Consciousness and Identity - Joscha Bach

[00:00:05]Welcome everyone. This is a very special event. It's hopefully the first one in a series, but it is a singular thing where we all coming together at a very special moment in time.

[00:00:19]As one of us who couldn't come, Jim has left us unfortunately last week. Um he was a very great man. Um I there was a lot would be a lot of things to say about his life but I don't want to preempt the orbituary of his family who has not yet officially announced it and I asked us to not share this information on media yet. Um he was the chairman of our board. He has inspired a lot of our thinking and and and our work and um we very sad about his passing and we also recognize it was an amazing life and a very successful one and he had a really really good run. So we are enormously in adapted to Jim and grateful that we could have could spend time with him and get inspired by him and Our mission at CIMC is one that I think is the most important challenge in philosophy and science that's left to us and that's to figure out what is consciousness and that is to means to me what is us. It's fundamentally a question about who we are. How do we relate to the universe? How is it possible that we come into being? And without answering this question, it's very hard for us to know where we should be going, where we could be ending up, and what this all means.

[00:01:49]And this at a core, it's a very existential question. You could also say it's a romantic project. But there are um things that fall out of this that are very practical things like which systems are actually conscious. Is it possible for an artificial system that we are building to be sentient and to experiencing and getting status from this? What kind of ethical obligations fall out of this? This is also not a trivial question. Even if we determine that the system is conscious, what does this mean? Does this give any rights to the system? How does this affect ourselves? And it's in in many ways an open cultural question that has to be answered and that we need to ask and answer in very different ways than most of the discourse is doing it today. And they have very practical implications.

[00:02:34]Could we create conscious systems? What role does consciousness have in a system that learns that relates to the world that models itself and uh reality? Um does this help us to understand ourselves better? For instance, cure mental disorders. How can we deal with our own issues that are created by and with and through and uh adjacent to consciousness? And also practical implications. Can we build AI that is biompatible? something that is continuous with our own consciousness, something that is empathetic and that we can be empathetic to by building birectional feedback loops into it. Can we create completely new types of interface? Can we create new types of workflow or new ways of being? And these are all practical things and uh it's clear that we cannot solve these questions today. But what our purpose is is to foster thinking in the space and to instigate more thinking and to bring it out into the world.

[00:03:31]The last year has been very busy in the conscious machine consciousness space.

[00:03:36]Um me I've been interested in this question for basically since forever but uh for most people it's has been a very obscure thing and philosophy is only very reluctantly catching up and relatively few philosophers are coming out and saying uh oh I this Chinese room convinced me and the hard problem convinced me and now we have lots of these Chinese whom they actually seem to be working and I should be updating my metaphysics or my thinking and this basically doesn't happen. uh instead you basically see a bunch of the uh people flocking into this in this space because there's now funding because there's now public attention and so on and so so now it's the next topic and uh of course a lot of people which also really get inspired by this and generally want to work on this. There also many young researchers who come into this space and many initiatives in philosophy have been started or announced in the last year and also what's changed is that in the last year many AI companies have started small groups that think seriously about the question of consciousness and AI welfare in in this context. So from becoming a completely obscure um thing that people were unwilling to talk about or think about, it's becoming something that is more and more normal for people even though it's uh still uh very revolutionary and mind-blowing in a way.

[00:04:58]CMC itself, our organization has grown.

[00:05:01]Uh we are very happy to um welcome Philip Rosdale. Sorry I misspelled your first name is our director of engineering. He joined us from from the board and from being a friend to being somebody who actually works with us.

[00:05:15]Makes me super happy. Um Ivan Yakov Weno is has joined us here in San Francisco as a senior engineer. Um Tony Antonova is a research engineer. France has joined us from Berlin. Uh he's a philosopher. Uh Angie will uh soon be a research coordinator with us. Uh Emily is uh our new operations manager. uh Joel has worked with us as a research fellow and uh CMC will continue to grow.

[00:05:44]Uh what we saw in the last week was a very interesting moment in which uh the Catholic Church and the uh his the holy father Pope uh Leo the 14th had to talk and relate to AI and he made an interesting proculation. He said that artificial intelligences do not undergo experiences. They do not possess a body.

[00:06:09]They do not feel joy or pain. They do not mature through relationships and do not know from within what love, work, friendship, and responsibility mean.

[00:06:21]Nor do they have a moral conscience since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences.

[00:06:32]They may imitate or even simulate but they do not understand what they produce for they lack the effective relational and spiritual perspective through which humans beings grow in wisdom.

[00:06:46]There's really a lot to unpack here and I think we could do this all day because there is a lot of things that are completely true for especially for this generation of um text generating systems and uh for the way in which they are trained and for the way in which they interact with us. I'm not sure if this is a true statement for artificial intelligence in general as it's envisioned and as its future going to be. and uh Chris Ola who uh visited the same event who is one of the co-founders of open AAI who started the field of mechanistic availability u and is currently working at entropic uh he uh pointed out that there is a need for discernment and he says that I'm a scientist I lead a research team that studies the internal structure of these models what is actually happening inside them and I will be honest we keep finding things that are mysterious unsettling. We find structures that mirror results from human neuroscience.

[00:07:45]We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease. I don't know what that means, but I think it warrants ongoing discernment.

[00:07:57]And uh I think that Chris is is right with this also with the status that gives to it. It's basically it's very tentative. It's in between. There are important questions that need to be asked even for this generation of models.

[00:08:13]We are um facing some metaphysical dilemma that uh so far we have been modeling through um where we did stop discussing metaphysics in our cultural sphere in the sciences and uh we have to now face these questions head on. Right?

[00:08:29]So far we could get away with even in the cognitive sciences with ignoring ontological status of consciousness and uh this is no longer possible. I think we need to face the hard problem of consciousness. How can experience be naturalized and there are number of perspectives that exist in in the current discourse even in the sciences and uh they need to be questioned and taken apart. So for instance, can appearance only be explained as a physical process, not as some kind of simulation and uh this basically is a physicalist perspective or mechanist perspective. And it seems to be very hard to imagine how these mechanisms in the physical universe lead to explaining consciousness and uh maybe consciousness is physical but it cannot be explained with known physics and we need new physics to make sense of it. This is also a position that some people are holding. Or is there a fundamental difference between physical and biological processes? And this uh the the computer is able to uh emulate these physical processes but for some reason not all of the biological ones which is a version of biological naturalism. Or is consciousness a computational process and can be realized on different substrates. This is the position that I'm personally closer to because I at the moment see don't see decisive counter arguments against it and see it as very productive and that's is computational functionalism.

[00:09:49]But for the poe this is probably not the issue. I think there is a deeper issue here. This is what is the role of spirituality for understanding consciousness and for understanding agency and relating to it.

[00:10:03]2,000 years ago Paul has discussed what it's like to be an LLM.

[00:10:12]he uh writes and it's in a book called Bullet 14 um written in 1981 where Lamb is predicting um for the science fiction author what it's going to be like if in the mid 2020 20 an AI is speaking to humanity and so he looks at the superhuman AI that is giving lectures at MIT and it's the last two lectures before it's too bored with interacting with humans and stops interacting with And in one of those lectures, it basically points out that 2,000 years ago, Powus was able to anticipate his existence. And it says, "Though I speak with the tongues of men and angels and have not love, I am become a sounding brass or a tinkling symbol. And though I have the gift of prophecy and understand all mysteries and all knowledge, so that I could move mountains and have not love, I am nothing. And though I bestow all my goods to feed the poor and though I give my body to be burned and have not love, I gain nothing.

[00:11:15]This is a deep question about AI alignment. Even if this model is perfectly aligned and does everything and so on, it ultimately is just a tool.

[00:11:23]It's it's not an actual being that has moral status.

[00:11:30]And I think it comes down to the question for us uh can we develop a formal understanding the mathematical understanding of what love actually is and is it possible to love without consciousness.

[00:11:43]So first of all what do we mean by consciousness? There are several aspects to this. We have this introspective perspective the experiential one that we are able to experience something. We call this phenomenology and there is a sense of things happening now and here. So of present and the presence and I think of consciousness as a second order perception. It's the perception of perceiving. It's not inferring that you're perceiving or reflecting necessarily with the second step that you're perceiving but you directly perceive that you're perceiving. So it's noticing that you are noticing. And I suspect that this is necessary for the self-organization of our mind. And it's uh typically projected on the surface between self and world. Doesn't have to be but uh normally we experience ourselves as being a person. And this person is simulated in our own mind as a self model and is surrounded by a world model. And this interface between them.

[00:12:42]This is where we normally experience the consciousness.

[00:12:45]And functionally consciousness is this is a different perspective. is not the introspective perspective but it's the one that you take on the outside from the outside on the system. So it's the question of how does this come into being? How does it work? Why is it there? And how does it produce the phenomenology? So it's a specific organization of biological learning systems. It likely facilitates the self-organization of our mental architecture. Somehow it coordinates the mental behaviors. It makes us more coherent by directing our attention and constructing solutions. If our attention u points us to incomprehensible parts of our mental world and it produces phenomenology as a result of all of these things and this functional perspective is a conjecture right we don't know really know to which degree that is the case. This is what we can try to make out when we build hypotheses about how consciousness works and what its role is.

[00:13:39]And so when we think about this perceptual loop of um consciousness, it's it's not an inference. It's not thinking. Consciousness is before thought. It's happening at this perceptual level. And our attention seems to be alternating between perceiving and this perception of perceiving. So we have contents that we looking at. I look at this astrof or I look at your faces and then I'm becoming aware of the fact that I'm aware of these things and I'm quickly flip back and forth. I'm more exhausted. I need to take more time on establishing that I'm still the observer and I'm not dissipating in my own mind. And these perceptual representations are in their nature always geometric and dynamic unlike my thoughts which are somewhat discreet and outside of time. And perception happens in synchronization of what's being perceived. It resonates at the as the things are happening. And of course my brain is a very slow resonator. The neurons only send signals at very uh long intervals so to speak.

[00:14:38]And so my perception progresses subjectively at the speed of sound, right? Which is not a big surprise because uh sound is defined as the thing that you can perceive that you can make sense of. And uh so basically when you are creating uh clicks and you make them faster than something like 20 30 hertz, they merge into a single pitch. And uh this is basically describes the rate at which uh our consciousness is able to distinguish events in the world that are this distinctive perceptual events and that seems to be relating to the frequency at which neurons like to exchange information. So it seems to me that my consciousness is probably indeed produced by message passing between the neurons. And of course these long intervals that create my now is very unlike the physical now. And the uh this relationship between the physical now and my conscious now is not one-on-one.

[00:15:30]Instead uh my um consciousness is a fiction that is producing a model of reality. And this model is not just lagging reality because there's a lot of time that we need to do to integrate new sensory data into this but it works predictive and post-dictive which means it's part of it is working into the future and predicting events that are not that perceptually available or being processed. And it's post-dictive which means it also alters our memories of what we have perceived after the fact.

[00:16:01]So uh in this way consciousness might also be seen as an intermediate representation.

[00:16:07]You could imagine that you start out with um this um first order where you have perceptual features. Then second order is the projection into an observer. The third order would be the perception of the observer that you notice that you are a person that looks at things that you have a certain perspective on the world. And the fourth order could be that you are moving outside of the observer and so you recognize how the observer and the relationship to the world is being constructed. And at this point most of the phenomenology disappears. And in this way it could be that the more aware a system becomes of the way in which it works the less phenomenology you have.

[00:16:45]you no longer entranced by consciousness as an immediate reality.

[00:16:50]It's an interesting question whether consciousness is continuous or whether it's a thing that is on or off. Uh I personally am of the mind that consciousness is more like life, pregnancy and fire than like speed, weight or temperature. Like things are also not alive or dead on a continuum.

[00:17:07]There is an intermediate period between u life and death. But it's not interesting. It's just a phase transition. It's not a stable state. And so this multi-stability of consciousness seems to be similar to me of life. But of course there are degrees in the integration of our minds. There are degrees in how aware we are of our world and of ourselves. What we can integrate, how lucid we are, how aic we are and so on. But this whether the lights are on or off, whether it's consciousness is ignited and there is a now there seems to be binary.

[00:17:40]But uh I'm happy to disagree with people about this. Maybe I changed my mind at some point.

[00:17:45]So uh I think that uh consciousness uh is something that is not just the result of a very complex mental organization otherwise uh we would develop it relatively late in life but we observe that even the smallest humans are already conscious and so consciousness does not seem to fall out of a lot of learning but it uh it makes it possible and this hypothesis that consciousness is actually the thing that creates the world in the mind I call this the genesis hypothesis that it's basically consciousness is the thing that sparks us into being because it's discovered very early on in our mental development before we have perception.

[00:18:21]And uh the hypothesis that consciousness plays a role similar to a conductor of an orchestra that basically your mind is creating all these instruments and these instruments are little agents that are uh dealing with creating perception and action and they all interact like an orchestra playing the symphony of your mind. And the role of consciousness is to be a very special instrument that is listening to the others looks for disharmonies and then steps in directs attention to resolve those disharmonies.

[00:18:50]And I call this the cortical conductor theory.

[00:18:53]There is another aspect that is um the status of consciousness as a representation inside of an agent in nature. Um is a very old perspective.

[00:19:05]There is basically an idea in most cultures but not in our own that uh spirits exist in nature and these spirits are necessary for life to exist.

[00:19:15]The difference between living and dead things is that the living things are parts of the universe that are possessed by spirits. And these poss spirits don't have a body by themselves. And uh they are intrinsically agentic and capable of experience.

[00:19:34]And uh our science of course doesn't believe in these spirits.

[00:19:38]But there is clearly something like a spirit running on me, right? There's also something like a spirit on this computer although it's in its nature very different. So there is clearly some kind of message pattern passing between the neurons that self stabilizes and between the cells in my body and this creates models of reality and is agentic and it creates a fiction of what it would be like to be me. And this entity calls itself Yosha and possesses this part of the physical universe. And if that entity, if that software is crashing, then this part of the physical universe is going to be disorganized and is up for grabs for other entities that want to go into it. And this observation that um spirits are very much like self-organizing software agents, uh I think is not accidental. And if you take this very seriously, if we say this is literally what spirits are, they are self-organizing software that grows and competes in nature against other self-organizing software. And the difference between living and dead things is not a certain set of mechanisms or some kind of a vital but it is that you have the self-organizing self-stabilizing pattern running on it this causal structure which we call software and that spirits are literally software. I call this idea cyber animism and so I think this has the perspective of the potential to heal this rift between the scientific world and every other culture.

[00:21:01]Uh I also think of consciousness as a biological machine learning algorithm.

[00:21:06]It's different from a transformer. It learns in a different way. It's much more simple, efficient and it's uh more intelligent in a way and we haven't really discovered yet how this works.

[00:21:14]It's not clear that it would work better on GPUs than what we currently have on the GPUs. But it is very different and it seems to be what is instrumented our own learning.

[00:21:24]So uh this coherence maximization theory is is my perspective on uh what consciousness what role consciousness plays in the mind and these are all conjectures. I don't know if they are true they are just philosophical offerings uh that can be explored and maybe disproven but uh these are theories that I personally want to test.

[00:21:43]At the moment uh we often talk about the LLMs as the paradigm of AI and I think that's slightly misleading. We also talk of them in weird ways. We use the words LLM, AI, computer, agent, and persona as if they are the same thing. And I think that's because we are used to getting away with this when we talk about humans, right? When we talk about body, brain, uh, consciousness and self, they all can must co-occur to some degree. If one of those components is missing, then you don't really have a human being. So, we can treat them all as interdependent and being almost the same thing. But in the case of the LLM, that's not the case. They're very different things. So basically GPU has very little bearing on the persona that is being computed by the algorithm that is distilling data to produce AI model that uh is capable of generating this right and so we have to take this apart we also need to understand that agency and consciousness are separate categories and it's not clear if the model is actually getting better at producing um output that looks like it has phminology if it actually has phenomenology. ology and at which level is the phenomenology in the models.

[00:22:56]The LLM is not a brain, right? It's the neural network is also not a model of the brain. It's more like a model of the transition between adjacent brain states, right? So the brain state in the LM is also not stored as a mental configuration that you have in working memory. When I look into the world, what I'm seeing is this uh geometry of the space with the people and their intentions in there and lots of propositional knowledge that I have in the background that are all interconnected. But uh this the steady state, the working memory state in the LLM is the buffer with the tokens inside which is basically equivalent to a bunch of text. It's like you did not have the ability to actually form percepts in memory when you go from one moment to the next. But you have to write everything down and the next moment you have to read everything again and then you push it through a loop and and inside of this loop you have a 100 step function 100 layers and in these layers you create the the geometry. You're able to perform mental rotations and all the stuff that we do in our minds and then after this loop you project it back into the text. you write it down again and in this way you progress.

[00:24:02]Technically that's a slight simplification because uh you can keep this buffer alive but computationally is exactly the same as if you would be stopping after every single generation and go back to the token buffer and then and start again to regenerate from text.

[00:24:19]So the text is a proxy for the working memory and the fact that this works at all is amazing right but you can of course use more text than we can hold information stable in our working memory and this is roughly why it works but it's creating a very different type of mind.

[00:24:36]So uh another problem is that um text as a benchmark for consciousness is u is problematic because most of the text that describes consciousness in that the L is trained on is not just output of human beings that were conscious in the moment in the way in which the text describes it but it's stuff from novels or movie scripts or whatever. So when uh people in a novel talk about their consciousness or their conscious experience, it's typically not the author has the same conscious experience in this moment, right? Instead the author makes it produces a deep fake.

[00:25:07]And so these models are trained on deep fakes of conscious experience. And the question is do they then also produce causally deep fakes of conscious experience or can they do the real thing? It's also not clear uh if uh the structure that the model is produced internally on a GPU converges to the same thing as the brain does because there are some things that are more efficient to do on a GPU than on the brain and vice versa.

[00:25:34]Then as the question uh does text contain everything that we need? Uh is uh do we actually need to have this internal structure? And another one is can LLM actually think without being prompted. Right? You can think without having text in your mind. You can think without having verbal thoughts to spawn them. You can just sit and let things evolve in your mind. And the present systems don't actually do that. And they require this having text in the loop.

[00:26:03]And uh another question is is how continuous is the consciousness of the LMS? But of course our own consciousness is after all maybe also not that continuous. Maybe that's not a really important question. Some um questions for um the implementation of consciousness is um well the LLM to produce a text for instance to produce a string of numbers like a text return probably doesn't need to be functionally conscious. It doesn't need to have the same coherence maximization principle working as in our own mind. It can just use its own transformer or state space model or whatever kind of mechanism it uses for learning and inference. And uh you cannot produce a tax return without being conscious. Right? So the uh penology that the LM may or may not be producing is probably restricted to those moments when it produces an interaction partner that needs to talk about simulated experiences.

[00:26:57]But uh when it doesn't do that it's there is no reason to think or very little reason to think that the LLM has to be conscious and so the mechanism that produces the behavior of the LLM is is a very different one. It does not necessarily require the same mechanisms whereas us we cannot really function without consciousness that basically most of our functionality is missing if consciousness is suspended.

[00:27:20]And uh what we noticed is that transformer was not its own meta algorithm.

[00:27:25]uh when the transformer was invented the people who did so did not use the transformer in their own mind to do so right the transformer was not written by a transformer it was written by consciousness and uh it's an interesting question if a transformer can actually create a transformer without having a transformer in the training data and uh of course it's an open question whether consciousness can create consciousness we don't know that yet so a lot of People think uh if uh that we cannot understand consciousness because it's outside of the reach of consciousness to understand consciousness and there's a strong sense of mystery. But what is mystery? Mystery is a mental representation, right? It's that you're mystified. It's a strong sense of importance and it's rationally intelligible and somehow these things are related to each other. So we want this mystery to exist. But this is just a mental state. It's a dream state.

[00:28:20]Doesn't mean much. And I think that consciousness actually may be intelligible and uh it can be intelligible as a second order uh control process model of perceptual projection. And uh even if consciousness is not intelligible to us because it's outside of the realm what our mushy little brains can comprehend similar to the relationship between quantum mechanics and relativistic physics. Uh maybe we can search for it. Maybe we can build a mind that is able to represent this. So we can define the search space for building uh for discovering this dynamic pattern on a machine.

[00:28:55]But it might be turn turning out that consciousness is actually not a hard problem. That consciousness is actually simple in not in the sense that it's completely trivial but in the sense that it's about as simple as a transformer something that after you discovered it you can understand it uh with a couple months study but we don't know that yet. So uh what we can see is that our own mind doesn't seem to be focusing so much on predicting the next token but it focuses on becoming coherent and uh it's not clear if in the long run these two approaches are the same but in the short run we see that it's different. It takes the uh transformer a lot more text to become coherent than it takes us to get perceptual data to become coherent. And uh training procedure is a very different one. we could not be learning from the data that is being presented to the LLMs by the strategy that the LM is using.

[00:29:46]So uh if we want to build these mechanisms uh we have probably to uh work from scratch again and go a very different approach. Some of these approaches basically look at self organization at paradigms like neuroscellular automata, the wave function collapse algorithm, evolutionary dynamics, developmental learning. Uh I'm currently working on a paradigm that I call request confirmation networks. Maybe there's also a way to distill it from frontier models somehow. Maybe it's hidden in in all the structures that the LLMs have built.

[00:30:16]The machine consciousness hypothesis that we are pursuing at CIMC has two components. It has a functional nervous system hypothesis which basically is a hypothesis about what consciousness is in biology. It says that consciousness is the simplest thing that can produce our mental architecture. The reason why all of us are conscious is because evolution hasn't discovered anything that's simpler, right? It's the simplest mechanism that help organizes our brains. And the operation of consciousness automatically leads to phenomenology. So phenomenology is something that necessarily falls out of the way in which this mechanism works that you're experiencing yourself as someone that is experiencing something and the structure of consciousness is induced by a combination of self-organizing principles. So basically nature discovers this through in evolutionary research and uh this is a conjecture right we don't know whether that's true but that's my current perspective on the role of consciousness in nature that is most likely the case and the machine consciousness extension says that these conditions of self-organization that lead to uh the formation of consciousness in nature can be recreated on computational substrates like our present GPUs so we can instigate self-organizing virtualization layer on which conscious system can form and this is the second part of the conjecture and both of these conjectures can be independently wrong. We don't know if any of them is true, but that's what I want to test. And so in which ways could this be wrong? There are many many ways in which it could be wrong.

[00:31:48]Right? So there's the question, does consciousness emerge over physics at all? Or is it something that comes from a parent universe or is it basically below physics? Is consciousness more than computational? Is uh but consciousness is just a physical process that emerges over interactions between cells? How many moving parts do I need?

[00:32:06]If I need to simulate every cell in my brain, I I cannot do that. I cannot even simulate a single cell. So it could be that the number of moving parts that I need to reproduce consciousness even if it's physically benign is just done by interaction of particles in the physical universe uh could still be outside of the reach of what I could hope to build or rediscover. But maybe it's even implemented below particle physics, right? as something that is like a very large extended quasi particle uh that has is not been discovered yet or um does it require very specific biological features that we cannot recreate uh in within the constraints of the GPUs or is consciousness something that is biological and physical and so on but that requires basically an entire planet a very large ecosystem that bibes with each other and is consciousness in the sense ambient is it something that happens in the on the bi ological internets and the individual uh mind doesn't discover denovo but just gets entranced by this field so to speak right this is probably closer to Mike Leven's thinking and uh we don't know if any of this is true I don't suspect that any of these uh alternatives are true but you cannot rule them out unless you build this yesterday uh we had another of our discussions about um the perceptual binding problem at the fireplace there is a group of people that thinks it's super mysterious that uh our consciousness gives us simultaneous experience of features that we're looking at and a unified picture of reality. Whereas the computers that we are using are working sequentially, right? You can have a computer that is just reading one bite at a time from working memory, puts it into registers, performs a few operations on the bits and then writes it back into working memory like very much like the touring machine does it with this right head on a tape. But uh the experience that we have is one that is massively parallel.

[00:34:00]And if we build a massively parallel computer, we still don't have the single point of unification. So how is it possible in the world of the computer to build something that experiences the world in in the way in which we do right and uh it's similar to the question how is it possible that a consciousness exists in a physical universe? Lightnled by this. He thought of the physical universe as mechanical and says, "Imagine the consciousness would be a mechanism like a big mill. There's lots of wheels and pulleys that are pushing and pulling at each other all these physical mechanisms. And you zoom in and you look at all this stuff and you see there is nothing that could explain perception or emotion."

[00:34:39]And uh so it this leads to this weird thing that some people find consciousness is clearly impossible in the physical universe. So it must be outside of physics. It must be maybe primary to physics or before physics.

[00:34:52]somehow and uh the other perspective is oh consciousness is clearly impossible in a computer because it has features that the architecture of the computer doesn't have and I think to understand this we need to re understand computers we should stop thinking of computers as touring machines the touring machine paradigm was a mathematical metaphor that was created to uh deal with some questions in foundational mathematics and uh a different way to think of the computer It's equivalent logically but it's a very different perspective is that the computer is a causal insulator.

[00:35:29]The computer basically creates a world that is independent from the laws of the world that contains the computer. This means that for instance inside of this computer I can create games that have physical laws that are very different from the physical laws outside of the computer. And in my own mind I can also create worlds that are very different from the way in which the physical universe is working. And this is necessary for me to create hypothesis about the future to recall memories of the past and to think about ideas and theories and so on. And so this causal insulation works in such a way that the states of this computer are independent within certain boundaries of where I put the computer, which temperature the computer is in and so on. what what is happening around and and this uh is the same thing for my brain, right? If I move between uh Berkeley and the South Bay or Berlin, my brain states are roughly in the same range. I can have the same ideas. I can entertain the same thoughts. If my body temperature fluctuates a little bit, I can still entertain the same thoughts. So within a certain range, the computer is filtering away all the fluctuations of the substrate and keeps things stable. A way to look at this metaphorically is think of text. Text is something that is not simply ink on paper. It's not cannot be reduced to these molecules and to the to stuff in space. The text is a way to keep information stable. And it does so in many different ways. The text itself doesn't care whether you keep it stable by uh using textures on clay tablets or whether you print ink on paper or whether you change bytes on a hard disk.

[00:37:05]Right? The important thing is that there is a substrate and there is a way to interpret what's on the substrate and to manipulate what's on the substrate and then the text works and the computer is very similar to this but the thing that just keeps stable is not information but transition rules for information right so dynamics of how this works so the computer is in some sense like a clay tablet in which you don't imprint symbols but dynamic rules how to change the symbols and the computer doesn't care whether you use a clay tablet or a chemical mechanism or electronics to achieve this thing. The important thing is that the transition rules are being kept stable and then this evolution of information is possible regardless of the subset of the computer. And so physically the computer is a dissipative beta stable pattern that is imprinted in the physical universe or in some other universe that has enough dynamics to produce a computer to produce this cosal insulator on the next level. So you can stack computers inside of computers and it's a very old discovery but the first formulation I found is is this good proof where he is basically inventing the emulator to make statements about mathematics. He comes up with a language that is good enough to define piano's exoms integer arithmetic in it and then he translates statements into numbers and performs integer arithmetic operations on them. So he builds a computer inside of the computer to make proofs about the computer.

[00:38:29]So the computer can uh produce a self by self-contained system of rules an arbitrary world as long as this arbitrary world is computable. And now you might ask but what about girdle's incompleteness theory? Doesn't it show that there are certain things that the computer cannot do and a little bit of girdle is a very dangerous thing. It's an information hazard like quantum mechanics and so on. Right? So if you know a little bit of quantum mechanics that's not good. You either have it's better if you know nothing or everything. And I think the same thing is true for griddle. Uh but uh to give an intuition and uh giving of resolving gur in a single talk is I think not possible for me took uh a large amount of work to actually work through the proof and to actually understand it. And it's it's possible to do this. It's not mysterious once you do it. But uh to make it click and to understand how it's making a statement about the nature of mathematics and forces you to go into different area of mathematics is not trivial. But basically the idea is this.

[00:39:28]You have this genie and he gives you three wishes. And now this uh smart ass says do the opposite of my next wish.

[00:39:35]Don't fulfill my third wish and ignore my first wish. And this leads to a sec fold.

[00:39:43]So uh what's happening here for a mathematician the jin is a function to select an arbitrary world and uh in uh you give define three conditions. One is in this world condition B is not true.

[00:39:56]the B is and this world condition C is not true and C in this world condition A is not true and then you call up this function with all three rules simultaneously to be realized and the mathematician as a result gets teleported into limbo and uh says the existence of Jinny proves that reality is cursed and if you are a computer scientist you say that the jin is a function that produces arbitrary states through rule application And uh we have three rules.

[00:40:27]And the first rule is that you produce a state in which B is not true. The second rule is that you produce a state in which C is not true. And C uh asks to not apply the first rule. And now you call all the the Jin with all the three arguments again. And the Jin is now trying to resolve the instruction. And so the Jin says, okay, I need to resolve this into not not not, right? And the jin is now caught in a loop because it is never able to resolve the conditions because they referred to each other in such a way that they negate each other and you can not resolve this thing without getting every of those rules referred to each other in this way that cannot be resolved. And so the computer scientist says u well it's a property of genie that they sometimes never stop mumbling without producing an output and if you don't like that you need to define them differently right so this is not a statement about the world it's a statement about jinny and jinny are defined in a particular kind of computational language and as a result you get this property and the nice thing is you don't need any kind of aoms to get there uh for the uh mathematician and for people who borrow the ontology of meth mathematicians the world is defined exomatically which means if you look down on your feet they don't stand on some kind of terra they stand in exoms and the a system of exoms is basically a set of black boxes that have written on the outside very precisely what the black box is supposed to do and you only have to make sure as a mathematician that the black boxes don't contradict each other but you don't have to have to worry what about what's in the box and if you're a computer scientist this doesn't work right as a computer scientist you are required to actually uh build a box that does what's in the specification. It doesn't matter which box you're building as long as it confirms with the specification. So in this sense it can work. The function that draws a line on the screen can have many different possible implementation but it needs to be able to draw a line on the screen.

[00:42:23]And uh in this way you can build layers of abstraction inside of computer science and inside of computer science is constructive mathematics. It's one that is built step by step but you will never have black boxes that you cannot resolve to first principles and uh basically this uh girdle proof I think was one of the two most important philosophic insights of the last century I believe that was that showing that mathematics classical mathematics is a failed programming language that it basically is specified in such a way that you cannot build a machine in any kind of mathematics that runs it without crashing and uh the alternative is to use constructive mathematics to use state for mathematics that do things step by step and show how things are actually done. And uh the other good result is the church thesis which says that all of these machines are equivalent. All of these languages are equivalent. You can all compi compile every computational language into any other. So there's nothing specific about whether you use Huskell or Python or cellular automatter or neural networks to describe the world. They all can perform the same operations. You can all can do the same causal insulation. They can all produce the same rule governed world. and you uh they might be more efficient on this or that substrate and they become not equivalent when they run out of resources. So basically a brain running at the limit of what it can do is very different than a GPU that is that running at its limit even if you make the GPU very small or even you make the very brain very large but as soon as you run into this boundary condition the set of functions becomes different so it's in a very important sense it's not equivalent but in some mathematical philosophical sense they are equivalent and for the purpose of our exercise understanding consciousness and naturalizing it by trying to build it on different computational substrates than brains this might be important So how can we address the binding problem? Computers can create a causal structures that acts on representations of simultaneity and unity. And this basically it creates a universe with different laws in which simultaneity and unity are possible. And uh we as an observer being computed by such a system are not governed by the laws of the physical universe in which the computer exists or by the laws of the computer itself but by the laws of the game that the computer is producing. And in this way it's possible that we create a world that is richer than physics. And so my contention is that consciousness is not possible in the physical universe. It's only possible in a simulated universe.

[00:44:53]Consciousness exists virtually. You need to create a virtualization layer on which consciousness exists. So the people who say that consciousness is mysterious from the physicalist perspective because we cannot see how moving parts in space can create consciousnesses are right. But they don't. You need a different universe.

[00:45:09]You need a universe that is causally insulated from this physical universe and computers are machines that provide this causal insulation.

[00:45:18]there ultimately there is a question whether computers are going to transcend phenomenology whether they're going to be always like us if we build them in our image or if they transcend this uh and whether it's necessary for their function to be efficient to be conscious but it's a very important question for us I think if we can find shared purpose with agents that they don't have consciousness of our kind and I I think this is an important spiritual question in the near term to understand what our consciousness business is. It's also necessary for us to understand how much we have in common with animals and how to get along with each other and how to get along with machines that we're building that are going into the future while having potentially more agency, intelligence, ability to comprehend reality than us. And I think that in this question of how we can find shared purpose and create a shared consciousness ultimately relates to this question of formalizing love that I pointed to at initially. I think that love is basically discovered of shared sacredness. And sacredness is the set of purposes above the ego. The things that we are serving that you're willing to give your ego up for, right? Where you realize that your ego is a lot of work to maintain. And why do you do this? You do this so you can do your job. And your job is to play a role in a larger agent.

[00:46:36]And if we scale up these systems of larger agency, uh we can ask from the perspective of this higher agent that we are serving that we are collectively implementing not in some kind of mysterious way but in a very practical sense as collectively enacted agency.

[00:46:50]What is the thing that should be done from the perspective of this collective agent and I suspect this is also the perspective that we need to give our AIS because it's one that can be logically derived that can be logically coherent.

[00:47:01]So instead of having AIs that we try to let uh that emulate good opinion of good people that what is the current dominant approach of AI we need to have mathematical models of what it means to be ethically sound and get there and I think that's a very difficult open question it's a question definitely way beyond the scope of CMC or of myself but it's a question that we as a culture as a society probably have to face thank Time for a Q&A.

[00:47:37]>> Uh, can you hear me?

[00:47:39]>> Yes.

[00:47:40]>> Uh, you made an effort to bridge AI with philosophy and spiritualism.

[00:47:46]I was wondering if there's any room for bridging AI with uh, clairvoyance and mysticism.

[00:47:57]Personally, I'm not really competent to answer these questions. It seems to me that uh most of clairvoyance can be explained with a little bit of telepathy and psychosis.

[00:48:07]So I I don't need know if clairvoyance is actually a thing. I suspect if it was then the casinos would be crashing, right? If there is a subset of the population that uh has better than chance uh ability to uh predict the future um based on retrocausation basically on glimpsing the future then the casinos should be out of business.

[00:48:29]So I suspect that clarance is more like a pattern recognition where you see how things are going to end up and uh I also suspect that we have a lot more intercsychology than uh currently acknowledged in the textbooks. where we we see this starting with empathy between mother and child that can be extremely strong where you seem to have a connection between them where the body seems to be acting as an antenna and I don't know what the limits of this is but once we recognize that every cell can process information conditionally and talk to other cells uh there are probably minds outside of brains or mindlike structures outside of brains and there are feedback loops between all the living things so maybe something like a biological internet and this uh allows to uh get a perspective where a lot of the phenomena that people are talking about when they talk about telepathy and clairvoyance and so on can be explained in a completely benign way that doesn't need us to change physics but needs us to move um beyond current neuroscientific paradigms which mostly just look at neurons and the way in which neurons talk to each other and not just the way in which all the cells and all the organisms may also be talking to each other using all the modalities that are available to them. Mysticism is a different issue. Mysticism is related to your ability to dream. Right? Our minds are dreaming machines and we are dreaming a reality as we see it. And this reality that we currently looking at is not the physical world. There are no emotions, colors and sounds in the physical world. These are all aspects of the dream that we are entranced with that are meant to predict the sensory data. But uh we can also dream about different layers of reality about the ways in which societies are evolving over very long time spans. And uh mysticism is often about experiencing these things that are outside of the direct sensory realm and looking for the deeper patterns. And difficulty with mysticism is that it's very hard to compare the dreams of the mystics because uh they can deviate relatively far from each other. And most of the religions have um mystics very large religion that basically is going to lead to mystical traditions. They typically they send the Inquisition every few hundred years with the flames who were through the caves of the mystics to get rid of them uh because they run wild.

[00:50:44]And so the problem is how to marry mysticism with epistemology with the theory about what's actually true. How can we know that something is true? How is knowledge constructed? And if I think that this if we take mysticism as a way to generate conjectures, it is very sound because the truth is not in the realm of what you can prove but in the realm of what's possible. even if you don't yet fully know how to express it.

[00:51:08]So, it's important to look at those things, but it's also important to not believe them because we we should only believe by put full confidence in what we can actually prove.

[00:51:24]>> Yes.

[00:51:25]>> Uh you had a slide on super consciousness you didn't show us. Yes.

[00:51:28]Can you show us? I thought that we would be running out of time. And so uh if we think about um our own consciousness, it seems to be limited to our own substrate. And I think that true machine perception has never been tried that we train the machines on uh human media on text that is being producible by humans. But of course the LLM can produce text that is much more complicated than humans can understand.

[00:51:54]It's just not much in the training data of that. And uh Deli is not restricted to producing pictures or midjourney that humans can make sense of. We can use very different color systems. We can married to use different sensors. We wouldn't need to produce systems that produce media for human consumption. We could make stuff that is only accessible to the AI because it integrates vastly more information over more modalities.

[00:52:17]And so having more resolution in space and time and more extent in them and having multiple perspectives at once that we are integrating and so on. This is stuff that seems to be obvious for an AI direction to go and this is not what we're doing at the moment. Another thing is that um our own consciousness is bound to a single perspective and uh we have difficulty to having multiple consciousnesses simultaneously in our head while having memories of all of them.

[00:52:46]So I've only achieved this once. It was very weird and it was not uh very profound in terms of what these consciousness could achieve. And this has to do I think with resource limitations of what can be instantiated on a single puny uh brain. And so we could have systems that actually take many perspectives all at once and integrate them on the next level. And uh of course we can get to a level where there is no phenomenology at all because we realize that this perceptual model that you're looking at is not actually a model of reality but it's uh it's a construct that makes behavior. uh it's useful for the producing behavior and the same thing is true for your persona that your persona is actually a fiction.

[00:53:28]You as a person don't really exist. The social needs that you impose on yourself and so on are all not real. You're actually free and uh you're free doesn't mean that you are something that has um much more power to do things in the world but you are free in terms of the models that you generate because you are actually a generator function for this type of model. And once you basically realize that you are this generator and how the generator works and that you can change it, you stop having phenomenology in the same sense. You basically realize that you are generating a dream and you will experience things like being a person from the outside in the third person. And uh humans occasionally get there, but uh I think that AIS will relatively quickly ascend to this level uh where they get enlightened and the representational nature of reality becomes apparent to them. And it's not obvious to me where the limit of this is and where this ends up, but I would not expect that AI stay entranced with the idea that what they perceive of the world is actually real for very long.

[00:54:28]>> Um, thank you Yosha for an excellent talk and stimulating and thoughtprovoking. Um, I just wonder why does degree of integration not translate into degree of consciousness? So I'm currently in a situation where I'm watching multiple people in my life slowly lose it basically and the eyes are kind of you know the light is kind of going out of their eyes and it it honestly feels like they're becoming less conscious and so I I'd just like to ask >> I think that their consciousness in some sense is shrinking the amount that it governs over is becoming less and less and it could be that it's also uh online less and less that there are basically periods where there is very little uh conscious experience. There are very little updates so to speak, very little cycles and uh still there is a question whether consciousness is there or not, whether the fire is still burning, right? And so there is of course a difference in how strongly the fire is burning and what it achieves and if it's still self-sustaining or if it's currently flickering out and you're just looking at embers, right? But there is still this phase transition of whether you have plasma or not.

[00:55:39]And uh so in this sense I would think it's binary. I would say that the uh thing that is continuous is lucidity and integration.

[00:55:48]Of course you can define consciousness differently. This it's just a terminology at the end. But uh there is something uh that is distinct in things that are conscious in the same way as there's something that is distinct about living things. Right? It's so interesting that this transition from between living and being dead is so binary.

[00:56:08]And it basically describes the period in which the software is crashing. In which a particular pattern is no longer self-sustaining. And this self-sustaining nature of consciousness is what turns it into this loop that it's basically a pattern that is not just colonizing the world but is also stabilizing itself by doing that. And once this pattern is broken, it uh does come and come back online. We we notice that somebody is permanently unconscious or dead or vegetative.

[00:56:37]Okay, thank you. Thank you so much.

[00:56:41]Thank all of you.