If you compare "thinking" to "text generators", don't you need to understand what "thinking" actually is, before you can conclude that these are unequal?
For a lot of us, LLM behavior and thinking are so plainly and obviously dissimilar that the endless comparisons look somewhere between naive or manipulative, depending on who’s making them.
But what's obvious to you is not obvious to everyone.
Even though you don't have to mathematically prove a goldfish doesn't walk for me to believe you, that's because we can both agree on a very good definition of walking. If it came down to it, I am sure we could sit down with pen an paper and some textbooks and agree upon a physical definition as robust as any. We're just skipping that step because it's been done before by others.
I am positive that we can't agree on a robust definition of thinking, because the definition of thinking is beyond human understanding.
We have to keep an open mind. I don't believe it's likely that any LLMs are experiencing anything we'd call thought, but without knowing how thought works, it would be foolish of me to say it's impossible. The problem in AI discussions is not the positions people are taking but the certainty with which they are taking them.
In that case, shouldn't the people who say LLMs can think do the hard work of explaining what "thinking" means and why they er think it's done by LLMs?
Surely the default position should be that LLMs don't think because they don't belong to the class of entities that we know can think. If that default assumption is wrong, well, then, someone has to do the hard work of rejecting it. But just claiming that we think they think because who knows what thinking is, is just an excuse to not do the work.
We can be sure humans think. Does a crow think? We don't know. Does an LLM think? We don't know.
We can talk about more specific phenomena in contexts that demand it, if we have the data. But there is no way to say right now that an LLM does or does not think. All we can say is it seems unlikely.
Constantly moving the goal posts with this incessant "but what IS intelligence, though?!?" gets us nowhere. By your own argument we can never define intelligence, so we can never actually discuss it. LLMs consistently and constantly fall down on tasks that we would not expect a human to fail at, but you and people like you insist we cannot use this as evidence because you refuse to accept any defined terms.
I mean, remain as hopeful and uncritical as you want. But you're basically asking the rest of us "Who are you gonna believe, me or your lyin' eyes?" and that's gonna go about as well for you as it typically does.
I'm not talking about getting scammed by SBF or whatever. I'm saying that there doesn't exist a definition of "thinking" that we can use to include or exclude LLMs from the group "beings that think". It's counterproductive to say they do or don't and is a distraction from useful conversation.
To say I'm not being critical when I clearly said I don't think it's likely that LLMs think feels disingenuous and, frankly, uncritical.
> By your own argument we can never define intelligence
In no way does that extend from my argument. Intelligence and thinking are separate phenomena, and I didn't say "never".
If you turn off the power to the computer hosting the LLM, you can also be sure it doesn't think.
There's no reason to presume we all think the same way; rather the opposite, given how many ways in which humans are already known to think unalike to each other.
"The mode of action of paracetamol has been uncertain, but it is now generally accepted that it inhibits COX-1 and COX-2 through metabolism by the peroxidase function of these isoenzymes."
Furthermore, the active ingredient in Tylenol is a very simple molecule from a synthetic perspective, and it was developed while trying to overcome the toxicity of an even simpler molecule called acetanilide. Even though it was developed 150 years ago, there was a systematic understanding of organic chemistry and the interaction with the human body already at that time.
[1] https://link.springer.com/article/10.1007/s10787-013-0172-x
No? Maybe the analogy you are suggesting isn’t the most accurate.
An easy way to prove the two are different is 'What is the equivalent to the endocrine system for an LLM and why doesn't it get tired like a mind?'
The logical problems that a lot of people run into are twofold:
1) At a very basic level, we modeled these architectures on brains (and named them after our brains), so at a very very basic metaphorical level, you can say 'these things are acting like our brains act.' This ignores the complexity of the brain.
2) We like to think that inconveniences of our minds (like the need for sleep) are "problems" that we can engineer away, rather than intrinsic components of the process. People won't like hearing this, but there's no formal reason why 'the need to sleep' is not a necessary criteria for 'being conscious,' because as far as we know, the only conscious beings out there also sleep. It's just human nature to assume we can engineer the 'good' parts of a biological system while avoiding the 'bad' while still essentially replicating the system, but that's usually not the case.
Are you suggesting that the brain is the only implementation that can “think”? Or that an endocrine system is required?
The difference isn’t the interesting, or necessarily useful, part; the practical similarities of the output are. If we can make a system that appears to "think", I can’t understand how it can be so easily dismissed when we don’t know what’s going on in it or us.
What else does (please do not mention any deterministic counting machines like semiconductors - neurons are not deterministic and thought isn’t a deterministic set of calculations).
I would argue that the difference in process is extremely important, as evidenced by how easily we are fooled by optical illusions.
Just because we think two things appear the same does not mean they are, and understanding why they appear the same but are different relies on a study of the “how”
To put your argument differently - “if we can make drawings that appear to be three dimensional to our eyes, why bother understanding the difference between 2-D and 3-D space?”
I have to mention it, because it's physics, and related to the current implementation of LLMs: random numbers are possible, and used to break determinism. Intel CPUs use thermal noise to generate random numbers [1]. With silicon, randomness is a free choice, not an impossibility. LLMs front ends, like anything from OpenAI, use random numbers to get non-deterministic output, which is also the input of the next word, and the context of the next response, resulting in output that's not deterministic, with broad divergence [2]. Both systems are somewhat bound by the "sensibility"/logic of the output though, of course.
> as evidenced by how easily we are fooled by optical illusions
This isn't neccesarily unique to humans [3]. Do you have a specific illusion in mind? Many are related to active "baselining", and other time response things, that happens in our eyes, with others being an incorrect resolution of real ambiguity, from a two sensor system, that any sensor system will struggle with.
> Just because we think two things appear the same does not mean they are
I don't think anyone is suggesting they're the same, but I see many people suggesting, with seemingly undue confidence, that they're completely unrelated, which would require an understanding of either system that we don't have.
> To put your argument differently - ...
My argument is, if it's 3d to your eyes, then that understanding, and relation, between 2d and 3d space already exists in the system, to some degree.
[1] https://www.intel.com/content/www/us/en/developer/articles/g...
[2] https://www.coltsteele.com/tips/understanding-openai-s-tempe...
[3] https://blog.frontiersin.org/2018/04/26/artificial-intellige...
A dot product run on Intel CPUs are intended to always be the same no matter what. The heat signature stuff isn’t changing the way circuits do math.
As to optical illusions, the point I was making was that “the way humans practically perceive something” is not a sufficient way to measure the similarness of two things, since our perceptions are so often tricked (LLMs are literally trained specifically to trick you into thinking you’re talking to sentience).
I also don’t think they are completely unrelated at all. As I said I know we designed one to be like the other. It’s just by no means “the same.” AI may one day do convergent evolution toward how brains think, but it’s still important to recognize that’s convergent evolution.
> It’s just by no means “the same.”
The implementation is not the same. Everyone agrees with that. The concept being compared is "thought", not "implementation of thought". Maybe I'm lost.
I would argue expanding the definition to include the kind of things that semiconductors do really dilutes the meaning of "thinking" to be near meaningless.
Maybe to put it succinctly - no computer has ever done anything without a human input (even if it's millions of layers abstracted), but thinking just happens spontaneously.
If that's not a sufficient condition to differentiate 'thinking' from 'calculating,' then IDK what 'thinking' even means then.
It’s a common experience that candy tasted sweeter to you as a child, right?
How does that work in a transformer?
These threads are always massive, embarrassing trainwrecks of naive takes on cognition and consciousness.
It's like HN just woke up and refuses to acknowledge the relevance of the extensive prior literature on these subjects. Instead, people seem mostly fine with making shit up.
Embarrassing.
It's fine. We're social animals, we discuss topics even if we lack perfect expertise, at our present level of understanding. I've certainly had, for example, many opinions on programming that were later superseded by deeper understanding or evolving priorities.
It's just important to remember there's probably a bigger and better expert out there, and stay humble and interested, and let yourself be corrected.
Among the other technical and scientific topics often discussed here, there is a basic general expectation of grounded reference to existing research and bodies of knowledge. Deviations from this standard are usually met with corrections.
With the topic of cognition and consciousness, this is not so (generally). The standard is to absolutely not reference existing research and to basically pretend it doesn't exist.
Instead, the vacuum of ignorance is filled with whatever comes to mind and threads engage in freeform speculation.
This is also ironic since these topics are usually about how LLMs don't understand and just fill the vacuum of its ignorance with whatever happens to be next on the statistical chain, regardless of veracity.
With physics, for example, people have an understanding that there's a large existing body of work, and that physics has been remarkably successful at building on earlier knowledge for decades without having to debunk a lot of earlier results. It's a familiar story.
With AI, however, the popular narrative is that earlier approaches to the problem were all laughing stocks that went nowhere. There's not a good sense for how old some the ideas that work now actually are, and how many of the older theoretical and thought frameworks are still perfectly valid, in particular at the interfaces toward other branches of science and engineering. There's some more nuanced and balanced primers on the topic (e.g. the Wooldridge book), but an awful lot more writing on AI Winters and what not.
This snowballs into a lack of awareness of the overlap between "AI" and, say, the body of knowledge in statistics.
In other words: Sure, I agree - HN collectively is less knowledge on this than on some other topics (myself included!), and it's worth taking stock of that, I suppose.
The onus [1] should go the other way.
____________
[1] Today's conversation-derailing trivium: "onus" means ass in Greek. As in donkey.
A colleague at my first job was working on some software that wrapped around other software, and he dubbed it "Burrito". When his project turned out not to be as useful as expected, he was pleasantly surprised that he could also use the original meaning of the word burrito, being "little donkey" in Spanish [1].
These silly generalisations about HN are pointless. HN has a pretty diverse set of worldviews, approaches, and levels of expertise across a range of topics. There is zero point in lumping people into a single, or even a majority, behaviour.
The idea that generalities can't be true or useful is pretty insidious, since it prevents reasoning about groups.
This idea was not mentioned, whether or not you attest that it was.
This means we have to be lucky to have some expert with current knowledge join in. Given that most people here are billionaires with a passion for dynamic typing, and not philosophers of mind, I am actually fairly surprised at the reasonable level of discussion.
Also, I doubt that referencing prior literature in the context of cognition and consciousness would be of much help. HN writes a lot of nonsense about it, but so have many professional philosophers. After sifting through all of that, one still has to add a lot of convincing arguments to advance an unpopular opinion.
A friend once suggested to build forum software that would allow discussions to reach a certain level of trustworthiness or truth. After several lengthy discussions I still doubt that such a thing is feasible. But it might spark your interest? :)
What does it mean to understand something? To think? How is a human any different than predicting the next “thing” given a series of inputs?