[robbrown451]

I agree with Hinton, although a lot hinges on your definition of "understand."

I think to best wrap your head around this stuff, you should look to the commonalities of LLM's, image, generators, and even things like Alpha Zero and how it learned to play Go.

Alpha Zero is kind of the extreme in terms of not imitating anything that humans have done. It learns to play the game simply by playing itself -- and what they found is that there isn't really a limit to how good it can get. There may be some theoretical limit of a "perfect" Go player, or maybe not, but it will continue to converge towards perfection by continuing to train. And it can go far beyond what the best human Go player can ever do. Even though very smart humans have spent their lifetimes deeply studying the game, and Alpha Zero had to learn everything from scratch.

One other thing to take into consideration, is that to play the game of Go you can't just think of the next move. You have to think far forward in the game -- even though technically all it's doing is picking the next move, it is doing so using a model that has obviously looked forward more than just one move. And that model is obviously very sophisticated, and if you are going to say that it doesn't understand the game of Go, I would argue that you have a very, oddly restricted definition of the word, understand, and one that isn't particularly useful.

Likewise, with large language models, while on the surface, they may be just predicting the next word one after another, to do so effectively they have to be planning ahead. As Hinton says, there is no real limit to how sophisticated they can get. When training, it is never going to be 100% accurate in predicting text it hasn't trained on, but it can continue to get closer and closer to 100% the more it trains. And the closer it gets, the more sophisticated model it needs. In the sense that Alpha Zero needs to "understand" the game of Go to play effectively, the large language model needs to understand "the world" to get better at predicting.

[lsy]

The difference is that "the world" is not exhaustible in the same way as Go is. While it's surely true that the number of possible overall Go game states is extremely large, the game itself is trivially representable as a set of legal moves and rules. The "world model" of the Go board is actually just already exhaustive and finite, and the computer's work in playing against itself is to generate more varied data within that model rather than to develop that model itself. We know that when Alpha Zero plays a game against itself it is valuable data because it is a legitimate game which most likely represents a new situation it hasn't seen before and thus expands its capacity.

For an LLM, this is not even close to being the case. The sum of all human artifacts ever made (or yet to be made) doesn't exhaust the description of a rock in your front yard, let alone the world in all its varied possibility. And we certainly haven't figured out a "model" which would let a computer generate new and valid data that expands its understanding of the world beyond its inputs, so self-training is a non-starter for LLMs. What the LLM is "understanding", and what it is reinforced to "understand" is not the world but the format of texts, and while it may get very good at understanding the format of texts, that isn't equivalent to an understanding of the world.

[famouswaffles]

>The sum of all human artifacts ever made (or yet to be made) doesn't exhaust the description of a rock in your front yard, let alone the world in all its varied possibility.

No human or creature we know of has a "true" world model so this is irrelevant. You don't experience the "real world". You experience a tiny slice of it, a few senses that is further slimmed down and even fabricated at parts.

To the bird who can intuitively sense and use electromagnetic waves for motion and guidance, your model of the world is fundamentally incomplete.

There is a projection of the world in text. Moreover training on additional modalities is trivial for a transformer. That's all that matters.

[lsy]

That's the difference though. I know my world model is fundamentally incomplete. Even more foundationally, I know that there is a world, and when my world model and the world disagree, the world wins. To a neural network there is no distinction. The closest the entire dynamic comes is the very basic annotation of RLHF which itself is done by an external human who is providing the value judgment, but even that is absent once training is over.

Despite not having the bird's sense for electromagnetic waves, I have an understanding that they are there, because humans saw behavior they couldn't describe and investigated, in a back-and-forth with a world that has some capacity to disprove hypotheses.

Additional modalities are really just reducible to more kinds of text. That still doesn't exhaust the world, and unless a machine has some ability to integrate new data in real time alongside a meaningful commitment and accountability to the world as a world, it won't be able to cope with the real world in a way that would constitute genuine intelligence.

[famouswaffles]

>I know my world model is fundamentally incomplete. Even more foundationally, I know that there is a world, and when my world model and the world disagree, the world wins.

Yeah this isn't really true. There's not how humans work. For a variety of reasons, Plenty stick with their incorrect model despite the world indicating otherwise. In fact, this seems to be normal enough human behaviour. Everyone does it, for something or the other. You are no exception.

And yes LLMs can in fact tell truth from fiction.

GPT-4 logits calibration pre RLHF - https://imgur.com/a/3gYel9r

Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback - https://arxiv.org/abs/2305.14975

Teaching Models to Express Their Uncertainty in Words - https://arxiv.org/abs/2205.14334

Language Models (Mostly) Know What They Know - https://arxiv.org/abs/2207.05221

The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets - https://arxiv.org/abs/2310.06824

Your argument seems to boil down to "they can't perform experiments" but that isn't true either.

[user_named]

It is a very basic fact that LLMs have no concept of true or false, it only has an ability to look up what text data it has seen before. If you do not understand this you are in no position to discuss LLMs.

[SpicyLemonZest]

I really don't know what people mean when they say this. We routinely instruct computer chips to evaluate whether some condition is true and take action on that basis, even though the chip is "just" a selectively doped rock. Why would the details of an LLM's underlying architecture mean that it can't have a concept of true or false?

[xcv123]

One of the most ridiculous comments I have read about LLMs here.

The ~100 layer deep neural networks infer many levels of features over the text, including the concept of true and false. That is trivial for an LLM.

Are you completely unaware these are based on deep neural networks?

Convolutional Neural Networks don't operate by "look up" of text data.

[robbrown451]

It certainly doesn't "look up" text data it has seen before. That shows a fundamental misunderstanding of how this stuff works. That's exactly why I use the example above of Alpha Zero and how it learns to play Go, since that demonstrates very clearly that it's not just looking things up.

And I have no idea what you mean by saying that it has no concept of true or false. Even the simplest computer programs have a concept of true or false, that's kind of the simplest data type, a boolean. Large language models have a much more sophisticated concept of true and false that has a lot more nuance. That's really a pretty ridiculous thing to say.

[jeffparsons]

They have no inherent concept of true or false, sure. But what are you comparing them to? It would be bold to propose that humans have some inherent concept of true or false in a way that LLMs do not; for both humans and LLMs it seems to be emergent.

[mattigames]

In all these arguments its implied that this "genuine intelligence" is something humans all have, and nothing could be farther from the truth, that is why we have flat earthers or religious people and many other people beliving for decades easily refutable lies.

[astrange]

There is no such thing as a world model, and you don't have one of them. This is a leftover bad psychological concept from the 70s AI researchers who never got anywhere. People and other creatures do very little modeling things, they mostly just do stuff.

[xcv123]

World model means inner representation of the external world. Any organism with a functioning brain has a world model. That's what brains do.

If you don't have a world model then you are a vegetable and could not be replying on HN.

[astrange]

If you close your eyes, how long can you navigate in the environment without hitting something? Not long, because you didn't model it.

If you're taking out the recycling, do you take the time to identify (model) each piece of it first? No, because that's not necessary.

[jofla_net]

I do agree, but more importantly love this part of the argument! Its when all the personality differences become too much to bear and suddenly people are accused of not even knowing themselves. Been there before, what a wild ride!

[Dylan16807]

> suddenly people are accused of not even knowing themselves

It's not some desperate retort. People don't know themselves very well. Look at the research into confabulation, it seems to be standard operating procedure for human brains.

[pizza]

Kant would like a word with you about your point on whether people themselves understand the world and not just the format of their perceptions... :)

I think if you're going to be strict about this, you have to defend against the point of view that the same 'ding an sich' problem applies to both LLMs and people. And also whether if you had a limit sequence of KL divergences, one from a person's POV of the world, and one from an LLM's POV of texts, what it is about how a person approaches better grasp of reality - and likewise their KL divergence approaches 0, in some sense implying that their world model is becoming the same as the distribution of the world - that can only apply to people.

It seems possible to me that there is probably a great deal of lurking anthropocentrism that humanity is going to start noticing more and more in ourselves in the coming years, probably in both the direction of AI and the direction of other animals as we start to understand both better

[tazjin]

The world on our plane of existence absolutely is exhaustible, just on a much, much larger scale. Doesn't mean that the process is fundamentally different, and for the human perspective there might be diminishing returns.

[kubiton]

What if we are just the result of a ml network with a model of the world?

[user_named]

We're not.

[wbillingsley]

LLMs are very good at uncovering the mathematical relationships between words, many layers deep. Calling that understanding is a claim about what understanding is. But because we know how the LLMs we're talking about at the moment are trained, it seems to have more problems:

LLMs do not directly model the world; they train on and model what people write about the world. It is an AI model of a computed gestalt human model of the world, rather than a model of the world directly. If you ask it a question, it tells you what it models someone else (a gestalt of human writing) is most likely say. That in turn is strengthened if user interaction accepts it and corrected only if someone tells it something different.

If we were to define that as what "understanding" is, we would equivalently be saying that a human bullshit artist would have expert understanding if only they produced more believable bullshit. (They also just "try to sound like an expert".)

Likewise, I'm not convinced that we can measure its understanding just by identifying inaccuracies or measuring the difference between its answers and expert answers - There would be no difference between bluffing your way through the interview (relying on your interviewer's limitations in how they interrogate you) and acing the interview.

There seems to be a fundamental difference in levels of indirection. Where we "map the territory", LLMs "map the maps of the territory".

It can be an arbitrarily good approximation, and practically very useful, but it's a strong ontological step to say one thing "is" another just because it can be used like it.

[robbrown451]

"LLMs do not directly model the world; they train on and model what people write about the world"

This is true. But human brains don't directly model the world either, they form an internal model based on what comes in through their senses. Humans have the advantage of being more "multi-modal," but that doesn't mean that they get more information or better information.

Much of my "modeling of the world" comes from the fact that I've read a lot of text. But of course I haven't read even a tiny fraction of what GPT4 has.

That said, LLMs can already train on images, as GPT4-V does. And the image generators as well do this, it's just a matter of time before the two are fully integrated. Later we'll see a lot more training on video and sound, and it all being integrated into a single model.

[voitvodder]

We could anthropomorphize any textbook too and claim it has human level understanding of the subject. We could then claim the second edition of the textbook understands the subject better than the first. Anyone who claims the LLM "understands" is doing exactly this. What makes the LLM more absurd though is the LLM will actually tell you it doesn't understand anything while a book remains silent but people want to pretend we are living in the Matrix and the LLM is alive.

Most arguments then descend into confusing the human knowledge embedded in a textbook with the human agency to apply the embedded knowledge. Software that extracts the knowledge from all textbooks has nothing to do with the human agency to use that knowledge.

I love chatGPT4 and had signed up in the first few hours it was released but I actually canceled my subscription yesterday. Part because of the bullshit with the company these past few days but also because it had just become a waste of time the past few months for me. I learned so much this year but I hit a wall that to make any progress I need to read the textbooks on the subjects I am interested in just like I had to this time last year before chatGPT.

We also shouldn't forget that children anthropomorphize toys and dolls quite naturally. It is entirely natural to anthropomorphize a LLM and especially when it is designed to pretend it is typing back a response like a human would. It is not bullshitting you though when it pretends to type back a response about how it doesn't actually understand what it is writing.

[SkiFire13]

> One other thing to take into consideration, is that to play the game of Go you can't just think of the next move. You have to think far forward in the game -- even though technically all it's doing is picking the next move, it is doing so using a model that has obviously looked forward more than just one move.

It doesn't necessarily have to look ahead. Since Go is a deterministic game there is always a best move (or moves that are better than others) and hence a function that goes from the state of the game to the best move. We just don't have a way to compute this function, but it exists. And that function doesn't need the concept of lookahead, that's just an intuitive way of how could find some of its values. Likewise ML algorithms don't necessarily need lookahead, they can just try to approximate that function with enough precision by exploiting patterns in it. And that's why we can still craft puzzles that some AIs can't solve but humans can, by exploiting edge cases in that function that the ML algorithm didn't notice but are solvable with understanding of the game.

The thing is though, does this really matter if eventually we won't be able to notice the difference?

[bytefactory]

> It doesn't necessarily have to look ahead. Since Go is a deterministic game there is always a best move

Is there really a difference between the two? If a certain move shapes the opponent's remaining possible moves into a smaller subset, hasn't AlphaGo "looked ahead"? In other words, when humans strategize and predict what happens in the real world, aren't they doing the same thing?

I suppose you could argue that humans also include additional world models in their planning, but it's not clear to me that these models are missing and impossible for machine learning models to generate during training.

[SkiFire13]

> If a certain move shapes the opponent's remaining possible moves into a smaller subset, hasn't AlphaGo "looked ahead"?

You're confusing the reason why a move is good with how you can find that move. Yeah, a move is good due to how it shapes the opponent remaining moves, and this is also the reasoning we make in order to find that move, but it doesn't mean you can only find that move by doing that reasoning. You could have found that move just by randomly picking one, it's not very probably but it's possible. AIs just try to maximize such probability of picking a good move, meanwhile we try to find a reason a move is good. IMO it doesn't make sense to try to fit the way AI do this into our mental model, since the middle goal is fundamentally different.

[xcv123]

> Since Go is a deterministic game there is always a best move

The rules of the game are deterministic, but you may be going a step too far with that claim.

Is the game deterministic when your opponent is non-deterministic?

Is there an optimal move for any board state given that various opponents have varying strategies? What may be the best move against one opponent may not be the best move against another opponent.

[SkiFire13]

Maybe "deterministic" is not the correct term here. What I meant is that there's no probability or unknown in the game, so you can always know what are the possible moves and the relative new state.

The opponent's moves may be considered non-deterministic, but you can just assume the worst case for you, that is the best case for the opponent, which is the opponent will always play the best move too.

[user_named]

At every point in time there are a range of moves with different levels of optimality. That range changes at the next point in time following the opponent's move.

[xcv123]

The opponents strategy is an unknown variable not determined by the current board state.

Therefore the best move cannot be determined by the current board state, as it cannot be determined in isolation from the opponents strategy.

[janalsncm]

The optimal strategy can be determined from the current state. This is the principle behind minimax.

In a perfect information zero sum game, we can theoretically draw a complete game tree, each terminal node ending with a win, loss, or draw. With a full understanding of the game tree we can make moves to minimize our opponent’s best move.

[xcv123]

I stand corrected. Thanks for that explanation.

[jon_richards]

> to play the game of Go you can't just think of the next move. You have to think far forward in the game -- even though technically all it's doing is picking the next move, it is doing so using a model that has obviously looked forward more than just one move.

While I imagine alpha go does some brute force and some tree exploration, I think the main "intelligent" component of alpha go is the ability to recognize a "good" game state from a "bad" game state based on that moment in time, not any future plans or possibilities. That pattern recognition is all it has once its planning algorithm has reached the leaves of the trees. Correct me if I'm wrong, but I doubt alpha go has a neural net evaluating an entire tree of moves all at once to discover meta strategies like "the opponent focusing on this area" or "the opponent feeling on the back foot."

You can therefore imagine a pattern recognition algorithm so good that it is able to pick a move by only looking 1 move into the future, based solely on local stone densities and structures. Just play wherever improves the board state the most. It does not even need to "understand" that a game is being played.

> while on the surface, they may be just predicting the next word one after another, to do so effectively they have to be planning ahead.

So I don't think this statement is necessarily true. "Understanding" is a major achievement, but I don't think it requires planning. A computer can understand that 2+2=4 or where to play in tic-tac-toe without any "planning".

That said, there's probably not much special about the concept of planning either. If it's just simulating a tree of future possibilities and pruning it based on evaluation, then many algorithms have already achieved that.

[theGnuMe]

The "meta" here is just the probability distribution of stone densities. The only way it can process those is by monte Carlo simulation. The DNN (trained by reinforcement learning) evaluates the simulations and outputs the top move(s).

[klodolph]

> As Hinton says, there is no real limit to how sophisticated they can get.

There’s no limit to how sophisticated a model can get, but,

1. That’s a property shared with many architectures, and not really that interesting,

2. There are limits to the specific ways that we train models,

3. We care about the relative improvement that these models deliver, for a given investment of time and money.

From a mathematical perspective, you can just kind of keep multiplying the size of your model, and you can prove that it can represent arbitrary complicated structures (like, internal mental models of the world). That doesn’t mean that your training methods will produce those complicated structures.

With Go, I can see how the model itself can be used to generate new, useful training data. How such a technique could be applied to LLMs is less clear, and its benefits are more dubious.

[Jensson]

A big difference between a game like Go and writing text is that text is single player. I can write out the entire text, look at it and see where I made mistakes on the whole and edit those. I can't go back in a game of Go and change one of my moves that turned out to be a mistake.

So trying to make an AI that solves the entire problem before writing the first letter will likely not result in a good solution while also making it compute way too much since it solves the entire problem for every token generated. That is the kind of AI we know how to train so for now that is what we have to live with, but it isn't the kind of AI that would be efficient or smart.

[bytefactory]

This doesn't seem like a major difference, since LLMs are also choosing from a probability distribution of tokens for the most likely one, which is why they respond a token at a time. They can't "write out' the entire text at a time, which is why fascinating methods like "think step by step" work at all.

[Jensson]

But it can't improve its answer after it has written it, that is a major limitation. When a human writes an article or response or solution, that is likely not the first thing the human thought of, instead they write something down and works on it until it is tight and neat and communicates just what the human wants to communicate.

Such answers will be very hard for an LLM to find, instead you mostly get very verbose messages since that is how our current LLM thinks.

[bytefactory]

Completely agree. The System 1/System 2 distinction seems relevant here. As powerful as transformers are with just next-token generation and context, which can be hacked to form a sort of short-term memory, some time of real-time learning + long-term memory storage seems like an important research direction.

[xcv123]

> But it can't improve its answer after it has written it, that is a major limitation.

It can be instructed to study its previous answer and find ways to improve it, or to make it more concise, etc, and that is working today. That can easily be automated by LLMs talking to each other.

[ewild]

that is true and isnt. GPT4 has shown itself to halfway through a answer say "wait thats not correct im sorry let me fix that" and then correct itself. For example it stated a number was prime and why, and when showing the steps found it was divisible by 3 and said "oh i made a mistake it actually isnt prime"

[Someone]

> There may be some theoretical limit of a "perfect" Go player, or maybe not, but it will continue to converge towards perfection by continuing to train

I don’t think that’s a given. AlphaZero may have found an extremely high local optimum that isn’t the global optimum.

When playing only against itself, it won’t be able to get out of that local optimum, and when getting closer and closer to it even may ‘forget’ how to play against players that make moves that AplhaGo never would make, and that may be sufficient for a human to beat it (something like that happened with computer chess in the early years, where players would figure out which board positions computers were bad at, and try to get such positions on the board)

I think you have to keep letting it play against other good players (human or computer) that play differently to have it keep improving, and even then, there’s no guarantee it will find a global optimum.

[theGnuMe]

Alphazero runs monte carlo tree search so it has a next move "planning" simulator. This computes the probability that specific moves up to some distance lead to a win.

LLMs do not have a "planning" module or simulator. There is no way the LLM can plan.

Could build a planning system into an LLM? Possibly and probably, but that is still open research. LeCunn is trying to figure out how to train them effectively. But even an LLM with a planning system does not make it AGI.

Some will argue that iteratively feeding the output embedding back into the input will retain the context but even in those cases it rapidly diverges or as we say "hallucinates"... still happens even with large input context windows. So there is still no planning here and no world model or understanding.

[eviks]

The issue with Alpha Zero analogy extremes is that those are extremely constrained conditions, so can't be generalized to something infinitely more complicated like speech

And

> When training, it is never going to be 100% accurate in predicting text it hasn't trained on, but it can continue to get closer and closer to 100% the more it trains.

For example, it can reach 25% of accuracy and have an math limit of 26%, so "forever getting closer to 100% with time" would still result in a waste of even infinite resources

[icy_deadposts]

> there isn't really a limit to how good it can get.

> it will continue to converge towards perfection

Then someone discovered a flaw that made it repeatably beatable by relative amateurs in a way that no human player would be

https://www.vice.com/en/article/v7v5xb/a-human-amateur-beat-...

[user_named]

It's not planning ahead, it is looking at the probabilities of the tokens altogether rather than one by one.

[anothernewdude]

> You have to think far forward in the game -

I disagree. You can think in terms of a system that doesn't involve predictions at all, but has the same or similar enough outcome.

So an action network just learns patterns. Just like a chess player can learn what positions look good without thinking ahead.

[huytersd]

Next word generation is one way to put it. The key point here is we have no idea what’s happening in the black box that is the neural network. It could be forming very strong connections between concepts in there with multi tiered abstractions.

[theGnuMe]

It is certainly not abstracting things.

[gilbetron]

If LLMs are just glorified autocompletion, then humans are too!

[notjoemama]

> I would argue that you have a very, oddly restricted definition of the word, understand, and one that isn't particularly useful.

Is it just me or does this read like “here is my assumption about what you said, and now here is my passive aggressive judgement about that assumption”? If you’re not certain about what they mean by the word “understand”, I bet you could ask and they might explain it. Just a suggestion.

[SpicyLemonZest]

I've asked that question in the past and I've never gotten an answer. Some people sidestep the question by describing something or other that they're confident isn't understanding; others just decline to engage entirely, asserting that the idea is too ridiculous to take seriously. In my experience, people with a clear idea of what they mean by the word "understand" are comfortable saying that ML models understand things.

[greenthrow]

This is absolute nonsense. The game of Go is a grid and two colors of pieces. "The world" here is literally everything.