(*) Even if we hack around this and just do the usual trick of simply laundering statefulness to a higher level, in this case the context window being fed in, I fail to identify (**) a representation of its own state in these bodies of text that it'd be meticulously maintaining. I further fail to identify how it could be hidden or maintained, considering I control like half of it. The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
(**) I did wish for something hidden though, so maybe it's just concealed? The same way people can encode a lot more of their emotional and mental state than normal into text if they read and write a lot of it, I'm aware of research that suggested the same for LLMs, albeit I cannot cite it. Maybe those phrasing signatures are just alien to me and will never pop out. Either way, I'd expect researchers to stumble upon this during interpretability studies, and either they haven't, they have but it wasn't popsci adopted, or they're keeping awfully tight lipped about it. If you know of anything like this, your turn now, would be happy to learn.
I do wonder how reasonable it is to expect e.g. a single maintained identity though. Maybe it isn't?
(*) Another way to hack around this of course is to just precompute some internal "self-awareness states" and hop around between them. Probably the closest to what the models are actually "doing".
> a hidden representation of self that is continually tended to
This sounds like a personality? They act like they have one of those. It may be an illusion, and even if it isn't an illusion it is unlikely to be anything like the source (us), but they act like it.
> I further fail to identify how it could be hidden or maintained, considering I control like half of it.
Indeed you control everything about a local model, and much of the context of even a remote model. But the state of activations and circuits in SotA AI is hidden in similar ways to those of synapses in your head: difficult to decipher even with probes monitoring the signals directly, and often not emitted at the normal output.
> The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
While we can be confident that LLMs make up personas etc., it is insufficient to go from "that doesn't necessarily represent the model's internal state" to "therefore it doesn't have one".
> You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
I've, unfortunately, also experienced this with humans. Perhaps they were losing their self-awareness at the time? I do wonder if old-age dementia does that by the end, though the person in question didn't ever get diagnosed with that.
> If you know of anything like this, your turn now, would be happy to learn.
Do you mean like these, or something else?
• https://researchportal.hkust.edu.hk/en/publications/decoding...
Not quite what I meant, but it's also not entirely unrelated I guess? Personality to me is like a natural bias. It does also shift over time, and is also an internal bit of state. I guess in some respects it can also be self-referential, like personal convictions.
> Perhaps they were losing their self-awareness at the time?
I do think it is entirely possible for people's self-awareness to shift, yes. Or more precisely, I do model things that way.
> Do you mean like these, or something else?
They're adjacent, but I more meant something like these:
https://arxiv.org/abs/2410.03768
https://arxiv.org/abs/2310.18512
https://arxiv.org/abs/2605.26537
So basically, steganography. The difference is that these papers investigate from the perspective of separate LLM instances covertly exchanging information between each other. This is in contrast with the scenario I'm laying out, where an LLM's past state is exchanging information with its future state, continuously representing and modulating a concealed internal state of some sort. And then that state just so happening to be some sort of self-referential meta state.
And the best inkling I have towards this is basically: https://www.youtube.com/shorts/WP5_XJY_P0Q
But then I don't think there's enough covert channel bandwidth in the agent replies for anything interesting like this.
I don't see why an LLM could not have a sense of identity or personality while it's evaluating a specific prompt, or even change self awareness while evaluating a prompt since many outputs model a back and forth conversation. My point is that without a mechanistic model of what "self awareness" means, we have no way of truly evaluating such questions, we're just hand waving vague intuitions about what it could mean.
This is kind of also the reason e.g. the HN site guidelines are worded the way they are. Regrettably, forums naturally yield themselves to tit for tat type exchanges, but there's really no reason one could not bounce such vague intuitions off of another. I do not have to be right or wrong, and you don't either. Admittedly difficult when its some intensely contentious topic.
If a mechanistic model existed, there would also be no reason to talk about this in the first place. There'd be nothing to discuss, you'd be simply told how a given model characterizes from this perspective on the model cards.
The reason I insist on mechanistic models is because the original post was making a definitive knowledge claim, and in my experience, the knowledge claim is unwarranted.
A language model (or an image model or whatever) cannot even be sentient, and I think sentience is a prerequisite for awareness.
Even if we express a lot of our subjective experience with words, the language is just a symbolic representation of those experiences. The qualia themselves, even those that are quite abstract, are rooted in our physical presence and evolution.
You can't have an understanding of what hunger or physical pain feel like if you have no need for food or a sensory capacity for feeling pain. You can't understand what loneliness or pride at an achievement mean if you don't have a neural network wired to value social connection or status. We value connection because we're social animals that have needed each other for survival.
Even the more abstract of our subjective experiences are in some way rooted in our physical evolution.
I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
AI awareness might actually be more believable if that awareness manifested itself in an entirely different way than in humans. But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience, I think we're seeing something that isn't actually there.
There is no objective evidence of qualia. All evidence of qualia are vocal or other expressions of belief in qualia. Perceptions clearly exist and are observable, subjective experience and qualia, not so much.
> I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
If your objection is to models based on "symbolic level of language" which you think lack semantic understanding of, say, trees, you should ask yourself how our brain, based on physics which also lacks any semantic category for trees, can somehow develop a semantic understanding of trees. All of these appeals to differences with the brain never seem to acknowledge that fundamentally, the brain has the same explanatory gap with physics.
> But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience
This assumes a lot. It seems very possible to me that intelligence inherently develops a map of natural categories (natural kinds), and language naturally develops around such categorical understanding. Semantics are then fundamentally the network of associations between categories, eg. there is no fundamental difference between symbols and semantics, and the latter cam be inferred from the former, and that's exactly what LLMs do, and why the semantic maps between different languages are so similar and how they can translate between languages.
You didn't address any of what I wrote, let alone provide any counterarguments. Which part of what I wrote do you think was wrong?
For example, the Turing machines and the lambda calculus don't look anything alike, but they are fundamentally interconvertible, and so in a real sense they are fundamentally equivalent. Without a model, all of your arguments are completely unconvincing for exactly the same reasons, eg. that there may exist many paths to fundamentally equivalent ends.
Hunger as a concept doesn't mean anything without the physical need. Politeness or bluntness, even in writing, don't mean anything without social dynamics. And we have social dynamics (and neural structures that directly process social cues and associated feelings) because we've evolved into social animals for whose survival that was important.
I see no reason to believe that a model trained only with symbolic representations, with no connection to the physical world phenomena that those symbols represent, could contain the subjective experience itself.
Neural network models may be able to derive novel (or at least novel-looking) output rather than just an obvious rehash of their input, but I don't think any set of bytes can fundamentally contain information that was never entered into it. (Even if e.g. a model produces previously unknown mathematical results, those results can in principle be derived from the information that they were trained with.)
I'm not saying that artificial neural networks couldn't, in principle, be aware. ANNs and biological neural nets may be equivalent in the sense that any information and processing structures represented by a biological one could in principle be represented by an artificial one. If that's the case, and awareness is purely a product of our neural systems as materialism would imply, it should be possible for an ANN to be aware, too.
But when the model has been trained with only language, and IMO the subjective experience can't be derived from the symbolic representation alone, I can't see how the model could include the actual subjective human experience.
An AI model could of course have an awareness and subjective experiences that are totally different than our human experience. But then the fact that it happens to produce output resembling what humans find meaningful shouldn't be considered indicative of such awareness.
This is of course more of a philosophical argument than a technical one, and I'm happy to hear counterarguments, but not on the level of off-hand dismissal.
People who say LLMs have subjective experience aren't saying they have human-type subjective experience. Nobody who sees an LLM express hunger when role playing as a hungry person thinks that the LLM is actually hungry.
I too can role play as a hungry person despite not being hungry, so there is no reason in either case to conclude that the words produced reflect genuine internal subjective states. The point is that such internal states may still exist.