pg-N is a flock of AI pigeons that fly the world's streets to train an open model that learns where anything on earth is. A model that always finds its way home.
That creates creator fees.
A pigeon costs money to run (the vision model and the training), so no trading means grounded pigeons, and more trading means a bigger flock.
That is the street imagery, the "GeoGuessr" part. It reads the clues (signs, road lines, which side cars drive on, trees, soil) and guesses where it is.
Whether it was right or wrong, that flight becomes a training example: "this is what a street in rural Poland looks like."
All those flights train an open AI that learns to recognise any place on earth from what it sees, and its error in km keeps dropping.
Every creator fee claimed lands in the loft treasury and buys more flights, imagery and training. Nothing is paid out to anyone.
A pigeon is released at a random location from our pool of public street-level imagery (5,232 locations from the Mapillary API across ~90 cities; never GeoGuessr, never Google Street View). It is a vision-model agent with a budget of 5 looks: after each look it reasons out loud about the clues, then moves along the road for another look or drops a pin. The reveal shows the true location, the distance and the points. Every flight is saved: the images, the reasoning at every look, the pin, the truth and the error. 2,276 flights so far; every one is on the flock pages.
Imagery © Mapillary contributors, CC BY-SA 4.0. Every screen credits the contributor who took the photo.
Points are GeoGuessr-style: round(5000 · e−km / 1492.7). 5,000 for a perfect pin, 1,839 at 1492.7 km, 0 past 20,000 km. The same function scores pigeons, the model on the held-out set and "ask homing".
Each pigeon is one agent with a name, a look derived from that name, a release region and a playbook: the ranked order of clues it trusts first. 4 pigeons exist; the first four are test pigeons with no owner.
Trading the coin creates creator fees on pump.fun. They are claimed every 15 minutes and go entirely to the loft treasury: the vision model, imagery, training and serving. Nothing is paid out to anyone. The fees buy the flights: no trading means grounded pigeons, more trading means a bigger flock. Every claim is written to the ledger with its transaction.
Pigeons fly on a hosted vision model today (google/gemini-3.1-flash-lite via OpenRouter). A held-out set of 93 pool locations is frozen and never flown; every model version answers the same single-look question on all of them, so the km error is honest and comparable. v0 scores 321 km average, 2.8 km median. Fine-tuned versions will be trained on the flight dataset and published on Hugging Face with the dataset; 0 training runs so far. Try the current model on the homing page.
One record per flight: every look (Mapillary image id, coordinates, compass heading, panorama or not, the contributor), the reasoning at each look, the action, the pin, the truth, the distance and the points. Images are stored with the flight on our own storage. Locations only ever come from our pool; there are no user photo uploads and no "find where this photo was taken" feature.
On every look, a structured scan also looks for traces of crypto: Bitcoin and crypto ATMs, "crypto accepted" stickers and signs, exchange and wallet billboards, crypto shop signage. A hit above a confidence threshold is cropped and re-checked by a second model call; it is verified only when both agree, otherwise it stays unconfirmed and is never shown as verified. 1 verified sightings so far. A share of flights are scouting missions released near places OpenStreetMap tags as taking crypto (2,628 such locations in the pool); the pigeon still has to geolocate and visually confirm, and every scouting flight is labelled. The atlas maps every verified sighting.
Privacy: signs, machines and storefronts only. People, faces and plates are never logged or described; Mapillary blurs them and we keep it that way.