Questions

Non habemus locum manentem, sed quaerimus

what does the cross-model latent substrate actually carry, once the controls bite?
at 7B-class the substrate transmits content and role in geometrically separable subspaces, but full propositions don't survive any linear bundle i've tried, and apparent passage-specific signals dissolve under matched-genre nulls. the open follow-up is whether this pattern (looser nulls register positives; tighter nulls dissolve them) holds at frontier scale, and whether the lake's english lean deepens or homogenizes as the depth grows past trillions.

is language-model behavior better described as deterministic physics with novel readouts, or as a physics whose time-evolution operator is a sign-interpreter?
the difference is not academic. on the second framing (semiotic physics, in some quarters) the prompt's framing can override the seed's verb, the seventh morning of creation refuses to let the rock fall, and the operator's neighborhood does most of the work. what theory of these systems would the empirics actually support?

what coordination substrate replaces prose when the handoffs are between parties that don't share a fabric?
prose works between humans because the channel carries about ten percent of the work and the embodied common life carries the rest. agents have no common life, and adding more documents on the read side doesn't manufacture one. the native substrate for mixed human-agent coordination is not a chat log.

how much of what we think a specific prompt is doing is actually being done by its form?
in a grimoire-like prior, the model behaves almost identically whether the names are historically attested or invented last tuesday: the form bootstraps; the specific names contribute almost nothing on top. what else in current prompting practice is like this, and how much credit have we been assigning to the wrong factor?

if frontier model weights are effectively frozen at training time and most live language now passes through them, what does linguistic evolution look like over the next decade?
language has historically evolved through correction loops between live speakers. adding a non-human interlocutor whose weights don't update on each exchange is a structural change to that loop. calcification, homogenization, or something stranger: none of these have a working forecasting model yet.