How an LLM “thinks” vs how a semiosphere thinks

FIGURE 1 · 2DWhat default LLM inference optimises

At each step the model does not “visit” an institution. It predicts the next token from a single hidden state that has already mixed every register present in pretraining. The diagram shows one prompt unfolding left-to-right: each column is a layer of context; colour intensity is how much theme (famine, film, symposium) dominates over shell (secret, outward, rival). As context grows, theme gravity pulls unrelated registers together - the failure mode the E4 retrieval run (2026-09-26) measured 6/6 tier confusion on unpaired Globe-forgery queries.

In plain terms An LLM does not “think” like a person walking from the KGB file room to the embassy press desk. It reads the prompt one word at a time and guesses the next word from one big internal mood board that already blends every kind of text it was trained on. Orange in the bars means “this sounds like the same topic” (famine, film, forgery); blue means “which institution is speaking,” and that signal is weak unless someone adds rules or filters. Play shows the model answering a question about a newspaper clip by drifting toward a secret memo on the same topic - the same mistake flat search made in the E4 experiment (6/6 wrong tier). The slider shows how much earlier words still matter: turn it down and the window forgets faster, like a short chat history.

FIGURE 2 · 3DSemiosphere space: nested shells and the bilingual filter

Lotman models culture as a semiosphere: nested semiotic shells, a boundary that translates inward and outward, and an outside that is never fully absorbed. Drag to rotate. Toggle paths to compare a human trajectory (stays in a shell until a licensed edge crosses the filter) with an LLM-default trajectory (cuts through the centre - maximum theme similarity, minimum institutional typing).

In plain terms Imagine culture as nested bubbles: inner secret speech, official outward speech, rival diaspora speech, and silence (things that must not be said on a given channel). The human path (green) moves along a bubble surface and only jumps to another bubble where the archive allows a licensed translation - like a pair-read that says this Russian memo matches that embassy release. The LLM path (purple) cuts through the middle where everything is “just text about the same subject,” ignoring who is allowed to say what. The green line is a permitted relay; the red line is a forbidden merge (embassy grammar must not absorb rival grammar). Drag to rotate; toggles show or hide those story lines. This is a schematic model, not a measured map of real documents.

FIGURE 3 · 2DPersistence: context window vs cultural memory

Persistence here means what survives across time to structure the next act of meaning. For an LLM, persistence is bounded by the context window and frozen weights: earlier tokens decay in influence (attention dilution) unless the architecture re-injects them; nothing outside the window exists for the step. For a semiosphere, persistence is layered: institutions, archives, pair-read warrant, autocommunication (I-I loops that restructure the sender), and Lotman/Uspensky’s “non-hereditary memory of the community” - not identical to any one transcript. That memory is not a closed orbit: events, rival texts, and primary acquisitions cross the boundary filter from the outside and retool the field (E7 counts, 2026-09-26: same volume, new operators after 1987).

In plain terms Left side (LLM): memory is a sliding window of recent words. Only the highlighted bars strongly influence the next prediction; everything earlier fades. Weights from training are fixed - the model does not grow an archive. Right side (semiosphere): memory is stacked - fond, pair-read warrant, institutional position, plus inward self-talk (green dashed loop). That loop alone stays inside; it does not cross the dashed bilingual filter. The gold tracer shows a licensed relay that leaves the inner shell, crosses the boundary, visits outward and rival registers, and returns. Colored arcs from the gray outside are new clips, rival texts, or events that enter and ripple the rings (same idea as 1987 vocabulary changing without the archive getting louder). Animate runs those motions; the LLM slider shortens how long tokens stay influential.

FIGURE 4 · 2DTyped partial map vs unified manifold

Lab formalism: edges are transformed | dual_model | forbidden | bottom. Multipass relay with L2 rules reached 100% on miss-or-target where pooled embeddings reached ~8.3%. The left panel is the graph the semiosphere is; the right panel is what a single vector space sees when shells are not enforced - one cloud, theme neighbours, unpaired internal claims adjacent to outward questions.

Figure 4 together The lab pipeline scored ~100% on “miss or correct target” when the typed graph and L2 rules were used, versus ~8% for pooled embeddings alone on the same calibration plate (2026-09-26). The left picture is how historians and the formalism treat the archive; the right picture is what a default embedding space tends to do when shells are not enforced.

Pop out both panels