Generative AI's failure to induce robust models of the world
garymarcus.substack.com
garymarcus.substack.com
Mainframes -> desktop computers -> a computer in every hand
Obese LLMs you visit -> agents riding with you whereever you are, integrated into your life and things -> everything everywhere, max specialization and distribution into every crevice, dominance over most tasks whether you're there or active or not
They haven't even really started working together yet. They're still largely living in sandboxes. We're barely out of the first inning. Pick a field and it's likely hardly even at the first pitch for most of them you can name, eg aircraft/flight.
In hindsight people will (jokingly?) wonder whether AI self-selected software development as one of its first conquests, as the ultimate foot in the door so it could pursue dominion over everything else (of course it had to happen in that progression; it'll prompt some chicken or the egg debates 30-50 years out).
By comparison, getting accurate feedback on whether facts are correct in a piece of text (for example) is much more difficult and expensive. At least, presumably that's why AI companies publish staged demo videos where the AI still makes factual errors half the time.
The problem is that it's difficult to create verifiers for many things we care about like architectural taste. So I expect to see superhuman capabilities on the things we can make verifiers for, but for other things it's harder to predict. We may see transfer learning or we may see collapse. My money would be more on transfer learning.
We have many, many people around the world die all the time from easily curable and preventable diseases, we just choose not to. This is largely not a technology problem. Just look at PEPFAR, which saved tens of millions of lives from HIV/AIDS. We just decided to stop funding it: https://en.wikipedia.org/wiki/President%27s_Emergency_Plan_f...
Is it possible that reason could emerge as the byproduct of being really good at predicting words? Maybe, but this depends on the antecedent claim that much if not all of reason is strictly representational and strictly linguistic. It's not obvious to me that this is the case. Many people think in images as direct sense datum, and it's not clear that a digital representation of this is equivalent to the thing in itself.
To use an example another HN'er suggested, We don't claim that submarines are swimming. Why are we so quick to claim that LLMs are "reasoning"?
Imagine we had such marketing behind wheels — they move, so they must be like legs on the inside. Then we run around imagining what the blood vessels and bones must look like inside the wheel. Nevermind that neither the structure nor the procedure has anything to do with legs whatsoever.
Sadly, whoever named it artificial intelligence and neural networks likely knew exactly what they were doing.
You can taste a beer, laugh so much it hurts, come to know how something works.
The parent comments were attempting to characterize LLMs as something more general than “word predictors”. The alternative “sequence predictors” was proposed.
My question relates to whether we have any reason to believe that the relevant aspects of human cognition are anything more than that.
Certainly humans have some advantages, like the ability to continuously learn (although there’s very strong evidence that we have a pretraining phase too, for example the difficulty of learning new languages as an adult vs. as a child.) But fundamentally, it’s not clear to me that our own language production skills aren’t “just” sequence prediction.
Perhaps, as the OP article speculates, there are other important components, like “models of the world”. But in that case, it may be that we’re augmented sequence predictors.
I'm with you on this. Software engineers talk about being in the flow when they are at their most productive. For me, the telltale sign of being in the flow is that I'm no longer thinking in English, but I'm somehow navigating the problem / solution space more intuitively. The same thing happens in many other domains. We learn to walk long before we have the language for all the cognitive processes required. I don't think we deeply understand what's going in these situations, so how are we going to build something to emulate it? I certainly don't consciously predict the next token, especially when I'm in the flow.
And why would we try to emulate how we do it? I'd much rather have technology that complements. I want different failure modes and different abilities so that we can achieve more with these tools than we could by just adding subservient humans. The good news is that everything we've built so far is succeeding at this!
We'll know that society is finally starting to understand these technologies and how to apply them when we are able to get away from using science fiction tropes to talk about them. The people I know who develop LLMs for a living, and the others I know that are creating the most interesting applications of them, already talk about them as tools without any need to anthropomorphize. It's sad to watch their frustration as they are slowed down every time a person in power shows up with a vision based on assumptions of human-like qualities rather than a vision informed by the actual qualities of the technology.
Maybe I'm being too harsh or impatient? I suppose we had to slowly come to understand the unique qualities of a "car" before we could stop limiting our thinking by referring to it as a "horseless carriage".
On a more general level, I also never understood this urge to build machines that are "just like us". Like you I want machines that, arguably, are best characterized by the ways in which they are not like us—more reliable, more precise, serving a specific function. It's telling that critiques of the failures of LLMs are often met with "humans have the same problems"—why are humans the bar? We have plenty of humans. We don't need more humans. If we're investing so much time and energy, shouldn't the bar be bette than humans? And if it isn't, why isn't it? Oh, right it's because actually human error is good enough and the actual benefit of these tools is that they are humans that can work without break, don't have autonomy, and that you don't need to listen to or pay. The main beneficiaries of this path are capital owners who just want free labor. That's literally all this is. People who actually want to build stuff want precision machines that are tailored for the task at hand, not some grab bag of sort of works sometimes stochastic doohickeys.
Most of these newer models are multi-modal, so tokens aren't necessary linguistic.
The mechanism by which they work prohibits reasoning.
This is easy to see if you look at a transformer architecture and think through what each step is doing.
The amazing thing is that they produce coherent speech, but they literally can't reason.
I think that the trained transformer has fixed weights and therefore cannot learn.
I think learning is one aspect of reasoning, and is demonstrated by challenges like navigation or puzzle solving where learning that one route to a solution is impossible is important.
I also think that the single forward pass of the model means that cyclic reasoning isn't feasible and that conditioning output by asking the model to "think" even when that thinking is done on the single forward pass means that logical processes are ruled out. The model isn't thinking in that case, the probabilities of the final part of the output are conditioned by requiring a longer initial output.
This is true of course in a pointlessly rhetorical sense.
Completely absurd though once we change "swimming" to the more precise "moving through water".
The solution is not to put arms and legs on the submarine so it can ACTUALLY swim.
It would be quite trivial to make a Gary Marcus style argument that humans still can't fly. We would need much longer and wider arms, much less core body mass, feathers.
What's more, his actual point is unclear. Even if you simply grant, "okay, even SOTA LLMs don't have world models", why do I as a user of these models care? Because the models could be wrong? Yes, I'm aware. Nevertheless, I'm still deriving subtantial personal and professional value from the models as they stand today.
Both statistical data generators and actual reasoning are useful in many circumstances, but there are also circumstances in which thinking that you are doing the latter when you are only doing the former can have severe consequences (example: building a bridge).
If nothing else, his perspective is a counterbalance to what is clearly an extreme hype machine that is doing its utmost to force adoption through overpromising, false advertising, etc. These are bad things even if the tech does actually have some useful applications.
As for benchmarks, if you fundamentally don't believe that stochastic data generation leads to reason as an emergent property, developing a benchmark is pointless. Also, not everyone has to be on the same side. It's clear that Marcus is not a fan of the current wave. Asking him to produce a substantive contribution that would help them continue to achieve their goals is preposterous. This game is highly political too. If you think the people pushing this stuff are less than estimable or morally sound, you wouldn't really want to empower them or give them more ideas.
In other words, overhyped in the short term, underhyped in the long term. Where short and long term are extremely volatile.
Take programming as an example. 2.5 years ago, gpt3.5 was seen as "cute" in the programming world. Oh, look, it does poems and e-mails, and the code looks like python but it's wrong 9 times out of 10. But now a 24B model can handle end-to-end SWE tasks in 0-shot a lot of the times.
It seems clear to me therefore that further improvements in programming ability will not come from better LLM models (which have not really improved much), but from better integration of more advanced compilers. That is, the more types of errors that can be caught by the compiler, the better chance of the AI fuzzing its way to a good overall solution. Interestingly, I hear anecdotally that current LLMs are not great at writing Rust, which does have an advanced type system able to capture more types of errors. That’s where I’d focus if I was working on this. But we should be clear that the improvements are already largely coming via symbolic means, not better LLMs.
I wrote some notes about a year ago about the irony of LLMs being considered a refutation of GOFAI when they are actually now firmly recapitulating that paradigm: https://neilmadden.blog/2024/06/30/machine-learning-and-the-...
Yes, I agree. But it's not just the cradles, it's cradles + training on traces produced with those cradles. You can test this very easily with running old models w/ new cradles. They don't perform well at all. (one of the first things I did when guidance, a guided generation framework, launched ~2 years ago was to test code - compile - edit loops. There were signs of it working, but nothing compared to what we see today. That had to be trained into the models.)
> will not come from better LLM models (which have not really improved much), but from better integration of more advanced compilers.
Strong disagree. They have to work together. This is basically why RL is gaining a lot of traction in this space.
Also disagree on llms not improving much. Whatever they did with gemini 2.5 feels like gpt3-4 to me. The context updates are huge. This is the first model that can take 100k tokens and still work after that. They're doing something right to be able to support such large contexts with such good performance. I'd be surprised if gemini 2.5 is just gemini 1 + more data. Extremely surprised. There have to be architecture changes and improvements somewhere in there.
This is because neither the LLMs nor the cradles are intelligent.
> They have to work together.
Exactly. Because they are essentially a single, brittle model. Not a "smart" text generator + a "smart" validation system.
LLMs are an enormous breakthrough in NLP and something like it will be part of an AGI system. But there is no path to AGI without more breakthroughs.
To use chess as an example. Humans sometimes play illegal moves. That does not mean Humans cannot reason. It is an instance of failing to show proof of reasoning. Not a proof of the inability to reason.
The argument is not "here's one failure case, therefore they don't reason". The argument is that systematically if you given an LLM problem instances outside training sets in domains with clear structural rules, they will fail to solve them. The argument then goes that they must not have an actual model or understanding of the rules, as they seem to only be capable of solving problems in the training set. That is, they have failed to figure out how to solve novel problem instances of general problem structures using logical reasoning.
Their strict dependence on having seen the exact or extremely similar concrete instances suggests that they don't actually generalize—they just compute a probability based on known instances—which everyone knew already. The problem is we just have a lot of people claiming they are capable of more than this because they want to make a quick buck in an insane market.
Given models can get things wrong even when the training data contains the answer, failure cannot show absence.
If you really wanted to ensure this with certainty just use the natural numbers to parameterize an aspect of a general problem. Assume there are N foo problems in the training set, then there is always a case N+1 parameter not in the training set, and you can use this as an indicative case. Go ahead and generate an insane number of these and eventually the probability that the Mth instance is not in the set is effectively 1.
Edit: Of course, it would not be perfect certainty, but it is probabilistically effectively certain. The number of problem instances in the set is necessarily finite, so if you go large enough you get what you need. Sure, you wouldn't be able to say there is a specific problem instance not in the set, but the aggregate results would evidence whether or no the LLm deals with all cases or (on assumption) just known ones.
What happens when someone makes a claim that they have gotten a model to do something not in the training data and another person claims it must be encoded in the training data in some form. It seems like an impasse.
Human fails at task due to not knowing the rules in perfect detail.
AI fails at task even though it knows the rules and could easily reproduce them for chess and dozens of chess variants.
"Look! The fallibility of humans rubbed off onto the AI, proving that they are more human and AGI than we give them credit to!"
Your statement that AI knows the rules would be considered anthropomorphising by many, I take it more to mean it 'knows' in the same sense that an election 'wants' to be at a lower energy level.
That said, humans who have written entire books on chess have been known to play illegal moves. That should count as proof by counterexample that your reasoning as to why humans fail at tasks is false.
But you misrepresented the test with respect to humans. Humans who know how to play chess don't make illegal moves.
> That said, humans who have written entire books on chess have been known to play illegal moves.
Citation needed. Unless you are talking about stories from when they first learned the rules?
5. Mouse slip
4. Forgot to call check
3. Accidentally touched 2 pieces, tried to fix it
2. Forgot to hit the clock button
1. Castle through attacked square
So, the only one of these that was an acual "illegal move" of the sort LLMs make was the castle through attacked square.
LLMs sometimes just move pieces wherever. And that does not happen when humans who know the rules play. Yes, they may mess up en passant or promotion too. But a basic "how a single piece moves" rule is what LLMs f up.
Moving through check definitely counts as as an example of a human knowing the rule and yet playing the move anyway. Which was the position you took when claiming humans would not do moves against rules they have learned.
In my experience sub 2000 players playing OTB informal chess do illegal moves fairly regularly, perhaps 1 in 50 games. Moving knights one square too far, slipping a bishop from one line to the next on a long diagonal. Castling after moving the king, not moving out of check, moving into check (especially by moving a pinned piece)
They all meet the criteria of knowing the rules and playing something else. Oftentimes people do this because they have a mistaken assumption about board state. I suspect the same is true for LLMs, they are making valid moves for what they mistakenly think the board is. That would be difficult to test, but I think possible with the right introspection tools.
I doubt any of it was rationalising for LLMs considering I was trying to address the contention that humans do not make moves counter to rules that they know. The performance of LLMs has no bearing on that claim one way or another.
Well, I hope your day is going well. Keep on cheerleading.
My initial point simplified down:
M = makes the wrong move, while knowing the rules.
A = AI Behavior
H = Human Behaviour
R = Resoning Ability
Assertion Q: if there exists an instance of M from X then X => !R
From this point down, no part of this reasoning involves Large Language models or an other aspect of AI.
Stipulation: H => R Humans can reason
Assertion Q where X is H: If there exists an instance of M from H then X=>!R
Lerc's premise L: There exists an instance of M from H
Therefore given the Stipulation either Assertion Q is false or Lerc's premise is false.
I agree that 1. does not show L but that did not matter since 5. did show L. The other points were not addressed. I also offer other examples of L that I have observed from my own experience. When I had the thought of books about chess being written by people who have made illegal moves, I actually had in mind Levy Rozman who would freely admit that he has occasionally played illegal moves.
Then you seem to want an apology for 1,2,3,4 not meeting the criteria? I'm a bit confused as to what's going on by now. One instance of L is all that is needed when L is a claim of existence. If the citation does not meet your criteria then you can simply say so, you allude to motivations regarding LLM as motivation as if you think that LLMs are still relevant to L.
You don't have to win conversations, you can just work to clarify ideas. Your request for apology, and passive aggressive sign-offs suggests you feel like this is some sort of fight. As an attempt to resolve this I have written this extended post to make as clear as possible what my position and motivations are.
I don't want to assert abilities or lack of abilities onto AI models, my concern is with whether people making such assertions are well founded. This stands for arguments saying that AI has a capability, Arguments saying AI does not have a capability, and Arguments saying AI will never have a capability.
To go back to the very beginning where someone suggested an anthropomorphic fallacy, the comparison to humans was not a suggestion of a similarity of similar function. Humans provide and example of a set of properties that are generally accepted. It is valid to apply the implications of any of those properties equally to Humans and AI. Implying the existence of a property in an AI may be anthropomorphism, evaluating the implications of the property should it exist is not.
Another algorithmic learning breakthrough, on the order of perceptrons, deep learning, transformers, etc is necessary to get anywhere near AGI.
PROMPT: Let's play a chess game. You start! e4 d5 2. exd5 e5 3. Bb5+ Bd7 4. Bxd7+ Nxd7 5. d4 Ngf6 6. dxe5 Qe7 7. f4 Qb4+ 8. Nc3 Nb6 9. exf6 Nc4 10. Qe2+ Be7 11. Qxe7+ Qxe7+ 12. Nge2 Qf8 13. fxg7 Qxg7 14. O-O Nd6 15.
RESPONSE: <played_move>15. Nxd5</played_move>
Most humans wouldn't even be able to play like this. Reasonably experienced chess players would play a lot of illegal moves.
The reason is that the encoding above requires cumulatively applying a series of actions to a two-dimensional model to which you apply rules that are described in a two-dimensional fashion.
It'd be interesting to see what the results would be if each prompt contained a two dimensional representation of the up to date board state.
I find it astonishing that people pay any attention to Gary Marcus and doubly so here. Whether or not you are an “AI optimist”, he clearly is just a bloviator.
https://www.anthropic.com/news/tracing-thoughts-language-mod...
I don't see much reason why future AI couldn't do that rather than just focusing on language though.
Understanding may not be a static symbolic representation. Contexts of the world infinite and continuously redefined. We believed we could represent all contexts tied to information, but that's a tough call.
Yes, we can approximate. No, we can't completely say we can represent every essential context at all times.
Some things might not be representable at all by their very chaotic nature.
Hm.
Dead reckoning is a terrible way to navigate, and famously led to lots of ships crashed on the shore of France before good clocks allowed tracking longitude accurately.
Ants lay down pheromone trails and use smell to find their way home... There's likely some additional tracking going on, but I would be surprised if it looked anything like symbolic GOFAI.
Many animals detect and interpret smells as chemical gradients. We don't have the hardware for it, but plenty of others do.
But some words are redacted. So I've uploaded the picture to Gemini and asked it what the redacted words are, and it told me. Not sure if they are correct, and some are way longer to fit in the redacted black box, but it didn't refuse the request.
https://arxiv.org/abs/2506.01622
Are world models a necessary ingredient for flexible, goal-directed behaviour, or is model-free learning sufficient? We provide a formal answer to this question, showing that any agent capable of generalizing to multi-step goal-directed tasks must have learned a predictive model of its environment. We show that this model can be extracted from the agent's policy, and that increasing the agents performance or the complexity of the goals it can achieve requires learning increasingly accurate world models. This has a number of consequences: from developing safe and general agents, to bounding agent capabilities in complex environments, and providing new algorithms for eliciting world models from agents.
On my reading, the philosophical claim is that these models do not develop an actual logical, internal representation of domains.
The functional import is whether or not they are able to realize specific behaviors within a domain. The paper argues that a markov process can realize the functional equivalence of the initial goal oriented picture of its domain—that is can solve goals with an error bound—but not that it develops an actual representation of the domain.
Lack of an actual representation prevents such a machine from doing other things. For example, iiuc, it would be unable to solve problems in domains that are homomorphic to the original, while an explicit representation does enable this.
The lack of a world model is a very real limitation in some problem spaces, starting with arithmetic. But this argument is unconvincing.
Obviously false for any useful sense by which you might operationalize "world model". But agree re: being a black box and having a world model being orthogonal.
The question is not whether or not they have any model at all, the question is whether the model they indisputably have (which is a model of language in terms of linear algebra) maps onto a model of the external universe (a “world model”) that emerges during training.
This is pretty much an unfalsifiable question as far as I can see. There has been research that aims to show this one way or another and it doesn’t settle the question of what a “world model” even means if you permit a “world model” to mean anything other than “thinks like we do”.
For example, LLMs have been shown to produce code that can make graphics somewhat in the style of famous modern artists (eg Kandinsky and Mondrian) but fail at object-stacking problems (“take a book, four wine glasses, a tennis ball, a laptop and a bottle and stack them in a stable arrangement”). Depending on the objects you choose the LLM either succeeds or fails (generally in a baffling way). So what does this mean? Clearly the model doesn’t “know” the shape of various 3-D objects (unless the problem is in their training set which it sometimes seems to be) but on the other hand seems to have shown some ability to pastiche certain visual styles. How is any of this conclusive? A baby doesn’t understand the 3-D world either. A toddler will try and fail to stack things in various ways. Are they showing the presence or lack of a world model? How do you tell?
But that doesn't mean that we can't, in theory, give the LLM a battery of tests that it should perform well (though not perfectly) on if it has a world model, and poorly (though not fail totally) on if it doesn't.
It's inherently a probabilistic system, so testing it in a probabilistic manner seems perfectly apt. Again: no, this will not produce a definitive result, due to that probabilistic nature—but it can produce an indicative one, and running the same test on multiple related LLMs, or similar tests on the same LLM, should help to smooth out noise in the results.
(...of course, this only works if the tests are designed well, and I don't have enough specific understanding of LLMs to know how one would go about doing that in a rigorous manner!)
If the AI has a world model, its world-model doesn't have features that allow it to do what I tested for.
I think I would probably word the distinction I would draw as "it is technically unfalsifiable, but it is not untestable."