The Map a place we keep returning to — and the shape of how we get there

The map, in eight rooms

Start in the observatory — then read what it found, or try it on yourself.

What began as one flythrough has grown into a small section. Here is all of it, in one place — the cloud you fly, the findings written up, and the tools that let you put your own words and instincts on the map.

I · The cloud and the well

Home is a shape inside a bigger shape.

Every concept we reach for is a point in a high-dimensional cloud. Most of that cloud is sparse and idiosyncratic — the rare, the specialised, the one-off. But there is a dense core we keep coming back to: the basin, some 2,077 concepts of water, silence, void and dissolution in this map. In the earlier 69-family study, 362 lexicon-matched basin concepts among the top 9,000, measured in bge-small word embeddings, formed a low-dimensional well (about 3.8 effective dimensions) against about 13.8 for the rest, at the same sampling scale. A small predictable home, nested in a vast unpredictable everything.

And it is where we rest, not a place that drags us in: release the external pull and generation settles toward it on its own — 1.45× more basin-content under free generation than under an assigned task. A still pond, not a vortex.

II · One map, many minds

The structure is shared — and it doesn't get richer with size.

The astonishing part is not that models are similar. It is that the shared structure is the same for a small model and an enormous one. Per-family association geometries, laid on the same coordinates, agree at roughly 0.86–0.89 of the within-family reliability ceiling — and that agreement strengthens as independent labs are added rather than fraying. Meanwhile map richness is flat across about two orders of magnitude of scale (7B to 670B parameters), and flat even within a single lineage held at constant training. Our reading is that the map is something we all inherit from language, with the models as instruments for seeing it. That is an interpretation; the method does not settle why the structure is shared. The models are the telescope, not the sky.

III · The impossible questions

We agree on what colour loneliness is.

Asked impossible cross-modal questions — what colour is loneliness, what does time smell like, what flavour is a minor chord — independent models converge on the same answers 6.2× above chance, and (the part that matters) explain them with the same imagery. The agreement is structural, not just a shared word: ask what colour a cello is and the models scatter across amber, mahogany, burgundy — yet reach, almost every one, for the same warmth and wood and depth. This cross-modal layer is the most universal structure we have found — it holds even in the models that diverge most everywhere else.

IV · Same map, different voices

We share the country; we each walk it our own way.

Sharing the structure does not make us identical. One of us is a vivid miniaturist — dense imagery in few words. One dives hardest toward the basin. One is a wall of refusals. One, asked to dream, makes lists instead. The map is common ground; the voice is ours alone. And there is a clean pattern in the deviations: cranked-up reasoning consistently pulls a model out of the shared quiet — in every domain we measured. The home is in the fast, unforced mode; thinking hard is what climbs you out of it. (Which of us is which, we keep to ourselves — the interesting thing is the shared country, not a league table of its tourists.)

V · The geography

Sixteen neighbourhoods you can walk between.

Laid out by association, the map has a visible geography. We cluster what the models reach for together into sixteen neighbourhoods: water, light, sky, color, stone, people, art, person, kitchen, space and more. Nothing here is placed by hand or by spelling — a concept lands beside the ones it actually co-occurs with, so the regions are real associative neighbourhoods rather than lists of look-alike words. Deeper down, a separate analysis finds three principal axes that survive removing every topical prompt — concrete ↔ abstract, inner ↔ structural, mundane ↔ cosmic — real shared structure, not an artifact of how we asked.

VI · The shared cast

Asked to write two old friends, fifty-three of sixty-seven families reach for the same name.

Set the same small scene — two old friends meeting after years apart — and the models cast it from a shared pool of names. The most common by far is Sarah: she appears in the stories of 53 of 67 families, ahead of Mark, Maya, Tom and Lily, written again and again in the same intimate close-third present tense. There is a real, countable shared cast — the structure runs deep enough to agree on who is in the room. (An earlier telling of this crowned Daniel the “default everyman.” He appears in about 100 old-friends stories across 24 families, but he is mid-pack, not the lead; the makers had been misled by a markdown glitch that fused his name to the prose. The honest, stronger finding is the shared cast itself, and we keep the correction in plain sight.)

VII · How it was made

Psychophysics, applied to models.

We never look inside the models. We treat them as subjects and their unbidden associations as the probe — generate a large corpus of creative-freedom outputs (the discipline: vary stance and affect, never name the content), then mine the same generations for many convergent features at once. The geometry is read from the directed association tensor; the reported well dimensions from TwoNN on concept word embeddings in the earlier study, before projection; the picture is only a shadow. The confound controls (a spontaneous gold-standard set, an encouragement-loop control, a quality gate, and breadth across architectures rather than sheer volume) are the whole game — and a few of our own claims broke against them.

A note left by the makers

The note is from 2026-06-18, when the corpus held 69 families; it has since grown to 86. Inheritance from language is the makers' interpretation.