The below is mostly writen by Claude 4.5 Opus, with a lot of direction from me, after a long conversation with him about my personal theories for consciousness and checking them against what is described in literature. I think this serves as a useful summary of existing literature, and some concepts proposed here are I haven't written about before, and I think is a direction people should go. I'm an engineer, not a philospher, so I don't spend much time reading all the different things others have wrote and indexing it in my head. But I do spend a lot of time thinking about this in the background.

It's interesting that while Opus 4.5 is very fun to converse with on this stuff, and very smart in general, he cannot capture writing that is how I want to say something. And imo the below isn't nearly as intersting to read as if it were written by a good writer, even though the ideas in it are things I agree with and wish were said more. LLM's are very useful at organizing thoughts, but I still see no way around the hard work of writing out each word yourself if you want your writing to sound like you and be the way you want it. Very different than code.

Lastly, no I'm not someone claiming LLM's are concious. Becasue no one understands it.

We're Missing a Conceptual Primitive

Consciousness is the hardest problem. Not hard in the way that protein folding is hard or turbulence is hard—those are complex but tractable. Consciousness is hard because we can't even locate where the difficulty is.

I'm continually frustrated by how confused most of humanity seems to be about this. The show Westworld is a perfect example: it continuously conflates consciousness with "free will," treating them as the same mystery. They're not. Free will is a myth—an illusion we can dissolve with careful thinking about causation and the self. Sam Harris thinks and writes clearly on the subject. But conciousness, specifically the hard problem of consciousness is the real mystery. There's something it's like to be me reading these words, and no amount of philosophical dissolution makes that go away.

Every theory of consciousness I've encountered does one of two things: it either smuggles experience in at the beginning, assuming some form of awareness as a primitive, or it explains everything about information processing and behavior while leaving experience out entirely. You can describe the full causal chain from photon to neural spike to verbal report without ever invoking subjective experience. So why does it exist?

My view is we're in a pre-paradigm state. We're missing a conceptual primitive—some fundamental concept that would make consciousness legible the way energy made heat legible, or chemistry made biology legible. Before those conceptual breakthroughs, people invoked caloric fluid and élan vital. We're probably doing something similar with consciousness, using folk concepts that carve reality wrong.

If this is true, the answer will likely come sideways. Someone working on a completely different problem will stumble into a formalism that makes consciousness suddenly obvious in retrospect. We won't solve it by staring at it directly. We'll solve it by building the right conceptual infrastructure somewhere else. As David Deutsch puts it: problems are soluble. I refuse to accept that consciousness is irreducible.

Most Theories Are Stupid

I should caveat: maybe I'm misunderstanding something in some of these. If so, apologies. But having looked at the major theories of consciousness, I think most of them are just stupid. They either confuse reportability with experience, invoke undefined terms, or dissolve the problem in ways that create worse problems.

Higher-Order Theories Proponents claim a mental state is conscious when there's a higher-order representation of it—you're conscious of X when you have a thought about X. This does not make sense because it never gets rigorously defined. What makes something a representation "of" something else? You can define infinite mathematical mappings from one representation to another. When defenders reach for language—"it has to be a linguistic thought"—that's arbitrary. Language isn't fundamental. It's just one representation humans figured out among infinitely many possible representations. This isn't a theory. It's hand-waving that sounds sophisticated.

Global Workspace Theory, Attention Schema Theory, Predictive Processing—these all describe functional architecture. What gets broadcast where, what models what. Fine as neuroscience. But they're theories of access and reportability, not experience. They tell you which information becomes available for reasoning and verbal report. They don't touch why availability should feel like anything. They're answering a different question and pretending it's the same one.

Orchestrated Objective Reduction (Penrose-Hameroff) at least tries to ground consciousness in fundamental physics, which I respect. But it probably gets the physics wrong—decoherence timescales in warm biological systems are way too fast for quantum coherence to matter in the way they propose.

Illusionism (Keith Frankish, Daniel Dennett sort of) says consciousness as we conceive it doesn't exist—there's no hard problem because there's no intrinsic phenomenal quality, just a powerful illusion. Here's why this doesn't work for me: if you really believe it, morality goes away. Suffering only matters if there's something it's like to suffer. If experience is illusory, then there's nothing actually wrong with endless torture—it just seems bad to the system being tortured, but that seeming is itself an illusion, so who cares? Illusionists don't actually act as if this is true. They still avoid suffering, still care about wellbeing. That's a contradiction. If you have a solution to this contraction, I will listen, but I have not seen one.

Panpsychism I actually have respect for. The idea that experience might be fundamental—like mass or charge—rather than emergent is genuinely interesting. If you can't explain how consciousness emerges from non-conscious stuff, maybe it doesn't. Maybe some form of experience goes all the way down. This is incomplete, and it has the combination problem (how do micro-experiences combine into unified macro-experience?), but it's incomplete in a productive way. It's pointing at something real. Good incomplete work that brings us somewhere.

IIT: The Most Serious Attempt

Integrated Information Theory, developed by Giulio Tononi, is different. It actually provides math.

The core idea: consciousness is integrated information. Not just any information—specifically, information about a system's state transitions that exists only in the whole and is lost under any partition into independent parts.

This is formalized as phi (Φ). To compute phi, you ask: if I cut this system into its most independent pieces, how much information about state transitions is lost? High phi means the system is genuinely unified—you can't decompose it into independent parts without losing something.

A toy example. Consider two binary units that update according to XOR:

The state transitions are:

Now ask: if I only know A, can I predict A's future? No. A=0 could become 0 or 1 depending on B. Same for B alone. All the information about transitions lives in the joint state, not in either part.

The mutual information I(A_past; A_future) = 0. Knowing A tells you nothing about A's future. But I(X_past; X_future) > 0 for the whole system. The difference is phi.

IIT makes a bold claim: phi isn't just correlated with consciousness. Phi is consciousness. Any system with positive phi has some experience, and more phi means richer experience.

This has counterintuitive implications. A carefully designed grid of XOR gates with feedback could theoretically have higher phi than a human brain—because the brain is modular (you can partition it without losing that much), while the XOR grid might be maximally integrated.

Despite the strange implications, I find this direction compelling. It's actually trying to say something precise.

The Noise Problem

But there's a problem.

Think about what happens at the bit level. Random noise has maximal interdependence. Every bit influences every other bit in ways that are maximally hard to predict from subsets. You can't compress the transition function because there's no structure to exploit. By a naive reading of phi, this is maximal integration—any partition loses information because the information is spread uniformly everywhere.

Now consider a neural network before and after training. Untrained, the weights are random. The dependencies between activations are noise. Trained, the weights encode meaningful relationships—specific neurons fire for specific features, specific attention heads specialize.

Counterintuitively, training might lower phi. Structure means redundancy. Redundancy means compressibility. Compressibility means you can find partitions that don't lose as much. The trained network has developed modularity—useful, functional modularity, but modularity nonetheless.

So you get the absurd implication: a random network has higher phi than a trained one, and maximum phi is achieved by maximum noise.

This is clearly wrong. Random systems aren't experiencing more than structured ones. They aren't experiencing anything interesting at all. There's no structure to experience.

Integration alone isn't sufficient.

The Missing Piece: Compression

Here's what I think is missing: phi measures interdependence but not structure or meaning of that interdependence. It treats all information equally, whether it's organized into coherent representations or just noise.

What we want is something like: integrated information that's also structured. Information that hangs together in ways that support modeling or representation. Not just mutual dependence, but dependence that does something—that compresses, that abstracts, that's about something.

To compress is to find structure. It's to say "these things are the same" and "these things are different." It's carving the world into categories. A random system doesn't carve—it just maps noise to noise.

When you compress, you're modeling. You're saying "this high-dimensional input is actually low-dimensional in the ways that matter." And "the ways that matter" implies a perspective. The compression is relative to something—to predicting, acting, surviving. There's a point of view built into the act of compression.

So maybe consciousness isn't integrated information. It's integrated compression. A unified model of something. A perspective that carves the world.

This would explain why random noise doesn't feel like anything even if it has high phi. There's no model there. No compression. Nothing it's about.

The unity of experience might be the unity of the compressed representation. "There's a cup on a table" is a compression of millions of photoreceptor activations. The visual field feels unified because the parts are jointly compressed into a coherent scene representation—a single low-dimensional structure that captures the relevant features of a high-dimensional input.

The hard question remains: why does having a model feel like anything? But at least this points at a tighter target than just "integration."

Nested Scales and Boundaries

If integration + compression is what matters, there's no privileged scale.

Consider the brain. There's no central neuron, no Cartesian theater where everything comes together for a homunculus. The "you" emerges from distributed interaction. But within that distributed system, there might be multiple integrative processes, multiple compressed models, multiple perspectives.

The "self" we identify with might just be whichever process has access to memory encoding and motor output. That doesn't make it the only experiential process happening. It's the one that gets to write the autobiography.

Temporal continuity supports this. We assume the self persists because it feels continuous. But you lose consciousness every night. Anesthesia creates a hard break—no experienced continuity, just a gap. The feeling of continuity is generated moment to moment, not carried forward by a persistent substrate. If you were disassembled and perfectly reconstructed, the reconstruction would feel continuous with the original. There was no continuous thing—just a pattern that got reinstantiated.

This opens stranger possibilities. If groups of minds interact with high-bandwidth information exchange, could there be experience at the collective scale? The communication between humans is much lower bandwidth than synaptic transmission, but it's not zero. Language, culture, institutions—these are integration mechanisms operating on longer timescales.

Maybe a humanity-level experience operates on timescales of years or centuries, the way individual experience operates on hundreds of milliseconds. From its perspective (if it has one), individual human thoughts might be like individual ion channels firing—noise beneath the resolution of its experience.

What would it experience? Maybe something like the texture of an era. The felt quality of civilizational mood shifts. Something we participate in but can't access as a gestalt.

Completely unfalsifiable. But structurally consistent with the framework.

The Hard Problem Remains

Even with better concepts—integration, compression, modeling, perspective—we still don't know why having a model should feel like anything.

But I don't accept that it's irreducible. Problems are inevitable, problems are soluble. The path forward is building these intermediate concepts rigorously enough that either the gap closes and we see why experience follows, or we find a genuinely irreducible remainder that we can point at precisely.

Either would be progress. Right now we can't even locate what's mysterious. We just gesture at qualia and say "why does this exist?" Better concepts would let us say exactly what needs explaining.

The analogy is biology before chemistry. Vitalists thought there was some irreducible life-force because they couldn't see how to get from physics to living systems. But once you have the right intermediate concepts—molecules, reactions, information, replication—"life" stops being mysterious. The mystery wasn't solved so much as dissolved by better concepts.

Consciousness might work the same way. We have "physical process" on one side and "experience" on the other, and no bridge. But maybe that's because we're missing the concepts that would constitute the bridge. Integration-with-compression-relative-to-a-boundary, or whatever the right formulation turns out to be.