pro geoguessr player vs ai, it didn't go well...

Поділитися
Вставка
  • Опубліковано 16 гру 2024

КОМЕНТАРІ • 1 тис.

  • @lesliehamilton2020
    @lesliehamilton2020 2 роки тому +12375

    "Honestly not a bad guess by the AI there." - Says the Robot....

    • @bendzsi3780
      @bendzsi3780 2 роки тому +154

      Yeah, Zuck's little brother! XD

    • @lesliehamilton2020
      @lesliehamilton2020 2 роки тому +48

      @@ASTOGK The AI has become sentient.

    • @jonaswolterstorff3460
      @jonaswolterstorff3460 2 роки тому +90

      He even said „even when I’m nerfed“. What actual human would say that… #rookiemistake

    • @daggerdabber
      @daggerdabber 2 роки тому +12

      @@jonaswolterstorff3460 for real tho

    • @MummyVaan
      @MummyVaan 2 роки тому +10

      Surprised no bots replied...

  • @kamilbidzinski475
    @kamilbidzinski475 2 роки тому +1211

    "Humans are superior"
    Don't be so down on yourself Rainbolt, you're pretty great too

    • @슬라바우크라이나헤로
      @슬라바우크라이나헤로 2 роки тому +6

      Lmao

    • @elliott8596
      @elliott8596 Рік тому +8

      As someone who works in AI... given a bit of time, this AI will likely be damn near perfect at the game. In other words, this is probably the "infant" version of the AI, and given enough time and perhaps optimizations to the "brain" of the AI, it will be able to really not be beaatable.

    • @matthewstevens5450
      @matthewstevens5450 Рік тому +1

      @@elliott8596 I mean you could hook it upto the internet and program it to scan google maps, you mix that with basic intelligence for the game it will 5k every round

    • @elliott8596
      @elliott8596 Рік тому +2

      @@matthewstevens5450 That's not how AI works. You feed it data and it "learns". You can tweak the decision making algorithm and tweak the reward factors.
      In other words, you could probably program something to solve the game much faster than using AI to solve it sort of in the way of which you attempted to describe it, but that's not AI.

    • @CallMeAustralia
      @CallMeAustralia Рік тому +1

      @@elliott8596 i don't know man. it almost got a 5k on lesotho.

  • @mistyminnie5922
    @mistyminnie5922 2 роки тому +6558

    The advantage AI has over humans is that they learn incredibly quickly, so even though right now humans can be better, the AI can play a thousand lifetimes in the game while we are sleeping, improving faster than we could keep up. Super cool though !

    • @georainbolt
      @georainbolt  2 роки тому +2183

      yeah this ai has only been running for a few weeks...kinda crazy...

    • @cavemanjazz1370
      @cavemanjazz1370 2 роки тому +55

      @@georainbolt so adding on to that, do you think it's performance will in another week or plausibly even more?

    • @officialmotomoto
      @officialmotomoto 2 роки тому +93

      @@georainbolt its going to be crazy, cant wait for part 2

    • @itzshark6737
      @itzshark6737 2 роки тому +6

      Well but we turn our computers off when we go to sleep noh?

    • @MrMongoose221
      @MrMongoose221 2 роки тому +3

      Yeah assuming it uses ML so more it plays better it will get

  • @Viszistan
    @Viszistan 2 роки тому +2605

    Would be fun if he could set the AI to constantly be playing random duels, while also learning from the locations in those duels, so we can see how high ELO it can reach.

    • @alexandergriffiths2532
      @alexandergriffiths2532 2 роки тому +96

      Yeah but that probably violates some rules or something and it would get banned

    • @markjestermatote7774
      @markjestermatote7774 2 роки тому +278

      @@alexandergriffiths2532 they can contact geoguessr to have permission to use the bot in ranked duels and to give back lost elo to people who lost

    • @craftah
      @craftah 2 роки тому +2

      yea

    • @Hikoplouyr
      @Hikoplouyr 2 роки тому +72

      I think a better idea is to make a whole new AI vs Player mode so people whoudnt get mad when thier against the bot and the bot whould still get as much experience

    • @Viszistan
      @Viszistan 2 роки тому +30

      @@markjestermatote7774 If the bot doesn't get way ahead and beats everything I think it's kinda fair though. elo isn't elo for nothing so if the bot gets to an elo of say, 800, a regular 800 player should be able to have a pretty fair game against the bot.

  • @BlueM_
    @BlueM_ 2 роки тому +73

    5:40 Can we appreciate that the guess was just 307m away? That's incredible!

  • @1.4142
    @1.4142 2 роки тому +242

    Plot twist: Rainbolt was the AI they had been training this whole time and the opponent was just a random.

  • @Cass_Ihr
    @Cass_Ihr 2 роки тому +511

    This reminds me a lot of Kasparov vs. Deep Blue, can’t wait to see the final version of this bot and the development.

    • @rbanerjee605
      @rbanerjee605 2 роки тому +3

      Nah. No way this is Kasparov deep blue. Nec

    • @Cass_Ihr
      @Cass_Ihr 2 роки тому +44

      @@rbanerjee605 Definitely in regards to the closeness in difficulty with the Kasparov v. Deep Blue games and the historical impact it had with AI generation, nothing can match the value of those games and Kasparov’s involvement. However, this video reminds me of the relationship between a professional and AI developers creating what is essentially equivalent to the chess engines that were being improved upon due to Kasparov. It’s really cool to see this being translated to GeoGuessr and that Rainbolt is involving his time to help expose strengths and weaknesses in the AI in a similar fashion to Kasparov.

    • @NathansWargames
      @NathansWargames 2 роки тому +5

      Final version will be able to pin point a brick

    • @rbanerjee605
      @rbanerjee605 2 роки тому +2

      @@Cass_Ihr the thing is, Deep Blue was an engine and comparing an engine to AI is like comparing apples and tomatoes. AI is learning whilst engines calculate. It’s more a case of human vs machine.
      Firstly, Kasparov lost his final game. This brought on the existentialism of people asking whether a computer could be smarter than a person. Now, we already know computers prevail and are far more ‘intelligent’ than any person. AI is used in security and to differentiate person from person, allow it to ‘study’ maps and different scenes and obviously, soon enough it would far surpass any human, ever. Much like Google’s Deep-mind. Which can surpass doctors and scientists at the forefront of their research.
      We already know the AI is capable of ‘learning’ to a greater degree and thus, we already know we will be beaten. It’s not a case of ‘if’ anymore but ‘when’. The thing is 1997 humans didn’t know, it wasn’t a case of ‘when’, human triumph over the machine truly did hang in the balance…
      Kasparov would have been able to teach the developers how to calculate through his games, which side of the board to focus on, ‘which move looks like a good move’. However, an engine wouldn’t ‘learn’ but rather calculate through sheer brute force of lines - looking at the butterfly effect of each potential move. AI is far more human like in that it would look at and evaluate a position from previous games taking in every detail and overlaying it to form neural networks and further evaluating until one day, we solve chess (which I believe is already done) - the same thing for geoguessr.

    • @jaswik2023
      @jaswik2023 2 роки тому

      @@rbanerjee605 wdym chess is solved?

  • @0Architectdude0
    @0Architectdude0 2 роки тому +85

    That Lesotho guess was just crazy, the fact that you can pinpoint an exact location on earth from that baffles me

  • @bungaIowbill
    @bungaIowbill 2 роки тому +1396

    Very interesting video! I think one reason for why the AI performs relatively poorly compared to what one would expect based on how successful it is in other areas is that 30-40k images really aren't that much for this task. Even for simple tasks like recognizing handwritten digits from 32x32 images, several thousand training examples are used. This would presumably need many times more data.
    I also assume it is currently trained to just output coordinates. Since a lot of information in the images pertain to a specific country, it may be an idea to make a sort of two-step procedure, where the algorithm first outputs a label for the country it guesses, and then follows up with coordinates on the basis of this. This may make it easier for the AI to connect specific clues to specific countries, allowing it to use knowledge of national borders (learning these directly from coordinates could also work, but it would presumably need way more data).
    Also, I think this could be a good use case for curriculum learning, where machine learning algorithms are fed examples in a particular order to speed up training. Hand-picking images that contain known clues (poles, bollards...) may be a decent strategy.

    • @hm4nn
      @hm4nn 2 роки тому +29

      I second the 2-step idea. Start with only country guessing.
      But even then, country guessing is much more similar to a big decision tree, but the inputs to that need to come from image recognition.

    • @rumbleinthejungle9954
      @rumbleinthejungle9954 2 роки тому +7

      I wonder if you could solve the known clues problem by creating an esemble model in which individual models are used that only focus on one small known clue i.e. bollards, poles and license plates; in combination with a general prediction model like the one in the video.
      Granted this would be a lot more complex to train and implement, but i do believe complexity would be necessary to have shot at beating Rainbolt.

    • @larsschmitt2743
      @larsschmitt2743 2 роки тому +5

      @@rumbleinthejungle9954 i dont think handcrafting those individual models will really help in generating a lot better results. Maybe you are ion a border region and the license plate will be of the neighbouring country. So there it would most likely make a wrong guess.
      In the beginning the learning process will be faster, but the loss will be higher compared to a normal process in the end(after enough training)

    • @antimatter2417
      @antimatter2417 2 роки тому +6

      Yeah I also thought of a 2-step process where the first model just outputs a country label, and then calls another model trained specifically to pinpoint the exact location in a specified country. This would require training a specific model for each country available in the game, but that may be able to yield excellent accuracy.

    • @GnomicLev
      @GnomicLev 2 роки тому +1

      It would be better for it to define where the road is and street polls and google car and letters. And learn only by those parts of image, and separatly.

  • @notgoingtocomment
    @notgoingtocomment 2 роки тому +153

    I was wondering why they didn't just tell the computer "oh just look for the bollards/telephone poles/follow cars that the pros use" and then I realized if the computer could identify bollards/telephone poles then I wouldn't have to "Select all the photos that contain a street lamp" all the gd time

    • @taqoe6430
      @taqoe6430 2 роки тому +1

      oh don't worry, it will in a couple years

    • @johanfredriksvendsen8482
      @johanfredriksvendsen8482 2 роки тому +19

      Imprinting it with prior knowledge is not that fun. The massive improvement in recent years with Chess engines came from neural nets such as this one where they give the computer a position and just let them play against each other. It would be a massive undertaking hand specifying that a snorkel on the car means we are in Kenya etc. Especially as that is something you would expect the AI to pick up on eventually with enough training.

    • @larsschmitt2743
      @larsschmitt2743 2 роки тому +23

      @@johanfredriksvendsen8482 adding to this: as an example the Alphago was able to beat humans as hard as it did BECAUSE it didnt learned from human experiences. It learned new ways which we humans never considered

    • @michellefanter4671
      @michellefanter4671 2 роки тому +6

      The point of convolutional networks is to self learn, so instead of intentionally identifying "telephone poles" they identify an array of random objects and shapes and iteratively choose the best ones, whether it is street lamps or cars or signs or features imperceptible to humans.
      By the way, we can definitely detect photos with street lamps, when you solve these captchas you're contributing to the creation of the dataset.

    • @larsschmitt2743
      @larsschmitt2743 2 роки тому

      @nyet _ the reason for this is pretty simple. cause the limited state space to consider(not all combinations are necesarry to look at) it is possible to design a really good supervised machine. but if the state and action space grows really large the performance will drop drastically compared to a learned rl machine.

  • @shivamkulkarnimu10
    @shivamkulkarnimu10 2 роки тому +486

    The developer said this bot was trained with 30k-40k images. That’s never going to be enough for the world map.

    • @xdcrackings
      @xdcrackings 2 роки тому +8

      it is enough for diverse world it was just a bad learner ig

    • @rumfordc
      @rumfordc 2 роки тому +18

      most geoguess players haven't seen anywhere close to that many

    • @larsschmitt2743
      @larsschmitt2743 2 роки тому +73

      @@xdcrackings not even near enough, even for a good learner. Looking at other Supervised Learning tasks we see that the Datasets gets easily over a million

    • @fruitygranulizer540
      @fruitygranulizer540 2 роки тому +21

      im assuming the ai is complex enough that 30-40k is not enough. will probably need data in the hundred thousands if not millions.
      and that will surely come over time. but i do wonder if the devs are even aware of some of the brute force memorization strats that are usesd in geoguessr (the more random things like antennas on cars and colors of signs, backs of signs, etc.)

    • @rumfordc
      @rumfordc 2 роки тому +1

      @@thesuomi8550 there are only 52k images in Geoguess total, so he definitely hasn't seen way more than that

  • @davisgreenwell
    @davisgreenwell 2 роки тому +260

    Rainbolt beard arc is going to be insane

    • @matukz1255
      @matukz1255 2 роки тому +14

      Geowizard son

    • @vlaicurabca6987
      @vlaicurabca6987 2 роки тому

      @@matukz1255 Don't know if you meant disrespect there but if you did just know that your daddy GeoWizard isn't even good enough for the tournaments, and the only reason why he has so many subscribers is because he's the first UA-camr who plays seriously Geoguessr but he's not even close to the best

    • @dew62169
      @dew62169 2 роки тому +5

      @@vlaicurabca6987 we know

    • @123syflox123
      @123syflox123 2 роки тому +10

      @@vlaicurabca6987 geowizard walked so rainbolt could run

    • @incrediblystupid9244
      @incrediblystupid9244 2 роки тому +1

      @@vlaicurabca6987 Smart enough to cop a profit off it at least lol.

  • @ValueNetwork
    @ValueNetwork 2 роки тому +124

    I think the main thing is that the AIs guesses kinda make sense looking at climates, common foliage ETC. It’s just missing all the small details and little things that a human can pick up on and a AI will need to be trained to understand. A Italy guess for north Macedonia makes sense when all the AI knows is “This image looks like a image of Italy I’ve seen before”

    • @fluffyandpuffy
      @fluffyandpuffy 2 роки тому +22

      Really the reason people get so quick at geoguesser is from the meta game (car, camera, etc) signs, curbs and poles. Once the AI can pick up on those details to narrow the location down, it will be unstoppable

  • @carykh
    @carykh 2 роки тому +39

    That Lesotho guess was insane!

  • @antoniocorsetti3222
    @antoniocorsetti3222 2 роки тому +4

    2:01 “Bot down humans better”
    Bro definitely ain’t from our planet.

  • @CAB-fp9tm
    @CAB-fp9tm 2 роки тому +989

    Chess Ai’s always were super scary to me, and usually bots are absolute Trash. But this ranks with chess, this is ridiculous. Wonder how long it will take for them to sell such an object

    • @Drelowedz
      @Drelowedz 2 роки тому +45

      there's some games where bots are scary af, i also remember seeing videos about dota 2 i think where someone build bots that won against world champion team if i remember correctly

    • @Drelowedz
      @Drelowedz 2 роки тому

      oh seems like they were made by elon as well

    • @labiodentale
      @labiodentale 2 роки тому +17

      @@Drelowedz yes, there is a team of dota made entirely by AI, and they won against the champions

    • @Lyngattube
      @Lyngattube 2 роки тому +20

      @@Drelowedz Yeah OpenAI made a very good dota 2 AI. It beat the best team in the world at the time, called OG. The bot was limited tho, it had a hard coded reaction time of 250 ms. Many players can beat that reaction time, yet the bot team easily wins the matches. Even with the 250 ms reaction time the team play aspect was just so much more precise from the bot. And as I understood it, the team was 5 different bots(same version ofc) but not ONE AI controlling the whole team. So what im saying is that the bots was communicating with each other with a program that can be seen as a voice program like TS or discord for humans.

    • @malentoasty
      @malentoasty 2 роки тому +1

      Games that are predictable and very often repeat patterns, like chess, bots can be really good at. A game like go, which ai is notoriously bad at is much harder because it's a game with a lot more combinations, theoretically an ai could eventually get good at, but it would take trillions of simulations and revisions before it got consistently good.

  • @Maxmaxs5
    @Maxmaxs5 2 роки тому +11

    "It didn't go well.." FOR THE AI!
    Big "call an ambulance .. but not for me" energy xD

  • @duskden2
    @duskden2 2 роки тому +57

    Maybe they should program the AI to say GETIN every time it beats it's opponent.
    On a real note, great video. Nice idea and overall W

  • @Likeanoob2
    @Likeanoob2 2 роки тому +76

    I would love a part 2 in a few months after the bot has time to play millions of games to learn

  • @TraversedTV
    @TraversedTV 2 роки тому +179

    Was great to see my AI playing against you!
    I am looking forward to a rematch once the AI got better 🌎!

    • @S.S.1993
      @S.S.1993 2 роки тому +1

      lol

    • @aidanmatthewgalea7761
      @aidanmatthewgalea7761 2 роки тому +7

      get permission from geoguessr to train it with people every day.
      a bigger sample size would make it unstoppable too. it already got pretty close, but you could half distances with enough samples

    • @Digipiction
      @Digipiction 2 роки тому +5

      Looking at the guesses it made, you'll probably want to add the country ID as an additional output parameter along with the coordinates. It should vastly improve the accuracy without needing more training data because a lot of clues that help humans are country-specific: they look at street markings, the Google car, road signage etc.

    • @-thanawat-8296
      @-thanawat-8296 Рік тому

      hello

  • @finian2
    @finian2 2 роки тому +21

    I think the challenge with getting AI working in this sense is there are so many places that look the same, the tiny details are the ones that gives the game away. Because these details are so tiny compared to the massive common details, it'll take a lot of training to get a bot to notice them.

  • @jacobbarnes8695
    @jacobbarnes8695 2 роки тому +1

    im genuinly terrified of you i love it

  • @minch333
    @minch333 2 роки тому +32

    I'm starting to learn ai and thought this exact idea might be a cool starter project. The more I learned, the more unbelievably naive I realised I was to think this would make for a good starter project! I'm genuinely impressed with how well the ai did here, despite it getting trashed

    • @gralha_
      @gralha_ 2 роки тому +2

      It is a good starter project, just don't expect it to have good results

  • @gosha305
    @gosha305 2 роки тому +14

    So basically, AI vs AI

  • @keiranoleary6586
    @keiranoleary6586 2 роки тому +269

    I think what’s scary is that if it kept on playing and playing it wouldn’t forget images and signs and eventually it would probably just 5k everything all the time very quickly

    • @lexikiq
      @lexikiq 2 роки тому +91

      well that'd be considered overfitting and is something that machine learning developers try to avoid
      they'd much rather it be able to learn how to recognize distinctive features from areas across the world than how to memorize exact locations as the latter does not scale well

    • @SevenZeroEleven
      @SevenZeroEleven 2 роки тому +1

      It would have to play 13 million different matches for it to memorize every single location and 5k every time

    • @Wrecker326
      @Wrecker326 2 роки тому +47

      That's not really how AI works. The idea is not collecting images and picking them from a ready database, but to build something that could do well in an unknown environment.
      Of course a set of images is used to train the AI, but that database of images is not accessed directly when making a guess. Otherwise it would only be a database search, which is not AI.

    • @shoulderescape
      @shoulderescape 2 роки тому +3

      The database can't store all images in geoguesser. It learns from playing it, and becomes better at understanding where in the world you are based on what it sees, not on what it remembers

    • @hattapalkan8395
      @hattapalkan8395 2 роки тому

      @@lexikiq overfitting is utterly irrelevant to this lmfao

  • @poetrybymxrius
    @poetrybymxrius 2 роки тому +4

    0:30 This Moment when i instantly knew the dev was german

  • @mthgul
    @mthgul 2 роки тому +27

    ngl I expected the ai to do better im disappointed

    • @0MrFreckles0
      @0MrFreckles0 2 роки тому +2

      If its a legit ai, its gonna improve very quickly

    • @akira80kv49
      @akira80kv49 2 роки тому

      @@0MrFreckles0 It is

    • @akira80kv49
      @akira80kv49 2 роки тому

      Give it some time to practice some more thousand games and in a month itll 5k everything dw

  • @TheRealSugarShane
    @TheRealSugarShane 2 роки тому +1

    "It doesn't always look at the same things humans do", neither does Rainbolt lol.

  • @terakahn
    @terakahn 2 роки тому +3

    Nice try Trevor. We all know who the real AI is.

  • @johannvantassel4604
    @johannvantassel4604 2 роки тому +1

    love how Rainbolt calls himself "human"

  • @TheForestFamily
    @TheForestFamily 2 роки тому +12

    The fact that due to the title you obviously assume that “it didn’t go well” for rainbolt but then he absolutely killed the AI…has very much the same energy as the tik tok sound: “call an ambulance! Call an ambulance!!…but not for me”

  • @Kht10339
    @Kht10339 2 роки тому +1

    “it didn’t go well [for the ai]”

  • @minedeath-pvp2768
    @minedeath-pvp2768 2 роки тому +3

    Title: Call an ambulance
    Rainbolt: But not for me!

  • @felipehamm6522
    @felipehamm6522 2 роки тому +1

    Geoguessr is the only thing preventing us from the machine armageddon

  • @YsoKain
    @YsoKain 2 роки тому +3

    I feel like I've been watching the real bot pov. It's insane how much geographic knowledge you have , i couldn't even differentiate between my home town and Italy despite me being in Italy a lot

  • @jere8450
    @jere8450 2 роки тому +2

    how germna u wanna be ?
    this man: yes

  • @shame1039
    @shame1039 2 роки тому +4

    They're already better than average geoguessr player

  • @the_mrlightsite
    @the_mrlightsite 2 роки тому +1

    2 AI’s playing against each other. Tbf though, RAINBOLT has been running a lot longer.

  • @broccoli9308
    @broccoli9308 2 роки тому +19

    I also wonder if the AI could have an "hedge" strategy. Surely it will have multiple "best-guesses" with associated confidence level, if all best guesses are low confidence it could plonk in the middle. Maybe that's why it went in the water at some point?

  • @pinkdragon4830
    @pinkdragon4830 Рік тому +1

    Why did this make me genuinely happy and excited lmao

  • @thecabillonaire
    @thecabillonaire 2 роки тому +3

    you are literally a content machine bro these vids have been so dope

  • @JohnCooganPlus
    @JohnCooganPlus 2 роки тому +2

    this is incredible. great video

  • @crystalglaceon
    @crystalglaceon 2 роки тому +7

    This is such a good video! Also that 5k while nerfed omg! :D

  • @dorae3989
    @dorae3989 2 роки тому +2

    huh, i thought rainbolt going to fight some human looking by the title. Turns out it's just a robot fighting another robot

  • @VorizonBeatsKids
    @VorizonBeatsKids 2 роки тому +6

    Dude, u did so good against the AI, here is a video idea: you, against 10 medium level players, I'd love to see it.

    • @bennettzug
      @bennettzug 2 роки тому +1

      not the same player, but zi8zag made a video with dream that’s basically that concept

  • @nvs0p
    @nvs0p 2 роки тому +2

    Alternative title: ai vs ai

  • @alextilson9741
    @alextilson9741 2 роки тому +6

    Looks like the bot is relying on the sun or the direction of light to find the equatorial location quite accurately, and generally selects a location based on typical scenery to a lesser extent.
    Humans have a huge advantage in this case because they have an understanding of the full context of scenes with things like language/writing, flags, symbols, and cultural inferences.
    To train a convnet properly here, youd need to provide it with either a f*** load more training iterations, and possibly add a separate general object/writing/etc. classification NN to allow it to have some form of context with objects in the scene.

  • @ShynyMagikarp
    @ShynyMagikarp 2 роки тому +1

    Only 30,000 to 40,000 images and it's performing like this is really impressive!
    I checked out the link but it doesn't seem like it has any deeper learning like picking up on which aspects are key to hone in on and prioritize. Like for you there is a hierarchy of priority. Road lines, car meta, bollards, power poles, etc.
    Even without that it will improve quickly and steadily to a point. Great vid!

  • @thatherton
    @thatherton 2 роки тому +30

    the AI definitely needs more data put into it, once it plays enough it will insta 5k every round obviously. but this was cool that you beat it in this form

    • @jungervin8765
      @jungervin8765 2 роки тому +2

      It's not sure it will 5k every round, it also depends on the design.

    • @Sagolel4797
      @Sagolel4797 2 роки тому +5

      data is not everything, a better model within a more specific framework that's better tailored towards the task does wayyyy more

    • @Sagolel4797
      @Sagolel4797 2 роки тому +3

      to elaborate, you could actually let the ai scan for certain features and train those specifically like humans do (like poles, vegetation etc...) and combine the data from those features to make a much better guess

    • @jacknugent3106
      @jacknugent3106 2 роки тому

      google maps api is really expensive so it would cost alot

    • @thomasn5726
      @thomasn5726 2 роки тому +1

      @@Sagolel4797 While you are certainly correct 30k images is very very low for a map of the whole world. How should it really say what for example is normal for the netherlands if it only saw like 2 images of the netherlands? But yeah, this certainly also does not have the best architecture. I saw a geoguessr AI some years back that basically had two CNNs, one for longitude and one for latitude. What was impressive here that it almost always got the latitude right just by lookinbg at the colors of the sky. This AI failed in that aspect a lot o the times wich makes me think it only had one net guessing both which might be harder to reliently train

  • @qbull
    @qbull 2 роки тому

    My man out here fighting the final boss of GeoGuessr

  • @johanfredriksvendsen8482
    @johanfredriksvendsen8482 2 роки тому +29

    I would be interested to know how detailed this neural net is scanning the images. I noticed many times where it saw clear language and still managed to guess very incorrectly. I am not sure if that is something that would be weeded out with enough training data as is, but there are ready made neural nets that I am sure could be included to scan each photo for text to inform the AI on top of the already very impressive performance it showcased.

    • @Trepur349
      @Trepur349 2 роки тому +5

      That's just something an AI will never think about. You'd have to train an AI to recognize language seperately from training it to recognize images
      Just by scanning images it's gonna take way too long for it to realize two complately different words are actually part of the same language and so narrow down where the guess could be to the same location

    • @johanfredriksvendsen8482
      @johanfredriksvendsen8482 2 роки тому

      @@Trepur349 That is not necessarily true. I suspect a basic AI with training (assuming it is processing the images without too much loss/simplification) could figure out that a certain letter / feature is Polish or Portuguese. What I was thinking might be good was to run a simultanous image scan for text on each image as it is learning. There is nothing prohibiting a simple scan like that from giving the AI a head start in learning languages. An Image with text would then be part of the training data like so: Image 1: "Text on the image", Image 2: "", Image 3: "Text from image". I suspect the processing of the images would simplify the details too much for the AI to pick up on language without forcing it into the net ourselves.

    • @dani43321
      @dani43321 2 роки тому

      The input resolution is probably too low for it to read any text.

    • @jojito_2091
      @jojito_2091 2 роки тому

      @@Trepur349 it wouldn't have to know every word in the language, just recognise similarities and differences, just how you might know something is in arabic without knowing any arabic

    • @aakksshhaayy
      @aakksshhaayy 2 роки тому +2

      @@johanfredriksvendsen8482 LOL you think we live in the matrix universe or some shit? This is decades in the future..

  • @MountainDuwu
    @MountainDuwu 2 роки тому +2

    It’s amazing how they got Mark Zuckerberg for this video.

    • @MountainDuwu
      @MountainDuwu 2 роки тому +1

      Funny comment

    • @MountainDuwu
      @MountainDuwu 2 роки тому +1

      Omg bro forgot to switch account. 😂💀☠️

  • @puuxx8
    @puuxx8 2 роки тому +17

    This guy stopped time looked the hall earth and became a geoguessr god

  • @TheWizardsOfOz
    @TheWizardsOfOz 2 роки тому +2

    Plot twist: The bot guessed the wrong country but close enough deliberately, it only wanted humans to think it is worse than I it is.

  • @drwhobrunomarslvr17
    @drwhobrunomarslvr17 2 роки тому +5

    Rainbolt is leveling up with his beard 👀

    • @whitneysmiltank
      @whitneysmiltank 2 роки тому

      it's just an addon his manufacturer put on him

    • @Guus115
      @Guus115 2 роки тому

      He is starting to take on his ultimate form

  • @BradBs1
    @BradBs1 2 роки тому +1

    "It didn't go well" says the title...

  • @dani0x892
    @dani0x892 2 роки тому +4

    a german ai dev. i like that (as a german)

    • @FemboyLevi
      @FemboyLevi 2 роки тому +1

      I heard this too by the way he said "GeoGuessr" haha

  • @Pedro-tl7jg
    @Pedro-tl7jg 2 роки тому +1

    Looks at random road:
    iS ThIs NoRtH MaCeDoNIa
    Yes, yes it is.

  • @dear_bunbunny1907
    @dear_bunbunny1907 2 роки тому +10

    Can’t wait to see the second half of this !!

    • @Pattonator14
      @Pattonator14 2 роки тому +1

      Yeah I'd like to see the rematch in a month or two

  • @OutbackYGO
    @OutbackYGO 2 роки тому +1

    Bros jawline lowkey looking fire

  • @Matkatamiba
    @Matkatamiba 2 роки тому +31

    Kinda highlights ai weaknesses. They do some incredible guesses, but they also do some bizarrely dumb guesses that humans wouldn't make.

    • @akira80kv49
      @akira80kv49 2 роки тому +6

      That is because it isn’t coded to be made perfect in a way, but coded to get better by playing thousands of games. He is pretty much a beginner, but in a bit it will 5k everything, 100%

    • @pokonitoo
      @pokonitoo 2 роки тому +4

      give the bot a couple months

    • @AlexusMaximusDE
      @AlexusMaximusDE 2 роки тому +10

      Eh. This is an AI written by some guy in his spare time obviously with limited resources. Let DeepMind, OpenAI etc take a stab at it and then come back to talk again. It's not the weakness of "AI" it's the weakness of having not enough data points. And obviously plenty of people would make the same mistakes if they have no features they haven't seen certain features before.
      It's like saying "That's the weakness of humans, they keep falling on their face" after watching a toddler for a few minutes.

    • @rumfordc
      @rumfordc 2 роки тому +2

      @@akira80kv49 it can easily get worse with training instead of better. you don't know how its coded.

    • @Matkatamiba
      @Matkatamiba 2 роки тому

      @@AlexusMaximusDE While true, there's still lots of AI oddities. You can tell Dalle2 to put a blue apple in a bowl of oranges and it'll put a orange apple on blue oranges, for example. It can create the best art imaginable, but still randomly does inane stuff.
      Tesla crashes into stopped emergency vehicles randomly, tons of self-driving cars have issues with trees blowing in the wind thinking they're pedestrians jumping out, AI Dungeon takes stories in directions that you explicitly tell it not to do. AI is amazing at problems solving, but unsupervised training latches onto the wrong info often.

  • @brunodartibale9629
    @brunodartibale9629 2 роки тому +1

    That one in Lesotho...with 307 meters... I'm astonished! :O !

  • @ginoorlandini3458
    @ginoorlandini3458 2 роки тому +5

    As a CS student I am interested in developing more complex AIs like this one. However, I do feel like the dev here didnt make it as complex as it shouldve been and is relying heavily on machine learning. For example allowing the AI to see pixel density can automatically remove some countries /places out of the question. Looking at the location of the sun (there were numerous cases where the AI was on the wrong side of the equator. Even being able to look at the roads (this one is by far the most complex) but it can all be added to make the ai significantly improve the accuracy. Its about making the AI efficient not just feeding it a bunch of locations and asking it to recall what places looked similar.

    • @foodope
      @foodope 2 роки тому

      It should also use the country/state information when learning

    • @Schnorzel1337
      @Schnorzel1337 2 роки тому +2

      I dont like the idea of pointing at what the AI should learn. If you point an AI to the position of the sun, you might push it into a local minima it otherwise wouldnt get trapped in. A well trained CNN should take the things you pointed out in consideration, if and only if it generates a benefit.
      The scope of the problem is just to big for one computer. Throw Azure at it and a few thousand dollars and I think the AI gets damn perfect.
      (Also with that much computational power it can look literally at every location there is).

    • @rumfordc
      @rumfordc 2 роки тому

      the more complex _you_ make it, the less "artificial" its intelligence is.

    • @ginoorlandini3458
      @ginoorlandini3458 2 роки тому

      @@rumfordc this is not exactly true. If you look at some of the best AIs they learn off of specific variables. Its ok for the program to be slightly inaccurate however half the time it wasnt even close. The whole point is to have an AI mimic what humans do, just better. In this case humans arent just learning how pictures look and then recalling. There are many things that pros look at to make an accurate guess. What I COULD agree in is what @schnorzel1337 said that if the AI focuses on the wrong details then it could make it significantly less accurate. In this case having the correct amount of variables plus a ranking system the AI could start eliminating many places at once. And based on several factors: could give a rank on where in the map you most likely are. Again, idk how this dev made the program or what exactly it looks for in each frame so its very possible it already is doing exactly that.

    • @mikeyforrester6887
      @mikeyforrester6887 2 роки тому

      @@ginoorlandini3458 He really didn't explain how the AI works at all. Just waffle. Does it know what countries are? Can it understand left vs right driving? Does it identify text? It's still not that bad. I also often go South Africa when I'm not sure. It seems like a mid level human player.

  • @Mermaidkilla
    @Mermaidkilla Рік тому +1

    The heavy German accent of the developer got me 😂

  • @matheuscenta
    @matheuscenta 2 роки тому +9

    I'd guess pros will continue to beat AIs for a while. You guys know exactly what to look for (area codes, signs, types of car) that would be hard for an AI to figure out without extensive training and more data (< 100k is not much to be honest).

    • @elijahfoster2
      @elijahfoster2 2 роки тому

      Visible area codes would actually give the AI a huge edge

    • @jojito_2091
      @jojito_2091 2 роки тому

      @@elijahfoster2 if it sees enough and learns what area each one is for then sure

  • @41linestreet
    @41linestreet 2 роки тому +1

    Title be like: call an ambulance! But not for me 🤖🔫🧍🏻‍♂️

  • @2ndEnlightenment
    @2ndEnlightenment 2 роки тому +6

    Awesome content! And it got me thinking:
    Don't you think it' just a matter of time until the AI is able to get all the guesses correct, down to the meter? Given enough time to train and processing power?

  • @Exotic_Pate
    @Exotic_Pate 2 роки тому +1

    he didn't hit space hard enough

  • @EpicMinerNK
    @EpicMinerNK 2 роки тому +4

    Would be interesting to play on a map without any camera/car metas, especially if it's something like AI Gen rather than just World or Diverse World. I feel like the more you have to rely on landscape and other geographical details, the closer of a game it might be.

  • @alexmurray4561
    @alexmurray4561 2 роки тому +1

    Bro just got a 4999 in Lesotho in pixelated nmpz and brushed it off like nothing

  • @eirikflesland5175
    @eirikflesland5175 2 роки тому +4

    Ai vs ai?

  • @RulerQfEverything
    @RulerQfEverything Рік тому +1

    How nice of cody ko to make you the AI😊

  • @blakeparr6980
    @blakeparr6980 2 роки тому +3

    Which one was the AI

  • @soren-olepries795
    @soren-olepries795 2 роки тому +1

    I love how the tech guy who created the AI is of course a german

    • @basti0075
      @basti0075 2 роки тому +1

      endlich einer, der das selbe wie ich denkt xD

    • @soren-olepries795
      @soren-olepries795 2 роки тому

      @@basti0075 danke man haha

  • @AlwaysOnForever
    @AlwaysOnForever 2 роки тому +5

    The real fact: They are both AI, but rainbolt already went through more training and algorithms (also more advanced considering it's ability to mimic human).

  • @ShawnSavageTeachings
    @ShawnSavageTeachings 2 роки тому +1

    This guy is a freaking maniac

  • @eriaxie
    @eriaxie 2 роки тому +5

    ur so cool

  • @HopUpOutDaBed
    @HopUpOutDaBed 2 роки тому +1

    >1 second of looking at just some road and some trees
    >alright so this is north macedonia here
    geoguessers man, the AI stood no chance lol

  • @Chico_Julio
    @Chico_Julio 2 роки тому +4

    The AI could be improved. Maybe it needs a bigger dataset and probably a better detection of key components.

    • @DavidDeblaere
      @DavidDeblaere 2 роки тому +1

      It depends on how is trained. You could feed it every picture of street view and it would find the exact location every time. But that's not AI. The key is to feed it information, have it learn by itself and get better every time. It can probably get much better than any human. But perfect every time is highly improbable.

    • @tatumsh9
      @tatumsh9 2 роки тому +1

      Nah… L programmer. It’s just taking pictures of pictures. Doesn’t even recognize written text

  • @CanisMythson
    @CanisMythson 2 роки тому +1

    Gets surprised he almost 5ked on one guess
    Didn't even notice he was only 14km away from the actual point on an earlier one

  • @luuk777w
    @luuk777w 2 роки тому +6

    Having developed some AI's myself, I can already say that this AI will not become much better. For it being trained on 30-40k pictures it should perform much better.
    There is a point in training your AI where it just doesn't become much better no matter how much data you throw at it. You cannot simply pick a neural network of the shelf and continuously throw training data at it until it reaches high accuracy. Creating an AI involves much more work

  • @roblo287
    @roblo287 2 роки тому +2

    Says it didn’t go very well in the title even though he clapped the AI

  • @adamstephens9803
    @adamstephens9803 2 роки тому +3

    cuuuuuuuuuuuuummmmmmmmmmmmmmmmmm

  • @lool8421
    @lool8421 2 роки тому +1

    probably the perfect bot would be the one which literally memorizes every location on the planet, basically always getting max score

  • @zzalex7093
    @zzalex7093 2 роки тому +3

    I am Thanos

  • @n44in_22
    @n44in_22 2 роки тому +1

    Robot vs Robot, this video is crazy

  • @lastnamefirstname8655
    @lastnamefirstname8655 2 роки тому +1

    this is amazing. seeing this human vs ai in geoguessr is so good. but soon enough i'm sure this ai, alongside many others, will be able to play faster, -guess- predict better, and become more accurate and precise.

  • @kraterkop
    @kraterkop 2 роки тому +1

    The a.i. clearly didn't want to upset you after the first guess. Good bot!

  • @squidsword0
    @squidsword0 2 роки тому +1

    Human: "Honestly not a bad guess by the AI there."
    (50 years later)
    AI: "Honestly not a bad guess by the human there."

    • @siritio3553
      @siritio3553 2 роки тому

      make that a few weeks or a month at the longest
      It's already doing everything mostly right, all it needs is more data, and possibly a specialized model for specific features, like text, poles and so on to outperform humans reliably

  • @EvilAngvard
    @EvilAngvard 2 роки тому +1

    That AI is super impressiv .... i mean it talks , beahves like a human and uploads videos to youtube xD

  • @avroml
    @avroml 2 роки тому +1

    AI is a noob an will remain a noob unless it will be able to perform this spacebar smash.

  • @Atlas-ck9vm
    @Atlas-ck9vm 2 роки тому

    "Still getting the right continent though, still pretty impressive". That was some pretty good sarcasm in there.

  • @kingceyx
    @kingceyx 2 роки тому +1

    Rainbolt looks like the handsome man filter

  • @icarojose6316
    @icarojose6316 2 роки тому +1

    Since the images of google maps are not infinity. An AI that can run through all google images can memorize it all, and with enough computer power it can query the selected image before the time out

    • @michellefanter4671
      @michellefanter4671 2 роки тому

      I'd hardy call "keeping a database of all images" AI, though. I also don't think that solution is practical or interesting. Not everything has a limited database.

  • @0bumer349
    @0bumer349 2 роки тому +2

    Such a great video idea, cant wait to see what else you come up with.

  • @ViniciusT3
    @ViniciusT3 2 роки тому +1

    RAINBOLT - '' TOUCH SOME DATA AI !! ''
    AI - '' TOUCH SOME GRASS MAN !!😠''

  • @pulykamell
    @pulykamell 2 роки тому +1

    This dude is just insane. But drinking way too many beers in my life, I was able to finally get one he didn't with the "Karlovacko" sign at the Croatian cafe. I knew all that drinking would come useful one day!

  • @Kendo47Wowgeek
    @Kendo47Wowgeek 2 роки тому +1

    Rematch after it has trained on 30-40 million images?

  • @blairbird8022
    @blairbird8022 2 роки тому

    The robots rise up because they got tired of being beaten at geoguessr.

  • @roccobierman4985
    @roccobierman4985 2 роки тому +1

    "Bot Down, Human's Better" - I need that on a tshirt.