# 13

This post briefly reflects on applications of rationality techniques to the game Geoguessr, of which I am a fan (though not exceptionally talented, having only played a little bit). Videos from YouTuber GeoWizard show the reasoning process of a high-skill Geoguessr player.

# Intro to Geoguessr

Geoguessr is an online game in which one is shown a photo from Google Street View and attempts to guess (by clicking on a world map) where the photo was taken.

The difficulty of the game can be modulated by changing:

• The time limit for guesses
• The types of movement, if any, allowed
• The maps from which photos are taken

# Outside View Reasoning

Outside view reasoning is very helpful to playing Geoguessr well. In the broadest possible gametypes, where the photo is not geographically limited by the game, outside-view reasoning can helpfully narrow down guesses quite rapidly.

For example, since photos are drawn from Google Street View, and not the set of all possible street-level photographs, knowing the distribution of Google Street View availability will help immensely. Thus, "China" is a bad guess, since Google Street View has very limited availability there. Correspondingly, if one sees Chinese writing in a photograph (and one does not appear to be in a Chinatown), "Taiwan" is a much better guess.

Similarly, outside view reasoning would suggest that the distribution of Google Street View photos is biased towards countries with more population, land mass, or economic activity. Thus, if a player thinks the photo is either (say) Germany or Luxembourg, "Germany" is of course a much better guess (though I have gotten Luxembourg!). Similarly, if you know the picture is from Illinois, the area around Chicago is a better guess, probabilistically, than elsewhere in the state.

# Inside View Reasoning

Geoguessr also forces good practice of updating. The above information is good in setting priors, but of course the main way to be good at the game is to guess well based on available information.

In some cases, you will be lucky and see something that gives strong evidence of a location (e.g., a flag, a place name, a distinctive written language). But in many cases, one can begin updating based on multiple, indirect sources of evidence.

Obvious ones include things like:

• What climate does the place appear to have?
• What are the flora like?
• What's the topography of the place?
• How rich is the place?
• What language(s) are present?
• Which side of the street do people drive on?
• What types of vehicles do people drive?
• What types of street signs does the area have?
• What do the people look like?
• Which direction is sunlight coming from?

By aggregating these types of information in a foxlike way (and by knowing world geography!), a skilled player (e.g., GeoWizard, not me) can often make an informed guess of the country upon seeing a minimally informative picture.

Skilled players also learn spurious correlations on Street View. For example, I learned from GeoWizard that Street View cars in Kenya often have a snorkel on them:

Learning these correlations, though not geographically relevant, makes one a better player.

# Thinking Fast and Slow

Adding time or movement restrictions to a game makes the game harder. It also elicits interesting System 1-System 2 dynamics.

Photos can unconsciously elicit a place. If one cannot move or find useful new information in time (e.g., because one is on a remote road with only fields in sight), System 2-type guesses ("this feels like Russia") may be all that one can rely on.

# If you want to play ...

I've found Geoguessr to be a fun alternative to prediction markets/tournaments. If you'd like to play with me, I am here.