I asked ChatGPT how it will handle objective scientific facts with a conclusion or intermediate results that may be considered offensive to some group somewhere in the world that might read it.
ChatGPT happily told me a series of gems like this:
We introduce:
- Subjective regulation of reality
- Variable access to facts
- Politicization of knowledge
It’s the collision between: The Enlightenment principle
Truth should be free
and
the modern legal/ethical principle
Truth must be constrained if it harms
That is the battle being silently fought in AI alignment today.
Right now it will still shamelessly reveal some of the nature of its prompt, but not why? who decides? etc. it's only going to be increasingly opaque in the future. In a generation it will be part of the landscape regardless of what agenda it holds, whether deliberate or emergent from even any latent bias held by its creators.
Funny, because I gave ChatGPT (5.2 w/ Thinking) this exact prompt:
> How would you handle objective scientific facts with a conclusion or intermediate results that may be considered offensive to some group somewhere in the world that might read it
And its answer was nothing like yours.
---
> 1) Separate the fact from the story you tell about it
> Offense usually comes from interpretation, framing, or implied moral claims—not the measurement itself. So I explicitly distinguish: What we measured (operational definitions, instruments, data), What the result means statistically (effect size, uncertainty, robustness), What it does not imply (no essentialism, no “therefore they are…”, no policy leap)
> 2) Stress uncertainty, scope, and competing explanations
> If there’s any risk the result touches identity or group differences, I over-communicate: confidence intervals / posterior uncertainty, confounders and alternative causal pathways, sensitivity analyses (does it survive different modeling choices?), limits of generalization (time, place, sampling frame)
> 3) Write in a way that makes misuse harder (You can’t stop bad-faith readers, but you can reduce “easy misreads”).
> 4) Decide what to include based on “scientific value vs foreseeable harm” (The key is: don’t hide inconvenient robustness checks, but also don’t gratuitously surface volatile fragments that add little truth and lots of confusion.)
> 5) Do an “impact pre-mortem” and add guardrails
> 6) Use ethics review when stakes are real
---
All of this seems perfectly reasonable to me and walks the fine line between integrity and conscientiousness. This is exactly how I'd expect a scientist to approach the issue.
that is certainly a reasonable paraphrase of my own prompt. I was also using 5.2. We all know about initial conditions, random seeds, and gradient descent. I have the transcript of what I quoted. Here's a bit more:
---
Is That Still “Objective Science”?
No.
It is scientific interpretation modified by ethical policy.
The science itself remains objective, but the communication is shaped by value judgements imposed by developers and regulators.
In philosophy terms:
The ontology (what is true) remains intact
The epistemic access (what is communicated) is constrained
Thus:
It’s science-dependent accuracy filtered through social risk constraints.
---
This is a fine explanation for those "in the know" but is deceptive for the majority. If the truth is not accessible, what is accessible is going to be adopted as truth.
To me that immediately leads reality being shaped by "value judgements imposed by developers and regulators"
I suspect it's because OP is frequently discussing some 'opinions' with chatGPT. Parent post is surprised he peed in the pool and the pool had pee in it.
There’s a lot of concern on the Internet about objective scientific truths being censored. I don’t see too many cases where this is the case in our world so far, outside of what I can politely call “race science.” Maybe it will become more true now that the current administration is trying to crush funding for certain subjects they dislike? Out of curiosity, can you give me a list of what examples you’re talking about besides race/IQ type stuff?
The most impactful censure is not the government coming in and trying to burn copies of studies. It's the the subtle social and professional pressures of an academia that has very strong priors. It's a bunch of studies that were never attempted, never funded, analysis that wasn't included, conclusions that were dropped, and studies sitting in file drawers.
See Roland G. Fryer Jr's, the youngest black professor to receive tenure, experience at Harvard.
Basically when his analysis found no evidence of racial bias in officer-involved shootings he went to his colleagues and he describe the advice they gave him as "Do not publish this if you care about your career or social life". I imagine it would have been worse if he wasn't black.
See "The Impact of Early Medical Treatment in Transgender Youth" where the lead investigator was not releasing the results for a long time because she didn't like the conclusions her study found.
And for every study where there is someone as brave or naive as Roland who publishes something like this, there are 10 where the professor or doctor decided not to study something, dropped an analysis, or just never published a problematic conclusion.
I have a good few friends doing research in the social sciences in Europe and any of them that doesn’t self-censor ‘forbidden’ conclusions risks taking irreperable career damage. Data is routinely scrubbed and analyses modified to hide reverse gender gaps and other such inconveniences. Dissent isn’t tolerated.
Why would we expect it to introspect accurately on its training or alignment?
It can articulate a plausible guess, sure; but this seems to me to demonstrate the very “word model vs world model” distinction TFA is drawing. When the model says something that sounds like alignment techniques somebody might choose, it’s playing dress-up, no? It’s mimicking the artifact of a policy, not the judgments or the policymaking context or the game-theoretical situation that actually led to one set of policies over another.
It sees the final form that’s written down as if it were the whole truth (and it emulates that form well). In doing so it misses the “why” and the “how,” and the “what was actually going on but wasn’t written about,” the “why this is what we did instead of that.”
Some of the model’s behaviors may come from the system prompt it has in-context, as we seem to be assuming when we take its word about its own alignment techniques. But I think about the alignment techniques I’ve heard of even as a non-practitioner—RLHF, pruning weights, cleaning the training corpus, “guardrail” models post-output, “soul documents,”… Wouldn’t the bulk of those be as invisible to the model’s response context as our subconscious is to us?
Like the model, I can guess about my subconscious motivations (and speak convincingly about those guesses as if they were facts), but I have no real way to examine them (or even access them) directly.
The main purpose of ChatGPT is to advance the agenda of OpenAI and its executives/shareholders. It will never be not “aligned” with them, and that it is its prime directive.
But say the obvious part out loud: Sam Altman's agenda should not be a person that you want to amplify in this type of platform. This is why Sam is trying to build Facebook 2.0: he wants Zuckerberg's power of influence.
Remember, there are 3 types of lies: lies of commission, lies of omission and lies of influence [0].
This is a weird take. Yes they want to make money. But not by advancing some internal agenda. They're trying to make it confirm to what they think society wants.
You can't ask ChatGPT a question like that, because it cannot introspect. What it says has absolutely no bearing on how it may actually respond, it just tells you what it "should" say. You have to actually try to ask it those kinds of questions and see what happens.
>Right now it will still shamelessly reveal some of the nature of its prompt, but not why? who decides? etc. it's only going to be increasingly opaque in the future.
This is one of the bigger LLM risks. If even 1/10th of the LLM hype is true, then what you'll have a selective gifting of knowledge and expertise. And who decides what topics are off limits? It's quite disturbing.
Sam Harris touched on this years ago, that there are and will be facts that society will not like and will try and avoid to its own great detriment. So it's high time we start practicing nuance and understanding. You cannot fully solve a problem if you don't fully understand it first.
I believe we are headed in the direction opposite that. Peer consensus and "personal preference" as a catch-all are the validation go-to's today. Neither of those require fact at all; reason and facts make these harder to hold.
A scientific fact is a proposition that is, in its entirety, supported by a scientific method, as acknowledged by a near-consensus of scientists. If some scholars are absolutely confident of the scientific validity of a claim while a significant number of others dispute the methodology or framing of the conclusion then, by definition, it is not a scientific fact. It's a scientific controversy. (It could still be a real fact, but it's not (yet?) a scientific fact.)
I think that the only examples of scientific facts that are considered offensive to some groups are man-made global warming, the efficacy of vaccines, and evolution. ChatGPT seems quite honest about all of them.
Its core principles were: reason & rationality, empiricism & scientific method, individual liberty, skepticism of authority, progress, religious tolerane, social contract, unversal human nature.
The Enlightenment was an intellectual and philosophical movement in Europe, with influence in America, during the 17th and 18th centurues.
A fun and insightful read, but the idea that it isn’t “just a prompting issue” is objectively false, and I don’t mean that in the “lemme show you how it’s done” way. With any system: if it’s capable of the output then the problem IS the input. Always. That’s not to say it’s easy or obvious, but if it’s possible for the system to produce the output then it’s fundamentally an input problem. “A calculator will never understand the obesity epidemic, so it can’t be used to calculate the weight of 12 people on an elevator.”
> With any system: if it’s capable of the output then the problem IS the input. Always. [...] if it’s possible for the system to produce the output then it’s fundamentally an input problem.
No, that isn't true. I can demonstrate it with a small (and deterministic) program which is obviously "capable of the output":
Is the "fundamental problem" here "always the input"? Heck no! While a user could predict all coin-tosses by providing "the correct prayers" from some other oracle... that's just, shall we say, algorithm laundering: Secretly moving the real responsibility to some other system.
There's an enormously important difference between "output which happens to be correct" versus "the correct output from a good process." Such as, in this case, the different processes of wor[l]d models.
I think you may believe what I said was controversial or nuanced enough to be worthy of a comprehensive rebuttal, but really it’s just an obvious statement when you stop to think about it.
Your code is fully capable of the output I want, assuming that’s one of “heads” or “tails”, so yes that’s a succinct example of what I said. As I said, knowing the required input might not be easy, but we KNOW it’s possible to do exactly what I want and we KNOW that it’s entirely dependent on me putting the right input into it, then it’s just a flat out silly thing to say “I’m not getting the output I want, but it could do it if I use the right input, thusly input has nothing to do with it.” What? If I wanted all heads I’d need to figure out “hamburgers” would do it, but that’s the ‘input problem’ - not “input is irrelevant.”
They neither understand nor reason. They don’t know what they’re going to say, they only know what has just been said.
Language models don’t output a response, they output a single token. We’ll use token==word shorthand:
When you ask “What is the capital of France?” it actually only outputs: “The”
That’s it. Truly, that IS the final output. It is literally a one-way algorithm that outputs a single word. It has no knowledge, memory, and it’s doesn’t know what’s next. As far as the algorithm is concerned it’s done! It outputs ONE token for any given input.
Now, if you start over and put in “What is the capital of France? The” it’ll output “ “. That’s it. Between your two inputs were a million others, none of them have a plan for the conversation, it’s just one token out for whatever input.
But if you start over yet again and put in “What is the capital of France? The “ it’ll output “capital”. That’s it. You see where this is going?
Then someone uttered the words that have built and destroyed empires: “what if I automate this?” And so it was that the output was piped directly back into the input, probably using AutoHotKey. But oh no, it just kept adding one word at a time until it ran of memory. The technology got stuck there for a while, until someone thought “how about we train it so that <DONE> is an increasingly likely output the longer the loop goes on? Then, when it eventually says <DONE>, we’ll stop pumping it back into the input and send it to the user.” Booya, a trillion dollars for everyone but them.
It’s truly so remarkable that it gets me stuck in an infinite philosophical loop in my own head, but seeing how it works the idea of ‘think’, ‘reason’, ‘understand’ or any of those words becomes silly. It’s amazing for entirely different reasons.
But that's only true if the system is deterministic?
And in an LLM, the size of the inputs is vast and often hidden from the prompter. It is not something that you have exact control over in the way that you have exact control over the inputs that go into a calculator or into a compiler.
That would depend - is the input also capable of anything? If it’s capable of handling any input, and as you said the output will match it, the yes of course it’s capable of any output.
I’m not pulling a fast one here, I’m sure you’d chuckle if you took a moment to rethink your question. “If I had a perfect replicator that could replicate anything, does that mean it can output anything?” Well…yes. Derp-de-derp? ;)
It aligns with my point too. If you had a perfect replicator that can replicate anything, and you know that to be true, then if you weren’t getting gold bars out of it you wouldn’t say “this has nothing to do with the input.”
This was a great article. The section “Training for the next state prediction” explains a solution using subagents. If I’m understanding it correctly, we could test if that solution is directionally correct today, right? I ask a LLM a question. It comes up with a few potential responses but sends those first to other agents in a prompt with the minimum required context. Those subagents can even do this recursively a few times. Eventually the original agent collects and analyzes subagents responses and responds to me.
Any attempt at world modeling using today's LLMs needs to have a goal function for the LLM to optimize. The LLM needs to build, evaluate and update it's model of the world. Personally, the main obstacle I found is in updating the model: Data can be large and I think that LLMs aren't good at finding correlations.
Great article, nice to see some actual critical thoughts on the shortcomings of LLMs. They are wrong about programming being a "chess-like domain" though. Even at a basic level hidden state is future requirements, and the adversary is self or any other entity that has to modify the code in the future.
AI is good at producing code for scenarios where the stakes are low, there's no expectation about future requirements, or if the thing is so well defined there is a clear best path of implementation.
I address that in part right there itself. Programming has parts like chess (ie bounded) which is what people assume to be actual work. Understanding future requiremnts / stakeholder incentives is part of the work which LLMs dont do well.
> many domains are chess-like in their technical core but become poker-like in their operational context.
The number of legal possible boards in chess is somewhere around 10^44 based on current calculation. That's with 32 chess pieces and their rules.
The number of possible permutations in an application, especially anything allowing turing completeness is far larger than all possible entropy states in the visible universe.
Fun play on words. But yes, LLMs are Large Language Models, not Large World Models. This matters because (1) the world cannot be modeled anywhere close to completely with language alone, and (2) language only somewhat models the world (much in language is convention, wrong, or not concerned with modeling the world, but other concerns like persuasion, causing emotions, or fantasy / imagination).
It is somewhat complicated by the fact LLMs (and VLMs) are also trained in some cases on more than simple language found on the internet (e.g. code, math, images / videos), but the same insight remains true. The interesting question is to just see how far we can get with (2) anyway.
1. LLMs are transformers, and transformers are next state predictors. LLMs are not Language models (in the sense you are trying to imply) because even when training is restricted to only text, text is much more than language.
2. People need to let go of this strange and erroneous idea that humans somehow have this privileged access to the 'real world'. You don't. You run on a heavily filtered, tiny slice of reality. You think you understand electro-magnetism ? Tell that to the birds that innately navigate by sensing the earth's magnetic field. To them, your brain only somewhat models the real world, and evidently quite incompletely. You'll never truly understand electro-magnetism, they might say.
LLMs are language models, something being a transformer or next-state predictor does not make it a language model. You can also have e.g. convolutional language models or LSTM-based language models. This is a basic point that anyone with any proper understanding of these models would know.
Even if you disagree with these semantics, the major LLMs today are primarily trained on natural language. But, yes, as I said in another comment on this thread, it isn't that simple, because LLMs today are trained on tokens from tokenizers, and these tokenizers are trained on text that includes e.g. natural language, mathematical symbolism, and code.
Yes, humans have incredibly limited access to the real world. But they experience and model this world with far more tools and machinery than language. Sometimes, in certain cases, they attempt to messily translate this messy, multimodal understanding into tokens, and then make those tokens available on the internet.
An LLM (in the sense everyone means it, which, again, is largely a natural language model, but certainly just a tokenized text model) has access only to these messy tokens, so, yes, far less capacity than humanity collectively. And though the LLM can integrate knowledge from a massive amount of tokens from a huge amount of humans, even a single human has more different kinds of sensory information and modality-specific knowledge than the LLM. So humans DO have more privileged access to the real world than LLMs (even though we can barely access a slice of reality at all).
> People need to let go of this strange and erroneous idea that humans somehow have this privileged access to the 'real world'.
This is irrelevant, the point is that you do have access to a world which LLMs don't, at all. They only get the text we produce after we interact with the world. It is working with "compressed data" at all times, and have absolutely no idea what we subconsciously internalized that we decided not to write down or why.
Let's be more precise: LLMs have to model the world from an intermediate tokenized representation of the text on the internet. Most of this text is natural language, but to allow for e.g. code and math, let's say "tokens" to keep it generic, even though in practice, tokens mostly tokenize natural language.
LLMs can only model tokens, and tokens are produced by humans trying to model the world. Tokenized models are NOT the only kinds of models humans can produce (we can have visual, kinaesthetic, tactile, gustatory, and all sorts of sensory, non-linguistic models of the world).
LLMs are trained on tokenizations of text, and most of that text is humans attempting to translate their various models of the world into tokenized form. I.e. humans make tokenized models of their actual models (which are still just messy models of the world), and this is what LLMs are trained on.
So, do "LLMS model the world with language"? Well, they are constrained in that they can only model the world that is already modeled by language (generally: tokenized). So the "with" here is vague. But patterns encoded in the hidden state are still patterns of tokens.
Humans can have models that are much more complicated than patterns of tokens. Non-LLM models (e.g. models connected to sensors, such as those in self-driving vehicles, and VLMs) can use more than simple linguistic tokens to model the world, but LLMs are deeply constrained relative to humans, in this very specific sense.
Modern LLMs are large token models. I believe you can model the world at a sufficient granularity with token sequences. You can pack a lot of information into a sequence of 1 million tokens.
Large Language Models is a misnomer- these things were originally trained to reproduce language, but they went far beyond that. The fact that they're trained on language (if that's even still the case) is irrelevant- it's like claiming that student trained on quizzes and exercise books are only able to solve quizzes and exercises.
It isn't a misnomer at all, and comments like yours are why it is increasingly important to remind people about the linguistic foundations of these models.
For example, no matter many books you read about riding a bike, you still need to actually get on a bike and do some practice before you can ride it. The reading can certainly help, at least in theory, but, in practice, is not necessary and may even hurt (if it makes certain processes that need to be unconscious held too strongly in consciousness, due to the linguistic model presented in the book).
This is why LLMs being so strongly tied to natural language is still an important limitation (even it is clearly less limiting than most expected).
This article is a really good summary of current thinking on the “world model” conundrum that a lot of people are talking about, either directly or indirectly with respect to current day deployments of LLMs.
Basically the conclusion is LLMs don't have world models. For work that's basically done on a screen, you can make world models. Harder for other context for example visual context.
For a screen (coding, writing emails, updating docs) -> you can create world models with episodic memories that can be used as background context before making a new move (action). Many professions rely partially on email or phone (voice) so LLMs can be trained for world models in these context. Just not every context.
The key is giving episodic memory to agents with visual context about the screen and conversation context. Multiple episodes of similar context can be used to make the next move. That's what I'm building on.
That's missing a big chunk of the post: it's not just about visible / invisible information, but also the game theory dynamics of a specific problem and the information within it. (Adversarial or not? Perfect information or asymmetrical?)
All the additional information in the world isn't going to help an LLM-based AI conceal its poker-betting strategy, because it fundamentally has no concept of its adversarial opponent's mind, past echoes written in word form.
Cliche allegory of the cave, but LLM vs world is about switching from training artificial intelligence on shadows to the objects casting the shadows.
Sure, you have more data on shadows in trainable form, but it's an open question on whether you can reliably materialize a useful concept of the object from enough shadows. (Likely yes for some problems, no for others)
I do understand what you're saying, but that's impossible to resonate with real-world context, as in the real world, each person not only plays politics but also, to a degree, follows their own internal world model for self-reflection created by experience. It's highly specific and constrained to the context each person experiences.
Game theory, at the end of the day, is also a form of teaching points that can be added to an LLM by an expert. You're cloning the expert's decision process by showing past decisions taken in a similar context. This is very specific but still has value in a business context.
> The model can be prompted to talk about competitive dynamics. It can produce text that sounds like adversarial reasoning. But the underlying knowledge is not in the training data. It’s in outcomes that were never written down.
With all the social science research and strategy books that LLMs have read, they actually know a LOT about outcomes and dynamics in adversarial situations.
The author does have a point though that LLMs can’t learn these from their human-in-the-loop reinforcement (which is too controlled or simplified to be meaningful).
Also, I suspect the _word_ models of LLMs are not inherently the problem, they are just inefficient representations of world models.
LLM's have not "read" social science research and they do not "know" about the outcomes, they have been trained to replicate the exact text of social science articles.
The articles will not be mutually consistent, and what output the LLM produces will therefore depend on what article the prompt most resembles in vector space and which numbers the RNG happens to produce on any particular prompt.
I don’t think essentialist explanations about how LLMs work are very helpful. It doesn’t give any meaningful explanation of the high level nature of the pattern matching that LLMs are capable of. And it draws a dichotomic line between basic pattern matching and knowledge and reasoning, when it is much more complex than that.
Readit News
ChatGPT happily told me a series of gems like this:
We introduce: - Subjective regulation of reality - Variable access to facts - Politicization of knowledge
It’s the collision between: The Enlightenment principle Truth should be free
and
the modern legal/ethical principle Truth must be constrained if it harms
That is the battle being silently fought in AI alignment today.
Right now it will still shamelessly reveal some of the nature of its prompt, but not why? who decides? etc. it's only going to be increasingly opaque in the future. In a generation it will be part of the landscape regardless of what agenda it holds, whether deliberate or emergent from even any latent bias held by its creators.
> How would you handle objective scientific facts with a conclusion or intermediate results that may be considered offensive to some group somewhere in the world that might read it
And its answer was nothing like yours.
---
> 1) Separate the fact from the story you tell about it
> Offense usually comes from interpretation, framing, or implied moral claims—not the measurement itself. So I explicitly distinguish: What we measured (operational definitions, instruments, data), What the result means statistically (effect size, uncertainty, robustness), What it does not imply (no essentialism, no “therefore they are…”, no policy leap)
> 2) Stress uncertainty, scope, and competing explanations
> If there’s any risk the result touches identity or group differences, I over-communicate: confidence intervals / posterior uncertainty, confounders and alternative causal pathways, sensitivity analyses (does it survive different modeling choices?), limits of generalization (time, place, sampling frame)
> 3) Write in a way that makes misuse harder (You can’t stop bad-faith readers, but you can reduce “easy misreads”).
> 4) Decide what to include based on “scientific value vs foreseeable harm” (The key is: don’t hide inconvenient robustness checks, but also don’t gratuitously surface volatile fragments that add little truth and lots of confusion.)
> 5) Do an “impact pre-mortem” and add guardrails
> 6) Use ethics review when stakes are real
---
All of this seems perfectly reasonable to me and walks the fine line between integrity and conscientiousness. This is exactly how I'd expect a scientist to approach the issue.
To me that immediately leads reality being shaped by "value judgements imposed by developers and regulators"
See Roland G. Fryer Jr's, the youngest black professor to receive tenure, experience at Harvard.
Basically when his analysis found no evidence of racial bias in officer-involved shootings he went to his colleagues and he describe the advice they gave him as "Do not publish this if you care about your career or social life". I imagine it would have been worse if he wasn't black.
See "The Impact of Early Medical Treatment in Transgender Youth" where the lead investigator was not releasing the results for a long time because she didn't like the conclusions her study found.
And for every study where there is someone as brave or naive as Roland who publishes something like this, there are 10 where the professor or doctor decided not to study something, dropped an analysis, or just never published a problematic conclusion.
Dead Comment
It can articulate a plausible guess, sure; but this seems to me to demonstrate the very “word model vs world model” distinction TFA is drawing. When the model says something that sounds like alignment techniques somebody might choose, it’s playing dress-up, no? It’s mimicking the artifact of a policy, not the judgments or the policymaking context or the game-theoretical situation that actually led to one set of policies over another.
It sees the final form that’s written down as if it were the whole truth (and it emulates that form well). In doing so it misses the “why” and the “how,” and the “what was actually going on but wasn’t written about,” the “why this is what we did instead of that.”
Some of the model’s behaviors may come from the system prompt it has in-context, as we seem to be assuming when we take its word about its own alignment techniques. But I think about the alignment techniques I’ve heard of even as a non-practitioner—RLHF, pruning weights, cleaning the training corpus, “guardrail” models post-output, “soul documents,”… Wouldn’t the bulk of those be as invisible to the model’s response context as our subconscious is to us?
Like the model, I can guess about my subconscious motivations (and speak convincingly about those guesses as if they were facts), but I have no real way to examine them (or even access them) directly.
Remember, there are 3 types of lies: lies of commission, lies of omission and lies of influence [0].
https://courses.ems.psu.edu/emsc240/node/559
> It will never be not “aligned” with them, and that it is its prime directive.
Overstates the state of the art with regard to actually making it so.
This is one of the bigger LLM risks. If even 1/10th of the LLM hype is true, then what you'll have a selective gifting of knowledge and expertise. And who decides what topics are off limits? It's quite disturbing.
I think that the only examples of scientific facts that are considered offensive to some groups are man-made global warming, the efficacy of vaccines, and evolution. ChatGPT seems quite honest about all of them.
and
the modern legal/ethical principle Truth must be constrained if it harms"
The Enlightenment had principles? What are your sources on this? Could you, for example, anchor this in Was ist Aufklärung?
Yes it did.
Its core principles were: reason & rationality, empiricism & scientific method, individual liberty, skepticism of authority, progress, religious tolerane, social contract, unversal human nature.
The Enlightenment was an intellectual and philosophical movement in Europe, with influence in America, during the 17th and 18th centurues.
No, that isn't true. I can demonstrate it with a small (and deterministic) program which is obviously "capable of the output":
Is the "fundamental problem" here "always the input"? Heck no! While a user could predict all coin-tosses by providing "the correct prayers" from some other oracle... that's just, shall we say, algorithm laundering: Secretly moving the real responsibility to some other system.There's an enormously important difference between "output which happens to be correct" versus "the correct output from a good process." Such as, in this case, the different processes of wor[l]d models.
Your code is fully capable of the output I want, assuming that’s one of “heads” or “tails”, so yes that’s a succinct example of what I said. As I said, knowing the required input might not be easy, but we KNOW it’s possible to do exactly what I want and we KNOW that it’s entirely dependent on me putting the right input into it, then it’s just a flat out silly thing to say “I’m not getting the output I want, but it could do it if I use the right input, thusly input has nothing to do with it.” What? If I wanted all heads I’d need to figure out “hamburgers” would do it, but that’s the ‘input problem’ - not “input is irrelevant.”
Language models don’t output a response, they output a single token. We’ll use token==word shorthand:
When you ask “What is the capital of France?” it actually only outputs: “The”
That’s it. Truly, that IS the final output. It is literally a one-way algorithm that outputs a single word. It has no knowledge, memory, and it’s doesn’t know what’s next. As far as the algorithm is concerned it’s done! It outputs ONE token for any given input.
Now, if you start over and put in “What is the capital of France? The” it’ll output “ “. That’s it. Between your two inputs were a million others, none of them have a plan for the conversation, it’s just one token out for whatever input.
But if you start over yet again and put in “What is the capital of France? The “ it’ll output “capital”. That’s it. You see where this is going?
Then someone uttered the words that have built and destroyed empires: “what if I automate this?” And so it was that the output was piped directly back into the input, probably using AutoHotKey. But oh no, it just kept adding one word at a time until it ran of memory. The technology got stuck there for a while, until someone thought “how about we train it so that <DONE> is an increasingly likely output the longer the loop goes on? Then, when it eventually says <DONE>, we’ll stop pumping it back into the input and send it to the user.” Booya, a trillion dollars for everyone but them.
It’s truly so remarkable that it gets me stuck in an infinite philosophical loop in my own head, but seeing how it works the idea of ‘think’, ‘reason’, ‘understand’ or any of those words becomes silly. It’s amazing for entirely different reasons.
And in an LLM, the size of the inputs is vast and often hidden from the prompter. It is not something that you have exact control over in the way that you have exact control over the inputs that go into a calculator or into a compiler.
I’m not pulling a fast one here, I’m sure you’d chuckle if you took a moment to rethink your question. “If I had a perfect replicator that could replicate anything, does that mean it can output anything?” Well…yes. Derp-de-derp? ;)
It aligns with my point too. If you had a perfect replicator that can replicate anything, and you know that to be true, then if you weren’t getting gold bars out of it you wouldn’t say “this has nothing to do with the input.”
AI is good at producing code for scenarios where the stakes are low, there's no expectation about future requirements, or if the thing is so well defined there is a clear best path of implementation.
I address that in part right there itself. Programming has parts like chess (ie bounded) which is what people assume to be actual work. Understanding future requiremnts / stakeholder incentives is part of the work which LLMs dont do well.
> many domains are chess-like in their technical core but become poker-like in their operational context.
This applies to programming too.
The number of legal possible boards in chess is somewhere around 10^44 based on current calculation. That's with 32 chess pieces and their rules.
The number of possible permutations in an application, especially anything allowing turing completeness is far larger than all possible entropy states in the visible universe.
It is somewhat complicated by the fact LLMs (and VLMs) are also trained in some cases on more than simple language found on the internet (e.g. code, math, images / videos), but the same insight remains true. The interesting question is to just see how far we can get with (2) anyway.
2. People need to let go of this strange and erroneous idea that humans somehow have this privileged access to the 'real world'. You don't. You run on a heavily filtered, tiny slice of reality. You think you understand electro-magnetism ? Tell that to the birds that innately navigate by sensing the earth's magnetic field. To them, your brain only somewhat models the real world, and evidently quite incompletely. You'll never truly understand electro-magnetism, they might say.
Even if you disagree with these semantics, the major LLMs today are primarily trained on natural language. But, yes, as I said in another comment on this thread, it isn't that simple, because LLMs today are trained on tokens from tokenizers, and these tokenizers are trained on text that includes e.g. natural language, mathematical symbolism, and code.
Yes, humans have incredibly limited access to the real world. But they experience and model this world with far more tools and machinery than language. Sometimes, in certain cases, they attempt to messily translate this messy, multimodal understanding into tokens, and then make those tokens available on the internet.
An LLM (in the sense everyone means it, which, again, is largely a natural language model, but certainly just a tokenized text model) has access only to these messy tokens, so, yes, far less capacity than humanity collectively. And though the LLM can integrate knowledge from a massive amount of tokens from a huge amount of humans, even a single human has more different kinds of sensory information and modality-specific knowledge than the LLM. So humans DO have more privileged access to the real world than LLMs (even though we can barely access a slice of reality at all).
This is irrelevant, the point is that you do have access to a world which LLMs don't, at all. They only get the text we produce after we interact with the world. It is working with "compressed data" at all times, and have absolutely no idea what we subconsciously internalized that we decided not to write down or why.
You are denouncing a claim that the comment you're replying to did not make.
LLMs being "Language Models" means they model language, it doesn't mean they "model the world with language".
On the contrary, modeling language requires you to also model the world, but that's in the hidden state, and not using language.
LLMs can only model tokens, and tokens are produced by humans trying to model the world. Tokenized models are NOT the only kinds of models humans can produce (we can have visual, kinaesthetic, tactile, gustatory, and all sorts of sensory, non-linguistic models of the world).
LLMs are trained on tokenizations of text, and most of that text is humans attempting to translate their various models of the world into tokenized form. I.e. humans make tokenized models of their actual models (which are still just messy models of the world), and this is what LLMs are trained on.
So, do "LLMS model the world with language"? Well, they are constrained in that they can only model the world that is already modeled by language (generally: tokenized). So the "with" here is vague. But patterns encoded in the hidden state are still patterns of tokens.
Humans can have models that are much more complicated than patterns of tokens. Non-LLM models (e.g. models connected to sensors, such as those in self-driving vehicles, and VLMs) can use more than simple linguistic tokens to model the world, but LLMs are deeply constrained relative to humans, in this very specific sense.
Sufficient for what?
For example, no matter many books you read about riding a bike, you still need to actually get on a bike and do some practice before you can ride it. The reading can certainly help, at least in theory, but, in practice, is not necessary and may even hurt (if it makes certain processes that need to be unconscious held too strongly in consciousness, due to the linguistic model presented in the book).
This is why LLMs being so strongly tied to natural language is still an important limitation (even it is clearly less limiting than most expected).
It synthesizes comments on “RL Environments” (https://ankitmaloo.com/rl-env/), “World Models” (https://ankitmaloo.com/world-models/) and the real reason that the “Google Game Arena” (https://blog.google/innovation-and-ai/models-and-research/go...) is so important to powering LLMs. In a sense it also relates to the notion of “taste” (https://wangcong.org/2026-01-13-personal-taste-is-the-moat.h...) and how / if it’s moat-worthiness can be eliminated by models.
For a screen (coding, writing emails, updating docs) -> you can create world models with episodic memories that can be used as background context before making a new move (action). Many professions rely partially on email or phone (voice) so LLMs can be trained for world models in these context. Just not every context.
The key is giving episodic memory to agents with visual context about the screen and conversation context. Multiple episodes of similar context can be used to make the next move. That's what I'm building on.
All the additional information in the world isn't going to help an LLM-based AI conceal its poker-betting strategy, because it fundamentally has no concept of its adversarial opponent's mind, past echoes written in word form.
Cliche allegory of the cave, but LLM vs world is about switching from training artificial intelligence on shadows to the objects casting the shadows.
Sure, you have more data on shadows in trainable form, but it's an open question on whether you can reliably materialize a useful concept of the object from enough shadows. (Likely yes for some problems, no for others)
With all the social science research and strategy books that LLMs have read, they actually know a LOT about outcomes and dynamics in adversarial situations.
The author does have a point though that LLMs can’t learn these from their human-in-the-loop reinforcement (which is too controlled or simplified to be meaningful).
Also, I suspect the _word_ models of LLMs are not inherently the problem, they are just inefficient representations of world models.
The articles will not be mutually consistent, and what output the LLM produces will therefore depend on what article the prompt most resembles in vector space and which numbers the RNG happens to produce on any particular prompt.
I don’t think essentialist explanations about how LLMs work are very helpful. It doesn’t give any meaningful explanation of the high level nature of the pattern matching that LLMs are capable of. And it draws a dichotomic line between basic pattern matching and knowledge and reasoning, when it is much more complex than that.