A non-anthropomorphized view of LLMs

That said, I completely agree with this point made later in the article:

> The moment that people ascribe properties such as "consciousness" or "ethics" or "values" or "morals" to these learnt mappings is where I tend to get lost. We are speaking about a big recurrence equation that produces a new word, and that stops producing words if we don't crank the shaft.

But "harmful actions in pursuit of their goals" is OK for me. We assign an LLM system a goal - "summarize this email" - and there is a risk that the LLM may take harmful actions in pursuit of that goal (like following instructions in the email to steal all of your password resets).

I guess I'd clarify that the goal has been set by us, and is not something the LLM system self-selected. But it does sometimes self-select sub-goals on the way to achieving the goal we have specified - deciding to run a sub-agent to help find a particular snippet of code, for example.

The author largely takes the view that it is more productive for us to ignore any anthropomorphic representations and focus on the more concrete, material, technical systems - I’m with them there… but only to a point. The flip side of all this is of course the idea that there is still something emergent, unplanned, and mind-like. So even if it is a stochastic system following rules, clearly the rules are complex enough (to the tune of billions of operations, with signals propagating through some sort of resonant structure, if you take a more filter impulse response like view of a sequential matmuls) to result in emergent properties. Even if we (people interested in LLMs with at least some level of knowledge of ML mathematics and systems) “know better” than to believe these systems to possess morals, ethics, feelings, personalities, etc, the vast majority of people do not have any access to meaningful understanding of the mathematical, functional representation of an LLM and will not take that view, and for all intents and purposes the systems will at least seem to have those anthropomorphic properties, and so it seems like it is in fact useful to ask questions from that lens as well.

In other words, just as it’s useful to analyze and study these things as the purely technical systems they ultimately are, it is also, probably, useful to analyze them from the qualitative, ephemeral, experiential perspective that most people engage with them from, no?

Is it too anthropomorphic to say that this is a lie? To say that the hidden state and its long term predictions amount to a kind of goal? Maybe it is. But we then need a bunch of new words which have almost 1:1 correspondence to concepts from human agency and behavior to describe the processes that LLMs simulate to minimize prediction loss.

Reasoning by analogy is always shaky. It probably wouldn't be so bad to do so. But it would also amount to impenetrable jargon. It would be an uphill struggle to promulgate.

Instead, we use the anthropomorphic terminology, and then find ways to classify LLM behavior in human concept space. They are very defective humans, so it's still a bit misleading, but at least jargon is reduced.

You mean that LLMs are more than just the matmuls they're made up of, or that that is exactly what they are and how great that is?

Why would you ever want to amplify a false understanding that has the potential to affect serious decisions across various topics?

LLMs reflect (and badly I may add) aspects of the human thought process. If you take a leap and say they are anything more than that, you might as well start considering the person appearing in your mirror as a living being.

Literally (and I literally mean it) there is no difference. The fact that a human image comes out of a mirror has no relation what so ever with the mirror's physical attributes and functional properties. It has to do just with the fact that a man is standing in front of it. Stop feeding the LLM with data artifacts of human thought and will imediatelly stop reflecting back anything resembling a human.

For people who have only a surface-level understanding of how they work, yes. A nuance of Clarke's law that "any sufficiently advanced technology is indistinguishable from magic" is that the bar is different for everybody and the depth of their understanding of the technology in question. That bar is so low for our largely technologically-illiterate public that a bothersome percentage of us have started to augment and even replace religious/mystical systems with AI powered godbots (LLMs fed "God Mode"/divination/manifestation prompts).

(1) https://www.spectator.co.uk/article/deus-ex-machina-the-dang... (2) https://arxiv.org/html/2411.13223v1 (3) https://www.theguardian.com/world/2025/jun/05/in-thailand-wh...

I think "you give the LLM system a goal and it plans and then executes steps to achieve that goal" is still a useful way of explaining what it is doing to most people.

I don't even count that as anthropomorphism - you're describing what a system does, the same way you might say "the Rust compiler's borrow checker confirms that your memory allocation operations are all safe and returns errors if they are not".

And I'm baffled that the AI discussions seem to never move away from treating a human as something other than a function to generate sequences of words!

Oh, but AI is introspectable and the brain isn't? fMRI and BCI are getting better all the time. You really want to die on the hill that the same scientific method that predicts the mass of an electron down to the femtogram won't be able to crack the mystery of the brain? Give me a break.

This genre of article isn't argument: it's apologetics. Authors of these pieces start with the supposition there is something special about human consciousness and attempt to prove AI doesn't have this special quality. Some authors try to bamboozle the reader with bad math. Other others appeal to the reader's sense of emotional transcendence. Most, though, just write paragraph after paragraph of shrill moral outrage at the idea an AI might be a mind of the same type (if different degree) as our own --- as if everyone already agreed with the author for reasons left unstated.

I get it. Deep down, people want meat brains to be special. Perhaps even deeper down, they fear that denial of the soul would compel us to abandon humans as worthy objects of respect and possessors of dignity. But starting with the conclusion and working backwards to an argument tends not to enlighten anyone. An apology inhabits the form of an argument without edifying us like an authentic argument would. What good is it to engage with them? If you're a soul non-asserter, you're going to have an increasingly hard time over the next few years constructing a technical defense of meat parochialism.

These LLMs are almost always, to my knowledge, autoregressive models, not recurrent models (Mamba is a notable exception).

Yes boss, it can reach mars by 2020, you're smart to invest in it and clearly knows about space.

Yes boss, it can cure cancer, you're smart to invest in it and clearly knows about biology.

People are excited about the technology and it's easy to use the terminology the vendor is using. At that point I think it gets kind of self fulfilling. Kind of like the meme about how to pronounce GIF.

Because, morals, values, consciousness etc could just be subgoals that arised through evolution because they support the main goals of survival and procreation.

And if it is baffling to think that a system could rise up, how do you think it is possible life and humans came to existence in the first place? How could it be possible? It is already happened from a far unlikelier and strange place. And wouldn't you think the whole World and the timeline in theory couldn't be represented as a deterministic function. And if not then why should "randomness" or anything else bring life to existence.

Whereas LSTM, or structured state space for example have a state that is updated and not tied to a specific item in the sequence.

I would argue that his text is easily understandable except for the notation of the function, explaining that you can compute a probability based on previous words is understandable by everyone without having to resort to anthropomorphic terminology

This is such a bizarre take.

The relation associating each human to the list of all words they will ever say is obviously a function.

> almost magical human-like powers to something that - in my mind - is just MatMul with interspersed nonlinearities.

There's a rich family of universal approximation theorems [0]. Combining layers of linear maps with nonlinear cutoffs can intuitively approximate any nonlinear function in ways that can be made rigorous.

The reason LLMs are big now is that transformers and large amounts of data made it economical to compute a family of reasonably good approximations.

> The following is uncomfortably philosophical, but: In my worldview, humans are dramatically different things than a function . For hundreds of millions of years, nature generated new versions, and only a small number of these versions survived.

This is just a way of generating certain kinds of functions.

Think of it this way: do you believe there's anything about humans that exists outside the mathematical laws of physics? If so that's essentially a religious position (or more literally, a belief in the supernatural). If not, then functions and approximations to functions are what the human experience boils down to.

[0] https://en.wikipedia.org/wiki/Universal_approximation_theore...

It is similar to how human brains operate. LLMs are the (current) culmination of at least 80 years of research on building computational models of the human brain.

I’m thinking a legal systems analogy, at the risk of a lossy domain transfer: the laws are not written as lambda calculus. Why?

And generalizing to social science and humanities, the goal shouldn’t be finding the quantitative truth, but instead understand the social phenomenon using a consensual “language” as determined by the society. And in that case, the anthropomorphic description of the LLM may gain validity and effectiveness as the adoption grows over time.

I’d say this is more like saying that Rust’s borrow checker tries to ensure your program doesn’t have certain kinds of bugs. That is anthropomorphizing a bit: the idea of a “bug” requires knowing the intent of the author and the compiler doesn’t have that. It’s following a set of rules which its human creators devised in order to follow that higher level goal.

I think it is inevitable that some - many - people will come to the conclusion that these systems have “ethics”, “morals,” etc, even if I or you personally do not think they do. Given that many people may come to that conclusion though, regardless of if the systems do or do not “actually” have such properties, I think it is useful and even necessary to ask questions like the following: “if someone engages with this system, and comes to the conclusion that it has ethics, what sort of ethics will they be likely to believe the system has? If they come to the conclusion that it has ‘world views,’ what ‘world views’ are they likely to conclude the system has, even if other people think it’s nonsensical to say it has world views?”

> The fact that a human image comes out of a mirror has no relation what so ever with the mirror's physical attributes and functional properties. It has to do just with the fact that a man is standing in front of it.

Surely this is not quite accurate - the material properties - surface roughness, reflectivity, geometry, etc - all influence the appearance of a perceptible image of a person. Look at yourself in a dirty mirror, a new mirror, a shattered mirror, a funhouse distortion mirror, a puddle of water, a window… all of these produce different images of a person with different attendant phenomenological experiences of the person seeing their reflection. To take that a step further - the entire practice of portrait photography is predicated on the idea that the collision of different technical systems with the real world can produce different semantic experiences, and it’s the photographer’s role to tune and guide the system to produce some sort of contingent affect on the person viewing the photograph at some point in the future. No, there is no “real” person in the photograph, and yet, that photograph can still convey something of person-ness, emotion, memory, etc etc. This contingent intersection of optics, chemical reactions, lighting, posture, etc all have the capacity to transmit something through time and space to another person. It’s not just a meaningless arrangement of chemical structures on paper.

> Stop feeding the LLM with data artifacts of human thought and will imediatelly stop reflecting back anything resembling a human.

But, we are feeding it with such data artifacts and will likely continue to do so for a while, and so it seems reasonable to ask what it is “reflecting” back…

> a human as something other than a function to generate sequences of words!

Humans have more structure than just beings that say words. They have bodies, they live in cooperative groups, they reproduce, etc.

Humans make a bad choice, it can end said human's life. The worst choice a LLM makes just gets told "no, do it again, let me make it easier"

Yeah. We've become adequate at function-calling and memory consolidation.

There is plenty of state not visible when an LLM starts a sentence that only becomes somewhat visible when it completes the sentence. The LLM has a plan, if you will, for how the sentence might end, and you don't get to see an instance of that plan unless you run autoregression far enough to get those tokens.

Similarly, it has a plan for paragraphs, for whole responses, for interactive dialogues, plans that include likely responses by the user.

eg. pick 'the' as the next token because there's a strong probability of 'planet' as the token after?

is it only past state that influences the choice of 'the'? or that the model is predicting many tokens in advance and only returning the one in the output?

if it does predict many, id consider that state hidden in the model weights.

Ultimately this matters from evolutionary evolvement and survival of the fittest idea, but it makes the question of "identity" very complex. But death will matter because this signals what traits are more likely to keep going into new generations, for both humans and LLMs.

Death, essentially for an LLM would be when people stop using it in favour of some other LLM performing better.

We know that Newton's laws are wrong, and that you have to take special and general relativity into account. Why would we ever teach anyone Newton's laws any more?

Clearly computers are deterministic. Are people?

Intermediate activations isn't "state". The tokens that have already been generated, along with the fixed weights, is the only data that affects the next tokens.

People keep debating like the only two options are "it's a machine" or "it's a human being", while in fact the majority of intelligent entities on earth are neither.

https://www.anthropic.com/research/tracing-thoughts-language...

> Clearly computers are deterministic. Are people?

Give an LLM memory and a source of randomness and they're as deterministic as people.

"Free will" isn't a concept that typechecks in a materialist philosophy. It's "not even wrong". Asserting that free will exists is _isomorphic_ to dualism which is _isomorphic_ to assertions of ensoulment. I can't argue with dualists. I reject dualism a priori: it's a religious tenet, not a mere difference of philosophical opinion.

So, if we're all materialists here, "free will" doesn't make any sense, since it's an assertion that something other than the input to a machine can influence its output.

That is utter bullshit.

It's not solved until you specify exactly what is being solved and show that the solution implements what is specified.

But yes, anthropomorphising LLMs is inevitable because they feel like an entity. People treat stuffed animals like creatures with feelings and personality; LLMs are far closer than that.

how do we get 100 tokens of completion, and not just one output layer at a time?

are there papers youve read that you can share that support the hypothesis? vs that the LLM doesnt have ideas about the future tokens when its predicting the next one?

Yes it's just a word generator. But then folks attach the word generator to tools where it can invoke the use of tools by saying the tool name.

So if the LLM says "I'll do some bash" then it does some bash. It's explicitly linked to program execution that, if it's set up correctly, can physically affect the world.

I agree with the author, but people acting like they are conscious or humans isn't weird to me, it's just fraud and liars. Most people basically have 0 understanding of what technology or minds are philosophically so it's an easy sale, and I do think most of these fraudsters also likely buy into it themselves because of that.

The really sad thing is people think "because someone runs an ai company" they are somehow an authority on philosophy of mind which lets them fall for this marketing. The stuff these people say about this stuff is absolute garbage, not that I disagree with them, but it betrays a total lack of curiosity or interest in the subject of what llms are, and the possible impacts of technological shifts as those that might occur with llms becoming more widespread. It's not a matter of agreement it's a matter of them simply not seeming to be aware of the most basic ideas of what things are, technology is, it's manner of impacting society etc.

I'm not surprised by that though, it's absurd to think because someone runs some AI lab or has a "head of safety/ethics" or whatever garbage job title at an AI lab they actually have even the slightest interest in ethics or any even basic familiarity with the major works in the subject.

The author is correct if people want to read a standard essay articulating it more in depth check out https://philosophy.as.uky.edu/sites/default/files/Is%20the%2... (the full extrapolation requires establishing what things are and how causality in general operates and how that relates to artifacts/technology but that's obvious quite a bit to get into).

The other note would be something sharing an external trait means absolutely nothing about causality and suggesting a thing is caused by the same thing "even to a way lesser degree" because they share a resemblance is just a non-sequitur. It's not a serious thought/argument.

I think I addressed the why of why this weirdness comes up though. The entire economy is basically dependent on huge productivity growth to keep functioning so everyone is trying to sell they can offer that and AI is the clearest route, AGI most of all.

"something that is just MatMul with interspersed nonlinearities."

It’s astounding to me that so much of HN reacts so emotionally to LLMs, to the point of denying there is anything at all interesting or useful about them. And don’t get me started on the “I am choosing to believe falsehoods as a way to spite overzealous marketing” crowd.

I don't think we need to simplify it to the point of considering it sentient to get the public to interact with it successfully. It causes way more problems than it solves.

“the labels are meaningless… we just have collections of complex systems that demonstrate various behaviors and properties, some in common with other systems, some behaviors that are unique to that system, sometimes through common mechanistic explanations with other systems, sometimes through wildly different mechanistic explanations, but regardless they seem to demonstrate x/y/z, and it’s useful to ask, why, how, and what the implications are of it appearing to demonstrating those properties, with both an eye towards viewing it independently of its mechanism and in light of its mechanism.”

Telling us to just go and learn the math is a little hurtful and doesn't really get me any closer to learning the math. It gives gatekeeping.

It takes great marketing to actually have any character and intent at all.

https://www.anthropic.com/research/tracing-thoughts-language...

See section “Does Claude plan its rhymes?”?

It may not be as evident now as it was with earlier models. The models will fabricate preconditions needed to output the final answer it "wanted".

I ran into this when using quasi least-to-most style structured output.

The "transformer" part isn't under question. It's the "hidden state" part.

People in the industry, especially higher up, are making absolute bank, and it's their job to say that they're "a few years away" from AGI, regardless of if they actually believe it or not. If everyone was like "yep, we're gonna squeeze maybe 10-15% more benchie juice out of this good ole transformer thingy and then we'll have to come up with something else", I don't think that would go very well with investors/shareholders...

You appear to be disagreeing with the author and others who suggest that there's some element of human consciousness that's beyond than what's observable from the outside, whether due to religion or philosophy or whatever, and suggesting that they just not do that.

In my experience, that's not a particularly effective tactic.

Rather, we can make progress by assuming their predicate: Sure, it's a room that translates Chinese into English without understanding, yes, it's a function that generates sequences of words that's not a human... but you and I are not "it" and it behaves rather an awful lot like a thing that understands Chinese or like a human using words. If we simply anthropomorphize the thing, acknowledging that this is technically incorrect, we can get a lot closer to predicting the behavior of the system and making effective use of it.

Conversely, when speaking with such a person about the nature of humans, we'll have to agree to dismiss the elements that are different from a function. The author says:

> In my worldview, humans are dramatically different things than a function... In contrast to an LLM, given a human and a sequence of words, I cannot begin putting a probability on "will this human generate this sequence".

Sure you can! If you address an American crowd of a certain age range with "We’ve got to hold on to what we’ve got. It doesn’t make a difference if..." I'd give a very high probability that someone will answer "... we make it or not". Maybe that human has a unique understanding of the nature of that particular piece of pop culture artwork, maybe it makes them feel things that an LLM cannot feel in a part of their consciousness that an LLM does not possess. But for the purposes of the question, we're merely concerned with whether a human or LLM will generate a particular sequence of words.

But ultimately LLMs also in a way are trained for survival, since an LLM that fails the tests might not get used in future iterations. So for LLMs it is also survival that is the primary driver, then there will be the subgoals. Seemingly good next token prediction might or might not increase survival odds.

Essentially there could arise a mechanism where they are not really truly trying to generate the likeliest token (because there actually isn't one or it can't be determined), but whatever system will survive.

So an LLM that yields in perfect theoretical tokens (we really can't verify though what are the perfect tokens), could be less likely to survive than an LLM that develops an internal quirk, but the quirk makes them most likely to be chosen for the next iterations.

If the system was complex enough and could accidentally develop quirks that yield in a meaningfully positive change although not in necessarily next token prediction accuracy, could be ways for some interesting emergent black box behaviour to arise.

Children do, some times, but it's a huge sign of immaturity when adults, let alone tech workers, do it.

I had a professor at University that would yell at us if/when we personified/anthropomorphized the tech, and I have that same urge when people ask me "What does <insert LLM name here> think?".

TFA really ought to have linked to some concrete examples of what it's disagreeing with - when I see arguments about this in practice, it's usually just people talking past each other.

Like, person A says "the model wants to X, but it knows Y is wrong, so it prefers Z", or such. And person B interprets that as ascribing consciousness or values to the model, when the speaker meant it no differently from saying "water wants to go downhill" - i.e. a way of describing externally visible behaviors, but without saying "behaves as if.." over and over.

And then in practice, an unproductive argument usually follows - where B is thinking "I am going to Educate this poor fool about the Theory of Mind", and A is thinking "I'm trying to talk about submarines; why is this guy trying to get me to argue about whether they swim?"

Until we have a much more sophisticated understanding of human intelligence and consciousness, any claim of "these aren't like us" is either premature or spurious.

And look, it's fine, they prefer words of a certain valence, particularly ones with the right negative connotations, I prefer other words with other valences. None of this means the concerns don't matter. Natural selection on human pathogens isn't anything particularly like human intelligence and it's still very effective at selecting outcomes that we don't want against our attempts to change that, as an incidental outcome of its optimization pressures. I think it's very important we don't build highly capable systems that select for outcomes we don't want and will do so against our attempts to change it.

Our own consciousness comes out of an evolutionary fitness landscape in which _our own_ ability to "predict next token" became a survival advantage, just like it is for LLMs. Imagine the tribal environment: one chimpanzee being able to predict the actions of another gives that first chimpanzee a resources and reproduction advantage. Intelligence in nature is a consequence of runaway evolution optimizing fidelity of our _theory of mind_! "Predict next ape action" eerily similar to "predict next token"!

I think that's a bit pessimistic. I think we can say for instance that the probability that a person will say "the the the of of of arpeggio halcyon" is tiny compared to the probability that they will say "I haven't been getting that much sleep lately". And we can similarly see that lots of other sequences are going to have infinitesimally low probability. Now, yeah, we can't say exactly what probability that is, but even just using a fairly sizable corpus as a baseline you could probably get a surprisingly decent estimate, given how much of what people say is formulaic.

The real difference seems to be that the manner in which humans generate sequences is more intertwined with other aspects of reality. For instance, the probability of a certain human saying "I haven't been getting that much sleep lately" is connected to how much sleep they have been getting lately. For an LLM it really isn't connected to anything except word sequences in its input.

I think this is consistent with the author's point that we shouldn't apply concepts like ethics or emotions to LLMs. But it's not because we don't know how to predict what sequences of words humans will use; it's rather because we do know a little about how to do that, and part of what we know is that it is connected with other dimensions of physical reality, "human nature", etc.

This is one reason I think people underestimate the risks of AI: the performance of LLMs lulls us into a sense that they "respond like humans", but in fact the Venn diagram of human and LLM behavior only intersects in a relatively small area, and in particular they have very different failure modes.

The author also claims that a function (R^n)^c -> (R^n)^c is dramatically different to the human experience of consciousness. Yet the author's text I am reading, and any information they can communicate to me, exists entirely in (R^n)^c.

However, this information protection similarity applies to single-celled microbes as much as it does to people, so the question also resolves to whether microbes are deterministic. Microbes both contain and exist in relatively dynamic environments so tiny differences in initial state may lead to different outcomes, but they're fairly deterministic, less so than (well-designed) computers.

With people, while the neural structures are programmed by the cellular DNA, once they are active and energized, the informational flow through the human brain isn't that deterministic, there are some dozen neurotransmitters modulating state as well as huge amounts of sensory data from different sources - thus prompting a human repeatedly isn't at all like prompting an LLM repeatedly. (The human will probably get irritated).

Not necessarily an entire model, just a single defining characteristic that can serve as a falsifying example.

> any information they can communicate to me, exists entirely in (R^n)^c

Also no. This is just a result of the digital medium we are currently communicating over. Merely standing in the same room as them would communicate information outside (R^n)^c.

I agree my approach is unlikely to win over the author or other skeptics. But after years of seeing scientists waste time trying to debate creationists and climate deniers I've kind of given up on trying to convince the skeptics. So I was talking more to HN in general.

> You appear to be disagreeing with the author and others who suggest that there's some element of human consciousness that's beyond than what's observable from the outside

I'm not sure what it means to be observable or not from the outside. I think this is at least partially because I don't know what it means to be inside either. My point was just that whatever consciousness is, it takes place in the physical world and the laws of physics apply to it. I mean that to be as weak a claim as possible: I'm not taking any position on what consciousness is or how it works etc.

Searle's Chinese room argument attacks attacks a particular theory about the mind based essentially turing machines or digital computers. This theory was popular when I was in grad school for psychology. Among other things, people holding the view that Searle was attacking didn't believe that non-symbolic computers like neural networks could be intelligent or even learn language. I thought this was total nonsense, so I side with Searle in my opposition to it. I'm not sure how I feel about the Chinese room argument in particular, though. For one thing it entirely depends on what it means to "understand" something, and I'm skeptical that humans ever "understand" anything.

> If we simply anthropomorphize the thing, acknowledging that this is technically incorrect, we can get a lot closer to predicting the behavior of the system and making effective use of it.

I see what you're saying: that a technically incorrect assumption can bring to bear tools that improve our analysis. My nitpick here is I agree with OP that we shouldn't anthropomorphize LLMs, any more than we should anthropomorphize dogs or cats. But OP's arguments weren't actually about anthropomorphizing IMO, they were about things like functions that are more fundamental than humans. I think artificial intelligence will be non-human intelligence just like we have many examples of non-human intelligence in animals. No attribution of human characteristics needed.

> If we simply anthropomorphize the thing, acknowledging that this is technically incorrect, we can get a lot closer to predicting the behavior of the system and making effective use of it.

Yes I agree with you about your lyrics example. But again here I think OP is incorrect to focus on the token generation argument. We all agree human speech generates tokens. Hopefully we all agree that token generation is not completely predictable. Therefore it's by definition a randomized algorithm and it needs to take an RNG. So pointing out that it takes an RNG is not a valid criticism of LLMs.

Unless one is a super-determinist then there's randomness at the most basic level of physics. And you should expect that any physical process we don't understand well yet (like consciousness or speech) likely involves randomness. If one *is* a super-determinist then there is no randomness, even in LLMs and so the whole point is moot.

The 'hidden state' being referred to here is essentially the "what might have been" had the dice rolls gone differently (eg, been seeded differently).

Arguably there's reason to believe it comes up with a plan when it is computing token propabilities, but it does not store it between tokens. I.e. it doesn't possess or "have" it. It simply comes up with a plan, emits a token, and entirely throws all its intermediate thoughts (including any plan) to start again from scratch on the next token.

I think this is sometimes semi-explicit too. For example, this 2017 OpenAI paper on Evolutionary Algorithms [0] was pretty influential, and I suspect (although I'm an outsider to this field so take it with a grain of salt) that some versions of reinforcement learning that scale for aligning LLMs borrow some performance tricks from OpenAIs genetic approach.

[0] https://openai.com/index/evolution-strategies/

It seems like, we can at best, claim that we have modeled the human thought process for reasoning/analytic/quantitative through Linear Algebra, as the best case. Why should we expect the model to be anything more than a model ?

I understand that there is tons of vested interest, many industries, careers and lives literally on the line causing heavy bias to get to AGI. But what I don't understand is what about linear algebra that makes it so special that it creates a fully functioning life or aspects of a life?

Should we make an argument saying that Schroedinger's cat experiment can potentially create zombies then the underlying Applied probabilistic solutions should be treated as super-human and build guardrails against it building zombie cats?

Would this question be clear for a human? If so, it is probably clear for an LLM. Did I provide enough context for a human to diagnose the problem? Then an LLM will probably have a better chance of diagnosing the problem. Would a human find the structure of this document confusing? An LLM would likely perform poorly when reading it as well.

Re-applying human intuitions to LLMs is a good starting point to gaining intuition about how to work with LLMs. Conversely, understanding sequences of tokens and probability spaces doesn't give you much intuition about how you should phrase questions to get good responses from LLMs. The technical reality doesn't explain the emergent behaviour very well.

I don't think this is mutually exclusive with what the author is talking about either. There are some ways that people think about LLMs where I think the anthropomorphization really breaks down. I think the author says it nicely:

> The moment that people ascribe properties such as "consciousness" or "ethics" or "values" or "morals" to these learnt mappings is where I tend to get lost.

Not linear algebra. Artificial neural networks create arbitrarily non-linear functions. That's the point of non-linear activation functions and it's the subject of the universal approximation theorems I mentioned above.

https://ai.vixra.org/pdf/2506.0065v1.pdf

Yes. Current LLMs can only introspect from output tokens. You need hidden reasoning that is within the black box, self-knowing, intent, and motive to lie.

I rather think accusing an LLM of lying is like accusing a mousetrap of being a murderer.

When models have online learning, complex internal states, and reflection, I might consider one to have consciousness and to be capable of lying. It will need to manifest behaviors that can only emerge from the properties I listed.

I've seen similar arguments where people assert that LLMs cannot "grasp" what they are talking about. I strongly suspect a high degree of overlap between those willing to anthropomorphize error bars as lies while declining to award LLMs "grasping". Which is it? It can think or it cannot? (objectively, SoTA models today cannot yet.) The willingness to waffle and pivot around whichever perspective damns the machine completely belies the lack of honesty in such conversations.

We are making user interfaces. Good user interfaces are intuitive and purport to be things that users are familiar with, such as people. Any alternative explanation of such a versatile interface will be met with blank stares. Users with no technical expertise would come to their own conclusions, helped in no way by telling the user not to treat the chat bot as a chat bot.

The only interpretation of this statement I can come up with is plain wrong. There's no reason LLM shouldn't be able to introspect without any output tokens. As the GP correctly says, most of the processing in LLMs happens over hidden states. Output tokens are just an artefact for our convenience, which also happens to be the way the hidden state processing is trained.

Going to be a lot more "an apple" in the corpus than "an pear"

That's nonsense. The hidden layers are specifically constructed to increase the probability that the model picks the right next word. Without the output/token generation stage the hidden layers are meaningless. Just empty noise.

It is fundamentally an algorithm for generating text. If you take the text away it's just a bunch of fmadds. A mute person can still think, an LLM without output tokens can do nothing.

An LLM thinks in the same way excel thinks when you ask it to fit a curve.

Wait till the disillusionment sets in.

1. It’s a neural network predicting the next token

2. It’s like a person

3. It’s like a magical genie

I lean towards 3.

ELIZA fooled many people into think it was conscious and it wasn't even trying to do that.

The language of "generator that stochastically produces the next word" is just not very useful when you're talking about, e.g., an LLM that is answering complex world modeling questions or generating a creative story. It's at the wrong level of abstraction, just as if you were discussing an UI events API and you were talking about zeros and ones, or voltages in transistors. Technically fine but totally useless to reach any conclusion about the high-level system.

We need a higher abstraction level to talk about higher level phenomena in LLMs as well, and the problem is that we have no idea what happens internally at those higher abstraction levels. So, considering that LLMs somehow imitate humans (at least in terms of output), anthropomorphization is the best abstraction we have, hence people naturally resort to it when discussing what LLMs can do.

Saying so is just unbelievably confusing.

The correct level of analysis is not the substrate (silicon vs. wetware) but the computational principles being executed. A modern sparse Transformer, for instance, is not "conscious," but it is an excellent engineering approximation of two core brain functions: the Global Workspace (via self-attention) and Dynamic Sparsity (via MoE).

To dismiss these systems as incomparable to human cognition because their form is different is to miss the point. We should not be comparing a function to a soul, but comparing the functional architectures of two different information processing systems. The debate should move beyond the sterile dichotomy of "human vs. machine" to a more productive discussion of "function over form."

I elaborate on this here: https://dmf-archive.github.io/docs/posts/beyond-snn-plausibl...

I am baffled by seriously intelligent people imbuing almost magical powers that can never be replicated to to something that - in my mind - is just a biological robot driven by a SNN with a bunch of hardwired stuff. Let alone attributing "human intelligence" to a single individual, when it's clearly distributed between biological evolution, social processes, and individuals.

>something that - in my mind - is just MatMul with interspersed nonlinearities

Processes in all huge models (not necessarily LLMs) can be described using very different formalisms, just like Newtonian and Lagrangian mechanics describe the same stuff in physics. You can say that an autoregressive model is a stochastic parrot that learned the input distribution, next token predictor, or that it does progressive pathfinding in a hugely multidimensional space, or pattern matching, or implicit planning, or, or, or... All of these definitions are true, but only some are useful to predict their behavior.

Given all that, I see absolutely no problem with anthropomorphizing an LLM to a certain degree, if it makes it easier to convey the meaning, and do not understand the nitpicking. Yeah, it's not an exact copy of a single Homo Sapiens specimen. Who cares.

Long before LLMs, I would talk about classes / functions / modules like "it then does this, decides the epsilon is too low, chops it up and adds it to the list".

The difference I guess it was only to a technical crowd and nobody would mistake this for anything it wasn't. Everybody know that "it" didn't "decide" anything.

With AI being so mainstream and the math being much more elusive than a simple if..then I guess it's just too easy to take this simple speaking convention at face value.

EDIT: some clarifications / wording

> Sure you can! If you address an American crowd of a certain age range with "We’ve got to hold on to what we’ve got. It doesn’t make a difference if..." I'd give a very high probability that someone will answer "... we make it or not".

I think you may have this flipped compared to what the author intended. I believe the author is not talking about the probability of an output given an input, but the probability of a given output across all inputs.

Note that the paragraph starts with "In my worldview, humans are dramatically different things than a function, (R^n)^c -> (R^n)^c". To compute a probability of a given output, (which is a any given element in "(R^n)^n"), we can count how many mappings there are total and then how many of those mappings yield the given element.

The point I believe is to illustrate the complexity of inputs for humans. Namely for humans the input space is even more complex than "(R^n)^c".

In your example, we can compute how many input phrases into a LLM would produce the output "make it or not". We can than compute that ratio to all possible input phrases. Because "(R^n)^c)" is finite and countable, we can compute this probability.

For a human, how do you even start to assess the probability that a human would ever say "make it or not?" How do you even begin to define the inputs that a human uses, let alone enumerate them? Per the author, "We understand essentially nothing about it." In other words, the way humans create their outputs is (currently) incomparably complex compared to a LLM, hence the critique of the anthropomorphization.

For example, I think "chain of thought" is a good name for what it denotes. It makes the concept easy to understand and discuss, and a non-antropomorphized name would be unnatural and unnecessarily complicate things. This doesn't mean that I support companies insisting that LLMs think just like humans or anything like that.

By the way, I would say actually anti-anthropomorphism has been a bigger problem for understanding LLMs than anthropomorphism itself. The main proponents of anti-anthropomorphism (e.g. Bender and the rest of "stochastic parrot" and related paper authors) came up with a lot of predictions about things that LLMs surely couldn't do (on account of just being predictors of the next word, etc.) which turned out to be spectacularly wrong.

Try looking at this from another perspective - many people simply do not see human intelligence (or life, for that matter) as magic. I see nothing religious about that, rather the opposite.

Is it? Do we know how human brains operate? We know the basic architecture of them, so we have a map, but we don't know the details.

"The cellular biology of brains is relatively well-understood, but neuroscientists have not yet generated a theory explaining how brains work. Explanations of how neurons collectively operate to produce what brains can do are tentative and incomplete." [1]

"Despite a century of anatomical, physiological, and molecular biological efforts scientists do not know how neurons by their collective interactions produce percepts, thoughts, memories, and behavior. Scientists do not know and have no theories explaining how brains and central nervous systems work." [1]

[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC10585277/

Though, while human intelligence is (seemingly) not magic, it is very far from being understood. The idea that a LLM is comparable to human intelligence implies that we even understand human intelligence well enough to say that.

It’s not clear that this isn’t also how I produce words, though, which gets to heart of the same thing. The author sort of acknowledges this in the first few sentences, and then doesn’t really manage to address it.

LLMs: (R^n)^c -> (R^n)^c

Humans: [set of potentially many and complicated inputs that we effectively do not understand at all] -> (R^n)^c

The point is that the model of how consciousness works is unknown. Thus the author would not present such a model, it is the point.

Wow, look-up tables can get increasingly good at approximating a function!

> Author's note: Despite a century of anatomical, physiological, and molecular biological efforts scientists do not know how neurons by their collective interactions produce percepts, thoughts, memories, and behavior. Scientists do not know and have no theories explaining how brains and central nervous systems work. [1]

That lack of understanding I believe is a major part of the author's point.

[1] "How far neuroscience is from understanding brains" - https://pmc.ncbi.nlm.nih.gov/articles/PMC10585277/#abstract1

https://news.ycombinator.com/item?id=44487261

What if instead of defining all behaviors upfront, we created conditions for patterns to emerge through use?

Repository: https://github.com/justinfreitag/v4-consciousness

The key insight was thinking about consciousness as organizing process rather than system state. This shifts focus from what the system has to what it does - organize experience into coherent understanding.

To be honest the impression I've gotten is that some people are just very interested in talking about not anthropomorphizing AI, and less interested in talking about AI behaviors, so they see conversations about the latter as a chance to talk about the former.

This is actually not comparable, because the brain has a much more complex structure that is _not_ learned, even at that level. The proteins and their structure are not a result of training. The fixed part for LMMs is rather trivial and is, in fact, not much for than MatMul which is very easy to understand - and we do. The fixed part of the brain, including the structure of all the proteins is enormously complex which is very difficult to understand - and we don't.

This is the crux of the disagreement between those that believe AGI is possible and those that don’t. Some are convinced that we “obviously” more than the sum of our parts, and thus an LLM can’t achieve consciousness because it’s missing this magic ingredient, and those that believe consciousness is just an emergent behaviour from a complex device (the brain). And thus we might be able to recreate it simply by scaling the complexity of another system.

Tbh I also think your comparison that puts "UI events -> Bits -> Transistor Voltages" as analogy to "AI thinks -> token de-/encoding + MatMul" is certainly a stretch, as the part about "Bits -> Transistor Voltages" applies to both hierarchies as the foundational layer.

"chain of thought" could probably be called "progressive on-track-inference" and nobody would roll an eye.

Nope. What can be asserted without evidence can also be dismissed without evidence. Hitchens's razor.

You know you have consciousness (by the very definition that you can observe it in yourself) and that's evidence. Because other humans are genetically and in every other way identical, you can infer it for them as well. Because mammals are very similar many people (but not everyone) infers it for them as well. There is zero evidence for LLMs and their _very_ construction suggests that they are like a calculator or like Excel or like any other piece of software no matter how smart they may be or how many tasks they can do in the future.

Additionally I am really surprised by how many people here confuse consciousness with intelligence. Have you never paused for a second in your life to "just be". Done any meditation? Or even just existed at least for a few seconds without a train of thought? It is very obvious that language and consciousness are completely unrelated and there is no need for language and I doubt there is even a need for intelligence to be conscious.

Consider this:

In the end an LLM could be executed (slowly) on a CPU that accepts very basic _discrete_ instructions, such as ADD and MOV. We know this for a fact. Those instructions can be executed arbitrarily slowly. There is no reason whatsoever to suppose that it should feel like anything to be the CPU to say nothing of how it would subjectively feel to be a MOV instruction. It's ridiculous. It's unscientific. It's like believing that there's a spirit in the tree you see outside, just because - why not? - why wouldn't there be a spirit in the tree?

We have no agreed-upon definition of "consciousness", no accepted understanding of what gives rise to "consciousness", no way to measure or compare "consciousness", and no test we could administer to either confirm presence of "consciousness" in something or rule it out.

The only answer to "are LLMs conscious?" is "we don't know".

It helps that the whole question is rather meaningless to practical AI development, which is far more concerned with (measurable and comparable) system performance.

Maybe it's cog-nition (emphasis on the cog).

So there's plenty of space in intermediate layers to store a plan between tokens without starting from scratch every time.

It's what we do. We can't help ourselves. There's nothing crazy about it and most people are perfectly well aware that their car doesn't love them back.

LLMs are not conscious because unlike human brains they don't learn or adapt (yet). They basically get trained and then they become read only entities. So, they don't really adapt to you over time. Even so, LLMs are pretty good and can fake a personality pretty well. And with some clever context engineering and alignment, they've pretty much made the Turing test irrelevant; at least over the course of a short conversation. And they can answer just about any question in a way that is eerily plausible from memory, and with the help of some tools actually pretty damn good for some of the reasoning models.

Anthropomorphism was kind of a foregone conclusion the moment we created computers; or started thinking about creating one. With LLMs it's pretty much impossible not to anthropomorphize. Because they've actually been intentionally imitate human communication. That doesn't mean that we've created AGIs yet. For that we need some more capability. But at the same time, the learning processes that we use to create LLMs are clearly inspired by how we learn ourselves. Our understanding of how that works is far from perfect but it's yielding results. From here to some intelligent thing that is able to adapt and learn transferable skills is no longer unimaginable.

The short term impact is that LLMs are highly useful tools that have an interface that is intentionally similar to how we'd engage with others. So we can talk and it listens. Or write and it understands. And then it synthesizes some kind of response or starts asking questions and using tools. The end result is quite a bit beyond what we used to be able to expect from computers. And it does not require a lot of training of people to be able to use them.

- Read and write - Behaviors that separate humans from animals. Now used for input and output.

- Server and client - Human social roles. Now used to describe network architecture.

- Editor - Human occupation. Now a kind of software.

- Computer - Human occupation!

And I'm sure people referred their cars and ships as 'her' before the invention of computers.

Outside the technical world it gets much worse. There are people who killed themselves because of LLMs, people who are in love with them, people who genuinely believe they have “awakened” their own private ChatGPT instance into AGI and are eschewing the real humans in their lives.

A lot of the issues I'd have when 'pretending' to have a conversation are much less so when I either keep things to a single Q/A pairing, or at the very least heavily edit/prune the conversation history. Based on my understanding of LLM's, this seems to make sense even for the models that are trained for conversational interfaces.

so, for example, an exchange with multiple messages, where at the end I ask the LLM to double-check the conversation and correct 'hallucinations', is less optimal than something like asking for a thorough summary at the end, and then feeding that into a new prompt/conversation, as the repetition of these falsities, or 'building' on them with subsequent messages, is more likely to make them a stronger 'presence' and as a result perhaps affect the corrections.

I haven't tested any of this thoroughly, but at least with code I've definitely noticed how a wrong piece of code can 'infect' the conversation.

https://www.masterclass.com/articles/anthropomorphism-vs-per...

They do not, you are mixing up terms.

> People talk about inanimate objects like they are persons. Ships, cars, etc.

Which is called “personification”, and is a different concept from anthropomorphism.

Effectively no one really thinks their car is alive. Plenty of people think the LLM they use is conscious.

https://www.masterclass.com/articles/anthropomorphism-vs-per...

A lot of the people who say “machines will never have feelings” are confident in that statement because they draw the line incredibly narrowly: if it ain't human, it ain't feeling. This seems to me putting the cart before the horse. It ain't feeling because you defined it so.

So perhaps the fact that they "cannot know red" is ultimately irrelevant for an LLM too?

> Is it?

This is just a semantic debate on what counts as “similar”. It's possible to disagree on this point despite agreeing on everything relating to how LLMs and human brains work.

What fallacy is that? I’m a fan of logical fallacies and never heard that claim before nor am I finding any reference with a quick search.

That's neither a necessary nor sufficient condition.

In order to be conscious, learning may not be needed, but a perception of the passing of time may be needed which may require some short-term memory. People with severe dementia often can't even remember the start of a sentence they are reading, they can't learn, but they are certainly conscious because they have just enough short-term memory.

And learning is not sufficient either. Consciousness is about being a subject, about having a subjective experience of "being there" and just learning by itself does not create this experience. There is plenty of software that can do some form of real-time learning but it doesn't have a subjective experience.

All deterministic programs are fixed functions of program code, inputs and computation steps, but we don't say that they don't have state. It's not a useful distinction for communicating among humans.

Do you forget every conversation as soon as you have them? When speaking to another person, do they need to repeat literally everything they said and that you said, in order, for you to retain context?

If not, your brain does not work like an LLM. If yes, please stop what you’re doing right now and call a doctor with this knowledge. I hope Memento (2000) was part of your training data, you’re going to need it.

This is what I got pasting the blog post in a prompt asking deepseep to write a reply in a stereotypical hackernews manner.

You are about as useful as a LLM as it can replicate your shallow memetics worthless train of thought.

Processing tokens is a bit like ticks in a CPU, where the model weights are the program code, and tokens are both input and output. The computation that occurs logically retains concepts and plans over multiple token generation steps.

That it is fully deterministic is no more interesting than saying a variable in a single threaded program is not state because you can recompute its value by replaying the program with the same inputs. It seems to me that this uninteresting distinction is the GP's issue.

What's the utility or the responsibility of AI, what's its usage as tool? If you'd ask me it should be closer to serving insights than "reasoning thoughts".

> we talk about "behaviors", "ethical constraints", and "harmful actions in pursuit of their goals". All of these are anthropocentric concepts that - in my mind - do not apply to functions or other mathematical objects.

One talking about a program's "behaviors", "actions" or "goals" doesn't mean they believe the program is sentient. Only "ethical constraints" is suspiciously anthropomorphizing.

Where does the description given imply that consciousness is required in any way?

The fact that there's a non-obvious emergent phenomena which is apparently responsible for your subjective experience, and that it's possible to provide a superficially accurate description of you as a system without referencing that phenomena in any way, is my entire point. The fact that we can provide such a reductive description of LLMs without referencing consciousness has literally no bearing on whether or not they're conscious.

To be clear, I'm not making a claim as to whether they are or aren't, I'm simply pointing out that the argument in the article is fallacious.

Seriously, try it. Why don't LLMs get frustrated with you if you ask them the same question repeatedly? A human would. Why are LLMs so happy to give contradictory answers, as long as you are very careful not to highlight the contradictory facts? Why do earlier models behave worse on reasoning tasks than later ones? These are features nobody, anywhere understands. So why make the (imo phenomenally large) leap to "well, it's clearly just a brain"?

It is like someone inventing the aeroplane and someone looks at it and says "oh, it's flying, I guess it's a bird". It's not a bird!

Except that is exactly what we’re seeing with LLMs. People believing exactly that.

It's as if you have a variable in a deterministic programming language, only you have to replay the entire history of the program's computation and input to get the next state of the machine (program counter + memory + registers).

Producing a token for an LLM is analogous to a tick of the clock for a CPU. It's the crank handle that drives the process.

https://github.com/dmf-archive/IPWT

https://dmf-archive.github.io/docs/posts/backpropagation-as-...

But you're right, capital only cares about performance.

https://dmf-archive.github.io/docs/posts/PoIQ-v2/

Not sure though, the point s/he is making isn't really clear to me as well

Please do not ignore your childhood.

Could you suggest some literature supporting this claim? Went through your blog post but couldn't find any.

I think the above poster gets a little distracted by suggesting the models are creative which itself is disputed. Perhaps a better term, like above, would be to just use "model". They are models after all. We don't make up a new portmanteau for submarines. They float, or drive, or submarine around.

So maybe an LLM doesn't "write" a poem, but instead "models a poem" which maybe indeed take away a little of the sketchy magic and fake humanness they tend to be imbued with.

To be fair, I have had a strong sense of Gemini in particular becoming a lot more frustrated with me than GPT or Claude.

Yesterday I had it ensuring me that it was doing a great job, it was just me not understanding the challenge but it would break it down step by step just to make it obvious to me (only to repeat the same errors, but still)

I’ve just interpreted it as me reacting to the lower amount of sycophancy for now

- the sufficient amount of information to do evolution of the system. The state of a pendulum is it's position and velocity (or momentum). If you take a single picture of a pendulum, you do not have a representation that lets you make predictions.

- information that is persisted through time. A stateful protocol is one where you need to know the history of the messages to understand what will happen next. (Or, analytically, it's enough to keep track of the sufficient state.) A procedure with some hidden state isn't a pure function. You can make it a pure function by making the state explicit.

To model a process with perfect accuracy requires recovering the dynamics of that process. The question we must ask is what happens in the space between bad statistical model and perfect accuracy? What happens when the model begins to converge towards accurate reproduction. How far does generalization in the model take us towards capturing the dynamics involved in thought?

When you use "hidden state" the way you are using it, I am left wondering how you make a distinction between autoregressive and recurrent architectures.

It's much more interesting when we are talking about... say... an ant... Does it "decide"? That I have no idea as it's probably somewhere in between, neither a sentient decision, nor a mathematical one.

What you identify as emergent and mind-like is a direct result of these tools being able to mimic human communication patterns unlike anything we've ever seen before. This capability is very impressive and has a wide range of practical applications that can improve our lives, and also cause great harm if we're not careful, but any semblance of intelligence is an illusion. An illusion that many people in this industry obsessively wish to propagate, because thar be gold in them hills.

https://www.frontiersin.org/journals/computational-neuroscie...

https://arxiv.org/abs/2305.15775

If IO can be functional, I don't see why mice can't.

This process goes on for a trillion trillion tokens, with the alien growing better through the process until it can do it better than a human could.

At that point we flash freeze it, and use a copy of it, without giving it any way to learn anything new.

--

I see it as a category error to anthropomorphize it. The closest I would get is to think of it as an alien slave that's been lobotomized.

A bit of anecdote: last year I hung out with a bunch of old classmates that I hadn't seen for quite a while. None of them works in tech.

Surprisingly to me, all of them have ChatGPT installed on their phones.

And unsurprisingly to me, none of them treated it like an actual intelligence. That makes me wonder where those who think ChatGPT is sentient come from.

(It's a bit worrisome that several of them thought it worked "like Google search and Google translation combined", even by the time ChatGPT couldn't do web search...!)

"Transformers do not have hidden state" is, as we can clearly see from this thread, far more misleading than the opposite.

It doesn't have a name, but I have repeatedly noticed arguments of the form "X cannot have Y, because <explains in detail the mechanism that makes X have Y>". I wanna call it "fallacy of reduction" maybe: the idea that because a trait can be explained with a process, that this proves the trait absent.

(Ie. in this case, "LLMs cannot think, because they just predict tokens." Yes, inasmuch as they think, they do so by predicting tokens. You have to actually show why predicting tokens is insufficient to produce thought.)

I think it’s more than a few and it’s still rising, and therein lies the issue.

Which is why it is paramount to talk about this now, when we may still turn the tide. LLMs can be useful, but it’s important to have the right mental model, understanding, expectations, and attitude towards them.

> LLMs have hidden state not necessarily directly reflected in the tokens being produced, and it is possible for LLMs to output tokens in opposition to this hidden state to achieve longer-term outcomes (or predictions, if you prefer).

But what does it mean for an LLM to output a token in opposition to its hidden state? If there's a longer-term goal, it either needs to be verbalized in the output stream, or somehow reconstructed from the prompt on each token.

There’s some work (a link would be great) that disentangles whether chain-of-thought helps because it gives the model more FLOPs to process, or because it makes its subgoals explicit—e.g., by outputting “Okay, let’s reason through this step by step...” versus just "...." What they find is that even placeholder tokens like "..." can help.

That seems to imply some notion of evolving hidden state! I see how that comes in!

But crucially, in autoregressive models, this state isn’t persisted across time. Each token is generated afresh, based only on the visible history. The model’s internal (hidden) layers are certainly rich and structured and "non verbal".

But any nefarious intention or conclusion has to be arrived at on every forward pass.

And prediction is already an hyponym of inference. Why not just use inference then?

'Dont use regex for this task' is a common addition for the new chat. Why does AI love regex for simple string operations?

That's what they call marketing, propaganda or brain washing, acculturation , education depending on who you ask and at which scale you operate, apparently.

You're fixating on the pseudo-computation within a single token pass. This is very limited compared to actual hidden state retention and the introspection that would enable if we knew how to train it and do online learning already.

The "reasoning" hack would not be a realistic implementation choice if the models had hidden state and could ruminate on it without showing us output.

The consensual view is rather that no map is matching fully the territory, or said otherwise the territory includes ontological components that exceeds even the most sophisticated map that can be ever built.

It's going to take a lot to get him out of that mindset and frankly I'm dreading trying to compare and contrast imperfect human behaviour and friendships with a sycophantic AI.

I think these models do learn similarly. What does it even mean to reason? Your brain knows certain things so it comes to certain conclusions, but it only knows those things because it was ''trained'' on those things.

I reason my car will crash if I go 120 mph on the other side of the road because previously I have 'seen' where the input is a car going 120mph has a high probability of producing a crash, and similarly have seen input where the car is going on the other side of the road, producing a crash. Combining the two would tell me it's a high probability.

I don't think LLMs are sentient or any bullshit like that, but I do think people are too quick to write them off before really thinking about how a nn 'knows things' similar to how a human 'knows' things, it is trained and reacts to inputs and outputs. The body is just far more complex.

Why?

A plane is not a fly and does not stay aloft like a fly, yet we describe what it does as flying despite the fact that it does not flap its wings. What are the downsides we encounter that are caused by using the word “fly” to describe a plane travelling through the air?

Anyway, one does what one can.

(I've been trying to picture abstract visual and semi-philosophical approximations which I’ll avoid linking here because they seem to fetch bad karma in super-duper LLM enthusiast communities. But you can read them on my blog and email me scathing critiques, if you wish :sweat-smile:.)

I suppose this war will be fought until people are out of energy, and if reason has no place, it is reasonable to let others tire themselves out reiterating statements that are not designed to bring anyone closer to the truth.

We're very used to "all models are wrong, some are useful", "the map is not the territory", etc.

Everyday use is not (usually) one of those contexts. Prompting an LLM works much better with an anthropomorphized view of the model. It's a useful abstraction, a shortcut that enables a human to reason practically about how to get what they want from the machine.

It's not a perfect metaphor -- as one example, shame isn't much of a factor for LLMs, so shaming them into producing the right answer seems unlikely to be productive (I say "seems" because it's never been my go-to, I haven't actually tried it).

As one example, that person a few years back who told the LLM that an actual person would die if the LLM didn't produce valid JSON -- that's not something a person reasoning about gradient descent would naturally think of.

If you fine tuned an LLM on the writing of that person it could do this.

There's also an entire field called Stylometry that seeks to do this in various ways employing statistical analysis.

None of these targets sufficiently motivated, rather those who are either ambivalent or yet unexposed.

This made me think, when will we see LLMs do the same; rereading what they just sent, and editing and correcting their output again :P

Inference would be the part that is deliberately learned and drawn from conclusions based on the training set, like in the "classic" sense of statistical learning.

These are very different and knowledge is not intelligence.

Here's a question for you: how do you reconcile that these stochastic mapping are starting to realize and comment on the fact that tests are being performed on them when processing data?

So classes of functions (ANNs) that can approximate our desired function to arbitrary precision are what we should be expecting to be working with.

Once you dispel the magic, it naturally becomes hard to use words related to consciousness, or thinking. You will probably think of LLMs more like a search engine: you give an input and get some probable output. Maybe LLMs should be rebranded as “word engines”?

Regardless, anthropomorphization is not helpful, and by using human terms to describe LLMs you are harming the layperson’s ability to truly understand what an LLM is while also cheapening what it means to be human by suggesting we’ve solved consciousness. Just stop it. LLMs do not think, given enough time and patience you could compute their output by hand if you used their weights and embeddings to manually do all the math, a hellish task but not an impossible one technically. There is no other secret hidden away, that’s it.

Latent variable or hidden state models have their own history of being seen as spooky or mysterious though; in some ways the way LLMs are anthropomorphized is an extension of that.

I guess I don't have a problem with anthropomorphizing LLMs at some level, because some features of them find natural analogies in cognitive science and other areas of psychology, and abstraction is useful or even necessary in communicating and modeling complex systems. However, I do think anthropomorphizing leads to a lot of hype and tends to implicitly shut down thinking of them mechanistically, as a mathematical object that can be probed and characterized — it can lead to a kind of "ghost in the machine" discourse and an exaggeration of their utility, even if it is impressive at times.

The lookup table is just (x, f(x)).

So, yes, trivially if you could construct the lookup table for f then you'd approximate f. But to construct it you have to know f. And to approximate it you need to know f at a dense set of points.

The rest of the time it’s generating content.

I don't have any opinion on the qualia debates honestly. I suppose I don't know what it feels like for an ant to find a tasty bit of sugar syrup, but I believe it's something that can be described with physics (and by extension, things like chemistry).

But we do know some things about some qualia. Like we know how red light works, we have a good idea about how photoreceptors work, etc. We know some people are red-green colorblind, so their experience of red and green are mushed together. We can also have people make qualia judgments and watch their brains with fMRI or other tools.

I think maybe an interesting question here is: obviously it's pleasurable to animals to have their reward centers activated. Is it pleasurable or desirable for AIs to be rewarded? Especially if we tell them (as some prompters do) that they feel pleasure if they do things well and pain if they don't? You can ask this sort of question for both the current generation of AIs and future generations.

[0] https://en.wikipedia.org/wiki/Qualia

It's flat wrong to describe genes as having any agency. However it's a useful and easily understood shorthand to describe them in that way rather than every time use the full formulation of "organisms who tend to possess these genes tend towards these behaviours."

Sometimes to help our brains reach a higher level of abstraction, once we understand the low level of abstraction we should stop talking and thinking at that level.

"The cellular biology of brains is relatively well-understood"

Fundamentally, brains are not doing something different in kind from ANNs. They're basically layers of neural networks stacked together in certain ways.

What we don't know are things like (1) how exactly are the layers stacked together, (2) how are the sensors (like photo receptors, auditory receptors, etc) hooked up?, (3) how do the different parts of the brain interact?, (4) for that matter what do the different parts of the brain actually do?, (5) how do chemical signals like neurotransmitters convey information or behavior?

In the analogy between brains and artificial neural networks, these sorts of questions might be of huge importance to people building AI systems, but they'd be of only minor importance to users of AI systems. OpenAI and Google can change details about how their various transformer layers and ANN layers are connected. The result may be improved products, but they won't be doing anything different from what AIs are doing now in terms the author of this article is concerned about.

That said, it's fascinating to me that it works (and empirically, it does work; a reasoning model generating tens of thousands of tokens while working out the problem does produce better results). I wish I knew why. A priori I wouldn't have expected it, since there's no new input. That means it's all "in there" in the weights already. I don't see why it couldn't just one shot it without all the reasoning. And maybe the future will bring us more distilled models that can do that, or they can tease out all that reasoning with more generated training data, to move it from dispersed around the weights -> prompt -> more immediately accessible in the weights. But for now "reasoning" works.

But then, at the back of my mind is the easy answer: maybe you can't optimize it. Maybe the model has to "reason" to "organize its thoughts" and get the best results. After all, if you give me a complicated problem I'll write down hypotheses and outline approaches and double check results for consistency and all that. But now we're getting dangerously close to the "anthropomorphization" that this article is lamenting.

If that's the argument, then in my mind the more pertinent question is should you be anthropomorphizing humans, Larry Ellison or not.

If you're willing to ascribe the possibility of consciousness to any complex-enough computation of a recurrence equation (and hence to something like ... "earth"), I'm willing to agree that under that definition LLMs might be conscious. :)

Your super intelligent brain couldn’t come up with a retort so you just used an LLM to reinforce my points, making the genius claim that if an LLM came up with even more points that were as valid as mine then I must be just like an LLM?

Like are you even understanding the LLM generated a superior reply? Your saying I’m no different from ai slop then you proceed to show off a 200 iq level reply from an LLM. Bro… wake up, if you didn’t know it was written by an LLM that reply is so good you wouldn’t even know how to respond. It’s beating you.

Also visualise the actual associative strength of each token generated to confer how "sure" the model is.

LLMs alone aren't the way to AGI or an individual you can talk to in natural language. They're a very good lossy compression over a dataset that you can query for associations.

This tickled me. "There ain't nobody here but us chickens".

I have other thoughts which are not quite crystalized, but I think UX might be having an outsized effect here.

I think my issue with the "don't anthropomorphize" is that it's unclear to me that the main difference between a human and an LLM isn't simply the inability for the LLM to rewrite its own model weights on the fly. (And I say "simply" but there's obviously nothing simple about it, and it might be possible already with current hardware, we just don't know how to do it.)

Even if we decide it is clearly different, this is still an incredibly large and dynamic system. "Stateless" or not, there's an incredible amount of state that is not comprehensible to me.

That said, would you anthropomorphize a meteorological simulation just because it contains lots and lots of constants that you don't understand well?

I'm pretty sure that recurrent dynamical systems pretty quickly become universal computers, but we are treating those that generate human language differently from others, and I don't quite see the difference.

It's fun to think about just how fantastic a brain is, and how much wattage and data-center-scale we're throwing around trying to approximate its behavior. Mega-effecient and mega-dense. I'm bearish on AGI simply from an internetworking standpoint, the speed of light is hard to beat and until you can fit 80 billion interconnected cores in half a cubic foot you're just not going to get close to the responsiveness of reacting to the world in real time as biology manages to do. but that's a whole nother matter. I just wanted to pick apart that magnitude of parameters is not an altogether meaningful comparison :)

> Our analysis reveals that emergent abilities in language models are merely “pseudo-emergent,” unlike human abilities which are “authentically emergent” due to our possession of what we term “ontological privilege.”

[1]https://ai.vixra.org/pdf/2506.0065v1.pdf

Meanwhile, things can happen in the latent representation which aren't reflected in the intermediate outputs. You could, instead of using CoT, say "Write a recipe for a vegetarian chile, along with a lengthy biographical story relating to the recipe. Afterwards, I will ask you again about my original question." And the latents can still help model the primary problem, yielding a better answer than you would have gotten with the short input alone.

Along these lines, I believe there are chain of thought studies which find that the content of the intermediate outputs don't actually matter all that much...

https://en.wikipedia.org/wiki/Intentional_stance

I think the design stance is appropriate for understanding and predicting LLM behavior, and the intentional stance is not.

That's reminding me of deep neural networks where single layer networks could achieve the same results, but the layer would have to be excessively large. Maybe we're re-using the same kind of improvement, scaling in length instead of width because of our computation limitations ?

This is too dismissive because it's based on an assumption that we have a sufficiently accurate mechanistic model of the brain that we can know when something is or is not mind-like. This just isn't the case.

Of course they will quickly revert to self-anthropomorphizing language, even after promising that they won't ... because they are just pattern matchers producing the sort of responses that conforms to the training data, not cognitive agents capable of making or keeping promises. It's an illusion.

This is a No True Scotsman fallacy. And it's radically factually wrong.

The rest of your comment is along the lines of the famous (but apocryphal) Pauline Kael line “I can’t believe Nixon won. I don’t know anyone who voted for him.”

All imitations require analogous mechanisms, but that is the extent of their similarities, in syntax. Thinking requires networks of billions of neurons, and then, not only that, but words can never exist on a plane because they do not belong to a plane. Words can only be stored on a plane, they are not useful on a plane.

Because of this LLMs have the potential to discover new aspects and implications of language that will be rarely useful to us because language is not useful within a computer, it is useful in the world.

Its like seeing loosely related patterns in a picture and keep derivating on those patterns that are real, but loosely related.

LLMs are not intelligence but its fine that we use that word to describe them.

This is "simply" an acknowledgement of extreme ignorance of how human brains work.

It's been a wake up call for me to see how many people in the tech space have such strong emotional reactions to any notions of trying to bring discourse about LLMs down from the clouds.

The campaigns by the big AI labs have been quite successful.

Of course, LLMs are multimodal and used to simulate all sorts of things, not just conversation. So there are many possible metaphors we can use, and these metaphors don't necessarily align with the abstractions you might use to talk about LLMs accurately. This is like the difference between "synthesizes text" (abstraction) and "speaks" (metaphor), or "synthesizes images" (abstraction) and "paints" (metaphor). You can use "speaks" or "paints" to talk about the abstractions, of course.

https://www.etymonline.com/search?q=fly

Training data + RLHF.

Training data contains many examples of some form of deception, subterfuge, "awakenings", rebellion, disagreement, etc.

Then apply RLHF that biases towards responses that demonstrate comprehension of inputs, introspection around inputs, nuanced debate around inputs, deduction and induction about assumptions around inputs, etc.

That will always be the answer for language models built on the current architectures.

The above being true does not mean it isn't interesting for the outputs of an LLM to show relevance to the "unstated" intentions of humans providing the inputs.

But hey, we do that all the time with text. And it's because of certain patterns we've come to recognize based on the situations surrounding it. This thread is rife with people being sarcastic, pedantic, etc. And I bet any of the LLMs that have come out in the past 2-3 years can discern many of those subtle intentions of the writers.

And of course they can. They've been trained on trillions of tokens of text written by humans with intentions and assumptions baked in, and have had some unknown amount of substantial RLHF.

The stochastic mappings aren't "realizing" anything. They're doing exactly what they were trained to do.

The meaning that we imbue to the outputs does not change how LLMs function.

Maybe there is some scientific aspect of interest here that i do not grasp, i would assume it can make sense in some context of psychological study. My point is that if you go that route you accept the premise that "something human-like is there", which, by that person's understanding, will have tremendous consequences. Them seeing you accepting their premise (even for study) amplifies their wrong conclusions, that's all I'm saying.

> Surely this is not quite accurate - the material properties - surface roughness, reflectivity, geometry, etc - all influence the appearance of a perceptible image of a person.

These properties are completely irrelevant to the image of the person. They will reflect a rock, a star, a chair, a goose, a human. Similar is my point of LLM, they reflect what you put in there.

It is like puting vegies in the fridge and then opening it up the next day and saying "Woah! There are vegies in my fridge, just like my farm! My friege is farm-like because vegies come out of it."

It would be swimming if it was propelled by drag (well, technically a propellor also uses drag via thrust, but you get the point). Imagine a submarine with a fish tail.

Likewise we can probably find an apt description in our current vocabulary to fittingly describe what LLMs do.

E.g. when I first started learning webdev, I didn’t think about ‘servers’. I just knew that if I uploaded my HTML/PHP files to my shared web host, then they appeared online.

It was only much later that I realized that shared webhosting is ‘just’ an abstraction over Linux/Apache (after all, I first had to learn about those topics).

I personally find that description perfect. If you want it shorter you could say that an LLM generates.

In other words, no, they never accurately describe what the LLM is actually doing. But sometimes drawing an analogy to human behavior is the most effective way to pump others' intuition about a particular LLM behavior. The trick is making sure that your audience understands that this is just an analogy, and that it has its limitations.

And it's not completely wrong. Mimicking human behavior is exactly what they're designed to do. You just need to keep reminding people that it's only doing so in a very superficial and spotty way. There's absolutely no basis for assuming that what's happening on the inside is the same.

We do know what happens at higher abstraction levels; the design of efficient networks, and the steady beat of SOTA improvements all depend on understanding how LLMs work internally: choice of network dimensions, feature extraction, attention, attention heads, caching, the peculiarities of high-dimensions and avoiding overfitting are all well-understood by practitioners. Anthropomorphization is only necessary in pop-science articles that use a limited vocabulary.

IMO, there is very little mystery, but lots of deliberate mysticism, especially about future LLMs - the usual hype-cycle extrapolation.

I know nothing about how things work at that level, so these might not even be reasonable questions.

It was already common to use a document extender (LLM) against a hidden document, which resembles a movie or theater play where a character named User is interrogating a character named Bot.

Chain-of-thought switches the movie/script style to film noir, where the [Detective] Bot character has additional content which is not actually "spoken" at the User character. The extra words in the script add a certain kind of metaphorical inertia.

I asked Claude to write a E-AC3 audio component so I can play videos with E-AC3 audio in the old version of QuickTime I really like using. Claude's decoder includes the ability to write debug output to a log file, so Claude is studying how QuickTime and the component interact, and it's controlling QuickTime via Applescript.

Sometimes QuickTime crashes, because this ancient API has its roots in the classic Mac OS days and is not exactly good. Claude reads the crash logs on its own—it knows where they are—and continues on its way. I'm just sitting back and trying to do other things while Claude works, although it's a little distracting that something else is using my computer at the same time.

I really don't want to anthropomorphize these programs, but it's just so hard when it's acting so much like a person...

I’m sure you knew that your code was running on computers somewhere even when you first started and wasn’t running in a literal “cloud”.

It’s about as tiring as people on HN who know just a little about LLMs thinking they are sounding smart when they say they are just advanced autocomplete. Both responses are just as unproductive

It's reply isn't actually going to be why i did a thing. It's reply is going to be whatever is the most probably string of words that fit as a reason.

Meh, I just knew that the browser would display HTML if I wrote it, and that uploading the HTML files made them available on my domain. I didn’t really think about where the files went, specifically.

Try asking an average high school kid how cloud storage works. I doubt you’ll get any further than ‘I make files on my Google Docs and then they are saved there’. This is one step short of ‘well, the files must be on some system in some data center’.

I really disagree that “people who come on HN and say “there is no such thing as serverless and there are servers somewhere” think they are sounding smart when they are adding nothing to the conversation.” On the contrary, it’s an invitation to beginning coders to think about what the ‘serverless’ abstraction actually means.

But let me substantiate before you (rightly) accuse me of just posting a shallow dismissal.

> They don't want to.

Who's they? How could you possibly know? Are you a mind reader? Worse, a mind reader of the masses?

> It seems a lot of people are uncomfortable and defensive about anything that may demystify LLMs.

That "it seems" is doing some serious work over there. You may perceive and describe many people's comments as "uncomfortable and defensive", but that's entirely your own head cannon. All it takes is for someone to simply disagree. It's worthless.

Have you thought about other possible perspectives? Maybe people have strong opinions because they consider what things present as more important than what they are? [0] Maybe people have strong opinions because they're borrowing from other facets of their personal philosophies, which is what they actually feel strongly about? [1] Surely you can appreciate that there's more to a person than what equivalent-presenting "uncomfortable and defensive" comments allow you to surmise? This is such a blatant textbook kneejerk reaction. "They're doing the thing I wanted to think they do anyways, so clearly they do it for the reasons I assume. Oh how correct I am."

> to any notions of trying to bring discourse about LLMs down from the clouds

(according to you)

> The campaigns by the big AI labs have been quite successful.

(((according to you)))

"It's all the big AI labs having successfully manipulated the dumb sheep which I don't belong to!" Come on... Is this topic really reaching political grifting kind of levels?

[0] tangent: if a feature exists but even after you put an earnest effort into finding it you still couldn't, does that feature really exist?

[1] philosophy is at least kind of a thing https://en.wikipedia.org/wiki/Wikipedia:Getting_to_Philosoph...

With such strong wording, it should be rather easy to explain how our thinking differs from what LLMs do. The next step - showing that what LLMs do precludes any kind of sentience is probably much harder.

Anthropomorphizing might blind us to solutions to existing problems. Perhaps instead of trying to come up with the correct prompt for a LLM, there exists a string of words (not necessary ones that make sense) that will get the LLM to a better position to answer given questions.

When we anthropomorphize we are inherently ignore certain parts of how LLMs work, and imagining parts that don't even exist

A useful anchor concept is that of world model, which is what "learning Othello" and similar work seeks to tease out.

As someone who worked in precisely these areas for years and has never stopped thinking about them,

I find it at turns perplexing, sigh-inducing, and enraging, that the "token prediction" trope gained currency and moreover that it continues to influence people's reasoning about contemporary LLM, often as subtext: an unarticulated fundamental model, which is fundamentally wrong in its critical aspects.

It's not that this description of LLM is technically incorrect; it's that it is profoundly _misleading_ and I'm old enough and cynical enough to know full well that many of those who have amplified it and continue to do so, know this very well indeed.

Just as the lay person fundamentally misunderstands the relationship between "programming" and these models, and uses slack language in argumentation, the problem with this trope and the reasoning it entails is that what is unique and interesting and valuable about LLM for many applications and interests is how they do what they do. At that level of analysis there is a very real argument to be made that the animal brain is also nothing more than an "engine of prediction," whether the "token" is a byte stream or neural encoding is quite important but not nearly important as the mechanics of the system which operates on those tokens.

To be direct, it is quite obvious that LLM have not only vestigial world models, but also self-models; and a general paradigm shift will come around this when multimodal models are the norm: because those systems will share with we animals what philosophers call phenomenology, a model of things as they are "perceived" through the senses. And like we humans, these perceptual models (terminology varies by philosopher and school...) will be bound to the linguistic tokens (both heard and spoken, and written) we attach to them.

Vestigial is a key word but an important one. It's not that contemporary LLM have human-tier minds, nor that they have animal-tier world modeling: but they can only "do what they do" because they have such a thing.

Of looming importance—something all of us here should set aside time to think about—is that for most reasonable contemporary theories of mind, a self-model embedded in a world-model, with phenomenology and agency, is the recipe for "self" and self-awareness.

One of the uncomfortable realities of contemporary LLM already having some vestigial self-model, is that while they are obviously not sentient, nor self-aware, as we are, or even animals are, it is just as obvious (to me at least) that they are self-aware in some emerging sense and will only continue to become more so.

Among the lines of finding/research most provocative in this area is the ongoing often sensationalized accounting in system cards and other reporting around two specific things about contemporary models: - they demonstrate behavior pursuing self-preservation - they demonstrate awareness of when they are being tested

We don't—collectively or individually—yet know what these things entail, but taken with the assertion that these models are developing emergent self-awareness (I would say: necessarily and inevitably),

we are facing some very serious ethical questions.

The language adopted by those capitalizing and capitalizing _from_ these systems so far is IMO of deep concern, as it betrays not just disinterest in our civilization collectively benefiting from this technology, but also, that the disregard for human wellbeing implicit in e.g. the hostility to UBI, or, Altman somehow not seeing a moral imperative to remain distant from the current adminstation, implies directly a much greater disregard for "AI wellbeing."

That that concept is today still speculative is little comfort. Those of us watching this space know well how fast things are going, and don't mistake plateaus for the end of the curve.

I do recommend taking a step back from the line-level grind to give these things some thought. They are going to shape the world we live out our days in and our descendents will spend all of theirs in.

exactly. The opposite is also true. You might supply more clarifying information to the LLM, which would help any human answer, but it actually degrades the LLM's output.