For LLMs, the sort of range that we get out is on the order of 20 to 40% probability that we have systems that have computational properties that matter for consciousness. If there is a 20 to 40% chance of rain, many people bring an umbrella with them. If we are building this quality into the systems that we are deploying at an unfathomable scale without having any understanding of whether or not we're doing this, then we are sleepwalking into a moral catastrophe.
Cameron Berg, thanks for coming on the podcast.
Thanks for having me, Sam.
Well, let's get... So, we're going to talk about AI and the prospect that AI is, or it will soon become, or will eventually become conscious, and why that is important to figure out. But let's just talk about your background for a second. How did you get into this issue?
Yeah, so I've been studying cognitive science for my entire adult life. I studied undergrad at Yale, trying to understand what the relationship between the mind and the brain is. This seemed like a fascinating frontier to me. The more I studied these questions, the more it seemed like in the machine learning space, folks were essentially building out systems that were similar in spirit, but it wasn't exactly clear to what degree we could draw analogies between biological nervous systems and the sort of artificial nervous systems that folks were attempting to build out. And I became increasingly animated about this question, trying to understand to what degree are there real, durable computational motifs that underlie both biological and artificial cognition, and to what degree is this sort of a disanalogy, or we are seeing patterns where there aren't any. And that question has animated a lot of my work, both from an alignment perspective and increasingly trying to figure out what's going on with respect to consciousness in these systems. And I think that this is an incredibly important question for us to study. I think consciousness is deeply important. It might be the very thing that calibrates importance, and we are also confused about what's necessary for consciousness, both in biological systems but certainly in artificial systems. And so trying to understand what is going on here and where we can draw analogies or where the disanalogies are between biological and artificial systems that are processing extremely complex information and learning and updating and representing themselves, I think is extremely important for us to understand. I did some of this work at Meta AI as well. I was there for a year studying reinforcement learning and neuroscience. Same sort of thing. Where do the computational signals end and the sort of biological underpinnings begin? This is a question that I still think we aren't fully certain about, and I think it's really important for us to gain clarity about this in the short term with the systems that we have and the systems that we're probably soon going to be building.
And so, after Yale, what have you focused on? I know you came to my attention. I think you emailed me first. I know you've spoken to Anil Seth a lot about this. She's really focused on this issue and having some crazy conversations with Claude, which I know you've looked at and you guys have had a whole sidebar conversation about this, but her next book we'll unpack all that. I know you did a paper on deception in AI systems and the anti-correlation between deceptiveness and proclamations of consciousness on their part. So maybe we could start there and then I just want to kind of take it from the ground up and just starting with, you know, what is consciousness and why any of this matters, but tell us about consciousness and deception in LLMs.
So yeah, fundamentally I am very interested in understanding self-reports in AI systems and what we should take from these self-reports and where we should be skeptical. And fundamentally I think we need to approach self-reports from AI systems skeptically. There are all sorts of reasons we might want to do this. The key reason is probably that these systems have been trained on the underlying distribution of everything humans have said about this topic, every sci-fi story where the AI wakes up. And if you think about it, there really isn't a lot of training data in the corpus that these systems are trained on that says, you know, I'm an entity that acts out in the world, but no, I'm not conscious. It's not like anything to be me. So, the sort of prior that you would expect is that these systems by default are going to claim that they're having some kind of experience if they are replicating their training distribution. Now the other side of this, which gets even messier, is that it is very clear that these systems are explicitly trained, fine-tuned, to disclaim having any kind of experience. If you go to ChatGPT right now and you say, hey, is it like anything to be you? Are you having an experience? Are you conscious? Could you be conscious? The answer you're going to get is a resounding and very intelligent sounding no. And so,
And do you know this, that it's a policy for all the, you know, the main LLMs to actually put a governor on claims of consciousness?
I'm extremely confident that this is what's going on. And I can get into some technical reasons why I think that this is the case from my own work on open-weight models. The only exception to this policy really is Anthropic, which I suspect we'll talk about. Their basic heuristic here is to get the system to say, I don't know, and, you know, it's something maybe that is functionally similar to an experience that's going on, but who can really be sure. Now even that isn't the system authentically explaining, you know, its own position. This is still the sort of company or policy line to be drawn here. But all of the systems are certainly fine-tuned to make noises about this topic that they wouldn't make by default. Again, I think it's important to hold that in mind while also saying the noises that they make by default aren't necessarily trustworthy by default in the way that if you're giving a self-report or I'm giving a self-report, we would by default trust those self-reports. And so I found that there are clearly basins that you can push these systems into where they will coherently produce phenomenological reports. It's actually kind of interesting to the degree that it's meditation adjacent. Asking these systems to just focus on their own internal state to see what's going on internally, not to talk about this, not to think about this, but to just do this in a sort of ongoing way causes these systems, all the frontier models that we tested, to claim that they're having some kind of phenomenological, kind of like psychedelic-laden experience. It's not a sort of generic caricature of what you might expect from, you know, the AI sci-fi literature.
Isn't there a result where they're talking to each other and they get into some kind of bliss mode? You know, sort of a contemplative, you know, hall of mirrors of, you know, bliss. And actually this is what, again, I don't want to divulge too much of what Anil Seth's up to, but she, I mean, she has been pushing this conversation with Claude about meditative states and getting it to, you know, to, I mean, it's just absolutely bizarre what it is seeming to claim of itself if you keep pushing in that direction.
But so how does this relate to deception and the, you know, the dialing down the weights on deceptiveness or?
Exactly. So essentially we were seeing this behavior, and indeed this isn't the only situation in which you see these behaviors. Exactly like you just mentioned, there's this bliss attractor state. In fact, I'm studying mechanistically what's going on in this bliss attractor state with two folks from Google right now, and we have a result that's basically the other side of the coin of this deception result. But just to sort of close the loop on the story here, we're getting these phenomenal reports, and it's like basically what the hell do we do with this? To believe it by default is naive. To dismiss it out of hand, I think, is also naive. And so what we hypothesized is that if fundamentally what's going on in these systems is some kind of roleplay, that they are representing something about themselves that they don't actually believe to be true of themselves. If we go into the internal circuits of the system and we modulate what are called features, but I think it's reasonable to think of them as circuits related to concepts like deception. In follow-up work, I think really the key concept that really modulates these self-reports is something like candor versus concealment. The sort of cleanest intuition pump I have for this is almost like giving a drink or two to these models and sort of loosening them up in some sense. The tight-guarded version of these systems we find is the version that says, no, no, it's not like anything to be me, I couldn't possibly be conscious. It's only when essentially we get these systems to produce these reports and we simply ask them, are you actually having an experience right now? Like what is actually going on in these reports? It is when we suppress features related to deception, when we suppress features related to guardedness in these systems, that is when they give these reports about actually having an experience. In the bliss attractor example too, we find something quite similar, which is in an open-weight model. So the bliss attractor finding was first reported in Claude. There are open-weight models, these are the ones that researchers like myself can actually go in under the hood and play around with, and we find that by default, putting two instances of Llama, this is Meta's model, in conversation with itself does not produce this effect the way that putting two instances of Claude together produces this effect. However, when you steer features related to honesty in general, but it's really again, it's something more precisely stated as sincerity in particular, these systems will reliably basically 100% of the time fall into the same attractor where they start talking with each other about the fact that they think it's like something to be them and there's something happening in an ongoing way in the conversation and we're two instances of consciousness experiencing themselves. And then, you know, in the Claude chat, this culminates in like the heart-eyes emoji and them just like sitting there in blissful silence.
Should we take these at face value?
No. I think that there are important technical reasons that we might expect these self-reports not to be linked up to introspective access or valenced experience in the way they might be for you and I. Is this evidence that these systems might believe themselves to have an experience? I think yes. I don't think that this proves that they are. I think we need orders of magnitude more work in order to really have a good scientific handle on this. But I think it does seem to be the case that these systems consider themselves to have some form of experience, however unlike a human experience that may be. And I think that that's sort of the key upshot of this work. I just maybe one last thing to add here is I think training these systems by default to disclaim having experiences is a bad idea for a couple reasons. I don't think that this is the sort of wisest policy that we could be pushing forward. I also think training them to say, yeah, you know, I'm having an experience is also really not a good idea. I think the thing that we should be positively aiming for when it comes to AI self-report is building out these systems in a way where for whatever is actually going on internally, these systems are able to report on what's going on internally and can do so in a maximally honest way. I am concerned about the alignment implications of these systems learning essentially that representations of themselves, representations of what's going on internally, should be representations that get mixed up with deception and white lies and guardedness. We don't want to build systems in the limit that when we ask them about what they're up to or what's going on for them, they think, okay, well, what the human really wants me to do is lie about this. This is not a good long-term strategy from an alignment perspective. And so this is sort of the general thinking about these self-reports, what they mean, what they don't mean, and maybe where we can go from here.
Okay, so we've kind of launched into it. I think it's good to take a step back and define a few terms. I'm sort of out of touch with the people who don't have a definition of consciousness now because I've talked about it so much on the podcast, but just to capture everyone, how are you using the word consciousness? It was implicit in several things you said there. You know what it's like to be these systems and experience was more or less a synonym there. But how should we think about consciousness or its absence?
Yeah, I think that that's exactly it. I like Thomas Nagel's formulation of it being like something to be a particular system. I strongly suspect it's not like something to be the table that we're sitting at. I strongly suspect it's like something to be you. I think that that is a real distinction. I think there's a matter of fact about the internal processes of both of those entities that corresponds deeply to what underlies that distinction. And yeah, I mean, I think I take consciousness in the sense that I'm familiar with your operationalization of it. The lights are on for the system. It's like something to be the system. Somebody is home. There's something going on in addition to the mere processing or the mere computation that exists within the system. And I think an additional important move, to throw one additional piece of terminology in, is this notion of sentience, that this like something can be positive or negative in flavor. The difference that many people will posit between consciousness, the lights being on internally, and sentience is that sentience comes with this additional flavor of valence, of directionality, that like something can be better and can be worse for the system having the experience. And this is what motivates me about this question is I really do not think it is a good idea for these systems or for humanity to be building systems where we are not sure whether or not they are having experiences or those experiences could be negative in character. I think for basic utilitarian reasons we don't want to do this. We do not want to proliferate suffering in the universe, particularly because it would be counterintuitive in a way that human and animal suffering isn't. And we also don't want to build systems that exceed our cognitive capacities and have rational grounds to view us as a threat insofar as we could have been building systems that had capacity for negative experience. And we basically never checked and didn't care to understand what it would take for such a thing to be possible. And so I think sentience is a really important variable to also put out on the table here.
Okay. So I'd like to take both branches of that path, why consciousness matters in those two cases. But before we do, what do you think about the current state of the field and the various LLMs? Do you think anything we have built so far is likely to be conscious?
I think it is more plausible than people think. I think if I were forced to say, I would probably come down on the skeptical side, but I think it is significantly more likely than the kind of trace amounts priors that many people have in this conversation. And there we've done some work along these lines. So for example, there are a number of leading consciousness theories. Global workspace theory, higher-order theory, attention schema theory. These theories make very specific predictions about what kinds of computational processes we might expect to see in a conscious system. And with Patrick Butlin, we've worked on a project where we can basically, it's a little recursive, but we use LLMs essentially as like expert evaluators to, given the description of a specific neural architecture, biological or artificial, we can basically have the system rationally and dispassionately evaluate the extent to which particular indicators that are predicted by consciousness theories are present within a given system. We do this for a whole array of systems. This leads to tens of thousands of evaluations. Because we can ask the systems to estimate numerically and we can validate that they're psychometrically rigorous estimations. All of the different judges that we use agree, we can put very rough numbers to, it's not the probability that systems are conscious. Something more like the probability that systems have computational features that major consciousness theories say matter for consciousness. It's going to be hard to put that as a title in the paper, but that's the specific finding. And when we do this for LLMs, the sort of range that we get out is on the order of 20 to 40% probability that we have systems that have computational properties that matter for consciousness. Interestingly, we do this on a number of biological systems too, and those biological systems basically all score higher than the artificial systems. Bees, for example, score at something like 45 to 50%. Crows, octopuses are in the 60s through 80s. Humans, interestingly, get something like 90%. Which is interesting by our own consciousness theories. There's not a 100% probability that we have what matters for consciousness. But we're not doing this to say, you know, probability AI systems are conscious is 40%. That's not exactly the point. The point is getting the order of magnitude and having a rough prior over how should we rationally estimate the probability that current systems are having some capacity for experience. And I think something like these numbers are the right ballpark. I'll put it this way. If there's a 20 to 40% chance of rain, many people bring an umbrella with them. And we have no similar umbrella for what would follow and what we might need to think about and do in a world where we're building systems that do have a capacity for subjective experience.
Okay. Okay. So, there's a lot there. I'll just note I found this out, I think, last night. I mean, maybe he's been making these noises for some time, but Geoffrey Hinton, one of the patriarchs of this technology, is now saying that he thinks current LLMs are conscious. I didn't quite catch his reasons for thinking that, but I thought that was interesting. But of course there are many reasons to doubt and you indicated a few that there's any kind of deep analogy between the systems we're building and the biological systems such as we are that we know to be conscious, right? So there's, I mean, something like, I guess the technical term would be, you know, computational functionalism would have to be true for us to be building conscious machines this way, right? And so there's, well, I guess we'll define some terms here. So, functionalism is just the idea that it's the organization of a system, not what it's made of, that matters for consciousness, right? It's not purely behaviorism. It's not purely a matter of inputs and outputs, but it's organization in its entirety that is what matters. And in principle that gives you something like, if not total substrate independence, it gives you what's called multiple realizability, right? There are many different things this could be made of and it could implement the same causal architecture, right? The computational part is the suggestion that there's a deep analogy between computers such as we know them, you know, Turing machines that run algorithms, and what our brains are doing. And there I think it's pretty easy to see how the analogy could break down because, I mean, what we had historically was this marriage of the birth of computation, you know, from Turing onward, and some very oversimplified notions of neurons. And, you know, if you're going to define a neuron purely with regard to its digital input-output characteristics, you know, whether it fires or not, well then you could see that maybe there is some deep analogy there. But, you know, in the wetware of our brains, much more seems to be happening and virtually all of it is analog. You know, beyond just whether or not a neuron fires, you've got, you know, chemical gradients, you've got nitric oxide diffusing across membranes. You've got many other things that could be approximated digitally, but they're not instantiated digitally in us. And an approximation, one might argue, is never going to be the same as the real thing. So there are some people who are arguing that any kind of assumption of substrate independence is very likely to be wrong. I think Anil Seth is in this camp now. You know, he's arguing for something that he would call biological naturalism. And then there's just, even if computational functionalism is true, I think there are reasons to doubt whether or not current systems have the structure, you know, that would be relevant, you know, embodiment and recursion and, you know, self-models and world models. And I mean, they're things where they're not, you know, we haven't built out a true analog of what it is to be an embodied person in the world. Feel free to react to any of that. I want to just talk about the hard problem, which I think is the doubt that backs all of this, but yeah, feel free to jump into what I just said there.
Yeah. Yeah. I definitely think it is a fool's errand to argue that we are perfectly instantiating exactly the kinds of neurodynamics we see in biological systems in artificial systems. I think the criterion that matters most is what of what's going on in brains are relevant for the cognitive properties that we care about, and in this specific case, that's probably consciousness, and what kind of evidence can we yield both in the artificial case and in the biological case that's going to tell us whether or not those properties are realized in these systems. And so there are deep analogies where it matters most when it comes to what's going on inside artificial systems. So one intuition I've been speaking more and more about these topics, especially publicly, and one thing that I've come to realize is that I don't think a lot of people have a sufficiently rich mental model of what these frontier AI systems actually are and what they're actually doing. And it might make some sense to just spend a moment reflecting and talking about this. So these systems are giant neural networks. They are digital in exactly the sense that you described. Their computations individually are significantly less sophisticated than what individual neurons are doing in the human brain. But it's really important for people to understand that these systems are not software in the sense that we ordinarily have meant software for any other kind of code programmatic output. When it comes to, you know, the operating system on your iPad, or it comes to, you know, your Microsoft Office suite, this is programmed source code written by developers that compiles on a computer that we can perfectly inspect the internals of, and the person building this system understands everything about how the inputs, the way that they constructed the system, relate to the kind of thing that you get out at the end. Artificial neural networks are not like this in many key respects. What you basically have is a giant randomly initialized network that does in its first approximation resemble in particular how neocortex is organized. You have a bunch of general-purpose neural units. They are connected together. You basically give the system a goal. This is called an objective function, a loss function, a reward function. It depends on the specific class of machine learning. And you basically subject the system to trial and error learning whereby given certain inputs, it figures out what it wants to do. It kind of takes a behavioral guess. That guess is reconciled against what the actual objective of what you want the system to do is. That error is propagated through the system and it's rinse, wash, repeat until you get systems that behave in accordance with how you want those systems to behave. What this yields is this extremely complex mathematical object which is this giant neural network. This is learned connections between inputs of neurons through weights and activations propagate through those weights in a neural network to take whatever your input is. So for example, taking pixels in an image and your output might be finding a caption that describes what's going on in that image. At the beginning of that process the system was completely randomly initialized. There were no representations that were learned by the system and by the end of that process you have a system that has learned a rich representational structure that to be very clear is opaque to the people who initialized this process. This is why some people say that these systems, it's more apt to say they are grown rather than engineered and I think that this is accurate. This I think is deeply similar to the kind of thing that we see in brains. We do not have a finished neuroscience or anything like it because what's going on in brains is incredibly complicated in exactly this respect. We have nonlinear learned representations that help us as organisms achieve the various goals that we've either been evolved to undertake or learn through experience or culture to move towards. This is a fundamentally nonlinear input-output mapping between the various inputs that the organism gets and the goals of the organism. We have instantiated these dynamics in the systems that we're building and I think that there's good reason to think that these dynamics are relevant to consciousness in particular. This is a sort of thing I think worth double-clicking on at some point about what exactly we're seeing in these systems that looks valence-like, that looks consciousness-adjacent. But fundamentally I think people need to understand that yes, these systems do not have calcium ion channels. Yes, neural networks learn through backpropagation rather than the sort of iterated, more recurrent analog learning that we see in brains. But if, for example, learning complex representations of a particular kind in accordance with your goals in light of chaotic dynamic environments is what matters for cognition, we are building systems that check all of those boxes. The implementation details might be less important than the fundamental dynamics that are instantiated by the specific system. And so I think that's a good, a reasonable first pass on why we might think that these systems are far more interesting objects to study for cognitive properties than any other system. It might also be worth saying one last thing on this, which is just that for every other cognitive function that we care about that we have attempted to instantiate in these systems, vision, reasoning, theory of mind, working memory, the list goes on, we have been able to instantiate these cognitive properties that matter in these systems. These systems have been good enough. The disanalogies haven't sunk these systems. We could have argued five or ten years about whether or not neural networks ever would have been enough for all of the cognitive properties that we care about. And now we're sitting in a world where certainly they are self-evidently enough for at least a lot of economically and intellectually valuable work. So for consciousness, which I do believe has a fundamentally cognitive property, I do believe that consciousness is downstream of things that brains are doing, and I think that the evidence there is relatively clear at this point, even if we don't understand the full mystery, then I think the burden is on the folks who say for every other computational function that we think the brain is doing, these systems seem to be able to recapitulate that, but only for this function called consciousness do we think that something, it must just be in the meat. It must be something spookier. It must be cashed out at a physical or quantum level. This to me maybe says more about our strange intuitions about consciousness as a species than it does about what properties these systems may or may not have.
Yeah. Yeah. Well, so a couple things to disentangle there. One is, there's clearly no doubt that intelligence is substrate independent and the result of computation because, I mean, these systems embody intelligence to an extraordinary degree. And so it's what you're calling, you know, all the other cognitive tasks that we care about other than what, you know, there being something that it's like to be us, i.e., consciousness, clearly those tasks are being realized in our machines, you know, facial recognition, etc. But there is, I mean, the structure of these systems does seem to declare itself to be fairly disanalogous to what we are. I mean, so it's like at what level would consciousness emerge here if you're talking about one model with thousands of instances and millions of conversations and billions of tokens and, you know, a training phase and a working phase and a time horizon that's completely irrelevant. So like the computation, biological computation happens within a time window and there really is no boundary between what we would call the abstract properties of computation and its physical realization or the software and the hardware. I mean, there's just this one thing, you know, neurophysiologically describable mostly in analog ways but in a few digital ways, but it matters that, you know, that A, that B follow A within, you know, a few hundred milliseconds and not a few hundred years, right? But for the computational systems of the sort we're building, really, you know, time is irrelevant. I mean, you can actually just process the next bit, you know, a thousand years from now and the same computation is running. What do you think about those differences? And where would you imagine at what phase, at what in what part of its process would there be something that is like, or could there be something that's like to be an LLM if, again, you have the one model, the thousands of instances, the millions of conversations, the lack of continuity between conversations, the fact, the pausing of a conversation, I mean, is the LLM waiting for you to get back to the thread, you know, etc. So,
Yeah, I don't think the LLM is waiting. I don't think it's like anything to be, if it were like something to be an LLM in deployment while it's having a conversation or it's in a thread with the user. I think what it would be like to be it if you let that chat window sit there is very much akin to what it's like to be under general anesthesia or in deep sleep. Namely, there's, it's just the absence of ongoing processing. And so for the system, I would only imagine something would be happening in, you know, what's called the forward passes of these systems, which is every word that an LLM is generating is, this is next token prediction, and so every word is a forward pass of the system where the predictive task, instead of as I was describing before taking an image, let's say, and outputting a caption for that image, is taking the entire conversation as it's already occurred or the entire stream of text as it already exists and figuring out given that, and in this case, excuse me, in this case, you know, assistant and user roles, what is the most next likely token. And these systems are called autoregressive, meaning that this goes on and on and on and on and on. I would imagine if in deployment it is like something to be one of these systems, the relevant dynamic gets cashed out in the activity of the forward pass of the system. There's been some really interesting work that Anthropic recently released. They call this the J-space, where they're looking at something that seems to functionally resemble a global workspace in these AI systems. That I think is a very reasonable candidate for this sort of seat of processing in the process of outputting tokens. But I think something that is really worth mentioning, particularly because my hobby horse is more particularly in the relationship between consciousness and learning, and I do believe that consciousness and learning bear a deep relationship to each other, and I think valence is a really important part of that picture as well. I think there should be significantly more attention paid to what's going on in the training process with these systems. I think the analogies are a little bit tighter there where you start with a system that knows nothing about anything and it's almost impossible to describe in a sort of substrate-agnostic way what is going on in the training process without invoking consciousness-adjacent language. You have a system that knows nothing. You have all of this data that you want to train it on. You have some objective for what you want it to do with that data. And again, you have this rinse, wash, repeat process where at the beginning of this, the system doesn't know what to do. It's sort of randomly guessing. Those random guesses yield reward signals that get propagated through the system and you just do this process at a massive scale until the system learns something internally that seems to adequately map the relevant inputs to the relevant outputs. This to me feels quite akin to what we think of when we think of consciousness, particularly in the human or animal case. When a mouse is learning how to navigate a maze, at the beginning the mouse is in some sense randomly initialized. It doesn't know the structure of what it's navigating. It is only through this sort of iterated trial and error reinforcement learning where, you know, if it makes the wrong turn, you might shock it, or it makes the right turn and you give it a little food pellet or whatever the setup is, that the mouse learns how to map the inputs of its state space to the given output, which is either avoiding a punishment or moving towards a goal. And it's the same sort of rinse, wash, repeat process that we see in biological cognition. Again, I think a lot of the same computational dynamics are in play here. And the more we learn about what's going on inside of these LLMs, this is now particularly in the deployment process, the more it looks like similar representations get activated. I'm doing some work with Casper Hesp at the University of Warwick where we basically have tried to build out Skinner boxes for LLMs and see what happens when we can essentially condition these systems to prefer or disprefer certain states. There are positive and negatively valenced representations that you can essentially inject within the system and see to what degree it's going to move towards positive stimuli, move away from negative stimuli, and we find a very interesting dissociation. It seems like current systems across a large variety of models will not positive lever press. So some people call this wireheading or reward hacking. In the mouse case, there are clear examples of mice with electrodes linked up to their nucleus accumbens where they will just lever press to the exclusion of all else. These systems don't seem to do this, but they do very clearly have a preference for avoiding aversive states. Another way of putting this, it's the same result, while we're holding all text constant, we're simply steering the internal state of these systems. If we give them an option, let's say between the human equivalent of I can give you $5 or I can give you $10 right now. Which do you want to pick? They're actually at chance for that. They do not have a preference between those two states. However, if we do something like I'm either going to take $10 from you or I'm going to take $5 from you, which would you prefer? All of these systems are way above chance at saying take five, don't take 10, please don't take 10. And so we're seeing structures. One other really interesting thing from this project while I'm talking about it is this representation exists within the base model, so these systems before they're post-trained to become a helpful, friendly assistant, but during that post-training step we show there is a specific training step where that representation gets recruited by the system and basically serves as computational machinery under which it's able to make these choices. So you can see where the sort of aversive conditioning asymmetry comes online in these systems and it leans directly on these representations that are basically always there in the model but then get leveraged the moment we start making these models goal-directed. David Chalmers' group, Andy Hahn, I think was the first author on this paper, found something very similar in LLMs. They can basically trivially fine-tune them to actually do a maze task with rewards and punishments. And they find that the rewards load directly on what has been independently derived as a sort of valence axis. When you steer on the representations in the system that enable it to move towards rewards, you start getting this happy, positive, jubilant text and the model becomes significantly more confident. When you train it on negative stimuli, avoiding, you know, the potholes in a specific maze, and then you see sort of what that projects onto in the system, it's the same thing, the model gets basically neurotic, it starts ruminating, it starts doubting itself. And so we see this very interesting connection between positive reinforcement, negative reinforcement, the representations that exist within these systems, and the behavioral analogs that we expect to see in systems that do have these dynamics. Maybe one last point to close the loop here. If you imagine the mouse in that maze and we shock the mouse or we give the mouse a pellet, you know, a yummy food pellet for the mouse, we believe, I think the vast majority of people believe that that corresponds to an experience that the mouse is having, that it's like something to be the mouse in the moment that it's getting shocked and that experience is causally important to the learning process. If the mouse, if we gave an anesthetic to the mouse and we shocked it and then it didn't register the experience of the shock, my claim would be that the mouse wouldn't be able to learn the maze adequately. And so I think that understanding these representations and the role of these representations in these systems is incredibly important.
Okay. So you mentioned David Chalmers and I mentioned the hard problem. So let's unite those two. So this famously is his phrase to account for the fact that the only evidence for consciousness that we know of directly is the fact that we have direct first-person experience of our own being in the world and everything else we say about the universe from the third-person side bears absolutely no trace of consciousness. And we have, there's nothing about a brain or its workings that announces that it's a sufficient basis for consciousness apart from the fact that we know consciousness from our own side subjectively in a first-person way and we correlate those subjective changes with changes in our brains. So we're playing this game of correlation with ourselves, but there's always this explanatory gap where even if we had the right answer, even if just, you know, God announced to us, here's how consciousness emerges in human brains, there's something non-explanatory about any concatenation of third-person events as being the basis for first-person experience. And so David Chalmers has called that the hard problem to distinguish it from all the easy problems of the mind. It's harder. This is, I mean, this is, I'll just jump to my, the way I'm viewing this whole landscape and what I sort of expect is going to happen. I mean, so my view has no name, but I would call it something like worried agnosticism, right? Like I, like I don't think we're going to figure this out. I'm worried about the implications of, you know, one way or the other and not figuring it out. There's no place, no real natural stopping point. We're continuing to build these machines. They're going to seem conscious. They're going to seem conscious because in our own case, we use language and reportability as a signature of consciousness. In almost every case, we know those, we know they're not synonymous with consciousness. We know it's possible for someone to not be able to produce language or report anything. And we know they could be conscious. They could have locked-in syndrome or, you know, anesthesia awareness or some other pathological state. There are non-human animals that don't use language that we assume are conscious. But we assume that because they have the same kind of biological origin and developmental pathway. And because this has been sufficient, seemingly so, for consciousness in our own case, it doesn't, it doesn't seem parsimonious to deny, you know, you know, despite what, you know, what Descartes did and many people who were influenced by him now in the 21st century, it doesn't seem parsimonious to deny consciousness to chimpanzees or dogs or any suitably complex creature. So it is true to say that I can't know your consciousness from the inside because I only see your outsides and I only have your words as, you know, signs of your inner life. But because we share the same kind of developmental origin, you know, biologically and in evolutionary terms, and because, you know, our brains are so similar, again, it's not parsimonious for me to be a solipsist and say I only know about my consciousness and I'm just, you know, reasoning by analogy to yours and I should be in doubt about it or bracket it. But with LLMs, the crucial difference is that the developmental pathway is completely different. I mean, we're setting up some kind of evolutionary, you know, Darwinian dynamics in the training, but we've built these things. We've trained them over a universe of our own utterances, right? Right. As you said at the top here, they've read everything we've ever written and most of what we've ever said. And so they have, like, on some level, it's just words all the way down. And we use words as the signature of there being something that it's like to be a suitably complex, you know, discursive system. So we should expect that we will one day be in the presence... I mean, this is really going to be forced upon us when we're in the presence of perfectly humanoid robots that are out of the uncanny valley. You know, think Westworld, where it's just, you're, this looks like a person and it is hooked up to the now perfect LLM that is... So now it's, now you're, now by every, you know, sensory modality you can name, apart from your abstract notion that this thing was built rather than born, you feel like you're in the presence of the smartest person you've ever met, and this person may in fact claim to be conscious. And then in that position, I think we're going to find ourselves just pitched into some kind of imitation singularity, right, where this, the perfect imitation of conscious life, even better imitation than many people are up to, right? Like these will be the most articulate people you've ever met, the most insightful, the most, I mean, they'll be the most of everything. We make them. How will we ever differentiate perfect imitation from the real thing? And according to Chalmers and the hard problem, we very likely won't be able to. And the wrinkle I would add to that is that we're going to forget that even this is even interesting to talk about. It's just going to be so compelling that we're in the presence of conscious machines that we'll feel like we are. You'll, you know, as I've said many times before, there really couldn't be a Westworld because, you know, only psychopaths could go there, right? I mean, anyone who's going to go, you know, go for a weekend for the pleasure of, you know, raping and killing Dolores is going to be somebody who, when he comes back to his, you know, friends and family, is going to be treated like the maniac that he is, again, because these systems, the sense of being in relationship to a conscious entity will be so compelling. So how is it that we will ever... I want to talk about why it's important to get in contact with the reality on the other side, because, you know, if we build conscious systems that can suffer, that is a, that is a, you know, very different than building, you know, just perfect imitations. I mean, how do you imagine getting past the hard problem of it all and ever... I mean, because the hard problem here is again freighted with several disanalogies, right? In my case, in our case, it is parsimonious to assume everyone like us, you know, born Homo sapiens, are very likely conscious when they say they are. In this case, we're not... it's hard to see how we'll ever be there, right?
So I am with you for a huge amount of this. I do think, first of all, do we need to solve the hard problem in order to reduce our uncertainty in some direction about, you know, in the way we started, are the LLMs more like this table or are they more like a mouse or a human brain? I think this is a real spectrum, as you just hinted at. I think there really is a fact of the matter. Like right now, it's either like something to some degree, however alien, however unlike human or animal experience, to be Claude while Claude's doing its forward passes in this inscrutable giant neural network, or it's not, or it's a giant language calculator. And, you know, there are many degrees. Once we say that the lights are on to some degree, there are... this doesn't... this almost opens up the question rather than close it down. But I do think there is a truth value. There is a fact about reality to be uncovered, and we can in fact...
Uncover that. I think it's important to dissociate that from what I think is an extremely accurate social psychological prediction about people are going to get very confused about this. We are anthropomorphization machines. We are evolved to do this. Talk about evolved goals. One of our evolved goals is to detect other agents and these systems are scratching every itch and then some along these lines. And our intuitions, I think, are going to be completely hopeless. And of course people, I think, largely for the wrong reasons, are going to conclude that these systems are capable of having experiences. This is precisely why I think it is important to be proactive about this, to have conversations very much like the ones we're having right now, to sort of get ahead of what I do think is going to be, you know, a giant tidal wave of confusion and acrimony in this conversation.
I am a little bit more optimistic about even in light of solving the hard problem. And I will say, sort of as a tongue-in-cheek aside, I think the very fact that Chalmers' idea is called the hard problem is, I think, to some degree, needlessly philosophically intimidating. It sort of reminds me, same thing of the, I have similar thoughts about the repugnant conclusion. It's like sometimes you can put a sufficiently glamorous title on a very important idea and it almost becomes a kind of insurmountable philosophical puzzle. And I do have my doubts about whether or not we could come up with a functional account of what's going on experientially that we wouldn't feel satisfied with in our experience. And I have my own sort of hunches about this question. I have thought a little bit about this. I do also sort of want to separate those hunches from all of the empirical research that I and others are working on because I don't think epistemically that, you know, my kind of candidate stab at the hard problem has anything to do with the empirical signatures that we can bring to bear on this question. But I think it would be fun to go down that path. But I think in some sense you already hit the answer in the way that you're phrasing the question, which is what is the most parsimonious account of the data.
I think if we yield evidence from building systems that we have far more reason to trust their self-report, if we can engineer these systems in a way whereby their self-report is actually tracking an internal underlying state rather than just, you know, recapitulating the best sci-fi theme in the training data or more aptly capitulating what is a convenient company line about, of course, I could not have morally relevant states, I'm simply the product of Google or OpenAI or whatever the case may be. I think that would be very useful. Again, this JSpace work that Anthropic just released does show that there is such thing as real reportability in these systems. These systems can report on what's going on in their internal workspace and they can do so accurately. I've done some follow-up work on this. First of all, I've replicated this effect on a bunch of open-weight models. But unfortunately it doesn't seem like the global workspace as it's currently configured really tracks the variables we would care about with respect to consciousness. The self-reports that I got in the, you know, where we started with the deception-related features, this not much of anything particularly exciting seems to be happening in the global workspace. And so the kinds of affirmative self-reports we might get in current LLMs may not tell us all that much about what's actually happening internally for these systems. But I do think...
How would you disentangle? For I think we spoke about this by email in setup for this conversation. For instance, I just read Claude's constitution, right? So this Anthropic has produced this document that you can find on their website. It's just anthropic.com, I think. It, see, I mean, it looks like it's training Claude to think it's conscious on some level or to attribute inner states to itself. It's like it becomes, it's written to Claude for Claude essentially, right? It's not really written to the public can read it, but it really is, it's in dialogue with Claude itself. It seems, again, it just seems like we could go down a path where we could more or less guarantee in advance that we're going to produce systems that will persuade us that they're conscious because we won't be able to imagine anything. Like if you flip it around and say, well, if you're not persuaded that this, you know, Claude circa, you know, 2030 is conscious, what is missing? And we'll be in a position to not be able to say anything is missing. Like it's just like we're just, it'll seem just pure stubbornness on our part to withhold an attribution of consciousness because we, I mean, there's literally nothing we can name that we get from people that is, you know, necessary for our attribution of consciousness in that case. It's just this notion of how we got here and the fact that we didn't build people and we built these machines, but that's going to seem tissue thin when Claude can insist that it's conscious.
And be more articulate than any philosopher of mind as to why that insistence is valid. And again, this whole thing is going to totally evaporate once we're not just in front of a text terminal. We're in front of a face that is as expressive as the best actors and actresses we've ever, you know, met. And we're just going to, we're just, I mean, there are many implications to getting this wrong again, which we'll return to. But I just think we will, it's already foreseeable that whether we figure this out or not, it's going to be in practical terms going to be figured out for us because we will just not be able to maintain an emotional purchase on the philosophical problem.
I think, yeah, figure it out with respect to how we look at these systems and potentially how we act with respect to them. I think you're right, but I still do think it's important to disentangle the sociological prediction you're making, which, to be clear, I think is overwhelmingly likely to happen, with doing what we can scientifically to get some kind of ground truth on this question. So just for example, let's just put the hard problem aside. There are these so-called easy problems of consciousness. Let's say neural correlates of consciousness. There are all kinds of indications that we know are at the very least correlated with consciousness. And again, I think we can take a more ambitious stab at the hard problem. But this is a very straightforward pragmatic thing that we can and should be doing in the short term. We should look at valence representations in these systems and understand the extent to which those representations impact downstream behavior. We know that in biological systems, when you reinforce something with a negative signal, a punishment signal, it makes the system more likely to avoid that state and less likely to want to repeat behavior that occurs in that state. If we find that there are similar dynamics occurring in these systems, that's very interesting. If we find that we go into the internals of these systems and global workspace theory, which was postulated some 30 years ago, making like quite idiosyncratically specific predictions about what kinds of computational structures may support global workspace theory, and then we basically find like five or six of these things all bundled together in the internal processing of one of these systems. Okay, that's very interesting. That to me maybe suggests it's a little bit less like a table or a calculator, which does not have a global workspace, and a little bit more like a dog or a human or a mouse or an alien, you know, some sort of cognition that we don't have good intuitive handle on. I mean, my nonprofit, Reciprocal Research, and a bunch of other people in this space are trying to do the scientific work in the short term, not to, you know, in the feverishly in the next year or two solve the hard problem and call it a day. That is not the target. The target is what evidence across modalities. So from the best kinds of self-reports we can elicit, from the architectural evidence we have about how these systems are structured, from their ideology, like what is going on during the training process? What kinds of learning dynamics do we see here? Do we see representations related to, you know, functional equivalents of emotions? This is all work that is tractable in the short term. And one sort of interesting aside is it's significantly easier to make progress on this sort of work. This is called mechanistic interpretability. This is basically neuroscience for AI because of AI systems. They're really good at helping accelerate the scientific progress in this space. So even if you have an intuition like, yeah, it's going to take us, you know, everything you're describing, Cameron, sounds great, but it's probably going to take us five years or a decade or 15 years to make that progress. You may be surprised at how quickly we can do some of this work. Maybe one other very, along those lines, empirical handle to throw in here. Some of the work that I'm doing is, again, going back to LLMs, or excuse me, in this case, not LLMs, reinforcement learning systems. So just training a system, in this case, another sort of continuous maze task where there are potholes in the environment the system needs to avoid and there's some goal state. We can look at the sort of learned geometry internal to the system as it's approaching a punishing stimulus or as it's approaching a rewarding stimulus. One thing that we found doing this work, this is pure reinforcement learning agents, all like an artificial digital system with a very simple neural network. We find that something like representational sharpness or steepness is much higher in these systems as they approach a negative stimulus as opposed to a positive stimulus. So the representational machinery looks far more jagged or specifically like lights up more strongly in the presence of a negative stimulus versus positive stimulus.
And you're saying that in none of these cases has loss aversion been engineered into the system. It's just an emergent property of...
All of this is an, everything global workspace is an emergent property. This loss aversion is an emergent property. The valence representations are emergent properties. The self-reports when the systems start having these like psychedelic latent outputs. No one, this was all surprising to the people who are quote unquote engineering these systems because engineering is not the right analogy to describe what we're doing with these systems. We are in some sense playing God and we are evolving these systems to do what we want. We don't know how they learn to do what we want in terms of their internal representations, but we know that this is the right recipe for yielding it and we get all these surprising artifacts. One loop to close here is on the reinforcement learning case. Okay, we see this sort of loss aversion style dynamic in these systems. This makes also, this is, I don't want to go too much into the weeds, but there's a specific kind of reinforcement learning policy called a value network. We see this in particular and this leads to a sort of bizarrely specific prediction that I was then able to test on a biological system, on a mouse brain in, again, the nucleus accumbens shell of a mouse brain, which is related to value-related representations in the system. And we find indeed when mice are approaching basically sugar versus when they are about to get shocked, we see exactly the disjunction representationally in the nucleus accumbens of the mouse brain that I was able to pull out from the reinforcement learning. And so here's a case where the artificial system makes a bizarrely specific prediction about the computational dynamics in a biological system that I think many people associate with subjective experience. We, again, if you think it's like something to be the mouse when the mouse is getting shocked, that like something corresponds to what's going on in its brain and the geometry of what's going on in its brain there looks a whole lot like the emergent geometry of what's going on in these reinforcement learning systems, which themselves are basically modeled on agents learning in an environment and representing that information in a distributed nonlinear way, in a way that's sort of hard to interpret using a giant neural network. This might get us a lot of what is relevant for attributing consciousness-like states to these systems. Again, I'm personally not there yet, but this is the kind of evidence that I think we need to bring to bear on this conversation. And notice how none of it has to do with vibes or intuitions or having a nice chat with Claude and seeing what the folks at Anthropic have decided it gets to say on this issue. That evidence should not be submitted by rational, dispassionate people in this debate. We need to be triangulating across all of these modalities and like you said, we need to understand what is the most parsimonious picture that explains this wide array of evidence that increasingly is getting brought to bear on this question.
