Recognition games
I think of MetGuessrĀ as a kind ofĀ recognition game.
These games pose a recognition task over items sampled from a large and varied set of elements in some domain.
They are a bit like the high-dimensional classification tasks posed by object detection benchmarks such as Imagenet - except you’re the classification function being evaluated.
Recognition games can also be a way toĀ exploreĀ andĀ discoverĀ things. If the domain is inherently rich and interesting, recognition games can be quite rich and interesting experiences themselves.
GeoGuessr
GeoGuessr is a recognition game with over 50 million registered users. In GeoGuessr, you’re dropped into a location sampled from Google Street View. Your task is to guess where exactly on the earth you are by zooming and clicking on a Google maps interface. Your guess is scored by a function that maps the distance of your clicked spot from the exact answer location into a 0 - 5000 point range.
At first, this seems like an impossible task. And when you play, rounds will come up that are frustratingly difficult. Like when you land on a dirt road in an open field with nondescript fauna, no signs, and no obvious topological features.
Just as often as you’re stumped, though, you surprise yourself.
Much of the game’s magic lies is in the way you can win points solely based on vibes and intuition. Whether it’s because you’ve been there, or because you’ve internalized enough representations of a place via media consumption, if you’re reading this blog post, chances are good you have an uncanny, inexplicable feel for when you’re in rural Scandinavia, or urban Mexico, or in the winding backroads of an island nation in… southeast Asia?
Because the scoring function leaves ample room for partial credit, the game rewards and hones this vibes-based recognition ability. However, it also rewards close reading and deductive sleuthing. You can zoom into signs to see scripts and languages.
(((more about logical / deductive approach here)))
More games
Many recognition games follow a design pattern established by Wordle.
Wordle is technically a recognition game; the set of items from which it samples are “4 letter words.” The skill it hones is recalling more 4 letter words that meet the properties implied by the feedback the game has already provided given your failed attempts.
While fun and beloved by many, Wordle’s core task seems slightly arbitrary to me. It’s not working a recognition muscle that has relevance outside of improving your ability to play Wordle, save perhaps playing other games like crossword puzzles. The domain it helps you explore is not particularly rich or interesting.
InĀ Tradle, you’re asked to recognize a mystery country given a visualization of its trade export data. within 6 guesses, providing feedback for incorrect guesses.
Hexcodle gives you 5 attempts to guess the hex code of a given color, providing feedback to tune your guess along the way.
Worldle follows the Wordle model, but the task is to recognize a country based on its geographical shape. The developers seem to have built over 10 other games. Some follow the geoguessr model with slight twists, while others hew closer to Wordle.
Heardle is an ad-ridden attempt at a music recognition game, but poorly executed. Bandle takes a more technical approach, exposing different instrumental layers of a song. Neither of these are particularly compelling for a domain with so much potential…
InĀ graphs.world, you’re shown a data visualization and asked to guess which dataset the graph represents, given a few multiple choice options. Playing encourages you to think critically about the shape of data generating processes over time - a kind of social scientific intuition that spans a variety of domains and social phenomena.
Font games displays a short text rendered in a particular font. You’re given two font names to choose from. The game seems to sharpen an intuition about the relationship betwen the names of fonts and their appearance.
There are a lot of recognition games left to build
Virtually every domain that is constituted by a high-cardinality set of artifacts could support recognition games.
Here are some I’m eying for my own purposes:
- Architecture
- Typography & Graphic design
- Color theory
- Music
- Trees
Shadow curricula
For most of these games, game play is essentially equivalent to taking a test. Repeatedly taking tests does seem to grow our recognition skills. Rainbolt, the viral GeoGuessr wizard, seems to have played the game in the way that a typical Juilliard student practices - 4-5 hours per day as a baseline, sometimes playing as much as 12 hours per day.
But it’s not clear that playing for hundreds of hours really the most efficient route to improving our recognition skills in a domain. In interviews, Rainbolt also mentions study sheets, flashcards, and resources created by the community. GeoGuessr itself does not
TODO
- Check out Song Trivia