Towards Consciousness Engineering By Max Hodak

These are my detailed notes of the talk Towards Consciousness Engineering by Max Hodak. It’s written with some help of Gemini 2.5

I. Definition of Consciousness

The term “consciousness” is notoriously slippery.

An Operational Definition of Consciousness

Consciousness is the thing that is modulated by anesthetics or psychedelics.

This definition deliberately grounds the phenomenon in biochemistry. Psychedelics, in particular, demonstrate that this thing we call consciousness can be reliably and repeatedly perturbed by pharmacological agents. There is a real, physical process that can be explained with science and affected by chemistry.

Guiding Assumptions & Key Questions

To build a theory, we need a starting point:

• Fundamental Property: We’re looking for something as fundamental as mass or charge, not an incidental byproduct.
• Computational Basis: It is related to the computation and transformation of information. The “bit” is a useful unit.
• Emergence from Ordinary Matter: The brain, the seat of our consciousness, is composed only of ordinary matter arranged according to the rules of chemistry. There is no “special stuff”
• Orthogonal to Intelligence & Agency: Pure experience (qualia) can exist without intelligence, and agency (the ability to act) is a separate concept.

This framework leads to a set of core questions that any good theory should be able to address:

1. The Spike Dilemma: The brain is encased in a dark skull. It only receives information as electrochemical “spikes” through a few fiber bundles (e.g., optic nerve, vestibulocochlear nerve). How does it know that some spikes should be seen and others should be heard?
2. The Origin of Modes: Why do we have the specific sensory modes that we do (vision, hearing, monologue) and not others?
3. The Passage of Time: What determines how quickly time passes phenomenally? What defines the length of “the moment”?
4. The Individuation Problem: Why do I experience my vision and my hearing, and not my vision and your hearing, if it’s all just computation in the universe?
5. The Classic Hard Problem: Is my experience of “red” the same as your experience of “red”?

II. The Brain as a Symmetry Prism

The answer to the first two questions lies in the structure of the information the brain receives. The brain acts like a “structural prism,” splitting information into different modes based on the symmetries inherent in the data streams.

The Ferret Experiment & The Sea Star’s Lesson

Classic experiments demonstrate this principle (For this better watch a video around 8th minute). In one, a ferret’s optic nerve was rerouted to its auditory cortex and the auditory cortex began to develop tuning properties characteristic of the visual cortex. This suggests that cortex is a flexible substrate that learns to process information based on the intrinsic structure of the input signal.

A simpler example comes from the brittle sea star. One species has melanin-filled chromatophores that act as tiny lenses, shielding them from light in specific directions. Another species has the same number of light-sensitive cells but lacks these lenses.

• First receives a signal that is a pure scalar intensity—it is invariant to the rotation of the light source. This is analogous to a simple auditory signal.
• Second breaks this rotational symmetry. Its signal now contains spatial information, allowing it to “see” and orient itself.

The key insight is that with the same amount of information (same number of cells), breaking a symmetry in the input data creates an entirely new sensory mode.

The Core Conjecture

The brain acts as a structural prism, splitting information into distinct modes based on shared symmetries. A “visual group” has different invariant properties than an “auditory group.” Because these representations are learned in a disentangled way (i.e., they don’t share embeddings), the brain experiences them as separate phenomenal modes.

This applies even to our own senses. A bat uses acoustic waves, but its brain constructs such rich, spatial representations from echolocation that it is effectively “seeing” with sound. The mode is defined not by the physical medium (light vs. sound), but by the symmetries of the world model the brain constructs.

III. The Physics of Experience: Representations, Traces, and Forms

To go deeper, we need to distinguish between three concepts:

Definition

• Object: An abstract concept with no physical existence (e.g., the idea of a “face”).
• Representation: A learned map of an object space, like the manifold of faces learned by a neural network or the face-selective neurons in the primate brain
• Trace: The physical instantiation of a representation, the firing of neurons, the state of bits in a computer, a physical record that requires energy to create and leaves a mark on the universe.

Every logical bit has a physical existence. Storing it requires energy (bounded by Landauer’s limit) and creates a “trace” that resists thermalization for some period of time. However, most traces are passive and will eventually be eroded by entropy. The crucial step for creating stable, complex conscious experiences is feedback.

Form

A Form is a trace that is actively stabilized by a feedback controller. This controller burns energy to continuously counteract entropy and maintain the trace’s information content, constraining it to a specific manifold (e.g., the “face” manifold).

The brain is full of these feedback loops, from thalamocortical loops to the default mode network. These loops are constantly burning energy to maintain and update the representations that constitute our world model.

The Binding Hypothesis

Understanding how to draw a border around a Form is equivalent to solving the binding problem. The experience of a unified percept (e.g., a red cup) is bounded by the feedback controller that is expending energy to stabilize that specific bundle of information. Where the feedback controller’s influence ends, the phenomenal experience ends.

IV. The Architecture of a Moment (Personally the most interesting part)

This leads to the question of time. What is the “present”?

The Moment

A block of observation within which we cannot temporally order its constituent parts

We can distinguish between:

• Proper Time: Physical time, measured in seconds.
• Ego Time: Phenomenal time, measured in moments per second.

Where a shorter moment grants higher temporal resolving power (like a cat seeing a snake’s strike), but the moment itself is informationally less rich, and a longer moment allows for more complex informational structures to be perceived (e.g., longer wavelengths of sound can be mapped to a phenomenal pitch), but at the cost of temporal resolution.

This is tied to brain oscillations. Fast gamma oscillations (30-60 Hz) might handle local integration within a single sensory mode, while the slower alpha rhythm (~10 Hz) binds different modes together into the unified “moment” that we experience. This is a very interesting thought, which also connects to the different time experiences under locked in mode, psychedelics, boredom, meditation etc. Crucially, LLMs don’t have this kind of temporal structure, which may explain why they lack consciousness.

V. The Speculative Leap: A Field Theory of Qualia

This is where the talk moves from established neuroscience and information theory into more speculative territory.

1. The Premise: The dynamical system of neural activity in the brain can be described by the mathematical machinery required for a classical field theory.
2. The Link: The symmetries of this system are defined by the fiber bundle structure of qualia.
3. The Question: What happens if you take this classical field theory of brain dynamics and run it through canonical quantization?

The Hypothesis

Canonical quantization of this brain-specific field theory could yield a non-Poincaré invariant gauge field whose excitations are, in fact, qualia.

I got a bit stuck here. Some analogies:

• A gauge symmetry in the “consciousness field” would mean you could swap all my “red” experiences for your “blue” experiences (a “relabeling”) without changing the underlying structure of my color perception. The relationship between red, green, and blue would remain the same. This “freedom to relabel” is the gauge symmetry of the consciousness field.
• Poincaré invariance is a fundamental principle stating that the laws of physics are the same for every observer, no matter where they are in space or time, or how they are moving (as long as they aren’t accelerating) This is a radical idea. It proposes that qualia are not just emergent properties but are fundamental excitations of a physical field.
• Non-Poincaré invariant proposed consciousness field would violate this principle. Its laws would not be the same for everyone. They would be fundamentally perspectival and observer-dependent.

Under this theory, the “Is your red my red?” question gets a precise answer: Yes, up to a gauge transformation. The local gauge symmetry of this proposed field is the freedom to relabel qualia consistently. The underlying structure is the same, even if the specific “labels” (our internal representations) might differ. Then the underlying field isn’t about specific qualia like “red,” but about the fiber space itself—the space of all possible modes. When a specific type of brain (like ours) couples to this field, it causes a spontaneous symmetry breaking, where a generic “bit” of potential experience collapses into a specific “quail” of actual experience.

VI. The Engineering Endgame

If this theory holds any truth, it moves consciousness from the realm of philosophy to engineering:

• If consciousness is an ontologically real physical field, it might exert a “downward causation” force, influencing neural activity in a way that is energy-conservative but informationally significant.
• A validated theory would provide the equations to design conscious machines that can replicate our phenomenal content.
• Then we could potentially design systems that allow us to experience entirely new qualia—seeing in 11 dimensions, for instance—by coupling novel computational structures into our own phenomenal moment.
• If binding is determined by feedback control, could we create a network that integrates the feedback controllers of two separate brains? This would create a merged, co-experienced moment—a true brain-to-brain interface!

VII. Shorter Term: LLMs

Lets see how we can apply these ideas to improving the current version of AI (LLMs):

• They already have some forms of qualias: for example, the internal representations of concepts like “red” or “circle” are learned and stabilized through training and can be represernted in high-dimensional vector spaces (embeddings).
• But these embeddings are stuck in time, they are stopped during training, never updated and never integrated into a moment with feedback loops.
• Current attempts to fix it is memory via RAGs, but in practice they work badly as models don’t know what’s important and what’s useless. You can see it in your chatgpt when it suggests new things based on 1 year old ticket search results.
• What if we can somehow simulate different oscillations in the model, to create different “moments” that can integrate information over time, and bind different modes together? There were some attempts to do it via Oscillatory Neural Networks, via Spiking Neural Networks, or via Artificial Kuramoto Oscillatory Neurons, but they are not mainstream yet and do not work as well as transformers. But maybe we can create a better short term-long term memory system based on these ideas?