There's an episode in which he competes with an AI, but it turns out to be narrowing down locations using data most (if not all) humans couldn't. Imagine for example that the camera in a certain locale had some debris on it, but it's a unique shape and location on the lens. Or perhaps there's rain on the lens in various patterns. The AI neatly organizes these locations by their coordinates, then if it sees any other images with similar terrain and the same lens anomalies, it will guess with extremely high accuracy.

It left me wondering how effective it would be with current technology if it couldn't "cheat" in this way. I put cheat in quotations because in any situation where any metadata like this would be useful for location identification, it wouldn't matter how it worked. But, strictly using geological data, how would it perform?

> Imagine for example that the camera in a certain locale had some debris on it, but it's a unique shape and location on the lens.

I believe Facebook patented a method for using lens imperfections (dust, scratches, etc.) - see https://news.ycombinator.com/item?id=18835377

The patent office is going to change it so if a generic neural network put to a task learns the same technique as a human patent, the patent fails the ordinarily skilled in the art test.

Human geoguessr players use cheats like like less defects or other accidents with the captures.

You’re right, I totally neglected to point that out. I more so meant the AI is far better and doing this than humans are, and it’s wrong to say most/all humans can’t do it. We just can’t do it as well or as thoroughly as the AI can.

Rainbolt also uses these tricks wherever possible. Eg pro geoguesser players know how high the camera angle is on street view in each country, and which countries the edge of the Google car is visible on shot.

Definitely, that’s worth mentioning and I was arguably wrong to say most or all humans couldn’t do it. Many of the things the AI did really are things humans do. I think what I meant to say is that the depth and breadth in which the AI can do it is superhuman. Like, tiny bits of dust on the lens a person couldn’t really see become clearly and reliably identifiable features of locales. We can use aspect, seams, and other more obvious features of camera images in specific areas, but the AI can go quite a lot further using minute and sometimes almost imperceptible details.

This reminds me of the Tank neural net urban legend:

https://gwern.net/tank

He has a few videos of him vs AI (that is specifically trained to play geoguessr) and it's quite impressive already just at that game-focused scale.

> The scary part of this being possible is that I would think AI could do what he is doing even better than him

Not a given.

Grass is incredibly distinctive, I can usually pick out if a photo is taken in australia from the grass.

It was so (irrelevantly) annoying that in a recent forza game based in Australia that they got the grass so wrong and it was so obvious to me.

> This guy is accurate to a degree which I didn't think was even possible before seeing it.

People forget geobolt is a performer.