Comment by seethishat
12 hours ago
LLMs are trained to produce outputs that humans would find reasonable. That's it. They don't do anything else.
To say they have feelings and that they can think/reason is insane.
12 hours ago
LLMs are trained to produce outputs that humans would find reasonable. That's it. They don't do anything else.
To say they have feelings and that they can think/reason is insane.
Reminds me of the Edsger W. Dijkstra quote:
Which has often been paraphrased as
Some might disagree with the relevance but it's a cute analogy.
The whole essay [1] is actually somewhat relevant to modern LLM use for other reasons. He spends a lot of time lamenting the industry's obsession with "user friendly" interfaces which obscure the underlying complexity of the problem at hand.
[1]: https://www.cs.utexas.edu/~EWD/ewd08xx/EWD898.PDF
I think this quote is often misunderstood. It isn't really about machines, but about English definition of swimming (and thinking): a living creature moving in water. That makes debating whether submarines can swim rather pointless. You can replace it by "can airplanes fly?" and the meaning changes because we colloquially say that yes, airplanes fly. In Russian, for example, submarines definitely swim (I believe Dijkstra himself mentioned it at some point). I personally believe that defining thinking as something only humans are capable of is rather narrow.
How are you making this confident of a claim? Do you have insight into how consciousness is a thing, allowing you to definitely say it doesn't exist in this case?
In living beings, consciousness is embodied and can not be reproduced by discrete steps. You can not sit down with pen and paper, and reproduce by hand exactly what is going on inside someone’s brain.
But you could replicate exactly the calculations happening as a LLM computes the next token. Would that be conscious? Where would that consciousness reside?
My view is that LLMs are designed and trained to autocomplete narratives, and that as part of that process they develop an emergent narrative about themselves. Whatever personality traits they exhibit is part of that narrative. Perhaps this is not that different from how humans learn to move in social contexts, but it is not consciousness as it happens in living beings.
Because it is common sense.
This is software trained to produce outputs. It doesn't feel anything.
How is this "common sense"? The question of what consciousness is, what it emerges from, etc. literally have no proved answer at the moment.
Take this statement: "consciousness emerges from sufficient complexity in any system". Neither you or I can prove it right or wrong, but we can both make an assumption that if it was right, then there would be a way for an LLM to be conscious.
I'm not arguing that they are, I'm saying being this confident that they are not is the exact same as being confident that they are.
4 replies →
You could just as much argue that a stone has consciousness. Some people believe that. Insane describes it quite well.
Why is it insane? We have no idea how or where consciousness arises, or even if it's a coherent concept.
10 replies →
All you’re doing is shutting down any inquiry because you’ve decided it’s not worth discussing
So why keep discussing?
You're disparaging a lot of people in your last sentence.
Not a great argument. You're describing what LLMs are trained to do, not necessarily every capability or property produced by that training. We do not know what conditions are sufficient for consciousness. Any statements saying it's not possible for consciousness to emerge (or that it hasn't already) in such a system is no more established than claiming that it has
When John Searle first published his so-called "Chinese Room" argument (https://web-archive.southampton.ac.uk/cogprints.org/7150/1/1...) he included his responses to several counter-arguments that he had been presented with in prior discussions: the systems reply, the robot reply and so forth. Interestingly (as it turns out) no-one seems to have argued that this was a task that could be completed without any understanding, but now that AI technology has reached the point of being able to implement the Chinese Room, it has become the fallback position of those who insist that true AI is impossible!
This is not an argument for true AI being possible, let alone already achieved; it is a warning about depending on arguments from incredulity, which are particularly susceptible to future developments.
The problem with philosophy is that someone can present an argument that is pretty dumb, everyone can point out it's dumb, people can show evidence that it's dumb, reason through how dumb it is, and publish articles about the many different ways in which it's dumb, and still it gets trotted out as if it counts as evidence for some argument. Searle's Room was debunked immediately upon being suggested: the "Systems Reply" is the debunking. His response was: "but if you take all the knowledge out of the system and put it in the person, then the person still doesn't understand Chinese!" Which is obviously false and immediately collapses his entire argument down to nothing. He starts from the premise "this person doesn't understand Chinese; he is just following instructions that are written down for him" and then immediately rejects his own premise and say "actually he does understand all of Chinese and everything he's doing, but he still doesn't, you know, understand it, see?" It's completely embarrassing.
Searle's Rooms is a fully rejected and rebutted argument, and Philosophy will never be (and never has been) a serious discipline until it figures out how to actually let go of failed arguments.
Huh? How does the person having all the instruction books memorized and whatnot imply that the person understands Chinese? Memorizing instructions to the point that one can mentally follow them doesn’t imply understanding why they work. If the person doesn’t understand the correspondence between the Chinese written words and things in the external world, then surely they don’t understand Chinese in the sense we generally mean when we talk about whether a person understands Chinese. (I think this is related to the robot reply?)
That’s not to say that I think the Chinese Room Argument is convincing. I don’t think it is. Or, I guess it might demonstrate something, but not what it aims to show.
I would argue that if, say, the person who has memorized this program and mentally executes it, and if this program also includes actuating some levers to control a mech or something, and the program (running via the person mentally going through the instructions by rote) takes actions to pursue some goals (e.g. purchasing some wood by conversing in Chinese and building a tower with that wood), while the person on whom the program is running wouldn’t know how to order a cheeseburger in Chinese,
Then, the program running in that person’s head might well understand Chinese even though the person doesn’t. (Whether or not the program actually understands Chinese, I think probably depends on the specific of the program in ways that I don’t think anyone knows. I don’t expect that any program that would understand could be small enough that a human person could memorize it and execute it by rote in this way, but this is a thought experiment and such issues can be disregarded I think.)
4 replies →
Searle's main argument seems to be that meat is magic and silicon is not.
The person who memorized the book doesn’t understand Chinese. In fact they don’t even understand that Chinese isn’t a language. Searle’s response stands.
5 replies →
It's becoming increasingly head-in-the-sand to make statements like this. Not that I'm saying LLMs are conscious. But to completely ignore the possibility of emergent properties in systems this complex is kind of simple.
i think there's plenty of evidence of emergent behavior in llms already. This is the driving factor for just more parameters and bigger training sets, "If we just make it bigger maybe more unexpected behaviors will emerge".
idk how legit this is but the summary is my point https://arxiv.org/abs/2206.07682
Not if you believe in a computational/functional theory of mind, and very smart people like Daniel Dennett believed exactly that!
https://plato.stanford.edu/entries/computational-mind/
Sorry, but they can think and reason.
They can have empathy too. Meaning, you can describe a circumstance with many parties involved, with different motivations and conflicting goals -- and it will tell you how these humans would likely feel about each other and the situation at hand.
Being able to describe how humans likely feel is not evidence of the model feeling empathy.
And what would be evidence?
1 reply →
By that logic, a book has empathy.
Not the book, but the author, yes.
2 replies →
They can provide text which contains thoughts, reasoning, and empathy, but is that evidence of them having thoughts, reasoning, or empathy?
To paraphrase the Chinese room argument: Put me in a room with infinite time, paper, pencils, filing cabinets, and give me an English language list of instructions for doing all the neural-net calculations required to run an LLM. Slide your prompt under the door, and, after doing all the calculations, I will slide the next word out, and so I will repeat until the response is done.
In this thought experiment, I'm just following the instructions. So, in this case, what is doing the "thinking" or "reasoning" or "empathy" here?
The inputs going through the model. That act is what is doing the thinking and reasoning and empathizing. Same as when your brain feeds your sensory inputs through its neural network.
Your brain is just a big set of weights/neurons organized into layers that do back and forward propagation, turns out. You are just statistics surrounded by flesh and bones.
A sociopath can also pantomime empathy for their own selfish reasons simply by observing how others behave in similar situations, while feeling nothing special at all.
The author is anthropomorphizing a language-tamagotchi.
LLMs can reason, reasoning is logic, and as you said yourself it is derived from what humans belief is reasonable. However, there is more to the structure of language, the way mathematics is also a language that ultimately describes the universe.