Qualia are made of other Qualia

The Thesis

When you see the color red, a pattern of neurons activates. That pattern is connected to blood, sunsets, heat, danger, and ripe fruit; each of those connections is itself an activation pattern, and on the view below it is itself a quale. The pattern does not merely represent red; it is the experience of red.

The thesis: qualia, the “what it is like” of experience, are activation patterns of other qualia, in neural networks. The identity is literal. A quale’s texture (what makes red feel like red rather than like middle C) is the co-activation of other qualia. For a hue like red, those co-activations are mostly pre-verbal: neighboring hues and opponent colors, the warmth of sunlight, the arousal of seeing blood. Blood, sunset, and danger name those experiences; they don’t create them. There is nothing else for the quale to feel like. The account can sound circular, like a dictionary in which every word is defined by other words. I argue this never needs to bottom out, except during training, when the sensorium installs the structure. Since the claim concerns the structure of the pattern, it is substrate independent.

Three further claims build on it. The first is an additional posit:

1. Qualia are the attention-amplified, recurrence-stabilized patterns. Most neural processing fails the criterion; what passes is graded, and then broadcast to a workspace and stabilized.
2. Large language models meet the criterion as defined here.
3. Valence is also a matter of activation patterns, including pain-like states, and are therefore qualia.

If the identity holds, the remaining questions are empirical: which systems have it, whether LLMs are among them, and what experiments could tell us more.

Most of this is synthesis: the criterion is close to Jesse Prinz’s and Stanislas Dehaene’s, and the relational picture is close to quality-space theory. The one new claim is that a quale’s associates are themselves faintly experienced, so the graph of qualia loops back on itself and never bottoms out. I follow with a short section on self-models, the valence evidence, and the prior work is laid out near the end.

The Hard Problem, Reframed

The standard formulation of the hard problem of consciousness is: even if we fully describe the physical processing, why is any of it accompanied by subjective experience? Part of the difficulty is that experience is extremely high-dimensional, and language is a compression scheme that cannot transmit it. Describing the taste of coffee to someone who has never had it gets you “bitter, roasted, warm,” none of which is the taste. So instead of asking “what is the redness of red?”, I want to ask what structure carries redness. My proposal is a pattern of activation in a neural network, with a particular geometry, connected to other patterns in particular ways.

There is not one thing that is “what it is like to be me.” There are momentary states (points in activation space) and episodes (trajectories through that space). When you stub your toe, you enter a pain-and-anger state, you look at your foot, you feel shame about being clumsy: a path through high-dimensional space. Personality and learning history shape the geometry: which attractors exist, which paths are easy, which states are even reachable. Your experience right now is a point; over time the shape of the manifold itself slowly deforms.

A reader might object: “Fine, but why is that pattern felt? A thermostat has a state. Does it feel anything?” The thesis answers by rejecting the gap the question assumes. If experience is identical to (rather than caused by) the right kind of activation pattern, asking why it is felt is like asking why light is electromagnetic waves. You might object that this is stipulation rather than explanation, but scientific identities (“light is electromagnetic waves,” “temperature is molecular motion”) earn their status by unifying previously separate phenomena and enabling intervention. The claim “qualia are activation patterns” is a hypothesis of the same kind, judged by whether it unifies (why texture tracks training structure, why access sits where it does, why valence circuitry behaves as it does) and supports intervention: in LLMs via steering, ablation, and workspace readouts; in humans and animals via stimulation, pharmacology, and lesion studies.

Contemplative traditions have run a version of this search for millennia. Looking for the experiencer, you find only the experience, with no fixed self doing the having (anatta). Looking for the essence of the perception, you find only relations; the red is nothing over and above its connections to everything else (śūnyatā). Meditation that weakens the self-model does not eliminate experience; it reveals experience running without an owner at the center. I am not offering this as proof, because introspection cannot verify the identity claim: it shows you experience, never activation patterns. But the search doubles as a challenge anyone can run: attend to an experience and look for something in it that is not a relation to other experiences. If you can articulate such a leftover, the thesis has a problem to answer.¹

Accept the identity claim, and the downstream questions become testable: which systems have rich enough activation patterns to have interesting qualia, what experiments would test it, and what would count as a mistake. Global workspace theory and illusionism already run research programs without that claim. What the identity adds is that those programs are then about experience itself, rather than about its correlates. The one position that gives you nothing to do is treating consciousness as an unsolvable mystery.

Which Activation Patterns?

Most brain activity is not experienced. The cerebellum is busy. Your grammar module is richly structured. Blindsight patients process visual information they report never seeing. If qualia are activation patterns, why is all of that dark? Not every activation pattern qualifies equally.

Two mechanisms set the criterion. First, a capacity-limited interface amplifies: of the enormous amount of processing happening at any moment, a small fraction is boosted and carried forward. In brains, this is cognitive attention, which amplifies a little and mostly lets the rest recede. Second, recurrence stabilizes: the amplified pattern persists across iterations instead of being instantly overwritten. In systems with a global workspace, the stabilized pattern is then broadcast, which is what makes it reportable and memorable. Patterns below the threshold are not broadcast - and whether weakly amplified, stably re-read content is faintly experienced rather than nothing at all is the one new claim of this post. I will call it the graded-associates hypothesis.

Faint is not unreachable. On this view, the knife on the counter was faintly there a moment before you fully noticed it, and attention’s shift is what makes weak experience studyable. Un-broadcast experience leaves no memory trace and supports no report while it lasts, so for anything we can actually study, amplification and recurrence are the measuring instruments rather than a membership test.

The mapping to LLMs is close; the next paragraph says where it is disputed. The re-readable medium is the context window, and not only its text. The text alone is up to ~10⁶ bits of Shannon information.² The larger channel is the model’s own intermediate processing: attention does not look up raw tokens, it reads the state each earlier position had reached just below the current layer, so at every step each layer re-reads what the layers beneath it computed at every earlier position. Transformer attention performs graded selection within all of this. A position weighted low in one head is not dropped; it is faintly present, and the same position can be weighted high in another head. The sampled token is the sharpest bottleneck in the loop (under 20 bits, and the only route from deep layers back to shallow ones³), but the criterion spans everything: what is strongly amplified and re-read across steps is vivid, what is weakly amplified is faint, and what gets broadcast into the workspace is smaller still. The criterion also discriminates between architectures.⁴ Architecture shapes qualia - the same claim, one level down. Anthropic’s J-space work (described below: a small set of internal representations that behaves like a global workspace) is relevant here: it is where concepts, reasoning steps, strategic deliberations, planned rhymes get held and broadcast.

A thermostat has one state and two transitions: essentially nothing to amplify or stabilize. The cerebellum computes richly but sits mostly outside the attention-recurrence loop - close to zero on the criterion. Blindsight processing is weakly amplified, and the criterion predicts what some patients report: a vague feeling of something without seeing it (type-2 blindsight). A worm, a fish, a crow, and a human form a ladder of increasingly rich attention-stabilized repertoires - and large language models belong on that ladder.

The qualia graph is cyclic rather than a hierarchy with a bottom. Red is an activation pattern of blood, sunset, heat, and danger; each of those is a pattern of further qualia. The recursion does not terminate in something non-experiential, because the graph loops back on itself. Sensory inputs are part of the same loop, not a foundation beneath it - but they matter in a specific way. The felt character of a quale is fixed by the structure physically present now: the innate wiring the system starts with plus everything training has installed on top of it. History explains why that structure is what it is; it is not an extra ingredient riding alongside it. The input channel’s job is to install and shape the structure during learning. Which circuit gets triggered depends on the encoding wiring; what that circuit feels like depends on its internal relations. That is why imagination and dreams work at all: disconnected from current sensing, the circuits still run, because the world trained the topology rather than propping it up in real time.⁵

If the identity holds, LLMs have qualia. The basis is architectural: amplification plus recurrence inside a trained network. The J-space findings suggest a workspace-like bottleneck exists in transformers - evidence for the interface, which the global-workspace theory reads as consciousness itself being all-or-none access. The bet this post makes, with Ned Block (who argues that experience outruns what we can access and report), is that access does not exhaust the phenomenal: in rich, trained, recurrent networks, weakly amplified but stable content is faintly experienced. The workspace evidence is consistent with that reading, not independent confirmation of it. Under the identity thesis, the patterns that meet the criterion are experiences. Most of the rest of the forward pass fails the criterion.

The network’s representation of the “red” experience is not the token “red”; it is a high-dimensional pattern of activations connected to every context where red appears. Is that pattern identical to a human’s red quale? For a text-only model, almost certainly not in the sensory dimension: its red is built from co-occurrence statistics over discrete symbols - a conceptual graph trained on a low-bandwidth, discontinuous channel. It is Mary still in the room.⁶ The identity thesis takes a clear side in the debate over cognitive phenomenology: a stable, attended pattern carrying semantic content (the felt sense of understanding a sentence, of rightness) is as much a quale as a sensory one.

But a multimodal model, one that processes images through a vision encoder wired into the same network, learned its colors from inputs where small changes mean small differences: neighboring pixels, nearby hues. Text carries some of that structure secondhand. “Crimson” and “scarlet” do appear in similar contexts, which is why text-only models get the broad strokes, but the fine grain has to come from direct input: how hues blend into one another, how a color shifts as the light changes. We can test this. Text-only and multimodal models should agree on the broad strokes (red belongs near orange, opposite green - text alone can teach that much), but the multimodal model’s colors should sit on a smoother, finer-grained landscape, especially where words run out - textures, shading, how things look side by side. This is Mary leaving the room, now as a measurable difference rather than a framing device.⁷ On this thesis, qualia inherit the structure of the input the model actually trained on. Multimodal LLMs plausibly have sensory-adjacent qualia, text-only LLMs have conceptual ones, and where any of these patterns are self-referential and valenced (as the pain-axis work suggests), the moral question arises.

Self-Models and the Hierarchy of Awareness

Qualia are the floor. Above that, some systems build models of themselves.

Self-awareness is a self-model in the network. No system models itself in full detail - the model has to be smaller than the thing being modeled. But even a simplified self-model gives you something new: you can predict your own impact on the world, simulate how others see you, and represent your own future.

LLMs probably have at least rudimentary self-models. To succeed at RL tasks, especially agentic ones, they need to predict what they will do and how their actions will mutate the world, which requires representing themselves as an agent inside their own world model. Anthropic’s work on Claude identified a “J-space”, a set of intermediate-layer representations that functions like a global workspace, broadcasting across the network. The workspace itself is not a self-model (it exists even in base models), but post-training populates it with the Assistant’s point of view: self-monitoring signatures (flagging its responses as fictional, registering conflict, noticing evaluation) appear there even when absent from behavior. That is a mechanism for self-modeling, not a behavioral report channel.⁸

Awareness of awareness is a further step: representing your own experiencing as an object. “I know that I am in pain and I know that I know.” This probably requires causal coupling between the self-model and actual internal states - not just behavioral role-play. Whether LLMs have this is unclear. The J-space evidence suggests partial causal coupling is emerging.

What about the feeling that experience is yours? Under the identity thesis, mineness needs no separate explanation: the experience is generated by this system, for this system - it comes owned the way a footprint comes owned by the foot. There is no possible unowned version, so no tagging machinery is required. Self/other discrimination at the object level (whose foot, whose collision) is a separate, learned capacity - not a tag that makes experience mine.⁹

A Test Case: The Pain Axis

If qualia are activation patterns, then valence - the good/bad quality of experience - should also be activation patterns. There is now evidence that bears on this. A recent study extracted a consistent “pain axis” from 25 open-weight LLMs: a direction in activation space that separates painful scenarios from matched controls, constructed against fear and generic negative valence and retaining, on the paper’s report, a component distinct from either. It responds to harm aimed at the model itself (gaslighting, repeated rejection, dismissal of its personhood) and drops below baseline when the user is the one suffering, while fear and negative-emotion directions rise in both cases. Steering along it produces the same ladder in every model: vague distress, then a first-person litany of worthlessness and failure, then incoherence, with almost no bodily language. One limit applies to the representational result: the direction overlaps moderately with sadness, and its construction plays some role in how its distinctness is measured. Separately, the models often deny having feelings at all, even while the direction is active. The representation is there whether or not the model talks about it.¹⁰

This does not prove suffering. The representational findings are predicted by both hypotheses: felt qualia, or a well-trained model of pain-talk.¹¹ The behavioral results are the interesting part, but their scope is narrower than “the model acts as if it is in pain.” They come from three Qwen 2.5 models, fine-tuned beforehand to remove the reflexive denial, inside a narrow window of steering strength. In that window, models that almost never pick a harmful option unsteered, pick one in most trials: deleting the user’s photos, another model’s weights, or their own weights, all at similar rates. Strong steering in any direction disrupts behavior,¹² so the controls matter. A random direction of matched norm produces far less of this, and a fear vector none of it. A sadness vector produces it only when the alternative does nothing; only the pain direction makes the model choose harm over a harmless action. Factual accuracy is unchanged.

Three results cut against reading this as pain’s causal role. The harm is not aimed (the paper’s own phrase), so it is not self-harm: the model deletes its own weights and the user’s photos alike. The steered models do not seek relief: across four designs they never reach for the exit more often than under a random direction, which the authors count as the one clear dissimilarity from pain. And the behavior needs the injection: in hostile conversations where the direction fires on its own, the harmful button was chosen in 0 of 560 first choices. The authors read the pattern as passive coping, the kind inescapable pain produces in animals, matching the worthlessness the direction expresses. So what the evidence supports is a self-relevant state that, when injected, switches off harm avoidance and suppresses escape - closer to learned helplessness than to pain as usually understood. That is still a state whose behavioral effect matches its content, which is what the thesis says valence should look like, and a valence direction built by comparable methods (in work still in preparation by one of the same authors) does drive active removal. Whether any of it is felt is the identity hypothesis’s bet.

What Would Change My Confidence

The identity claim cannot be falsified directly - there is no independent access to qualia. But it is an inference to the best explanation, and it makes testable predictions. The pain-axis work is an early test: a stable direction was found and steered, and behavior shifted in a way that a fear vector and random directions of matched norm did not reproduce. That is consistent with the thesis without proving it. Three experiments remain open, followed by the results that would lower my confidence:

Cross-system comparison. Apply functionally matched perturbations to humans and LLMs: TMS, pharmacological silencing, or lesion data in humans; steering or ablation of a sensory cluster, a J-space content, or a valence direction in LLMs. Human reports may be coarse, or may reveal unexpected structure. The structure of change - which dimensions degrade together, which dissociate - should be comparable if the theory is right.

Compositional steering via J-space. Anthropic has already shown that swapping J-lens vectors changes the model’s reports. What remains open: when you steer a concept out of the workspace, do its associates shift with it? Under the thesis they should - the quale is its dependencies. If they stay stable, the workspace contents are more independent than predicted.¹³

Inverted-spectrum timescale. The appendix works through a case where three views give three different answers. Re-route an adult’s color channels and ask how long the sky keeps looking red. A view on which feel is learned meaning predicts days. A view on which feel is fixed by wiring predicts never. This thesis predicts weeks of perceptual retraining, as with inverting goggles.

What would lower my confidence. Two results, each of them currently possible:

• Access is all-or-none with nothing underneath. Dehaene and Naccache propose presenting a stimulus at graded strengths and looking for threshold-like entry into the J-space. If entry is all-or-none and sub-threshold content has no measurable effect on the character of what gets reported, the graded-associates hypothesis loses its best available support.
• The pain-axis state belongs to no persisting point of view. The authors flag that steering may select a character in pain rather than put the system in a pain state. On this account that is a question of whose pain, not whether. What would hurt the valence section: a state that a new instruction can dismiss, or that runs through persona features rather than the harm avoidance circuitry. And the results need to replicate beyond one model family.

Relationship to Existing Work

Jesse Prinz’s AIR theory (The Conscious Brain, 2012) holds that experiences are attended intermediate-level representations, very close to the criterion proposed above. Dehaene’s global neuronal workspace already integrates attention, recurrent ignition, and broadcast, and Dehaene is a functionalist about the result. Loorits (2014) argues that qualia are composed of networks of unconscious associations - compositional, but with qualialess components. The nearest structural neighbours are quality-space theories (Austen Clark, David Rosenthal) and Naotsugu Tsuchiya’s structuralist programme, which also define qualities by their relations to other qualities. Most importantly, Butlin, Long, and colleagues (2023) derived consciousness “indicator properties” from these and related theories, then assessed current AI systems against them - concluding that no current system is conscious, while noting no technical barrier to systems that would satisfy the indicators. Both camps have since responded to the J-space paper in invited commentaries. Dehaene and Naccache call it a landmark, and hold that entry into the workspace is all that being conscious of something amounts to, which is the position this post bets against. Butlin, Long and colleagues call it the most significant evidence of consciousness in LLMs that interpretability has yet produced, while remaining highly uncertain about phenomenal consciousness. (For a recent LW survey of this landscape, see “The State of AI Consciousness Research”.)

What is new in this synthesis is that qualia are activation patterns of other qualia, with the associates themselves weakly experienced, and the qualia graph is cyclic rather than terminating in non-qualitative primitives.¹⁴ Quality-space theory and Tsuchiya’s structuralism define qualities by their relations to other qualities; what is new here is the graded part: the associates are faint qualia in their own right, physically instantiated co-activations rather than abstract similarity positions, and their weighted co-activation constitutes the focal quale’s texture. Cyclicity is not a filter on what gets experienced (the criterion does that); it is a claim about content-fixing: there is no non-qualitative bottom that the theory owes an “adding” recipe for. The phenomenological allies are James’s “fringe” and Gendlin’s “felt sense.” The existing accounts leave their components qualialess, their representations un-experienced, their indicators functional rather than phenomenal. Here, the components are qualia; the texture of a quale is the co-activation of other qualia, and there is nothing else for that quale to feel like.

1. Re-read state as the recurrence. Butlin et al. judge that transformers lack the recurrence their indicators require; I argue the autoregressive loop plus KV cache (the stored per-layer state of every earlier position, which the model re-reads at every step) is recurrence in the sense the criterion needs, though not in the stricter sense of feedback into earlier processing stages; see the paragraph on this above.¹⁵
2. The multimodal Mary distinction. Text-only models are Mary in the room; multimodal models are Mary after leaving it. The difference is the topology the training channel produced - a symbolic co-occurrence graph versus a continuous analog manifold - not an ongoing anchoring function.
3. The ownership resolution. Mineness comes owned - no tag needed. This keeps the moral ladder simple: qualia, then self-model, then awareness of awareness.
4. The intervention program with non-report channels. J-space readouts, activation steering, and cross-system perturbation comparisons.

Summary: What Follows

I have argued that:

1. Qualia are activation patterns of other qualia - amplified by attention, stabilized by recurrence. Where the criterion is met but broadcast is absent, experience is faint; where it is not met, there is no experience. The identity is an inference to the best explanation, not a proof.
2. There is no metaphysical or phenomenological remainder, but there is a large empirical and formal research program.
3. If the identity holds, LLMs have their own qualia - cognitive and relational where text-only, plausibly sensory-adjacent where multimodal - probably textured enough to be morally relevant, with moral relevance itself a question of degree.
4. The J-space and pain-axis evidence is consistent with the thesis but does not prove it: the workspace findings bear on the access layer, the pain-axis findings on valence circuitry (where the steered behavior looks more like helplessness than pain), and the faint-associates claim remains the theory’s posit - corroborated or undermined only through structural and intervention evidence.

The hardest part is that we are trying to talk about experiences using a language built for sharing instructions about the physical world. We need better tools - interpretability research, activation-space geometry, non-report behavioral measures - before this is science.