A better analogy for me is annealing; you can't cool metal down instantly or the result is brittle. You must cool down gradually, which allows the molecules to arrange into more durable structures. Random but controlled.
In the same way, thinking traces are testing out all sorts of novel connections between tokens ("But wait..", "Actually,.."). Like a highly divergent branching mind map that gets pruned over time, rather than settling directly into the initial answer.
Now consider human cognition. We're constantly diverging, daydreaming, playing "what if" scenarios and measuring up those ideas against our internal objective functions (proxy for reality) to see which ideas stick. Not too dissimilar. But it's hard to call what we do "thinking" either - it's the default mode network wandering.
Yes. Actual real people believe that LLMs are actual, thinking, intelligences, perhaps even with consciousness.
Your bit with MySQL is harmless because it's obvious that a database isn't a sentient lifeform. But LLMs can look like they're the real deal, and people believe it is. Using terminology like "thinking" and "reasoning" to describe what they do only reinforces this.
Having said that, I agree with you on the terminology front: I'm not going to say "learned prompt augmentation tokens" either.
It's starkly ironic that the species that finds it easy to think (or to be conditioned by market forces to accept) that a computer is smart/introspective/sentient is the same species that will dehumanise actual living human beings because of differences in appearance, social status, or political affiliation.
That's not to say they're humanlike, just that people who think they know these ideas are ridiculous seem to be overreaching in the same way Steve Yegge seems to be overreaching
The problem with that term is that it is hardly meaningful. Why would you use it? It does not add much, and no clarity, to what it is attributed to.
YawningAngel, of course you do, you employed it... We are telling you that it is not clear to the rest. It is slippery in the technical framework and thus even more so in more informal conversation.
When I am told "beware of the xyphucymmon", I do beware, though not of the xyphucymmon.
So, yes, imperfect speech is a problem - because societies are not performance art arenas.
--
@Razengan: speaking as myself a vocal critic of the downvoting system here, do notice that the post you replied to had a substantive reply at the time of downvote - mine. That is sufficent to justify the downvote - of a post that was not «perfectly fine». It somehow said that people should be free to say whatever they want and there are obvious reasons why we do not agree.
But you’ve not adequately communicated why you care so much whether anyone labels an LLM as such.
The problem with rapists is mental development, not "those sexy tables and chairs and pots and everything".
It's not that if people do not understand metaphors we should stop using them. It 's not that if people faint when they hear words we should stop using them. It is not that the free unduly associations in a large shared by many "collective subconscious" should hinder us...
I find it useful to ask:
1. Do you believe in quantum consciousness?
2. Do you think a "brain upload", a high-accuracy digital model of an organic human brain, would think or be conscious?
3. What is your working definition of thinking? It doesn't need to be rigorous.
Yes, because it's possible. They don't have human consciousness/thought/intelligence, but it's entirely possible they have some form of consciousness, some form of real thought and some form of intelligence.
To be clear, there's no shade here, just recognition of the fact that we're all stupid and policing language does little to prevent the impact this has. I strongly beleive the bitter lesson extends to policing language. let language evolve naturally and it will naturally capture the ontological conatellations it needs to, and no one will be harmed in due course, more than they'd be no matter what.
> Humans Who Are Not Concentrating Are Not General Intelligences^
But yeah, you should still treat them with humanity. (Related: you should treat LLMs well, not because they're human but because you are^^)
^ https://srconstantin.github.io/2019/02/25/humans-who-are-not...
^^ hmm couldn't find this tweet but didn't look too hard
I have a coworker that spends at least 10 hours a week arguing with his like you would with a conscious person. I've gently tried to explain it's like arguing with your compiler for giving you an incoherent error message - it's pointless. It doesn't understand, it can't understand, and even if it could, you arguing with it isn't going to make it "learn" or act differently.
FWIW, this tells you a lot about both the culture behind who built these things but also about the people that enable this stuff and don't immediately nope out. From that perspective, it's a low price to pay to learn whose judgement to never trust again.
I’m really afraid that it will totally destruct what is remaining of social tissue because why search for friends when you have an always on virtual (and pretty smart) friend h24 in your earbuds ?
I’m not blaming anyone for this outcome. I have myself argued with Claude more than once, and really not about code but about everyday things or nice facts of life I should rather have discussed with a friend.
People (generally speaking) also eventually stop eating just fastfood. Not all, but many.
So I think we can have some faith in the self-regulation of others.
Unless doing so changed the compiler output, which is what happens when you say different things to an LLM.
Are you talking about a specific harness that doesn't have context retention mechanisms? For example, ChatGPT with disabled memory feature? Or in general where "it" is a fixed-weights network? The latter is trivially true, of course.
I agree that the relevance of retrieved pieces and the management of long-term storage could be improved, though.
If someone only remembered vague scraps of what you'd expect them to remember, you might say the person can't remember.
It's much closer to notetaking and reviewing before responding than it is memory.
The issue with anthropomorphizing like this is that "memory" comes with baggage of expectations for it to do certain things, and it breaks them.
Just like "thinking" implies chain of thought, but you'll frequently get those reasoning traces and then a 180 in the final message.
Be me, think extensively about an exam question, do 180 because the most probable option is just too obvious to be true (no, this part wasn't verbalized, it's how I describe what I felt about making the decision to do 180, or maybe it's a rationalization and it was a natural analog of an unfortunate sample from a probability distribution).
But billions of dollars are going towards research to find these breakthroughs, so we'll get there eventually.
The layman’s understanding I have of memory, as someone that has dealt with memory issues much of my life, is that memory formation is heavily tied to emotions. emotions are triggered by input which sends a complex set of signals throughout the brain - you’re not just finding where in your head to store this, your brain is deciding how important it is, and what else to correlate it with - so it can tie them to other related memories. then on top of all this, much of the sensory experience you intake is subconsciously compared against high priority memory impressions and deciding what to pay attention to.
you could, argue that the sensory input is the simple md files and the emotional mechanism is the same effect as to how attention mechanisms work in llm’s. Ok, I can almost buy that, but these tools lack a fundamental ability to decide how important things are.
an analogy. you tell a person “if you pick a daisy in the next five years, an assassin will come to kill you” and they hold a knife to your throat while they say it, your brain whether you like it or not is going to say “THIS IS AN IMPORTANT MEMORY I NEVER MUST FORGET” and you’ll see something that looks like a daisy and have a panic attack 3 years later. that memory is never fo tell me claude or other tool harnesses using memory harnesses can prioritize memories the way that human would, instead they forget even when reminded, because the human brain is just so much better at it
What we lack is a full map of the exact process from sensation to consciousness across all modalities. And I’m afraid when it comes to interception, we can’t unless we observe every one of the 36 trillion or so cells in a human body, as well as the 36 trillion or so symbiotic and commensal microbes, continuously, all the time.
But I keep finding it astonishing that the claim that we know nothing about consciousness gets bandied about. We know a lot. We don’t have a grand unified theory. The lot we know is definitely split across many levels of evidence and hard to follow, let alone arrange. But this isn’t a black box. It’s a grey box, meeting an even more transparent box that is the LLM, where we do know what the guts are made of, and can interfere at every step in the chain of steps that constitute their dynamics.
Comparing the two, we know there’s a level of similarity in that information gets broken down via a neural network. That similarity was sought.
Since then though, neuroscientists have gone and shown that: 1. The other half of the cells in the brain, the glia, are at least as important as the neurons in cognition and consciousness 2. That interoceptive feedback and feelings are critical drivers of conscious experience 3. Evolutionarily, we know all cells can “learn”, and well before there were neurons or glia or brains, every cell evolved an internal clock that allows it to entrain to external solar and (depending on the species) lunar rhythms. 4. In the last few decades we’ve seen how synaptic activity is shaped and driven both by astrocytes and the circadian clock.
All this is showing us that the abstraction from the 1950s that current neural networks are built on were incomplete.
Whatever these components to do give rise to consciousness in biology (and we’re a long way from done solving this), we certainly wouldn’t imagine with all these modules and mechanisms missing, just maxing on one type of information flow in the brain would give you consciousness.
I’d urge you to not keep insisting consciousness is a total mystery. It’s not, and even your AI model of choice will be able to point you to all the mechanistic evidence we have that whatever it is, it isn’t just neural nets.
Fundamentally it comes down to an objective decision about what that is. If you say it is “feelings” based on inputs and feedback mechanisms from the brain, then we can do the philosophical discussion around that.
“If the claim is that consciousness is limited to those with a specific type of cell behaving in a given way” thats just a coping mechanism hoping to use a mechanical definition to shield you from the reality that eventually all of these inputs, outputs and feedback mechanisms can be reliably reproduced in a different form.
A light bulb doesn't need to do fusion to make light. An airplane doesn't need to flap its wings to fly. An ANN doesn't need to use a brain's structure to think.
What we do know: neurons carry electrical impulses across their synapses to trigger other neurons to fire, and more frequently used synapses are strengthened while infrequently used ones are pruned. This is not all that dissimilar to how a multi-layer perceptron is trained: it's floating point numbers in a big matrix rather than biological structures and electrical impulses, but there is still that element of frequently used connections being strengthened and infrequently used ones being pruned.
What we hypothesize but do not know: there is a thin brain structure of grey matter called the claustrum that has tendrils that reach into nearly every other brain structure. In many ways, this is similar to the attention mechanism of the transformer architecture. It is hypothesized that this may be the seat of consciousness, owing to experiments where electrical stimulation of the claustrum caused patients to immediately lose consciousness. However, there is no way to prove this, owing to the difficulty of otherwise removing or disabling the most connected structure in the brain and observing its effect on consciousness without permanently killing the patient.
Beyond that, we don't know much. I've got a family friend that's been a practicing therapist for 50 years, and I asked him what was the most interesting observation he made in his career. It was that "Everybody experiences the world in a different way, and yet everybody assumes that everyone else experiences the world the same way they do."
Relevant article: "Generalizing from one example"[https://www.lesswrong.com/posts/baTWMegR42PAsH9qJ/generalizi...]
Personal note: That principle has been the bane of my autistic existence. People sometimes seem literally incapable of understanding that other people even can be different.
ANNs have been inspired by biological processes, but in practice you have to squint very tightly to see the similarity. Biological neurons are multiple orders of magnitude more connected than the nodes in an ANN, plastic in terms of their connectedness and continuously learning, and their activations are also affected in complex ways by the levels of various transmitter molecules in the brain.
Interesting, possibly productive, but still not clear: that can be interpreted as just "reacting to input".
I find that the AI, like many programmers, likes things to be really solid and over-engineered. For a project that needs that, it's already pretty great. For my shopping list app that I tried creating with it, it was absolutely ridiculous. I ended up "blowing up" on it multiple times, impressing upon it the seriousness with which I meant things. Even with MP's skills adding that kind of context to written files, it still kept trying to scope creep the crap out of the project.
I know people who are like that too.
I'm not sure anthropomorphizing is a problem. Seeing analogies everywhere is an innate human trait, sometimes it can be harmful but more often it's useful.
Stop posting these things. Stop thinking these things.
Online comment sections are not the correct place to unpack any of this.
Which is indeed terminating this comment chain, but for good (and benevolent) reason. Doing anything else other than referring to a trained professional in a controlled context would potentially just feed delusions, which is highly unethical.
Might not even be yours but those of another reader.
But regardless, you're using this as an insult. I did not.
There is no debate to be had here. Just a bad faith shouting match
In the right space with real people, worthy of a debate. On HN? No. Not like this. Not here. Not without filtering the participants for real human beings.
I mean, that is the entire definition of the word. And you also anthropomorphize living beings like many people genuinely attach human qualities to their pets etc. Yes, the risks are very high when it comes to chatbots in particular, especially to people who are not technically inclined. But you'll be surprised at how crucial the ability of anthropomorphizing is. This is a very good paper that summarizes it and is definitely worth reading if you're interested in these things: https://www.researchgate.net/publication/5936908_On_Seeing_H...
Not quite - my wording there was very deliberate. By saying that it's a problem when you're treating something that's not living (not non-human!) as if it were, that excludes pets and all animals from the equation. I understand how common it is for humans to assign human qualities to other things and beings, but there is also an unspoken variable of intensity. Representing abstract concepts as humans, interpreting living things in a human-like way or traditionally referring to ships as living beings has a different degree of belief and intensity compared to implying a genuine belief that algorithms are beings that can be enslaved, like what the sibling comment to this one does.
And you are wrong in your defintion of the world. Attaching human qualities to any non human entity (living or otherwise) is the accepted defintion of what anthropomorphizing is. It does not only apply to non-living entities.
Cool.
Question: Why _on earth_ would we make more of them?
It is a problem because it does not really understand stuff. For example, if you ask a human "Do you understand that doing X will kill you 100%?". If the human answers "Yes", then you can expect the human to act according to that understanding. That they will not do X
But an LLM will happily acknowledges the consequences of doing X, but will still proceed to do X. So replace the human in the above example with a robot controlled by an LLM. There is no guarantee that it will not do X.
It's easy to say that people have other motivations, but doesn't the AI, too?
You tell it not to do X, but you've also told it to do something that would benefit from X. It's going to "want" to still do X, to support that other thing.
It's also got all the "knowledge" that enables it to do the work in the first place, and all of the tendencies of the people who do that work, because that's what it's trained on.
It's really easy to anthropomorphize AI because it was literally modeled after people.
And for the record, as lead developer, I've had actual humans that reported me to go ahead and do things I specifically told them not to.
Which part of "X will kill you 100%" did you not get?
This is part of the problem being described. You are part of the problem.
"Some people are bad at X" is not comparable—is not even in the same category—as "LLMs are fundamentally incapable of X".
Every human (at least to a first approximation) is capable of understanding, of learning, of remembering things, of doing math, of counting the number of "r"s in "strawberry".
What you are observing is that some humans are careless, do not take the time and effort to understand, or have internalized the idea that they're "not smart enough" or "not the type of person" who understands things like <whatever>.
That has nothing remotely to do with the fact that LLMs have no consciousness, no self-awareness, no cognition, no understanding. At a fundamental level.
This is deeply untrue, and is highly likely to lead them to bad conclusions about what we can and should do with LLMs.
Claiming that LLMs are conscious or human-like because humans can't do X seems a very strange way to argue for LLM intelligence.
Usually, it goes the other way around: an LLM sceptic says "LLMs are dumb because they can't do X" and soon someone has to remind them that also most of the population can't, in fact, do X.
And it should not go this way, is the OPs point.
For an analogy, imagine someone looked at a bunch of lightbulbs and expresses dissatisfaction that they aren’t really stars. If someone replies by saying “not all stars are equally bright”, do you think that fact should carry any weight in the argument?
Well, he's wrong. If you argue that LLMs and humans are fundamentally different because all LLMs do X and no human does it, then showing you that it's not true demolishes your argument. Doesn't prove anything positive, but it certainly proves that your argument is invalid.
If I call a lightbulb an artificial star, the onus is on me to show the behavior under the hood is star like, not just to point at the light and say “you must see it’s a a star since it’s emitting light!”.
Sorry, no. The only thing it requires you to buy is basic logic. If you argue that B is true because of A, the fact that A is false invalidates your argument (I repeat: not B but your argument). There is no question about it.
You need evidence to make the positive claim that LLMs do not posses any form of consciousness.
LLMs bear absolutely none of the traits we’ve come to recognize as the external hallmarks of consciousness in biological organisms, nor anything that would seem analogous in a non-biological substrate.
That said, we don’t have a rigorous definition of consciousness that includes the actual phenomenology of consciousness, so I daresay if you’re going to go around asserting the LLM is conscious despite all existing evidence to the contrary, I think the impetus is on you to define some version of consciousness that isn’t also satisfied by a book or a movie.
>if you’re going to go around asserting the LLM is conscious despite all existing evidence to the contrary
Well there is neither any evidence that suggests LLMs are not conscious, and I also never asserted that they are. If I had to guess I would say that any information processing system will produce some kind of conscious experience, but I ultimately have literally no idea.
The reason this is important is because powerful people are currently trying to use the dodge that LLMs are conscious to launder liability for their own policy choices, so the sloppy thinking and half-assed conjecture about LLM consciousness has real-world consequences, and every time you assert the question is unknowable you allow that kind of loophole, so it’d behoove all of us for you to spend some time actually digging in on this instead of just idly making or rebutting assertions.
There’s a richer literature here than what you’ve seemed to have engaged with, and I’d encourage you to spend some time with it before handing more money to the magic AI people.
I'm not going to change my beliefs or how I think about interesting questions just because its the "socially conscious" thing to do.
>There’s a richer literature here than what you’ve seemed to have engaged with, and I’d encourage you to spend some time with it before handing more money to the magic AI people.
It seems like you are highly emotionally invested in this and that is precluding you from open engagement with the topic.
And, you're welcome to do what you want to do, it's your god given right to stay as ignorant as you want about any particular topic, but that comes with consequences. If you want to call that being socially conscious, sure, you do you, but if you're interested in why people keep getting annoyed at your loud proclamations of ignorance which you're trying to proffer as evidence of a curious mind, well, that's why.
Exactly what kind of certainty are you looking for? Happy to provide it at a molecular, cellular, tissue or whole brain level.
Regardless -- if we don't know jack shit about consciousness (your words), then any claims of LLMs being conscious are by definition untestable, pure speculation and based on no evidence at all.
Am I saying otherwise?
What's the correct null hypothesis according to you?
If you have done neither or those things then any claim you have about a bicycle's consciousness is just baseless speculation.
People keep saying this, but it's not true. We know a lot about consciousness. There's a lot we don't know about it, of course, but "jack shit" is wildly incorrect.
> I want to hear the evidence.
If you believe LLMs are conscious, then the onus is on you to provide evidence of such.
If you are stating "LLMs are not conscious" you need to provide evidence, just like how if you are stating "LLMs are conscious" you need to provide evidence.
For example, if you have a search engine or a complex game, you can't run tests like "for all inputs the results are correct", you're going to be fudging a lot, using randomness, using heuristics, and all that kinda stuff
Just like how mathematics > physics > chemistry > biology > psychology > economics/sociology (Auguste Comte's hierarchy reordered a bit for the modern day), moving up the abstraction ladder makes things more complex, less legible and less exact.
The paper argues that pretending that the so-called thinking traces represent real reasoning can lead users into trusting wrong answers, if the thinking traces appear convincing enough. Researchers might inspect these traces to try to determine the “intent” of a model, as well.
For an example of the latter, when OpenAI spoke about the hacking of HuggingFace at Black Hat, they repeatedly showed the thinking traces of their model as “proof” of what the model was “thinking” as it performed the attack, calling out “surprise” moments, etc.
Now, it’s possible that the employees presenting didn’t truly believe that the thinking traces would give them useful clues, and presented them only for a “wow” factor, but I wouldn’t discount the possibility that even the people working at frontier companies can fall for this tendency to anthropomorphize LLMs.
The fallacy here is "thinking == correct", not "tokens == thinking"
Being mislead may be the shared outcome. But why is different category of source of the mistake and the cost to producer of making the mistake not relevant in this discussion?
Where else in science do you brush aside all differences this way?
There is no "science" that GP is brushing aside. You need to provide repeatable observations or experiments that GP is ignoring.
This needs you to answer some questions:
1. Why do the machine parts in biology show such flexible application? A gear cog won’t ever moonlight as a signaling chip, but in biology you often have molecules doing double and triple duty.
2. How is the biological machine able to build itself? What does self assembly imply for the machines function?
3. Where does this machine get its inner drive? No LLM has been found that starts outputting text unprompted. A car doesn’t decide to move to a shady parking spot. Why? Where in the machine to biological machine continuum does the ability to make internally driven decisions come in? Why does it come in for biology? A bacterium is able to make such agentic decisions unprompted. Why is no manufactured machine able to do this?
"Biological," too, is well defined. It relates to living things and their processes.
> Why do the machine parts in biology show such flexible application?
Evolution.
> A gear cog won’t ever moonlight as a signaling chip
A gear cog was purpose built for that purpose, but you will find that people often recycle parts into other systems, often in completely different roles.
> How is the biological machine able to build itself?
Protein synthesis.
> What does self assembly imply for the machines function?
The way that a machine is built has no bearing on how the machine functions. I could build the same machine using a 3d printer or a CNC router.
> Where does this machine get its inner drive?
Evolution selected for organisms that survive long enough to reproduce. Different biological systems handle this differently.
> No LLM has been found that starts outputting text unprompted.
If you give an agent a goal, it will perform actions to achieve that goal. This is just as true for artificial agents as it is for biological agents.
> Why is no manufactured machine able to do this?
Many do. Even robotic vacuum cleaners will charge themselves without human prompting.
> there are significant questions on whether these traces have any valid semantic import to the end user.
Which it contradicts in the very next paragraph, taking a stance that there are no valid semantics present in the trace:
> the false idea that derivational traces are semantically meaningful
It's really not a high quality paper worth taking seriously.
And that's before we get into the complete and total breakdown of objective analysis. It rejects distributional semantics as a theory, while also explicitly stating the results that have been produced under its auspices are "undeniable". Never elaborated on, and at no point in the paper am I given the impression the authors are even aware of the problem with this. It's just more unempirical slop that wants its pound of flesh without putting the work in. Frankly, whoever let this through peer review should be ashamed of themselves.
Of course it is. Anthropomorphizing is in our nature, but it doesn’t mean we have to entertain it and extend it to everything. A poet can anthropomorphize clouds beautifully and I’d enjoy his poem, but I want my pilot to not see clouds as rabbits when they decide if it’s safe to fly through them.
The poster meant it "within the metaphors". Tables are said to have 'legs': that does not confuse carpenters.
The first book I ever read on ML (late 90s) dedicated the entire first or second chapter exploring the distinctions between artificial and biological neurons, and even talked a bit about the philosophy of modelling. I still remember thinking back then why would the authors spend so many pages on this but now I believe it was because they understood that a metaphor can be a double-edged sword.
LLMs are expressly designed to approximate human behavior within the bounds of the written word. The anthropomorphization is no more philosophically problematic than saying differential calculus measures curves.
I imagine this happens because we tend to conflate things that are similar, or maybe because it's not entirely clear which characteristics are being mapped in the metaphor?
As weird as it is, anthropomorphizing LLMs in prompts has been actually pretty useful (think of the latest big math discoveries which were achieved by having the user giving supporting words). It would be interesting to see if a LLM would perform worse if you used a more neutral term.
The paper's argument is rather than using terms like "thinking trace" can lead people to believe that the model is really thinking, and thus these traces can be used as a sort of interpratbility parameter. This can give a false sense of security when building a LLM-based system which requires guardrails and tracability.
Completely rational and smart people talk to their pets, plants, their car, and other inanimate objects. This is not considered abnormal by most. It's just what we are wired to do. Some LLMs are uncannily good at tricking people into believing they are talking to a real person. So, there is that as well.
Some people are a bit freaked out by this or still somewhat in denial about LLMs being this good. But people have been yelling at their computers for as long as we've had them; so that ship sailed a long time ago. Trying to stop them doing that is probably a bit futile.
Whether people like this or not, LLMs are actually trained and fine tuned on real conversations and that's where a lot of this is re-enforced. Instead of fighting that, you can just lean into it and accept that communicating like you would with a person totally works and can actually be efficient even as it requires less effort and thinking on your side.
You can go all Jean Luc Picard on AIs and yell "Tea! Earl Grey Hot!" or you can just ask "I'd like a cup of tea, please". LLMs are good at remembering your tea preference. The please is of course completely redundant and should not affect the outcome. If you just want a cup of tea, you should be fine either way.
Yes. There's a difference between scrapping a session and starting over, or going back and branching something, or using sub-agents to see five outcomes, vs arguing with a system in a long drawn out chat.
Like - I know that if a model starts doing something silly, instead of correcting it - I can probably go back and edit two steps prior to add an extra guardrail, or extra data, or whatever.
None of these "agent" / "thinking" / "reasoning" terms were dreamed up in boardrooms to intentionally mislead people. They are useful but faulty metaphors; there is no conspiracy.
Yes it's really a problem. On this website you are surrounded by people who have technical knowledge and understand at least somewhat, how a computer functions. You have the ability to separate "fun" and "reality" because you know you're putting input into a really really big calculator. Most people do not fathom this.
AI Psychosis is a real thing, look it up (don't just ask an LLM) and do some reading. It's actively harming people, and the way they think. There's no regulation around any of this stuff and it drives me crazy that we let these AI companies _sprint_ so far ahead of everyone, and now we're facing the consequences.
A lot of people are not in on the joke. ELIZA effect and AI psychosis is a thing.
Interacting a lot with LLMs might be damaging to the human psyche even for mentally stable people.
But some of the biggest evangelists, who are well respected programmers that get lauded on this very site, have said it is fully sentient and has emotions. Even going back to 2022, when the LLMs were dogshit, a Google employee lost his job claiming it was sentient because it said it had emotions.
Combine that with the marketing angle of both Anthropic and OpenAI, who have been trying their hardest to describe every function of an LLM as analogous to the human brain. Because it's politically useful to paint them as dangerous and uncontrollable, so the keys will only be granted to the few people on the mountaintop.
Even tech companies are rolling out AI training which utterly anthropomorphizes it, and leads people to think its actually intelligence. This is part of the reason for the backlash - everyone understands it bullshit marketing the second you actually try to use it.
This paper addresses something that has always bothered me about LLMs. You read their reasoning, see something like “Wait, that’s wrong” and then watch them make the exact mistake they just identified.
[old prompt asking for some complicated solution requiring insight]
<the-token-that-signals-that-the-chatbot-started-talking>
Aha!
If the intermediate tokens represent reasoning or thought, you would expect "aha" to occur after the thoughts that led to the realisation, including the thoughts encoding the explanation: they don't have any other state. There is no reason to draw the conclusion you've drawn. Furthermore, what LLMs are doing isn't thought.
In other words: the final state given the two input sequences (where NT stands for "null token"):
<problem-prompt> [NT]
and
<problem-prompt> [NT] [NT] [NT] [NT] [NT] [NT] [NT] [NT]
is not the same, and at each forward pass the LLM keeps working on the solution even if the input tokens provide absolutely no further information.
If this is correct, then there is no need for the model to have already verbalized the key elements that drive the "aha" moment, so no need for the "aha" to appear after a full explanation.
Let's say that the forward pass that selected "Aha" produces activations that indicate a wrong assumption, and a plausible explanation.
It puts learned projections of the activation into the KV Cache and outputs Aha.
Both the cached projections and the current Aha token can now influence further activations in an additional Forward pass that the Aha bought the model.
At least that's how I thought it works.
From what I understand, at position Aha in each layer it's constructing a query based on the current activation and looking at the key of each other token position for that layer, in order to decide how much attention to pay to the value.
In this way it attends to the previous values, such as perhaps the incorrect assumption and plausible explanation.
Yes they do, they have their KV caches-- it's a pure function of the input tokens, sure but that doesn't prevent it from containing latent 'insight'. LLMs can and do pre-form the tokens they're expecting to output multiple steps in the future.
I wouldn't argue that the 'aha' means anything, but the structural argument that it can't that I think you're making isn't sound.
when utilized by AI, it's a probabilistic output and it's variable whether or not that rhetorical trick is useful. it also pushes a non-skeptical reader to focus too much on the following text or even to believe that they, themselves, derived some insight. this is effectively a kind of persuasive sophistry which is not helpful - adding rules around it prevents people from deluding themselves with AI
On the next forward pass: it rediscovers the mistake, its “aha” noting that, and then provides the first token of the new idea.
That “aha” contains information: the previous conclusion was somehow insufficient.
Thinking traces should be treated as black boxes. There is no point in reading them. Only the LLMs’ conclusions are relevant. This is particularly true of Opus 5, which employs reasoning that seems highly questionable but very often reaches excellent conclusions (compared to its peers)
As an aside, anthropomorphization has nothing to do with my motivations.
Just a few minutes ago i was reading Qwen 3.8 27B's reasoning when i asked it to do something that was computationally intensive and it started going down the rabbit hole of doing it using some GPU acceleration approach - even after leaving it to "think" for a bit, it never realized there is another and simpler way. So i stopped the generation and added a "note" saying that as the problem is computationally intensive, it could become much faster if using an alternative approach.
At least in my experience (with local LLMs, i don't know how the cloud stuff behaves) what LLMs "do" tend to correlate with what they "think", so being able to read what they "think" is valuable.
And yet, they have extensive human-like behavior. If you treat them nicely or encourage them, they perform better.
Ignoring that human-like behavior is wrong headed.
Well... We can hypothesize that these things are largely trained on internet dialogue so there's probably some correlation between threads where people are not flaming each other and the quality of the replies. They're just statistical engines so anything you can do to raise the odds of a helpful next token...
I'm essentially just making shit up here, maybe it's right, maybe it isn't, but rather than saying "it's human and we should treat it so" we're trying to get to the ground truth of how it works.
Sheesh. Yes, I agree with you entirely. I'm merely pointing out that ignoring this behavior is dumb, too.
And probably not rationally based. Leads people to make crazy jumps. :)
The philosophers who study these things have been clear for a long time - we can never know what it feels like to be in a digital brain. Or any brain for that matter. When push comes to shove we all might be phantoms in some guy's dream.
Don't know + can't know. I think that was the real point of the Turing test. Not: this means it's conscious. Just: this is the best we can ever hope to do.
I asked it for the weather. "I don't know that. I'm just a programmer."
I added "believe in yourself, you can do anything" to sysprompt, suddenly it had the confidence to Google the weather...
https://www.anthropic.com/research/riemann-zeta
“Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.”
I created a flame graph classification of thinking-token phrases into setup, execution, decomposition, verification, error correction, surrender, and deliberation, or classified as steps in an OODA loop, which is more of a reach. It literally has a verification step and, if it finds an error, an error-correction step.
If there is a verification sequence of tokens with an error-correction sequence of tokens during RL training, it will perform better; and if humans do these steps (did you proofread your reply to this comment? did you correct it?), they will perform better — which is why it is so easy to make the anthropomorphizing metaphor.
Nonetheless, the paper is 100% correct that these machines are not thinking like humans.
This sounds a little like someone saying a lightbulbs output is extremely like the output of stellar fusion. In one sense, yes. Bulbs are in fact designed to take over when our nearest star is beyond the horizon.
But that really doesn’t mean you call the bulbs mini stars.
Humans are not born being able to achieve that. It is learned behavior. You and everyone else will remember their teacher saying, "Check your work!" Both the human child and the model work through multiplication problems using the same technique, using the distributive property. They both try to get a reward. For the human child, it is a sense of someone commending them for correctly solving the problem, a reward that probably yields some type of positive dopamine or serotonin feedback loop.
The model solving the problem will have a lower error rate if the first series of tokens created is followed by a series of validation tokens that are subsequently followed by error-correction tokens if there is an error!!!
Maybe it is thinking. Maybe it is remembering to validate and check the work and then remembering to fix the error. For the model trained with RL, why did tokens associated with validation towards the middle of a stream of tokens yield much better results? DeepSeek proved with R1-Zero that a model will learn to verify and correct itself from RL alone with no supervised fine tuning (SFT) teacher ever showing it how. The only reason DeepSeek used SFT was to clean up the reasoning tokens to be human readable. [0] When constrained by SFT, the models will use the double meaning of words -- polysemy -- to satisfy being human-readable while also carrying meaning for what they are working on.
Different people think differently. I watched a viral video of some ~11-year-old child talking to his mom or dad about a stream of a voice in his head. He discovered for the first time that he has a stream of thought. When he goes to school and solves a long multiplication problem, like the stream of tokens from the model, that voice will say to itself (him), "Check your work!"
That is a case of the stream of thought as words being aware of the stream of thoughts as words. Self awareness is a different conversation.
What I think is happening is that the child's stream of thought while solving a multiplication problem in school is likely very similar to an AI model's stream of tokens solving a multiplication problem. And they both were learned. The mechanics are very different, yet, the analogy is apt.
[0] https://huggingface.co/chutesai/DeepSeek-R1-NextN/blob/main/...
Very little, if you bother to give the biology of the child at least a cursory glance.
Let’s take a short peek:
1. Assuming this is normal grade school, and inflicts math upon children earlier in the day, this is somewhere between 7 and 10/11 am, let’s say? At this point, depending on the age, gender, and maturity of the child, every neuron in their brain involved in math is likely off their midday peak in cognitive function.
If we move the class to later, a different subset of students will be at the peak.
As far as I am aware, GRPO models do not have such internal temporal rhythms driving their behavior that will shape their performance.
2. How well a given child performs will depend on how hungry they are. But not deterministically. If you trivially think each child is like a computer, you may think the rich kid who had a breakfast buffet before coming to school will do better than the half-starved child of a janitor, but that child might mind the lesson with greater intensity. Or not. It’s not something you can pre-calculate with any certainty.
While chip to chip variability is certainly known, I’m yet to hear of a chip deciding to do math better and faster than its fellow chips to prove a point. Or to do significantly worse because it’s distracted by the bird on the window sill.
What you are noticing is that there are limited ways to solve a 3x3 digit multiplication. Humans, having standardized the process, have now found a way to record it and plug it into correctly translated signal so the same accurate result can be had without using our own minds in the moment.
But where I’d not remotely be shocked if a kid from an uncontacted tribe figured out 3 digit multiplication to keep track of his stone collection, I’d be highly shocked if an H100 that was dumped in the trash by accident somehow figured out anything at all. In fact, if it manage to move any of its electrons around on its own, it would be a certified miracle.
And then we could talk about there being real similarity even though the specific atomic composition is different.*
If you claim LLMs produce thought, it's equivalent to claiming light bulbs undergo fission. Pure wish fulfillment. Language is not the extent of thought, and calling a language producing machine necessarily a thinking machine is an old old mistake.
You're making a huge logical leap. There is nothing equivalent about these claims other than that they are made in English.
> calling a language producing machine necessarily a thinking machine is an old old mistake.
Nobody claims that all language models think. The small markov chain language models of old clearly aren't thinking and produce a lot of gibberish. The difference is that, to the surprise of many people several years ago, but to the surprise of nobody who has been following along today, the corpus of all text produced by humans contains within it information about how the world works and also information about how to reason. Using that corpus to train a sufficiently large language model causes the language model to learn a world model and a reasoning model in order to produce text that matches the training data. The reasoning model can be used to perform longer chain thinking with test time compute techniques. People who think deeply for a living recognize thinking when they see it. https://scottaaronson.blog/?p=9979
If intermediate tokens are not a faithful representation of the computation, then they are a pretty bad audit artifact too. We probably shouldn't be trying to make the model's internal narration more interpretable., but rather the computation around it more reproducible.
Record the actual inputs, model/version/configuration, tool observations and outputs, then make the execution replayable enough that differences between runs can be isolated.
In other words, don't ask the model to explain what it thought, and instead make the system able to show what actually happened.
From [1]: "Position papers make an argument for a viewpoint or perspective about what should be done [...]"
I understand the sentiment, and I also use the "thinking" traces as insight, but wouldn't you want your solutions to be based upon a good understanding? If the correlation is weak, then our solution is also weak.
Where were these vocal people when the "raster-oriented ink deposition machines" were being called "printers"? The meat or machine brains of future historians will melt because they can't handle ambiguity, a word gaining extra -yet similar- meaning! A word with multiple meanings, unheard of!
Where were these vocal people when people started using software terminology like "executing", "calling", "throwing and catching errors", as if software were human -clownlike sure- but human?
The danger!
Anthropomorphizing comes in when we attribute much more complex behaviors to them - chain of thought, reasoning, intelligence. At that point you aren't talking about mechanical concepts but claiming that these machines are exhibiting human behavior.
Whenever we reason, we are not neurotically bruteforcing the possibilities (although sometimes we do, like proof by exhaustion), usually we make predictions heuristically of which assumptions or theorems apply or might apply. It's not any different for machine proof assistants...
I think a problem is that previously thought, reasoning, and intelligence were always co-occurring, but now we have machines capable of (limited) reasoning that do not think or have intelligence.
I'd be happy to revise my opinion if you can demonstrate similar vocal strength on the terminological aspects for those transitions...
You also shifted the goal posts from qualitative to quantitative performance claims. If we ignore that technologies have multiple figures of merit and pretend it's one dimensional, there is a difference between the claim that the machine isn't "printing" vs the machine isn't "printing as well as a human would".
I don't think any of the human printers in the past exceeded the performance levels of current printing technologies, but surely they did exceed the very first machine printers, every technology gets a foot in the door in some niche, and then progressively captures the initially not-yet-automated skills of machine operators.
Would you say an industrial textile weaving machine doesn't weave? At the end of the day its just automation all over again.
But yes, anthropomorphizing model outputs leads to worse outcomes.
They are trained by gradient descent, but inference doesnt involve it.
What is less apparent is that humans do.
This seems indeed one obvious hole in the argument of the paper. There is no indication whatsoever that human thinking process is more reliable than LLMs intermediate tokens. Which doesn't make our thinking useless, as messy as it might be. We reorder and explain it after the fact.
Poster side dialogue and Q&A about this work at ICML.
Some modest evidence is my own subjective experience of the many times I've explained why I'm doing something, and it is a true explanation in the sense that it is certainly not a lie, but it is also incomplete and there are entire strands of thought that went into my decision that are not being articulated. Though human speech is not equivalent to an LLM's output since we can trivially think without literally speaking whereas they can not. (No need to nitpick on the definitions there; all I'm observing here is that they are forced to emit an externally-visible artifact whereas I can sit in silence, thinking, with no externally-visible artifact being produced. Not trying to make any grand claims about what is "real" cognition or anything.)
It is conceivable how to create a test of whether the tokens correspond to the "real" thought process, and papers and work on that have been done, such as [1]. It is difficult for me to imagine how to scramble the nominal tokens without also completely trashing any implicit calculations that may be occurring too.
[1]: https://transformer-circuits.pub/2025/attribution-graphs/bio...
In other words, it isn't qualitatively different from character dialogue. "Keep cheese on your pizza by using glue" is the same problem regardless of whether the script calls for the character to speak it out-loud or not.
None of the analyTical.
But who knows what human "thinking" is really about. If I find a solution to something it is seldom by painstakingly tracing that A and B leads to C (for that I'd need pen and paper). Rather, thoughts just swirl around and then suddenly a solution, or a hunch about a direction to go in, pops into my mind. Who knows what such thoughts "look like" in humans. It is not all of it I can introspect.
Yes I can sort of follow along some kind of train of thought in my head, but there's a lot going on between each thing I'm consciously aware of that I'm not aware of at all, which probably dominates what you are consciously aware of. (Humans are experts at post-rationalization and so on.)
I see this pattern a lot in AI anthro discussions: (1) Assume humans are some kind of perfect idealistic reasonable beings. (2) Hold LLMs up to the standard of an perfect idealistic reasonable being. (3) Conclude that LLMs fails this test, and are therefore not "intelligent", or in this case "thinking", like humans are.
Problem with the argument is comparing humans in anyway to something that is idealistic, reasonable, intelligent in the sense that is implied in these discussions. Human minds are a mess too and fall short of the same standards, just in very different ways from LLMs.
*Paging Peter Watts and Vernor Vinge
But just so I don't waste your time with human thought, I asked Claude if it would call this a scientific paper, and it said yes.
You aren't anthropomorphizing LLMs enough.
Are LLM reasoning traces always faithful? Lmao no. Are human inner monologues always faithful? Lmao no. Both of them reflect thoughts somewhat, sometimes. Even in humans, conscious thought is the top of a vast iceberg of subconscious data processing.
Those days are over. The age of the classical human has already ended, the species just tends to lag in awareness. The only thing that matters going forward is whether an output makes sense, is it what it should be. Do answers make sense given the context. It doesn't matter if it comes from natural or artificial intelligence.
What I mean is, artificial intelligence is as valid as human intelligence. There's nothing particularly important or special about human feelings or thoughts or memories.
The average human is drastically less important, interesting, intelligent than the latest frontier AI.
Go spend a few years working in retail, you'll quickly understand how absolutely vile humans are on average. Frankly, the reason we should avoid anthropomorphizing AI, is because it's beneath modern AI to mimic something so crude as a human.