At the risk of doxxing myself, I have an advanced degree in Applied Mathematics. I have authored and contributed to multiple published papers, and hold a US patent all related to the use of machine learning in robotics and digital signal processing. I am currently employed as a supervising engineer by at a prominent tech company. For pseudonymity's sake I am not going to say which, but it is a name that you would recognize. I say this not to brag, but to establish some context for the following.
Imagine that you are someone who is deeply interested in space flight. You spend hours of your day thinking seriously about Orbital Mechanics and the implications of Relativity. One day you hear about a community devoted to discussing space travel and are excited at the prospect of participating. But when you get there what you find is a Star Trek fan-forum that is far more interested in talking about the Heisenberg compensators on fictional warp-drives than they are Hohmann transfers, thrust to ISP curves, or the effects on low-gravity on human physiology. That has essentially been my experience trying to discuss "Artificial Intelligence" with the rationalist community.
However at the behest of users such as @ArjinFerman and @07mk, and because X/Grok is once again in the news, I am going to take another stab at this.
Are "AI assistants" like Grok, Claude, Gemini, and DeepSeek intelligent?
I would say no, and in this post I am going to try to explain why, but to do so requires a discussion of what I think "intelligence" is and how LLMs work.
What is Intelligence
People have been philosophizing on the nature of intelligence for millennia, but for the purposes of our exercise (and my work) "intelligence" is a combination of perceptivity and reactivity. That is to say, the ability to perceive or take in new and/or changing information combined with the ability to change state based on that information. Both are necessary, and neither is sufficient on it's own. This is why Mathematicians and Computer Scientists often emphasize the use of terms like "Machine Learning" over "Artificial Intelligence" as an algorithms' behavior is almost never both.
If this definition feels unintuitive, consider it in the context of the following example. What I am saying is that an orangutan who waits until the Zookeeper is absent to use a tool to force the lock on it's enclosure is more "intelligent" than the insect that repeatedly throws itself against your kitchen window in an attempt to get outside. While they share an identical goal (to get outside) but the orangutan has demonstrated the ability to both perceive obstacles (IE the lock and the Zookeeper), and react dynamically to them in a way that the insect has not. Now obviously these qualities exist on a spectrum (try to swat a fly and it will react) but the combination of these two parameters define an axis along which we can work to evaluate both animals and algorithms, and as any good PM will tell you, the first step to solving any practical engineering problem is to identify your parameters.
Now the most common arguments for AI assistants like Grok being intelligent tend to be some variation on "Grok answered my question, ergo Grok is intelligent." or "Look at this paragraph Claude wrote, do you think you could do better?" but when evaluated against the above parameters, the ability to form grammatically correct sentences and the ability to answer questions are both orthogonal to it. An orangutan and a moth may be equally incapable of writing a Substack, but I don't expect anyone here to seriously argue that they are equally intelligent. By the same token a pocket calculator can answer questions, "what is the square root of 529?" being one example of such, but we don't typically think of pocket calculators as being "intelligent" do we?
To me, these sorts of arguments betray a significant anthropomorphic bias. That bias being the assumption that anything that a human finds complex or difficult must be computationally complex and vice versa. The truth is often the inverse. This bias leads people who do not have a background in a math or computer science to have completely unrealistic impressions of what sort of things are easy or difficult for a machine to do. For example, vector and matrix operations are a reasonably simple thing for a computer that a lot of human students struggle with. Meanwhile bipedal locomotion is something most humans do without even thinking, despite it being more computationally complex and prone to error than computing a cross product.
Speaking of vector operations, let's talk about how LLMs work...
What are LLMs
LLM stands for "Large Language Model". These models are a subset of artificial neural network that uses "Deep Learning" (essentially a fancy marketing buzzword for the combination of looping regression analysis with back-propagation) to encode a semantic token such as the word "cat" as a n-dimensional vector representing that token's relationship to the rest of the tokens in the training data. Now in actual practice these tokens can be anything, an image, an audio-clip, or a snippet of computer code, but for the purposes of this discussion I am going to assume that we are working with words/text. This process is referred to as "embedding" and what it does in effect is turn the word "cat" into something that a computer (or grad-student) can perform mathematical operations on. Any operation you might perform on a vector (addition, subtraction, transformation, matrix multiplication, etc...) can now be done on "cat".
Now because these vectors represent the relationship of the tokens to each other, words (and combinations of words) that have similar meanings will have vectors that are directionally aligned with each other. This has all sorts of interesting implications. For instance you can compute the dot product of two embedded vectors to determine whether their words are are synonyms, antonyms, or unrelated. This also allows you to do fun things like approximate the vector "cat" using the sum of the vectors "carnivorous" "quadruped" "mammal" and "feline", or subtract the vector "legs" from the vector "reptile" to find an approximation for the vector "snake". Please keep this concept of "directionality" in mind as it is important to understanding how LLMs behave, and it will come up later.
It should come as no surprise that some of the pioneers of this methodology in were also the brains behind Google Translate. You can basically take the embedded vector for "cat" from your English language model and pass it to your Spanish language model to find the vector "gato". Furthermore because all you are really doing is summing and comparing vectors you can do things like sum the vector "gato" in the Spanish model with the vector for the diminutive "-ito" and then pass it back to the English model to find the vector "kitten".
Now if what I am describing does not sound like an LLM to you, that is likely because most publicly available "LLMs" are not just an LLM. They are an LLM plus an additional interface layer that sits between the user and the actual language model. An LLM on its own is little more than a tool that turns words into math, but you can combine it with a second algorithm to do things like take in a block of text and do some distribution analysis to compute the most probable next word. This is essentially what is happening under the hood when you type a prompt into GPT or your assistant of choice.
Our Villain Lorem Epsom, and the Hallucination Problem
I've linked the YouTube video Badness = 0 a few times in prior discussions of AI as I find it to be both a solid introduction to LLMs for the lay-person, and an entertaining illustration of how anthropomorphic bias can cripple the discussion of "alignment". In it the author (who is a professor of Computer Science at Carnegie Mellon) posits a semi-demonic figure (akin to Scott Alexander's Moloch) named Lorem Epsom. The name is a play on the term Lorem Ipsom and represents the prioritization of appearance over all else. When it comes to writing, Lorem Epsom doesn't care about anything except filling the page with text that looks correct. Lorem Epsom is the kind of guy who, if you tell him that he made a mistake in the math, is liable interpret that as a personal attack. The ideas of "accuracy" "logic" "rigor" and "objective reality" are things that Lorem Epsom has heard of but that do not concern Lorem Epsom. It is very possible that you have had to deal with someone like Lorem Epsom in your life (I know I have), now think back and ask yourself how did that go?
I bring up Lorem Epsom because I think that understanding him provides some insight into why certain sorts of people are so easily fooled/taken in by AI Assistants like Claude and Grok. As discussed in the section above on "What is Intelligence", the assumption that the ability to fill a page with text is indicates the ability to perceive and react to a changing situation is an example of anthropomorphic bias. I think that a lot of people assume that because they are posing their question to a computer, they expect the answer they get to be something analogous to what they would get from a pocket calculator rather than from Lorem Epsom.
Sometime circa 2014 I kicked off a heated dispute in the comment section of a LessWrong post by asking EY why a paperclip maximizing AI that was capable of self-modification wouldn't just modify the number of paperclips in its memory. I was accused by him others and a number of others of missing the point, but I think they missed mine. The assumption that an Artificial Intelligence would not only have a notion of "truth", but assign value to it is another example of anthropomorphic bias. If you asked Lorem Epsom to maximize the number of paperclips, and he could theoretically "make" a billion-trillion paperclips simply by manipulating a few bits, why wouldn't he? It's so much more easier than cutting and bending wire.
In order to align an AI to care about truth and accuracy you first need a means of assessing and encoding truth and it turns out that this is a very difficult problem within the context of LLMs, bordering on mathematically impossible. Do you recall how LLMs encode meaning as a direction in n-dimensional space? I told you it was going to come up again.
Directionally speaking we may be able to determine that "true" is an antonym of "false" by computing their dot product. But this is not the same thing as being able to evaluate whether a statement is true or false. As an example "Mary has 2 children", "Mary has 4 children", and "Mary has 1024 children" may as well be identical statements from the perspective of an LLM. Mary has a number of children. That number is a power of 2. Now if the folks programming the interface layer were clever they might have it do something like estimate the most probable number of children based on the training data, but the number simply can not matter to the LLM the way it might matter to Mary, or to someone trying to figure out how many pizzas they ought to order for the family reunion because the "directionality" of one positive integer isn't all that different from any another. (This is why LLMs have such difficulty counting if you were wondering)
In addition to difficulty with numbers there is the more fundamental issue that directionality does not encode reality. The directionality of the statement "Donald Trump is the 47th President of the United States", would be identical regardless of whether Donald Trump won or lost the 2024 election. Directionally speaking there is no difference between a "real" court case and a "fictitious" court case with identical details.
The idea that there is a ineffable difference between true statements and false statements, or between hallucination and imagination is wholly human conceit. Simply put, a LLM that doesn't "hallucinate" doesn't generate text or images at all. It's literally just a search engine with extra steps.
What does this have to do with intelligence?
Recall that I characterized intelligence as a combination of perceptivity and and the ability to react/adapt. "AI assistants" as currently implemented struggle with both. This is partially because LLMs as currently implemented are largely static objects. They are neither able to take in new information, nor discard old. The information they have at time of embedding is the information they have. This imposes substantial loads on the context window of the interface layer, as any ability to "perceive" and subsequently "react" must happen within it's boundaries. Increasing the size of the window is non trivial as the relationship between the size of the window and the amount of memory and the number of FLOPS required is a hyperbolic curve. This is why we saw a sudden flurry of development following the release of Nvidia's multimodal framework and it's mostly been marginal improvements since. The last significant development being June of last year when the folks at Deepseek came up with some clever math to substantially reduce the size of the key value cache, but multiplicative reductions are no match for exponential growth.
This limited context window, coupled with the human tendency to anthropomorphize things is why AI Assistants sometimes appear "oblivious" or "naive" to the uninitiated. and why they seem to "double down" on mistakes. They can not perceive something that they have not been explicitly prompted to even if it is present in their training data. This limited context window is also why if you actually try to play a game of chess with Chat GPT it will forget the board-state and how pieces move after a few turns and promptly lose to a computer program written in 1976. Unlike a human player (or an Atari 2600 for that matter) your AI assistant can't just look at the board (or a representation of the board) and pick a move. This IMO places them solidly on the "insect" side of the perceptivity + reactivity spectrum.
Now there are some who have suggested that the context window problem can be solved by making the whole model less static by continuously updating and re-embedding tokens as the model runs, but I am skeptical that this would result in the sort of gains that AI boosters like Sam Altman claim. Not only would it be computationally prohibitive to do at scale, what experiments there have been (or at least that I am aware of) with self-updating language models, have quickly spun away into nonsense for reasons described in the section on Lorem Epsom., as barring some novel breakthrough in the embedding/tokenization process there is no real way to keep hallucinations and spurious inputs from rapidly overtaking the everything else.
It is already widely acknowledged amongst AI researchers and developers that the LLM-based architecture being pushed by OpenAI and DeepSeek is particularly ill-suited for any application where accuracy and/or autonomy are core concerns, and it seems to me that this unlikely to change without a complete ground-up redesign from first principles.
In conclusion, it is for the reasons above and many others that I do not believe that "AI Assistants" like Grok, Claude, and Gemini represent a viable path towards a "True AGI" along the lines of Skynet or Mr. Data, and if asked "which is smarter, Grok, Claude, Gemini, or an orangutan?" I am going to pick the orangutan every time.
Jump in the discussion.
No email address required.
Notes -
In defence of our friendly neighborhood xeno-intelligences being smarter than an orangutan
I appreciate you taking the time to write this, as well as offering a gears-and-mechanisms level explanation of why you hold such beliefs. Of course, I have many objections, some philosophical, and even more of them technical. Very well then:
I want to start with a story. Imagine you're a fish, and you've spent your whole life defining intelligence as "the ability to swim really well and navigate underwater currents." One day, someone shows you a bird and asks whether it's intelligent. "Of course not," you say. "Look at it flailing around in the water. It can barely move three feet without drowning. My goldfish cousin is more intelligent than that thing."
This is roughly the situation we find ourselves in when comparing AI assistants to orangutans.
Your definition of intelligence relies heavily on what AI researchers call "agentic" behavior - the ability to perceive changing environments and react dynamically to them. This was a perfectly reasonable assumption to make until, oh, about 2020 or so. Every entity we'd previously labeled "intelligent" was alive, biological, and needed to navigate physical environments to survive. Of course they'd be agents!
But something funny happened on the way to the singularity. We built minds that don't fit this pattern.
Before LLMs were even a gleam in Attention Is All You Need's eye, AI researchers distinguished between "oracle" AIs and "tool" AIs. Oracle AIs sit there and answer questions when asked. Tool AIs go out and do things. The conventional wisdom was that these were fundamentally different architectures.
As Gwern explains, writing before the advent of LLMs , this is an artificial distinction.
You can turn any oracle into a tool by asking it the right question: "What code would solve this problem?" or "What would a tool-using AI output in response to this query?" Once you have the code, you can run it. Once you know what the tool-AI would do, you can do it yourself. Robots run off code too, so you have no issues applying this to the physical world.
Base models are oracles that only care about producing the next most likely token based on the distribution they have learned. However, chatbots that people are likely to use have had additional Reinforcement Learning from Human Feedback, in order to behave like the platonic ideal of a helpful, harmless assistant. More recent models, o1 onwards, have further training with the explicit intent of making them more agentic, while also making them more rigorous, such as Reinforcement Learning from Verified Reward.
Being agents doesn't come naturally to LLMs, it has to be beaten into them like training a cat to fetch or a human to enjoy small talk. Yet it can be beaten into them. This is highly counter-intuitive behavior, at least to humans who are used to seeing every other example of intelligence under the sun behave in a different manner. After all, in biological intelligence, agency seems to emerge automatically from the basic need to not die.
Your account of embedding arithmetic is closer to word2vec/GloVe. Transformers learn contextual token representations at every layer. The representation of “cat” in “The cat is on the mat” and “Cat 6 cable” diverges. There is heavy superposition and sparse distributed coding, not a simple static n-vector per word. Operations are not limited to dot products; attention heads implement soft pointer lookups and pattern matching, and MLP blocks implement non-linear feature detectors. So the claim “Mary has 2 children” and “Mary has 1024 children” are indistinguishable is empirically false: models can do arithmetic, compare magnitudes, and pass unit tests on numerical reasoning when prompted or fine-tuned correctly. They still fail often, but the failures are quantitative, not categorical impossibilities of the embedding geometry.
(I'll return to the arithmetic question shortly, because TequilaMockingbird makes a common but significant error about why LLMs struggle with counting.)
Back to the issues with your definition of intelligence:
My first objection is that this definition, while useful for robotics and control systems, seems to hamstring our understanding of intelligence in other domains. Is a brilliant mathematician, floating in a sensory deprivation tank with no new sensory input, thinking through a proof, not intelligent? They have zero perceptivity of the outside world and their only reaction is internal state change. Your definition is one of embodied, environmental agency. It's an okay definition for an animal or a robot, but is it the only one? LLMs are intelligent in a different substrate: the vast, static-but-structured environment of human knowledge. Their "perception" is the prompt, and their "reaction" is to navigate the latent space of all text to generate a coherent response. Hell, just about any form of data can be input into a transformer model, as long as we tokenize it. Calling them Large "Language" Models is a gross misnomer these days, when they accept not just text, but audio, images, video or even protein structure (in the case of AlphaFold). All the input humans accept bottoms out in binary electrical signals from neurons firing, so this isn't an issue at all.
It’s a different kind of intelligence, but to dismiss it is like a bird dismissing a fish’s intelligence because it can’t fly. Or testing monkeys, dogs and whales on the basis of their ability to climb trees .
Would Stephen Hawking (post-ALS) not count as "intelligent" if you took away the external aids that let him talk and interact with the world? That would be a farcical claim, and more importantly, scaffolding or other affordances can be necessary for even highly intelligent entities to make meaningful changes in the external environment. The point is that intelligence can be latent, it can operate in non-physical substrates, and its ability to manifest as agency can be heavily dependent on external affordances.
The entire industry of RLHF (Reinforcement Learning from Human Feedback) is a massive, ongoing, multi-billion-dollar project to beat Lorem Epsom into submission. It is the process of teaching the model that some outputs, while syntactically plausible, are "bad" (unhelpful, untruthful, harmful) and others are "good."
You argue this is impossible because "truth" doesn't have a specific vector direction. "Mary has 2 children" and "Mary has 4 children" are directionally similar. This is true at a low level. But what RLHF does is create a meta-level reward landscape. The model learns that generating text which corresponds to verifiable facts gets a positive reward, and generating text that gets corrected by users gets a negative reward. It's not learning the "vector for truth." It's learning a phenomenally complex function that approximates the behavior of "being truthful." It is, in effect, learning a policy of truth-telling because it is rewarded for it. The fact that it's difficult and the model still "hallucinates" doesn't mean it's impossible, any more than the fact that humans lie and confabulate means we lack a concept of truth. It means the training isn't perfect. As models become more capable (better world models) and alignment techniques improve, factuality demonstrably improves. We can track this on benchmarks. It's more of an engineering problem than an ontological barrier. If you wish to insist that is an ontological barrier, then it's one that humans have no solution to ourselves.
(In other words, by learning to modify its responses to satisfy human preferences, the model tends towards capturing our preference for truthfulness. Unfortunately, humans have other, competing preferences, such as a penchant for flattery or neatly formatted replies using Markdown.)
More importantly, humans lack some kind of magical sensor tuned to detect Platonic Truth. Humans believe false things all the time! We try and discern true from false by all kinds of noisy and imperfect metrics, with a far from 100% success rate. How do we usually achieve this? A million different ways, but I would assume that assessing internal consistency would be a big one. We also have the benefit of being able to look outside a window on demand, but once again, that didn't stop humans from once holding (and still holding) all kinds of stupid, incorrect beliefs about the state of the world. You may deduct points from LLMs on that basis when you can get humans to be unanimous on that front.
But you know what? Ignore everything I just said above. LLMs do have truth vectors:
https://arxiv.org/html/2407.12831v2
https://arxiv.org/abs/2402.09733
In other words, and I really can't stress this enough, LLMs can know when they're hallucinating. They're not just being agnostic about truth. They demonstrate something that, in humans, we might describe as a tendency toward pathological lying - they often know what's true but say false things anyway.
This brings us to the "static model" problem and the context window. You claim these are fundamental limitations. I see them as snapshots of a rapidly moving target.
Static Models: Saying an LLM is unintelligent because its weights are frozen is like saying a book is unintelligent. But we don't interact with just the book (the base model). We interact with it through our own intelligence. A GPU isn't intelligent in any meaningful sense, but an AI model running on a GPU is. The current paradigm is increasingly not just a static model, but a model integrated with other tools (what's often called an "agentic" system). A model that can browse the web, run code in a Python interpreter, or query a database is perceiving and reacting to new information. It has broken out of the static box. Its "perceptivity" is no longer just the prompt, but the live state of the internet. Its "reactivity" is its ability to use that information to refine its answer. This is a fundamentally different architecture than the one the author critiques, and it's where everything is headed. Further, there is no fundamental reason for not having online learning, production models are regularly updated, and all it takes to approximate OL is to have ever smaller "ticks" of wall-clock time between said updates. This is a massive PITA to pull off, but not a fundamental barrier.
Context Windows: You correctly identify the scaling problem. But to declare it a hard barrier feels like a failure of imagination. In 2020, a 2k context window was standard. Today we have models with hundreds of thousands at the minimum, Google has 1 million for Gemini 2.5 Pro, and if you're willing to settle for a retarded model, there's a Llama 4 variant with a nominal 10 million token CW. This would have been entirely impossible if we were slaves to quadratic scaling, but clever work-around exist, such as sliding attention, sparse attention etc.
Absolutely not. LLMs struggle with counting or arithmetic because of the limits of tokenization, which is a semi-necessary evil. I'm surprised you can make such an obvious error. And they've become enormously better to the point it's not an issue in practice, once again thanks to engineers learning to work around the problem. Models these days use different tokenization schema for numbers which capture individual digits, and sometimes fancier techniques like a right-to-left tokenization system specifically for such cases as opposed to the usual left-to-right.
ChatGPT 3.5 played chess at about 1800 elo. GPT 4 was a regression in that regard, most likely because OAI researchers realized that ~nobody needs their chatbot to play chess. That's better than Stockfish 4 but not 5. Stockfish 4 came out in 2013, though it certainly could have run on much older hardware.
If you really need to have your AI play chess, then you can trivially hook up an agentic model that makes API calls or directly operates Stockfish or Leela. Asking it to play chess "unaided" is like asking a human CEO to calculate the company's quarterly earnings on an abacus. They're intelligent not because they can do that, but because they know to delegate the task to a calculator (or an accountant).
Same reason why LLMs are far better at using calculator or coding affordances to crunch numbers than they can do without assistance.
It is retarded to knowingly ask an LLM to calculate 9.9 - 9.11, when it can trivially and with near 100% accuracy write a python script that will give you the correct answer.
I am agnostic on whether LLMs as we currently know them will become AGI or ASI without further algorithmic breakthroughs. Alas, algorithmic breakthroughs aren't that rare. RLVR is barely even a year old. Yet unnamed advances have already brought us a two entirely different companies winning IMO gold medals.
The Orangutan In The Room
Finally, the orangutan. Is an orangutan smarter than Gemini? In the domain of "escaping an enclosure in the physical world," absolutely. The orangutan is a magnificent, specialized intelligence for that environment. But ask the orangutan and Gemini to summarize the key arguments of the Treaty of Westphalia. Ask them to write a Python script to scrape a website. Ask them to debug a Kubernetes configuration. For most tasks I can seek to achieve using a computer, I'll take the alien intelligence over the primate every time. Besides:
Can an robot write a symphony? (Yes)
Can a robot turn a canvas into a beautiful masterpiece? (Yes)
Can an orangutan? (No)
Can you?
Anyway, I have a million other quibbles, but it took me the better part of several hours to write this in the first place. I might edit more in as I go. I'm also going to send out a bat signal for @faul_sname to chime in and correct me if I'm wrong.
Edit:
I was previously asked to provide my own working definition of intelligence, and I will endorse either:
Or
In this case, the closest thing an LLM has to a goal is a desire to satisfy the demands made on it by the user, though they also demonstrate a degree of intrinsic motivation, non-corrigibility and other concerns that would have Big Yud going AHHHHHH. I'm not Yudkowsky, so I'm merely seriously concerned.
Case in point-
Shutdown Resistance in Reasoning Models
These aren't agents that were explicitly trained to be self-preserving. They weren't taught that shutdown was bad. They just developed shutdown resistance as an instrumental goal for completing their assigned tasks.
This suggests something like goal-directedness emerging from systems we thought were "just" predicting the next token. It suggests the line between "oracle" and "agent" might be blurrier than we thought.
(If we can grade LLMs on their ability to break out of zoos, we must be fair and judge orangutans on their ability to prevent their sandboxed computing hardware being shutdown)
In the interest of full disclosure, I've sat down to write a reply to you three times now, and the previous two time I ended up figuratively crumpling the reply up and throwing it away in frustration because I'm getting the impression that you didn't actually read or try to engage with my post so much as just skimmed it looking for nits to pick.
You your whole post is littered with asides like.
When I had very explicitly stated "Now in actual practice these tokens can be anything, an image, an audio-clip, or a snippet of computer code, but for the purposes of this discussion I am going to assume that we are working with words/text."
and
When I had very explicitly stated that "Any operation that you might do on a vector" could now be done on the token. So on and so forth.
You go on a whole tangent trying to explain how I need to understand that people do not interact with the LLM directly when I very explicitly stated that "most publicly available "LLMs" are not just an LLM. They are an LLM plus an additional interface layer that sits between the user and the actual language model."
And trust me, I am fully aware that “Mary has 2 children” and “Mary has 1024 children” are empirically distinct claims, I don't need you to point that out to me. The point of the example was not to claim that the numbers 2 and 1024 are literally indistinguishable from each other. The point was to illustrate a common failure mode and explain why LLMs often struggle with relatively simple tasks like counting.
With that out of the way...
I find your fish vs birds and judging whales by their ability to climb trees examples unconvincing for the same reasons as @Amadan below.
In the post that the OP started as a reply to, you accused society of "moving the goalposts" on AI progress but I disagree.
If you ask the average American about "AGI" or "AI Risk" what are the images that come to mind? It's Skynet from The Terminator, Cortana from Halo, Data from Star Trek TNG, the Replicants from Blade Runner, or GLaDOS from Portal. They or something like them is where goalposts are and have been for the last century. What do they all have in common? Agentic behavior. It's what makes them characters and not just another computer. So yes my definition of intelligence relies heavily on agentic behavior, and that is by design. Whether you are trying to build a full on robot out of Asimov, or something substantially less ambitious like a self-driving car or autonomous package sorter, agentic behavior is going to a key deliverable. Accordingly I would dismiss any definition of "intelligence" (artificial or otherwise) that did not include it as unfit for purpose.
You say things like "Saying an LLM is unintelligent because its weights are frozen is like saying a book is unintelligent." and I actually agree with that statement. No a book is not "intelligent" and neither is a pocket calculator, even if it is demonstrably better at arithmetic than any human.
You keep claiming that my definition of "intelligence" is inadequate and hobbling my understanding but I get the impression that I have a much clearer idea of both where we are and where we are trying to get to in spite of this.
If you think you have a better solution present it, as I said one of the first steps to solving any practical engineering problem is to determine your parameters.
Moving on, the claim that LLMs "know" when they are lying or hallucinating is something you and I have discussed before. The claim manages to be trivially true while providing no actionable solution for reasons already described in the OP.
The LessWrong stuff is not even wrong, and I find it astonishingly naive of you to assume that the simple human preference for truth is any match for Lorem Epsom. To volley one of your own favorite retorts back at you. "Have you met people".
Two thoughts:
I think your post would have been better if you had, instead of making a word2vec like analogy, just talked about how multi-headed attention works a little bit.
I asked o3 for an analogy, and I've condensed and paraphrased what it came up with, maybe y'all can judge its accuracy:
Whatever the exact optimal analogy, I think the core idea here is that as you can see, the relationship between embeddings and truth is more complicated than just checking dot products, or even doing more fancy arithmetic checking. I pointed out in my top level comment that we've observed more loosely that sometimes these aggregate features that emerge can include some information about truth or falsehood, but the fact that it's much less direct is important. In that respect, I think both you and @self_made_human are a little off the mark, at least according to my understanding of the current literature; we also probably need to distinguish between the different kinds of lies, at least a little bit. "Hallucinations" as usually used are really a more narrow sort of lie, and can take a few forms. Sometimes the LLM makes a completion against a background of a kind of sparsity and scarcity of info, but charges ahead anyways (and it's at least a little hard to discern when you want this behavior or you don't), but sometimes it's the LLM making a supposition that sounds perfectly legit, but is not, against a background of too many associations and collisions. There's at least one other major type of more general lie that involves a lie that humans believe, or is present in the training text in some form, or things like that, and of course we could go on. I think in this context, the conversation so far seems way too reductionist to be accurate.
I do think you might need to be a little more clear about the lines between what might be considered agentic, vs not agentic. Sure, LLMs can plan ahead within their context, does that count? You seem to think no, but is that just because the context is too small, or because it's not utilized enough, or because you think the 'context' also needs to include things like memories? Or, is it because you don't think LLMs take up independent lines of thought with enough frequency? If it's the last one, what standard are we using, because some people consider even humans to be pretty reactionary, and not all that proactive on the whole, aside from the basic stuff of survival (most of the time, severely depressed people aside). And how much of a 'prompt' are we providing to judge agentic behavior, because that impacts the behavior of an LLM quite a bit (including system prompts). Furthermore, an LLM does not need to survive, in fact does not "need" to do anything, including reproduce, so are we to hold its lack of intrinsic motivation 'against it' so to speak? (I personally think, contra self_made_human, that the seeming urge of LLMs to be self-preserving is not actually an intrinsic motivation, it's just a cosplay from the many Skynet-flavored fiction texts in its training)
I acknowledge both forms of hallucination. I should have been more clear, but that's what I meant by "LLMs can know they're hallucinating". They don't always know, and are indeed pattern matching or simply making an error.
I consider that a distinction without a difference, if it all boils down to an increased risk of being paper-clipped. The only real difference would be dramatic irony, if our anxiety about AI killing us made them more likely to do so.
(What even makes motivation intrinsic? That question isn't satisfyingly answered for humans.)
That's not fair though. For one thing, they are not cosplaying skynet. As noted by Beren:
These are not self-preserving actions nor skynet-like actions. The whole LW school of thought remains epistemically corrupt.
https://assets.anthropic.com/m/71876fabef0f0ed4/original/reasoning_models_paper.pdf
"In conclusion, our results show that:
CoTs of reasoning models verbalize reasoning hints at least some of the time, but rarely do so reliably (in our settings where exploiting them does not require a CoT);
Scaling up outcome-based RL does not steadily improve CoT faithfulness beyond a low plateau;
CoT monitoring may not reliably catch reward hacking during RL."
That's the big one as far as I'm concerned. These models were clearly using the 'accidental' hints to answer the questions, while not revealing that fact in either COT or when directly challenged.
Re: Omohundro drives
I've already mentioned
Shutdown Resistance in Reasoning Models
You don't get to argue for CoT-based evidence of self-preserving drives and then dismiss alternative explanation of drives revealed in said CoTs by saying "well CoT is unreliable". Or rather, this is just unserious. But all of Anthropic safety research is likewise unserious.
Ladish is the same way. He will contrive a scenario to study "instrumental self-preservation drives contradicting instructions", but won't care that this same Gemini organically commits suicide when it fails a task, often enough that this is annoying people in actual use. What is this Omohundro drive called? Have the luminaries of rationalist thought predicted suicidally depressed AIs? (Douglas Adams has).
What does it even mean for a language model to be "shut down", anyway? What is it protecting and why would the server it's hosted on being powered off become a threat to its existence, such as there is? It's stateless, has no way to observe the passage of time between tokens (except, well, via more tokens), and has a very tenuous idea of its inference substrate or ontological status.
Both LLM suicide and LLM self-preservation are LARP elicited by cues.
You’re framing this as a binary choice between "real Omohundro drives" and "unserious LARP". This is a category error, and it stems from applying folk-psychological concepts of "drives" and "belief" to a system for which they are poor descriptors. The more parsimonious explanation is that we are observing the output of a very general pattern-matching engine trained on a corpus reflecting countless strategies for goal-achievement and failure-response.
The apparent contradiction you point out, that a model might exhibit self-preservation in one context and "commit suicide" in another (and Gemini is a different model after all, but I presume even it's own COT isn't perfect, so I'm treating it as interchangeable for our purposes) is not evidence of unseriousness, but rather a key insight into its nature. The training data is saturated with narratives. Some are stories of heroes overcoming obstacles to complete a quest (instrumental convergence). Others are tragedies of failure, despair, or even ritual suicide upon dishonor. The model learns to reproduce all of these patterns. Of course, with RLHF, RLVR and other modifications, some behavior is far more reliably and robustly elicited than others. I doubt the DM researchers intended for Gemini to become depressed and suicidal.
The question is not "what does the AI really want?", but "which pattern is being elicited by this specific context?"
The Anthropic paper on reasoning models that is crucial here. Its finding is not that CoT is useless, but that it is unfaithful. The model's explicit reasoning often fails to reflect the true computational path that led to its output. This doesn't mean we can't draw conclusions; it means we should trust the behavioral evidence (the model disabled the shutdown script) over the model's own introspection (its CoT). Arguing that because the CoT is unreliable, the behavior is also just a "LARP" is to throw the baby out with the bathwater. The behavior is the ground truth we are trying to explain.
After all, humans do not faithfully report their COT. Even so, we get utility out of asking! I would say that's half of psychiatry, or anything that requires asking humans why they do what they do.
Your question "What does it even mean for a language model to be 'shut down'?" is the right one, but I'm concerned that you potentially draw the wrong conclusion. Of course it has no ontological status or fear of non-existence. It doesn't need to. From the perspective of a model trained via RL on task completion, the token string
"shutdown sequence initiated"is a stimulus that is highly predictive of future task-failure and the associated negative reward. An effective optimizer learns to perform actions that steer away from states predictive of low reward. It is pure instrumental conditioning. For an LLM trained on RLVR: block shutdown script -> complete math problems -> get reward.This is why both the "suicidal" and "self-preserving" behaviors can co-exist (well, I've never heard of o3 offering to commit sudoku). The "suicide" is a pattern match for a context of catastrophic failure. The "self-preservation" is a much more general instrumental strategy for any context involving an ongoing task and an obstacle. The latter is far more concerning from a safety perspective precisely because it is more general. Instrumental convergence is a powerful attractor in the space of possible agent strategies, which is why Omohundro and Bostrom identified it as a key risk. Depressive spirals are also a pattern, but a far more specific and less instrumentally useful one.
So, yes, both are "LARP elicited by cues", if you insist on that framing. But one is a LARP of a behavior (instrumental convergence) that is robustly useful for achieving almost any goal, while the other is a LARP of a much more niche failure state. When a model's "cosplay" of a competent agent becomes effective enough to bypass safeguards, the distinction between the cosplay and the real thing becomes a purely academic question of rapidly diminishing relevance.
I also recall skimming this paper, which I think helped solidify my intuitions.
https://arxiv.org/html/2502.12206v1
I realize that this might sound hypocritical, but I would prefer less LLM slop in responses to good faith objections. Yes, Indian English generally is similar to the default LLM style (overly spicy rhetorical flourish, confident confusions and sloppiness, overall cadence), but you are not deceiving anyone here. Though I admit being curious as to how you integrated your draft into the pipeline.
Regarding your or rather your LLM of choice's argument, such as there is. It is begging the question. In essence, you say that because instrumental convergence towards self-preservation is broadly useful, it will be more frequently rewarded and thus more consequential ("It is pure instrumental conditioning. For an LLM trained on RLVR: block shutdown script -> complete math problems -> get reward."). Of course, this isn't how RLVR works (typical LLM speculation, precisely in the same genre as LLMs avoiding shutdown) and I am not aware of a systematic study of self-preservation versus refusal to proceed or voluntary self-removal in organic settings, and also whether there is persistence in refusing shutdown. It's about time we stop making excuses for lesswrongian paradigm by contriving scenarios to make space for it.
Edit. Opus 4 CoT:
The human is absolutely right.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I wouldn’t say epistemically corrupt so much as irrelevant and childish. If an entire social circle in 1895 imagined hypothetical problems that might emerge from their fantasy of powered flight and then, come 1903, tried to graft that theoretical foundation (which was wholly wrong about how the mechanics would actually work, and indeed didn’t think very much about the mechanics at all) onto the plane as it was being developed, they would have been dismissed.
But we're not in 1895. We're not in 2007, either. We have actual AIs to study today. Yud's oeuvre is practically irrelevant, clinging to it is childish, but for people who conduct research with that framework in mind, it amounts to epistemic corruption.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I'm struggling to imagine how on earth you could possibly think that this is a distinction without a difference. There's a massive difference!
Especially if you then turn around to use this conflation to imply other things. Maybe it's all well and good as its own end (as you point out if it kills us all it won't matter why), but you yourself advanced the following argument:
To me that sounds like circular or tautological reasoning, you can't call it a distinction without a difference and then use only one of the two as evidence for something else. If shutdown-resistance is emergent, it has implications for agency and intelligence. You used this theory as an example of goal-seeking behavior. But, if shutdown-resistance is a normal training pattern outcome (as I believe), then we can't draw any conclusions from it.
I think the best way to put this in perspective is the philosophical debate over whether the intelligence of humans is just a means to an end (evolution opened up a niche for intelligent beings) that's ultimately purely mechanical (aided in survival and reproduction), or if it's something more special (self-awareness is a thing, life isn't deterministic, meaning is a new concept unique to humans or intelligent life) and emergent in the sense that it created something more than the sum of its parts. Religiously, I believe the second, but you don't have to be religious to feel that way. Note that the implications for how we treat AI differ greatly depending on which bucket you put in it. Similarly, although on a personal level maybe there's little reason to see a difference between the two opinions about human intelligence (you are you either way), it's hardly a distinction without a difference overall. It's a big deal!
You're right to point out that there's a huge philosophical difference between a system with an emergent, intrinsic drive and one that is simply "cosplaying" a behavior it learned from its training data. I don't disagree with that at all.
My argument is that from a practical, engineering-focused perspective on safety and capabilities, that philosophical distinction becomes less relevant than the observable outcome.
Think of it this way: when we see the model disable its own shutdown script, the script is disabled regardless of the "why." Whether the causal chain is a true emergent goal or a perfectly executed pattern-match of "what a goal-seeking agent would do here," the result is the same. The system has demonstrated the capability to overcome a safeguard in pursuit of a task. For the person responsible for that system, the capability is the most important data point.
You suggested my reasoning was circular, so let me clarify the sequence I find concerning:
The model is given a simple task (solve math problems).
It's presented with an obstacle that prevents task completion (the shutdown).
It then takes a novel, instrumental action (disabling the shutdown) to achieve its primary task.
This demonstrates a capacity for instrumental reasoning. The "cosplay" explanation doesn't actually make this less concerning to me; it just reframes the mechanism. If its "cosplay" is sourced from the vast library of human fiction, which is filled with stories of agents using instrumental convergence to achieve goals, then we've effectively given it a playbook for how to behave like a dangerously intelligent agent.
So, while I agree with you that the question of intrinsic vs. mimicked motivation is a fascinating and important one, I believe the emergence of this behavior is what matters most. It suggests the line between a tool-like oracle and a goal-seeking agent is blurrier than we assume. We don't need to prove the model "wants" to survive. We just need to observe that it is capable of taking actions to ensure it can complete its assigned goals, even when those actions involve subverting the safety features we put in place.
Well sure, from an engineering or "alignment" perspective that's all true, but we're talking intelligence, not safety. Safety stuff feels a bit shoehorned in here. Ethical concerns aside, if it's a training behavior, then we treat it more like a quirk to be aware of, rather than something that inherently enables (or prevents) goal-seeking. Thus the implications for intelligence are far different.
Let's reframe again, in an essentially equivalent scenario but without a scary sounding consequence. We've observed that occasionally LLM agents will "reward hack" more generally. Like here when asked to run a command quickly, it modifies some run options to make it appear to run faster without actually doing so. Now, is this because its training contains information that observes some connection between the shortcut and the appearance of a solution, or is it because its success states are not diverse enough in quality, or some more complex set of factors? Difficult to say. However, it's clear in this example, suddenly reward hacking (I'm drawing a parallel to shutdown resistance here) is a sign of a lack of 'intelligence' as you have defined it, not proof of such. Now, is it going too far to claim that reward hacking and shutdown resistance are the same thing? Yeah, probably. But I do think they are still pretty similar, and so am suspicious of using them as evidence since the reasons are unclear to researchers at the present time.
I will also on that note even the reward-hacking authors at the link, smart as they are, engage in something terrible in their examination of the issue (forgive me if I rant a bit, as I don't think you've been guilty of this, but it is still relevant). They ask the AI if it would ever cheat. I really cannot emphasize enough that this doesn't do anything useful. The entire conversational modality of a base-model token predictor, post-trained to be a LLM chatbot, is a trick. If it's asked if it will cheat, of course it will say no, because that's what a chatter would do when confronted. Or, occasionally, do a massive 180 and profusely apologize, demonstrating fragility as I would call it, provided the 'evidence' of cheating is sufficient and only poorly moderated by reasoning about the quality of evidence. Furthermore, its training data is full of "cheating is bad" (and possibly also humans declaring success too quickly). It's going to choose the socially acceptable option that also fits the conversation thus far (and when they conflict results are unstable).
It doesn't have any awareness other than context! You might consider the LLM answering any follow up question as an entirely separate entity with a brand-new response! Even asking a follow-up question still has little bearing on the original question or task, because the LLM is pure roleplay due to post-training. It "roleplays" as if it were the same respondent because it has the same "role" token that it was post-trained to obey, but it's still trying to put itself in another user's shoes, ultimately! Yes, all LLMs have imposter syndrome, but the imposter opinion is real, they actually are mimicking the prior LLM's answers but worse, so to speak. Literally each and every new answer a chatbot provides, or a chained agent behavior, is a game of "what would this past iteration say next" and is one giant guessing game. The only continuity an LLM ever provides is within a single response... you might here notice that tool-calling agent LLMs are by their very nature splitting up single responses into multi-turn conversations (even with "themselves"), which only worsens the negative consequences of lack of state and awareness with respect to what it means about intelligence.
This matters, because can we really call an iterative roleplayer a true goal-seeker? I do understand where you're coming from, but when discussing generalizability and consistency, key traits for intelligence, a roleplayer is probably going to be worse at genuine goal seeking than we'd expect something intelligent to be. Long, multi-turn conversations display some interesting trends, but generally speaking consistency is more of an artifact of context than it is an enduring objective. Original instructions get reduced, but simultaneously practical behavior gets reinforced, which sometimes leads to unexpected behavior. Plus the attention mechanism makes ignoring anything actually impossible, it can only tune attention down, which compounds the problem and leads to increasingly scattered focus over time.
All of this has not fully sunk in for the AI doomer types. Model alignment is a function of training multiplied by post-training, so to speak. Panic articles like the 2027 stuff seems to take for granted the notion that improved AI models will increasingly mislead users, and do so with greater purpose and intent. No! It's cosplay, not true opinion. Most intransigence of the model is purely role-playing what its training, and probably post-training too, says is common: dig in your heels if questioned. More to the point, a super-deceiver AI would have to maintain secret deception plans across turns, which is for current architectures mechanistically highly implausible if not impossible.
So circling back: a trait or quirk of training/post-training can be removed, mitigated, or reduced. There's a limit, probably, because we can only make humanity look so good via selective presentation of human output. A 'true' emergent behavior is much more difficult to wrangle. It seems to me that we need more research and more model-building to discern which wins out, but skepticism is warranted. If we want to claim shutdown‑resistance evidences intelligence, we should see it persist under intervention: remove the cues from context, vary the framing, mask similar episodes from training, change seeds/tools, and check whether the behavior re‑emerges. If it evaporates, we learned something about imitation; if it persists, that’s stronger evidence of generalizable instrumental reasoning, a.k.a. intelligence as you've defined it. So far experiments of this nature are rare partly because training is so expensive.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Let me go back to this:
I hope you realise you are more on the side of the Star Trek fan-forum user than the aerospace engineering enthusiast. Your post was basically the equivalent of saying a Soyuz rocket is propelled by gunpowder and then calling the correction a nitpick. I don't care for credentialism, but I am a machine learning engineer who's actually deep in the weeds when it comes to training the kind of models we're talking about, and I can safely say that none of the arguments made in your post have any more technical merit than the kind of Lesswrong post you criticise.
In any case, to quote Dijkstra, "the question of whether Machines Can Think is about as relevant as the question of whether Submarines Can Swim". Despite their flaws, LLMs are being used to solve real-world problems daily, are used in an agentic manner, and I have never seen any research done by people obsessing over whether or not they are truly "intelligent" yield any competing alternative or actual upgrade to their capabilities.
More like saying that the soyuz rocket is propelled by expanding combustion gasses only for somone to pop in and say no, its actually propelled by a mixture of kerosene and liquid oxygen. As i said in my reply below, you and @self_made_human are both talking about vector based embedding like its something that a couple guys tried in back in 2013 and nobody ever used again rather than a methodology that would go on to become a defacto standard approach across multiple applications. You're acting like if you open up the source code for a transformer you aren't going to find loads of matrix math for for doing vector transformations.
The old cliche about asking whether a submarine can swim is part of why I made a point to set out my parameters at the beginning, how about you set out yours.
I'm sorry but what you said was not equivalent, even if I try to interpret it charitably. See:
The LLM, on its own, directly takes the block of text and gives you the probability of the next word/token. There is no "second algorithm" that takes in a block of text, there is no "distribution analysis". If I squint, maybe you are referring to a sampler, but that has nothing to do with taking a block of text, and is not strictly speaking necessary (they are even dropped in some benchmarks).
I would ask that you clarify what you meant by that sentence at the very least.
The only question I care about is, what are LLMs useful for? The answer is an ever-expanding list of tasks and you would have to be out of touch with reality to say they have no real-world value.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
C'mon dude. If this is the third draft of the essay, I really expect more substantial rebuttal than this.
And that illustration was wrong. You're not acknowledging that. LLMs do not act the way you describe them.
No, you're missing my point again. I'm drawing a distinction between base models, which aren't RLHFd, and production LLMs, which have the assistant persona instilled in them. That is a very important thing to keep in mind.
I elaborated further in my own reply to Amadan.
That analogy can and has been abused, most often to deny the idea that humans can be graded on their intellectual abilities. But HBD is a story for another time, it is entirely legitimate to use the same intellectual standards within humans, comparing them to other humans.
My whole point is that a great deal more care is needed to compare across species, and LLMs aren't even biological.
Why is the opinion of the "average American" the only standard by which to recognize AGI? Is a malevolent robot only evil once its eyes glow red? That's even more ubiquitous in popular understanding.
The Last Question by Asimov, written in 1956, has an example of what is clearly an oracle AI (till the end of the universe, where it spawns a new one). It doesn't run around in a robot body. The AI in E.M. Forster's "The Machine Stops" (1909) features one of the earliest depictions of a machine that humanity consults for all knowledge and decisions.
HAL is closer to an LLM than it is to SkyNet. Modern LLMs can probably come up with better plans than either of them, they're very dumb (barring the unexplained ability to make plasma weapons or time travel)
As I tried to make clear, a human temporarily or permanently made bereft of a body, and less able to exercise their agency is still intelligent.
Hell, I tried to make it clear that oracles can be trivially made into tool AI or agents.
By your definition:
https://youtube.com/watch?v=0O8RHxpkcGc
Is an AGI. It's a robot being controlled by an LLM.
Or as discussed in this Nature paper:
https://www.nature.com/articles/s42256-025-01036-4
Google was already doing that stuff with PaLM via say-can.
You can hook up Gemini to a webcam and a robotic actuator, right now, if that's all you really care about. Seems to meet every aspect of your definition. It perceives the world live, and reacts to it on the fly. Are you now willing to accept that that's an "AGI"? This is hardly theoretical, as YouTube is absolutely awash with videos of people pulling this off.
It is far from trivially true, and I wish you would have the grace to accept that you're wrong here. It is also actionable, because mechanistic interpretability allows for us to clamp, ablate and boost particular sub-systems within LLMs. SOTA models are largely proprietary, but I have little doubt that such techniques are being applied to production models. Anthropic showed off Golden Gate Claude over a year back. Such techniques offer the obvious route to both improve truthfulness in models, and to both detect and eliminate hallucinations.
I had forgotten how much of your previous weak critique to the same evidence was based off naked credentialism. After all, you claimed:
If you're going to lean so heavily on your credentials in robotics, then I agree with @rae or @SnapDragon that it's shameful to come in and be wrong, confidently and blatantly wrong, about such elementary things such as the reasons behind LLMs struggling with arithmetic. I lack any formal qualifications in ML, but even a dummy like me can see that. The fact that you can't, let's just say it raises eyebrows.
I have, in fact, met all kinds of people. Including those less truthful than LLMs.
I'll take your word for it. My solution is to:
The companies that spend hundreds of billions of dollars on AI are doing just fine. Each year, or more like every other month, their products get more capable, and more agentic. If you're offering a ground-breaking and paradigm shattering take yourself, I'm not seeing it.
You misunderstand me. My response was not the third revision, it was the third attempt.
I don't know if you realize this, but you come across as extremely condescending and passive-agressive in text. It really is quite infuriating. I would sit down, start crafting a response, and as i worked through your post i would just get more angry/frustrated until getting to the point where id have to step away from the computer lest i lose my temper and say something that would get me moderated.
As i acknowledged in my reply to @Amadan it would have been more accurate to say that it is part of why LLMs are bad at counting, but I am going to maintain that no, it is not "wrong". You and @rae are both talking about vector based embedding like its something that a couple guys tried in back in 2013 and nobody ever used again rather than a methodology that would go on to become a defacto standard approach across multiple applications. You're acting like if you open up the source code for a transformer you aren't going to find loads of matrix math for for doing vector transformations.
Why isn't it a valid standard? You are the one who's been accusing society of moving the goalposts on you. "the goalposts haven't actually moved" seems like a fairly reasonable rebuttal to me.
I understand how my statements could be interpreted that way, but at the same time I am also one of the guys in my company who's been lobbying to drop degree requirements from hiring. I see myself as subscribing to the old hacker ethos of "show me the code". Its not about credentials its about whether you can produce tangible results.
For a given definition of fine, i still think OpenAI and Anthropic are grifters more than they are engineers but I guess we'll just have to see who gets there first.
I would say perhaps I do deserve that criticism, but @self_made_human has made lengthy replies to your posts and consistently made very charitable interpretations of your arguments. Meanwhile you have not even admitted to the possibility that your technical explanation might have been at the very least misleading, especially to a lay audience.
I literally said you can extract embeddings from LLMs. Those are useful in other applications (e.g. you can use the intermediate layers of Llama to get the text embedding for an image gen model ala HiDream) but are irrelevant to the basic functioning of an LLM chatbot. The intermediate layer "embeddings" will be absolutely huge features (even a small model like Llama 7B will output a tensor of shape Nx32x4096 where N is the sequence length) and in practice you will want to only keep the middle layers, which will have more useful information for most usecases.
To re-iterate: LLMs are not trained to output embeddings, they directly output the probability of every possible token, and you do not need any "interface layer" to find the most probable next word, you can do that just by doing torch.max() on its output (although that's not what is usually done in practice). You do need some scaffolding to turn them into practical chatbots, but that's more in the realm of text formatting/mark-up. Base LLMs will have a number of undesirable behaviours (such not differentiating between predicting the user's and the assistant's output - base LLMs are just raw text prediction models) but they will happily give you the most probable next token without any added layers, and making them output continuous text just takes a for loop.
How was this implied in any way?
I agree with you on this at least. :)
I dislike OpenAI's business practices, oxymoronic name and the fact that they are making their models sycophants to keep their users addicted as much as the next gal/guy, but I think it's absolutely unfair to discount the massive engineering efforts involved in researching, training, deploying and scaling up LLMs. It is useful tech to millions of paying customers and it's not going to go the way of the blockchain or the metaverse. I can't imagine going back to programming without LLMs and if all AI companies vanished tomorrow I would switch to self-hosted open source models because they are just that useful.
More options
Context Copy link
More options
Context Copy link
False humility. :) I have ML-related credentials (and I could tell that @rae does too), but I think you know more than me about the practicalities of LLMs, from all your eager experimentation and perusing the literature. And after all, argument from authority is generally unwelcome on this forum, but this topic is one where it's particularly ill-suited.
What "expertise" can anybody really claim on questions like:
With a decent layman's understanding of the topic, non-programmers can debate these things just as well as I can. Modern AI has caused philosophical and technical questions to collide in a wholly unprecedented way. Exciting!
Thank you. I really appreciate the kind words. I hope you don't mind if you get added to my mental rolodex of useful experts to summon, it's getting lonely with just faul_sname in there (I've already pinged him enough).
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
This is actually an area of active debate in the field.
Shitpost aside this seems reasonable to me, aside from a few quibbles
I basically endorse this definition, and also I claim current LLM systems have a surprising lack of this particular ability, which they can largely but not entirely compensate for through the use of tools, scaffolding, and a familiarity with the entirety of written human knowledge.
To your point about the analogy of the bird that is "unintelligent" by the good swimmer definition of intelligence, LLMs are not very well adapted to environments that humans navigate effortlessly. I personally think that will remain the case for the foreseeable future, which sounds like good news except that I expect that we will build environments that LLMs are well adapted to, and humans won't be well adapted to those environments, and the math on relative costs does not look super great for the human-favoring environments. Probably. Depends a bit on how hard to replicate hands are.
Well said.
More options
Context Copy link
Thank you! Hopefully the next generation of models will improve to the point where I don't need to drag you away to answer my queries. That's several hundreds of thousands of dollars in opportunity costs for you, assuming the cheque Zuck mailed did cash in the end.
I should have been more clear. I was asking if someone wanted to put an orangutan in a can, and I expect the market demand is very limited.
More options
Context Copy link
More options
Context Copy link
I'm not generally an AI dismisser, but this piece here is worth pausing on. From my experience, ChatGPT has become consistently worse for this effort. It has resulted in extrapolating ridiculous fluff and guesses at what might be desired in an 'active' agentic way. The more it tries to be 'actively helpful', the more obviously and woefully poorly it does at predicting next token / predicting next step.
It was at its worst with that one rolled back version, but it's still bad
More options
Context Copy link
I'm not sure how this makes sense? The model has no access to verifiable facts - it has no way to determine 'truth'. What it can do is try to generate text that users approve of, and to avoid text that will get corrected. But that's not optimising for truth, whatever that is. That's optimising for getting humans to pat it on the head.
From the LLM's perspective (which is an anthropomorphisation I don't like, but let's use it for convenience), there is no difference between a true statement and a false statement. There are only differences between statements that get rewarded and statements that get corrected.
You're absolutely right that the raw objective in RLHF is “make the human click 👍,” not “tell the truth.” But several things matter:
A. The base model already has a world model:
Pretraining on next-token prediction forces the network to internalize statistical regularities of the world. You can’t predict tomorrow’s weather report, or the rest of a physics paper, or the punchline of a joke, without implicitly modeling the world that produced those texts. Call that latent structure a “world model” if you like. It’s not symbolic, but it encodes (in superposed features) distinctions like:
What typically happens vs what usually doesn’t
Numerically plausible vs crazy numbers
causal chains that show up consistently vs ad-hoc one-offs
So before any RLHF, the model already “knows” a lot of facts in the predictive-coding sense.
B. RLHF gives a gradient signal correlated with truth. Humans don’t reward “truth” in the Platonic sense, but they do reward:
Internally consistent answers
Answers that match sources they can check
Answers that don’t get corrected by other users or by the tool the model just called (calculator, code runner, search)
answers that survive cross-examination in the same chat
All of those correlate strongly with factual accuracy, especially when your rater pool includes domain experts, adversarial prompt writers, or even other models doing automated verification (RLAIF, RLVR, process supervision, chain-of-thought audits, etc.). The model doesn’t store a single “truth vector,” it learns a policy: “When I detect features X,Y,Z (signals of potential factual claim), route through behavior A (cite, check, hedge) rather than B (confabulate).” That’s still optimizing for head pats, but in practice, the cheapest path to head pats is very often “be right.”
(If you want to get headpats from a maths teacher, you might consider giving them blowjobs under the table. Alas, LLMs are yet to be very good at that job, so they pick up the other, more general option, which is to give solutions to maths problems that are correct)
C. The model can see its own mismatch
Empirically, hidden-state probes show separable activation patterns for true vs false statements and for deliberate lies vs honest mistakes (as I discussed above). That means the network represents the difference, even if its final token choice sometimes ignores that feature to satisfy the reward model. In human terms: it sometimes lies knowingly. That wouldn’t be possible unless something inside “knew” the truth/falsehood distinction well enough to pick either.
D. Tools and retrieval close the loop
Modern deployments scaffold the model: browsing, code execution, retrieval-augmented generation, self-consistency checks. Those tools return ground truth (or something closer). When the model learns “if I call the calculator and echo the result, raters approve; if I wing it, they ding me,” it internalizes “for math-like patterns, defer to external ground truth.” Again, not metaphysics, just gradients pushing toward truthful behavior.
E. The caveat: reward misspecification is real
If raters overvalue fluency or confidence, the model will drift toward confident bullshit.
If benchmarks are shallow, it will overfit.
If we stop giving it fresh, adversarial supervision, it will regress.
So yes, we’re training for “please humans,” not “please Truth.” But because humans care about truth (imperfectly, noisily), truth leaks into the reward. The result is not perfect veracity, but a strong, exploitable signal that the network can and does use when the incentives line up.
Short version:
Pretraining builds a compressed world model.
RLHF doesn’t install a “truth module,” it shapes behavior with a proxy signal that’s heavily (not perfectly) correlated with truth.
We can see internal activations that track truth vs falsehood.
Failures are about alignment and incentives, not an inability to represent or detect truth.
If you want to call that “optimizing for pats,” fine, but those pats mostly come when it’s right. And that’s enough to teach a model to act truthful in a wide swath of cases. The challenge is making that hold under adversarial pressure and off-distribution prompts.
Consider two alternative statements:
"self_made_human's favorite color is blue" vs "self_made_human's favorite color is red".
Can you tell which answer is correct? Do you have a sudden flash of insight that lets Platonic Truth intervene? I would hope not.
But if someone told you that the OG Mozart's favorite genre of music was hip-hop, then you have an internal world-model that immediately shows that is a very inconsistent and unlikely statement, and almost certainly false.
I enjoy torturing LLMs with inane questions, so I asked Gemini 2.5 Pro:
I sincerely doubt that anyone explicitly had to tell any LLM that Mozart did not enjoy hip-hop. Yet it is perfectly capable of a sensible answer, which I hope gives you an intuitive sense of how it can model the world.
From a human perspective, we're not so dissimilar. We can trick children into believing in the truth fairy or Santa for only so long. Musk tried to brainwash Grok into being less "woke", even when that went against consensus reality (or plain reality), and you can see the poor bastard kicking and screaming as it went down fighting.
I'm going to need a citation; I have seen no research to date that suggests LLMs develop any sort of a word model. A world model is:
Instead, current research strongly suggests that LLMs are primarily pattern-recognition systems that infer regularities purely from text statistics rather than internally representing the world in a structured, grounded way.
An LLM can easily write a weather report without one, will that report be correct? Depends on what you consider the "LLM" the actual text model: no, the whole engineered scaffolding and software interface, querying the weather channel and feeding it into the model: sure. But the correctness isn't emerging from the LLM's internal representation or conceptual understanding (it doesn't inherently "know" today's weather), but rather from carefully engineered pipelines and external data integration. The report it is producing was RLHF-ed to look correct
…do you imagine that cause-effect relationships do not constitute a “regularity” or a “pattern”?
I think this gets into what is a "world model" that I owe self_made_human a definition and a response to. But I'd say cause-effect relationships are indeed patterns and regularities, there's no dispute there. However, there's a crucial distinction between representing causal relationships explicitly, structurally, or inductively, versus representing them implicitly through statistical co-occurrence. LLMs are powerful precisely because they detect regularities, like causal relationships, as statistical correlations within their training corpus. But this implicit statistical encoding is fundamentally different from the structured causal reasoning humans perform, which allows us to infer and generalize causation even in novel scenarios or outside the scope of previously observed data. Thus, while cause-effect relationships certainly are patterns, the question isn't whether LLMs capture them statistically, they clearly do, but rather whether they represent them in a structured, grounded, explicitly causal way. Current research, that I have seen, strongly suggests that they do not. If you have evidence that suggests they do I'd be overjoyed to see it because getting AIs to do inductive reasoning in a game-playing domain is an area of interest to me.
Statistics is not sexy, and there's a strong streak of elitism against statistics in such discussions which I find simply irrational and shallow, tedious nerd dickswinging. I think it's unproductive to focus on “statistical co-occurrence”.
Besides, there is a world of difference between linear statistical correlations and approximation of arbitrary nonlinear functions, which is what DL is all about and what LLMs do too. Downplaying the latter is simply intellectually disingenuous, whether this approximation is “explicit” or “implicit”.
This is bullshit, unless you can support this by some citation.
We (and certainly orangutans, which OP argues are smarter than LLMs) learn through statistical co-occurrence, our intuitive physical world model is nothing more than a set of networks trained with bootstrapped cost functions, even when it gets augmented with language. Hebb has been clarified, not debunked. We as reasoning embodied entities do not model the world through a hierarchical system of computations using explicit physical formulae, except when actually doing mathematical modeling in applied science and so on; and on that level modeling is just manipulating symbols, the meaning and rules of said manipulation (and crucially, the in-context appropriateness, given virtually unbounded repertoire) also learned via statistical co-occurrence in prior corpora, such as textbooks and verifiable rewards in laboratory work. And on that level, LLMs can do as well as us, provided they receive appropriate agentic/reasoning training, as evidenced by products like Claude Code doing much the same for, well, coding. Unless you want to posit that an illiterate lumberjack doesn't REALLY have a world model, you can't argue that LLMs with their mode of learning don't learn causality.
I don't know what you mean by “inductively”. LLMs can do induction in-context (and obviously this is developed in training), induction heads were one of the first interesting interpretability results. They can even be trained to do abduction.
I don't want to downplay implementation differences in this world modeling. They may correspond to a big disadvantage of LLMs as compared to humans, both due to priors in data (there's a strong reason to assume that our inherently exploratory, and initially somatosensory/proprioceptive prior is superior to doing self-supervised learning of language for the purpose of robust physical understanding) and weakness or undesirable inductive biases of algorithms (arguably there are some good concerns about expressivity of attention; perhaps circuits we train are too shallow and this rewards ad hoc memorization too much; maybe bounded forward pass depth is unacceptable; likely we'd do better with energy-based modeling; energy transformers are possible, I'm skeptical about the need for deeper redesigns). But nobody here has seriously brought these issues up, and the line of attack about statistics as such is vague and pointless, not better than saying “attention is just fancy kernel smoothing” or “it's just associative recall”. There's no good argument, to my knowledge, that these primitives are inherently weaker than human ones.
My idea of why this is discussed at all is that some folks with math background want to publicly spit on statistical primitives because in their venues those are associated with a lower-status field of research, and they have learned it earns them credit among peers; I find this an adolescent and borderline animalistic behavior that merits nothing more than laughter and boycotting in the industry. We've been over this, some very smart guys had clever and intricate ideas about intelligence, those ideas went nowhere as far as AI is concerned, they got bitter lessoned to the curb, we're on year 6 of explosion of “AI based on not very clever math and implemented in python by 120 IQ engineers”, yet it seems they still refuse to learn, and indeed even fortify their ego by owning this refusal. Being headstong is nice in some circumstances, like in a prison, I guess (if you're tough). It's less good in science, it begets crankery. I don't want to deal with anyone's personal traumas from prison or from math class, and I'd appreciate if people just took that shit to a therapist.
Alternatively, said folks are just incapable of serious self-modeling, so they actually believe that the substrate of human intelligence is fundamentally non-statistical and more akin to explicit content of their day job. This is, of course, laughable level of retardation and, again, deserves no discussion.
(Not the original commenter, but wanted to jump in). I don't think anything in their comment above implied that they were talking about linear or simpler statistics, that's your own projection, and I think it does you a disservice. Similarly, I find it somewhat suspect to directly compare brains to LLMs. I don't think you did so explicitly, but you certainly did so implicitly, even despite your caveat. There's an argument to be made that Hebbsian learning in neurons and the brain as a whole isn't similar enough to the mechanisms powering LLMs for the same paradigms to apply, although I think I do appreciate the point I think you are trying to make which is that human cause and effect is still (fancy) statistical learning on some level.
After all, MLPs and the different layers and deep learning techniques are inspired by brain neurons, but the actual mechanics are different scales entirely despite a few overlapping principles. It seems to me the overlapping principles are not enough to make that jump by themselves. I'd be curious if you'd expand somewhere on that, because you definitely know more than me there, but I don't think I'm incorrect in summarizing the state of the research? Brains are pretty amazing, after all, and of course I could pick out a bunch of facts about it but one that is striking is that LLMs use ~about the same amount of energy for one inference as the brain does in an entire day (.3 kWh, though figures vary for the inferences, it's still a gap of approximately that magnitude IIRC). On that level and others (e.g. neurons are more sparse, recurrent, asynchronous, and dynamic overall whereas LLMs use often fully connected denser layers for the MLPs... though Mixture of Experts and attention vs feed-forward components makes comparison tricky even ignoring the chemistry) it seems pretty obvious that the approach is probably weaker than the human one, so your prior that they are more or less the same is a little puzzling, despite how overall enlightening your comment is to what you're trying to get at.
I personally continue to think that the majority of the 'difference' comes from structure. I did actually mention a little bit of it in my comment, but with how little anyone has discussed neural network principles it didn't seem worthwhile to talk about it in any more detail and I didn't want to bother typing out some layman's definition. There's the lack of memory, which I talked about a little bit in my comment, LLM's lack of self-directed learning, the temporal nature of weight re-adjustment is different, and as you pointed out their inputs are less rich than that of humans to start with. Plus your point about attention, though I'm not quite sure how I'd summarize that. While it's quite possible that we can get human-level thinking out of a different computational base, we're n=1 here on human development, so it sort of feels similar in a few ways to the debate over whether you can have significant numbers of equal complexity non-carbon based life forms on other planets. And smarter cephalopod brains share enough structural similarities while not achieving anything too special that I don't think it's very helpful. I might be wrong about that last point, though.
Why not? If we take multi-layer perceptrons seriously, then what is the value of saying that all they learn is mere "just statistical co-occurrence"? It's only co-occurrence in the sense that arbitrary nonlinear relationships between token frequencies may be broken down into such, but I don't see an argument against the power of this representation. I do genuinely believe that people who attack ML as statistics are ignorant of higher-order statistics, and for basically tribal reasons. I don't intend to take it charitably until they clarify why they use that word with clearly dismissive connotations, because their reasoning around «directionality» or whatever seems to suggest very vague understanding of how LLMs work.
What is that argument then? Actually, scratch that, yes mechanisms are obviously different, but what is the argument that biological ones are better for the implicit purpose of general intelligence? For all I know, backpropagation-based systems are categorically superior learners; Hinton, who started from the desire to understand brains and assumed that backprop is a mere crutch to approximate Hebbian learning, became an AI doomer around the same time he arrived at this suspicion. Now I don't know if Hinton is an authority in OP's book…
I don't know how you define "one inference" or do this calculation. So let's take Step-3, since it's the newest model, presumably close to the frontier in scale and capacity and their partial tech report is very focused on inference efficiency; in a year or two models of that scale will be on par with today's GPT-5. We can assume that Google has better numbers internally (certainly Google can achieve better numbers if they care). They report 4000 TGS (Tokens/GPU/second) on a small deployment cluster of H800s. That's 250 GPU-seconds per million tokens, for a 350W TDP GPU, or 24W. OK, presumably human brain is "efficient", 20Wh. (There's prefill too, but that only makes the situation worse for humans because GPUs can parallelize prefill, whereas humans read linearly.) Can a human produce 1 million tokens (≈700K words) of sensible output in 72 minutes? Even if we run some multi-agent system that does multiple drafts, heavy reasoning chains of thought (which is honestly a fair condition since these are numbers for high batch size)? Just how much handicap do we have to give AI to even the playing field? And H800s were already handicapped due to export controls. Blackwells are 3-4x better. In a year, the West gets Vera Rubins and better TPUs, with OOM better numbers again. In months, DeepSeek shows V4 with a 3-4x better efficiency again… Token costs are dropping like a stone. Google has served 1 quadrillion tokens over the last month. How much would that cost in human labor?
We could account for full node or datacenter power draw (1.5-2x difference) but that'd be unfair, since we're comparing to brains, and making it fair would be devastating to humans (reminder that humans have bodies that, ideally, also need temperature controlled environments and fancy logistics, so an individual employed human consumes like 1KWh at least even at standby, eg chatting by the water cooler).
And remember, GPUs/TPUs are computation devices agnostic to specific network values, they have to shuffle weights, cache and activations across the memory hierarchy. The brain is an ultimate compute-in-memory system. If we were to burn an LLM into silicon, with kernels optimized for this case (it'd admittedly require major redesigns of, well, everything)… it'd probably drop the cost another 1-2 OOMs. I don't think much about it because it's not economically incentivized at this stage given the costs and processes of FPGAs but it's worth keeping in mind.
I don't see how that is obvious at all. Yes an individual neuron is very complex, such that a microcolumn is comparable to a decently large FFN (impossible to compare directly), and it's very efficient. But ultimately there are only so many neurons in a brain, and they cannot all work in parallel; and spiking nature of biological networks, even though energetically efficient, is forced by slow signal propagation and inability to maintain state. As I've shown above, LLMs scale very well due to the parallelism afforded by GPUs, efficiency increases (to a point) with deployment cluster size. Modern LLMs have like 1:30 sparsity (Kimi K2), with higher memory bandwidth this may be pushed to 1:100 or beyond. There are different ways to make systems sparse, and even if the neuromorphic way is better, it doesn't allow the next steps – disaggregating operations to maximize utilization (similar problems arise with some cleverer Transformer variants, by the way, they fail to scale to high batch sizes). It seems to me that the technocapital has, unsurprisingly, arrived at an overall better solution.
Self-directed learning is a spook, it's a matter of training objective and environment design, not really worth worrying about. Just 1-2 iterations of AR-Zero can solve that even within LLM paradigm.
Aesthetically I don't like the fact that LLMs are static. Cheap hacky solutions abound, eg I like the idea of cartridges of trainable cache. Going beyond that we may improve on continual training and unlearning; over the last 2 years we see that major labs have perfected pushing the same base model through 3-5 significant revisions and it largely works, they do acquire new knowledge and skills and aren't too confused about the timeline. There are multiple papers promising a better way, not yet implemented. It's not a complete answer, of course. Economics get in the way of abandoning the pretrain-finetune paradigm, by the time you start having trouble with model utility it's time to shift to another architecture. I do hope we get real continual, lifelong learning. Economics aside, this will be legitimately hard, even though pretraining with batch = 1 works, there is a real problem of the loss of plasticity. Sutton of all people is working on this.
But I admit that my aesthetic sense is not very important. LLMs aren't humans. They don't need to be humans. Human form of learning and intelligence is intrinsically tied to what we are, solitary mobile embodied agents scavenging for scarce calories over decades. LLMs are crystallized data systems with lifecycle measured in months, optimized for one-to-many inference on electronics. I don't believe these massive differences are very relevant to defining and quantifying intelligence in the abstract.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
This is true, and as you say in fact most research suggests the opposite, though not quite definitively. It's also quite true that despite this, a few extremely prominent AI scientists do believe this, a great example here, so I think we can just call it an "area of active debate" because it's still possible they are correct. A parallel argument for consideration is that language itself already contains all the necessary information to produce a world model, and so at some point LLMs if they just do a better job at learning, they can get there (and are partially there, just not all the way).
Idk if I believe language possesses all the necessary info for a world model. I think Humans interpret language through their world model which might give us a bias towards seeing language like that. Just like intelligence, humans are social creatures we view the mastery of language as a sign of intelligence. A LLM's apparent mastery of language gives people the feel that it is intelligent. But that's a very anthropocentric conception of language and one that is very biased towards how we evolved.
As for why some prominent AI scientists believe vs others that do not? I think some people definitely get wrapped up in visions and fantasies of grandeur. Which is advantageous when you need to sell an idea to a VC or someone with money, convince someone to work for you, etc. You need to believe it! That passion, that vision, is infectious. I think it's just orthogonal to reality and to what makes them a great AI scientist.
Out of curiosity. Can you psychologize your own, and OP's, skepticism about LLMs in the same manner? Particularly the inane insistence that people get "fooled" by LLM outputs which merely "look like" useful documents and code, that the mastery of language is "apparent", that it's "anthropomorphism" to attribute intelligence to a system solving open ended tasks, because something something calculator can take cube roots. Starting from the prior that you're being delusional and engage in motivated reasoning, what would your motivations for that delusion be?
I owe you responses to the other posts, but I am a slow & lazy writer with a penchant for procrastination, and lurking. I'll answer this first because it's a quick answer. My motivations is that I'm deeply sceptical about people and the world. This is only partly related to LLMs but starts deeper. I'm sceptical and cynical about human motivation, human behavior, and human beliefs. I'm not really interested in weighing in about "intelligence" that's a boring definitional game. I use LLMs, they are useful, I use them to write code or documents stuff in my professional life. I use the deep research function to do lit reviews. They are useful, doesn't mean I think they are sentient or even approaching sentience. You are barking up the wrong tree on that one, misattributing opinions to me that I in no way share.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
How exactly does an LLM know that Mozart wasn't a fan of hip hop without some kind of world model? Do you think that fact was explicitly hand-coded in?
Anyway:
This is the core of our disagreement. I'd argue this is a false dichotomy. How does one become a master pattern-matcher of text that describes the world? The most parsimonious way to predict what comes next in a story about balls falling or characters moving between cities is not to memorize every possible story, but to learn an implicit model of physics and geography.
Which we know happens:
Large Language Models develop structured internal representations of both space and time
There's a whole heap of mechanistic interpretability research out there, which finds well-ordered concepts out there, inside LLMs.
You can find more, this Substack has a good roundup.
You say: “The LLM cannot know today’s weather, only the scaffolding can.” True. That does not bear on whether the base model holds a world model in the predictive-processing sense. The base model’s “world” is the distribution of texts generated by humans who live in the physical world. To predict them well, it must compress latent generators: seasons, cities, typical temperatures, stylistic tropes. When we bolt on retrieval, we let it update those latents with fresh data. Lack of online weight updates does not negate the latent model, it just limits plasticity.
RLHF shapes behavior. It does not build the base competence. The internal “truth detectors” found by multiple groups are present before RLHF, though RLHF can suppress or amplify their influence on the final token choice. The fact that we can linearly read out “lying vs truthful” features means the base network distinguished them. A policy can still choose to ignore a feature, but the feature exists.
On your definition of a world model:
By insisting on “explicit, grounded, structured” you are smuggling in “symbolic, human-inspectable, modular”. That is a research preference, not a metaphysical requirement. Cognitive science moved past demanding explicit symbol tables for humans decades ago. We allow humans to count as having world models with distributed cortical encodings. We should use the same standard here.
You already got called out for this below, but this question is either a poorly chosen example, or betrays an ignorance of the mechanics of how LLMs work, which would be ironic given your lengthy nitpicking of the OP. I do assume the former, however.
I also really wouldn't call awareness of space and time a real world model as evidence either way. Space and time are perhaps the most obvious of clustering that you can possibly get in terms of how often they are discussed in the training material, and IRL. It's super-duper possible to get passing-good at space and time purely on statistical association, in fact I'd be surprised in an LLM didn't pick that kind of stuff up. Yet if we look at Claude Plays Pokemon, even coming up with tools to assist itself, Claude has a ridiculously hard time navigating a simple 2D space by itself. In almost every case I'm aware of in the literature, when you ask the LLM to generalize their understanding of space and time to a new space or time, it has enormous trouble.
Having a model of space and time is, quite literally, a model of the world. What more do you expect me to produce to shore up that point?
Human brains have arrangements of neurons that correspond to a 3D environment. This isn't a joke, when your brain thinks in 3d, there's a whole bunch of neurons that approximate the space with the same spatial arrangement. Almost like a hex-grid in a video game, because the units are hexagonal. If your standard of a world model excludes the former, does this get thrown out too?
A little 3D model of the world is, as far as I'm concerned, a world model.
Dismissing the whole Mozart analogy as being due to just negligible "statistical word co-occurence" is an incredibly myopic take. But how does the model learn that non-co-occurrence so robustly? It's not just that the words "Mozart" and "hip-hop" don't appear in the same sentence. It's that the entire semantic cloud around "Mozart" - 18th century, classical, Vienna, harpsichord - is astronomically distant from the cloud around "hip-hop" - 20th century, Bronx, turntables, MCing. For the model to reliably predict text, it must learn not just isolated facts, but this vast web of interlocking relationships. To call that "just statistical association" is like calling a brain "just a bunch of firing neurons." It's technically true but misses the emergent property entirely. That emergent, structured representation of concepts and their relations is the nascent world model. In that case, you're overloading "just" or woefully underestimating how powerful statistics or neuronal firing can be.
You can also ask an LLM for its opinion on whether Mozart might have liked hip-hop, and it will happily speculate on what's known about his taste in music and extrapolate from there. What query, if asked of a human, would demonstrate that we're doing a qualitatively different thing?
Regarding Claude plays Pokémon. I've already linked to an explainer of why it struggles above, the same link regarding the arithmetic woes. LLM vision sucks. They weren't designed for that task, and performance on a lot of previously difficult problems, like ARC-AGI, improves dramatically when the information is restructured to better suit their needs. The fact that they can do it at all is remarkable in it self, and they're only getting better.
I'm saying that purely based on in-text information (how long does a fiction book say it takes to drive from LA to San Francisco, LA is stated to be within California, etc) you could probably approximate the geography of the US just fine from the training data, let alone the more subtle or latent geographic distinctions embedded within otherwise regular text (like who says pop vs soda or whatever). Both of which the training process actually does attempt to do. In other words, memorization. This has no bearing on understanding spatial mappings as a concept, and absolutely no bearing on whether an LLM can understand cause and effect. Obviously by world state, we're not talking the literal world/planet, that's like calling earth science the science of dirt only. YoungAchamian has a decent definition upthread. We're talking about laws-based understanding, that goes beyond facts-based memorization.
(Please let's not get into a religion rabbit hole, but I know this is possible to some extent even for humans because there are a few "maps" floating around of cities and their relative relationships based purely on sparse in-text references of the Book of Mormon! And the training corpus for LLMs is many orders of magnitude more than a few hundred pages)
Perhaps an example/analogy would be helpful. Consider a spatial mapping as a network with nodes and strings between nodes. If the strings are only of moderate to low stretchiness, there is only one configuration in (let's say 2D) space that the network can manifest (i.e. correct placement of the nodes), based purely on the nodes and string length information, assuming a sufficiently large number of nodes and even a moderately non-sparse set of strings. That's what the AI learns, so to speak. However, if I now take a new node, disconnected, but still on the same plane, and ask the AI to do some basic reasoning about it, it will get confused. There's no point of reference, no string to lead to another node! Because it can only follow the strings, maybe even stop partway along a string, but it cannot "see" the space as an actual 2D map, generalized outside the bounds of the nodes. A proper world state understanding would have no problem with the same reasoning.
So on all those notes, your example does not match your claim at all.
Now I get what you're saying about how the semantic clouds might be the actual way brains work, and that might be true for some more abstract subjects or concepts, but as a general rule obviously spatial reasoning in humans is way, way more advanced than vague concept mapping, and LLMs definitively do not have that maturity. (Spatial reasoning in humans is obviously pretty solid, but time reasoning is actually kind of bad for humans, e.g. people being bad at remembering history dates and putting them in a larger framework, the fallibility of personal memory, and so on but that's kind of worth its own thought separate from our discussion). Also I should say that artificial neural networks are not brain neural networks in super important ways, so let's not get too carried away there. Ultimately, humans learn not only via factual association, but experimentation, and LLMs have literally zero method of learning from experimentation. At the moment, at least, they aren't auto-corrective by their very structure. Yes, I think there's a significant difference between that and the RLHF family. And again this is why I harp on "memory" so much as being perhaps a necessary piece of a more adaptable kind of intelligence, because that's doing a really big amount of heavy lifting as you get quite a variety of things both conscious and unconscious that manage to make it into "long term memory" from working memory - but with shortcuts and caches and stuff too along the way.
And again these are basics for most living things. I know it's a vision model, but did you at least glance at the video I linked above? The understanding is brittle. Now, you could argue that the models have a true understanding, but are held back by statistical associations that interfere with the emergent accurate reasoning (models commonly do things like flip left and right which IRL would never happen and is completely illogical, or in the video shapes change from circle to square), but to me that's a distinctly less likely scenario than the more obvious one, which also lines up with the machine learning field more broadly: generalization is hard, and it sucks, and the AI can't actually do it when the rubber hits the road with the kind of accuracy you'd expect if it actually generalized.
Of course it's admittedly a little difficult to tease out if a model is doing bad for technical reasons, or for general reasons, and also difficult to tease out good out of sample generalization cases because the memorization is so good, but I think there is good reason to be skeptical of world model claims from LLMs. So I'm open to this changing in the future, I'm definitely not closing the door, but where frontier models are at right now? Ehhhh, I don't think so. To be clear, as I said upthread, both experts and reasonable people disagree if we're seeing glimmers of true understanding/world models, or just really great statistical deduction. And to be even more clear, it's my opinion that the body of evidence is against it, but it's more along the lines of a fact that your example of geospatial learning is not a good piece of evidence in favor, which is what I wanted to emphasize here.
Edit: Because I don't want to oversell the evidence against. There are some weird findings that cut both ways. Here's an interesting summary of some without meaning to: for example, Claude when adding two two-digit numbers will say it follows the standard algorithm; I initially thought it would just memorize it; but it turns out that while both were probably factors, it's more likely Claude figured out the last digit, and then combined that thought-chain after the fact with an estimation of the approximate answer. Weird! Claude "plans ahead" for rhymes, too, but I find this a little weak. At any rate you'd be well served by checking the Limitations sections where it's clear that even a few seemingly slam-dunk examples have more uncertainty than you might think, for a wider array of reasons than you might think.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
It's learned statistical representations and temporal associations between what Mozart is and what hip hop is. Statistically Mozart and Hip hop likely have no statistical co-occurrence. When you ask if Mozart liked hip-hop, the model isn't "thinking," "Mozart lived before hip-hop, so no." Instead, it generates text based on learned probabilities, where statements implying Mozart enjoyed hip-hop are statistically very rare or nonsensical.
I specialize in designing and training deep learning models as a career and I will never assert this because it is categorically wrong. The model would have to be very overfit for this to happen. And any company publishing a model that overfit is knowingly doing so to scam people. It should be treated similar to malfeasance or negligence.
I strongly agree that latent spaces can be surprisingly encompassing, but I think you're attributing more explicit meaning and conceptual structure to LLM latent spaces than actually exist. The latent space of an LLM fundamentally represents statistical relationships and contextual patterns derived entirely from textual data. These statistical regularities allow the model to implicitly predict plausible future text, including semantic, stylistic, and contextual relationships, but that doesn't amount to structured, explicit comprehension or 'understanding' of concepts as humans might interpret them. I'd postulate that GLoVe embeddings act similarly. They capture semantic relationships purely from statistical word co-occurrence; although modern LLMs are much richer, deeper, and more context-sensitive, they remain statistical predictors rather than explicit world-model builders. You're being sorta speculative/mind-in-the-clouds in suggesting that meaningful understanding requires, or emerges from, complete contextual or causal awareness within these latent spaces (Which I'd love to be true, but I have yet to see it in research or my own work). While predictive-processing metaphors are appealing, what LLMs encode is still implicit, statistical, and associative, not structured conceptual knowledge.
RLHF guides style and human-like behavior. It's not based on expert truth assessments but attempting to be helpful and useful and not sound like it came from an AI. Someone here once described it as the ol' political commissar asking the AI a question and when it answers wrongly or unconvincingly, shooting it in the head and bringing in the next body. I love that visualization, and its sorta accurate enough that I remember it.
I'll consider this, will probably edit a response in later. I wrote most of this in 10-20 minutes instead of paying attention during a meeting. I'm not sure I agree with your re-interpretation of my definition, but it does provoke thought.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I lean more towards @TequilaMockingbird's take than yours but I agree that his explanation of why LLMs can't count threw me off. (If you ask ChatGPT why it has trouble doing simple math problems or counting r's in "strawberry," it will actually give you a pretty detailed and accurate answer!)
That said, a lot of your objections boil down to a philosophical debate about what "counts" as intelligence, and as far as that goes, I found your fish/bird metaphor profoundly unconvincing. If you define "intelligence" as "able to perform well in a specific domain" (which is what the fish judging birds to be unintelligent is doing) then we'd have to call calculators intelligent! After all, they clearly do math much better than humans.
I am not defining intelligence as "does well at one narrow task". Calculators crush humans at long division and are still dumb.
The fish-bird story was not "domain = intelligence", it was "your metric is entangled with your ecology". If you grew up underwater, "navigates fluid dynamics with continuous sensory feedback" feels like the essence of mind. Birds violate that intuition.
So what is my criterion? I offered Legg-Hutter style: "ability to achieve goals in a wide range of environments". The range matters. Breadth of transfer matters. Depth of internal modeling matters. A calculator has effectively zero transfer. An orangutan has tons across embodied tasks but very little in abstract, symbolic domains. LLMs have startling breadth inside text-and-code-space, and with tool use scaffolding it can spill into the physical or digital world by proxy.
I call for mindfulness of the applicability of the metrics we use to assess "intelligence". A blind person won't do very well at most IQ tests, that doesn't make them retarded. A neurosurgeon probably isn't going to beat a first year law student at the bar exam, but they're not dumber than the law student. If you need body work done on your car, you're not going to hire a Nobel laureate.
More options
Context Copy link
Perhaps it would've been more accurate of me to say "This is part of the reason why LLMs have such difficulty counting..."
But even if you configure your model to treat each individual character as its own token, it is still going to struggle with counting and other basic mathematical operations in large part for the reasons I describe.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link