@@3446 That 2 month long project relies on the progress made by thousands of super smart scientists in the field of AI spending their lifetimes in research over the last century.
@@joshwhitley8969 Huh? No, a chess computer is extremely simple by comparson. It's literally just analysing every single possibility, even the really stupid moves. Chess has very few rules, and a single clear objective, and can easily be reduced to a few simple instructions. It's a very easy thing for a machine to be good at: Just brute-forcing all the possibilities and possible responses to responses to responses, down to a given depth.
@@ashscott6068no just no Until today no one has built a bot that knows every position in chess There is more position than atoms in earth you couldn't multiply it because the knight moves different the bishop does different pawn does different rook does different The highest bot today in chess in stockfish 16 the newest realease just beating stockfish 15
Thats how we used to beat the cheating extension people used to use a couple years back. It could see the data behind the pic and 5k it but it needed like 20 seconds, so if u guess within the first 15 it just didnt guess lol
That could be a way. Large AI models are pretty slow and need time. Just look at how long Bing sometimes takes to answer and they have a lot more power than some students I would guess.
@@mip4422 the zero shot means it guesses things it has never seen before without examples and the open domain shit means it can guess the location of anywhere in the world, not limited to specific places
@@Nova0Rock AIs can play more games of chess against itself than a grandmaster can play against any oponent, really, during a lifetime. I don't think so, if this AI can be trained using images, it will use basically every image available.
@@Nova0Rock He has been using geoguesser longer but an AI has basically no concept of time the way we do, they could train so much more in a couple minutes than we ever could in years. They said they showed the AI 250,000 images to train. There's no way rainbolt has seen over 250,000
You could just train an AI by making it play a billion times and do basic image recognition. You don't even need an AI for that, just save every image you see and map each of them to a set of coordinates.
I suppose it's a bit different, because every round in geoguessr has a 100% correct answer, whereas in chess, the best moves are calculated by the engine. It would be comparable if you wouldn't get the solution to a round in geoguessr and had to verify your guess manually
Main advantage probably has very little to do with the AI and more what they were talking about with sort of pretraining it by feeding it politically/geographically relevant cells that they picked out manually
17:08 - only in Belgium parking signs MUST be on an orange pole (that was only changed in traffic law in April 2023, so most will still be orange). Also the street surface, the sidewalks and the architecture are unmistakably Belgian (definitely NOT Danish).
The stupid kids don’t realize that they’re building technology for the current billionaires of the world who are use this technology to suppress the rest of society for the rest of human existence I mean, how dumb are these kids? They’ve created some thing that allows the police to track somebody from a simple photo I mean do they really think they’re helping society because the only ones they’re helping are the 0.01% that’s the only ones that they are helping I mean read the Bible this is demonic technology here this is the type of stuff the devil wants for complete global control .
The smudge thing makes me wonder if the camera can pick up on stuff like a region having 2% less bright photos than another similar region because it's an hour later in the day. If it hyperfocuses on such inhuman metas it should be pretty bad at transfering the skill to photos that imitate the google street car but are not original geoguessr footage.
@@richhobo1216 mhm I said post-grad. No doubt that undergrads would get a lot of offers though. Particularly with the state the comp sci industry is in rn, most people's dog could get a job somewhere if it promises to learn about machine learning.
@@zigotina I feel like you have to know how it's working to make it... The only way they might not know how it's working is that they don't know exactly what it looks for and exactly what "weighting" it puts on the nodes i.e. how important are telephone poles compared to how important grass colour is.
I do a bit of AI for fun, and I can almost guarantee that a huge amount of the accuracy comes from meta info. It might pick up on a tiny scratch on the camera that the human eye can't even see and instantly know that it can only be within the region where the mapping car had a scratch in that exact spot. It only needs a few almost impercievable clues like that to beat people every time. I'd be curious to see how it would perform on images taken with a mapping car that was never used in the training set
@Some Guy I’m not referring to actual image metadata, I mean meta info from a geoguessr sense. Referring to small hints in images such as an antenna showing up in a certain country or a rift in the sky due to improper image stitching in another
@Some Guy yup it’s pretty terrifying. I’m trying to keep my skills honed enough to where I can keep up when AI pair programming is industry standard, but it’s gonna get messy. I’m 25 so it’s not like I can hope to be retired before the real changes hit lol
Yeah would love to give it a photo I take from out on the street and see how accurate it is - I imagine it would be, but interesting to see what might start slipping it up
17:01 Quick tip: In Belgium, all poles that hold trafic signs concerning parking a vehicle, have to be in the orange color. This rule will change soon but they won't remove the old orange poles once it isn't a rule anymore. As far as I know, Belgium is the only country with this unique rule.
It's also kind of sad that one of the best humans stands no chance against the AI. A year ago he would win easily, now he's barely winning in Laos. This is just the beginning.
@@Ruzzky_Bly4t Why would it be sad? It's just the natural advancement of technology. They are programmed to be better at handling specific tasks than human beings and yet without human beings, AIs are dumb as fuck.
@@Ruzzky_Bly4t How is it sad? If anything, it's mighty impressive that a single human can be talented enough to almost stand shoulder to shoulder against an AI which is built using the collective knowledge provided by lifetime's worth of research by thousands of scientists who are most likely geniuses themselves!
This AI could be actually useful in some life or death situations where you need to locate someone as an example from an image but at the same time the AI could be used for bad things, like, stalking.
I am just an average computer scientist at almost 40 y/o, but seeing this young exceptionally bright folks really brightened my day. Back in that age (well it was a bit of a different time but still) I wouldn't even be CLOSE to be able to form such well structured thoughts and programs. Congrats and I am pretty sure all those young folks have a bright future ahead! But I am sure they already know.
Actually cs graduates back then were much better in average. Because now there's a lot of people flooding in cs in hope of getting a easy well paying tech job
Using the smudges on the camera is insanely clever, the same car will have the same smudges for quite a long time, perhaps even over multiple routes so by knowing them you can pinpoint the region. Absolutely insane, and a good example of how these models doesn’t “think” like we do.
It’s called meta and humans heavily rely on meta too, just different one. It’s really rarely about soil, vegetation and road signs or architecture. I watched competitions, they all notice first camera angles, camera types, reflections and distortions, post processing, weather, lighting and season, contrast and color temperature and all of that and judge location based on that. Dirt on camera used in some location, just humans are jot very great at noticing small ones, we are all trained to ignore them when we look through windows and screens
When it cut to the first shot of the two students leaning against the wooden wall, I legitimately said to myself, that one on the right is some bad CGI
Imagine a generation where people are using the same methods AI use to locate images. There's a lot from both worlds that look interesting in the future.
im so impressed with the mechanics of the AI...my roommate at college is a comp sci major and as a psych major even the lower level stuff looks crazy to learn!!! that being said this was a really awesome vid :) rainbolt's humor and the editing makes me laugh out loud
My biggest question is how well does the AI handle modifiers like blur, half screen, no panning or super low time. Also you should try and use the AI for the locating real photos series you do
Actually, on thinking about it, I think only blur would be an issue that would throw off the AI al the time. The others like half screen, no panning, and other things like upside down and mirrored and even black and white or scrambled could probably work out of the box, since the AI picks up on small things and I don't think it's bothered by where in the picture that detail is too much, or what colour.
Blur/half screen might work because the model wasn't trained on that. Panning wouldn't change much, it would just take 1 screenshot instead of 4. 'Super low time' doesn't really change anything because the AI takes a screenshot which is fed to the model. But even with blur/half screen it's just a matter of training the model on such images, which the creators can do very easily because they already have the image dataset that they can apply blur to or anything they want.
@@waltkowalsky4344 I think you're underestimating the impact that 4 pics vs 1 has. you saw in the video how much worse it did with 3 pics instead of 4 in the glitched round rainbolt almost won
@@rydenkaye9735 Well it wasn't just 3 pictures, it was 3 pictures plus an incorrect one. There is a saying that the only thing worse than no information is misinformation.
the fact the AI learns so quickly from going basically not knowing which continent to being able to 5k quite easily is astounding, but also scary as fuck. At some point you can no longer even keep up.
Its pretty much what happened to chess, we're now at the tipping point where AIs are starting to beat the best geo players, and eventually they'll be good enough to discover new metas that allow people to guess more accurately than they ever could now.
@@seventeen777 What lol chess engines first beat the world champion in 1997... A chess engine ran on my toaster would 10-0 Magnus Carlson with ease. Engines are no longer a competition they're the answer to what you should've done.
@Asdasdasa exactly what I said. I was using chess as an example of how geo ai will likely play out in the next couple decades, since so far it has followed the exact same path as chess
From my understanding of Ai fun fact: even if it has seen these locations before, it wouldn't be able to just 5000 them, because it doesn't retain data on individual locations, it only learns from them to improve its overall ability, it adjusts the patterns it looks for based on the information the locations provide
I don't know much about neural networks either, but from what I've understood overfitting and therefore making a model remember a dataset too well is definitely a thing.
@@klauspeter2199 Yes, but also could be highly inefficient, depending on what you want it to do and how much you want to brute-force it. Not to mention that it could make more mistakes if something doesn't totally fit its data. It's why most neural network AIs don't do that and they try to generalize from the beginning. Otherwise, you get an AI that learns to go to the green pixel instead of the exit sign.
I have a master's degree in CS / Machine learning... What they have done is incredibly impressive. I believe detecting the camera smudges though is an unfair example of over-fitting on the training data. The purpose of an AI model like this would be to take any photo of the countryside and geolocate it. Relying on the specific google street view car camera lens dirt, won't help it solve the general problem. I'm sure it can still do really well with other photos. But their explanation that it had a really small chance to have ever seen any of those specific photos is somewhat broken when a much larger percentage of those photos will have similar smudges.
Agree, it would be really interesting (in the sense of being useful for humans to learn from) if they trained on random geotagged photos, not specifically google street view photos, so that it can't learn any kind of meta.
they need to make this AI output the strongest feature combinations which led to its decision. Thus, the AI can train you to discover more metas, and then you will stand a better chance against it.
2K likes attained, so expecting the rematch of the human pros vs. the AI!!! Your reaction on every round was priceless, clinging to the hope you'd win at least one game!
@@casualfool122once they publish the paper it wouldnt take much to read it and re-implement it yourself if you have decent CS knowledge and had a good GPU. Once one person does that itll be readily available.
@@Pharoah2 maybe, but the chances of it spreading enough to where its readily available for stalkers wouldn't be extremely rapid by any means. and that's only if it actually gets spread at all
Color balance is probably something AI cues off of that is just impossible for humans to match. Some cameras might have a slight blue hue to them, or might have different levels of saturation, etc etc
@Some Guy actually, the military's technology isn't any better. mostly because developing AI like this costs a lot, takes a lot of time and has limited military uses. China has the facial recognition stuff, that's probably the closest equivalent after all, scientists figured out HOW to make an atom bomb years before they found the real method
I've got a feeling that it's subtly picking up on camera discrepancies to be very accurate country-wise. Maybe it can identify the noise pattern of the specific camera sensors, and has a large enough sample of locations that it's seen something from almost every Google street car.
Laos has very small coverage, and you need a relatively large dataset to train an AI that can classify images with this precision. Based on this, I assume the AI has seen almost all Laos locations and failed on the ones that weren't in the training set. Still very impressive, tho
I gotta say i am slightly disappointed by the fact that the AI used the streetview camera imperfections to guess. It kind of defeats the whole purpose of geoguessr and geolocation from a single impression, and means that the AI would perform much worse when used with photos taken with random cameras when trying to for example locate missing people.
I am fascinated and somewhat afraid of AI’s capabilities. Who knows, how advanced it could get. Something tells me AI will be coding AI soon. But overall a very well made and interesting video rainbolt. Loving the content and to see you and your friends so humble and happy.
I don't think people quite understand how crazy it is that this guy is taking rounds off of an AI, that is mindblowing, this guys brain needs to be looked at
If the AI is using camera smudges as a meta, that's a bummer cuz then it's really picking up on image artifacts from those specific google streetview photosets instead of understanding the actual location and scenery on a deeper level, so the application of the AI is only limited to the game and not wider uses like digital forensics and such. That said, I'm sure it has other metas that don't include just photoset specific artifacts, but I wonder to what extent.
This happens in most AI problems involving images, its understanding differences in imaging data noise signatures that a human would not be able to detect
I propose they name the Pigeon AI "Lockbird" both in honor of Stockfish 15 (best Chess AI) *and because the bird was more locked-in than any human I've seen.* This new AI has the potential to be just like Stockfish which helps Chess players learn new things about chess via post-game analysis of their online or OTB games. Great video, and great work Stanford students! [Geogusser AI fan-boy warning]: Lockbird will probably many times better than any human by Lockbird 15 [if development continues for as long as Stockfish has been developed that is] just like Stockfish 15 already is.
Assuming this will also work on custom images and not just street view images, I wonder if these students realized that they have just invented a way for anyone to locate the origin of pretty much any outdoor picture or video since the beginning of time. It will take a lot of tweaking, and the farther back you go to the harder the model training will get, but that would be the next logical step for this technology. It’s honestly hard to fathom.
this truly was so amazing to see. ai has improved and is capable of doing so much. this is very impressive. well done as well, though, rainbolt! humans vs ai in geoguessr must continue, defend humanity's honor!
I do have to say, if you want to get into neural networks, this would be the perfect project to start out on. It is incredibly simple to get to work (maybe not to work well, but to work at all) Basically, you have an input image, like in many neural networks you can see online, and you output an x and y coordinate. And even better: you could not ask for a better dataset for training than what google or microsoft have to offer, I mean there are billions of images, it's amazing. So perfect starter project!
for Brazil, I don't know if is feasible or not to learn but the area code from phone numbers are a two digit number prefix right at the start. So here at 3:46 bellow "Racco" there's the 61 area code from the phone number, which means this is from Goiás/Federal District area.
Wow! It's like watching Magnus Carlsen play stockfish! He could give a nice fight but all can he hopes is to make a draw! And following this analogy, the last game is like you started the game with a piece odd! So fassinating and beautiful! Can't wait for the next video!
5:31excellent call -- also worth noting, if you ever see a stop sign that says "STOP ARRET" on it, you are guaranteed to be in New Brunswick specifically
It's insane how the brain works. At 16:55, RAINBOLT is stumped, and then there's me, somebody who barely plays geoguessr going: "This looks oddly familiar... Maybe Belgium?" I can't explain how I knew, my brain just fired a few neurons when the image showed lol.
You can always improve performance by just scaling up the model, but using CLIP is probably the best architecture they could use at the moment so no obvious ways to create big jumps. (Besides giving it more images than 4 to guess from and just training it a lot more)
would be interesting to see a rematch not using geogessr, but just photos of different places. then the AI training data might become much less useful (like the smudges), so it would be interesting to see how well that knowledge transfers.
Maybe I heard wrong, but didn't they say at some point that they didn't based it on Geoguessr but on random Google Streetview? That would mean that it would do even better
Thanks I really needed the label for which one was which in the thumbnail considering rainbolt does have the stature of Arnold Schwarzenegger. Love the video ❤
1997: Deep blue beats Garry Kasparov at chess
2023: Stanford AI beats Rainbolt at geoguessr
It was a 2 month long project. Rainbolt took years.
@@3446 That 2 month long project relies on the progress made by thousands of super smart scientists in the field of AI spending their lifetimes in research over the last century.
Not even a comparison. Gary chess way more impressive.
@@joshwhitley8969 Huh? No, a chess computer is extremely simple by comparson. It's literally just analysing every single possibility, even the really stupid moves. Chess has very few rules, and a single clear objective, and can easily be reduced to a few simple instructions. It's a very easy thing for a machine to be good at: Just brute-forcing all the possibilities and possible responses to responses to responses, down to a given depth.
@@ashscott6068no just no
Until today no one has built a bot that knows every position in chess
There is more position than atoms in earth you couldn't multiply it because the knight moves different the bishop does different pawn does different rook does different
The highest bot today in chess in stockfish 16 the newest realease just beating stockfish 15
pressuring an AI model to push out a prediction by inputting your answer quickly is the funniest shit I have ever heard.
I mean if the AI needed more time to upload and make the inference then yes it would work, but yeah not in the "pressure" way lol
Thats how we used to beat the cheating extension people used to use a couple years back. It could see the data behind the pic and 5k it but it needed like 20 seconds, so if u guess within the first 15 it just didnt guess lol
Is just limiting the ai processing time.
@@gabrielamaral978 18 sec ago
That could be a way.
Large AI models are pretty slow and need time.
Just look at how long Bing sometimes takes to answer and they have a lot more power than some students I would guess.
No way Tom Holland and Shawn Mendes collabed to make an AI to curb stomp your ass in Geoguessr
fr 😍😍
Mf this is so funny😂
😂
tom holland from ohio
The tom holland-lookin guy has such a look, I love it. He looks like someone straight out of a movie.
for anyone wondering, the paper is called "Learning Generalized Zero-Shot Learners for Open-Domain Image Geolocalization"
thanks bro
what is "learning generalized zero-shot learners" even supposed to mean?
I think they made that up just to make it sound cool😂
@@mip4422 the zero shot means it guesses things it has never seen before without examples and the open domain shit means it can guess the location of anywhere in the world, not limited to specific places
@@deadshot_ed1tsit's genuine
I've never seen this guy before but the fact that he could almost beat the AI trained on so many pictures is absolutely insane.
At this point in time, he has probably seen more ;)
@@Nova0Rock AIs can play more games of chess against itself than a grandmaster can play against any oponent, really, during a lifetime. I don't think so, if this AI can be trained using images, it will use basically every image available.
@@Nova0Rock He has been using geoguesser longer but an AI has basically no concept of time the way we do, they could train so much more in a couple minutes than we ever could in years. They said they showed the AI 250,000 images to train. There's no way rainbolt has seen over 250,000
He has actually beaten AI before.
The game was rigged against the AI there lmao
AI really just said "I recognise that dead insect on the camera"
It has found its own obscure meta and Trevor wants it to share the meta :)
😂😂
😂😂
Avdol
@@normal5389 was waiting for this lol
So much to learn from analysing what the AI is actually looking at in the image
meta players smh /s
Definitely gonna be getting a a bunch of new dirt on camera metas
The ai actually just scanning the place thats it 💀
You could just train an AI by making it play a billion times and do basic image recognition. You don't even need an AI for that, just save every image you see and map each of them to a set of coordinates.
@@Jonassoe and when new coverage comes out you're fucked
Dang, geoguessr is about to have engine prep just like chess
That’s what I was thinking. Pigeon is the stock fish of geoguesser
This is actually crazy. It’s scary if malicious people get this tech to track people.
Ohh wow! I never thought of that😮
I suppose it's a bit different, because every round in geoguessr has a 100% correct answer, whereas in chess, the best moves are calculated by the engine. It would be comparable if you wouldn't get the solution to a round in geoguessr and had to verify your guess manually
@@gingeral253 Shit you got a point, this might seems actualy malicious tool hidden behind harmless fun.
4:34 im convinced dude with the glasses is also an AI just how he sits there like that
Bro's glasses are wiggling too
It’s truly impressive that it takes a specially trained AI to beat this guy
How do you think ai is made at all. You have to train it
@@starxdreamz. I think they meant that it was specifically made for that purpose
The AI got so much better since the last AI video, this is really impressive. Good job to those guys!
Yo
and how much better did this guy get ;D
Main advantage probably has very little to do with the AI and more what they were talking about with sort of pretraining it by feeding it politically/geographically relevant cells that they picked out manually
@@gavinjenkins899That's how most child prodigies are made.
I love Rainbolt but honestly watching him just get decimated and seeing him spiral is actually really funny 🤣🤣🤣
now he knows the feeling of being a mortal like us lmao
@@njbrx woooooooooooooah
for real lmao, especially during the cambodia game
@@soyanshumohapatra no u
@@KianTheeGreat what?
17:08 - only in Belgium parking signs MUST be on an orange pole (that was only changed in traffic law in April 2023, so most will still be orange). Also the street surface, the sidewalks and the architecture are unmistakably Belgian (definitely NOT Danish).
2 am moment
@@georainbolt as you always say, "cope" 😂
🤓
also the sign that says "20m" is only in belgium
The stupid kids don’t realize that they’re building technology for the current billionaires of the world who are use this technology to suppress the rest of society for the rest of human existence I mean, how dumb are these kids? They’ve created some thing that allows the police to track somebody from a simple photo I mean do they really think they’re helping society because the only ones they’re helping are the 0.01% that’s the only ones that they are helping I mean read the Bible this is demonic technology here this is the type of stuff the devil wants for complete global control .
21:24 AI knew exactly where it was but chose not to be on the bridge
The smudge thing makes me wonder if the camera can pick up on stuff like a region having 2% less bright photos than another similar region because it's an hour later in the day. If it hyperfocuses on such inhuman metas it should be pretty bad at transfering the skill to photos that imitate the google street car but are not original geoguessr footage.
These guys are going to go far . Tech companies will be all over comp sci graduates with AI knowledge . Straight to the top
True, but I think tech companies are all over a lot of post-grad/PhD Stanford students lmao
@@jjpswfc yeah just graduating from Stanford with a bachelors would get a ton of job offers, much less being a graduate/PhD student
they don't even know how is it working
@@richhobo1216 mhm I said post-grad. No doubt that undergrads would get a lot of offers though. Particularly with the state the comp sci industry is in rn, most people's dog could get a job somewhere if it promises to learn about machine learning.
@@zigotina I feel like you have to know how it's working to make it... The only way they might not know how it's working is that they don't know exactly what it looks for and exactly what "weighting" it puts on the nodes i.e. how important are telephone poles compared to how important grass colour is.
I do a bit of AI for fun, and I can almost guarantee that a huge amount of the accuracy comes from meta info. It might pick up on a tiny scratch on the camera that the human eye can't even see and instantly know that it can only be within the region where the mapping car had a scratch in that exact spot. It only needs a few almost impercievable clues like that to beat people every time. I'd be curious to see how it would perform on images taken with a mapping car that was never used in the training set
It should work on custom images. So you can find out what regions it only guesses on meta va what it knows. Their has to be 1 region that is pure meta
@Some Guy I’m not referring to actual image metadata, I mean meta info from a geoguessr sense. Referring to small hints in images such as an antenna showing up in a certain country or a rift in the sky due to improper image stitching in another
@Some Guy yup it’s pretty terrifying. I’m trying to keep my skills honed enough to where I can keep up when AI pair programming is industry standard, but it’s gonna get messy. I’m 25 so it’s not like I can hope to be retired before the real changes hit lol
I mean that's what he showed in the end of the video
Yeah would love to give it a photo I take from out on the street and see how accurate it is - I imagine it would be, but interesting to see what might start slipping it up
Its crazy that the only time we can humanize rainbolt is when he is against an AI 🤣
The sheer innate confidence of “we don’t really see any improvements we need to make right now”
no he said that they dont see any immediate improvements available. theyve reached near the peak of what their methodology is capable of
It doesn't mean the ai can't be better, they know it can be better they just don't see any obvious ways how to make it better yet.
"I am the ultimate lifeform"
17:01 Quick tip: In Belgium, all poles that hold trafic signs concerning parking a vehicle, have to be in the orange color. This rule will change soon but they won't remove the old orange poles once it isn't a rule anymore. As far as I know, Belgium is the only country with this unique rule.
The Spain guess when it was Italy at 12:25 is a perfect represation of never doubt yourself or whatever the saying was
"trust your gut"
@@miniepicness thanks i already forgot what it was
@@ahhkaraj You should've trusted your gut
Nah but his gut was also going to Spain to more that’s why he voted if he knew it was Italy more then he would’ve guessed it
It’s still impressive how good rainbold is. He can hold a dual with a fucking AI super pc, even if he loses.
It's also kind of sad that one of the best humans stands no chance against the AI. A year ago he would win easily, now he's barely winning in Laos. This is just the beginning.
@@Ruzzky_Bly4t Why would it be sad? It's just the natural advancement of technology. They are programmed to be better at handling specific tasks than human beings and yet without human beings, AIs are dumb as fuck.
We can compare this like Gary Kasparov and Blueray Supercomputer chess match
@@Ruzzky_Bly4t How is it sad? If anything, it's mighty impressive that a single human can be talented enough to almost stand shoulder to shoulder against an AI which is built using the collective knowledge provided by lifetime's worth of research by thousands of scientists who are most likely geniuses themselves!
Rainbold😮💨
This AI could be actually useful in some life or death situations where you need to locate someone as an example from an image but at the same time the AI could be used for bad things, like, stalking.
Not really. It average 44km which is way too much to pinpoint someone. Maybe if it gets to under 3km
@@Skyl3t0n that's better than not knowing the general area to search in
@@Skyl3t0n getting within 44km of osama bin laden would have been very helpful
@@Skyl3t0n And it's also only using 4 images each guess with a very strict time limit to process them
Like seriously the application of this outside of just geoguessr is just massive
I am just an average computer scientist at almost 40 y/o, but seeing this young exceptionally bright folks really brightened my day. Back in that age (well it was a bit of a different time but still) I wouldn't even be CLOSE to be able to form such well structured thoughts and programs. Congrats and I am pretty sure all those young folks have a bright future ahead! But I am sure they already know.
Actually cs graduates back then were much better in average. Because now there's a lot of people flooding in cs in hope of getting a easy well paying tech job
Using the smudges on the camera is insanely clever, the same car will have the same smudges for quite a long time, perhaps even over multiple routes so by knowing them you can pinpoint the region. Absolutely insane, and a good example of how these models doesn’t “think” like we do.
I'd call it a side effect. The AI doesn't "know" what it's doing
@@frankman2they know what they doing. they have objective. and only care about that
It’s called meta and humans heavily rely on meta too, just different one. It’s really rarely about soil, vegetation and road signs or architecture. I watched competitions, they all notice first camera angles, camera types, reflections and distortions, post processing, weather, lighting and season, contrast and color temperature and all of that and judge location based on that. Dirt on camera used in some location, just humans are jot very great at noticing small ones, we are all trained to ignore them when we look through windows and screens
Are you sure the students aren't robots themselves?! I mean they were completely deadpan, confident in the powers of the monster they'd built!
Bro don't do people on the spectrum that dirty😢
@@OfficialCANVASlol
@@OfficialCANVAS why are you assuming smart people are automatically autistic? that's a harmful stereotype
When it cut to the first shot of the two students leaning against the wooden wall, I legitimately said to myself, that one on the right is some bad CGI
literslly man... the one kid with the screen to himself looks completely like an animation
Thats no AI. Its your lost twin that they have kidnapped and trapped in the basement.
Okay, this is it! 😂
Imagine a generation where people are using the same methods AI use to locate images. There's a lot from both worlds that look interesting in the future.
I find it hilarious and extremely talented how most the time your first waypoint and guess is dead on but you move it second guessing
im so impressed with the mechanics of the AI...my roommate at college is a comp sci major and as a psych major even the lower level stuff looks crazy to learn!!! that being said this was a really awesome vid :) rainbolt's humor and the editing makes me laugh out loud
My biggest question is how well does the AI handle modifiers like blur, half screen, no panning or super low time. Also you should try and use the AI for the locating real photos series you do
I don't expect they could adapt to that out of the box.
Edit: spelling
Actually, on thinking about it, I think only blur would be an issue that would throw off the AI al the time. The others like half screen, no panning, and other things like upside down and mirrored and even black and white or scrambled could probably work out of the box, since the AI picks up on small things and I don't think it's bothered by where in the picture that detail is too much, or what colour.
Blur/half screen might work because the model wasn't trained on that. Panning wouldn't change much, it would just take 1 screenshot instead of 4. 'Super low time' doesn't really change anything because the AI takes a screenshot which is fed to the model.
But even with blur/half screen it's just a matter of training the model on such images, which the creators can do very easily because they already have the image dataset that they can apply blur to or anything they want.
@@waltkowalsky4344 I think you're underestimating the impact that 4 pics vs 1 has. you saw in the video how much worse it did with 3 pics instead of 4 in the glitched round rainbolt almost won
@@rydenkaye9735 Well it wasn't just 3 pictures, it was 3 pictures plus an incorrect one. There is a saying that the only thing worse than no information is misinformation.
Sorry I can't tell why but those two AI guys look like some supervillains
ikr, I was waiting for the dude with the glasses and piled up hair to steeple his fingers and say "soon the world will be mine, bwahahahaha!"
why is their skin so perfect too
i hate them so much
ai filter@@buildintotrains
Glasses and the sweatshirt
the fact the AI learns so quickly from going basically not knowing which continent to being able to 5k quite easily is astounding, but also scary as fuck. At some point you can no longer even keep up.
Its pretty much what happened to chess, we're now at the tipping point where AIs are starting to beat the best geo players, and eventually they'll be good enough to discover new metas that allow people to guess more accurately than they ever could now.
@@seventeen777 What lol chess engines first beat the world champion in 1997... A chess engine ran on my toaster would 10-0 Magnus Carlson with ease. Engines are no longer a competition they're the answer to what you should've done.
@Asdasdasa exactly what I said. I was using chess as an example of how geo ai will likely play out in the next couple decades, since so far it has followed the exact same path as chess
@@seventeen777 Sorry thought you meant that the tipping point was in chess and not geo
Yeah, this will happen with literally everything. Human supremacy is coming to an end, we’re being replaced in every aspect aside from manual labor
Does anyone else think the dudes look like they have ai faces?
Idk why but seeing Rainbolt make a better guess than the AI is so exciting
You don’t know why
16:20 is literally the house I grew up in. How spooky is that!
BRO
Those last rounds are basically Terminator 3, the Termanitor teaming up with the humans against a newer, better version of itself. :D
The students faces are also AI generated, nice
Lmao my thought exactly
nah thats just how germans look
I thought the same thing lol
From my understanding of Ai fun fact: even if it has seen these locations before, it wouldn't be able to just 5000 them, because it doesn't retain data on individual locations, it only learns from them to improve its overall ability, it adjusts the patterns it looks for based on the information the locations provide
I don't know much about neural networks either, but from what I've understood overfitting and therefore making a model remember a dataset too well is definitely a thing.
Wrooooooooong
@@TheJulianmc it is? Could you elaborate?
@@klauspeter2199 Yes, but also could be highly inefficient, depending on what you want it to do and how much you want to brute-force it. Not to mention that it could make more mistakes if something doesn't totally fit its data. It's why most neural network AIs don't do that and they try to generalize from the beginning. Otherwise, you get an AI that learns to go to the green pixel instead of the exit sign.
Nah, but making an AI that can make an educated guess from the smudges on the camera is genuinely insane, props to those guys for making such a beast
But then it's not geoguessing anymore, it's camera smudge guessing
Now take the 3 AIs you've played and let them team up in a pro tournament.
Nah just 1. The AI is cracked and need the nerf
i feel like the AI would do a frighteningly good job with the dirt maps
I liked the bit where the AI went "It's intelligencing time" and intelligenced all over Rainbolt.
Artificially of course.
Why do the researchers look like they were generated by AI
they don't
@@regd6060they do honestly
@@regd6060 they do
Thats crazy fr
@@regd6060 they do
I have a master's degree in CS / Machine learning... What they have done is incredibly impressive. I believe detecting the camera smudges though is an unfair example of over-fitting on the training data. The purpose of an AI model like this would be to take any photo of the countryside and geolocate it. Relying on the specific google street view car camera lens dirt, won't help it solve the general problem. I'm sure it can still do really well with other photos. But their explanation that it had a really small chance to have ever seen any of those specific photos is somewhat broken when a much larger percentage of those photos will have similar smudges.
Agree, it would be really interesting (in the sense of being useful for humans to learn from) if they trained on random geotagged photos, not specifically google street view photos, so that it can't learn any kind of meta.
@@jonbbbb Thanks, glad someone read my comment.
@@jonbbbb I agree.
"my geoguessr career is obviously over now so if you guys have any ideas for what i could do..." LMAO
they need to make this AI output the strongest feature combinations which led to its decision.
Thus, the AI can train you to discover more metas, and then you will stand a better chance against it.
2K likes attained, so expecting the rematch of the human pros vs. the AI!!! Your reaction on every round was priceless, clinging to the hope you'd win at least one game!
15:29 "If you can't beat em', Join em' Germany." - Rainbolt 2023
😂
bro got cooked the frustration on his face when he loses the cambodia match to pigeon🤣
That AI has a future as a tool for detective work. That's crazy.
It also has a future for stalking purposes.
@@Ruzzky_Bly4t probably not for public use if I was to guess bro
@@Ruzzky_Bly4t i mean u can get the person's location from exif data, u don't even need AI i.e. recognition to get someone's location off his pics
@@casualfool122once they publish the paper it wouldnt take much to read it and re-implement it yourself if you have decent CS knowledge and had a good GPU. Once one person does that itll be readily available.
@@Pharoah2 maybe, but the chances of it spreading enough to where its readily available for stalkers wouldn't be extremely rapid by any means. and that's only if it actually gets spread at all
"Do I feel good about that? No that feels like such an underhand pitch."
*Kiwis looking over at the Ozzies* 😂
Watching trevor slowly being more upset/sweaty is so funny i dont even know how lol
It’s funny till it hits ur workplace
Color balance is probably something AI cues off of that is just impossible for humans to match. Some cameras might have a slight blue hue to them, or might have different levels of saturation, etc etc
this can easily translate into crime investigation. missing persons, etc. trying to find someones location with only an image, or video. amazing work.
The unfortunate part about this is that this probably could be the end of AIs in tourneys, which was really fun, but now they're too good...
Could still be doable with a team of humans so you get a little hedging going.
@@Gameboygenius just give up theyve replaced us 🤖🦾
@@Gameboygenius Give the AI another 2 years, and it will probably destroy 10 pro players at once.
@Some Guy actually, the military's technology isn't any better. mostly because developing AI like this costs a lot, takes a lot of time and has limited military uses.
China has the facial recognition stuff, that's probably the closest equivalent
after all, scientists figured out HOW to make an atom bomb years before they found the real method
well it has been the case for a few years now already (cf Deep Blue for example)
I've got a feeling that it's subtly picking up on camera discrepancies to be very accurate country-wise. Maybe it can identify the noise pattern of the specific camera sensors, and has a large enough sample of locations that it's seen something from almost every Google street car.
Laos has very small coverage, and you need a relatively large dataset to train an AI that can classify images with this precision. Based on this, I assume the AI has seen almost all Laos locations and failed on the ones that weren't in the training set. Still very impressive, tho
Dude it's an AI AND it's 4 very smart minds putting that ai together against 1. You did amazing bro
Saying that today for this kind of test shows how long this technology came. No one would believe it possible a couple decades ago.
I gotta say i am slightly disappointed by the fact that the AI used the streetview camera imperfections to guess. It kind of defeats the whole purpose of geoguessr and geolocation from a single impression, and means that the AI would perform much worse when used with photos taken with random cameras when trying to for example locate missing people.
wanna see stique playing against it
IN Albania :D
I am fascinated and somewhat afraid of AI’s capabilities. Who knows, how advanced it could get. Something tells me AI will be coding AI soon. But overall a very well made and interesting video rainbolt. Loving the content and to see you and your friends so humble and happy.
Training the AI on 1 in 1000 street view images definitely does not count as "virtually impossible" that it's ever seen any of these rounds before.
I don't think people quite understand how crazy it is that this guy is taking rounds off of an AI, that is mindblowing, this guys brain needs to be looked at
If the AI is using camera smudges as a meta, that's a bummer cuz then it's really picking up on image artifacts from those specific google streetview photosets instead of understanding the actual location and scenery on a deeper level, so the application of the AI is only limited to the game and not wider uses like digital forensics and such. That said, I'm sure it has other metas that don't include just photoset specific artifacts, but I wonder to what extent.
That's what I was thinking too, it can win at geoguesser but it may depend on these meta informations instead of the actual location
Its objective is to get the highest score I think so its going to go meta just like a human would.
This happens in most AI problems involving images, its understanding differences in imaging data noise signatures that a human would not be able to detect
I propose they name the Pigeon AI "Lockbird" both in honor of Stockfish 15 (best Chess AI) *and because the bird was more locked-in than any human I've seen.*
This new AI has the potential to be just like Stockfish which helps Chess players learn new things about chess via post-game analysis of their online or OTB games. Great video, and great work Stanford students!
[Geogusser AI fan-boy warning]:
Lockbird will probably many times better than any human by Lockbird 15 [if development continues for as long as Stockfish has been developed that is] just like Stockfish 15 already is.
This is wild... really great work from these guys honestly that must feel hella rewarding
11:35 bro the AI will get revenge for that later
"guys I sat there and walked every single street" 🤣🤣 I'm rolling
Assuming this will also work on custom images and not just street view images, I wonder if these students realized that they have just invented a way for anyone to locate the origin of pretty much any outdoor picture or video since the beginning of time. It will take a lot of tweaking, and the farther back you go to the harder the model training will get, but that would be the next logical step for this technology. It’s honestly hard to fathom.
I really want to see a game with like 10 pros against this AI. Would be the ultimate showdown
These guys probably have a ready-to-go company when they're done with school
We're going to need to assemble the Geoguessr Avengers to take on this Ultron.
5:39 "pigeon, is a great guess there"
gotta love this dude..
His knowledge, skill, however you may like to name it, it's somehow frightening
this truly was so amazing to see. ai has improved and is capable of doing so much. this is very impressive. well done as well, though, rainbolt! humans vs ai in geoguessr must continue, defend humanity's honor!
5 GeoGusesr Pros vs 1 AI is such a great idea
11:20 dude really trashtalked AI 💀
Yo that was crazy when he picked Spain over Italy and lost, I've lived in Spain most of my life and felt 100% confident that was Spain 😮
8:56 rainbot does not know how to comprehend losing geoguessr
'Could be finland but I like sweden' that's how to guess the location right there
Love the trash talking of the AI LOL
I don't know why but I laughed so hard when he said that they were "quivering in their boots" lmao
I do have to say, if you want to get into neural networks, this would be the perfect project to start out on. It is incredibly simple to get to work (maybe not to work well, but to work at all)
Basically, you have an input image, like in many neural networks you can see online, and you output an x and y coordinate. And even better: you could not ask for a better dataset for training than what google or microsoft have to offer, I mean there are billions of images, it's amazing. So perfect starter project!
for Brazil, I don't know if is feasible or not to learn but the area code from phone numbers are a two digit number prefix right at the start. So here at 3:46 bellow "Racco" there's the 61 area code from the phone number, which means this is from Goiás/Federal District area.
“This is just gonna be North Turkey, surely… like… where else would this be?” -Geo 😂
probs people said already but looking at what the ai sees for tough countries could unlock tons of new metas!
Wow! It's like watching Magnus Carlsen play stockfish! He could give a nice fight but all can he hopes is to make a draw!
And following this analogy, the last game is like you started the game with a piece odd!
So fassinating and beautiful! Can't wait for the next video!
Good for Kelvin from Sons of the Forest to be a part of this brilliant team. Some intelligence between them.
5:31excellent call -- also worth noting, if you ever see a stop sign that says "STOP ARRET" on it, you are guaranteed to be in New Brunswick specifically
It's insane how the brain works.
At 16:55, RAINBOLT is stumped, and then there's me, somebody who barely plays geoguessr going: "This looks oddly familiar... Maybe Belgium?" I can't explain how I knew, my brain just fired a few neurons when the image showed lol.
11:28 lmao he lost three games in a row and AI got one wrong he said AI is so dumb its so funny
You can always improve performance by just scaling up the model, but using CLIP is probably the best architecture they could use at the moment so no obvious ways to create big jumps. (Besides giving it more images than 4 to guess from and just training it a lot more)
The whole video I sat there nail biting and hoping Rain would at least get one W!
Rainbolt is a genuine cool nice polite well spoken dude he deserves a ton and glad he’s getting it.
'This is north Switzerland, obviously.' Man, watching me fumble around with this game would make you cry in frustration. And the AI is even scarier.
would be interesting to see a rematch not using geogessr, but just photos of different places. then the AI training data might become much less useful (like the smudges), so it would be interesting to see how well that knowledge transfers.
Maybe I heard wrong, but didn't they say at some point that they didn't based it on Geoguessr but on random Google Streetview? That would mean that it would do even better
@@Kochen51 geoguessr is streetview
@@ORCIproductions yes but it's a database of some of the locations, not every point in the world, right? The spawn points
I’m sorry but this dude at 1:34 looks like an AI human 😂
Pretty sure this is the part where Rainbolt goes into hiding forever…
Rainbolt is the kind of man, that if you were to kidnap him, he would immediately say "uhm, I'm going to guess Costești Romania locking it in."
Thanks I really needed the label for which one was which in the thumbnail considering rainbolt does have the stature of Arnold Schwarzenegger. Love the video ❤