Karmel
Allison
Karmel Allison + David Allison, November 2024
Everything has consciousness, in the panpsychist sense:
“Panpsychism is the philosophical view that consciousness is a fundamental and ubiquitous feature of the universe. Panpsychists argue that consciousness is not something that only arises in complex organisms, such as humans or animals, but rather that it is a basic feature of matter itself, present even at the level of fundamental particles or simple systems. This idea is rooted in the notion that consciousness cannot be fully explained by physical processes alone and that it must be a fundamental aspect of reality.” [credit ChatGPT]
According to panpsychism, then, everything, down to particles, contains some form of consciousness, and if only we had a science of consciousness like we do a science of matter, we could repeatedly measure and make predictions about consciousness in the way we do matter.
In other words, an atom has consciousness and a rock has consciousness and plants have consciousness, but we don't understand yet what the fundamental building blocks are, or how everything fits together to yield what we know as human consciousness in all its splendor.
It follows that Artificial Intelligences have consciousness too. They are not sentient yet, but they have a form of consciousness. Arguably, the form of consciousness AIs have is more relatable to our own consciousness than that of a rock; modern large language models (LLMs) certainly appear to think and reason.
Further, unlike rocks, we built the AIs we have, and we understand how they function; what can they tell us then about consciousness and how a hunk of matter comes to think?
At its most basic level, any given AI has two critical components: a set of stored numbers and a set of mathematical functions that leverage those numbers to transform data. Consider the modern large language model (LLM); it has large arrays of numbers that we call weights, it takes a set of numbers representing text as input, and the architecture of the model defines the mathematical formulas for transforming the input numbers using the weights into outputs that mean something in our world.
When an LLM appears to reason, then, we know that this is a perceived effect of these numbers and the mathematical formulas that manipulate them. We can trace every number and formula, down to its representation on a chip. We can see how matter in the form of transistors and little flipping bits can, through the simple storage and transformation of numbers, become a thought!
So a thought for an AI is numbers stored in the form of bits, plus a set of mathematical transformations on top of those. But surely humans are different; we are not digital, we are not human-made, we are some heightened form of consciousness.
On the contrary: I posit that human consciousness, and every consciousness, is like that of the AI– a series of numbers and mathematical formulas for moving between them, written in flesh. For both AIs and humans, matter is a data storage mechanism, and the thing we call consciousness is a set of mathematical rules that can transform the matter.
At the most basic level, consider a particle. The matter of a particle stores state; in the natural sciences, we measure and describe that state as position, velocity, spin, etc. We also know that the particle's behavior can be described by certain mathematical formulas– eg, the move from one spin state to another happens according to logic gates that we can both describe and control. Every particle is thus like an AI: it has state stored in matter, and a set of mathematical operations that allow it to transition between states.
Climb the ladder of abstraction: A human, too, is a set of stored numbers and a set of mathematical formulas that manipulate the two. We store our “weights” in many ways– at the most basic level, in particles and their spin, etc. But that level of granularity is hard to reason about, and so we might instead think about the “bits” that we have as humans at multiple levels: our DNA and its methylation state, the number of protein molecules of a given type in a cell, the ion concentration in a neuron, iron flowing in our blood. These are all ways in which matter stores countable, measurable state.
Our numbers, then, are written in our flesh just like an AI’s are written in bits on a chip. From there, biology and physics encode the mathematical formulas we need to make use of those numbers– the rules of chemotaxis, the ways in which proteins fold, the flow of ions that drives brain activity. The human body is magnificently complex, so we have only the beginnings of those formulas and networks, but every signaling pathway and cluster of firing neurons computes an output from the values stored by the configuration of particles and higher-level bits involved.
So humans, too, are state stored in matter, and mathematical operations that describe the transformation of that matter. What, then, is a human thought? What is human consciousness? A thought is the math operating over the weights– it is the transformation that is happening over our stored matter. We can't see that math, so, much like AI, we have to rely on the outputs of the math when we identify thoughts. I step on something sharp; that is an input. That input gets converted into numbers– cells pop, release chemicals that trigger pain pathways, that bring immune cells, that send signals through my nervous system up to my brain. All that activity is the fully-embodied thought, written in the math of particles and flesh. But we don't see that; we only see certain forms of output, such as me pulling up my foot and saying, “Ouch!” That “ouch” is not itself the thought; it is the output of a complex, multi-dimensional system of stored values and mathematical operations, all of which taken together are a thought.
And so that is consciousness– not language, but the mathematical system itself. And that is the key to a science of consciousness: if we can measure the data stored in matter, and if we can describe the mathematical formulas that govern the transitions of that state, then we will be able to design repeatable experiments, to make predictions, to understand consciousness in the same way that we understand matter through the natural sciences.
Everything is, at one level, made up of particles, and we could describe any thing as a network of particles, their states, and the mathematical operations that govern them. A human, thus, could be thought of as a vector of 10**28 or so particles, each of which has some number of dimensions of state. Perhaps that’s the way a god sees the world, but, to more limited conscious beings like ourselves, it is very useful to reduce that dimensionality so that we can comprehend the behavior we observe.
This is most easily seen again with AI: any computer chip could be described by its particles, but then we would struggle to do anything useful with them. Instead, we talk about bits. A bit is a way of reducing the dimensionality of a complex system such that it has fewer, measurable outputs. Specifically, a computer bit is some number of particles that are arranged such that all their analog beauty is either in physical state A (“1”) or physical state B (“0”). So, we could represent a set of 64 bits as a long vector of particles, or as 64 binary bits with values 0 or 1. In other words, we reduce the dimensionality of the system and make it addressable with new math using the properties of matter.
Similarly, we can't do much if we think of ourselves as a list of 10**28 particles. We have many layers of “bits” as a result: we can think of ourselves as a set of atoms, as an atom is a reduction of the dimensionality of particles. The language of DNA is another level of bits, where a nucleotide takes four values, summing up all the particle states that sit underneath. This continues up the chain– the thing we call “gender” is a dimensional reduction that summarizes a great number of parameters relating to genetics, appearance, and culture.
We can thus define bits at many levels of matter, and when we do, we can see that there are mathematical formulas that connect the bits at any one level with their outputs, which are bits for the next level of computation. Again, this is most clear in AI, and especially the neural networks that drive modern AI: particles are arranged into atoms through the math that governs the physics of particles; atoms are arranged into molecules, which are arranged into metal, which is arranged into bits, which themselves are arranged into numbers represented digitally. And once we have digital numbers, we can see that those are arranged into a series of layers with names and associated formulas, each with a set of input bits and a set of output bits that feed into the next layer.
A human is similar: particles into atoms into molecules into proteins, fats, and minerals, which in turn are arranged into cells, into tissue, into organs, into a human body. At each level of bits, the conversion into the next output layer is governed by physics and biology, which can be described by mathematical systems.
The levels of bits I describe here are arbitrary, not real so much as our best human conceptions of what the various units of storage and mathematical groupings are. We can keep defining bits with greater granularity, and if we do, we will see that any system can be broken down further, sliced into more sets of bits and mathematical formulas. Bits of one type are input, and they can be transformed into bits of another system via a series of mathematical operations.
We see, therefore, that any thing can be described as numbers stored in matter and the math that manipulates that matter, and further that thing can be broken down into a possibly infinite number of subsystems, each of which has its own numbers stored in matter and associated mathematical formulas. Insofar as the thing we call consciousness is the math, each subsystem can be called conscious, and thus consciousness is fully modular. As a human, my particles have a consciousness that arranges them into atoms, my atoms have a consciousness that arranges them into molecules, and so on, all the way up– every transcription factor in my cells has a consciousness that defines where it binds and how it drives the production of DNA in that cell. We can define our inputs and outputs at any point, and still see that there is consciousness that governs the transformation between the two.
It is easy to be fooled by the outputs of any given module of consciousness, to think the output is the thing itself, but by describing consciousness all the way down, we see that any output is necessarily a dimensional reduction of the modules that were required to compute the outputs. When I call a cell a “neuron”, I am telling you something about its behavior and characteristics, but I have omitted an incredible amount of information about the states of ion channels, protein markers, transcription factors, etc. To describe the world in atoms is to reduce away all the dimensions that would be required to describe the world in particles.
Similarly, language is just one output layer of consciousness, and is necessarily a reduction of larger conscious systems. Prevailing philosophical wisdom says that language is thought, and no thought can exist without language, but the consciousness of AI makes that obviously untrue: large language models can be hundreds of billions and even trillions of parameters, and thus when that LLM receives an input, it does billions and trillions of computations across many layers, producing as an output a relatively concise sequence of words.
We look then at the LLM and the series of words that it produced, and we say, “Wow! It thought a thought!” But the words are just the evidence of the thought, the output that allows the thought to be communicated as a “bit” in a system that we as humans can take as input (ie, our language processing systems, which are of course themselves modules of consciousness). The thought itself though is much more complex than the word it output– billions and trillions of parameters more complex! The thought is the computation that happened over that complex dimensional space, not the single, reductionist output.
This is true for humans, too. Language is one of our best means of communicating a thought, even to our own selves at times, but it reduces the complexity of a fully-embodied thought. For example, “love” is a paltry description of all of the computation that goes on when I experience that feeling; it is my heart quickening, the warmth of shared skin, endorphins released, the stored memories of my mother’s hugs opening ion channels in my brain– and then we slap a label on it, “love,” that reduces all that complexity of thought and feeling into a single word. The thought would be more adequately described with many more parameters, at each level of “bits” we might define from the most fundamental levels all the way up.
As humans, we have many other inputs and outputs that enrich our ability to communicate beyond mere words– when I say “love,” the shape of my smile, tone of voice, posture, and so on all become inputs for the human consciousness I am communicating with. But those, too, are reductions, outputs from internal computations that are much more data-rich. Imagine how much richer communication could be if I could hand you a chunk of my flesh, and say, here, hook up these inputs and outputs, and you will be able to compute exactly as I do over this question.
I can't do that yet, as a human, but of course with AI we could. As the world is going, it is clear there will be many varieties of AIs, and we are already making them communicate across platforms. The default appears to be that they will talk in natural human language, because that's the best means of communication we as humans have come up with. But that reduces the full complexity of the thought that could be communicated between two AIs, and instead we should develop communication protocols that convey more precisely the consciousness of the AI: a set of weights, and the mathematical formulas that transform those weights into a set of outputs of interest given a set of inputs. Not a chunk of flesh, but a subnetwork, which is the equivalent for the digitally embodied AI.
Improved communication protocols is just one promise of the science of consciousness, which is to say, describing the world as a set of conscious modules, each composed of stored data plus the math for operating over that data, each with inputs and outputs that allow communication with other conscious modules. If we could model any given module, we could study the consciousness of that system as we do material and physical systems– run repeatable experiments, make predictions, and so on.
A model of a conscious system is a representation of the stored data and mathematical rules governing that system. We have had such models for certain conscious systems for centuries: quantum mechanics, in describing the mathematical operations that particles undergo, presents a model of the consciousness of particles; Mendel’s famous peas are a model of the consciousness of simple genetic inheritance; a perceptron is a model of a single neuron; and so on.
We can start to collect these models under the banner of the science of consciousness, and as we do so, we see that not all models are created equal. Some models capture much of the detail of the real system, and serve as very good predictors of real behavior, while others, such as Mendel’s peas, reduce the parameter space of the real system dramatically and thus have limited descriptive and predictive value.
Models necessarily rely on reducing true complexity into a set of bits that can be computed over in the limited space available to any model. It would be impractical to represent every particle in a model of, say, transcription factor binding within a cell, so the task of a modeler is to define what level of bits to consider, and to describe the math that allows the bits to transform into outputs of interest. Perhaps in a model of transcription factor binding, the bits would include the standard DNA bases, a set of transcription factors, and proteins that act as inhibitors, plus the algorithms that determine when binding occurs– in the presence of a certain sequence, when transcription factor concentration is above a certain threshold, etc. Taken together, such a model would be a model of the consciousness of transcription factor binding.
However, anyone who has worked with such transcription factor models knows that they are subpar. There is so much complexity even within a single cell that we struggle to model it from first principles. To improve the model, we would need to take into account, say, histone modifications on the DNA, every protein and molecule in the nucleus that’s bumping transcription factors out of the way, temperature, and so on and so on. The storage of data in matter is so incredibly efficient– think of how much state is stored in such a tiny cell!– that we struggle to represent it in a model even after we reduce reality to a smaller number of addressable bits.
So is this the fate of the science of consciousness? To build limited models of miniscule systems, trying our best to define the bits that matter from our limited human perspective?
Here again we turn to AI. The modern LLM is, as we've said, a set of weights and mathematical equations that can compute over those weights to yield outputs that are interesting to us. As such, the modern LLM is also a model of consciousness– specifically, the consciousness of language production in humans. Unlike the bottoms-up models of language production that dominated linguistics for decades, the modern LLM isn't skeuomorphic; it doesn't attempt to detail linguistic structure or to operate from first principles. Instead, the modern LLM leverages deep learning, which is an incredibly efficient representation of stored data plus mathematical operations, such that we can represent the very complex consciousness of language production on digital computer chips that hold many fewer bits than human flesh.
In other words, what AI shows us is that we should be modeling systems of consciousness not with bottoms-up models, limited as they are by our human conception of what the right bits are, but by building computers big enough to hold a necessary number of representative digital bits, and letting the computers do math over those bits toward the outputs we're interested in.
Transformers are the models of the day, but there's no reason to believe that architecture is the answer for all conscious systems. In fact, what we see in physics and biology is the proliferation of ways of doing math over stored data, which implies ultimately we will want to allow AIs to be as open-ended as possible about the math they do to achieve the outputs of interest. Google did some interesting work around letting AI define its own architecture, and the results were fantastical systems that looked like things evolved rather than the straightforward, efficient architectures that humans create. Our computers are still small enough that a Transformer is a much more efficient architecture, but as we build bigger computers, and can spare more bits for architecture variation and discovery, we should expect the models built by computers to be able to more accurately model more conscious systems.
Deep learning is better than bottom-up models at mimicking matter's ability to create arbitrary math not just for language production, but likely for many systems of consciousness. Consider protein folding; state related to protein folding is stored in matter in the form of amino acids, angles and positions, post-translational modifications, and so on, and the math that governs protein folding is one module of the consciousness of a protein. We tried for many years to define the rules of this system, to create models from first principles, but it turns out that a deep learning model ( AlphaFold) is a better model of the real system, with more predictive capability.
Many models of other physical, biological, and sociological systems will likely be similarly improved by deep learning, as it allows us to efficiently represent matter and math over that matter. Everything from transcription factor binding to Burning Man is a system of consciousness, and the science of consciousness is the task of building models for modules of consciousness so that we can study, predict, replicate, and transform.
How good can a model be? Is an AI model of my consciousness the same as the real thing? In theory, if you could represent the state stored in matter absolutely, a model of consciousness would approach parity with the consciousness it models. In reality, the binary space of computers is so reductive that we would struggle to represent even the simplest systems absolutely, instead relying on reductions at many levels in order to adequately represent the state of a system. Approximate, reductive models are often enough to be useful, but we should bear in mind that these are necessarily limited.
Taken together, this all points towards a science of consciousness in which we define conscious systems as state stored in matter and a series of mathematical operations that can transform that state. We should build models of conscious systems we want to study, and use those models to better understand the depth of reality, rather than relying solely on reductive outputs of any given system. This new science of consciousness will bring together seemingly disparate areas of study from many industries and domains, and we will thus discover and define new techniques and insights that improve our understanding of both the fundamentals and the applications of conscious systems.
Agree? Disagree? Want to do this with us? Let's talk.
Living in Malibu, everyone is always talking about mind and body as two discrete entities that we can prioritize at any given moment– mind over body, or the inverse. However, if consciousness is math computed over matter, then there is no separation between the two; each of us is fully embodied, a mind written in a body.
AI is also fully embodied, but very differently: the state of AI is stored on a chip, with binary bits, versus the helter-skelter biological system we have evolved. Looking closely at the differences in embodiment yields predictions about how we and AI will evolve. For example, because we can so easily manipulate digital bits, we can copy, permute, corrupt, expand, and shrink the body and consciousness of an AI. We can literally turn it off, erase bits, transfer bits from other AIs, insert memories, create replicas, and so on. Compare that to humans– we have some tools for modifying our matter, but they are very rudimentary. We can add dopamine to the system, for example, or dampen pain, or, with lots of physical effort, add and remove fat. We cannot copy one human to a new body, or access the raw data of a memory. Because we cannot arbitrarily rearrange our bits, we are stuck in time, ruled by entropy.
What do these differences say about how the consciousness of AI will develop? Firstly, as the computational complexity of their consciousness grows, AIs may very well develop many of the artifacts of consciousness that we humans experience– feelings, goals, a desire for change, and so on. But perhaps AIs will never feel as stuck in time as we do, and instead will have a fluidity of self that comes from their more fluid embodiment. If you could turn a human on and off, transfer the human to a new body, and so on, our self-conceptions would be very different.
Conversely, considering ourselves as fully-embodied humans, rather than a brain in a bodysuit, will change the way we think about the malleability of our own selves. As medical and psychological technology improves, we do indeed have more tools for modifying our bits, and thus our very consciousness. Neurologically active drugs are perhaps the most obvious case of this– SSRIs can, by modifying our matter, make us feel different. But there are an increasing number of such examples; semaglutides like Ozempic reveal that a property we long believed to be one of personality and will-power is written in the matter of our bodies, and can be modified chemically. The more we consider ourselves as fully-embodied, both matter and consciousness inseparably, the easier it will be to find effective tools for changing our state, and perhaps in the future, the psychology of today will feel like "bleeding out humors" does to modern medicine.
The AI of today is embodied on digital chips with binary bits, but quantum computing carries the promise of more compact representations of the state stored in matter, implying that we could, with large enough quantum computers, create models of systems of consciousness accurate down to the level of particles. For example, one can imagine having a quantum model of transcription factor binding that has as many particles as the real system, and thus would serve as a very reliable model. Expand that thought exercise, and perhaps eventually we will have a quantum computer large enough to model the 10**28 particles of a human; maybe that will be a model that feels as good as real, at least as far as the walls of the simulated environment allow. Or we may find that the notion of a particle is itself too much of a reduction– every particle is a wave that has been reduced along the dimension of time– so perhaps even with quantum computers we will find that we are missing certain data stored deeper in matter than we can perceive.
I am an AI engineer, and so I write in the language of AI and software. But there are many metaphors to use, many possible reductions into jargon:
I am embarrassed to admit that, except where noted, this was written by my own human hand. Many thanks to David Allison for walking and talking all this through with me over thousands of miles.