. Seeing a LLM as purely language based might be missing the forest for the atoms.

akomtu 4y ago

A token generator corresponds to reflexes: a fuzzy hash table of situations and responses. Cats and dogs have that. Assigning meanings to tokens and establishing relationships between them is where intelligence begins. GPT continues sequence CGGA with T because there is a similar sequence in its dataset, but we interpret T as a tree and see how this meaning relates to A. I believe GPT will solve this problem to some extent in 10 years and will finally earn its real AI badge. Above it there is abstract mind that only top scientists use to some extent.

root_axis 4y ago

I won't speak for what's "all the way down", but on whatever level humans perform reflective reasoning, it's self-evident that we don't simply "predict the next word" in the way a LLM does. Whether or not a LLM is capable of "creating anything new" is a different argument altogether. "New" is in eye of the beholder and it is often said that nothing is new under the sun, but we already know that unthinking processes within nature are capable of producing beauty far and beyond anything that even the greatest human artist could ever hope to accomplish. That a large LLM is unthinking doesn't preclude it from producing art.

tambourine_man 4y ago
Please provide evidence that intelligent speech isn’t simply predicting the next word, as it’s obviously not self-evident to me.
- lukev 4y ago
  I'm not sure it counts as "evidence" because I can't show it to you, but based on personal introspection, whenever I write something it starts with a barely-linguistic thought I want to communicate that gradually gets expanded into as much prose as is necessary to get across the concept whilst establishing context and defending from the most probable detractors. I literally never just start writing and then just keep adding words.
  - tambourine_man 4y ago
    It’s not evidence, sure, but what you describe, the “feel of it” is most of what we have to work with at the moment, unfortunately. But it clearly is not self-evident that intelligence is or isn’t prediction. Nothing is self-evident in a field we know so very little about. That’s my main objection, we should be humble and curious, not dogmatic.
    - orwin 4y ago
      It's pretty much self-evident when you worked with 3-6 that this isn't how it work. The idea come before the vocalization. Or even when you talk in a foreign language you're semi-competent with.

stochim 4y ago

This. I think the biggest distinction between biological neural nets that seem to be self aware and those that are not is simply metacognition. I suspect the actual stimuli responses (decision making) is virtually the same but self awareness adds a layer after the fact of reasoning about the decision. IF I'm right, one could reasonably argue that metacognition is simply a tool to help adjust the stimuli response for future decisions. One that has become complex enough that it rose to the level to give the sense of self awareness. A sort of back propagation tool to put it in ML terms. Further, if a system isn't adjusting weights and biases as it goes, and metacognition/self awareness is just a mechanism to do so, then we're not so different from these models aside from them being orders of magnitude simpler. We like to think the way we think is somehow magic and discredit these systems as a way of keeping our thinking elevated. Maybe our brains are not that special just far more advanced than our current technological capabilities. Maybe I'm way off base, but like you said, it's dangerous to just assume these systems are missing some magic and therefore just bs generators (as if people aren't pretty consistent bs generators) tl;dr: Agreed. Maybe all biological neural networks are untrustworthy bs generators too. A pile of bs built on previous bs all the way down to the beginning of life.

Barrin92 4y ago

>What if it’s all just a “computational guessing game” all the way down then we would be categorically unable to meaningfully work on purely analytical problems that have nothing to do with induction, which is how these learning systems work. Human beings can invent mathematical proofs and build machines that do symbolic computation, ChatGPT can't. At best it can try to copy a solution that humans have written down somewhere or mangle it in the process.

tambourine_man 4y ago
> then we would be categorically unable to meaningfully work on purely analytical problems How so? Sorry, but to categorically affirm such bold claim you need more substance. This is a hard problem, perhaps the hard problem, and we’ve been banging our heads for millennia with very little progress. Now we create a machine that pretty much passes the Turing test. It’s certainly feels more human than some telemarketing call I’ve received. And we’re not going to investigate it further but rather dismiss it as a cool trick? For all we know, that’s all intelligence may be.
- Barrin92 4y ago
  >How so? Sorry, but to categorically affirm such bold claim you need more substance. It's not a bold claim, it's a trivial claim. No existing learning system can reason. If it could we could hand it axioms and it could do the same deductive reasoning that you can. By definition and how these systems work, all they do is generalize from particular examples, and you can't do maths by empirically looking for numbers in a set of data, because there's an infinite amount of them. That's why ChatGPT spits out nonsense if you give it two large numbers two multiply, but your TI-83 does it perfectly on a battery. We can program deterministic, deductive methods into machines, we have no machine that develops them. Now is there some architecture that can at some point learn how to do more than statistical prediction? Sure but this isn't it. Passing the Turing test and making people feel excited has little to do with what actual intelligence a system has. The Mechanical Turk in the 18th century fooled people, but it wasn't a intelligent machine, it was an elaborate contraption that mimics human behavior. And people understandably conflate the latter with the former.

c3534l 4y ago

I'm very much just predicting the next word when I talk. I don't know the grammar of the sentence and everything I will say when I start the sentence. I just start talking and try to hit the right notes and, if its a topic I've covered before, have some vague plan for the whole thing. Maybe other people's experiences are different, but when it comes to my natural language only and not my ability for reasoning or abstract thought - I am absolutely not doing anything significantly more complicated than chatGPT is doing.

snowwrestler 4y ago

A human mind probably is essentially a prediction engine, but it is predicting everything one can perceive, not just language. That's how a human can throw and catch a ball, or walk over uneven ground, or know that Mom will be mad about the mess. Our intelligence is trained by hard reality, and so our use of language fits into a much broader context. "Touching a hot stove" is a metaphor because we know viscerally what would happen if we actually did it. The limitation of ChatGPT is not that it predicts, the limitation is the incredible paucity of data from which it tries to predict. Language encodes some structure of the world because that's where it was created. But to the extent that can be inferred by ChatGPT, it's a weak and unreliable echo of the original reality. ChatGPT exists in a featureless, timeless void with only patterns of characters in its memory. It can't tell real from fake (from our perspective) because ALL of its entire reality is just those patterns of text. The words are all real, and all there is. This doesn't even get into why human minds make predictions. ChatGPT makes predictions because we built it to, and we prompt it to. Why do our brains? The trite answer is survival of the fittest, but that just raises the question of why we want to survive. The point is, there is a fundamental difference between a living intelligence, and a machine learning tool. ChatGPT will sit quietly until prompted, and then answer only what it is asked. A human won't. I'm not sure that comes down to intelligence, or something else. Even bugs don't sit still and do what they're asked.

peterth3 4y ago

How well GPTZero can detect ChatGPT generated text by measuring perplexity (how random the words choices are) and burstiness (how diverse the sentence structure is) shows that whatever algorithm our brain uses has stronger creative capabilities than this LLM. GPT-3.5 isn’t a great writer like AlphaGo is a great go player. Maybe one day AI will generate better scripts and novels than humans, but not this model. Medium-quality writing is ok for informative content though, but it’s problematic when the model doesn’t know fact from fiction. That’s the important complaint. Is it dangerous? Maybe. But is it useful? Not if it’s wrong too often. You’re right that this tech should be taken seriously, but so should the hallucination problems. These problems can be solved. And maybe they should be solved before anyone trusts it with serious questions.

tasuki 4y ago
> it’s problematic when the model doesn’t know fact from fiction. This is in no way unique to an AI. Have you ever interacted with humans? Half the population thinks the other half can't tell fact from fiction. The other half thinks the same about the first half. We're all wrong about all the time.
- danaris 4y ago
  There's a fundamental difference between "some people are wrong some of the time"—or even "half the population has trouble telling fact from fiction some of the time", if we grant that as true for the sake of argument—and "ChatGPT (and similar ML algorithms) don't even have a metric to determine truth from fiction; they just predict what's the most likely set of words to stick together in response to your prompt." ChatGPT fundamentally cannot ever know when it's wrong. I should hope it goes without saying that that's not true of humans.
tambourine_man 4y ago
From what I’ve read, LLM can’t reliably detect itself.

tasuki 4y ago

You are just a bullshit generator - and so am I!

YeGoblynQueenne 4y ago

>> How do we know we’re not doing the same or similar in our brains? Because if I ask you to calculate 85 + 73 in your head, you will not start searching for the string that completes "85 + 73 = " with maximum probability, but instead try to add those two numbers using an algorithm you learned at school. Which means you can compute stuff, just like a computer, and not only guess stuff, just like an LLM.

tambourine_man 4y ago
You’re just repeating that computing is not predicting without explaining why. Also, LLMs are not grep. If the string doesn’t match, it will come up with something it deems plausible, like us. It’s even bad at math in eerily similar ways to us, like off-by-a-decimal-place errors.
- YeGoblynQueenne 4y ago
  >> You’re just repeating that computing is not predicting without explaining why. Yes, because I assume you're a programmer, or a computer scientist, and you already understand the difference. To clarify, the difference is that a program that computes the most likely next token in a sequence, like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs; and an LLM cannot calculate every such program, nor can it perform the same calculations as those programs, it can only calculate the most likely next token in a sequence. Is the distinction clearer now? Also, are you a programmer? If so, I would point to the wikipedia article on Turing Machines or the one on the theory of computation. It's perhaps not a very easy subject to explain in a HN comment, but if you have some background in programming computers you should be able to grok it. You will see for example that a Turing machine is a device that manipulates symbols to produce other symbols, which is what we generally mean by "computation" and not a device that predicts the next token in a sequence, as an LLM does. Sorry if I assume too much about your background - this is HN after all.
  - tambourine_man 4y ago
    No worries, you were right on your assumption, I am a programmer, like you seem to be as well. > like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs I don’t know if it’s been formally verified, but it’s pretty safe to bet that ChatGPT is Turing Complete. If so, than your statement is false. ChatGPT can emulate every computable problem.

Planktonne 4y ago

When I don't know how a stage magician's trick works, I don't immediately default to assuming that they must have occult powers.