Not OP, but I think the argument here would be not that LLMs "are not smart" but that smart is just the wrong category of thing to describe an LLM as.
A calculator can do very complex sums very quickly, but we don't tend to call it "smart" because we don't think it's operating intelligently to some internal model of the world. I think the "LLMs are AGI" crowd would say that LLMs are, but it's perfectly consistent to think the output of LLMs is consistent/impressive/useful, but still maintain that they aren't "smart" in any meaningful way.
> "we don't think it's operating intelligently to some internal model of the world"
Okay, but you have to actually address why you think LLMs lack an "internal model of the world"
You can train one on 1930s text, and then teach it Python in-context.
They've produced multiple novel mathematical proofs now; Terrance Tao is impressed with them as research assistants.
You can very clearly ask them questions about the world, and they'll produce answers that match what you'd get from a "model" of the world.
What are weights, if not a model of the world? It's got a very skewed perspective, certainly, since it's terminally online and has never touched grass, but it still very clearly has a model of the world.
I'd dare say it's probably a more accurate model than the average person has, too, thanks to having Wikipedia and such baked in.
I should say that quote was referring to a calculator - I wasn't trying to stake a position on LLMs in that comment, more just pointing out that I think its consistent to think they're helpful without thinking they have AGI.
There's obviously a lot more of a case for suggesting LLMs are generally intelligent than a calculator, but for me, I think the key point is that understanding them as "next token generators" is a lot more helpful to explain things like hallucinations and some of the other issues/loops they get into.
For me, if understanding models as "generally intelligent agents operating with an internal model of the world" explained their behaviour better than "next token generators", I'd think calling them "smart" would have some justification[0]. I'm just a person on the internet though, and defining intelligence is pretty rarely clear, even without bringing LLMs into the mix.
I would analogize LLMs to physics simulations in software. Game engines, for example, simulate physics enough to provide a good enough semblance of real-world physics for suspension of disbelief but we would never mistake it for real world physics. Complicated enough simulations, e.g. for weather forecasting, nuclear weapons, or QCD, can provide insights and prove physics theories, but again, experts would never mistake it for real world physics and would be able to explain where the simulation breaks down when trying to predict real world behavior.
Now we have these LLMs that provide some simulation of reasoning merely through prediction of token patterns and that is indeed unexpected and astonishing. However, the AI promoters want to suggest that this simulation of reasoning is human-level reasoning or evolving toward human-level reasoning and this is the same as mistaking game engine physics for real physics. The failure cases (e.g. the walk vs drive to a car wash next door question or the generating an image of a full glass of wine issue), even if patched away, are enough to reveal the token predictor underneath.
Intelligence can be defined as an optimization problem: "find X which maximizes F(X, Y)" where X is the solution, Y is constraints, and F is optimality/fitness criterion. Most other definitions are inane. E.g. "invent an aircraft" can be described as optimization over possible build instructions under given constraints for base materials which optimizes its ability to fly. Absolutely any invention can be formulated as an optimization problem.
It's not like a calculator because LLM can solve very broad classes of problems - you'd struggle to define problems which LLM can't solve (given some fine-tuning, harness, KB, etc).
All this talk about "smartness" isn't even particularly cute...
> It's not like a calculator because LLM can solve very broad classes of problems
I definitely buy this, as least somewhat. Personally I think it'd be a lot more helpful to talk about how "generalisable" a tool is, rather than "general intelligence". LLMs can definitely solve a much broader class of problems than a calculator.
I don't know that "artificial general intelligence" or even "general intelligence" has a very good definition, personally I feel like "solving problems generally" doesn't seem to capture what I mean when I use those kinds of terms. For one, it makes a swiss army knife seem more intelligent than a cat, which personally seems the opposite of what I'd want a good definition of general intelligence to do.
They aren’t smart, they approximate language constructs. They don’t have believes, ideas, etc. but have a few rounds of discussions with any LLMs and you see how they are probabilistic autocompletes based on whatever patterns from rounds of discussions you feed them
At what point does autocomplete stop being "just autocomplete"?
Clearly there's a limit. For example, if an alien autocomplete implementation were to fall out of a wormhole that somehow manages to, say, accurately complete sentences like "S&P 500, <tomorrow's date>:" with tomorrow's actual closing value today, I'd call that something else.
You can call it however you want. The point of using the term autocomplete is to make the underlying technology relatable and remove the mystic from it. In any case, your alien autocomplete wouldn’t be an LLM if it can predict the future
> At what point does autocomplete stop being "just autocomplete"?
> The point of using the term autocomplete is to make the underlying technology relatable and remove the mystic from it.
I think it fails to do that. It's the wrong level of abstraction. Or is it helpful to model an ISA as the individual atoms making up a CPU implementing it?
I use LLMs vastly differently from the actual auto-complete in my phone's messaging app. The comparison doesn't seem very informative. You can't do much with it.
You can always redefine "intelligent" so that humans meet the requirements but AIs don't.
A better model to use is this: LLMs possess a different type of intelligence than us, just like an intelligent alien species from another planet might.
A calculator has a very narrow sort of intelligence. It has near perfect capability in a subset of algebra with finite precision numbers, but that's it.
An old-school expert system has its own kind of intelligence, albeit brittle and limited to the scope of its pre-programmed if-then-else statements.
By extension, an AI chat bot has a type of intelligence too. Not the same as ours, but in many ways superior, just as how a calculator is superior to a human at basic numeric algebra. We make mistakes, the calculator does not. We make grammar and syntax errors all the time, the AI chat bots generally never do. We speak at most half a dozen languages fluently, the chat bots over a hundred. We're experts in at most a couple of fields of study, the chat bots have a very wide but shallow understanding. Etc.
Don't be so narrow minded! Start viewing all machines (and creatures) as having some type of intelligence instead of a boolean "have" or "have not" intelligence.
Would you say that a display and a printer are a perfect painter because they can render images? And a speaker is a very good musician because they can produce sound?
The LLM tasks is to produce a string of words according to an internal model trained on texts written by humans (and now generted by other LLMs). This is not intelligence.
I wouldn’t say it’s a general definition, but the consensus (according to my opinion) is that intelligence is being able to define problems (not just experience them), discern the root cause, and then solve that.
Where it fails is generally the first step. It’s kinda like the old saying “you have to ask the right question”. In all problem solving matters, the definition of problem is the first step. It may not be the hardest (we have problems that are well defined, but unresolved), but not being able to do it is often a clear indication of not being able to do the rest.
> What would convince you that you're wrong?
Maybe when I can have the same interaction as with my fellow humans, where I can describe the issue (which is not the problem) and they can go solve it and provide either a sound plan to make the issue disappear. Issue here refer to unpleasantness or frustrating situation.
Until then, I see them as tools. Often to speed up my writing pace (generic code and generic presentation), or as a weird database where what goes in have a high probability to appear.
> Maybe when I can have the same interaction as with my fellow humans, where I can describe the issue (which is not the problem) and they can go solve it and provide either a sound plan to make the issue disappear.
I don't know what LLMs are you using, but frontier models do this regularly for me in programming.
Without prodding it along and giving it “hints”? And monitoring it like a baby trying their first steps? If yes, please give me the name of the model so I can try it too.
Yes, mostly without those things. I regularly use Claude Opus 4.6/4.7, Gemini 3.1 Pro and GPT-5.4/5.5. For diagnosing and planning, I always use the highest thinking setting, perhaps with the exception of GPT, where xHigh is pretty costly and slow, so I tend to use High unless the problem is really hard. After the plan is done, for implementation I often use cheaper models, like Sonnet 4.6.
> A calculator has a very narrow sort of intelligence.
Have you ever heard anyone refer to a calculator as intelligent?
These companies have a vested interest in making the product appear more human/smart than it is. It's new tech smeared with the same ole marketing matter.
That's the sorcery mentioned in the GP, the issue comes when people believe it to be smart however in reality it is just a next word prediction. Gives the impression it's actually thinking, and this is by design. Personally I think it's dangerous in the sense it gives users a false sense of confidence in the LLM and so a LOT of people will blindly trust it. This isn't a good thing.
I'm curious how you think "word predictor" meaningfully describes an instruct model that has developed novel mathematical proofs that have eluded mathematicians for decades?
edit:
You cannot predict all the actions or words of someone smarter than you. If I could always predict Magnus Carlsen's next chess move, I'd be at least as good at chess as Magnus - and that would have to involve a deep understanding of chess, even if I can't explain my understanding.
I can't predict the next token in a novel mathematical proof unless I've already understood the solution.
I think that's more of a limitation in how people think about word predictors
If you can predict the words a bright person will say about X... Isn't that some truly astounding tool? That could be used in myriad useful ways if one is a little creative with it
Since it's also "alien" it can also detect and explore paths that we simply haven't noticed since their biases aren't quite the same as ours
Magnus Carlsen understands chess, a machine designed to simply predict his next move would not necessarily understand chess. This is essentially the Chinese Room experiment.
So I think "word predictor" makes sense here. A word predictor can be really really cool.
It’s an incoherent argument, or a meaningless semantic distinction.
There is no design of such a machine that does not encode a very deep understanding of the game.
Leela Chess Zero does understand chess. She plays at roughly 2300 strength with a search depth of 1 ply - purely on the strength of her gestalt evaluation of the position. Humans have learned a lot about chess from studying her (and AlphaZero’s) games. General, transferable knowledge she developed herself about - for example - the long term value of early rook pawn advances.
“Understanding” doesn’t imply anything about personhood or self reflection or awareness.
It's not, and unfortunately you cannot just dismiss perhaps the greatest refutation of functionalism as being incoherent, you have to actually address the argument.
Take a person (Fred) with no experience of knowledge of chess. They don't know how the game works, how the pieces move, or any of the rules. They memorise an algorithm, say how Leela does its search and evaluation, and they can then look at a position on a board, run the calculations, and come up with a move. Fred can now play chess really strongly, and simultaneously has no understanding of chess. Now in the original experiment it was a room with a person, and the person used a book to reply in Chinese. But the same idea applies.
There is no algorithm that can be memorized. Leela's understanding is in the weights, not neural net algorithms.
I'm familiar with the Chinese room argument and I've never accepted it because what it describes isn't real. It imagines some algorithm for which there is no evidence. Show me this process running and then ask me if it understands Chinese.
To me this is as philosophically dubious as the notion of p-zombies.
You can actually do calculations of LLMs and models like Leela on paper or in your head if you had enough time (and patience)! It's basically just a whole lotta matrix multiplication. It's a thought experiment and its validity does not rest on the ability for someone to actually do these calculations in a suitable timespan. The specifics of the algorithm have no relevance.
If you did see the process running, when asked would you say it understands Chinese?
You can do a thought experiment about an invisible pink dragon, that doesn't mean I have to take a position on it. "Suppose" is doing all the lifting. My position is that experiment can't happen as described.
There is no algorithm for manipulating abstract symbols in a manner that "speaks chinese" without "understanding" it. The experiment bakes in the conclusion from the beginning.
> My position is that experiment can't happen as described.
Say you are the room and are passed symbols on paper, like the suits of playing cards. You use a book (lookup table) to transform series of symbols into a new symbol, and pass it out of the room to the observer.
You get passed ♠ + ♣ and you return ♢. Do you have an understanding of the underlying concept? If so, reply and tell me what it is! But if you don't know what the underlying concept is, how could you argue that the person in the room does?
What does "understand" even mean here? So many people arguing about this seem to assume they can just use words and everyone must accept that because the words have a certain connotation, their argument must be true.
I have no idea how Magnus Carlsen "understands" chess. Neither does anyone else. His brain is giant neural net, taking inputs, sending signals around, and coming out with an output. We think we understand the mechanics of this, but we do not understand exactly why or how sending these signals around produces such good outputs.
So to argue you know for certain that an LLM is not intelligent because it is "just" a next token predictor, without knowing if that is how the human brain operates, is thinking too highly of yourself.
I don't have to try and imagine how Magnus Carlsen understands chess, since I also understand chess, and I operate with the assumption that other people are not zombies and possess a similar form of consciousness. My comment works regardless of the skill of the player.
Imagine you have never played chess, you have no concept of the rules or how the game is played, yet you've learned the entirety of Stockfish's algorithms and can dutifully run them step by step on a piece of paper when you look at a chess position. You would be the strongest chess player ever, and yet you would have less understanding of the game than even a beginner. Just because you can take an input and produce an intelligent output does not mean there is any sort of underlying understanding. This is really just a modification of Searle's Chinese Room Argument, and one of the most famous refutations of functionalism.
"In almost any other application, the biggest Achilles heel of AI is that it makes unverifiable mistakes. But in mathematics, almost uniquely, you can automatically check the output — at least if the output is supposed to be the proof of a theorem, although that is not the only thing mathematicians do. So, AI companies have recognized that their most unambiguous successes — if they’re going to have any — are going to come from mathematics.
In my opinion, there are many use cases of AI that are risky and controversial. In mathematics, the downsides are much more limited"
AI successes in mathematics don't generalize to successes in other fields as the AI promoters want to suggest.
That only explains why the post training is much more efficient - and thus where we have seen the most gains. It says nothing to support the notion that a stochastic parrot has “predicted” an original result.
I knew how LLMs work since 2019 and I've been testing their capabilities. I believe they actually are smart in every meaningful way.
"Next word prediction" just means that answer is generated through computation. I don't think computation can't be smart.
If you believe that LLMs are probabilitic and humans aren't, how do you explain randomness in human behavior? E.g. people making random typos. Have you ever tried to analyze your own behavior, understand how you function? Or do you just inherently believe you're smarter than any computation?
What's the difference between "smart" and "next word prediction", at this point? Back when they first came out, sure, but now they can write code and create art.
What would it take for you to concede a future model was smart?
My personal take would always be that it produces something that isn't in the training set, ie: Demonstrable Creativity, or innovation.
For example, it's training set it purely engineering and code with general language data set, would be "aware" what art is, but has never seen an artistic image, aware what colours are and able to create something it never saw before.
Like a child with a paintbrush, there is an intuitive behavior that happens.
Can you name any examples of a human doing this? I learned about colors, color theory, and so forth in school. I've definitely seen artistic images before.
They can already create something they've never seen - you can prompt ChatGPT to generate images, and there's a few dedicated models for it: https://chatgpt.com/images/
It’s not about them being smart or not. It’s about giving anthropic/openai/google the power to handle our future. Haven’t we learned anything about tech giants so far?
To me they seem to be pretty damn smart, to put it mildly. They sometimes do stupid things - but so do smart people!