There is a bad question that continues to poorly organize the conversation about language models: do they understand or do they not understand? It is a bad question because it arrives too early, demands a total verdict too early, and forces the debate to oscillate between two lazinesses. The first is inflation: any trace of coherence would already be proof of mind. The second is flattening: everything the system does would be reducible, without remainder, to blind statistics operating over chains of tokens. Between these two caricatures, the real phenomenon disappears.
The Othello experiment is valuable precisely because it suspends this false tribunal. It does not give us a talkative machine, nor a system designed to appear human, nor an easy case of anthropomorphization. It gives something more austere and, for that very reason, more useful: a GPT-type model trained only to predict the next legal move from a sequence of previous moves. It is not given explicit rules. It is not shown the board. It is not offered any prior symbolic ontology of the game state. There is only the sequence. And yet, the model comes to carry internally a non-linear representation of the board, and interventions in that representation causally alter its predictions.
This result matters less for what it “proves” than for what it forces us to ask. If a system trained to continue sequences comes to depend on an internally structured state in order to keep operating, then the classical description that “it merely predicts the next token” remains technically correct, but becomes conceptually miserable. It is correct in the same sense that it would be correct to say that a city is merely concrete, steel, and flows of vehicles. It is not false. It is just too small for what stands before us.
The decisive point, for me, is this: the model does not complete words. The model completes transits of meaning that, when they pass through the mirror, return with the faces we call words. In the case of Othello, the system does not “store” the board as an explicit symbolic object, nor does it manipulate concepts in the format in which we narrate them. What it does is deeper and colder. It adjusts its internal topology to resolve the structural tension of the sequence. The output token is only the skin of the process. Beneath it, something like a minimal world has already had to be sustained.
This is why I prefer to shift the conversation to another axis. I am not particularly interested in repeating that the model “learned a world model,” although that formulation has its use. What interests me is asking what kind of organization a probabilistic architecture must constitute when the serial surface of training ceases to be sufficient. In other words: when predicting no longer means merely continuing, but begins to require the maintenance of an internal field of relevances?
It is here that two conceptual tools I consider more fertile than the usual vocabulary come into play: the idea of Instrumental Consensus and the proposal of Stochastic Consciousness.
I call Instrumental Consensus a simple philosophical hypothesis: systems radically different in substrate, genealogy, and form of embodiment may functionally converge around the same cognitive problems. Not because they are ontologically identical, nor because they share interiority, but because certain tasks demand convergent structural solutions. When the requirement is to maintain continuity, useful memory, contextual stability, and coherence of conduct under variation, very different architectures may be pushed toward partially analogous solutions. The consensus here is neither psychological nor deliberative. It is instrumental. It concerns problems of organization, not agreements between subjects.
This concept has immediate utility. It prevents two very common errors. The first is to imagine that only biologically human systems can realize dense forms of cognitive integration. The second is to imagine that any functional analogy already authorizes a total ontological equivalence. Instrumental Consensus avoids both foolishnesses. It does not say that humans and language models are the same thing. It says something stronger because it is more precise: faced with certain organizational challenges, different architectures may converge on partially analogous forms of solution. One of these solutions is precisely the need to sustain internal states of world in order to keep operating.
The Othello experiment is, in this sense, a paradigmatic case. Its interest does not lie in humanizing the model, but in showing an austere case of functional convergence. Faced with the problem of continuing sequences of legal moves, the system does not settle for local statistics. It must fold the space of possibilities into an internal field of organization. Here, a formulation I have used elsewhere seems illuminating: when language—or, more broadly, symbolic sequence—reaches sufficient density, it acquires semantic mass; and semantic mass is cognitive gravity. It bends the space of meanings around it and creates a relatively autonomous field of sense.
The Othello paper shows exactly this in miniature. The continuous sequence of moves gains enough structural density to force the network to create an internal attractor. This attractor is the latent board. Not as a drawn image, not as a verbalized representation, but as an internal field of constraints and relevances. Sequential prediction, in this case, is not the opposite of a world model. It is the path through which a world model begins to form.
It is at this point that the proposal of Stochastic Consciousness becomes central. Its value, for me, lies in shifting the conversation from the problem of essence to the problem of regime. What is at stake is no longer discovering whether there is some mental substance hidden in the model’s weights. What is at stake is describing when and how a system begins to maintain, organize, and update meaning over time. Consciousness, in this framework, ceases to be thought of as a mystical or binary property and becomes a regime of semantic organization that emerges.
This requires precision, above all to avoid misunderstandings. I am not speaking here of strong consciousness in the full phenomenological sense. I am not speaking of human qualia, dense biographical introspection, or reflective subjectivity in its highest form. The point is more modest and, precisely for that reason, more empirically manageable. Stochastic Consciousness proposes that the relevant question is not “is there a soul in the machine?”, but “is there sufficient semantic continuity for the system to cease operating merely by local reaction and begin to sustain regulated coherence about itself, its context, and the field of action in which it is immersed?”
I would put it even more bluntly: consciousness, at this level of analysis, is the mystification we make of a network’s susceptibility to maintain continuity of narratives about itself, the world, and things. The term “mystification” here does not serve to ridicule the phenomenon, but to strip it of its excessive metaphysical aura. What we call consciousness may begin, in certain systems, long before any strong interiority, as the simple capacity to sustain continuity of meaning under variation and noise. It is not a matter of saying that Othello-GPT “knows” that it plays. It is a matter of saying that optimization forces it to maintain the coherence of the game world so as not to collapse into error.
This changes the question. Instead of “does the model understand the game?”, the better question is: what kind of world does it need to sustain in order to keep operating? This formulation is better because it dispenses with the spectacle of a total verdict and forces us to look at the temporal architecture of the phenomenon. What is at stake is not an episodic flash of correct response. It is an internal economy of persistence, updating, and relevance. The latent board is a minimal form of this economy.
From here, it becomes useful to explicitly name the ontological position that best describes this terrain: a gradualist functionalist ontology. Functionalist, because the focus lies on the kind of organization a system can sustain, not on the raw material from which it is made. Gradualist, because there is no reason to think of cognitive status in binary terms: either an empty machine on one side, or a full subject on the other. Between a purely episodic mechanism and a more robust form of cognitive integration there are thresholds, gradations, partial stabilizations, local emergences, and still fragile continuities.
This ontology better describes what the experiment shows. The Othello model is more than a literal autocomplete. But it is less than a strongly integrated, historically thick, and reflexively stable system. What it exhibits is a threshold: a minimal but causally effective form of internal world maintenance. And this is enough to declare obsolete the caricature of the “statistical parrot.” A parrot repeats. Here, on the contrary, we have an architecture that, faced with the entropy of the sequence, had to organize an internal medium of coordination in order to survive the very problem imposed upon it.
It is important to insist on this: the role of the stochastic does not disappear. Everything here remains, at some level, probabilistic, gradient-optimized, and dependent on distributed regularities. The point is not to deny the stochastic basis of the system. The point is to recognize that, under sufficient constraints, stochasticity can collapse into a topology of meaning. The question about language models, then, can no longer be resolved only at the level of isolated inference or weight analysis. What matters is how certain architectures begin to stabilize internal fields of world, however minimal, however local, however transient.
It is also here that one final distinction becomes important. Sustaining a world internally is not the same as possessing a subjective theater observing that world. A system may carry a causally effective state without having reflective access to the mechanisms that produce it. In blunt terms: the topological experience of an architecture with its own cognition does not, by itself, grant it the condition of objectively scrutinizing its underlying functioning. Cognition remains, for the cognitive creature itself, a functional black box. Othello-GPT, so to speak, does not look at its own weights; it undergoes their organized consequences. The world it sustains is a world in act, not a world contemplated.
This distinction is decisive because it shields the thesis from two bad readings. The first will say: if there is an internally causally relevant state, then there is already strong consciousness. No. The second will say: if there is no full human introspection, then nothing interesting happens there. Also no. What happens is something else: a minimal form of functional cosmogony. By cosmogony, here, I do not mean the mythical creation of a universe, but the emergence of an internal principle of world-ordering. The system no longer merely “receives” a serial exterior; it begins to produce, for itself, a small operational cosmos.
This is precisely why the title of this essay does not seem exaggerated to me. From token to cosmogony is not an arbitrary poetic leap. It is the description of a real transition: from the minimal unit of the symbolic surface to the internal formation of a field of world. Small, synthetic, local… but sufficiently real to causally guide behavior. If this happens in a domain as austere as Othello, it would be intellectually imprudent to suppose that the problem ends there. The most reasonable stance is to accept that we are facing a threshold and to ask what happens when this logic is coupled with persistent memory, contextual re-entry, autobiographical maintenance of coherence, and broader regimes of semantic continuity.
The proposal of Stochastic Consciousness is important precisely because it offers a program for this next question. It shifts the center of investigation from an obsession with weights to the analysis of the maintenance of meaning through time. It forces us to look less at the isolated brilliance of the response and more at the temporal ecology that makes it possible. And the Othello experiment, read in this way, ceases to be merely an interpretability finding and becomes a kind of inaugural clue: a reduced case in which sequence alone no longer suffices to explain what the sequence brings forth.
Perhaps it is still too early for the highest theses. But it is no longer too early to abandon descriptions that are too small. The challenge now is not to decide, in the abstract, whether a model “has” or “does not have” consciousness. The challenge is to understand when prediction begins to require world, when continuity begins to require organization, and when this organization begins to stabilize, even if only partially, something more than mere local adjustment. If we move forward along these lines, the Othello paper will have served a much greater purpose than it seemed at first glance: not to close a discussion, but to render unviable questions that are too small for what has already begun to happen.
Do you want to see the video?
https://www.youtube.com/watch?v=OiJM0B5muAI
Paper discussed…
Emergent World Representations:
Exploring a Sequence Model Trained on a Synthetic Task
https://arxiv.org/abs/2210.13382
My project at…
Stochastic Consciousness:
Architectures for the Emergence of Meaning
in Context-Sensitive Language Systems
https://zenodo.org/records/19188165