4 ms·

It doesn't. LLMs don't think.

by cratermoon 1y ago

TheNewsIsHere 1y ago
It’s insane that this needs to continually be said.
- anonzzzies 1y ago
  It is, but half of humanity cannot think either so that doesn't help evaluating AI.
JKCalhoun 1y ago
Put "think" in quotes then. I'm only 30 minutes into the interview and find it fascinating — learning some esoterica about how LLMs work. A lot of counter-intuitive stuff too about how larger models keep concepts more closely together, how we may be an order of magnitude or two away from the size of a human brain…
simonw 1y ago
I'll bite. Define "think". (And, to be fair, I try to always use the terms "reasoning" and "thinking" in scare-quotes when I'm writing about LLMs. But honestly I mostly do that to avoid tedious arguments about how "they're not actually thinking"!)
- cratermoon 1y ago
  I'd love for the AI hypesters to give their definition. They don't because any definition of "thinking" that would apply to their constructions would be laughed out of the room.
  - arghwhat 1y ago
    I'd love to hear your definition of human thinking. No one else knows how that works, so it would most certainly be worth sharing, especially what the discriminator for whether something "thinks" would be. While avoiding being "laughed out of the room", of course.
  - esafak 1y ago
    The ability to extrapolate from facts. I don't think self awareness, emotions, or senses are required. You should give your definition too, so we can debate where computers fall short. See also: Levels of AGI for Operationalizing Progress on the Path to AGI (https://arxiv.org/abs/2311.02462 https://arxiv.org/abs/2311.02462)
  - bigyabai 1y ago
    I'll steelman the other side, even as someone that detests AI and marketing hype. There isn't any harm in calling AI-generated text "thought" in the same way we aren't debasing humans by saying that crows or worms think too. It doesn't confer sophistication, sentience or even intelligence onto the thing that is thinking. It just means that some sort of process is taking an input and creating a distinct output as a product of consideration. There's no rule anywhere that says thought has to be intangible, or that it can't be represented as an algebraic formula. We can peer very easily into your train of thought here. You won't produce evidence to logically justify your stance, you can't take a position of authority on the subject, and basically rely purely on pathos to sell a fear of LLMs. Aristotle would call your rhetoric AI-generated if he lived to see the modern age.
- loudmax 1y ago
  Along the same lines, it's often useful to think of an LLM's "understanding" a question or a context. Sometimes responses from smaller models show a shallower "understanding" of the question than you'd get from a bigger model. Obviously, we shouldn't anthropomorphize too much here, and even the most powerful LLMs don't "understand" or "reason" or "think" the way humans do. But whatever they're doing, it's at least analogous to what we do. These concepts are genuinely useful for making better use of these tools.
  - simonw 1y ago
    Yeah, that's my position as well. I hesitated on using the term "reasoning", but it's honestly a really good shortcut for describing that thing where the recent models can "think step by step" about a problem before providing their final answer.
- emp17344 1y ago
  No one can provide a technical definition of what it means to think. That doesn’t mean LLMs are thinking. This style of argument, which often boils down to “you can’t solidly define the thought process, therefore LLMs can think!” is a fallacy. Why would we expect LLMs to be capable of thought in the first place? The fact that these models have devoured the contents of the entire internet and still aren’t AGI is an indication that LLMs won’t develop a capacity for cognition.
  - nullstyle 1y ago
    > Why would we expect LLMs to be capable of thought in the first place? "We" didn't have the expectation, but upon interacting with LLMs, we can see and feel and hear that thought is taking place, and have realized that maybe the special place we held for human thought isn't so special as we had hoped. > The fact that these models have devoured the contents of the entire internet and still aren’t AGI is an indication that LLMs won’t develop a capacity for cognition. Why do you assume that training on the internet would lead to AGI?
    - AstroBen 1y ago
      > upon interacting with LLMs, we can see and feel and hear that thought is taking place uh.. no we can't?
      - nullstyle 1y ago
        Then you aren't paying attention, IMO. An example for the dense: A friend of mine used his conversations with ChatGPT to get through a tough spot in his marriage. The distinction between what happened in ChatGPT's internal processing and what would have happened in the internal processing of the brain of a family therapist is a meaningless distinction. My friend was able to benefit from the thoughtwork of ChatGPT.
        AstroBen 1y ago
        Calling people dense isn't helping them engage with you
- keybored 1y ago
  3000 BC to now: All vexing philosophical problems on things we do as easily as breathe but can’t seem to conceptualize rigourosly Now: If you can’t come up with a good definition on the spot I don’t see why my graph dump can’t do it as well
- whytaka 1y ago
  Even if one were to deny humans the will to "act", to think - as a verb, as some action being committed by a willing being or an automaton - by my definition, requires some weighing or surveying of possible answers. If I were to ask you a question and you were to blurt out an answer by reflex or by unguided stream of consciousness, I can accuse you of not thinking. The kind of thinking I'm referring to you here is one where you take pause to let go of prejudices and consider alternatives before answering. I'd say that LLMs are at best simulating reflexive streams of consciousness. Even with chain-of-thought, it never pauses to actually "think". But maybe even our own pauses are just internal chains of thought. Look at me be a reflexive stream of consciousness.
sebzim4500 1y ago
Obviously he's using a different definition of think than you are. Your comment is as useless as me correcting an American for using a different spelling of 'colour' than I am used to.
- keybored 1y ago
  Metaphors taking on a life of their own makes programmers feel like they understand linguistics (my programming language is like a language) and cognitive science (look, the advanced chat bot is chatting like a thinking person). That’s not the same as using one wowel too much when writing words.
monkaiju 1y ago
The hype machine just loves humanizing these little toys...
icedchai 1y ago
They don't "think" like a human, but on the other hand, they don't "think" like a calculator, either.

It doesn't. LLMs don't think.

TheNewsIsHere 1y ago

It’s insane that this needs to continually be said.

anonzzzies 1y ago
It is, but half of humanity cannot think either so that doesn't help evaluating AI.

JKCalhoun 1y ago

Put "think" in quotes then. I'm only 30 minutes into the interview and find it fascinating — learning some esoterica about how LLMs work. A lot of counter-intuitive stuff too about how larger models keep concepts more closely together, how we may be an order of magnitude or two away from the size of a human brain…

simonw 1y ago

I'll bite. Define "think". (And, to be fair, I try to always use the terms "reasoning" and "thinking" in scare-quotes when I'm writing about LLMs. But honestly I mostly do that to avoid tedious arguments about how "they're not actually thinking"!)

cratermoon 1y ago
I'd love for the AI hypesters to give their definition. They don't because any definition of "thinking" that would apply to their constructions would be laughed out of the room.
- arghwhat 1y ago
  I'd love to hear your definition of human thinking. No one else knows how that works, so it would most certainly be worth sharing, especially what the discriminator for whether something "thinks" would be. While avoiding being "laughed out of the room", of course.
- esafak 1y ago
  The ability to extrapolate from facts. I don't think self awareness, emotions, or senses are required. You should give your definition too, so we can debate where computers fall short. See also: Levels of AGI for Operationalizing Progress on the Path to AGI (https://arxiv.org/abs/2311.02462 https://arxiv.org/abs/2311.02462)
- bigyabai 1y ago
  I'll steelman the other side, even as someone that detests AI and marketing hype. There isn't any harm in calling AI-generated text "thought" in the same way we aren't debasing humans by saying that crows or worms think too. It doesn't confer sophistication, sentience or even intelligence onto the thing that is thinking. It just means that some sort of process is taking an input and creating a distinct output as a product of consideration. There's no rule anywhere that says thought has to be intangible, or that it can't be represented as an algebraic formula. We can peer very easily into your train of thought here. You won't produce evidence to logically justify your stance, you can't take a position of authority on the subject, and basically rely purely on pathos to sell a fear of LLMs. Aristotle would call your rhetoric AI-generated if he lived to see the modern age.
loudmax 1y ago
Along the same lines, it's often useful to think of an LLM's "understanding" a question or a context. Sometimes responses from smaller models show a shallower "understanding" of the question than you'd get from a bigger model. Obviously, we shouldn't anthropomorphize too much here, and even the most powerful LLMs don't "understand" or "reason" or "think" the way humans do. But whatever they're doing, it's at least analogous to what we do. These concepts are genuinely useful for making better use of these tools.
- simonw 1y ago
  Yeah, that's my position as well. I hesitated on using the term "reasoning", but it's honestly a really good shortcut for describing that thing where the recent models can "think step by step" about a problem before providing their final answer.
emp17344 1y ago
No one can provide a technical definition of what it means to think. That doesn’t mean LLMs are thinking. This style of argument, which often boils down to “you can’t solidly define the thought process, therefore LLMs can think!” is a fallacy. Why would we expect LLMs to be capable of thought in the first place? The fact that these models have devoured the contents of the entire internet and still aren’t AGI is an indication that LLMs won’t develop a capacity for cognition.
- nullstyle 1y ago
  > Why would we expect LLMs to be capable of thought in the first place? "We" didn't have the expectation, but upon interacting with LLMs, we can see and feel and hear that thought is taking place, and have realized that maybe the special place we held for human thought isn't so special as we had hoped. > The fact that these models have devoured the contents of the entire internet and still aren’t AGI is an indication that LLMs won’t develop a capacity for cognition. Why do you assume that training on the internet would lead to AGI?
  - AstroBen 1y ago
    > upon interacting with LLMs, we can see and feel and hear that thought is taking place uh.. no we can't?
    - nullstyle 1y ago
      Then you aren't paying attention, IMO. An example for the dense: A friend of mine used his conversations with ChatGPT to get through a tough spot in his marriage. The distinction between what happened in ChatGPT's internal processing and what would have happened in the internal processing of the brain of a family therapist is a meaningless distinction. My friend was able to benefit from the thoughtwork of ChatGPT.
      - AstroBen 1y ago
        Calling people dense isn't helping them engage with you
keybored 1y ago
3000 BC to now: All vexing philosophical problems on things we do as easily as breathe but can’t seem to conceptualize rigourosly Now: If you can’t come up with a good definition on the spot I don’t see why my graph dump can’t do it as well
whytaka 1y ago
Even if one were to deny humans the will to "act", to think - as a verb, as some action being committed by a willing being or an automaton - by my definition, requires some weighing or surveying of possible answers. If I were to ask you a question and you were to blurt out an answer by reflex or by unguided stream of consciousness, I can accuse you of not thinking. The kind of thinking I'm referring to you here is one where you take pause to let go of prejudices and consider alternatives before answering. I'd say that LLMs are at best simulating reflexive streams of consciousness. Even with chain-of-thought, it never pauses to actually "think". But maybe even our own pauses are just internal chains of thought. Look at me be a reflexive stream of consciousness.

sebzim4500 1y ago

Obviously he's using a different definition of think than you are. Your comment is as useless as me correcting an American for using a different spelling of 'colour' than I am used to.

keybored 1y ago
Metaphors taking on a life of their own makes programmers feel like they understand linguistics (my programming language is like a language) and cognitive science (look, the advanced chat bot is chatting like a thinking person). That’s not the same as using one wowel too much when writing words.

monkaiju 1y ago

The hype machine just loves humanizing these little toys...

icedchai 1y ago

They don't "think" like a human, but on the other hand, they don't "think" like a calculator, either.