https://osf.io/preprints/thesiscommons/wehmg_v1
Abstract
How To Build Conscious Machines
How to build a conscious machine? For that matter, what is consciousness? Why is my world made of qualia like the colour red or the smell of coffee? Are these fundamental building blocks of reality, or can I break them down into something more basic? If so, that suggests qualia are like an abstraction layer in a computer. A simplification. Some say simplicity is the key to intelligence. Systems which prefer simpler models need fewer resources to adapt. They ‘generalise’ better. Yet simplicity is a property of form. Generalisation is of function. Any correlation between them depends on interpretation. In theory there could be no correlation and yet in practice, there is. Why? Software depends on the hardware that interprets it. It is made of abstraction layers, each interpreted by the layer below. I argue hardware is just another layer. As software is interpreted by hardware, hardware is by physics. There is no way to know where the stack ends. Hence I formalise an infinite stack of layers to describe all possible worlds. Each layer embodies policies that constrain possible worlds. A task is the worlds in which it is completed. Adaptive systems are abstraction layers are polycomputers, and a policy simultaneously completes more than one task. When the environment changes state, a subset of tasks are completed. This is the cosmic ought from which goal-directed behaviour emerges (e.g. natural selection). ‘Simp-maxing’ systems prefer simpler policies, and ‘w-maxing’ systems choose weaker constraints on possible worlds. I show w-maxing maximises generalisation, proving an upper bound on intelligence. I show all policies can take equally simple forms. Simp-maxing shouldn’t work. To explain why it does, I invoke the Bekenstein bound. It means layers can use only finite subsets of all possible forms. Processes that favour generalisation (e.g. natural selection) will then make weak constraints take simple forms. I perform experiments. W-maxing generalises at 110-500% the rate of simp-maxing. I formalise how systems delegate adaptation down their stacks. I show w-maxing will simp-max if control is infinitely delegated. Biological systems are more adaptable than artificial because they delegate adaptation further down. They are bioelectric polycomputers. As they scale from cells to organs, they go from simple attraction and repulsion to rich tapestries of valence. These tapestries classify objects and properties that cause valence, which I call causal-identities. I propose the psychophysical principle of causality arguing qualia are tapestries of valence. A vast orchestra of cells play a symphony of valence, classifying and judging. A system can learn 1ST, 2ND and higher order tapestries for itself. Phenomenal ‘what it is like’ consciousness begins at 1ST-order-self. Conscious access for communication begins at 2ND-order-selves, making philosophical zombies impossible. This links intelligence and consciousness. So why do we have the qualia we do? A stable environment is a layer where systems can w-max without simp-maxing. Stacks can then grow tall and complex. This may shed light on the origins of life and the Fermi paradox. Diverse intelligences could be everywhere, but we cannot perceive them because they do not meet preconditions for a causal-identity afforded by our stack. I conclude by integrating all this to explain how to build a conscious machine, and a problem I call The Temporal Gap.
Authors
Goal-directed Behaviour
Bennett defines goal-directed behaviour as an emergent property of systems acting to preserve their own existence, a drive that stems from a fundamental principle the author calls the Comic Ought.
Simp-maxing Systems
A simp-maxing system operates on the principle of Occam’s Razor, favouring the simplest explanation or model. The goal is to maximise the simplicity of form, meaning choosing the policy that can be described with the least amount of information (e.g. the shortest program).
Bennett argues that this approach is not optimal. While it often works in practice, he proves it is neither necessary nor sufficient for maximising the probability of successful adaption (generalisation).
W-maxing Systems
A w-maxing system, the proposed alternative, operates on what he calls Bennett’s Razor: “Explanations should be no more specific than necessary”. The goal is to maximise the weakness of constraints on function. A weaker policy is one that is more general. It is a correct policy for the largest amount of possible tasks.
Cosmic Ought
Bennett acknowledges the philosophical problem and states that to build a system with purpose or goal-directed behaviour, he needs a foundational, universal ought that isn’t just an arbitrary human value. He can’t simply say a machine ought to do something based on a description of its current state (is).
He proposes that this fundamental “ought” comes from the universe itself, which he calls the cosmic ought. In his thesis, this is defined as follows:
The universe is in constant change, and time is equated with change. As the universe changes from one state to the next, some things (or “statements”) persists while others perish. This process of persistence acts like a form of natural selection on everything, not just living things. Whatever persists is what the universe, by its own physical laws, has “selected”.
Stack Theory
Bennett rejects the concept of computational dualism, the idea that software (“mind”) is a separate entity that runs on hardware (“body”). He argues that software is simply a physical state o hardware, and you cannot understand an AI by looking at its code alone without considering the hardware and environment that interpret it.
The proposed stack theory is a framework for understanding reality and intelligence by describing everything as a series of nested abstraction layers, where each layer is a physical state of the layer below it, extending from software to hardware, all the way down to the fundamental laws of physics and potentially beyond.
Causal Identities
Bennett defines causal identities as embodied policies that a system learns in order to classify specific objects, properties, or agents that cause it to feel attraction or repulsion (what the author calls “valence”). In simpler terms, a causal identity is how an organism learns to recognise a “thing” in the world by the effect it has.
Causal identities are prelinguistic classifiers. Before an organism has a word for “fire”, it develops an internal policy - a causal identity - that allows it to recognise fire as the cause of a specific feeling of repulsion (heat / pain)…
Bennett argues that a system will only form a causal identity for an object if two conditions are met:
- Incentive: The object must be relevant to the system’s goals, like survival.
- Scale: The system’s own physical structure must be complex enough to represent and discriminate the object.
If these preconditions are not met, the object or entity effectively does not exist for the system; it is perceived as random noise.
A self is a causal identity constructs for its own agency. Bennett categorises them into three orders of selves:
First-Order Self
This is the most basic form of self. It is the causal identity an organism learns for its own interventions in the world. It allows the organism to distinguish between events it has caused (e.g. “I moved”) and events it merely observed (e.g. “The world moved around me”).
Second-Order Self
This is a more complex causal causal identity built upon the first. It represents an organism A’s prediction of another organism B’s prediction of A’s first-order self. In simpler terms, it’s “what I think you think of me”. This order of self is necessary for true communication.
I show how 2ND-order-selves are necessary for communication as described by Grice. Grice argued that if I am speaking to you, my meaning is what I intend. You have understood me if you infer my intended meaning. A 2ND-order-self lets me predict what you think I think. I can use that to predict what you will think I intend. Hence I can anticipate what I need to express to bias your inference toward my intended meaning. Conversely, if I want to know what you mean, I can abduct that from my prediction of your prediction of my prediction of you.
Third-Order Self
This is an even higher-order causal identity where an organism becomes aware of its own self-awareness. It represents a prediction of its own second-order selves. This allows for more complex planning and social reasoning.
Conscious vs. Unconscious Processing
Bennett distinguishes between the small amount of information we are consciously aware of (like the words being written) and the vast majority of information that happens “in the dark”, without our awareness (like muscle atrophy).
He mentions two problems of conscious processing:
- Why are we consciously aware of some information and not others?
- Why does phenomenal consciousness exist at all if all the necessary information processing could theoretically happen “in the dark”?
Bennett argues that phenomenal consciousness is the function of consciousness in the first person and that natural selection demands that there exists a self to be subject to sensation.