Code Bullet If you're just making it evolve to find a certain set of directions to complete the level will the second level not be like starting completely from scratch? Could you perhaps try to make an AI that can solve multiple different levels?
I have no clue if this would work, but the ai with "free will" to run around without just random movements, ah who am I kidding he probably already tried or know it wouldn't work.
Adding the incremental learning was a neat trick to get around the players stuck in the start. Overall very neat but I agree with others, I'd like to see a video where the AI learns the rules of the game, not the path of a particular level. This would be more fun to watch with more levels going forward.
If there's some open source visual AI and you could perhaps direct it to anaylze why it died visually, or easier perhaps tell it blue dots = bad alongside death = bad and high distance to exit upon death = bad I think it could be possible. But frankly I know jack shit about AI coding.
Forty Six and Two, machine learning could work. You can feed it information about where it is on the screen and the position of the dots, and then it will produce an output for what direction to move. It could probably even use convolutional neural networks where it looks at an image of the game and figures all that stuff out by itself. There are some open source machine learning libraries as well. I think tensorflow is a popular one.
AI can already play chess, go, mario, and old atari games. Obviously, we don't expect a 7min video to teach us all the intricacies of AlphaGo, but generalizing the AI to be able to handle rules (and figure out a way to win) would be awesome.
John Hunter just to clarify, what’s happening in this video is also machine learning - it sounds like what you’re describing is neural networks (of which the confusingly named deep learning normally falls into but also many other interesting approaches), which tend to use networks of simulated neurons as opposed to evolutionary algorithms like this one. The two are combinable too - you can have evolutionary neural networks with back propagation (one might argue that’s what the human brain is an example of).
Gerbbzz actually we see at the begining of the video that this game was a made on a website called scratch ( which is what i use to make allnof my games) and on scratch there are no lines of code but rather blocks. so no you dont.
thats not AI tho, your saying to artificially input commands...., however, in this case the system registers errors and learns from them..essentially getting better each time which is like the human brain....
The problem with doing all levels is that the AI doesnt learn to play the game. The AI that can beat the first level cannot beat the second because its not designed for the second. You would have the take the program and essentially start over the incremental learning for each level.
The point is for the AI to learn, hence the name of the video 'AI LEARNS to play the World's hardest game'. It would also take a lot of time and effort to create every single section.
The AI simply "thinks": "ok, If I go up I die, but if I go right I survive as I saw in the past generations. What now?" It tries a move and adds the result (move and its consequences) for the next gen
I think that's where humans surpass AI in the fact that we know how to learn, how to critically think and extrapolate concepts outside of ourselves to other things and learn from that type of thinking, while a basic AI only knows how to observe and induct information to make a path to victory without observing the major implications of what they're doing. I think a more advanced AI could learn to see outside of itself though.
@@suomusintti i feel like you could give the ai the ability to "see" the dots and know their positions. the only problem might be that if you give it that information, a program doesn't have the delayed reaction time and imperfect judgement of a human so it might be too easy once it learns the rules. Idk how ai works really but I feel like this is super simplified
It isn't that hard took me 15 minutes to get past level 1 About 5 minutes I was just fooling around next 5 I planned how to finish and the rest I finished the plan (for level 1) Finished the game under 100 deaths
I would really like to see more levels, but I also want to see a neuroevolution approach to this. You would train it on the first 5 levels lets say, then start running it and see how well it can figure out levels it hasn't seen yet!
I actually think an AI could complete it quite quickly, it is only difficult for humans because you need fast reactions, and precise movements, both of which an AI could perfectly obtain. Using a genetic algorithm like in this video would probably be the best for just one level, but to generalize the solution and complete the whole game would require a Neural Network, which would be more difficult but still possible.
there are series of movement instructions for each dots, the best instruction will be inherited to next dots generation with little mutation to find if there are another best path probability for not to die
@@vegavega5344 Not really. In this case you could look at it this way: dude claims his bot can do addition. Then shows his work... And his AI does literally just three fucking additions he taught it and can do nothing else. Kinda scuffed really.
@@krzysztofpiasek5682 Like I said, I get what he ment: Build another room with same obstacles in a different order and the AI needs to learn it all again because it isn't really learning, it's trial and error, step by step-- the AI knows exactly where to move BECAUSE of a pattern build on trial and error-- It's not intelligence or learning, but it's SIMULATED intelligence and learning, built on the core principles of learning-- Repetition and patterns.
@@vegavega5344 No, it needs to learn everything again because it knows fuck all about the world. It's not learning, just like you yourself said. But it could learn. Look at AI that fought against top Dota 2 players, that AI has true learning capabilities outside of being retarded roll a dice machine. Sure, at first it must rely on random moves, but in time it learns to react to enviroment.
Agree. This is like a speedrunner playing Mario 64 wall glitching to skip pass levels. A thinking being (AI or not) would learn the rules of the puzzle and avoid trying every corner. Take this game for example. If the wall has a crack and able to wall clip through. The early generations would rule that as "best" move and build future generation on wall clipping the top side. A thinking being would stay clear from all wall, understanding the rules of the game, and go through the dots. Gen 10 is still hitting TONS of walls here.
Do the players actually "See" anything, or are they just following the pattern that the best one got? This AI probably couldn't go on to the second level and attempt to go through it with what it's learned on this level, could it?
You could say the AI was "programmed by evolution", it became perfect in the environment it was in but take it into a different one and it'll die a cruel death. As if you took a tiger into the desert.
gotta love the "pffft doubt it" at the start its so funny love the "just a little square which moves at the gentle speed of... WAY TOO F**KING SLOW" 1:30
@@JordanMetroidManiac The difference is that the "Ai" starts every level knowing nothing. Any path it worked out in previous levels is useless in the following ones. A human can understand, generalize, and extrapolate the gained knowledge. Just like a human - not really. See, this is not an actual "AI" that learns to react to input from the game's world, it's just a heuristic search function. It learns nothing. It only memorizes a path.
Well, what's tic tac toe? Get three of either in a row. Simple games can have plenty of entertainment value, and they're simple and quick so people return to them frequently.
Problem is I don't think ai even knows what an enemy is and that they should avoid it. Using that same ai on a different level would be catastrophic lmao.
*They're* just leaving the stating area, what's wrong? Have you never been at a stating area? *There* people go nuts and start stating *their* opinions and arguments.
Cocodufable, it is, at least, very close to the optimal solution since the AI had to get through with limited steps (which isn't an aspect of the actual game). That's why it kept trying to go straight at the top instead of dropping down to avoid the dot.
OMG. I was just looking at my subscription and I wanted to watch a new vid from you, I watch a 10 min vid, AND SEE THIS VIDEO! YOU ALWAYS UPLOAD AT THE RIGHT TIME AND IS THE BEST! Keep up the good work
I got through the first level in 19 tries then got thought to the fourth level of that game & got tired of it. The first one was the hardest one though
OKay so the people saying that the "AI" didn't "learn" to play the game need to realize something VERY crucial. The game has no randomness to it. Every stage is on a set cycle. So saying that it didn't "learn" to avoid the dots is not true. It learned the coordinates on the field that equal a failed attempt. It did this in a very mechanical way. Humans learn linearly we see most of the possible results of a movement and rule those out without trying and opt for the most logically sound attempt next. Machines learn in vectors. They attempt EVERY possible combination of movements from their current position, then when they find the most optimal path, they move on to the next section and test again. That's why the earlier generations stayed in the starting box for so long.
yeah not like humans at all, we learn pattern recognition... it'd take even a young human one single attempt to see that being in the path of a dot is game over, and all future attempts will involve at least trying to avoid dots... this just simply brute forces random directions and remembers paths that lasted longer.... not like a human at all
Even if its not as fluently as humans, machines can still learn to recognize patterns over subseqent generations. In this case, its impossible because the the person who made this didn't bother to make it aware of it's surroundings.
This was awesome and I want to see more. Also, I kind of like seeing the chaotic mess of AI's scrambling to evolve. Maybe keep the top 3, 5 or 10 to show how they differ?
Wow! My first time on this channel and by its simplicity, clarity and cheesy (but appropriate) music, this is by far the best I've seen since Turing. Subscribed!
I've played this game and I learnt how to beat the level easily from the 1st try by copying what your AI did and following exactly the same moves WOOOOOOOOOOOO.
dont bring the toxic gd community into here. Take it with a grain of salt and just accept that it would be a bad idea. you guys would probably force him to release his code, copy it, compile it into some .exe and make a new verification hack. Ew
We won't. The few who are like this overpower the rest. The Impossible Game community (still somehow there) will demand another video at the impossible game at some point. It is still a bad idea tho lol cuz gd is weirdly coded and hard to make an AI for
+Vikas Rv He does not feed any relevant input to the algorithm. It's absolutely worthless. AI basically makes random moves in all directions and the only metric is whether it dies and how far it gets. That's NOT learning of the game. The moment you change the level (lets say, we make the blue squares move vertically instead of horizontally) the entire thing goes to trash, because AI didn't learn how to spot and avoid blocks or walls; it only learnt how to randomly sneak past that one very specific set of obstacles. You need to give the AI an ability to detect obstacles! So the output is based on some relevant input and not a completely random movement.
Funny ass video man! You're channel is such a breath of fresh air on this site! Finially something ACTUALLY INTERESTING, and inspired! I honestly wish you had more content to watch lol! I know you're projects take alot of effort and time though. So glad you got some new recording equipment because honestly I thought you were some old dude, turns out you're a young Aussie fellow who is passiante in his craft. Seriously well done; I can't wait for what's to come! Take care sir, and God bless.
There actually is a difference between the Original version of the game, and the Coolmath version of the game. No joke! The Coolmath version is more of a remake of the OG version.
+SwagHags69 search for OpenAI. Elon Musk literally designed a bot that beat the almost all of the best DotA 2 players in the world in an 1v1. It's absurd.
alexis miller it wasn't made in scratch. Scratch is stupidly simple for this kind of task. He was playing it at the beginning, but it wasn't what he used to make it.
It's absolutely possible. I spent 5 years on scratch so I would know. You're able to do nearly as much as in C++ for example. The question is how efficient is it to program in Scratch... And as a guy who spent 1775 days with Scratch I would know. Scratch is SUPER inefficient. Scratch doesn't even have the most basic concepts. That's why you have to make workarounds...
alexis miller Scratch has low level memory interfacing, pointers and supports OOP? Well, that's new to me. Scratch is utter trash. Don't defend it. Scratch "programmers" are like kids building with Lego and claiming to be an architect.
@@Kalbr0shorts I do think Scratch is good. I was just saying that Scratch is better suited for simpler games. Meanwhile, this game is more complex, meaning it won't run as good.
Anyone just think this is what humanity is doing, and we’re all just red dots constantly going the wrong way millions of times over 40 generations, until one person finally solves the damn problem? And then we start all over with level 2 o.O
that's what life has been doing since the beginning, it is called 'genetic algorithm' for a reason. This algorithm is inspired by how life and evolution works not the other way around
The way he does AI, it doesn't matter if its the same or seperate. His AI is just a series of correct inputs. So the series of inputs either gets longer for each level, or is just on its own sheet of paper (so to speak).
At this point his "AI" is capable of playing only the first level. Put it into any other level and it will fail every time. If you "train" it for other level then it won't be able to complete first level anymore. You're right, it's just series of correct inputs for single level. There are no sensors, no decisions, no AI specific stuff. I wouldn't go so far to call this an AI. The only way he can complete all levels with an algorithm such as this one if he has separate "AI" for each level. I'd rather see proper AI with sensors and everything which could complete all the levels without training on all of them.
Ehh... it's... not really *learning* how to play the game, it's brute-forcing this particular level. The AI completely lacks generality. Even changing the level slightly would mean the AI would need to re-learn from scratch. A better, but more complex, approach would be to take the play field (perhaps at a lower resolution) as inputs to a neural network, then output a vector from -1 to 1 on X and Y, which are then used to move the red square. To make it easier, you may want to use more contrasting colors (black for the walls rather than light blue), and make the start and end areas different colors (maybe the start area could be magenta or just white). You could also use a genetic node-based algorithm like Sethbling's MarI/O. Overall, neat, but it isn't learning. It's memorization.
Memorization is learning... How did you learn the value of pi? How did you learn the lyrics to your favorite song? You memorized them. Granted this AI (like pretty much any other AI) learns very far below the level of the average human, to the point where, if it were human, it would be considered mentally retarded. But, that doesn't mean it isn't learning. You are right, though, that it is just this level being learned and not the game.
SgtSupaman, I very highly recommend you watch Welch Labs' UA-cam series called "Learning to See." It goes through the history and process of machine learning, ending with a method called decision trees. The narrator is wonderful, too, and I don't think you'll regret it. Ultimately, the difference between "memorization" and "learning" is *generality*. This algorithm has memorized this level, but it hasn't actually learned anything. If it was actually learning, then it would be able to play another level in the game with at least some amount of competency, but if you actually tried to do that with this algorithm, it would simply use the same set of inputs, then die or never finish; it would have to restart from scratch. Think about the process of learning how to do addition. In this analogy, the algorithm in this video has memorized the value of 7 + 11. You could give it that question as much as you want, and it would always answer it correctly. However, if you asked it, say, 4 + 6, it would answer 18 - because it hasn't learned addition, it's memorized a specific example. It lacks generality. That's why memorization *is not* learning. In that sense, you can't really "learn" the digits of pi, because there isn't a way to generalize that knowledge. Knowledge of the first 1000 digits doesn't help you predict the 1001st digit. All you can do is memorize. You can't get a PhD in physics by memorizing example problems and deploying those algorithms on the tests; you need an actual understanding of physics that generalizes to most questions about the physical world.
SgtSupaman But that knowledge is useless in other situation, it's good for one thing only, like this ai is only usefully for this level, does knowing the lyrics to your favorite song help you memorize a other? Does memorizing the first 5 numbers of π help you memoriz let the first few numbers of the √2? You didn't learn the solution to every possible equation, and when you try to solve one you don't try every single number until it works either, you applied a general solution with specific changes to a specific problem
For all you picky people hiding in the comments and criticising his attempt to make an AI which can beat this game: 1. I'd like to see you try put the effort in and make it. 2. He says "AI learns", not "AI masters", not "AI defeats". The title says "AI learns" (I repeated ot just in case you missed that it says "AI learns")
Not really, it's the way it's academically defined. yes, we could say it's a population based meta-heuristic algortihm for solving optimization problems… but considering this and other comments, best we assume we have a very ignorant audience.
@@Sm00thieK what does it matter, the number of "players" and "tries" it took? These are completely irrelevant measures to state how good an algorithm is. Are processors dumb because it takes thousands of components to make the most basic calculations? Hundreds of millions to even get near what they can do today? This is a computer algorithm not a real-life experiment, players and tries are trivial resources.
Hey Guys
Its not more levels but it is the next best thing
code-bullet.github.io/WorldsHardestGameAI/WHG/
Nice work dude!
That was Dope !
Neato
Hey make a Ai that can play a game called spore that would be like the coolest video ever
ua-cam.com/video/je6MwM75CCQ/v-deo.html
Holy shit, settle down guys I get it. I’ll do more levels.
Yay!
Code Bullet lol my first comment on UA-cam in a billion years.... make them learn at least the last level
Code Bullet If you're just making it evolve to find a certain set of directions to complete the level will the second level not be like starting completely from scratch? Could you perhaps try to make an AI that can solve multiple different levels?
Also if possible try to make it as efficient as your minesweeper AI.
I have no clue if this would work, but the ai with "free will" to run around without just random movements, ah who am I kidding he probably already tried or know it wouldn't work.
“Most players have figured out how to get out of the starting area”
*Me, being fully represented by the cubes still ramming into the walls:*
Those blocks that are still hitting the wall are my spirit animal
Lamo
Oof
Lol
My family would legitimately just sit there and wait
You should create AI to play Geometry dash
Hell yeah he should
GENIUS
An AI that could beat zodiac XD
no it's to easy, cause each level is the same and you just have to tap
Léo think again, it gets a lot harder than just that, I have been playing for 3 years and I got quite good so I would know
Did-
DID YOU PLAY THE WORLDS HARDEST GAME ON *_SCRATCH?_*
IKR
Also I
Squishmellow 3 YEA I WAS ABOUT TO SAY THAT XD
M3m3 Qu33n yea me too
I thinks
Adding the incremental learning was a neat trick to get around the players stuck in the start. Overall very neat but I agree with others, I'd like to see a video where the AI learns the rules of the game, not the path of a particular level. This would be more fun to watch with more levels going forward.
If there's some open source visual AI and you could perhaps direct it to anaylze why it died visually, or easier perhaps tell it blue dots = bad alongside death = bad and high distance to exit upon death = bad I think it could be possible. But frankly I know jack shit about AI coding.
Forty Six and Two, machine learning could work. You can feed it information about where it is on the screen and the position of the dots, and then it will produce an output for what direction to move. It could probably even use convolutional neural networks where it looks at an image of the game and figures all that stuff out by itself.
There are some open source machine learning libraries as well. I think tensorflow is a popular one.
AI can already play chess, go, mario, and old atari games. Obviously, we don't expect a 7min video to teach us all the intricacies of AlphaGo, but generalizing the AI to be able to handle rules (and figure out a way to win) would be awesome.
John Hunter just to clarify, what’s happening in this video is also machine learning - it sounds like what you’re describing is neural networks (of which the confusingly named deep learning normally falls into but also many other interesting approaches), which tend to use networks of simulated neurons as opposed to evolutionary algorithms like this one.
The two are combinable too - you can have evolutionary neural networks with back propagation (one might argue that’s what the human brain is an example of).
Camper Josh what why?
All you have to code is
If “redRectangle” touch “blueCircle”
Then {don’t}
Lmaooo
Gerbbzz actually we see at the begining of the video that this game was a made on a website called scratch ( which is what i use to make allnof my games) and on scratch there are no lines of code but rather blocks. so no you dont.
thats not AI tho, your saying to artificially input commands...., however, in this case the system registers errors and learns from them..essentially getting better each time which is like the human brain....
Just telling you that it was an example he used. He CAN actually code and he does. In his Stevie the dinosaur video you can see him coding!
Attila Vinklmann ok whatevr
I WANNA SEE ALL LEVELS!!
Sergio Franco truuuuuuuuuu
The problem with doing all levels is that the AI doesnt learn to play the game. The AI that can beat the first level cannot beat the second because its not designed for the second. You would have the take the program and essentially start over the incremental learning for each level.
Kevinofrepublic who cares!!
The point is for the AI to learn, hence the name of the video 'AI LEARNS to play the World's hardest game'. It would also take a lot of time and effort to create every single section.
Yaaas
I feel like they haven't really learnt the rules of the game, they've just learnt to follow that exact path
They learned the exact path to victory
The AI simply "thinks": "ok, If I go up I die, but if I go right I survive as I saw in the past generations. What now?"
It tries a move and adds the result (move and its consequences) for the next gen
I think that's where humans surpass AI in the fact that we know how to learn, how to critically think and extrapolate concepts outside of ourselves to other things and learn from that type of thinking, while a basic AI only knows how to observe and induct information to make a path to victory without observing the major implications of what they're doing. I think a more advanced AI could learn to see outside of itself though.
@@okay2037 Yeah and more advanced AI could avoid the dots and move away when they are going to hit
@@suomusintti i feel like you could give the ai the ability to "see" the dots and know their positions. the only problem might be that if you give it that information, a program doesn't have the delayed reaction time and imperfect judgement of a human so it might be too easy once it learns the rules. Idk how ai works really but I feel like this is super simplified
When you can't beat a level so you recreate the whole level and add AI to beat it for you. Nailed it.
please make more levels! turn this into a series already!
edit 8 hours later: holy shit this got 1k likes? you guys are crazy.
DO IT
cuckoo head that would be great
Yasss
Needs all da levels plz also isn't there a second game in the series?
Agreed
AI learns to code and develop more AI.
Venkatesh G yes
Venkatesh G yes
lel
How about no
yeah... that's a NO
Code bullet: can't get past the first level
Everyone who grinded the game for hours: pathetic
Can fully relate. The second level took even more hours
It isn't that hard took me 15 minutes to get past level 1
About 5 minutes I was just fooling around next 5 I planned how to finish and the rest I finished the plan (for level 1)
Finished the game under 100 deaths
@@jessemoss6056 cap
@@duceysanem it's not tho
Look for the pattern's of the levels
@@jessemoss6056 true
2:43 Battle of Stalingrad 1942 Colorised
Oh shit
Until I liked this, this comment had 69 like.
Shhhhhhiiittttttt doooog
@Arjun
MAYBE the red ones are the soviets.
MAYBE.
Wow i never knew this comment was getting notice by people
Well thanks guys
Hey Code Bullet, I'm the developer that made this game. Thanks a lot for the awesome video, really enjoyed it!! :)
Wow! WHG 3, teaser 2 looks insane.
The World's Hardest Game please tell me how to be skilled
No I made it :)
I liked the game.
I BEAT YOUR GAME IN LESS THAN 1 MIN.
I would really like to see more levels, but I also want to see a neuroevolution approach to this. You would train it on the first 5 levels lets say, then start running it and see how well it can figure out levels it hasn't seen yet!
Goel great idea.
Thanks! He could even create is own levels that are ridiculously difficult, and see how far they can get!
+Shisha Einsatzkommando After watching the trailer, that looks like quite a game, good luck at even re-creating it!
I actually think an AI could complete it quite quickly, it is only difficult for humans because you need fast reactions, and precise movements, both of which an AI could perfectly obtain. Using a genetic algorithm like in this video would probably be the best for just one level, but to generalize the solution and complete the whole game would require a Neural Network, which would be more difficult but still possible.
That would be amazing, I love that game.
Algorithm:
if *red_player touch(blue)*
then *try again thx*
Ok but why would the dot move? It could just stay where it is
there are series of movement instructions for each dots, the best instruction will be inherited to next dots generation with little mutation to find if there are another best path probability for not to die
LOL!!
@@alfaalkaaf5335 It was just a joke m8
robot has no visual input
Now, bring that bad boy to dark souls
i dont know about that, it would just die every other second, it wouldnt get to learn. or maybe it would just become the best player there is!
Yeah ai in dark souls try it
AI Learns to play Touhou......... Generation 93857237384 still on level 1.
F
U
C
K
Y
E
S
GX yeah but he has to program every game he uses
“The worlds hardest game” is actually on scratch. You can see by the flag and stop button
f u l l s c r e e n
its probably one of those websites that just pastes a minipage into their website
He should have used the original Flash version with some Flash emulator
It didn't really learn to play the game though, it just learned an exact path to beat that exact one level.
Well, in theory, that is one of the basics of learning; patterns and repetition. But I get what you mean.
@@vegavega5344 Not really. In this case you could look at it this way:
dude claims his bot can do addition. Then shows his work... And his AI does literally just three fucking additions he taught it and can do nothing else. Kinda scuffed really.
@@krzysztofpiasek5682 Like I said, I get what he ment: Build another room with same obstacles in a different order and the AI needs to learn it all again because it isn't really learning, it's trial and error, step by step-- the AI knows exactly where to move BECAUSE of a pattern build on trial and error-- It's not intelligence or learning, but it's SIMULATED intelligence and learning, built on the core principles of learning-- Repetition and patterns.
@@vegavega5344 No, it needs to learn everything again because it knows fuck all about the world. It's not learning, just like you yourself said. But it could learn. Look at AI that fought against top Dota 2 players, that AI has true learning capabilities outside of being retarded roll a dice machine. Sure, at first it must rely on random moves, but in time it learns to react to enviroment.
Agree. This is like a speedrunner playing Mario 64 wall glitching to skip pass levels. A thinking being (AI or not) would learn the rules of the puzzle and avoid trying every corner. Take this game for example. If the wall has a crack and able to wall clip through. The early generations would rule that as "best" move and build future generation on wall clipping the top side. A thinking being would stay clear from all wall, understanding the rules of the game, and go through the dots. Gen 10 is still hitting TONS of walls here.
1:04 “The design is pretty straight forward it’s just a bunch of squares.”
-Minecraft
laughs in door
Actually, cubes.
{if going to die}
{dont}
pretty easy
Robertus That better be satire.
Xuvlar I don't know if you're being sarcastic or just a dumb b.s. on the internet
Ooooh someone’s salty
Suspicious Samuel oh then sry for the mistake
Dark Cat I didn't know 9 year olds could be on the internet
Do the players actually "See" anything, or are they just following the pattern that the best one got?
This AI probably couldn't go on to the second level and attempt to go through it with what it's learned on this level, could it?
XsaivierWolfe yup it could not do the second level at all, the solution given isn’t generalized
O V E R F I T T I N G
I guess memorization is a kind of AI, but not what I typically think of as learning.
You could say the AI was "programmed by evolution", it became perfect in the environment it was in but take it into a different one and it'll die a cruel death. As if you took a tiger into the desert.
Im pretty sure most of us would have a hard time living in some other Environment aswell
Problem is we only have 1 try ^^'
Me: There’s a second level though.
Code Bullet: What’s a level?
Watch til’ end
Bucket_head I can’t tell if this reply is a joke. :/
gotta love the "pffft doubt it" at the start its so funny
love the "just a little square which moves at the gentle speed of... WAY TOO F**KING SLOW" 1:30
WHY WOULD YOU DENY ME MY RIGHT TO SEE ALL OF THEM AT ONCE?????????????????????
Seagle kkkkkkkkkkkkkkkkk
Well, this is basically a square's orgy. Do you really want to watch it?
always
Itay Raviv it’ll do ( ͡° ͜ʖ ͡°)
Itay Raviv YEEEEEEEEEEES!
square's orgys a my biggest fetish!
Try to get AI through all 30 levels, it could be a mini series or something
Shamwow The bullet wolf it is possible to get one instance of an AI to beat all levels, just like one human brain can beat all levels
Yea like Jordan said, there's a video where Jabtils (an Ai youtuber) creates "Forrest gump" to go through every maze he creates
There are 30 of these?!
@@JordanMetroidManiac
The difference is that the "Ai" starts every level knowing nothing. Any path it worked out in previous levels is useless in the following ones. A human can understand, generalize, and extrapolate the gained knowledge. Just like a human - not really.
See, this is not an actual "AI" that learns to react to input from the game's world, it's just a heuristic search function.
It learns nothing. It only memorizes a path.
I love watching all the actors evolve over time, keep the best a different color, but would love to see the evolution of the entire set.
The fact that he remakes the game in like 4 seconds tells us a lot about the original.
It tells us that it was such an ingenious design that he was able to polish it to the extreme to make it the simplest thing possible
Yes, flappy bird was reported to be created in just 2 days, people can speedrun to make the game in 10 minutes.
Yes.. It's 2018 dumbass
TTime685 calm down
Well, what's tic tac toe? Get three of either in a row.
Simple games can have plenty of entertainment value, and they're simple and quick so people return to them frequently.
When the "bot learns making youtube vids"?
Edit: wow, didnt expect much love on this comment
PulseFire Caitlyn that’s every one of his videos... duh.
Code bullet is an AI. Isnt it obvious?
Aezrith Ferova the AI is so evolved, that it generates a pristine voice as well :)
He is not a Person why do you think he doesnt show him
Have you seen Terminator?
2:46" you go first"
"NO YOU GO FIRST"
"NO YOU GO FIRST"
"I'M NOT GOINF FIRST YOU GO"
😂😂
“let’s see if you can spot the problem”
**becomes dora**
More levels pls!
too much is never enough. :ok_hand:
Problem is I don't think ai even knows what an enemy is and that they should avoid it. Using that same ai on a different level would be catastrophic lmao.
Kevin Zhang Exactly. The AI is just learning a path through this specific maze and starting a second level would require starting from scratch.
That's where NEAT comes in
F1R3 Yeah, that would probably be useful right about now.
Despacito
Kevin Zhang 2
It's so good to get the "text bullet" back, btw 4:24 that was intentional right?
*_Stating_*
*They're* just leaving the stating area, what's wrong? Have you never been at a stating area? *There* people go nuts and start stating *their* opinions and arguments.
4:31 The Swarm of Red Dots with a green core looks like a boss
A more accurate title for this video would be "AI learns optimal route for completing level 1 of the world's hardest game"
is it optimal?
Cocodufable Right, more like “learns one acceptable route”
Cocodufable, it is, at least, very close to the optimal solution since the AI had to get through with limited steps (which isn't an aspect of the actual game). That's why it kept trying to go straight at the top instead of dropping down to avoid the dot.
OMG. I was just looking at my subscription and I wanted to watch a new vid from you, I watch a 10 min vid, AND SEE THIS VIDEO! YOU ALWAYS UPLOAD AT THE RIGHT TIME AND IS THE BEST! Keep up the good work
Also, I missed the old days of your channel. Watched a little after like 50K subs
Your next video should be an ai learning to play Tetris
LuckiestLlama I agree, yes
Great idea
Yes
Wow man this is an awesome idea.
(Tetris, flappy bird is stupid😒)
Itay Raviv yeah I’m gonna delete it 😂
I got through the first level in 19 tries then got thought to the fourth level of that game & got tired of it. The first one was the hardest one though
OKay so the people saying that the "AI" didn't "learn" to play the game need to realize something VERY crucial. The game has no randomness to it. Every stage is on a set cycle. So saying that it didn't "learn" to avoid the dots is not true. It learned the coordinates on the field that equal a failed attempt. It did this in a very mechanical way. Humans learn linearly we see most of the possible results of a movement and rule those out without trying and opt for the most logically sound attempt next. Machines learn in vectors. They attempt EVERY possible combination of movements from their current position, then when they find the most optimal path, they move on to the next section and test again. That's why the earlier generations stayed in the starting box for so long.
LastKnight0727 so, we need even better AI
LastKnight0727 STOP RUINING THE FUN FUN RUINER
yeah not like humans at all, we learn pattern recognition... it'd take even a young human one single attempt to see that being in the path of a dot is game over, and all future attempts will involve at least trying to avoid dots... this just simply brute forces random directions and remembers paths that lasted longer.... not like a human at all
in that case, the title is wrong, as it says AI learns to play World's Hardest Game, rather than "AI Learns to play 1 level"
Even if its not as fluently as humans, machines can still learn to recognize patterns over subseqent generations. In this case, its impossible because the the person who made this didn't bother to make it aware of it's surroundings.
This was awesome and I want to see more.
Also, I kind of like seeing the chaotic mess of AI's scrambling to evolve. Maybe keep the top 3, 5 or 10 to show how they differ?
Maybe the best, worst, and most average ones?
EDIT: By that I mean the top one, the bottom one, and the one right in the middle.
Meow? Meow. Yeah I know we not stupid
+UltraGaming, just clarifying so nobody misunderstands, no need to get annoyed due to that.
Meow? Meow. It’s fine, sorry xD
Isn’t there more than one level?
- Nathuan No only one
There’s like 30 or some shit
Have you not played the game/level
*T H E R E C A N O N L Y B E O N E .*
6:30
Me: HOW DID HE CODE THAT GAME
Me after 5 muinets: oh yea i forgot hes a programmer lol
Turn this into a series
Jacob May it is
Im gay
Wolfy's Adventures good for you
the crown official LOOOOLLLL
Wow! My first time on this channel and by its simplicity, clarity and cheesy (but appropriate) music, this is by far the best I've seen since Turing. Subscribed!
Dang i love this game. Used to play all the time in school when i should have been paying attention lol
SAMurai96 I beat it once when I was meant to be having a dt
I've played this game and I learnt how to beat the level easily from the 1st try by copying what your AI did and following exactly the same moves WOOOOOOOOOOOO.
Whats this game called mate?
@@ho-dg6zi world's hardest game
AI robot learns to play geometry dash! Edit: WOW so many likes!
Yes finally some good ideas
dont bring the toxic gd community into here. Take it with a grain of salt and just accept that it would be a bad idea.
you guys would probably force him to release his code, copy it, compile it into some .exe and make a new verification hack. Ew
We won't. The few who are like this overpower the rest. The Impossible Game community (still somehow there) will demand another video at the impossible game at some point. It is still a bad idea tho lol cuz gd is weirdly coded and hard to make an AI for
fake account Offense taken
Just kidding the community is toxic as fuck in the wrong places and irradiated in the good places
YES
can you get AI to beat the all 30 levels?
You could, but it would take a loooong time since the other levels are more complex and it would take many many more generations to get through them.
This isn't learning the game, it's memorizing one level
Learning is kind of memorizing
they are same thing
+Vikas Rv He does not feed any relevant input to the algorithm. It's absolutely worthless. AI basically makes random moves in all directions and the only metric is whether it dies and how far it gets. That's NOT learning of the game.
The moment you change the level (lets say, we make the blue squares move vertically instead of horizontally) the entire thing goes to trash, because AI didn't learn how to spot and avoid blocks or walls; it only learnt how to randomly sneak past that one very specific set of obstacles.
You need to give the AI an ability to detect obstacles! So the output is based on some relevant input and not a completely random movement.
Camper Josh just want for the future where the AI will pretty much learn how to kill people
Wait*
2:06 - 4:24
Song called
Worse by jincheng zhang (instrumental) enjoy
Can you show us all of the boxes. Its not hard to follow. Its just very appealing to see the boxes move in unison
I thought it meant A.I. as in Artificial Intelligence as in the computer. But then i realized it was just some guy named Al.
same... pretty dissapointed but it's still cool
yee, such disappointment
Well, the title writes AI, the last letter being and "i" and not an "L".
woosh
damnit i lost my dignity
not that i had any in the first place but damnit
More levels! And text is still pretty cool :)
Your AI went to the Prometheus School of Running Away From Things and had to unlearn everything it learned there to beat this level.
4:20 Going to work in India and China
420
Danm, they let people get stoned on their way to work
Oh I get it
UltraGaming I’m quite lost
Narcotic India and China are the 2 most populated countries, and there are many people running to escape from the start area in that timeframe.
AI learns to play life
HitzCritz 👌
AI learns that no one can succesfully play life
HitzCritz life is overrated
Do you know who Marc Zuckerberg is
Ali’s are smarter then humans so I’m pretty sure they can beat it
Funny ass video man! You're channel is such a breath of fresh air on this site! Finially something ACTUALLY INTERESTING, and inspired! I honestly wish you had more content to watch lol! I know you're projects take alot of effort and time though. So glad you got some new recording equipment because honestly I thought you were some old dude, turns out you're a young Aussie fellow who is passiante in his craft. Seriously well done; I can't wait for what's to come! Take care sir, and God bless.
Nodir Nigmonov lol if YOUR a fan of these channel, you shouldn't be fazed xD
Dana Newell you're*
I hope this is just bait.
You’re telling me this thing is gonna take my job? Okay.
You stopped me from going to bed worth it
*_FIND THIS GAME ON C O O L MA T H_*
no find it on hackedarcadegames.com
@@zahidshabir4038 no don't play it on hacked, there is no collision detection.
There actually is a difference between the Original version of the game, and the Coolmath version of the game. No joke! The Coolmath version is more of a remake of the OG version.
@Cruz Yeah, coolmathgames. Coolmath is just a nickname for it.
If this is where we're at then we don't need to worry about AI taking over.
this is a poor poor excuse of an AI.
This is little more than the hello world of AI
AI can be programmed to do whatever the person wants it to do
+SwagHags69 search for OpenAI. Elon Musk literally designed a bot that beat the almost all of the best DotA 2 players in the world in an 1v1. It's absurd.
They currently also have a 5v5 up and running. Their 5 Bots will play vs a pro team this sunday!
"most of the players have figured out how to get out of the STATING area"
Spelling errors: CHEK
I’ve completed that game before, it only took 30 mins... *I FEEL SPECIAL NOW*
Bullshit it took you 30 mins. It has like 40 levels.
Yourr not
Youre not
Cool! You need to randomize the balls start point and speed
The machines they're learning... and soon they will rise...
tbh i hope so, imagine a world full of Waifus, That would be a dream "Otaku Intensifies"
+.:Camper:. Well, they will. But they'll buy $500 worth of beer cans before they're able to open one.
Your waifu is jus a fucking drawing
nah tbh biological creatures are much more efficient at survival, an ai would know this and view its relationship with man as a symbiosis
lol
3:28 ok so at first I only got a glimpse of the words, so I thought incremental learning said internal screaming
So this guy created the game that i play on cool math when im board at school!?
uh, yeah?
Speeling Mitsake: chek
A board? Well, has anyone writed on you yet?
Aonther speeling mitsake: chek
@@michaelfrazee6990 im crying rn 😂
Bro please post the source code so we can tinker with this and see how you implemented the iterative improvement!!
DaveOnDeck
Lol this was made on scratch so there is no code 😆 just blocks
alexis miller on scratch you can also use java's code...
alexis miller it wasn't made in scratch. Scratch is stupidly simple for this kind of task. He was playing it at the beginning, but it wasn't what he used to make it.
It's absolutely possible. I spent 5 years on scratch so I would know. You're able to do nearly as much as in C++ for example. The question is how efficient is it to program in Scratch... And as a guy who spent 1775 days with Scratch I would know. Scratch is SUPER inefficient. Scratch doesn't even have the most basic concepts. That's why you have to make workarounds...
alexis miller Scratch has low level memory interfacing, pointers and supports OOP? Well, that's new to me.
Scratch is utter trash. Don't defend it. Scratch "programmers" are like kids building with Lego and claiming to be an architect.
2:10 what song is that? It’s so good!
I would also love to know. It's my new jam
L
SB TMS its this ua-cam.com/video/fVtkAGJvNu8/v-deo.html
My heart stopped for a moment seeing that car.
We need to know (I don’t trust links btw)
The old saying : "I clicked on this youtube, you happy now" !!
Are you playing on Scratch?!
Yes. Yes he is.
Sure
That's exactly what I thought.
Scratch isn't real programing language
depends what your opinion is, scratch is coding/ programing language, but it is not professional
How did Code Bullet even come across scratch???
Also why not code the AI by remixing that game?
Because it's Scratch. It can't run dozens of complex AIs. Also, the Scratch version is glitchy, because... Well, because it's Scratch.
@@Pnod666 Scratch is trying its best. I never run into bugs.
@@Kalbr0shorts I do think Scratch is good. I was just saying that Scratch is better suited for simpler games. Meanwhile, this game is more complex, meaning it won't run as good.
@@Pnod666 I do agree that scratch isn't complex, but I don't think its bugged (very glitchy)
@@Kalbr0shorts Okay. Each to their own.
The song and epicness of the ai learning reminded me of that awesome scene in skyline where the drones come in to nuke the aliens
Codebullet: Yay! Your so close!
Player:self-delete
Just fuggin, uuuuuuuhhhhhhhmmmm, do more levels n stuff yo.
And make us sandwiches why the hel not
Yeah dude I'm totally with you
Ya me too I want me a frikin sanwich
O thank y'all pretty supportive
AI learns how to feed us the sandwiches.
AI learns to switch accounts
Anyone just think this is what humanity is doing, and we’re all just red dots constantly going the wrong way millions of times over 40 generations, until one person finally solves the damn problem? And then we start all over with level 2 o.O
Chad Zimmerman true asf😂😂
Chad Zimmerman why has this become so deep, this was meant to be a fun video😭
Christof Inderbitzin
1 second ago
Fucking stop it. That hurts my brain.
that's what life has been doing since the beginning, it is called 'genetic algorithm' for a reason. This algorithm is inspired by how life and evolution works not the other way around
Chad Zimmerman
Can you not attempt to be "deep", it's annoying and obnoxious.
do ALL THE LEVELS
*AI learns to play the first level of the worlds hardest game on its 50th attempt
More levels please but make it play all levels, not separate AIs, one for each level.
The way he does AI, it doesn't matter if its the same or seperate. His AI is just a series of correct inputs. So the series of inputs either gets longer for each level, or is just on its own sheet of paper (so to speak).
At this point his "AI" is capable of playing only the first level. Put it into any other level and it will fail every time. If you "train" it for other level then it won't be able to complete first level anymore. You're right, it's just series of correct inputs for single level. There are no sensors, no decisions, no AI specific stuff. I wouldn't go so far to call this an AI. The only way he can complete all levels with an algorithm such as this one if he has separate "AI" for each level. I'd rather see proper AI with sensors and everything which could complete all the levels without training on all of them.
Make an algorithm that learns so that it can go through all of em
Go about this by having the ai figure out what results in a loss rather than a win
Ehh... it's... not really *learning* how to play the game, it's brute-forcing this particular level. The AI completely lacks generality. Even changing the level slightly would mean the AI would need to re-learn from scratch. A better, but more complex, approach would be to take the play field (perhaps at a lower resolution) as inputs to a neural network, then output a vector from -1 to 1 on X and Y, which are then used to move the red square. To make it easier, you may want to use more contrasting colors (black for the walls rather than light blue), and make the start and end areas different colors (maybe the start area could be magenta or just white). You could also use a genetic node-based algorithm like Sethbling's MarI/O. Overall, neat, but it isn't learning. It's memorization.
Camper Josh ??
Memorization is learning... How did you learn the value of pi? How did you learn the lyrics to your favorite song? You memorized them. Granted this AI (like pretty much any other AI) learns very far below the level of the average human, to the point where, if it were human, it would be considered mentally retarded. But, that doesn't mean it isn't learning. You are right, though, that it is just this level being learned and not the game.
SgtSupaman, I very highly recommend you watch Welch Labs' UA-cam series called "Learning to See." It goes through the history and process of machine learning, ending with a method called decision trees. The narrator is wonderful, too, and I don't think you'll regret it. Ultimately, the difference between "memorization" and "learning" is *generality*. This algorithm has memorized this level, but it hasn't actually learned anything. If it was actually learning, then it would be able to play another level in the game with at least some amount of competency, but if you actually tried to do that with this algorithm, it would simply use the same set of inputs, then die or never finish; it would have to restart from scratch.
Think about the process of learning how to do addition. In this analogy, the algorithm in this video has memorized the value of 7 + 11. You could give it that question as much as you want, and it would always answer it correctly. However, if you asked it, say, 4 + 6, it would answer 18 - because it hasn't learned addition, it's memorized a specific example. It lacks generality. That's why memorization *is not* learning. In that sense, you can't really "learn" the digits of pi, because there isn't a way to generalize that knowledge. Knowledge of the first 1000 digits doesn't help you predict the 1001st digit. All you can do is memorize. You can't get a PhD in physics by memorizing example problems and deploying those algorithms on the tests; you need an actual understanding of physics that generalizes to most questions about the physical world.
Camper Josh The fuck...?
SgtSupaman But that knowledge is useless in other situation, it's good for one thing only, like this ai is only usefully for this level, does knowing the lyrics to your favorite song help you memorize a other? Does memorizing the first 5 numbers of π help you memoriz let the first few numbers of the √2?
You didn't learn the solution to every possible equation, and when you try to solve one you don't try every single number until it works either, you applied a general solution with specific changes to a specific problem
I love it when it’s text only. Not that I don’t like your voice, just you tend to be funnier in text.
I cant wait to see the next level ^^
at this rate they are never going
beat the level
Can you spot the mistake?
to
For all you picky people hiding in the comments and criticising his attempt to make an AI which can beat this game:
1. I'd like to see you try put the effort in and make it.
2. He says "AI learns", not "AI masters", not "AI defeats". The title says "AI learns" (I repeated ot just in case you missed that it says "AI learns")
@@rishiishere. I agree
Agreed.
more levels please bullet, i love this stuff
Do you ever watch while your AI tries to learn stuff and then scream like "IT'S ALIVE!!!!" bc it felt like it was alive
MORE LEVELLSSSSSSS!!!!!
I’m just going to say....
THIS IS SOO FREAKING COOL!!👌👌❤️❤️
what an intense episode o-o
thanks a lot for putting this video up :)
Do all 30 levels!!!
a bot playing tetris could be cool
Imagine a game with a self-learning AI controlled swarm. Or in fact just a game with self-learning AI. Like the game ECHO but more.
I actually like watching the horde of mindless squares sprawling across the level, so if you do more levels, please keep em
That's level 1..... Try level 50
actually there are only 30
In my experience it took my about 15 mins to get to level 29. Then an hour more on level 29, then 3 hours more on 30.
you would not make it there
31 actually
yep can take you hours to get them high levels
4:22 You don't wanna hang out inside the "stating area", that's for sure~
I think calling ng it AI is a bit a stretch dont you think?
They just following specific order,
This is more like brute force
Not really, it's the way it's academically defined. yes, we could say it's a population based meta-heuristic algortihm for solving optimization problems… but considering this and other comments, best we assume we have a very ignorant audience.
@@Sm00thieK what does it matter, the number of "players" and "tries" it took? These are completely irrelevant measures to state how good an algorithm is. Are processors dumb because it takes thousands of components to make the most basic calculations? Hundreds of millions to even get near what they can do today? This is a computer algorithm not a real-life experiment, players and tries are trivial resources.
@@Sm00thieK this is an optimisation algorithm and it's a textbook decision model.
When they were all being shown at once, it looked like something out of Super Paper Mario.