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"!)
(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"!)
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.
"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?
uh.. no we can't?
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.
That looks like thinking. If you want to argue it's not thinking you are absolutely welcome to, but you need to say more than just "no we can't?".
At some point the cry will change to, "It only looks like thinking from the outside!" And some time after that: "It only looks conscious from the outside..."
https://arxiv.org/abs/2505.13775
Just because these models are good at outputting text that appears to be a series of thoughts, doesn’t mean that LLMs are thinking.
'thinking' here needs a clearer definition. Even for us humans the language is just an after the fact attempt at describing what went on, or is going on inside our heads
Consider: https://www.unsw.edu.au/newsroom/news/2019/03/our-brains-rev...
Before it makes a move, a little LED labeled ‘thinking’ blinks for a while.
Looks like thinking, too.
Again, why should we believe that LLMs are capable of thought, other than the fact that they vaguely “seem” to communicate with conscious intent?
I believe this to be the case because the appearance of thoughtfulness is a smooth function, not a line in the sand. Why are you bringing consciousness into this discussion? Consciousness isn't thoughtfulness.
> Again, why should we believe that LLMs are capable of thought, other than the fact that they vaguely “seem” to communicate with conscious intent?
I define "to think" as the act of applying context to a model. What do you define it as? Given my definition, llm inference seems plainly to be an act of thought.
Answer my question, please. edit: let me rephrase to be more clear with what i'm asking. Why are you assuming that consuming the entire internet would lead to either "AGI is here" or "AGI will never be here"? New techniques for developing better intelligence are emerging all the time.
To quote N. Chomsky when he made an analogy: "When you have a theory, there are two questions you'd need to ask: (1) Why are things this way? (2) Why are things not that way?"[1]
I get it, this post-truth century is difficult to navigate. Engineering achievements get picked up through a mix of technical excitement and venture boredom, its fallacies are stylized by growth hackers as this unsolvable paradox that seems to fit just the right terminology for any marketing purposes, only to be piped through one hype cycle that just paints another picture of the doomsday of capitalism with black and white colors. /rant
That doesn't mean I think they can think myself. I just get frustrated at the quality of discussion every time this topic comes up.
> That doesn't mean I think they can think myself. I just get frustrated at the quality of discussion every time this topic comes up.
You’re arguing for the potential of a ghost in a machine based on people not being able to answer hard philosophical questions behind human behavior and properties.[1] That’s a kind of Vulgar Theism quality of argumentation.
[1] Not exclusively human. I’m sure killer whales can think.
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.
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.
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.
See also: Levels of AGI for Operationalizing Progress on the Path to AGI (https://arxiv.org/abs/2311.02462)
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.
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