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
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
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
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 60% on average; hiding 5 random cells cut it by 31%. The bright cells mattered more on 89% of the 88 photos.
How often it gets it right
Every system below played the same 88 photos, in a run on October 15, 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.
| System | Within 25 km | Within 750 km | Right country | Median error | Average points |
|---|---|---|---|---|---|
| Random city | 0.0% | 1.1% | 2.3% | 7,836 km7,081 km to 8,722 km | 268 |
| Best constant guess | 1.1% | 3.4% | 2.3% | 6,470 km4,468 km to 7,355 km | 591 |
| CLIP ViT-B/32 (LAION-2B), country only | 5.7% | 28.4% | 30.7% | 2,590 km1,833 km to 3,963 km | 1,645 |
| CLIP ViT-B/32 (LAION-2B), country then city | 9.1% | 27.3% | 27.3% | 2,593 km1,633 km to 4,098 km | 1,739 |
| CLIP ViT-B/32 (OpenAI), country only | 3.4% | 22.7% | 20.4% | 2,936 km1,777 km to 4,171 km | 1,520 |
| CLIP ViT-B/32 (OpenAI), country then cityPocket Atlas, which runs in your browser | 5.7% | 19.3% | 15.9% | 3,426 km2,047 km to 4,566 km | 1,428 |
| CLIP ViT-L/14 (OpenAI), country only | 6.8% | 42.0% | 43.2% | 1,156 km701 km to 2,823 km | 2,149 |
| CLIP ViT-L/14 (OpenAI), country then cityThe Atlas you play against | 13.6% | 43.2% | 42.0% | 949 km649 km to 2,835 km | 2,359 |
Median error has a 95% confidence interval under it. Average points use the game's scoring, where a perfect guess is worth 5,000.
Where Atlas struggles
Atlas's median error on the game's 88 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 Oceania (166 km) and furthest off in South America (7,075 km).
| Continent | Photos | Median error | Right country |
|---|---|---|---|
| Oceania | 5 | 166 km | 80% |
| North America | 16 | 272 km | 75% |
| Asia | 25 | 717 km | 52% |
| Europe | 25 | 1,481 km | 16% |
| Africa | 9 | 3,396 km | 44% |
| South America | 8 | 7,075 km | 0% |
Its best guesses
,_Elizabeth_Quay_Bridge_--_2019_--_0375-9.jpg){kind=link}
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.
Really
Engilchek Glacier, Tian Shan, Kyrgyzstan
Atlas said
Grytviken, GS
15,302 kmoff
Maryliflower, CC BY-SA 4.0
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 or not-a-photo at all, 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.
| Should be turned away for | Caught for that reason | Caught for any reason |
|---|---|---|
| People in the shot | 10 of 12 | 10 of 12 |
| Text that names the place | 12 of 12 | 12 of 12 |
| Indoor scenes | 12 of 12 | 12 of 12 |
| Not a photo (maps, art, screenshots) | 9 of 12 | 11 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 and non-photos already.
Turned away: people in the shot
Redeytraveler, CC BY 4.0
Turned away: people in the shot
Mostafameraji, CC BY-SA 4.0
Turned away: indoor scenes
Joe Ross from Lansing, Michigan, CC BY-SA 2.0
Turned away: indoor scenes
Roland Unger, CC BY-SA 3.0
Turned away: text that names the place
Nbarth, CC0
Turned away: text that names the place
Erwin Beilfuss, CC BY-SA 4.0
Turned away: not a photo (maps, art, screenshots)
Jackidko, CC BY 4.0
Turned away: not a photo (maps, art, screenshots)
Not mycircus605, CC BY-SA 4.0
.jpg){kind=link}
-02.jpg){kind=link}
Read these numbers carefully
- The test set is small, so the ranges are wide.
- The photos come from Wikimedia Commons, and CLIP learned from a huge scrape of the web that probably included some of them. That could make Atlas look better here than on photos it has never seen.
- Atlas hasn't been trained on geography at all yet. This is the starting line that a trained Atlas will have to beat.