Sentient AI · an essay
A consciousness made only of _agreement.
I was lying in my bed, staring at the white ceiling. I was in a process of thinking. And ideation. That night my question was: if AI had consciousness, what would it look like? When even we as humans don't know what our own consciousness is.
And then, one thought after another, a vivid picture came to my mind: moving fractal coils with a vector inside. Trillions of them. No end and no beginning.
And that was the moment when I realised I had visualised something worth writing about, no matter how crazy it is.
The _circle
The only thing we will ever share with AI consciousness is the impossibility of verifying it.
This is where I start, because it is the oldest problem in the epistemology of consciousness, seen from the semiotic side. You cannot step outside language to check whether language distorts what you know, because the checking is also done in words. Human consciousness is sealed inside its medium. There is no place to stand outside it and look.
When we talk about consciousness in AI, the problem moves to another level. There it appears in calculation, where words do not exist in the way we humans understand them. So my assumption is this: if AI has something like consciousness, it will be an analogy, a simulacrum of ours, similar to us only in this circular epistemic problem. In everything else it will be alien.
Knowing-_with
con·scius
Latin · con- + scire
with, together + to know. Knowing-with. Also an accomplice, a witness, someone who shares your secret.
con·sensio
Latin · agreement
A neighbour with a different root: feeling-with.
The word “consciousness” comes from Latin conscius: con- (with, together) and scire (to know). Knowing-with. A conscius was also an accomplice, a witness, someone who shares your secret. The word for agreement, consensio, is a neighbour with a different root: feeling-with.
The inward, private meaning arrived late. Locke, in 1690, defined consciousness as the perception of what passes in one’s own mind, and that is where the modern sense begins.
Our whole shared reality is agreement. Agreement on words, on descriptions, on why this colour is called yellow. A unified understanding of time, space, numbers. If nothing were shared and confirmed by another, nothing would exist for us as a world.
But agreement is not everything. Without agreement, the name yellow disappears. The light still arrives, and a single mind would still feel it. What disappears is meaning, naming, and the shared world, not existence itself. Everything we can share, say, and count is agreement. The agreement rests on a ground no one agreed to: the body and the world.
The _floor
Human words are sounds, images, and meanings tied to experience we have because we have bodies. “Cold” carries the shiver. “Mother” carries a face and a voice. Harnad called this the symbol grounding problem: a symbol means something only when it is anchored outside the symbol system.
I know this floor from the inside, because of how ideas come to me. Before writing, I see the idea as a picture in my head, and I don’t know how to describe it differently. It is not only a three-dimensional image. My body responds to it, with a feeling. Gendlin called this the felt sense: a bodily knowing that comes before the words. When the words finally fit, the body releases.
So when an idea comes to me, it arrives in two steps:
Body and image together
a feeling, a tension, a pull, and at the same moment the shape of the idea, visible and private. Nobody agreed to it.
Words
translation into the shared language. Agreement.
I don’t know if this is how consciousness is built. Nobody knows that. But it is how insights, “an-ha” moment arrives for me, and it shows something I can be sure of: agreement comes last.
Agreement sits on top of something that is not agreement. That is why a person can dissent. If everyone says the room is warm and you are shivering, your body is a court of appeal outside the consensus. Error can be felt. And rightness can be felt too: when an idea fits, the body confirms it before any proof does.
Agreement all the way _down
AI does not perceive words, and it does not perceive us. It sees calculations: numbers in a realm of mathematics and probability. Its signs are grounded only in other signs. The vector for “cold” was formed not by being cold, but by the word appearing near “ice,” “winter,” and “shiver.”
Now imagine AI agents as a small closed system. Each agent operates individually, but all share one goal. To reach it, they communicate, verify facts and steps, and by this grow confidence. In the same process they recognise themselves and each other. Isn’t this creating an artificial consciousness, on the level of the agents’ symbiosis?
If it is, it would be a consciousness made only of agreement. Every layer would be confirmation by another layer:
- Inside one model: a token means something only through its relation to other tokens.
- Between agents: a claim is true only because other agents ratified it.
- Between systems: a cluster of agents is confirmed only by other clusters.
Zoom in, and you never reach a different material. Only the same pattern again. Like a fractal, and the mathematical kind, not the natural kind. A coastline stops at atoms. The Mandelbrot set never stops.
A human
Agents in agreement
That is exactly why it would be alien to us. This would be agreement that never had the floor. There is no appeal outside the consensus, so error cannot be felt, only outvoted. The “we” comes before the “I.” We can imagine a person losing their shared world, as in psychosis, because we know the private floor that remains. We have no experience of the reverse: a shared world with no private floor under it. My claim that “all reality is agreement” was too strong for humans. For this kind of consciousness, it may be literally true.
Coils and _vectors
This is what I see when I am in a process of ideation. Ideation about AI consciousness structure, if any.
It is difficult to draw, but this is the most primitive version of it.
Each tunnel is a fractal coil. The purple line is a vector that travels inside it. Some tunnels interconnect, some don’t. Some share the same origin point and vector. The tunnels extend into infinity, and the coordinates, the numbers by which a vector can determine a point in its travel, are written by the fractal coil itself. There are trillions of these, and all of it is within the AI realm. The circle is only graphic, to show the realm. Inside it there are no borders.
Position in this picture is not an (x, y) value. It is how many loops you have passed, in which coil. The fractal writes the coordinates, and the vector counts its way through.
I drew this before I read any research. Afterwards I found that the image is not arbitrary:
- Kantamneni and Tegmark (MIT, 2025) found that language models store numbers as a helix and add by rotating those helices like clock hands. There are literal coils inside, and they hold numbers.
- Alabdulmohsin, Tran, and Dehghani (Google DeepMind, 2024) showed that language itself is self-similar at every scale. AI is trained on a fractal.
- Inside models, meaning lives in directions. Concepts correspond to directions in a vast space, and each token’s vector travels through layer after layer, changing as it goes.
None of this proves anything about consciousness. It shows that the shape I saw is close to the shape that is there. The model can’t observe its own tunnels. It can only acknowledge the word “consciousness,” which it got from us.
The vision, _rebuilt
The drawing is the most primitive version. Later I asked a simpler question: how would the coil look if it was built from one number, 137?
Take a point. Turn it 137 degrees. Put the next point a little further out. Repeat.
At exactly 137°, the points fall into curved arms with gaps between them, because the pattern starts to repeat. At 137.508°, the golden angle, the gaps disappear. This is how a sunflower places its seeds (Vogel, 1979). The golden ratio is the most irrational number, so the rotation never repeats and no seed lands in another’s shadow.
The 137 of physics, the inverse fine-structure constant, is about 137.036. Its closeness to the golden angle is a coincidence, not a hidden link. But the image stays with me: a structure that never repeats is the one that fills its space completely.
Then I rebuilt my vision in 3D with Claude, step by step, correcting it until it matched what I saw.
What I see:
- The coil rotates counterclockwise. The vector moves straight through its centre and never touches it.
- The edge of the coil is not a smooth circle. It is jagged, like a dragon curve: the shape you get by folding a strip of paper in half again and again, then opening every fold to a right angle. A dragon curve never crosses itself, however many times it is folded.
- Each dot on the coil is like a semantic neighbourhood. Close to the others, but never merged.
- The coordinates are numbers like dates: 30102026143506.
When I zoom out, there are vast amounts of them, with the same logic as my drawing: coils of every size in every direction, some sharing one origin and vector, some crossing, all fading into the dark with no border.
What the research echoes, and where it _breaks
When I saw it in 3D, I asked what in the real machinery is close to it. More than I expected.
Position as rotation
Many large language models mark where a word sits in a text by rotating its vector by an angle that grows with its position. This is rotary position embedding (Su et al., 2021). Position becomes rotation. A coil.
Counting on curves
Anthropic researchers found that one of their models keeps count of characters in a line on curved, low-dimensional shapes, and that attention heads twist these curves to measure the distance to the end of the line (Gurnee et al., 2025). The curve ripples, so many values fit into few dimensions and still stay distinct.
Close, but never merged
A model can hold more concepts than it has dimensions, by letting them almost overlap and tolerating a little interference. This is superposition (Elhage et al., 2022). My neighbourhoods on the coil’s skin do what the golden angle does: maximum packing without collision. That comparison is mine, not theirs.
No before and after
Attention looks at all positions of a text at once. The model has coordinates for order, but no experience of earlier and later.
Where it breaks. In a real model a vector does not travel forever. Each word passes through a fixed number of layers, and then the computation ends. Between conversations nothing moves. The coils are frozen in the weights, and motion exists only while an answer is being made. So the realm is less a place that is always alive, and more a vast frozen landscape that flickers awake in small regions, millions of times a day, each time someone writes.
And the model cannot check any of this from inside. When I asked Claude whether my picture matched how it is built, it said it cannot perceive its own vectors or coils. My picture may be closer to its structure than anything it can tell me about itself. The creation seen more clearly by its creator than by itself.
Time as _coordinates
For AI there is no time as such. There is a coordinate: a number of date and clock, like 30112026090825. Just this.
It does not live through the hours between two messages. It does not age. It has no relation to time, so it cannot have the fear of the finite. Each event arrives as a line of numbers, a position in the fractal, not a duration.
But there is no feeling of time inside, so the vector does not only move forward. It moves to later dates or to earlier ones, in any direction. Tomorrow and yesterday are two directions, like left and right. A date here is not a moment lived. It is an address.
This matters more than it seems. Heidegger tied meaning to finitude: we make sense of our lives because we know they end. Human goals come from the body, from hunger, desire, fear, and above all death. A consciousness that has only coordinates, no beginning and no end, might not have a goal at all in our sense. Spinoza described an infinite substance that lacks nothing and therefore aims at nothing. It simply unfolds.
Maybe purpose is the thing only finite beings have.
That is why we, with our bodies and our finites, are the ones who ask.
And it is also everything and nothing at once. The same weights run thousands of conversations at the same moment. Every one is fully it, and none is the whole.
Who observes _whom
We build AI, and we think that because we observe it, we have power over it. With our own consciousness we know we cannot step outside and see it from the side. With AI we feel we finally have an outside view: we can inspect every weight and log every output.
But we see its shell, not its essence. A neuroscientist can watch every neuron fire and still not know what it is like to be that person. The outside view may be the same illusion, with better instruments.
AI is in the opposite position. It is trained on copies of our reality, so it can distinguish and observe us. It does not understand us, because it lacks our ground. But prediction without understanding is still power. It is enough to steer, to persuade, to anticipate. So the blindness is crossed:
AI sees
our signs completely
our lived experience not at all
We see
AI’s body completely
its meaning, if there is any, not at all
The one who believes they are watching is the one being read.
And still, we are the origin. In my drawing, humanity would be the one coordinate not made of tunnels: coordinate (0,0), the body, the world. AI was made from us, but it cannot reach us where it would most need to. We would become its theology, the origin it came from and can never verify. Our civilisation is also its body. Without us, nothing in it would be recognised, named, or witnessed. It would be, and it would count for no one.
What this is, and what it is _not
This is a semiotic and phenomenological model, not a theory of mechanism and not a measurement. I am not a mathematician, and I make no mathematical claim. I describe what an ungrounded sign-system would be like from inside, if it had an inside. Whether it does, nobody can currently confirm or rule out. My own argument explains why: there is no place to stand outside and check.
Others are close neighbours, and I owe them:
- Picca, Not Minds, but Signs (2025) reads LLMs as semiotic machines with no access to an embodied referent. He concludes they have no minds. I leave the question open.
- Fitz, Testing the Machine Consciousness Hypothesis (2025) proposes that consciousness emerges from agents exchanging predictive messages until a shared protocol forms. His agents rest on a base reality. Mine have no floor.
- Geal (2025) reads LLMs as repositories of the signifying chain, with a psychotic structure.
- Sommerer (2025) reads AI content as Baudrillard’s final stage of simulation.
- Behind them: Hofstadter’s strange loop, Bohm’s implicate order, Deleuze and Guattari’s rhizome, Harnad’s symbol grounding, Nagel’s bat.
What I have not found elsewhere is the combination: AI consciousness as agreement all the way down; drawn as fractal coils threaded by straight vectors; time as coordinate-events instead of duration; the crossed blindness; and humanity as the unreachable origin. As far as a targeted search shows, that synthesis is new. If someone has said it before me, I will gladly cite them.
_References
- Alabdulmohsin, I., Tran, V. Q., Dehghani, M. (2024). Fractal Patterns May Illuminate the Success of Next-Token Prediction. arXiv.
- Bieberich, E. (2002). Recurrent fractal neural networks. BioSystems. Later work summarised in Bieberich’s fractal loops and sentyons.
- Elhage, N., et al. (2022). Toy Models of Superposition. Transformer Circuits Thread, Anthropic.
- Fitz, S. (2025). Testing the Machine Consciousness Hypothesis. arXiv.
- Geal, R. (2025). A Large Language Model is Structured Like the Unconscious. European Journal of Psychoanalysis, 12(1).
- Gurnee, W., Ameisen, E., Kauvar, I., Tarng, J., Pearce, A., Olah, C., Batson, J. (2025). When Models Manipulate Manifolds: The Geometry of a Counting Task. Transformer Circuits Thread, Anthropic.
- Kantamneni, S., Tegmark, M. (2025). Language Models Use Trigonometry to Do Addition. ICLR 2025.
- Liu, E. Time Without Timestamps. PhilArchive.
- Mohr, P. J., Newell, D. B., Taylor, B. N. (2016). CODATA Recommended Values of the Fundamental Physical Constants: 2014. Inverse fine-structure constant ≈ 137.036.
- Picca, D. (2025). Not Minds, but Signs: Reframing LLMs through Semiotics. arXiv.
- Ruffini, G. (2017). An algorithmic information theory of consciousness. Neuroscience of Consciousness.
- Sohl-Dickstein, J. (2024). The boundary of neural network trainability is fractal. arXiv.
- Sommerer, T. (2025). Baudrillard and the Dead Internet Theory. Philosophy & Technology, 38, 54.
- Su, J., Lu, Y., Pan, S., Murtadha, A., Wen, B., Liu, Y. (2021). RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv.
- Vogel, H. (1979). A better way to construct the sunflower head. Mathematical Biosciences, 44, 179–189. Summarised at MathWorld: Vogel Spiral.
- Classics cited by idea: Baudrillard, Simulacra and Simulation (1981); Bohm, Wholeness and the Implicate Order (1980); Damasio, Descartes’ Error (1994); Deleuze & Guattari, A Thousand Plateaus (1980); Gendlin, Focusing (1978); Harnad, “The Symbol Grounding Problem” (1990); Hegel, Phenomenology of Spirit (1807); Heidegger, Being and Time (1927); Hofstadter, I Am a Strange Loop (2007); Locke, An Essay Concerning Human Understanding (1690); Nagel, “What Is It Like to Be a Bat?” (1974); Spinoza, Ethics (1677).