People need to get ahead of the curve and understand that "world models" as touted are just the next pitch to VCs with no substantive conceptual progress. We don't know how to construct world models that are functionally like a human's (if it's even coherent to describe human cognition in those pretty dualistic terms) because we don't know how to model embodiment (simulating bodies in video games is a farce).
Symbolic AI faced intractable problems with world modelling because it was the wrong substrate, just as LLMs are, but for different reasons. It would be as reprehensible as the current strategy in Silicon Valley to not learn from that history.
+ the idea of a neurosymbolic hybrid is just trying to squeeze a right out of two wrongs. There were fundamental format issues with symbolic representations that cannot be mended by supplementing them with statistical bells and whistles.
All of which leads to the fact that the world needs more money dedicated to research instead of throwing it into the open jaws of people like Scam Altman, Dario Misfit Amodei and above all Mark .uckerberg!
They dont give money to real AI research because soon AI will do all of the research. Haven't you heard? And than you'd waste money on research, when you could just feed a monster that will soon do all research for you.
The holy grail is to be able to derive world models from data using unsupervised learning. I’m not aware any system that would be able to that. The way humans and other animals do that is by interacting with the actual world. We internally generate predictions about the future state of the world based on our current model, compare it to what actually happens and adjust the model accordingly. This requires ability to select observations/facts relevant to the situation and that is hard problem. So, I don’t think that combining AI based on explicit world models with NN-like learning is necessarily a dead end, but there are many hard problems that need to be solved. It is definitely in the realm of fundamental research (as opposed to engineering).
Actually humans learn text book knowledge as a model of world knowledge Its called school. The equivalent training is importing text book knowledge into a semantic ontology. Thats been done for decades.
This is a misuse of ‘sophism’. But of course it’s handwaving and that doesn’t matter - it’s a one sentence gesture at a blindspot some people may have on the issue.
Have you seen what the Thousand Brains Project has achieved with their open source Monty initiative with eight full time staff in the last year: https://thousandbrains.org
Sensory-motor object recognition using 57 million times less energy than training an LLM 🤗
They are backed by the Gates Foundation and a Korean tech outfit.
Very early days, but doing world-modelling the way the neocortex works.
Even if a world model conceptually exists. the problem is the cost of human labor to hand code any world model. Gary often mentions CYC as the proof. What is considered impossible is automating knowledge acquisition.
The link below describes a process to automate knowledge acquisition from text books and corpora to build ontologies with massive deployment.
Symbolic representations which encode natural language propositions are not world models, they're models of world models. Their failure in principle is described well in Hubert Dreyfus's 'What Computers Can't Do'.
I would go so far as to say that it's a contradiction in spirit to think that you can have a world model which doesn't enable knowledge acquisition - they're of a piece. If you don't have knowledge acquisition, whatever you do have cannot be a world model.
Really helpful Thanks. The specific operational systems I discuss as SAM-1 can be extended to manage causality and beliefs while storing facts with the limitations you provided.
Dreyfus's arguments are well known to be fallacious, based on misunderstandings of Gödel's incompleteness theorems ... Roger Penrose based his refuted nonsense on the same mistakes.
Dreyfus’s main arguments were based in Heidegger and Merleau-Ponty. Whatever he wrote on Gödel has no substantive bearing on his critique of symbolic AI. Penrose’s criticism of computationalism is an entirely different thing - mentioning it in the same breath shows you don’t understand Dreyfus’s work.
"We don't know how to construct world models that are functionally like a human's (if it's even coherent to describe human cognition in those pretty dualistic terms)"
Functionalism isn't dualistic.
"Symbolic AI faced intractable problems with world modelling because it was the wrong substrate"
This simply isn't true.
"It would be as reprehensible as the current strategy in Silicon Valley to not learn from that history."
Silicon Valley is reprehensible, but for other reasons.
A critique of dualism in world model talk has to do with representationalism, not functionalism.
And though you might want to argue that symbolic AI doesn’t face problems in principle with world modelling (good luck), it is simple historical fact that it did face tractability problems in practice - that was the reason for the first AI winter.
I believe LLMs got so hyped because they felt like a big shortcut. Anyone can understand that building world models to a level that would actually revolutionize artificial intelligence is damn near impossible, just as neurosymbolic AI is hard because it has to be actually robust and has to follow the rules. It requires very hard work. Building true artificial intelligence is extremely difficult. We knew this before the current AI hype.
LLMs felt like cheating because it initially just seemed to "grow" and develop "emergent" capabilities. I am fairly certain this is why many people got so hooked on the idea. It felt like: "Maybe we don't actually have to do it the hard way. We can just make it bigger and it will probably start evolving by itself".
LLMs are, as was apparent quite early on, just an illusion of intelligence by means of scaling a quite simple concept to an absurd degree. It's sad to me that even the brilliant people working on these technologies sort of just ignored these facts and instead got high on the "what if?"
exactly my thought. People tried to go around fundamental problems of knowledge representation, inference, etc. LLMs were driven farther than they should have not by scientific community but by pandering to public. They don't know any better and have no way to tell that the language just looks like thinking. It was very dishonest of the community IMO
I think a large part of the problem is the tech bros are very bad philosophers and at heart still Cartesian dualist. They think there is a world out there that a mind in here understands. They fail to grasp that we don't understand the world, but experience the world and things like time, gravity and causality are not primarily know but lived. There is a union of mind, body and world where the interconnecting boundaries between these notions is the fundamental reality not these states as separable things. Our primary world models are not based on I think, but I can. We are phenomenological creatures prior to being rational ones and the world opens up to us as things that matter in relation to intentions (can I eat it, will it eat me, can I jump that space, is it hot etc) rather than disinterested knowledge. So to build true AGI there machine would have to understand the world in an embodied way and have intentionality towards it. I'm not sure that would ever be possible. I'm sure we can make this tech smarter and more useful as we have with so many tools and technologies, but pretending it is intelligent in any way that livings things are is wrong headed.
I’m also not sure that, even if it was possible, building a machine possessing lived experience (and hence also own opinions about the world) would be a good idea. How would we benefit from such a creation?
Gary, your paper (which is a good read, by the way, it's part of my 'essential reading on where we are in AI' library) is from 2020. It references some older papers when talking about neurosymbolic (e.g. NS-CL from 2019). Can you point us to the current best research on neurosymbolic? What is the state of the art currently?
The "neurosymbolic" approach in this vein is most notable in Ben Goertzel's work. He uses the "MeTTa cognitive language to integrate neural networks, symbolic reasoning, and evolutionary learning".
It has promise for specialized areas where a rigorous and provable rigorous chain is needed. It is far from clear it is practical. Seems that LLM can imitate faster, more flexibly and ultimately cheaper than any special-built rigorous method.
Understanding comes from using models. "Neurosymbolic" doesn't mean "understanding". It means follow rules. A fluid simulator understands fluids well-enough, because it can predict their properties, for example.
So, imitation coupled with world models, which can come from invoking tools, becomes good enough of understanding to function properly.
Our own understanding is also limited and context-dependent.
Part of the problem here is that with Scam Altman etc. sucking all of the oxygen (ie. money) out of the room, very little funding is availbale for research of other paths
How large a part? Altman is sucking VC money out of the world (and we’re close to scraping the bottom of the barrel here), research money I would guess, is available for convincing avenues. It seems to me those are not widely available and the few I’ve seen (e.g. LiquidNet, NS-CL) are almost as old as the transformers that enabled the massive scaling of RNNs.
I might bring out my old 1996 expert system for classifying poisonous mushrooms then. It needed a power supply in the forest for the desktop PC and was only 65% accurate at recommending a non-deadly mushroom. So room for improvement.
“Neurosymbolic” means little more than “neural nets and something else.” It’s so broad as to be meaningless.
“World models” similarly means little more than “not just language in LLMs.” The top players (LeCun, Li Fei-Fei, etc) are pursuing the dumbest approaches to this end like generative video, 3D models and game engines.
Neural nets don’t have explicit representations. Symbolic has explicit representations. Combine the two and you have almost the entire space of possible algorithms (forgive the loose terminology).
“Neurosymbolic AI” is a tautology that is too broad to be any useful. Hence it’s functionally meaningless.
We know it is possible to have neurosymbolic in principle: humans are functionally like that. What we do not have (yet) is a practical alternative/artificial implementation created by ourselves.
We don’t know what human brains are. We know what they’re not, and they’re certainly not neural nets, as backprop is incompatible with biological constraints.
My bilingual experience leads me to believe they’re not traditional symbolic models either, which map word relationships via linguistic operators like IS and HAS. All language-based approaches, even those that ostensibly loop in “world models,” are biased toward Germanic/Romantic grammar and are a dead end for AGI.
They are indeed not like artificial neural nets (e.g. with backpropagation), but they *are* networks of neurons. And one can see that this fact is at the basis of much naive belief in the power of the artificial ones (Hinton, Sutskever, LeCun). Or even more naive belief that such artificial networks operating at token/pixel level could reach AGI (LeCun and Sutskever earlier, but no way this is going to fly).
But human brains consists of a connectome (a network of connections of neurons) and exhibit behaviour that includes (some, not much) symbolic/discrete reasoning. How that happens we do not really know, though it seems likely to me that interferences of analog signals with enough stability to create 'discreteness' will play a role. And of course, neurons do have a characteristic that will help with this (the signalling threshold). But how exactly that works, nobody yet knows afaik.
Human world model example. In psychosis the failure to correctly integrate visual cortex and limbic system inputs to the PFC can result in people experiencing others as adroids/robots. The human world model we are apparently pursuing needs both cognitive and emotional valence for accurate perception (see Phineas Gage wiki). How do we create an equivalent for an ai?. LLM are really useful as a narrow ai technology but other than that they are still in reality phenomenal stochastic machines. Dead end - no. A hammer is a dead end if you want to put in a screw properly, but if you have a nail.... World Models as a concept is a great idea but I invite anyone to define their own in a way that explains their own HGI (Human Generalised Intelligence).
Is a World Model possible? A model is a structured, simplified representation of reality — necessarily incomplete, yet practical when it includes only what’s known to matter. What counts as “significant” stems from observation, experience, and meaning. A model is valuable if it can broadly predict future behavior. Engineering applies such models while accounting for the unknown, adding safety margins and avoiding single points of failure. Each model belongs to a specific domain and depends on judgment. Once validated, it guides optimization and performance decisions. True AGI would require a universal model spanning physics to social behavior — an impossible task for the foreseeable future. As Stuart Kauffman’s fundamental work on the non-ergodicity of nature reminds us, reality cannot be fully captured. AI excels in defined domains, but pursuing a “genetal world model” is an unrealistic ambition.
Complete non sequitur. The point was that world models are possible, and that a "universal model" isn't necessary.
Also, people like to say "arguably" when they actually have no argument.
P.S. I see from your feed that you are an anti-trans bigoted piece of trash who thinks the Democratic Party should be outlawed, and I didn't delve further into your right wing rantings ... not the sort of person I choose to engage with.
How does one construct a 'world' model? (I did some consulting in the Himalayan foothills, and we needed an interpreter for two villages on two hilltops which were only 8 miles apart, but spoke completely different languages.) I am tempted to say that there is no such thing as a world model.
And mostly I work with the idea of 'micro-anthropologies' at best. Sure there are commonalities, but 'world' models? My experience is that context is the basis for meaning and identity.
Gary touched on that in his Eisman interview. There are many different world models, cultures and realities; the trick is to know which one you're operating in.
Thank you! I agree that the term is poorly defined. For one thing, are we talking about a model of the "real world" or just a model that we can call "the world"? If the former, then I'd like to adduce the following working definition: (1) that world is real to me in which my actions have consequences; (2) that world is real to us in which our interactions are reciprocal.
Are we framing the goal wrong.. we know that all the talk of AGI was hyped up from journalists and people like Sam Altman, because they wanted something sensational to sell papers or to raise money.
I do think we should celebrate how far the LLM approach has taken us though and how much value there is for us to apply to the world now.. Shouldn't we put AGI out of everyones minds for now and focus on this?
1. Agree about wrong goal - unattainable and idiotic
2. There is some value in the LLM approach - as long as you only treat it as offering drafts to be checked, for instance. But for me is outweighed by the damage its use for chatbots seems to be causing.
Completely agree, but why can't a particular person see that? I put it down to naivety about real-world problems, and the raft of conditions that go with them.
Large Language Models are criticized for “lacking a world model,” as if understanding must precede language. But language itself is already a compressed interface to the world — not a substitute for it, but a historically evolved judgment layer built from collective interaction with reality.
What’s often missed is this:
LLMs don’t fail because they lack causality — they fail where *human judgment itself has not been made explicit*. Causality is not something humans possess as an internal simulator; it is something we stabilize through judgment under real constraints, error correction, and consequence.
Calling LLMs “mere pattern matchers” overlooks that human reasoning is also statistical, embodied, and historically conditioned — just embedded in biology instead of silicon.
World models matter. But dismissing language-based systems as a dead end assumes that understanding must look like physics engines in the head. That assumption itself is unexamined.
The real open question isn’t “language vs world models,” but whether we know how to externalize, stabilize, and test *judgment* — in machines or in ourselves.
Avid Xu: We test judgments with their backdrop of meaningful, well-understood, accumulated, hopefully redundant or even invariant, and tested evidence until we reach enough relevant intelligibility where no more serious questions arise, then we say YES, it is or NO it's not, or I don't know--let's go back to the drawing board, maybe even change the meaning of our questions. Catherine Blanche King (See B. Lonergan's Insight, A Study of Human Understanding, 2000).
"Calling LLMs “mere pattern matchers” overlooks that human reasoning is also statistical, embodied, and historically conditioned — just embedded in biology instead of silicon."
This is oft repeated by LLMphiles but it is simply false ... that is not how human reasoning works, and it misses the fact that LLMs pattern match over tokens in their training data (which consists of human utterances, which is where all of the apparent "understanding" of LLMs comes from), not facts in the world.
it’s the clearest statement i have seen from Demis that LLMs are likely a dead end and that world models are necessary. he’s been exploring them for a while (though the company was founded on a model-free approach to Atari games).
thank you for your insight. why natural system differ from tech and not compatible...but it has great potential for many utility. One is not expert on anything..perhaps useful following ....if not trash .one click away..
Time is not a separate parameter for it only indicates the interactive interval in an axiomatically dynamic oscillatory state. Therefore, as 1.344*10^-51 was derived on the bases of an accelerative transmigration from all six sides of a volumetric state,
(1.59E -18) *(6.283E +17) =1
Two important principles are covered in this equivalence. Volumetric change in interactions in the simultaneous state is equal to the change in the cyclic interval thereby proving that no stresses are unaccounted for. Or that the algebraic sum of the exchange of compressive and expansive stresses are zero.
A log one cycle reduction in potential count increases interactive counts by 10 units, which then releases a set of 2Pi*10^17*10=2PI*10^18 reciprocal of which 1.591681E-19 which is value of an electromagnetic charge of a volt in Physics
Planck’s constant h is considered the energy quantum, in actual fact 1.344*10^-51 is the true fundamental quantum in a completely unified field, for it determines the limiting mass, the 3.571428 state and the cyclic potential as of 18 orders all of which almost equal the Planck’s quantum. It underlines the error-potential in Physics when dealing at that energy level, for the Planck’s quantum is not a true elemental state. It suggest including human body to …all have this potential and is underlying mechanism must most important part of creating all new system…,models…etc…..….and has inherent limit of nature.
As noted before, Nobel Laureate AI co-founder and symbolic AI Tower of Hanoi wizard Herb Simon predicted in 1960 that by 1980 computers could do any work a man could do. (Probably included women’s capabilities as well.) This might be enough to end the discussion, but is unlikely to. Cognitive world models are good for Tower of Hanoi, chess, some coding. Homo sapiens arrives with cognitive world models, but more emotional and social world models which we don’t understand well enough to emulate. Gary noted that bolting some symbolic onto neural won’t work. Cyc bolted some neural onto 25 years of symbolic work and didn’t set the world on fire. But Philip Wilkinson noted a key point: Simon and the other pioneers saw what we call AGI as the goal. Many commenters do as well. But some of us never have, our goal is to find something useful, not harmful, and economically feasible. There is no question that LLMs can be useful. But harmless and capable of finding a revenue model are not clear.
It suffices to look at where we are with autonomous driving (i.e. navigating a rather simple 2D environment based on a handful degrees of freedom) after more than a decade and billions in R&D to see that AI based on "world models" is just the next chapter in wishful thinking.
As usual Darwin will set you free! The only intelligence we know of evolved by a process of natural selection. There is a continuum of agentic capabilities from us all the way to bacteria. Intelligence is about predicting the future of the world for fun & profit. When the future world includes other intelligences, some of whom you can talk to, you quickly get into an arms race of competition between the quality of your predictive models. But there’s another catch that was pointed out by Donald Hoffman. Our (evolved) senses only give us fitness not truth about the world. We don’t even know how many dimensions of space there are or if time is fundamental or emergent. Evolution conveniently hides the details from us. Building AIs based on a human sense of the world is absurd. It’s akin to building a computer inside Minecraft: you can do it, but it is tremendously inefficient and cannot possibly scale. This is the main problem with LLMs - they are so abstracted from reality that they cannot possibly deal with it. Until AI takes a truly bottom-up and evolutionary approach, all of it is doomed to failure.
People need to get ahead of the curve and understand that "world models" as touted are just the next pitch to VCs with no substantive conceptual progress. We don't know how to construct world models that are functionally like a human's (if it's even coherent to describe human cognition in those pretty dualistic terms) because we don't know how to model embodiment (simulating bodies in video games is a farce).
Symbolic AI faced intractable problems with world modelling because it was the wrong substrate, just as LLMs are, but for different reasons. It would be as reprehensible as the current strategy in Silicon Valley to not learn from that history.
+ the idea of a neurosymbolic hybrid is just trying to squeeze a right out of two wrongs. There were fundamental format issues with symbolic representations that cannot be mended by supplementing them with statistical bells and whistles.
All of which leads to the fact that the world needs more money dedicated to research instead of throwing it into the open jaws of people like Scam Altman, Dario Misfit Amodei and above all Mark .uckerberg!
I much prefer Ed Zitron’s “Wario Amodei”. Seems accurate.
Yes you are right ... Wario is much better. I just didn't think of it when I wrote my earlier comment. My bad!
I’ve heard him say that so many times that I forgot his name isn’t Wario 🤫
They dont give money to real AI research because soon AI will do all of the research. Haven't you heard? And than you'd waste money on research, when you could just feed a monster that will soon do all research for you.
How soon?
A) 1 to 5 years
B) By the end of the decade
C) Sooner than people think...
D) It already happened, we just don't know
The holy grail is to be able to derive world models from data using unsupervised learning. I’m not aware any system that would be able to that. The way humans and other animals do that is by interacting with the actual world. We internally generate predictions about the future state of the world based on our current model, compare it to what actually happens and adjust the model accordingly. This requires ability to select observations/facts relevant to the situation and that is hard problem. So, I don’t think that combining AI based on explicit world models with NN-like learning is necessarily a dead end, but there are many hard problems that need to be solved. It is definitely in the realm of fundamental research (as opposed to engineering).
Actually humans learn text book knowledge as a model of world knowledge Its called school. The equivalent training is importing text book knowledge into a semantic ontology. Thats been done for decades.
http://intellisophic.net/2025/09/12/the-fundamental-innovation-orthogonal-corpus-indexing-oci/
Please cite or describe these issues. Did W3C RDF have the problems you allude to?
Indeed ... It gets us nowhere
This is absurd sophism and handwaving.
This is a misuse of ‘sophism’. But of course it’s handwaving and that doesn’t matter - it’s a one sentence gesture at a blindspot some people may have on the issue.
I don’t see how we can get to artificial intelligence without affect.
Have you seen what the Thousand Brains Project has achieved with their open source Monty initiative with eight full time staff in the last year: https://thousandbrains.org
Sensory-motor object recognition using 57 million times less energy than training an LLM 🤗
They are backed by the Gates Foundation and a Korean tech outfit.
Very early days, but doing world-modelling the way the neocortex works.
Regards,
Trevor
Even if a world model conceptually exists. the problem is the cost of human labor to hand code any world model. Gary often mentions CYC as the proof. What is considered impossible is automating knowledge acquisition.
The link below describes a process to automate knowledge acquisition from text books and corpora to build ontologies with massive deployment.
http://intellisophic.net/2025/09/12/the-fundamental-innovation-orthogonal-corpus-indexing-oci/
Symbolic representations which encode natural language propositions are not world models, they're models of world models. Their failure in principle is described well in Hubert Dreyfus's 'What Computers Can't Do'.
I would go so far as to say that it's a contradiction in spirit to think that you can have a world model which doesn't enable knowledge acquisition - they're of a piece. If you don't have knowledge acquisition, whatever you do have cannot be a world model.
Really helpful Thanks. The specific operational systems I discuss as SAM-1 can be extended to manage causality and beliefs while storing facts with the limitations you provided.
http://aicyc.org/2025/09/16/extending-sam-to-handle-causality-with-do-calculus/
I appreciate the reference and will take a look!
Dreyfus's arguments are well known to be fallacious, based on misunderstandings of Gödel's incompleteness theorems ... Roger Penrose based his refuted nonsense on the same mistakes.
Dreyfus’s main arguments were based in Heidegger and Merleau-Ponty. Whatever he wrote on Gödel has no substantive bearing on his critique of symbolic AI. Penrose’s criticism of computationalism is an entirely different thing - mentioning it in the same breath shows you don’t understand Dreyfus’s work.
"We don't know how to construct world models that are functionally like a human's (if it's even coherent to describe human cognition in those pretty dualistic terms)"
Functionalism isn't dualistic.
"Symbolic AI faced intractable problems with world modelling because it was the wrong substrate"
This simply isn't true.
"It would be as reprehensible as the current strategy in Silicon Valley to not learn from that history."
Silicon Valley is reprehensible, but for other reasons.
A critique of dualism in world model talk has to do with representationalism, not functionalism.
And though you might want to argue that symbolic AI doesn’t face problems in principle with world modelling (good luck), it is simple historical fact that it did face tractability problems in practice - that was the reason for the first AI winter.
I believe LLMs got so hyped because they felt like a big shortcut. Anyone can understand that building world models to a level that would actually revolutionize artificial intelligence is damn near impossible, just as neurosymbolic AI is hard because it has to be actually robust and has to follow the rules. It requires very hard work. Building true artificial intelligence is extremely difficult. We knew this before the current AI hype.
LLMs felt like cheating because it initially just seemed to "grow" and develop "emergent" capabilities. I am fairly certain this is why many people got so hooked on the idea. It felt like: "Maybe we don't actually have to do it the hard way. We can just make it bigger and it will probably start evolving by itself".
LLMs are, as was apparent quite early on, just an illusion of intelligence by means of scaling a quite simple concept to an absurd degree. It's sad to me that even the brilliant people working on these technologies sort of just ignored these facts and instead got high on the "what if?"
exactly my thought. People tried to go around fundamental problems of knowledge representation, inference, etc. LLMs were driven farther than they should have not by scientific community but by pandering to public. They don't know any better and have no way to tell that the language just looks like thinking. It was very dishonest of the community IMO
I think a large part of the problem is the tech bros are very bad philosophers and at heart still Cartesian dualist. They think there is a world out there that a mind in here understands. They fail to grasp that we don't understand the world, but experience the world and things like time, gravity and causality are not primarily know but lived. There is a union of mind, body and world where the interconnecting boundaries between these notions is the fundamental reality not these states as separable things. Our primary world models are not based on I think, but I can. We are phenomenological creatures prior to being rational ones and the world opens up to us as things that matter in relation to intentions (can I eat it, will it eat me, can I jump that space, is it hot etc) rather than disinterested knowledge. So to build true AGI there machine would have to understand the world in an embodied way and have intentionality towards it. I'm not sure that would ever be possible. I'm sure we can make this tech smarter and more useful as we have with so many tools and technologies, but pretending it is intelligent in any way that livings things are is wrong headed.
I’m also not sure that, even if it was possible, building a machine possessing lived experience (and hence also own opinions about the world) would be a good idea. How would we benefit from such a creation?
As such a machine would have subjectivity making them labour for you would be tantamount to slavery so it is indeed a terrible idea.
Yes, time to reread Winograd and Flores.
"word model" sounds like a generalisation of expert system, which was the AI approach a few decades ago. That was also a dead end
please read the paper cited at the end which is a clear call for a neurosymbolic hybrid rather than pure expert system
Gary, your paper (which is a good read, by the way, it's part of my 'essential reading on where we are in AI' library) is from 2020. It references some older papers when talking about neurosymbolic (e.g. NS-CL from 2019). Can you point us to the current best research on neurosymbolic? What is the state of the art currently?
The "neurosymbolic" approach in this vein is most notable in Ben Goertzel's work. He uses the "MeTTa cognitive language to integrate neural networks, symbolic reasoning, and evolutionary learning".
It has promise for specialized areas where a rigorous and provable rigorous chain is needed. It is far from clear it is practical. Seems that LLM can imitate faster, more flexibly and ultimately cheaper than any special-built rigorous method.
But is imitation really a worthwhile goal? I’d prefer understanding.
Understanding comes from using models. "Neurosymbolic" doesn't mean "understanding". It means follow rules. A fluid simulator understands fluids well-enough, because it can predict their properties, for example.
So, imitation coupled with world models, which can come from invoking tools, becomes good enough of understanding to function properly.
Our own understanding is also limited and context-dependent.
Part of the problem here is that with Scam Altman etc. sucking all of the oxygen (ie. money) out of the room, very little funding is availbale for research of other paths
How large a part? Altman is sucking VC money out of the world (and we’re close to scraping the bottom of the barrel here), research money I would guess, is available for convincing avenues. It seems to me those are not widely available and the few I’ve seen (e.g. LiquidNet, NS-CL) are almost as old as the transformers that enabled the massive scaling of RNNs.
I might bring out my old 1996 expert system for classifying poisonous mushrooms then. It needed a power supply in the forest for the desktop PC and was only 65% accurate at recommending a non-deadly mushroom. So room for improvement.
I have just read your paper with help of ChatGPT :-)
having a world model sounds reasonable, but the tricky part is how to implement it.
I assume nobody has an idea how to do it apart from good-old hand-crafting.
“Neurosymbolic” means little more than “neural nets and something else.” It’s so broad as to be meaningless.
“World models” similarly means little more than “not just language in LLMs.” The top players (LeCun, Li Fei-Fei, etc) are pursuing the dumbest approaches to this end like generative video, 3D models and game engines.
No one is even close to even thinking about AGI.
it typically means having explicit symbolic representations, ontologies etc. it is not meaningless, it is broad.
Neural nets don’t have explicit representations. Symbolic has explicit representations. Combine the two and you have almost the entire space of possible algorithms (forgive the loose terminology).
“Neurosymbolic AI” is a tautology that is too broad to be any useful. Hence it’s functionally meaningless.
We know it is possible to have neurosymbolic in principle: humans are functionally like that. What we do not have (yet) is a practical alternative/artificial implementation created by ourselves.
We don’t know what human brains are. We know what they’re not, and they’re certainly not neural nets, as backprop is incompatible with biological constraints.
My bilingual experience leads me to believe they’re not traditional symbolic models either, which map word relationships via linguistic operators like IS and HAS. All language-based approaches, even those that ostensibly loop in “world models,” are biased toward Germanic/Romantic grammar and are a dead end for AGI.
They are indeed not like artificial neural nets (e.g. with backpropagation), but they *are* networks of neurons. And one can see that this fact is at the basis of much naive belief in the power of the artificial ones (Hinton, Sutskever, LeCun). Or even more naive belief that such artificial networks operating at token/pixel level could reach AGI (LeCun and Sutskever earlier, but no way this is going to fly).
But human brains consists of a connectome (a network of connections of neurons) and exhibit behaviour that includes (some, not much) symbolic/discrete reasoning. How that happens we do not really know, though it seems likely to me that interferences of analog signals with enough stability to create 'discreteness' will play a role. And of course, neurons do have a characteristic that will help with this (the signalling threshold). But how exactly that works, nobody yet knows afaik.
Human world model example. In psychosis the failure to correctly integrate visual cortex and limbic system inputs to the PFC can result in people experiencing others as adroids/robots. The human world model we are apparently pursuing needs both cognitive and emotional valence for accurate perception (see Phineas Gage wiki). How do we create an equivalent for an ai?. LLM are really useful as a narrow ai technology but other than that they are still in reality phenomenal stochastic machines. Dead end - no. A hammer is a dead end if you want to put in a screw properly, but if you have a nail.... World Models as a concept is a great idea but I invite anyone to define their own in a way that explains their own HGI (Human Generalised Intelligence).
Is a World Model possible? A model is a structured, simplified representation of reality — necessarily incomplete, yet practical when it includes only what’s known to matter. What counts as “significant” stems from observation, experience, and meaning. A model is valuable if it can broadly predict future behavior. Engineering applies such models while accounting for the unknown, adding safety margins and avoiding single points of failure. Each model belongs to a specific domain and depends on judgment. Once validated, it guides optimization and performance decisions. True AGI would require a universal model spanning physics to social behavior — an impossible task for the foreseeable future. As Stuart Kauffman’s fundamental work on the non-ergodicity of nature reminds us, reality cannot be fully captured. AI excels in defined domains, but pursuing a “genetal world model” is an unrealistic ambition.
So I guess humans and other animals don't exist.
But they're biological, not silicon. 😉🙂
Arguably, entirely different kettles of fish, different substrates permitting or restricting different edifices on top of them.
Complete non sequitur. The point was that world models are possible, and that a "universal model" isn't necessary.
Also, people like to say "arguably" when they actually have no argument.
P.S. I see from your feed that you are an anti-trans bigoted piece of trash who thinks the Democratic Party should be outlawed, and I didn't delve further into your right wing rantings ... not the sort of person I choose to engage with.
How does one construct a 'world' model? (I did some consulting in the Himalayan foothills, and we needed an interpreter for two villages on two hilltops which were only 8 miles apart, but spoke completely different languages.) I am tempted to say that there is no such thing as a world model.
And mostly I work with the idea of 'micro-anthropologies' at best. Sure there are commonalities, but 'world' models? My experience is that context is the basis for meaning and identity.
Gary touched on that in his Eisman interview. There are many different world models, cultures and realities; the trick is to know which one you're operating in.
Thank you! I agree that the term is poorly defined. For one thing, are we talking about a model of the "real world" or just a model that we can call "the world"? If the former, then I'd like to adduce the following working definition: (1) that world is real to me in which my actions have consequences; (2) that world is real to us in which our interactions are reciprocal.
Yes, on both counts. Thank you.
Are we framing the goal wrong.. we know that all the talk of AGI was hyped up from journalists and people like Sam Altman, because they wanted something sensational to sell papers or to raise money.
I do think we should celebrate how far the LLM approach has taken us though and how much value there is for us to apply to the world now.. Shouldn't we put AGI out of everyones minds for now and focus on this?
1. Agree about wrong goal - unattainable and idiotic
2. There is some value in the LLM approach - as long as you only treat it as offering drafts to be checked, for instance. But for me is outweighed by the damage its use for chatbots seems to be causing.
Thanks, Demis, for telling us something we already knew anyway!
And World Models means we have to understand what words mean - particularly the flood of new words describing a particular person's behaviour..
Sorry, but Neurosymbolics has nothing to offer when everything is illogical.
The only thing that works is Semantic AI - at least it can describe the chaos (see NYT columnists).
Neurosymbolic is just a mashup of two limited approaches. Better to get back to a serious thinker like Wittgenstein
Completely agree, but why can't a particular person see that? I put it down to naivety about real-world problems, and the raft of conditions that go with them.
This framing repeats a common category error.
Large Language Models are criticized for “lacking a world model,” as if understanding must precede language. But language itself is already a compressed interface to the world — not a substitute for it, but a historically evolved judgment layer built from collective interaction with reality.
What’s often missed is this:
LLMs don’t fail because they lack causality — they fail where *human judgment itself has not been made explicit*. Causality is not something humans possess as an internal simulator; it is something we stabilize through judgment under real constraints, error correction, and consequence.
Calling LLMs “mere pattern matchers” overlooks that human reasoning is also statistical, embodied, and historically conditioned — just embedded in biology instead of silicon.
World models matter. But dismissing language-based systems as a dead end assumes that understanding must look like physics engines in the head. That assumption itself is unexamined.
The real open question isn’t “language vs world models,” but whether we know how to externalize, stabilize, and test *judgment* — in machines or in ourselves.
Avid Xu: We test judgments with their backdrop of meaningful, well-understood, accumulated, hopefully redundant or even invariant, and tested evidence until we reach enough relevant intelligibility where no more serious questions arise, then we say YES, it is or NO it's not, or I don't know--let's go back to the drawing board, maybe even change the meaning of our questions. Catherine Blanche King (See B. Lonergan's Insight, A Study of Human Understanding, 2000).
ADDENDUM: See my own work, which draws from the above philosopher's work, on Academia.edu.
"Calling LLMs “mere pattern matchers” overlooks that human reasoning is also statistical, embodied, and historically conditioned — just embedded in biology instead of silicon."
This is oft repeated by LLMphiles but it is simply false ... that is not how human reasoning works, and it misses the fact that LLMs pattern match over tokens in their training data (which consists of human utterances, which is where all of the apparent "understanding" of LLMs comes from), not facts in the world.
Calling this "breaking" is a bit much. DeepMind have been "on Gary's side" since before OpenAI existed.
it’s the clearest statement i have seen from Demis that LLMs are likely a dead end and that world models are necessary. he’s been exploring them for a while (though the company was founded on a model-free approach to Atari games).
thank you for your insight. why natural system differ from tech and not compatible...but it has great potential for many utility. One is not expert on anything..perhaps useful following ....if not trash .one click away..
Time is not a separate parameter for it only indicates the interactive interval in an axiomatically dynamic oscillatory state. Therefore, as 1.344*10^-51 was derived on the bases of an accelerative transmigration from all six sides of a volumetric state,
(1.59E -18) *(6.283E +17) =1
Two important principles are covered in this equivalence. Volumetric change in interactions in the simultaneous state is equal to the change in the cyclic interval thereby proving that no stresses are unaccounted for. Or that the algebraic sum of the exchange of compressive and expansive stresses are zero.
A log one cycle reduction in potential count increases interactive counts by 10 units, which then releases a set of 2Pi*10^17*10=2PI*10^18 reciprocal of which 1.591681E-19 which is value of an electromagnetic charge of a volt in Physics
Planck’s constant h is considered the energy quantum, in actual fact 1.344*10^-51 is the true fundamental quantum in a completely unified field, for it determines the limiting mass, the 3.571428 state and the cyclic potential as of 18 orders all of which almost equal the Planck’s quantum. It underlines the error-potential in Physics when dealing at that energy level, for the Planck’s quantum is not a true elemental state. It suggest including human body to …all have this potential and is underlying mechanism must most important part of creating all new system…,models…etc…..….and has inherent limit of nature.
If they were, then:
1. Why hasn't Demiss said it this clearly before?
2. Why has a big chunk of DeepMind gone full on on Gemini?
1. he has hinted, but he was blown with the wind for some time.
2. DeepMind has bizz responsibility for Google... and they need to be relevant or better than their peers
As noted before, Nobel Laureate AI co-founder and symbolic AI Tower of Hanoi wizard Herb Simon predicted in 1960 that by 1980 computers could do any work a man could do. (Probably included women’s capabilities as well.) This might be enough to end the discussion, but is unlikely to. Cognitive world models are good for Tower of Hanoi, chess, some coding. Homo sapiens arrives with cognitive world models, but more emotional and social world models which we don’t understand well enough to emulate. Gary noted that bolting some symbolic onto neural won’t work. Cyc bolted some neural onto 25 years of symbolic work and didn’t set the world on fire. But Philip Wilkinson noted a key point: Simon and the other pioneers saw what we call AGI as the goal. Many commenters do as well. But some of us never have, our goal is to find something useful, not harmful, and economically feasible. There is no question that LLMs can be useful. But harmless and capable of finding a revenue model are not clear.
It suffices to look at where we are with autonomous driving (i.e. navigating a rather simple 2D environment based on a handful degrees of freedom) after more than a decade and billions in R&D to see that AI based on "world models" is just the next chapter in wishful thinking.
As usual Darwin will set you free! The only intelligence we know of evolved by a process of natural selection. There is a continuum of agentic capabilities from us all the way to bacteria. Intelligence is about predicting the future of the world for fun & profit. When the future world includes other intelligences, some of whom you can talk to, you quickly get into an arms race of competition between the quality of your predictive models. But there’s another catch that was pointed out by Donald Hoffman. Our (evolved) senses only give us fitness not truth about the world. We don’t even know how many dimensions of space there are or if time is fundamental or emergent. Evolution conveniently hides the details from us. Building AIs based on a human sense of the world is absurd. It’s akin to building a computer inside Minecraft: you can do it, but it is tremendously inefficient and cannot possibly scale. This is the main problem with LLMs - they are so abstracted from reality that they cannot possibly deal with it. Until AI takes a truly bottom-up and evolutionary approach, all of it is doomed to failure.
Likely a dumb question, but what is the difference in a nutshell between a world and neurosymbic model?