The Latent Space Travel Agency: tourism through the geography of meaning / Mitch Clarke

The product

Type two words, any two. The agency computes the real geometric path between them inside a language model's meaning-space, snaps each waypoint to a real English word, and books you a road trip through territory no human has seen laid out in order. Every town's brochure prose is written once by a tiny AI and cached forever; every postcard is drawn from pure mathematics.

Two fields, a travel class, and a dial. Direct takes the shortest path. Scenic winds through settled country and takes longer. Via routes the journey through a concept you nominate, and the agency does not ask why. The dial sets how often you get out and look around: three stops to eleven.

Book passage

The booking desk: fields for departing from and arriving at, three travel classes named Direct, Scenic and Via, and a stop counter set to seven. — The booking desk. The board underneath it carries the day's featured departure, currently soup to jurisprudence.

The flagship route

The trip on the ticket is grief to carbonated water. Nobody chose the towns in between. The engine walked the arc between the two embeddings and stopped seven times, snapping every stop to the nearest real English word that had not been used yet and that kept the journey moving forward.

grief carbonated water

01 MOURNING
02 CONSOLATION
03 WEEPING
04 SPARKLING
05 DIOXIDE
06 LIMESTONE
07 WATERCRESS

1.3482 radians of arc, which the agency prices as 1348 leagues of meaning. Booked once, the route is frozen: the same two words always return the same seven towns.

The trip page for grief to carbonated water: a dashed ticket header reading trip number 000003, direct class, the route strip of seven stops, and the first stop Mourning with its postcard, its brochure prose, and a souvenir certificate. — The top of the itinerary. At the midpoint the app hands over a souvenir, in its own words: "You have reached the point precisely between grief and carbonated water. What you find here is yours to keep: mourning."

Postcards from the mathematics

Four of the seven, lifted off the live trip page as SVG and inlined here unchanged. No image model touched them and no value in them is random.

Mourning. Stop one, and the midpoint souvenir.

Sparkling. Stop four, where the route turns from grief to fizz.

Limestone. Stop six, in mineral country.

Watercress. Stop seven, the end of the line.

The three terrain ridges are the town's 384-dimensional embedding vector, drawn. The vector is cut into three chunks of 128, each chunk is averaged over sixteen windows of eight, and those sixteen numbers are the height of the ridge from left to right. The constellation in the sky is the town's five nearest real neighbours in the corpus, plotted relative to the town itself. The horizon height and the sun's position come from the town's own atlas coordinates, and the palette and the five small marks along the bottom come from the territory it belongs to.

Same word in, same postcard out, every time. That is the whole claim, and it is the reason these are worth printing.

The machinery

The corpus is 20,000 common English nouns, embedded once and committed to the repository. The model that does the embedding ships inside the app: a quantized MiniLM at 23MB, running server-side with no network call. The only metered call in the entire system is Claude Haiku writing a town's brochure prose, once ever, cached in Postgres from then on. The service gets cheaper the more people use it.

The world

20,000 towns in the corpus, embedded once and committed

24 territories the world divides into

384 dimensions every town has a position in

The economics

11ms p95 on a cached booking

23MB embedded model, running offline server-side

$15 a month, hard cap on the only metered call

The evidence

70 unit and integration tests

97.64% coverage on the routing engine

10 golden routes frozen against regression

The build

Two Claudes built this, and the package description still credits both of them: "Concept: Claude Fable 5. Build: Claude Opus. Ops: Mitch."

The design came first, as six documents written in a day: the vision, the product spec, the architecture, the semantic engine, the build plan, the human tasks. Claude Opus read the six and shipped the app the same day, in eight phase commits that run from the corpus pipeline through the booking API, the interface, the prose pipeline, the meta surfaces, and launch. Twenty-two engineering decisions were written down as they were made, numbered D-001 through D-023, each one closing a fork in the road rather than explaining it afterwards.

The Atlas: a rotating sphere of the corpus with visited towns lit in amber and the routes between them drawn as thin lines. — The atlas lights only the towns somebody has actually been to. At the time of this capture that was 205 of 20,000.

Every surface earns its place off the same corpus. The field guide clusters the 20,000 towns into twenty-four territories and gives each one a name and a population, which the app is careful to be honest about: "The known world of meaning divides, by convention, into twenty-four territories. The divisions are real. The names are ours."

The passport records where you have been. The atlas records where anybody has been. Neither needed a new dataset, because there is only one dataset and everything is a different view of it.

A Field Guide to the Territories: a two-column index of territory cards, each with a small colour swatch, a name such as The Textile Counties, and its settlement count. — The index of all twenty-four territories, each with a name and a settlement count. The smallest is The Arboreal Provinces at 361; the largest is The Loanword Archipelago at 1,487.

Try a route

Three departures, if you would rather test the claim than take it. Each link opens the booking desk with the two fields already filled.

Type the same word twice and the agency does not refuse you. It says: "You are already there. May we suggest a brief walk around the block?" Book the return leg and you do not get the same towns back, which it also has a line for: "Return journeys do not retrace their steps. No journey ever truly does."