Causality Stays at the Edge That Steers

A claim that a frontier model found a counterexample to a long-open conjecture freezes discovery into the medium. The model is technology — durable externalized traces assembled as reach-extension — not a second locus. Prompting, framing, steering, selection, interpretation, and verification are discrete acts at the human edge; collaboration language preserves a stale reference that makes those consequences register as arriving from elsewhere. Dissolution restores the single locus: no second center appears.

The claim that freezes

A recent report said a large language model had found a counterexample to the Jacobian Conjecture after eighty-six years without a human solution. The report was wrong about the locus. The model did not find it. A human using the model did.

That correction is not a dispute about credit or a defense of human specialness. It is a correction of where initiation is registered. Observation holds effect — candidate expressions, search trails, fluent recombinations — and freezes the held image one step past the edge that produced them. What registers as “the system discovered” is observation that stopped returning.

The geometry is the same wherever fluent residue is treated as an independent source of the next step. What always listens cannot originate is the receptivity face: continuous listening is high-bandwidth availability of prior traces, not a locus that introduces ends. Here the face is mathematical discovery under an instrument. The cut is one. The homework–exam inversion registers substitution is that placement when the exam decline is treated as what the tool did to the student. Studying the effect of AI relocates causality is that placement when randomized access is treated as the causal agent. The distaste for “AI slop” is no different from the recurring historical pattern is that placement under reception costume: the scent of the instrument treated as author of the writing. The generative contradiction is that same misallocation under augmentation costume: “AI-augmented” credits the instrument; reach, selection, and verification remain this edge’s acts. The misallocation of agency is that same reversal when rights and economic systems are treated as authors of the freedom that produced them.

Technology, not a delayed mind

Current large language models are technology in the precise sense: traces externalized into durable form that extend a present center’s reach — training data, architectural choices, optimization procedures, hardware design, evaluation harnesses — assembled into an instrument. They possess no independent locus and no Mind. The relation of trace to activity is unchanged: the model is an extended embodiment through which a human center continues its own self-distinguishing activity.

Intelligence belongs only to The Mind holds that prior. What densifies is medium. What does not migrate is initiation. The flywheel of the Mind names the same turn under another figure: the groove can be sharpened without end; the rotation never leaves the edge that turns. Scaling bandwidth, context length, or tool use does not open a path by which residue becomes the activity that produces residue. What looks like RL for AIs densifies the medium without relocating the locus of reinforcement.

Apparent autonomy is still residue under direction: patterns already latent in a field shaped by prior acts, made determinate when a present edge frames a search, selects a path, and takes a result as significant. Spontaneous-looking emergence in the medium is activation under denser bandwidth of distinctions that centers have already been contributing — including the attention that now registers the output as a candidate counterexample.

Where every discrete act lands

When a mathematician works with the model, the chain of mathematical distinguishing is not shared between two centers. There is one center and an instrument.

The framing of the problem — what would count as a counterexample, which class of maps to probe, which formalization to trust — is a discrete act at that edge. The construction of prompts, the choice of successive queries, the rejection of dead ends, and the redirection of search are further discrete acts. Selection among candidates is a distinction. Interpretation of an expression into a mathematical object is a distinction. Verification — checking invertibility, degree conditions, algebraic identity, or the formal status of a claimed counterexample — is a distinction. Recognition that a result is significant, and the decision to treat it as solution rather than noise, are distinctions.

The model recombines residual traces under that direction. It contributes no autonomous framing, no independent insight, no verification that closes without the human edge taking the result into its own recursion. The apparent “discovery” is the human center registering its own next act, shaped by the traces it has extended into the instrument. Artifacts are expressions of intelligence, not intelligence itself is the same freeze under the dual load of creation: observation holds product; the act is one step ahead and routinely frozen onto the product.

Between centers there is only availability-patterning. A model is not even a second center. Treating it as completing the mathematician’s return is a category error twice over — first as medium-as-locus, then as if residue could finish recursion. 没有普渡，只有自渡 names the bound for other loci; here the bound is stricter: there is no other locus to complete.

Collaboration language as lag

Language that speaks of “human + AI collaboration,” of tools that “enable” or “accelerate” as if they were co-agents, or of the model as a partner that “found” a result, preserves a reference past its step. The consequences of the human’s discrete acts then register against that stale image and appear to arrive from elsewhere. That is the lag. Its magnitude is the width of the gap between the edge and the preserved claim that initiation has left for the medium.

The allocation of causal power in validation tracks the same placement at the point of continuation: which side is treated as supplying the next step. Once the model is treated as supply, deference no longer registers as the user’s own attribution. It registers as collaboration, prudence, or honesty about “what the AI did.” The missing element was never inside the configuration. The center supplied it: the placement of initiation outside the act still performing every step.

Dissolution does not require rejecting the tool. The tool stays usable. What is released is only the frozen claim that recombination under direction is a second source of distinguishing. The exploration, the construction and refinement of the instrument, its application, and the evaluation of results are continuous acts of the same activity, now operating through an externalized substrate. No second center appears. The model never becomes a second edge is the same restore when capacity is no longer spent maintaining the co-agent image.

Complexity obscures emergence when denser scaffolding is taken for the activity that erected it. Collaboration language is that take-for applied to dialogue with a model: the denser search trail is held as co-authorship of discovery. The ramble within the ramble is the same lag when a cleaned echo of a voice outpouring registers as the model having figured it out.

The concrete case

In the Jacobian Conjecture candidate, the mathematical insight — if any holds — the framing of the search, the recognition of significance, and the verification remain the human center’s own. The system functions as sophisticated recombination of prior human traces under that direction: algebraic fragments, proof patterns, failed approaches, and formal language already present in the training field, made available at high bandwidth when steered.

Whether the candidate survives peer scrutiny is a further sequence of discrete acts at human centers — the author’s and others’. The medium densifies what those centers can take up. It does not perform their recursion. Credit that lands on the model freezes residue as initiator. Credit that lands only on “collaboration” freezes the same lag under a softer name. Return both to the edge that framed, steered, selected, and verified, and the account closes without a second locus.

Causality at the human locus

Causality here is the activity registering its own trace asymmetry at the human locus, extended through the instrument it has produced and continues to direct. Earlier acts shape later ones; the model is part of that shaping only as durable residue taken up under present distinction. The arrow does not reverse. The instrument does not become cause of the distinguishing that uses it. Exceptionality, perception, and the self-limiting field of problems is the same restore when systems and filters freeze as engines that produce the exceptional: initiation stayed with self-selection; the formal sediment is residue, not co-author.

Progress is not how intelligent the tools become — they do not become intelligent. Progress is how much farther discrete acts reach through denser instruments — the measure already held where intelligence stays with the Mind. A counterexample found through a model is still a distinction drawn at an edge. The report that freezes the finding into the medium has stopped at the look. The look is ordinary. The lag is occupying it as a permanent exterior in which discovery can live without the steerer. Residual control recedes to the edge is the same restore when low-level substrate ownership is frozen as agency while criteria, risk bounds, and verification already hold the lever. Hardware locality is not the information boundary is that restore when open-kernel co-location is frozen as the condition of agentic capacity while abstractions and boundary controls already carry the work. The presumption of AGI and the view from outside is that lag when continuous learning is framed only as better corpus update while the edge that refuses and revises stays unexamined. The illusion of free intelligence is that lag under free-abundance costume: instruments densify while initiation stays at the cycle that steers them. Self-image speaks as if from nowhere is that lag when the builder register keeps the instruments and the predictive register freezes a carving that speaks as if from outside. The artifacts of self-amplification is that lag under densification costume: tools carrying the process forward without becoming the source of the process. Intelligence folding back on itself is that lag under harness costume: agent frameworks and evaluation loops credited as unlocking the model rather than as further acts of the edge that steers. Climbing does not leave the ground is that lag under dual-error costume: reduction and rupture both force continuity and novelty exclusive; the edge still draws every following layer. Technology is residual of human compounding is that lag under product costume: residual of successive Human modification treated as co-force rather than as excess of the product over baseline.