Status: blue-sky reconnaissance. Part of the Representation, Not Language thread; sibling to Map Synchronisation.
Korzybski's formula — the map is not the territory — assumes three things: a territory that exists independently of any observer, a map that is a lossy reduction of it, and the possibility of checking one against the other. The interesting question is not whether the metaphor is true but which of those three assumptions survives when the mapper is an LLM rather than a person.
Durrell's Alexandria Quartet is the metaphor's ideal case. One real sequence of events in Alexandria; several observers; each producing a partial, angle-dependent account. Durrell built it on an explicit "relativity proposition" and denied that any perspective is the correct one — the first three volumes deliberately do not resolve into a canonical version.
Crucially the territory is still there. Darley's map of his affair with Justine is wrong in a specific, checkable way: Balthazar's Interlinear tells him Justine loved Pursewarden and used him as a decoy. That correction is possible only because there were facts of the matter to be wrong about.
So for humans: territory exists; maps are lossy and positioned; the territory can in principle arbitrate between competing maps, even if in practice nobody ever gets full access to it.
The zone claim in Representation, Not Language denies the first assumption. There is no neutral, pre-existing semantic territory that training discovers and represents imperfectly. The zone is grown, contingent on architecture, corpus and objective — not fitted to a ground truth sitting outside it.
Sharper version, and probably the cleanest statement of the human/LLM difference: humans map territory; LLMs map maps. Training data is not the world. It is the accumulated cartographic record — a corpus of other people's already-reduced accounts. The model's representation is a compression of maps, with no independent channel to the thing being mapped and no way to detect where the maps it was trained on were collectively wrong.
Under this reading, an LLM is closer to Balthazar reading Darley's manuscript than to a witness at the events: everything it has is testimony. The difference is that Balthazar was also present in Alexandria.
The obvious next move — "the internal map is a tensor" — is right about the substrate and wrong about the object. Weights and activations are literally tensors; for humans the equivalent is a distributed high-dimensional neural state, which is a metaphor rather than a literal claim. Either way the tensor is the medium, not the meaning.
Keep the distinction sharp, because "the map is a tensor" quietly reintroduces exactly the precise, point-like, arithmetic framing the zone argument rejected. Better: a concept is a zone in a tensor space. The tensor is the material; the zone is the shape. Saying the map is a tensor is like saying a country is paper.
• Is "maps of maps" a difference in kind or only in degree? Most of what any human knows also arrives as testimony — other people's maps — not direct contact. The claim needs to rest on the existence of a perceptual channel, however little used, rather than on how much of a given map came through it. Not yet settled.
• Does multimodality close the gap? Vision and audio encoders give something that looks like territory contact, but it is still a curated dataset — sensors pointed at a corpus, not at the world. Worth deciding whether that counts.
• Does "no territory" overreach into anti-realism? It should not. The claim is that there is no independent territory of meaning for a representation to be a map of. It is not a claim that there is no physical world. Keep the two apart or the thesis will be dismissed for something it does not say.
• If there is no territory, what does "wrong" mean for a model's representation? Darley can be corrected; can a zone be? This is the question that Map Synchronisation picks up.