I don't understand this view. How I see it the fundamental bottleneck to AGI is continual learning and backpropagation. Models today are static, and human brains don't learn or adapt themselves with anything close to backpropagation. World models don't solve any of these problems; they are fundamentally the same kind of deep learning architectures we are used to work with. Heck, if you think learning from the world itself is the bottleneck, you can just put a vision-action LLM on a reinforcement learning loop in a robotic/simulated body.
> I don't understand this view. How I see it the fundamental bottleneck to AGI is continual learning and backpropagation. Models today are static, and human brains don't learn or adapt themselves with anything close to backpropagation.
Even with continuous backpropagation and "learning", enriching the training data, so called online-learning, the limitations will not disappear. The LLMs will not be able to conclude things about the world based on fact and deduction. They only consider what is likely from their training data. They will not foresee/anticipate events, that are unlikely or non-existent in their training data, but are bound to happen due to real world circumstances. They are not intelligent in that way.
Whether humans always apply that much effort to conclude these things is another question. The point is, that humans fundamentally are capable of doing that, while LLMs are structurally not.
The problems are structural/architectural. I think it will take another 2-3 major leaps in architectures, before these AI models reach human level general intelligence, if they ever reach it. So far they can "merely" often "fake it" when things are statistically common in their training data.
Humans are notoriously bad at formal logic. The Wason selection task is the classic example: most people fail a simple conditional reasoning problem unless it’s dressed up in familiar social context, like catching cheaters. That looks a lot more like pattern matching than rule application.
Kahneman’s whole framework points the same direction. Most of what people call “reasoning” is fast, associative, pattern-based. The slow, deliberate, step-by-step stuff is effortful and error-prone, and people avoid it when they can. And even when they do engage it, they’re often confabulating a logical-sounding justification for a conclusion they already reached by other means.
So maybe the honest answer is: the gap between what LLMs do and what most humans do most of the time might be smaller than people assume. The story that humans have access to some pure deductive engine and LLMs are just faking it with statistics might be flattering to humans more than it’s accurate.
Where I’d still flag a possible difference is something like adaptability. A person can learn a totally new formal system and start applying its rules, even if clumsily. Whether LLMs can genuinely do that outside their training distribution or just interpolate convincingly is still an open question. But then again, how often do humans actually reason outside their own “training distribution”? Most human insight happens within well-practiced domains.
> The Wason selection task is the classic example: most people fail a simple conditional reasoning problem unless it’s dressed up in familiar social context, like catching cheaters.
I've never heard about the Wason selection task, looked it up, and could tell the right answer right away. But I can also tell you why: because I have some familiarity with formal logic and can, in your words, pattern-match the gotcha that "if x then y" is distinct from "if not x then not y".
In contrast to you, this doesn't make me believe that people are bad at logic or don't really think. It tells me that people are unfamiliar with "gotcha" formalities introduced by logicians that don't match the everyday use of language. If you added a simple additional to the problem, such as "Note that in this context, 'if' only means that...", most people would almost certainly answer it correctly.
Mind you, I'm not arguing that human thinking is necessarily more profound from what what LLMs could ever do. However, judging from the output, LLMs have a tenuous grasp on reality, so I don't think that reductionist arguments along the lines of "humans are just as dumb" are fair. There's a difference that we don't really know how to overcome.
Quoting the Wikipedia article's formulation of the task for clarity:
> You are shown a set of four cards placed on a table, each of which has a number on one side and a color on the other. The visible faces of the cards show 3, 8, blue and red. Which card(s) must you turn over in order to test that if a card shows an even number on one face, then its opposite face is blue?
Confusion over the meaning of 'if' can only explain why people select the Blue card; it can't explain why people fail to select the Red card. If 'if' meant 'if and only if', then it would still be necessary to check that the Red card didn't have an even number. But according to Wason[0], "only a minority" of participants select (the study's equivalent of) the Red card.
[0] https://web.mit.edu/curhan/www/docs/Articles/biases/20_Quart...
People in everyday life are not evaluating rules. They evaluate cases, for whether a case fits a rule.
So, when being told:
"Which card(s) must you turn over in order to test that if a card shows an even number on one face, then its opposite face is blue?"
they translate it to:
"Check the cards that show an even number on one face to see whether their opposite face is blue and vice versa"
Based on this, many would naturally pick the blue card (to test the direct case), and the 8 card (to test the "vice versa" case).
They wont check the red to see if there's an odd number there that invalidates the formulation as a general rule, because they're not in the mindset of testing a general rule.
Would they do the same if they had more familiarity with rule validation in everyday life or if the had a more verbose and explicit explanation of the goal?
Yeah maybe if you phrased it as "Which card(s) must you turn over in order to ensure that all odd-numbered cards are blue?" you'd get a better response?
Exactly. We invented rule-based machines so that we could have a thing that follows rules, and adheres strictly to them, all day long.
Im not sure why people keep comparing machine-behaviour to human's. Its like Economic models that assume perfect rationality... yeah that's not reality mate.
I've confidently picked 8+blue and is now trying to understand why I personally did that. I think that maybe the text of the puzzle is not quite unambiguous. The question states "test a card" followed by "which cards", so this is what my brain immediately starts to check - every card one by one. Do I need to test "3"? No, not even. Do I need to test "8"? yes. Do I need to test "blue"? Yes, because I need to test "a card" to fit the criteria. And lastly "red" card also immediately fails verification of a "a card" fitting that criteria.
I think a corrected question should clarify in any obvious way that we are verifying not "a card" but "a rule" applicable to all cards. So a needs to be replaced with all or any, and mention of rule or pattern needs to be added.
It also doesn't explain why people don't think it necessary to check the 3 to make sure it's not blue (which it would be if "if" meant "if and only if").
I think we're actually closer to agreement than it might seem.
You're right that the Wason task is partly about a mismatch between how "if" works in formal logic and how it works in everyday language. That's a fair point. But I think it actually supports what I'm saying rather than undermining it. If people default to interpreting "if x then y" as "if and only if" based on how language normally works in conversation, that is pattern-matching from familiar context. It's a totally understandable thing to do, and I'm not calling it a cognitive defect. I'm saying it's evidence that our default mode is contextual pattern-matching, not rule application. We agree on the mechanism, we're just drawing different conclusions from it.
Your own experience is interesting too. You got the right answer because you have some background in formal logic. That's exactly what I'd expect. Someone who's practiced in a domain recognizes the pattern quickly. But that's the claim: most reasoning happens within well-practiced domains. Your success on the task doesn't counter the pattern-matching thesis, it's a clean example of it working well.
On the broader point about LLMs having a "tenuous grasp on reality," I hear that, and I don't want to flatten the differences. There probably is something meaningfully different going on with how humans stay grounded. I just think the "humans reason, LLMs pattern-match" framing undersells how much human cognition is also pattern-matching, and that being honest about that is more productive than treating it as a reductionist insult.
Agree with much of your comment.
Though note that as GP said, on the Wason selection task, people famously do much better when it's framed in a social context. That at least partially undermines your theory that its lack of familiarity with the terminology of formal logic.
Maybe the social version just creates a context where "if x then y" obviously does not include "if not x then not y". Everyone knows people over the drinking age can drink both alcoholic and non-alcoholic drinks, so you obviously don't have to check the person drinking the soft drink to make sure they aren't an adult.
I for the life of me could not solve the <18 example from wikipedia. but the number/color one is super easy
As they say, "think about how smart the average person is, then realize half the population is below that". There are far more haikus than opuses walking this planet.
We keep benchmarking models against the best humans and the best human institutions - then when someone points out that swarms, branching, or scale could close the gap, we dismiss it as "cheating". But that framing smuggles in an assumption that intelligence only counts if it works the way ours does. Nobody calls a calculator a cheat for not understanding multiplication - it just multiplies better than you, and that's what matters.
LLMs are a different shape of intelligence. Superhuman on some axes, subpar on others. The interesting question isn't "can they replicate every aspect of human cognition" - it's whether the axes they're strong on are sufficient to produce better than human outcomes in domains that matter. Calculators settled that question for arithmetic. LLMs are settling it for an increasingly wide range of cognitive work. The fact that neither can flip a burger is irrelevant.
Humans don't have a monopoly on intelligence. We just had a monopoly on generality and that moat is shrinking fast.
The "God of the gaps" theory is a theological and philosophical viewpoint where gaps in scientific knowledge are cited as evidence for the existence and direct intervention of a divine creator. It asserts that phenomena currently unexplained by science—such as the origin of life or consciousness—are caused by God.
We are doing inversion of God of gaps to "LLM of Gaps" where gaps in LLM capabilities are considered inherently negative and limiting
It is not actually the gaps in capability, and instead it arises from an understanding of how it works and an honest acknowledgement of how far it could go.
The question is not if these things are actually intelligent or not. The question is if these things will be useful without an endless supply of training data and continuous re-alignment using it..
And the questions "Are these things really intelligent" is just a proxy for that.
And we are interested in that question because that is necessary to justify the massive investment these things are getting now. It is quite easy to look at these things and conclude that it will continue to progress without any limit.
But that would be like looking at data compression at the time of its conception, and thinking that it is only a matter of time we can compress 100GB into 1KB..
We live in a time of scams that are obvious if you take a second look. If something that require much deeper scrutiny, then it is possible to generate a lot more larger bubble.
> and that moat is shrinking fast..
The point is that in reality it is not. It is just appearance. If you consider how these things work, then there is no justification of this conclusion.
I have said this elsewhere, but the problem of Hallucination itself along with the requirement of re-training, the smoking gun that these things are not intelligence in ways that would justify these massive investments.
> If you added a simple additional to the problem, such as "Note that in this context, 'if' only means that...", most people would almost certainly answer it correctly.
Agreed. More broadly, classical logic isn't the only logic out there. Many logics will differ on the meaning of implication if x then y. There's multiple ways for x to imply y, and those additional meanings do show up in natural language all the time, and we actually do have logical systems to describe them, they are just lesser known.
Mapping natural language into logic often requires a context that lies outside the words that were written or spoken. We need to represent into formulas what people actually meant, rather than just what they wrote. Indeed the same sentence can be sometimes ambiguous, and a logical formula never is.
As an aside, I wanna say that material implication (that is, the "if x then y" of classical logic) deeply sucks, or rather, an implication in natural language very rarely maps cleanly into material implication. Having an implication if x then y being vacuously true when x is false is something usually associated with people that smirk on clever wordplays, rather than something people actually mean when they say "if x then y"
Your response contains a performative contradiction: you are asserting that humans are naturally logical while simultaneously committing several logical errors to defend that claim.
This comment would be a lot more useful with an enumeration of those logical errors.
commenter’s specific claim—that adding a note about the definition of "if" would solve the problem—is a moving the goalposts fallacy and a tautology. The comment also suffers from hasty generalization (in their experience the test isn't hard) and special pleading (double standard for LLM and humans).
When someone tells you "you can have this if you pay me", they don't mean "you can also have it if you don't pay". They are implicitly but clearly indicating you gotta pay.
It's as simple as that. In common use, "if x then y" frequently implies "if not x then not y". Pretending that it's some sort of a cognitive defect to interpret it this way is silly.
In the original studies, most people made an error that can't be explained by that misunderstanding: they failed to select the card showing 'not y'.
From my armchair this feels relevant:
> Decoding analyses of neural activity further reveal significant above chance decoding accuracy for negated adjectives within 600 ms from adjective onset, suggesting that negation does not invert the representation of adjectives (i.e., “not bad” represented as “good”)[...]
From: Negation mitigates rather than inverts the neural representations of adjectives
At: https://journals.plos.org/plosbiology/article?id=10.1371/jou...
> But then again, how often do humans actually reason outside their own “training distribution”? Most human insight happens within well-practiced domains.
Humans can produce new concepts and then symbolize them for communication purposes. The meaning of concepts is grounded in operational definitions - in a manner that anyone can understand because they are operational, and can be reproduced in theory by anyone.
For example, euclid invented the concepts of a point, angle and line to operationally represent geometry in the real world. These concepts were never "there" to begin with. They were created from scratch to "build" a world-model that helps humans navigate the real world.
Euclid went outside his "training distribution" to invent point, angle, and line. Humans have this ability to construct new concepts by interaction with the real world - bringing the "unknown" into the "known" so-to-speak. Animals have this too via evolution, but it is unclear if animals can symbolize their concepts and skills to the extent that humans can.
> Humans can produce new concepts and then symbolize them for communication purposes.
Sure, but the question is how often this actually happens versus how often people are doing something closer to recombination and pattern-matching within familiar territory. The point was about the base rate of genuine novel reasoning in everyday human cognition, and I don't think this addresses that.
> Euclid invented the concepts of a point, angle and line to operationally represent geometry in the real world. These concepts were never "there" to begin with.
This isn't really true though. Egyptian and Babylonian surveyors were working with geometric concepts long before Euclid. What Euclid did was axiomatize and systematize knowledge that was already in wide practical use. That's a real achievement, but it's closer to "sophisticated refinement within a well-practiced domain" than to reasoning from scratch outside a training distribution. If anything the example supports the parent comment.
There's also something off about saying points and lines were "never there." Humans have spatial perception. Geometric intuitions come from embodied experience of edges, boundaries, trajectories. Formalizing those intuitions is real work, but it's not the same as generating something with no prior basis.
The deeper issue is you're pointing to one of the most extraordinary intellectual achievements in human history and treating it as representative of human cognition generally. The whole point, drawing on Kahneman, is that most of what we call reasoning is fast associative pattern-matching, and that the slow deliberate stuff is rarer and more error-prone than people assume. The fact that Euclid existed doesn't tell us much about what the other billions of humans are doing cognitively on a Tuesday afternoon.
> Formalizing those intuitions is real work, but it's not the same as generating something with no prior basis.
> The fact that Euclid existed doesn't tell us much about what the other billions of humans are doing cognitively on a Tuesday afternoon.
Birds can fly - so, there is some flying intelligence built into their dna. But, are they aware of their skill to be able to create a theory of flight, and then use that to build a plane ? I am just pointing out that intuitions are not enough - the awareness of the intuitions in a manner that can symbolize and operationalize it is important.
> The whole point, drawing on Kahneman, is that most of what we call reasoning is fast associative pattern-matching, and that the slow deliberate stuff is rarer and more error-prone than people assume
David Bessis, in his wonderful book [1] argues that the cognitive actions done by you and I on a tuesday afternoon is the same that mathematicians do - just that we are unaware of it. Also, since you brought up Kahneman, Bessis proposes a System 3 wherein inaccurate intuitions is corrected by precise communication.
[1] Mathematica: A Secret World of Intuition and Curiosity
The bird analogy is actually a really good one, but I think it supports a narrower claim than you're making. You're right that the capacity to symbolize and formalize intuitions is a distinct and important thing, separate from just having the intuitions. No argument there. But my point wasn't that symbolization doesn't matter. It was about how often humans actually exercise that capacity in a strong sense versus doing something more like recombination within familiar frameworks. The bird can't theorize flight, agreed. But most humans who can in principle theorize about their intuitions also don't, most of the time. The capacity exists. The base rate of its deployment is the question.
On Bessis, I actually think his argument is more compatible with what I was saying than it might seem. If the cognitive process underlying mathematical reasoning is the same one operating on a Tuesday afternoon, that's an argument against treating Euclid-level formalization as categorically different from everyday cognition. It suggests a continuum rather than a bright line between "pattern matching" and "genuine reasoning." Which is interesting and probably right. But it also means you can't point to Euclid as evidence that humans routinely do something qualitatively beyond what LLMs do. If Bessis is right, then the extraordinary cases and the mundane cases share the same underlying machinery, and the question becomes quantitative (how far along the continuum, how often, under what conditions) rather than categorical.
I'll check out the book though, it sounds like it's making a more careful version of the point than usually gets made in these threads.
> Kahneman’s whole framework points the same direction. Most of what people call “reasoning” is fast, associative, pattern-based. The slow, deliberate, step-by-step stuff is effortful and error-prone, and people avoid it when they can. And even when they do engage it, they’re often confabulating a logical-sounding justification for a conclusion they already reached by other means.
Some references on that
https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow
https://thedecisionlab.com/reference-guide/philosophy/system...
System 1 really looks like a LLM (indeed completing a phrase is an example of what it can do, like, "you either die a hero, or you live enough to become the _"). It's largely unconscious and runs all the time, pattern matching on random stuff
System 2 is something else and looks like a supervisor system, a higher level stuff that can be consciously directed through your own will
But the two systems run at the same time and reinforce each other
In my naive understanding, neither requires any will or consciousness.
S1 is “bare” language production, picking words or concepts to say or think by a fancy pattern prediction. There’s no reasoning at this level, just blabbering. However, language by itself weeds out too obvious nonsense purely statistically (some concepts are rarely in the same room), but we may call that “mindlessly” - that’s why even early LLMs produced semi-meaningful texts.
S2 is a set of patterns inside the language (“logic”), that biases S1 to produce reasoning-like phrases. Doesn’t require any consciousness or will, just concepts pushing S1 towards a special structure, simply backing one keeps them “in mind” and throws in the mix.
I suspect S2 has a spectrum of rigorousness, because one can just throw in some rules (like “if X then Y, not Y therefore not X”) or may do fancier stuff (imposing a larger structure to it all, like formulating and testing a null hypothesis). Either way it all falls down onto S1 for a ultimate decision-making, a sense of what sounds right (allowing us our favorite logical flaws), thus the fancier the rules (patterns of “thought”) the more likely reasoning will be sounder.
S2 doesn’t just rely but is a part of S1-as-language, though, because it’s a phenomena born out (and inside) the language.
Whether it’s willfully “consciously” engaged or if it works just because S1 predicts logical thinking concept as appropriate for certain lines of thinking and starts to involve probably doesn’t even matter - it mainly depends on whatever definition of “will” we would like to pick (there are many).
LLMs and humans can hypothetically do both just fine, but when it comes to checking, humans currently excel because (I suspect) they have a “wider” language in S1, that doesn’t only include word-concepts but also sensory concepts (like visuospatial thinking). Thus, as I get it, the world models idea.
I remember reading about this in a book, 'The enigma of reason', basically it was saying that reasoning was exactly that, we decided and then we came up with a reason for what we had decided and usually not the other way around.
This is because, the 'reasoning' part of our brain came from evolution when we started to communicate with others, we needed to explain our behaviour.
Which is fascinating if you think of the implications of that. In the most part we think we are being logical, but in reality we are pattern matching/impulsive and using our reasoning/logic to come up for excuses for why we have chosen what we had already decided.
It explains a lot about the world and why it's so hard to reason with someone, we are assuming the decision came from reason in the first place, which when you look at such peoples choices, makes sense as it's clear it didn't.
> The story that humans have access to some pure deductive engine and LLMs are just faking it with statistics might be flattering to humans more than it’s accurate.
Your point rings true with most human reasoning most of the time. Still, at least some humans do have the capability to run that deductive engine, and it seems to be a key part (though not the only part) of scientific and mathematical reasoning. Even informal experimentation and iteration rest on deductive feedback loops.
The fact that humans can learn to do X, sometimes well, often badly, and while many don’t, strongly supports the conjecture that X is not how they naturally do things.
I can perform symbolic calculations too. But most people have limited versions of this skill, and many people who don’t learn to think symbolically have full lives.
I think it is fair to say humans don’t naturally think in formal or symbolic reasoning terms.
People pattern match,
Another clue is humans have to practice things, become familiar with them to reason even somewhat reliable about them. Even if they already learned some formal reasoning.
—-
Higher level reasoning is always implemented as specific forms of lower order reasoning.
There is confusion about substrate processing vs. what higher order processes can be created with that substrate.
We can “just” be doing pattern matching from an implementation view, and yet go far “beyond” pattern matching with specific compositions of pattern matching, from a capability view.
How else could neurons think? We are “only” neurons. Yet we far surpass the kinds of capabilities neurons have.
I don't disagree with any of that. My comment was only in relation to the question of human-specific capability that current LLMs may not be able to duplicate. I was not making the value judgments you seem to have read.
When people do math or rigorous deductive reasoning, are we sure they aren't just pattern matching with a set of carefully chosen interacting patterns that have been refined by ancient philosophers as being useful patterns that produce consistent results when applied in correctly patterned ways?
I've often wondered this. I suspect not, though I don't know. You're right that the answer matters to understanding LLM limitations relative to humans, though.
Brilliant insight. The success of LLM reasoning, ie “telling yourself a story”, has greatly increased my belief that humans are actually much less impressive than they seem. I do think it’s mostly pattern matching and a bunch of interacting streams analogous to LLM tokens. Obviously the implementations are different, because nature has to be robust and learn online, but I do not think we are as different from these machines as most people assume. There’s a reason Hofstadter et al. reacted as they did even to the earlier models.
This is why I also think humans being logical inference machines is mostly not true. We are seemingly capable of it, but there must be some cost that keeps it from being commonly used.
While humans did seemingly evolve socially very fast, with the tools we seem to have had for a few hundred thousand years it could have been far faster if there were not some other limitations that are being applied.
> Even with continuous backpropagation and "learning"
That's what I said. Backpropagation cannot be enough; that's not how neurons work in the slightest. When you put biological neurons in a Pong environment they learn to play not through some kind of loss or reward function; they self-organize to avoid unpredictable stimulation. As far as I know, no architecture learns in such an unsupervised way.
https://www.sciencedirect.com/science/article/pii/S089662732...
Forgive me for being ignorant - but 'loss' in supervised learning ML context encode the difference between how unlikely (high loss) or likely (low loss) was the network in predicting the output based on the input.
This sounds very similar to me as to what neurons do (avoid unpredictable stimulation)
So, I have been thinking about this for a little while. Image a model f that takes a world x and makes a prediciton y. At a high-level, a traditional supervised model is trained like this
f(x)=y' => loss(y',y) => how good was my prediction? Train f through backprop with that error.
While a model trained with reinforcement learning is more similar to this. Where m(y) is the resulting world state of taking an action y the model predicted.
f(x)=y' => m(y')=z => reward(z) => how good was the state I was in based on my actions? Train f with an algorithm like REINFORCE with the reward, as the world m is a non-differentiable black-box.
While a group of neurons is more like predicting what is the resulting word state of taking my action, g(x,y), and trying to learn by both tuning g and the action taken f(x).
f(x)=y' => m(y')=z => g(x,y)=z' => loss(z,z') => how predictable was the results of my actions? Train g normally with backprop, and train f with an algorithm like REINFORCE with negative surprise as a reward.
After talking with GPT5.2 for a little while, it seems like Curiosity-driven Exploration by Self-supervised Prediction[1] might be an architecture similar to the one I described for neurons? But with the twist that f is rewarded by making the prediction error bigger (not smaller!) as a proxy of "curiosity".
[1] https://arxiv.org/pdf/1705.05363
So can't you just use how real neurons learn as training data to to learn how to learn the same way?
I think people MOSTLY foresee and anticipate events in OUR training data, which mostly comprises information collected by our senses.
Our training data is a lot more diverse than an LLMs. We also leverage our senses as a carrier for communicating abstract ideas using audio and visual channels that may or may not be grounded in reality. We have TV shows, video games, programming languages and all sorts of rich and interesting things we can engage with that do not reflect our fundamental reality.
Like LLMs, we can hallucinate while we sleep or we can delude ourselves with untethered ideas, but UNLIKE LLMs, we can steer our own learning corpus. We can train ourselves with our own untethered “hallucinations” or we can render them in art and share them with others so they can include it in their training corpus.
Our hallucinations are often just erroneous models of the world. When we render it into something that has aesthetic appeal, we might call it art.
If the hallucination helps us understand some aspect of something, we call it a conjecture or hypothesis.
We live in a rich world filled with rich training data. We don’t magically anticipate events not in our training data, but we’re also not void of creativity (“hallucinations”) either.
Most of us are stochastic parrots most of the time. We’ve only gotten this far because there are so many of us and we’ve been on this earth for many generations.
Most of us are dazzled and instinctively driven to mimic the ideas that a small minority of people “hallucinate”.
There is no shame in mimicking or being a stochastic parrot. These are critical features that helped our ancestors survive.
> We can steer our own learning corpus
This is critical. We have some degree of attentional autonomy. And we have a complex tapestry of algorithms running in thalamocortical circuits that generate “Nows”. Truncation commands produce sequences of acts (token-like products).
Models don't care. They aren't alive. This is the source of the chasm between here and AGI. You have to fear death to reason about the world and how to behave in it.
I guess I just always thought it was obvious that you can't do better than nature. You can do different things, sure, but if a society of unique individuals wasn't the most effective way of making progress, nature itself would not have chosen it.
So in a way I think Yan is smart because he got money, but in a way I think he's a fucking idiot if he can't see just how very, very very far we are from competing with organic intelligence.
Not only that but people like this aren't actually interested in understanding the physical world. Because we don't understand it yet. If you care about understanding the world I think you become someone more like Jane Goodall than Yan LeCun
"You have to fear death to reason about the world and how to behave in it."
You're onto something there.
If everyone knew they were to die tomorrow, all of a sudden they'd choose to act differently. There is no logical thought process that determines that - it's something else. Something we can't concretely point toward as an object.
> They will not foresee/anticipate events, that are unlikely or non-existent in their training data, but are bound to happen due to real world circumstances. They are not intelligent in that way.
Can you be a bit more specific at all bounds? Maybe via an example?
I'm sure that if a car appeared from nowhere in the middle of your living room, you would not be prepared at all.
So my question is: when is there enough training data that you can handle 99.99% of the world ?
The main difference is humans are learning all the time and models learn batch wise and forget whatever happened in a previous session unless someone makes it part of the training data so there is a massive lag.
Whoever cracks the continuous customized (per user, for instance) learning problem without just extending the context window is going to be making a big splash. And I don't mean cheats and shortcuts, I mean actually tuning the model based on received feedback.
They can write to files then refer to them in a next session.
A bit like the main character played by Guy Pierce in the movie Memento (which doesn't work great for him to be honest).
Why not just provide more compute for say, 1 billion token context for each user to mimic continuous learning. Then retrain the model in the background to include learnings.
The user wouldn’t know if the continuous learning came from the context or the model retrained. It wouldn’t matter.
Continuous learning seems to be a compute and engineering problem.
Because that re-training is not strong enough to hold, or so it seems. The same dumb factual errors keep coming up on different generations of the same models. I've yet to see proof that something 'stuck' from model to model. They get better in a general sense but not in the specific sense that what was corrected stays put, not from session to session and not from one generation to the next.
My solution is to have this massive 'boot up' prompt but it becomes extremely tedious to maintain.
> Models today are static, and human brains don't learn or adapt themselves with anything close to backpropagation.
While I suspect latter is a real problem (because all mammal brains* are much more example-efficient than all ML), the former is more about productisation than a fundamental thing: the models can be continuously updated already, but that makes it hard to deal with regressions. You kinda want an artefact with a version stamp that doesn't change itself before you release the update, especially as this isn't like normal software where specific features can be toggled on or off in isolation of everything else.
* I think. Also, I'm saying "mammal" because of an absence of evidence (to my *totally amateur* skill level) not evidence of absence.
they can be continuously updated, assuming you re-run representative samples of the training set through them continuously. Unlike a mammal brain which preserves the function of neurons unless they activate in a situation which causes a training signal, deep nets have catastrophic forgetting because signals get scattered everywhere. If you had a model continuously learning about you in your pocket, without tons of cycles spent "remembering" old examples. In fact, this is a major stumbling block in standard training, sampling is a huge problem. If you just iterate through the training corpus, you'll have forgotten most of the english stuff by the time you finish with chinese or spanish. You have to constantly mix and balance training info due to this limitation.
The fundamental difference is that physical neurons have a discrete on/off activation, while digital "neurons" in a network are merely continuous differentiable operations. They also don't have a notion of "spike timining dependency" to avoid overwriting activations that weren't related to an outcome. There are things like reward-decay over time, but this applies to the signal at a very coarse level, updates are still scattered to almost the entire system with every training example.
You could have continual learning on text and still be stuck in the same "remixing baseline human communications" trap. It's a nasty one, very hard to avoid, possibly even structurally unavoidable.
As for the "just put a vision LLM in a robot body" suggestion: People are trying this (e.g. Physical Intelligence) and it looks like it's extraordinarily hard! The results so far suggest that bolting perception and embodiment onto a language-model core doesn't produce any kind of causal understanding. The architecture behind the integration of sensory streams, persistent object representations, and modeling time and causality is critically important... and that's where world models come in.
yes those are bottlenecks that world models don't solve. but the promise of world models is, unlike LLMs, they might be able to learn things about the world that humans haven't written. For example, we still don't fully know how insects fly. A world model could be trained on thousands of videos of insects and make a novel observation about insect trajectories. The premise is that despite being here for millenia, humans have only observed a tiny fraction of the world.
So I do buy his idea. But I disagree that you need world models to get to human level capabilities. IMO there's no fundamental reason why models can't develop human understanding based on the known human observations.
The fact that models aren't continually updating seems more like a feature. I want to know the model is exactly the same as it was the last time I used it. Any new information it needs can be stored in its context window or stored in a file to read the next it needs to access it.
> The fact that models aren't continually updating seems more like a feature.
I think this is true to some extent: we like our tools to be predictable. But we’ve already made one jump by going from deterministic programs to stochastic models. I am sure the moment a self-evolutive AI shows up that clears the "useful enough" threshold we’ll make that jump as well.
Stochastic and unpredictability aren't exactly the same. I would claim current LLMs are generally predictable even if it is not as predictable as a deterministic program.
No, but my point is that to some extent we value determinism. By making the jump to stochastic models we already move away from the status quo; further jumps are entirely possible. Depending on use case we can accept more uncertainty if it comes with benefits.
I also don’t think there is a reason to believe that self-learning models must be unpredictable.
Persistent memory through text in the context window is a hack/workaround.
And generally:
> I want to know the model is exactly the same as it was the last time I used it.
What exactly does that gain you, when the overall behavior is still stochastic?
But still, if it's important to you, you can get the same behavior by taking a model snapshot once we crack continuous learning.
It’s a feature of a good tool, but a sentient intelligence is more than just a tool
Unless you use your oen local models then you don't even know when OpenAI or Anthropic tweaked the model less or more. One week it's a version x, next week it's a version y. Just like your operating system is continuously evolving with smaller patches of specific apps to whole new kernel version and new OS release.
There is still a huge gap between a model continuously updating itself and weekly patches by a specialist team. The former would make things unpredictable.
Especially they will require even more compute to get anything close to usable output. Human brains are super efficient at learning and producing output. We will need exponentially more compute for real time learning from video + audio + haptic data.
I never understood why we believe humans don't backprop. Isn't it that during the day we fill up our context (short term memory) and sleep is actually where we use that to backprop? Heck, everyone knows what "sleep on it" means.
Brains are not doing linear algebra, and they don't follow a concise algorithm.
What LLM do is even farther away from what neural nets do, and even there - artificial neurons are inspired by but not reimplementing biological neurons.
You can understand human thought in terms of LLMs, but that is just a simile, like understanding physical reality in terms of computers or clockworks.
It's pretty simple... the word circle and what you can correlate to it via english language description has somewhat less to do with reality than a physical 3D model of a circle and what it would do in an environment. You can't just add more linguistic description via training data to change that. It doesn't really matter that you can keep back propagating because what you are back propagating over is fundamentally and qualitatively less rich.
LeCun is a researcher.
From his point of view, there are not much research left on LLM. Sure we can still improve them a bit with engineering around, but he's more interested in basic research.
I don't understand why online learning is that necessary. If you took Einstein at 40 and surgically removed his hippocampus so he can't learn anything he didn't already know (meaning no online learning), that's still a very useful AGI. A hippocampus is a nice upgrade to that, but not super obviously on the critical path.
> If you took Einstein at 40 and surgically removed his hippocampus so he can't learn anything he didn't already know (meaning no online learning), that's still a very useful AGI.
I like how people are accepting this dubious assertion that Einstein would be "useful" if you surgically removed his hippocampus and engaging with this.
It also calls this Einstein an AGI rather than a disabled human???
Hypotheticals fear him
He basically said that himself:
"Reading, after a certain age, diverts the mind too much from its creative pursuits. Any man who reads too much and uses his own brain too little falls into lazy habits of thinking".
-- Albert Einstein
I guess the sheer amount and also variety of information you would need to pre-encode to get an Einstein at 40 is huge. Every day stream of high resolution video feed and actions and consequences and thoughts and ideas he has had until the age of 40 of every single moment. That includes social interactions, like a conversation and mimic of the other person in combination with what was said and background knowledge about the other person. Even a single conversation's data is a huge amount of data.
But one might say that the brain is not lossless ... True, good point. But in what way is it lossy? Can that be simulated well enough to learn an Einstein? What gives events significance is very subjective.
Kinda a moot point in my eyes because I very much doubt you can arrive at the same result without the same learning process.
That's true. Though could that hippocampus-less Einstein be able to keep making novel complex discoveries from that point forward? Seems difficult. He would rapidly reach the limits of his short term memory (the same way current models rapidly reach the limits of their context windows).
It could possibly be useful but I don't see why it would be AGI.
Where does that training data come from?
If your model is poor, no amount of learning can fix it. If you don't think your model architecture is limited, you aren't looking hard enough.
I don’t understand your view. Reality is that we need some way to encode the rules of the world in a more definitive way. If we want models to be able to make assertive claims about important information and be correct, it’s very fair to theorize they might need a more deterministic approach than just training them more. But it’s just a theory that this will actually solve the problem.
Ultimately, we still have a lot to learn and a lot of experiments to do. It’s frankly unscientific to suggest any approaches are off the table, unless the data & research truly proves that. Why shouldn’t we take this awesome LLM technology and bring in more techniques to make it better?
A really, really basic example is chess. Current top AI models still don’t know how to play it (https://www.software7.com/blog/ai_chess_vs_1983_atari/) The models are surely trained on source material that include chess rules, and even high level chess games. But the models are not learning how to play chess correctly. They don’t have a model to understand how chess actually works — they only have a non-deterministic prediction based on what they’ve seen, even after being trained on more data than any chess novice has ever seen about the topic. And this is probably one of the easiest things for AI to stimulate. Very clear/brief rules, small problem space, no hidden information, but it can’t handle the massive decision space because its prediction isn’t based on the actual rules, but just “things that look similar”
(And yeah, I’m sure someone could build a specific LLM or agent system that can handle chess, but the point is that the powerful general purpose models can’t do it out of the box after training.)
Maybe more training & self-learning can solve this, but it’s clearly still unsolved. So we should definitely be experimenting with more techniques.
> Reality is that we need some way to encode the rules of the world in a more definitive way
I mean, sure. But do world models the way LeCun proposes them solves this? I don't think so. JEPAs are just an unsupervised machine learning model at the end of the day; they might end up being better that just autoregressive pretraining on text+images+video, but they are not magic. For example, if you train a JEPA model on data of orbital mechanics, will it learn actually sensible algorithms to predict the planets' motions or will it just learn a mix of heuristic?
Iirc LeCunn talks about a self organizing hierarchy of real world objects and imo this is exactly how the human brain actually learns
Who knows? Perhaps attention really is all you need. Maybe our context window is really large. Or our compression is really effective. Perhaps adding external factors might be able to indirectly teach the models to act more in line with social expectations such as being embarrassed to repeat the same mistake, unlocking the final piece of the puzzle. We are still stumbling in the dark for answers.
The reason LLMs fail today is because there’s no meaning inherent to the tokens they produce other than the one captured by cooccurrence within text. Efforts like these are necessary because so much of “general intelligence” is convention defined by embodied human experience, for example arrows implying directionality and even directionality itself.
Agents have the ability of continual learning.
Putting stuff you have learned into a markdown file is a very "shallow" version of continual learning. It can remember facts, yes, but I doubt a model can master new out-of-distribution tasks this way. If anything, I think that Google's Titans[1] and Hope[2] architectures are more aligned with true continual learning (without being actual continual learning still, which is why they call it "test-time memorization").
[1] https://arxiv.org/pdf/2501.00663
[2] https://arxiv.org/pdf/2512.24695
It really reminds me of the movie Memento - it has to constantly put notes down to remember who it is and what it should do after waking up without memory every n minutes.
I have had it master tasks by doing this. The first time it tries to solve an issue it may take a long time, but it documents its findings and how it was able to do it and then it applies that knowledge the next time the task comes up.
There is some things that just don't transfer really well without specific training. I tried to create diagrams in Typst with Cetz (a Processing and Tikz inspired graphing library), and even with documentation, GPT 5.2-thinking can't really do complex nice diagrams like it can in Tikz. It can do simple things that are similar to the shown examples, but nothing really interesting. Typst and specially Cetz is too new for any current model to really "get it", so they can't use it. I need to wait to the next batch of frontier models so that they learn Typst and Cetz examples during pre-training.