It doesn't think, it doesn't reason, and it doesn't listen to instructions, but it does generate pretty good text!
People constantly assert that LLMs don't think in some magic way that humans do think, when we don't even have any idea how that works.
It doesn't matter anyway. The marquee sign reads "Artificial Intelligence" not "Artificial Human Being". As long as AI displays intelligent behavior, it's "intelligent" in the relevant context. There's no basis for demanding that the mechanism be the same as what humans do.
And of course it should go without saying that Artificial Intelligence exists on a continuum (just like human intelligence as far as that goes) and that we're not "there yet" as far as reaching the extreme high end of the continuum.
Aircraft and submarines belong to a different category and of the same category, than AI.
Humans are not all that original, we take what exists in nature and mangle it in some way to produce a thing.
The same thing will eventually happen with AI - not in our lifetime though.
The proof burden is on AI proponents.
There's this very cliched comment to any AI HN headline which is this:
"LLM's don't REALLY have <vague human behavior we don't really understand>. I know this for sure because I know both how humans work and how gigabytes of LLM weights work."
or its cousin:
"LLMs CAN'T possibly do <vague human behavior we don't really understand> BECAUSE they generate text one character at a time UNLIKE humans who generate text one character a time by typing with their fleshy fingers"
Intelligent living beings have natural, evolutionary inputs as motivation underlying every rational thought. A biological reward system in the brain, a desire to avoid pain, hunger, boredom and sadness, seek to satisfy physiological needs, socialize, self-actualize, etc. These are the fundamental forces that drive us, even if the rational processes are capable of suppressing or delaying them to some degree.
In contrast, machine learning models have a loss function or reward system purely constructed by humans to achieve a specific goal. They have no intrinsic motivations, feelings or goals. They are statistical models that approximate some mathematical function provided by humans.
Even if it can extrapolate to some degree (altough that's where "hallucinations" tend to become obvious), it could never, for example, invent a game like chess or a social construct like a legal system. Those require motivations like "boredom", "being social", having a "need for safety".
> it could never, for example, invent a game like chess or a social construct like a legal system. Those require motivations like "boredom", "being social", having a "need for safety".
That's creativity which is a different question from thinking.
Yes, humans are also capable of learning in a similar fashion and imitating, even extrapolating from a learned function. But I wouldn't call that intelligent, thinking behavior, even if performed by a human.
But no human would ever perform like that, without trying to intuitively understand the motivations of the humans they learned from, and naturally intermingling the performance with their own motivations.
Humans invent new data, humans observe things and create new data. That's where all the stuff the LLMs are trained on came from.
> That's creativity which is a different question from thinking
It's not really though. The process is the same or similar enough don't you think?
LLMs aren't good at either, imo. They are rote regurgitation machines, or at best they mildly remix the data they have in a way that might be useful
They don't actually have any intelligence or skills to be creative or logical though
What LLMs do is using what they have _seen_ to come to a _statistical_ conclusion. Just like a complex statistical weather forecasting model. I have never heard anyone argue that such models would "know" about weather phenomena and reason about the implications to come to a "logical" conclusion.
In the same way a human might produce a range of answers to the same question, so humans are also drawing from a theoretical statistical distribution when you talk to them.
It's just a mathematical way to describe an agent, whether it's an LLM or human.
We don't just study it in humans. We look at it in trees [0], for example. And whilst trees have distributed systems that ingest data from their surroundings, and use that to make choices, it isn't usually considered to be intelligence.
Organizational complexity is one of the requirements for intelligence, and an LLM does not reach that threshold. They have vast amounts of data, but organizationally, they are still simple - thus "ai slop".
[0] https://www.cell.com/trends/plant-science/abstract/S1360-138...
In my opinion AI slop is slop not because AIs are basic but because the prompt is minimal. A human went and put minimal effort into making something with an AI and put it online, producing slop, because the actual informational content is very low.
And you'd be disagreeing with the vast amount of research into AI. [0]
> Moreover, they exhibit a counter-intuitive scaling limit: their reasoning effort increases with problem complexity up to a point, then declines despite having an adequate token budget.
[0] https://machinelearning.apple.com/research/illusion-of-think...
It does say that there is a maximal complexity that LLMs can have - which leads us back to... Intelligence requires organizational complexity that LLMs are not capable of.
That doesn't mean such claims don't need to made as specific as possible. Just saying something like "humans love but machines don't" isn't terribly compelling. I think mathematics is an area where it seems possible to draw a reasonably intuitively clear line. Personally, I've always considered the ability to independently contribute genuinely novel pure mathematical ideas (i.e. to perform significant independent research in pure maths) to be a likely hallmark of true human-like thinking. This is a high bar and one AI has not yet reached, despite the recent successes on the International Mathematical Olympiad [3] and various other recent claims. It isn't a moved goalpost, either - I've been saying the same thing for more than 20 years. I don't have to, and can't, define what "genuinely novel pure mathematical ideas" means, but we have a human system that recognises, verifies and rewards them so I expect us to know them when they are produced.
By the way, your use of "magical" in your earlier comment, is typical of the way that argument is often presented, and I think it's telling. It's very easy to fall into the fallacy of deducing things from one's own lack of imagination. I've certainly fallen into that trap many times before. It's worth honestly considering whether your reasoning is of the form "I can't imagine there being something other than X, therefore there is nothing other than X".
Personally, I think it's likely that to truly "do maths" requires something qualitatively different to a computer. Those who struggle to imagine anything other than a computer being possible often claim that that view is self-evidently wrong and mock such an imagined device as "magical", but that is not a convincing line of argument. The truth is that the physical Church-Turing thesis is a thesis, not a theorem, and a much shakier one than the original Church-Turing thesis. We have no particularly convincing reason to think such a device is impossible, and certainly no hard proof of it.
[1] Individual behaviours of LLMs are "not understood" in the sense that there is typically not some neat story we can tell about how a particular behaviour arises that contains only the truly relevant information. However, on a more fundamental level LLMs are completely understood and always have been, as they are human inventions that we are able to build from scratch.
[2] Anybody who thinks we understand how brains work isn't worth having this debate with until they read a bit about neuroscience and correct their misunderstanding.
[3] The IMO involves problems in extremely well-trodden areas of mathematics. While the problems are carefully chosen to be novel they are problems to be solved in exam conditions, not mathematical research programs. The performance of the Google and OpenAI models on them, while impressive, is not evidence that they are capable of genuinely novel mathematical thought. What I'm looking for is the crank-the-handle-and-important-new-theorems-come-out machine that people have been trying to build since computers were invented. That isn't here yet, and if and when it arrives it really will turn maths on its head.
And here's some more goalpost-shifting. Most humans aren't capable of novel mathematical thought either, but that doesn't mean they can't think.
As for most humans not being mathematicians, it's entirely irrelevant. I gave an example of something that so far LLMs have not shown an ability to do. It's chosen to be something that can be clearly pointed to and for which any change in the status quo should be obvious if/when it happens. Naturally I think that the mechanism humans use to do this is fundamental to other aspects of their behaviour. The fact that only a tiny subset of humans are able to apply it in this particular specialised way changes nothing. I have no idea what you mean by "goalpost-shifting" in this context.
If we knew that, we wouldn't need LLMs; we could just hardcode the same logic that is encoded in those neural nets directly and far more efficiently.
But we don't actually know what the weights do beyond very broad strokes.
we understand on this low level, but LLMs through the training converge to something larger than weights, there is a structure of these weights which emerged and allow to perform functions, and this part we do not understand, we just observe it as a black box, and experimenting on the level: we put this kind of input to black box and receive this kind of output.
Why? Team "Stochastic Parrot" will just move the goalposts again, as they've done many times before.
That doesn't mean much.
No it doesn't, this is an overgeneralization.
When you type the next word you also put a word that fits some requirement. That doesn't mean you're not thinking.
- We have a sense of time (ie, ask an LLM to follow up in 2 minutes)
- We can follow negative instructions ("don't hallucinate, if you don't know the answer, say so")
Until we see an LLM that is capable of this, then they aren't capable of it, period
That's the difference. AI cannot be held responsible for hallucinations that cause harm, therefore it cannot be incentivized to avoid that behavior, therefore it cannot be trusted
Simple as that
The general notion of passage of time (i.e. time arrow) is the only thing that appears to be intrinsic, but it is also intrinsic for LLMs in a sense that there are "earlier" and "later" tokens in its input.
Imagine a process called A, and, as you say, we've no idea how it works.
Imagine, then, a new process, B, comes along. Some people know a lot about how B works, most people don't. But the people selling B, they continuously tell me it works like process A, and even resort to using various cutesy linguistic tricks to make that feel like it's the case.
The people selling B even go so far as to suggest that if we don't accept a future where B takes over, we won't have a job, no matter what our poor A does.
What's the rational thing to do, for a sceptical, scientific mind? Agree with the company, that process B is of course like process A, when we - as you say yourself - don't understand process A in any comprehensive way at all? Or would that be utterly nonsensical?
It's like we're pretending cognition is a solved problem so we can make grand claims about what LLM's aren't really doing.
It turns out we didn't need a specialist technique for each domain, there was a reliable method to architect a model that can learn itself, and we could already use the datasets we had, they didn't need to be generated in surveys or experiments. This might seem like magic to an AI researcher working in the 1990's.
A lot of this is marketing bullshit. AFAIK, even "machine learning" was a term made up by AI researchers when the AI winter hit who wanted to keep getting a piece of that sweet grant money.
And "neural network" is just a straight up rubbish name. All it does is obscure what's actually happening and leads the proles to think it has something to do with neurons.