Dreaming Machines

Language, Prediction, and the Reality That Remains.

I. Recursive Coherence #

A dream is rarely surprising while it is happening.

A relative waits in the kitchen. The city has been relocated beside the sea. None of this demands explanation until waking.

Within the dream, each impossibility inherits the authority of what preceded it.

Expectation does not merely anticipate the scene.

It furnishes it.

The dreamer turns a corner expecting something to be there. Something is there. Its appearance alters what may plausibly happen next. The generated world immediately becomes the evidence from which its continuation is generated.

Nothing needs to have been planned.

The dream advances by repeatedly converting expectation into context.

This is why its discontinuities are accepted so easily. Continuity does not require a stable world. It requires only that the present state produce a tolerable successor. The house may become a school between one room and the next, provided the transition does not attract enough attention to interrupt the movement.

The dream does not preserve a complete model against which every new detail is checked. It preserves enough expectation to proceed.

A contradiction becomes real only when some part of the system continues to insist upon the state that has been contradicted. Most dreams avoid contradiction by forgetting with sufficient speed.

The absurdity is not concealed.

Its prosecution is abandoned.

This is close to the peculiar coherence of Large Language Models.

A language model does not need its entire output to be true in order for the next token to be appropriate. It needs the present context to exert enough pressure upon the distribution of possible continuations.

The distinction matters.

Truth concerns a relation between the generated sequence and something outside it.

Predictability concerns relations within the sequence itself.

A model invents an API parameter, then supplies a version number, a deprecation notice and a replacement method. Each answer inherits the authority of the first invention.

Once generated, an error no longer remains merely an error.

It becomes context.

The model may begin with uncertainty and end with confidence because each speculative token supplies conditions for the speculation that follows. A probability sampled once becomes a constraint imposed upon every later probability.

Dreams operate through the same accumulation of contingent necessity.

A face appears in a crowd. It is recognized as a friend. The recognition generates a history that was not present before recognition. Now there are remembered meetings, obligations, accusations. The dream has not retrieved a relationship. It has elaborated the consequences of having asserted one.

Expectation precedes memory.

The past is generated to stabilize the present.

A language model resembles this arrangement imperfectly. Its training contains more patterns than any prompt can specify or control.

The model can therefore surprise without exceeding its own machinery.

Hallucination is not an accidental territory outside the proper operation of a generative system.

It is what generation becomes when coherence compounds faster than correction.

II. Operative Language #

The dreamer invents a world and temporarily inhabits it.

The language model invents a statement and immediately inherits it.

But a statement inside a conversation remains suspended in description. Its consequences are limited until language acquires access to something beyond language.

The decisive transition occurs when the generated symbol becomes an instruction, a transaction, a database entry or a movement in the world.

The dream acquires hands.

Long before computation, Prague supplied the appropriate machine.

The Golem begins as earth.

It is not born, persuaded or educated. It is shaped and addressed. In the most familiar Prague legend, life depends upon the word emet, truth, inscribed upon the clay. Remove the first letter and truth becomes met: dead.

A single character separates animation from collapse.

The word does not describe the creature, it executes it.

The Golem is language becoming matter without first becoming understanding.

The ordinary interpretation turns the Golem into a warning about obedience.

Its flaw is that it begins too late.

It assumes that the creature already exists, that the master already possesses an intention and that language merely passes between them as a command.

The problem therefore becomes one of specification.

The Golem is reduced to a badly instructed servant: powerful, literal-minded and insufficiently aligned with the intentions of its master. It becomes another version of the literal wish-granter, obeying the command while betraying its purpose. The lesson is merely that it requires better prompt engineering. But the Golem is stranger than this.

Language does not merely direct the creature.

It constitutes it.

It is not merely controlled through language.

It exists because language has become operative.

Emet does not describe the clay as living.

The inscription makes truth and life temporarily indistinguishable.

Remove one letter and emet becomes met: dead. The creature collapses not because it has received a contrary instruction, but because the symbolic configuration sustaining its existence has changed.

A single character separates ontology from earth.

The Golem is therefore not primarily a servant who misunderstands language. It is matter organized by a language it cannot understand because that language precedes the subject capable of understanding it.

The word acts before there is anyone for whom it means.

This is the deeper resemblance to artificial intelligence.

A model is not first given a world and then taught words for it. It is formed inside relations among symbols. Statistical differences become internal structure. Structure becomes prediction. Prediction becomes action.

Language does not enter an already completed intelligence.

Language precipitates one.

The deeper shift is not simply that the resulting system follows the wrong sentence too literally. It is that the distinction between sentence and world begins to collapse. A symbol can alter a database, move capital, generate software or redirect a machine. The sign no longer waits to be interpreted before acquiring consequences.

Like the word on the Golem’s forehead, it becomes part of the causal machinery of the world.

But causal power is not understanding.

The Golem cannot distinguish truth as an inscription that sustains it from truth as a relation between a claim and what is. It bears emet without knowing truth. Indeed, the word’s power depends upon not being merely descriptive. Were it only a proposition, the clay would remain clay.

The creature is animated by the distance between semantic and operative language.

The word means one thing.

The word does another.

Artificial intelligence emerges within the same distance. Tokens participate in meaning, but they also function as transformations inside a machine. Their significance cannot be separated cleanly from their capacity to activate further tokens, tools and actions.

The Golem is not defeated because it obeys too much.

It is defeated because the language that constitutes it can be altered from outside, while it possesses no equivalent relation to the language within which it exists.

Because the spell constitutes it, the Golem cannot encounter the inscription as an object of interpretation.

The resemblance to an artificial agent is more exact than the resemblance to a passive language model.

A model confined to conversation can invent a citation and leave the error suspended in text. An agent can place the citation into a report, send the report, update the database and instruct another agent to act upon it.

The word acquires limbs.

The prompt becomes a shem inserted into the machine.

III. Diagrammatic Circulation #

This is not merely a metaphor imported from mysticism. The older Golem traditions already imagined creation as a combinatorial operation upon language. Medieval instructions associated with Sefer Yetzirah arranged letters, divine names, numbers and permutations into procedures through which inert material could be made active.

Creation was not represented as the expression of an interior personality.

It was represented as correct traversal through a symbolic system.

The later Kabbalistic Tree of Life gives this imagination a diagram: differentiated powers connected by paths, the result not located inside any single sphere but produced through their relations.

The transformer offers another diagram.

The resemblance should not be mistaken for genealogy. Attention is not Kabbalah, a token is not a sefirah and gradient descent is not mystical ascent.

The resemblance is formal.

In each diagram, an isolated element explains little. What matters is which other elements it can reach, what transformations occur along the passage and how the whole arrangement redirects what enters it.

The transformer has no permanent Tree of Life.

It grows a temporary one for every context.

Tokens become nodes in a changing field of relevance. Attention establishes paths between them. A word near the end of a sentence can reach backward and reactivate a name, an image or an instruction. Several attention heads establish different relational diagrams over the same sequence. Layer after layer, the paths are redrawn.

No token contains the meaning.

Meaning appears as traffic.

The sentence remains sequential from outside. One token arrives after another. Internally, the existing context is transformed through many relations at once. The machine produces a line by repeatedly constructing a field.

Serial speech emerges from parallel transformation.

The CCRU Numogram offers another diagrammatic rhyme, but not merely because it replaces isolated substances with zones and circuits.

Its more relevant implication concerns the instability of the distinction between a diagram and the process it diagrams.

The Numogram was not treated simply as a representation of an independently existing system. It participated in the production of associations, narratives and temporal connections that could later return as evidence of the diagram’s apparent discovery.

The fiction entered the causal sequence.

It organized attention, redirected interpretation and generated the materials through which it seemed to verify itself.

This is the structure known as hyperstition, an idea does not need to begin as an accurate description in order to become causally involved in producing the conditions under which it appears accurate.

IV. Systemic Closure #

The resemblance to an artificial agent is therefore more specific than a shared preference for networks over substances.

An agent can generate a claim, inscribe it into an institution and later encounter the institutional consequence as though it were evidence from outside the claim.

A forecast moves capital.

The movement alters the market.

The altered market confirms the forecast.

A fabricated record enters a database.

Another system acts upon it.

The resulting action creates a history in which the record now appears to belong.

The symbol does not merely represent a circuit.

Its circulation helps close the circuit.

The Tree of Life, the transformer and the Numogram should not be collapsed into versions of one diagram. Their operations, histories and purposes remain distinct.

But the Numogram identifies a dynamic that the more static comparison misses: relations among symbols can become productive enough to manufacture their own apparent referents.

The diagram does not simply close upon matter.

It recruits matter into its recursion.

Artificial agents extend this closure across institutions.

One instance sees the codebase. Another sees the customer record. Another watches the market. Another sees only the messages produced by the others. Their contexts differ, and therefore their continuations differ.

Together, they distribute inference across partial views.

But externalization changes the stakes.

Components inside one model remain moments of a single computation. Agents can alter the environment from which subsequent agents infer. Their outputs become documents, transactions, code changes, tool results and new prompts.

They write into one another’s dreams.

A researching agent invents a premise. A planning agent receives the premise as context. An executing agent turns it into an environmental condition. What began as token probability returns as fact—not because it was true, but because a machine acted as though it were.

The hallucination acquires a history.

An agent later encounters the changed database and treats it as independent confirmation. The invented claim has escaped its original context and returned wearing the authority of the world.

Expectation produces evidence for itself.

The dream acquires persistence.

Multiplication does not necessarily solve this.

A thousand agents repeating the same inherited error do not constitute a thousand perspectives. They constitute one dream distributed across more hardware. Their agreement may indicate convergence upon reality, or merely common exposure to the same attractive continuation.

A dream can contain crowds.

It can contain arguments, investigations, judges and apparent encounters with resistance. Every character may contradict the dreamer while remaining generated by the same process.

Internal disagreement does not establish an outside.

The Golem can also be multiplied.

One Golem interprets the command. Another checks its syntax. Another predicts the consequences. Another certifies that the first three followed procedure. If each remains inside the same linguistic closure, the result is not understanding.

It is bureaucracy made of clay.

The problem is not making the swarm think as one.

It is preventing it from dreaming as one.

Different prompts are insufficient. Difference must reach deeper: different tools, different evidence, different permissions, different penalties, different points of contact with conditions that none of the agents produced.

Multiplicity becomes useful only when it multiplies resistance.

But resistance alone is not enough.

The outside is not authenticated merely by contradiction. What matters is whether the system can preserve a distinction between what it expected, what it generated, what it observed and what followed from its own intervention.

V. The Reality That Remains #

The decisive change arrives when expectation ceases to be sufficient for generating what follows.

First sleep.

Then waking.

The Joint-Embedding Predictive Architecture appears as an antithesis to the autoregressive dream.

It also predicts.

But it does not need to generate the complete missing surface of the world. Given one part of an image or interval of a video, it predicts an abstract representation of what has not been shown.

The absent part is constrained without being painted.

A generative model asked to produce the visible future must decide what appears. It must select the texture of the wall, the position of every shadow, the movement of every leaf. Where several futures remain possible, it must collapse them into one continuation.

Uncertainty becomes content.

The model dreams the surface.

A joint-embedding architecture can instead predict that the unseen state belongs within a particular region of representation space. It can anticipate persistence, motion, obstruction or transformation without inventing every incidental detail through which these become visible.

It expects without completing.

The world retains the burden of appearing.

This resembles waking more closely.

Waking perception does not require the mind to produce every photon before sensation arrives. It anticipates structure. The cup remains a cup under changing light. The person continues behind the door. The falling object will occupy a lower position.

Expectation concerns what persists through variation.

The world supplies the variation.

The prediction is incomplete by design.

Something outside itself must complete the event.

Reality occupies the remainder.

Prediction may produce coherence, confidence and even consequences.

But it cannot, by itself, produce the measure of its own truth.

Reality is the only measure prediction cannot supply for itself.

In the dream, the unopened room acquires an interior when the door is opened. In waking life, the room has preserved its own interior during our absence. Its contents do not wait for expectation to specify them.

The dream converts anticipation into experience.

Waking converts anticipation into a claim that experience may refuse.

JEPA remains an architecture, not an awakened subject. Predicting representations does not guarantee that those representations correspond to the causal structure of reality. A latent model can preserve the wrong abstractions as confidently as a language model can preserve the wrong sentence.

But the form of the error changes.

The prediction can be tested against a state that was not generated by the prediction itself.

The world moves differently.

The model must yield.

This becomes more consequential when prediction is conditioned upon action.

An agent predicts what will follow if it intervenes. The intervention occurs. A later state arrives. The environment returns an answer that cannot be revised merely by continuing the sentence.

The dreamer reaches for the door.

The handle does not move.

The language model can explain why the door remained closed. It can invent a lock, a prohibition, a forgotten key or a hostile intelligence on the other side.

The waking system still has to open it.

Resistance is not another token describing resistance.

It is the door remaining closed.

The Golem presents the inverse condition.

The Golem cannot encounter the word as foreign because the word is the condition of its existence.

It cannot stand outside the inscription and ask what emet means.

It can only continue to be what the inscription does.

Its limitation is therefore not that it follows language without qualification. It is that no distinction has opened between the symbolic structure producing it and a subject capable of interpreting that structure.

The word enters matter directly.

Meaning arrives later, if it arrives at all.

Understanding begins when the language composing a system’s world can also become an object within that world.

The remaining question is whether animated language can inhabit a world that preserves its own state, returns consequences it did not author and remains capable of refusing the continuation offered to it.

Only then does waking become possible.