Whereabout

How Atlas guesses

Atlas is the AI you play against. Nobody has told it where any of these photos were taken. It reads each one with CLIP, a model trained to match pictures with descriptions, and narrows the whole planet down to one city in three steps.

  1. 1

    Read the photo

    CLIP turns the photo into a long list of numbers that captures what is in it: the light, the plants, the roofs, the road markings. Descriptions in words get turned into numbers the same way, so a photo and a sentence can be compared directly.

  2. 2

    Pick a country

    Atlas compares the photo with sentences like “a photo taken in Peru” for every country, in three phrasings averaged together, and turns the matches into probabilities that add up to 100%.

  3. 3

    Pick a city

    In its three most likely countries, Atlas compares the photo with “a photo taken in Cusco, Peru” for each country's 40 largest cities. Its guess is the city with the highest chance overall: how likely the country is, times how likely the city is within it.

The clues you see after each round, like “Snow and ice” or “No buildings”, come from the same trick: Atlas compares the photo with a few descriptions of the scene and only shows a clue when one clearly wins.

Where Atlas looks

CLIP only ever sees a square from the middle of each photo. To find out which parts of that square Atlas relies on, it is split into a 7×7 grid and Atlas guesses again with one cell greyed out at a time. Where hiding a cell makes Atlas much less sure of its country, that cell glows. It takes about fifty extra looks per photo, so it runs once, offline, alongside the guesses.

A glowing map is only worth something if it is true, so it is tested. Hiding the 5 cells that glow brightest cut Atlas's confidence in its country by 63% on average; hiding 5 random cells cut it by 32%. The bright cells mattered more on 92% of the 388 photos.

How often it gets it right

Every system below played the same 388 photos, in a run on October 22, 2026. Two models were compared, each with and without the city step, against two simple baselines. The most accurate one became the Atlas in the game.

How close each system got, on the same 388 photos
SystemWithin 25 kmWithin 750 kmRight countryMedian errorAverage points
Random city0.3%2.3%1.6%7,744 km7,207 km to 8,262 km350
Best constant guess2.3%7.2%9.5%6,529 km6,105 km to 7,205 km655
CLIP ViT-B/32 (LAION-2B), country only4.6%28.3%26.6%2,422 km1,911 km to 3,163 km1,677
CLIP ViT-B/32 (LAION-2B), country then city5.7%29.9%25.5%2,139 km1,561 km to 2,604 km1,814
CLIP ViT-B/32 (OpenAI), country only3.1%25.0%19.9%2,997 km2,290 km to 4,106 km1,502
CLIP ViT-B/32 (OpenAI), country then cityPocket Atlas, which runs in your browser3.6%23.2%15.7%3,998 km2,721 km to 4,885 km1,427
CLIP ViT-L/14 (OpenAI), country only8.3%43.6%42.0%972 km778 km to 1,120 km2,360
CLIP ViT-L/14 (OpenAI), country then cityThe Atlas you play against11.9%46.1%38.1%906 km707 km to 1,178 km2,487
CLIP ViT-L/14 (OpenAI), country then city, then nearest reference photos5.1%37.4%38.1%1,280 km1,012 km to 1,546 km2,210

Median error has a 95% confidence interval under it. Average points use the game's scoring, where a perfect guess is worth 5,000.

Atlas also tried moving its pin to the most similar of 4,322 geotagged reference photos (never game photos) in the country it chose. That moved the pin on 275 of 388 photos, and the median error went up by 374 km (95% interval +187 km to +600 km). It only plays once the whole interval is below zero, so the game keeps the Atlas without it for now.

Where Atlas struggles

Atlas's median error on the game's 388 photos, split by the continent each photo was really taken on. Some continents have only a handful of photos, so treat the small groups as hints, not verdicts.

Atlas is closest in Asia (512 km) and furthest off in Antarctica (15,144 km).

Asia: median error 512 km over 113 photos; right country 46%Asia113 photos · country 46%512 kmNorth America: median error 894 km over 66 photos; right country 52%North America66 photos · country 52%894 kmEurope: median error 910 km over 111 photos; right country 23%Europe111 photos · country 23%910 kmAfrica: median error 1,323 km over 35 photos; right country 43%Africa35 photos · country 43%1,323 kmOceania: median error 1,864 km over 24 photos; right country 50%Oceania24 photos · country 50%1,864 kmSouth America: median error 1,890 km over 38 photos; right country 26%South America38 photos · country 26%1,890 kmAntarctica: median error 15,144 km over 1 photo; right country 0%Antarctica1 photo · country 0%15,144 km
ContinentPhotosMedian errorRight country
Asia113512 km46%
North America66894 km52%
Europe111910 km23%
Africa351,323 km43%
Oceania241,864 km50%
South America381,890 km26%
Antarctica115,144 km0%

Its best guesses

And its worst

These are the clearest view of what a zero-shot model gets wrong, and the first thing a trained Atlas has to fix.

Choosing the photos

A good round needs an outdoor photo of a place, with no faces and no sign that simply names the town. To grow the game past hand-picked photos, CLIP screens each candidate first: it scores how much a picture looks like people, an interior, readable text, not-a-photo at all or a close-up of an animal or plant (nothing to locate), and turns it away if any score is too high. The bar is set so the hand-reviewed game photos get through (88 of 88 do), then tested on images that should be turned away.

How many unwanted images the photo screen catches
Should be turned away forCaught for that reasonCaught for any reason
People in the shot10 of 1211 of 12
Text that names the place12 of 1212 of 12
Indoor scenes12 of 1212 of 12
Not a photo (maps, art, screenshots)9 of 1211 of 12
Close-ups of animals and plants12 of 1212 of 12

The test images come from Wikimedia Commons categories, not from a person checking each one, so a few may be labelled wrong. It catches most people, text, indoor scenes, non-photos and close-ups already.

Read these numbers carefully

Play today's fiveTest Atlas on your photoCode, data and results on GitHub