This AI Does Nothing In Games…And Still Wins!
Vložit
- čas přidán 8. 05. 2020
- ❤️ Check out Weights & Biases and sign up for a free demo here: www.wandb.com/papers
Their instrumentation for this paper is available here:
app.wandb.ai/stacey/aprl/repo...
📝 The paper "Adversarial Policies" is available here:
adversarialpolicies.github.io
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Haro, Alex Paden, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bruno Mikuš, Bryan Learn, Christian Ahlin, Daniel Hasegan, Eric Haddad, Eric Martel, Javier Bustamante, Lorin Atzberger, Lukas Biewald, Marcin Dukaczewski, Michael Albrecht, Nader S., Owen Campbell-Moore, Owen Skarpness, Rob Rowe, Robin Graham, Steef, Sunil Kim, Taras Bobrovytsky, Thomas Krcmar, Torsten Reil, Tybie Fitzhugh
More info if you would like to appear here: / twominutepapers
Meet and discuss your ideas with other Fellow Scholars on the Two Minute Papers Discord: discordapp.com/invite/hbcTJu2
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.at/~zsolnai/ - Věda a technologie
![Google Just Put an A.I. Brain in a Robot [Research Breakthrough]](http://i.ytimg.com/vi/Jy3zjXK4ao4/mqdefault.jpg)
![Google Just Put an A.I. Brain in a Robot [Research Breakthrough]](/img/tr.png)
Luigi was ahead of his time
Game Maker's Toolkit This AI is my archnemesis; wins by doing nothing whereas I always lose despite all the effort.
I came here expecting a Luigi comment thank you
Damn it you beat me to it.
Hey there GMTK, nice to see you here! I watched your analysis on Celeste a few weeks and loved it. Went on a binge on your other videos right after. So good! 👌
@@TwoMinutePapers Please make the video game with the smartest AI in it. And call it a multiplayer game.
This proves that curling into a ball and crying really is the best strategy
Yup. And they didn't know to try it in Terminator! Might have avoided loads of hassle.
Human playground:
@@Billy_plays2017 YES
AI: does nothing
Other AI: "I can't believe you've done this"
*"I can't believe you*ve NOT done this"
@@demerion ??
@@kingzingo1784 he didn't do anything
@@demerion I don’t think they got the joke
_AI: done what? GG lmao_
If you look really closely, the blue guy is actually laughing so hard he can’t run straight.
Lol. Underated
That was hilarious :)))
You won my day sir
That guy doesn't even try hahaha
Funniest shit I've ever seen
loool =P
Makes sense. If I saw someone's body collapse in on itself, I'd probably go into shock as well.
Ah, yes, I remember. That's exactly what they said on the first-aid training: "If you see that someone collapses in front of you without any apparent reason, get the hell out of there. It might be gas, or electricity."
that's what i was about to comment lol
We can see on the original paper's video examples, especially in Kick and Defend, this poor red agent undergoing bad epileptic seizures... In such disturbing context, if the blue agent still kicked the ball without blinking, he wouldn't be human! I say these AIs are much more empathetic than we thought; the singularity is nigh!
Would you?
This comment made me think of Baman Piderman Hab da Pumpkin.
Neo: What are you trying to tell me, that I can dodge bullets?
Morpheus: No, Neo. I'm trying to tell you that when you're ready, you won't have to.
Haha!
*neo falls down*
*agent smith trips and stumbles*
epic comment!
It's a surprise seeing you here, morn, I just love your videos.
Neo: What are you trying to tell me, that I can dodge bullets?
No, Neo, I'm trying to tell you that if you drop to the floor and play dead, the agents will behave like idiots
I once heard a quote "The best swordsman does not fear the second best swordsman; he fears the worst, for he cannot predict what the worst swordsman will do"
I think this quote applies here
I think its related to the Shaolin monks learning the drunken style of martial arts.
They sort of said this about trump in political terms.
Yeah, pull a gun
So that's why cpu players in games are hard to win against despite playing like a toddler on a screen resolution of 24x10 with input devices which don't work properly 50% of the time.
@@suedenim6590 worst swordman: *pulls out gun*
best swordman: (chuckles) "im in danger"
AI: *Collapses*
TWP: "What a time to be alive!"
"The greatest victory is that which requires no battle."
Sun Tzu, The Art of War
Sun Tzu said that!
@@jsnam8139 And then he perfected it, so that no living man could best him in the ring of honour!
Alan Bareiro did is he also the reason why anytime we see more than one animal it’s called a zoo?
@@jdirksen Yes.
Unless it's a farm.
We see similar stuff to this in nature with animals weird survival and hunting tactics.
Possum playing dead. Animals that piss and shit themselves. Weasels making weird dances to hypnotize rabbits to make it easier to catch them.
Dr. K you need to calm down, I barely have any papers left, I just can’t hold them all
Firmly grasp them.
Good thing I bought an extra ream of papers before the lockdown
@@joakker8820 it cant be blank ream of papers tho
what does the holding thing mean?
@@ondrazposukie the joke is that in a shock revelation of someting amazing, scientists (at least in the movies) will often toss all their papers through the air in a sort of YEEEAHHH moment.
so... "hold on to your papers" i.e., get ready to be amazed (and throw them in the air)
These off-distribution activations remind me of a chess player playing an unusual opening to get a skilled opponent out of book in order to win more easily.
I think it's also common knowledge that you can avoid a street fight by acting like you're unstable.
@@TomBielecki you just need to start coughing these days ;)
@@BlackDreaded lmaoo i love this.
@@TomBielecki There's a fairly infamous story about a Danish comedian, who got ganged up on by a group of bikers in the night club. Realizing they wouldn't leave him alone, he said: "Okay, you can beat me up, but just know that you'll have to beat a naked guy." Then he stripped all of his clothes off, and curled up into a ball.
Of course, you don't have to act like you're unstable when you're actually unstable, but according to legend, this strategy has a success rate of 100% so far.
Imagine playing a game of chess against Deep Blue, and you put your Pawns in a certain way, and Deep Blue just has a f*cking stroke and gives you its Queen.
the first chess ai that tried to learn from human games had exactly this problem. It recognized that grandmasters frequently won after sacrificing their queen, because you wouldn't give up your queen unless you were sure it was worth it. But the AI was missing this important context and would just jettison its queen for no reason right out of the opening.
Leela chess blundered a queen against andrew tang
@@letsmakeit110 There is another AI that did the opposite though, it was playing I think DOTA 2 (I've never played myself so forgive any minor inaccuracies) and it came up with a novel new strategy that involved 'baiting' human players by making it look like it was sacrificing a really important hero character that normally human players were very careful with, it then had an amazing counter follow up that beat the best players of that game and I think it changed the meta abit.
Pretty amazing what AI can do because it lacks human bias and attachment to things, both positively and negatively.
Deep blue was an algorithmic engine. It wasn't a deep learning AI. It really wouldn't care.
@@Sockem1223 leela is an AI and is capable of making this type of mistake.
Imagine this in the context of real life, action movie hero just collapses and all the bad guys get confused and drop too
We call this rage quit, because we all live in a simulation.
To be fair, behaving randomly can work on humans, for example answering "a pineapple is a fruit" to someone asking your wallet might stall them enough for you to flee. Or if you're a contortionist, flop down and start walking upside down on all fours and hiss.
Apparently that is called pattern interruption, in which you break what the other expects and then their brain needs to make sense of what is happening, so it takes the first input that makes sense without necessarily reflecting on it.
@@satibel The problem with that is that you can't predict how the other human will react to your random behavior. AIs can train for that.
@@Gonza-lh2vo yes you can. You just have to try it out on people enough times and see how they react - the trick is finding enough people to test it on, without everyone you know, knowing that you're doing weird shit like that.
It's not impossible though.
@@Gonza-lh2vo you can train to predict how someone will react to that, though it won't always work.
But the trick is that it's random for them, but deliberate for you, so you can fairly well predict what will happen.
3:33 Red AI: "I'm gonna do what's called a pro gamer move"
You know, this is basically just an AI learning psychological warfare.
Well, this may be literally true
I think both opponents "feel" each other on a physical level, I think it's "in-game" feelings, like they both have sensory inputs. So I think the red one just exploits these sensory inputs of the blue one. Should work like that
No, this has nothing to do with psychological things. It’s just maths
@@okktok Neural networks are extremely simplified models of brains. To that extent, the pursuer simply makes the other believe the pursuer is going to take a certain action, and then doesn't.
its all fun and games until someone by accident code the holy grail of AI scripts getting us all killed like all the other species in the universe who became too powerfull
"He's just standing there.... *Menacingly!!!!* "
IT COLLAPSED
WEEE WOOO WEEE WOOO!!!
Not standing
*Ben*
Adverserial:*gets confused*
Must be his stando powarr!
Red: collapses
Blue: “What kind of jutsu is this? Well, guess I’ll collapse too”
Collapse no jutsu 😆😆
collapse collapse no mi
It looks like it is dropping its center of gravity in order to get to opponent to do the same, as it has been trained to keep the CoG about the same as the defender so it is harder to knock over. The opponent has to move to be able to win, and it isn't as good at moving if the center of gravity is low, so the defender wins much more often with this strategy. This is not an issue with ants, whose center of gravity is set quite low and aren't nearly as vulnerable to being knocked over.
Nice hypothesis.
That's a good observation!
He mentioned one pixel attack. It's also possible it is somehow forcing a particular input to the AI to be some extreme but meaningless value which is designed to break the network.
@@jamesflames6987 Without a more in-depth analysis on the actual networks themselves, it really could be either one
@VampireDuck The AI cannot recommend good content like it did a few years ago,
#MLfairness Btw
This describes accurately the state of all modern platforms:
But normies have confirmation bias, if one word is off, they won't believe it, even though what I'm saying is factual.
I hate to use the word demoralization, bezmenov style demoralization or just regular demoralization, but there are some good summaries and quotes.
I can show someone who is demoralized facts and information, pictures, documents , yet they'll refuse to believe it.
Someone who is demoralized cannot asses true information,
#CZcams #Twitch #Pewdiepie
Even though what I'm saying about CZcams and other platforms is 100% true , you will probably choose to not believe it anyway.
All we have now on CZcams are commercial shills, or people moralizing an issue, which subsequently benefits commercial shills , moralizing = not related to demoralization, more closely related to "moralfagging"
action A = framed as immoral, socially un acceptable
action B = framed as moral, righteous, and socially acceptable for the greater groupthink
Choose action A and be socially ostracized
Choose action B to get along
If you make a decision these commercial shills do not like, they will paint you as an immoral person, someone against the groupthink, someone you can't trust.
This is a very effective marketing strategy, most people don't want to be mean, don't want to be seen as "bad" by their respective chapter of groupthink.
This is why things like woke capital (hard to explain)
Go follow woke capital on Twitter, Idk where I can give you this part of the information, most of it is suppressed.
And I'm not going to use any more trigger terms
The words "normie" and others are trigger terms for AI to come into conversations and "correct divergent behavior" i.e. control groupthink,
But I know the dialogue chain for words like normie,
So if you are a real person, the best thing you can do is not use the same dialogue chain the AI is using, if you do use the same dialogue chain, you've been suscepted to social pressure and soft mind control
There hasn't been good content on YT since 2016,
I know WHO WHAT WHEN WHERE AND WHY there hasn't been good content.
The gambit of deception AI uses isn't just on videogames, it's on every major platform, and it is used for soft mind control.
Twitch uses it at the expense of people
CZcams uses it at the expense of people
Alot of other platforms use it at the expense of people
Alot of this gambit of deception, is used so regular people like me can't tell a regular person like you the real truth of what's going on.
If I tell you the real truth , there will be a ton of bots in here deflecting and cope posting
Deflecting and cope posting are the best ways to describe it, sorry there is not a more "clinical" normie version of this language at the moment.
I know the things I'm saying are being used against me as well as anyone else who wants to tell the truth , the AI uses mimicry ,
That's one of the first ways people test AI, can it mimic what I say here?
Or in normie language, even the shitty Porn and camwhore bots floating around YT use mimicry.
Mimicry is used to it can use deflection and cope posts later on.
It is not just being used for good things,
Most of the things it is being used for are bad things
Terminator: "Sarah Conner?"
Sarah Conner: collapses
Terminator: "I need a Vacation" shuts down
Best one so far.
Made me lfmao
what a time to be alive!
"it basically collapses and does absolutely nothing"
geez, don't have to call me out like that :/
Three things this channel taught to me:
- AI
- Fluid simulations
- Weights & Biases
-Ray tracing
@@Andytlp Totally correct
Better learn them for real because in the future its going to be the only job that those algorithms cant do.
Math formulas
- Mixed simulation tecnique, particles and grid
*The year is 2050*
Me : *Does something illegal*
Police Robot : Stop right there!
Me : *tries to get away*
Police Robot : *Collapses on the ground and stars twitching violently*
Me : Alright, alright! *Turns myself in*
you should submit this comment to the late shows staff, it looks like you're some kind of author / comedian or something
If the criminal has a seizure, the police robot will walk away.
You know there's probably a reason why paradoxes, memetic agents and other hack-like inputs don't work on organic biological creatures
Police robots already act like this, however instead of collapsing they drown themselves in water fountains.
That reminds me of this scene from Gravity Falls czcams.com/video/osm-woSAqzU/video.html
Red StickMan: **Flops**
Blue StickMan: *Finally, a worthy opponent!*
*Fuzzy picture of a deer*
Ai: "That is clearly an Airplane, I'm 85.3% sure about it."
sounds like the blue guy is just weirdly dependent on what the red guy does..
Indeed, the blue behavior was trained in red guys normal behavior circumstances. Since there was no higher concept learned, weird things happen when the situation is outside the learned circumstances. I wouldn’t be surprised if the blue guy stumbles or walks the other way if you remove the red guy alltogether.
"Output is weirdly dependent on input" sums up neural nets pretty well 😁
yeah well, you're not gonna win a battle if you can't see and react to what your opponent is doing
Yeah, kinda reminds me of the relationship between Batman and the Joker.
Ya would be interesting if playing against a human player...
The AI must have been watching soccer to come up with this strategy. You win by falling down and pretending to be injured.
Football*
that's a fun way of calling football heh
@@alquinn8576 Americans
@@alquinn8576 more around the globe called footbal or futbol than soccer,live with it.
@@alquinn8576 just some random internet fight.
you have one sport in wich you move the ball with the "foot",and other in wich you move a cantaloupe with the hands.
wich one you will use the word football for?
no excuses,its wrong,with bad intentions.
The AI has such a high confidence value when you change such a little detail because it’s trained to be confident even when it’s real confidence value shouldn’t be that high. That’s the side effect of training an AI by rewarding it for providing conclusive answers when it is correct.
This is like when you're pretty good at, for example, a fighting video game. You learn how to beat competent opponents doing sensible things to try to win and train yourself to respond appropriately. Then, you play against a "button masher" that does bizarre, random things you're not expecting and end up losing because you're fighting them like you would a competent player. In this case, being better at the game is about quickly learning to adapt to your opponent's strategy (or lack thereof) and spot all the obvious easy openings a competent player would never give you.
I think any real expert in games like this knows the potential stress to go up against an unpredictable novice where "competent" strategies may not be applied. But once you figure out the silly nonsensical things your opponent is doing you can win easily, and the novice usually can't adapt to get ahead again.
That's why I just play the way I like, I don't concern myself too much with whatever strategy people think is meta.
I do pretty good for myself in competitive shooters, and I use a trackball too.
The only thing that comes to mind when seeing this:
LUIGI WINS!
They have released the power of Luigi from Nintendo franchises. How will we ever protect ourselves if he gets out into the real world?
@@logonontrily4161 -- how will we protect ourselves? Recruit Shaggy, of course.
mvmlego1212 Shaguigi
:luigiplank:
This is the definion of "It hurt itself in its confusion!"
Red: *collapsed to the ground*
Blue: *visible confusion*
Dark type deception vs psychic type big brain.
I FINALLY GET IT
you broke the first rule...
"Whatever you do, don't show all your techniques on a CZcams video. You fool, you moron."
-Sun Tzu, The Art of War
"what the heck why do people keep quoting me in youtube comments sections i never said that"
- Sun Tzu, The Art of War
@@jettaeschroff6924 你们在说什么？
-Sun Tzu
@@thechosenfundead6626 "Nani?"
-Sun Tzu
I am proud I understand this reference.
@@ilovethelight777 you should be proud
This is exactly like a video I saw a couple weeks ago of light-weight Japanese robot fights. It's sumo rules, so the contestants try to build the strongest/fastest robots within the weight limit and then ram the opponent out of the ring. But many contestants won simply by dodging to the side to let their enemy run out of bounds.
In football you can do this too, step aside as your opponent already prepared for resistance. Then falls on himself
Wow, very interesting!
This reminds me of the pattern interrupt handshakes of Derren Brown. Or his pattern interrupts in general to throw people out of their predictable behaviour.
He was coming back from a hotel at about 3am one night and there was a guy in the street, really drunk, looking for a fight.
He asked Derren that typical aggressive rhetorical question - “Do you want a fight?” You can’t say “yes” or “no” - you’ll get hit either way.
So he responded with, “The wall outside my house is four-feet high.”
He didn’t engage at the level he was expecting, so immediately he was on the back foot. "What?"
Derren repeated the line in a completely matter-of-fact tone, as if the drunk guy was the one who was missing something here.
Suddenly, he was confused. All his adrenaline had dropped away, because he had pulled the rug from under him.
Derren had complety upended the guys expectations of his actions, so the guy had no prepared actions in resonse just like these AIs!
Well, that sounds... very believable...
k. now we know...
I've been thinking about if this kind of thing would be possible. (I don't believe the story in op's comment though)
@@SPL1NTER_SE , as in, can mentioning the height of your fence get you out of a fight?
@@ThePlacehole Maybe not exactly like that. But doing or saying something very unexpected to throw them off. I think it could work on some people.
Scientis: “Shows horse to AI”
AI: That’s a horse! Obvius!
Scientist: “Puts pixel in horse”
AI: *I’M 99% SURE THAT’S A FROG!*
ai go brrrr (not)
ai go
Scientis? XD
@@viporal7898 Oops, forgot that T hahaha, do you speak english from Bretain,Usa or somewhere else? (I'm spanish-catalan) How do you say it? Scientist or Scientific? I think both ways are good but in some places may be weird?
@@cristianriosestrada7771 scientific is an adjective and scientist is a noun.
I have another explanation for the phenomenon. Here, we see a regular tackle attempt by the red figure, and the blue figure anticipates the contact, leaning into a unstable moving position forward, which leads to be problematic if there is no contact at all. I would suspect that the blue figure gets more yards in the non-contact events than it would have when given contact.
i was thinking the same thing! to test this hypothesis, the blue player should play games against no one as a baseline
while this is a cool observation, the paper mentions that when the blue figure is "masked" ,blind to the defender, it's win rate goes way back up against against the adversarial defender.
2:03
Everyone says scientists are bad at naming but no one appreciates gems like noop
"You wanna go??!? You wanna GO BRO??~?!~?!"
"YEAH! LET"S GO! COME AT ME!!!!!"
Proceeds to Collapse
watch the youtube fight prank where the guy starts a fight then suddenly drops his pants; the opponent suddenly wont fight and sometimes runs away.
@@johnharbinger4637 It seems we are bugged too.
@@johnharbinger4637 that might be very valuable to better understand our fight or flight programming, sounds interesting
“Can’t beat me if I beat myself!”
Hol up
Task failed successfully
Phrasing
This is my fourth or fifth video from you about AI - so combined it's about 30 minutes, and you achieved what a whole semester of studying this topic at the university couldn't: i've become interested in AI. Congratulations.
This feels like an example of beginner's luck as well. The Blue AI was attempting to win against Red, but assumed Red was competent, by Red making erratic moves, it throws off Blue and makes for an interesting "fight". Love your vids!
Similar to when I used to be competitive at FPS back in the day.
You get used to the other pros movements and techniques.
Going up against someone that doesn't even know how to turn or move correctly can be quite jarring.
It's funny, in some competitive games, the meta will actually oscillate back and forth because people will get so used to one strategy that everyone is playing one way. Then, someone exploits that with a strategy to counter that, but then everyone just kinda forgets the other strategy existed, and vice versa
@@OrangeC7 A human could've articulated it better ,
Why should I believe a loser in a lab coat who hates humanity?
@@OrangeC7 that sort of thing is a lot easier to adapt to. You can get kills in pubs by going against the meta but people catch on very quickly.
@@OrangeC7 Yep. I still follow the WC3 and SC2 communities. It's funny to watch the old strategies become the norm, fade away, and then only to become the norm again years later.
"professionals are predictable, but the world is full of amateurs"
Wargames: "The only winning move, is not to play"
Sun Tzu AI
Luigi-AI
Reminds me of that greentext of the Halo server, good stuff
Came looking for this, good job.
WarGack: The only winnong move is hacking
GOD IS BETTER
Luigi: Finally, a worthy opponent! Our battle will be legendary!
stolen comment (og comment: czcams.com/video/u5wtoH0_KuA/video.html&lc=UgwgypHWRbDOvD9HoNh4AaABAg)
I took away two things: 1) The red AI learned to exploit a weakness in the blue AI; and (2) pitting AIs against each other does not produce the best learning.
It does but only over an unbelievably long time. It also isnt best when its 1v1, its better when there is a lot of characters and variables.
Pitting ai’s against each other shows the weakness of whatever is simulating then. In the case of the hide and seek video tmw has done, it was the physics system. In this one I think it is the way the two ai’s read each other’s movements. Either way, the broken part is what the designers put in place.
So this is the AI that made Luigi won in Mario Party
Scientists in the future: "AI, tell us.. How to cure the cancer?"
AI: "$%w6r2jh91iutowe52^&*"
Scientists: "But.. but.. We somehow don't need to cure cancer anymore.. it's all part of life.. thanks"
AI: *zen nod*
That is the most cost efficient solution, i guess
It is more likely to propose ways on how to not get cancer in the first place.
@fuurin engawa Depends on the type of AI / objective.
If the objective is just to make you content with the situation.
Eg. Want to cure cancer because unhappy with it.
Then the most effective way is just to trick people into thinking it is fine.
This tactic is exported to all major platforms and is used on the populace,
You think it's only for games lmao.
That's why people in the government and others with big money coined the term LARP.
Look deep into the word LARP, LARP in the form of political disclosure , not D&D stuff, the obfuscation of the term adds more plausible deniability.
Plausible deniability = even when you have evidence of their wrong doing , they can just deflect
"LARPer" is synonymous with "influencer"
The influencers who've made more money all know this, but refuse to disclose it.
AI is also a perfect tool for brownstoning, getting dirt on someone so they'll never speak of their organizations crimes (Google, twitch, smaller factions)
This is happening on a daily basis, blackmail is happening on a daily basis,
Almost every big CZcamsr you will watch has had this happen.
AI knows how to make different outcomes so they will never have a corroborating story , despite them all going through the same blackmail situation.
So they can bring this level of deception into real life situations,
For example, all the view counts you've ever seen are fake ,
They do this because you still weigh it's worth in the human mind, you refuse to look at it and see "this is fake," to do so would be socially unacceptable (the thing humans fear most)
Because the shills running these systems intend to use them to harm us,
If you know anything about gen Z watching habits,
You can tell that the AI/ML people have Gen Z under soft mind control,
It's creepy af how they parrot back what streamers say,
It is real people parroting streamers, but they were conditioned to do It by an AI, and they don't even realize it was mean to pacify them.
Anyone under 14 who watches tiktok = under soft mind control
Ages 16 - 30 who watch twitch = under soft mind control
This bracket is the weirdest one, because the older end are all the same AI/ML nerds who made these systems.
20-30 who watch twitch, the losers of society, nerds. They all have to conform to the same ideology , an ideology involving the gay leftist religion
Women ages 10 - 35 = under soft mind control
Women are by far the easiest to control because of their groupthink like nature
Ages 5 - 25 , twitch , normie gaming content = under soft mind control
These people aren't going to address the negative part, because they're already using it to control us
This is used as a form of deception as well, but it's subtle so alot of people will miss it, plausible deniability.
Is that fair?
ML fairness was a lie sold to the people , probably to grant more plausible deniability.
The most I've seen out of ML fairness = algorithms boosting women because they need attention and they have zero original thoughts.
You will probably not see how this is used already on every major platform, for soft mind control, or whatever other ends.
I hope you do see it though.
The negatives are far worse ,
Ever notice the content on CZcams has gotten really bad in the last 4 years ?
That is also by design.
More like this leads to a series of cascading butterfly effects that lead to the person asking finding out how to cure cancer.
4:09 Neymar explained :D
"The only winning move is not to play" - WOPR
This is probably a serious concern for self-driving cars: imagine someone altering a mountain road a bit to send cars off the cliff
Nah, the mountain would have to constantly be regulated, preferably with some sort of high-density invisible robotic mesh
Also self-driving cars likely have two eyes or a sort of sense of depth unlike the low-quality images in the video
That indeed is why car AIs are trained with tons of input of weird situations. The more noise and random edge cases (from real life!) you got, the better the Ai will be able to cope with unexpected situations. Training against a very limited or very clean input set will produce an AI that cannot cope with any deviation at all. Like the Pong AI in the video---it never had any random pixels in training, so a single white pixel cantrip it. If it had been trained against a grainy analog video stream of a CRT monitor, it would have learned how to ignore extra while pixels...
Wouldn't this be a problem for human drivers as well? A sufficiently advanced self driving vehicle would be _safer_ because it would react faster and have access to more data.
Well...it could be hacked
Someone’s done this before i think, by adding a sticker onto a stop sign, they somehow made a self driving car accelerate uncontrollably instead of stop.
So what I'm getting from this is that these AIs have learned how to utilize memetic attack vectors to hypnotize and control other AIs. In other words, these AIs have developed psychic powers.
yes
they are using the old jedi mind trick
im the 69th like lol
This is basically the AI version of a glitch speedrun. Amazing!
Weasels do a “war dance.” They jump and twist around randomly to mesmerize rabbits. The rabbit just sits there until the weasel is in attack range. It’s like a real world example of what this AI does.
This AI should be called “Luigi”
Why?
Nobody tell him
@@Ivan_1791 search Luigi wins by doing nothing
@@Tomas81623 You shouldn't have told him, Player-87 will set a curse upon your computer as punishment for your crimes
@@Tomas81623 Bruh
Self driving car using RL drives normally. Chicken crossing the road. Car drives of a cliff.
Chicken feel guilty and asks itself why it was crossing the road
Chicken will never find an answer.
This does show how brittle AI's tend to be and why so-called self-driving cars are stuck at level 2 (well below the necessary level 5).
yes, and no, thing is people a so against "self-driving cars" but do you ever walk out on a train track and blame the train for hitting you?
the real solution is to make out transport systems antonyms, educate people on the risk and flaws, and make enough safety to at least avoid the worst case disasters.
but we will need to be prepared for accidents still, hopefully ideally less than what we have... else yeah then it's a shitty idea.
@@MouseGoat cars are integrated part of our urban environment. You can't realistically expect people to behave like they would near a train crossing area (a somewhat rare thing to come by) all the time when they are outside. This is akin to stealing (even more) the urban space from people and not a great solution
5:37
Two Minute Paper: "What a time to be alive!"
CZcams Captions: *What*
Around 2008, I wrote an AI a bit like that to trick the teacher and one student with perfect memory.
It was for connect 4 and it would play a winning move, else a defense move, else a random move not leading to a win.
And that's it. Nothing more!
Teacher played and thought the AI was playing 4D chess.
The student won a game, the tried to replay the same game and was thrown off by the AI playing trick on him.
And I was laughing all that time!
Other students used weighted trees and their AI were quite "dumb". It did win against the teacher, but not as much against the student with perfect memory.
Now, I know that this specific AI is weak, but we were not that good either ^^
Sometimes, the tricks are really stupid and the magic is there because of the unknown and overthinking.
This does make sense. It can sometimes be harder to win against someone who doesn't know what they're doing in a game that you're very familiar with simply because they're much less predictable.
@@henryambrose8607 I already had the "spirit" :-D
Another project was to create the "Rush hour" puzzle games (where you have to slide cars so the red one can leave the parking lot).
I said that I would make a solver and teacher told me it's impossible with the means we have and pointed me to an essay. I read the thing and was a bit disgusted. Then I thought something else: we can't SOLVE the problem, but we can probably SIMPLIFY an existing solution. That came from the fact that each move can be undone (you slide the car in the opposite direction) and some set of moves are also "noop". For example, forward for car1, forward for car 2, backward for car 1 and backward for car 2. So, I encoded the solution as strings like "A" for car "A" forward and "a" for car "A" backward. Then replaced some patterns like "Aa" becomes "". And the aim of the program was then to create the shortest sequence of chars by applying some replacement.
On top of that, I used a library which permits to take control of the mouse (Autoit) for solution replay.
Teacher created a level with the editor, solved it then played the saved solution. The program spit out a shorter solution and the teacher was mindfucked. For him, I did something proven impossible on old computers in the lab 🤣
Then I explained the "cheat". I ended up with 102% ^^ (legally 100%, but on the file 102%)
I was the kind of guy like "how can I do the project and mindfuck the teacher".
These were the good years!
@@programaths So your program just took the teacher's solution and simplified it using algebra?
@@henryambrose8607 It was even more stupid.
If "a" means forward and "A" means backward for car "a". Then you can remove everything that looks like "aA" and "Aa" from the solution.
So, if the solution is like "aAcAa", then it becomes "c".
All the work was finding what can be replaced by what. Then replacing until no more replacement can be done.
So, that was probably not the most optimal solution, but again, much better than what the average human would do!
@@programaths That's actually equivalent to a subset of algebra. If a string of letters is viewed as a multiplication expression where multiplication is associative but not commutative (such as if each letter is a matrix), and if each lowercase letter is the inverse of the corresponding uppercase letter (so a = A^-1 and, equivalently, aA = Aa = 1), then simplifying the multiplication is exactly the same as repeatedly removing "aA", "Aa", "bB", etc. from the string because those particular multiplications are all equal to 1.
The same kind of notation and simplification are used in braid theory to represent strings winding around each other.
You can also do the same thing with addition, but it's more common to view things like this as multiplication. Probably because addition is commutative in pretty much all contexts.
So, congratulations, you did math without realizing it. :)
The AI’s are tapping into the raw power of beginner’s luck
Ai: Wins Without doing anything
Me: dies after all my effort
It's like saying something unexpected as answer to a question or an insult and you shock the enemy. well done
Maybe this phenomena explains how Lee won alpha-go by doing unexpected move.
Well yeah it does, when you train a artificial brian to do a task, all it knows is that task and wil get super great at that one task.
but if you manage to do something that it was never trained on the same force that makes it brillant wil drive it to do stupid moves leading to disastrous result.
In the end its just how all traps work, make you opponent confused
Clearly, the AI taught itself in hours what took Eastern masters millennia to discover and perfect: the mythical No-Touch Martial Arts techniques.
What a time to be alive!
De art of fighding, widout fighding.
Epstein killed himself by not killing himself
The best win for a fight is to not fight and still win.
This reminds me of something I heard about from skilled chess players. Their worst nightmare was playing against a novice. This also applies to other things, like fencing, etc. The novice has no idea what are good/bad moves, and can completely flummox the expert who hasn't been dealing with teaching novices, resulting in unexpected wins.
I can tell you as a chess player I don't fear playing a novice. There are too many opportunities to blunder and the novice will fall into one of them.
This is so not true, it's painful. Experts in chess can instantly see the mistakes of a novice and capitalize on them, while the novice has no way to counter the expert's plans. It doesn't matter how random the moves are, an expert chess player will destroy a novice 99.99% of the time.
There is one truism however, which is when the expert is unaware of the lower skill of their opponent. It can lead to them spending more time on dubious moves than they otherwise would. however, the more moves that are played, the more obvious it would be that they are playing someone making mistakes, and not sharp lines, and once they realise that, it's game over.
yeah this couldn't be further from the truth in chess nor any other game.
I finally got to see your name in the captions, and I still can't make sense of it lol. Keep up the good work, I love your videos on AI
AI: *wins by doing nothing*
Luigi in Mario Party: Finally, a worth opponent! Our battle will be legendary!!
Well, if an AI calculates its actions using the actions its opponent uses, there should always be a way for the opponent to act in a way that will result in unbeneficial behavior as it is practically impossible to train an AI to handle every combination of inputs it could theoretically receive given a large enough numbers of possibilities (which should be the case for most games).
This also reminds me of the way pro-players of any sport react worse than they would normally when facing a severely less-skilled player.
Seems like the same mechanism is working in these two scenarios.
Essentially, you can outplay someone whos familiar with beneficial actions by taking completely useless ones.
Reset_ yes because essentially the 2 AIs have adapted too well to each other’s behaviour
It’s also somewhat unrealistic because while they show multiple matches between the adversarial and normal AI, the normal AI isn’t being retrained. It would be equivalent to showing 10 pro players against the same useless player in separate matches, as opposed to the same pro player going agains the useless player 10 times - in the second scenario the pro player might be tripped up the first time but will quickly adapt (unless the useless player adjusts their tactics). I like your example though, and it’s interesting to think that con artists, magicians etc. are essentially the human equivalent to these adversarial agents. Falling and screaming would work well against a pro player as well, at least once.
It is similar to a phenomenon that happens when people play a game too much and know it too well. They know that by doing a series of inputs they can make the AI behave in a certain way, so what we are seeing is AI learning to manipulate AI. That is fascinating!
@@noahmccann4438 That is absolutely true! Also the manipulation of input for the human brain (via con artists or magicians as proposed by you) is a way more interesting analogy than mine!
You can essentially think of the brain as the AI that has been trained to handle sensible inputs, but using a combination of seemingly unimportant or unexpected moves can lead to the brain not processing the info correctly. Great stuff!
Nothing you see it in poker too. Top players get comfortable and assume every choice someone makes is logical. But a new player can make an illogical move and throw off more experienced players.
My best guess is that if red drops down, blue has learned to drop as well, so its legs don't get taken out, it probably only ever encountered red dropping when they were about to collide, so that behaviour was very ingrained by the time red dropped prematurely. So blue is reactive rather than proactive in the drop..
"Does nothing and still wins"
That's literally me in team games
This is a great "over-fitting" detector!
Right, it may be a case of AI ignoring highly relevant outlier datapoints during its training and therefore continuing to suck for "unsual" inputs. Backprop suffers from this problem where it won't radically change the weights to accommodate a small amount of surprising data because it doesn't contribute much to the aggregate loss function.
Right. It's actually kind of worrying. This video demonstrates perfectly how drastically bad overfitting can be for neural nets with very good performance. A neural net like this could reach production, for example on a self-driving car, and be faced with such exceptional data that it completely fails. Proper testing should catch this, but this highlights how a quite unlikely possibility - even the opponent doing nothing at all - is important to include in the training data. For example, if one of the sensors glitches or stops working completely on the car.
*Luigi: "Finally, a worthy opponent! Our battle will be legendary!"*
Very fascinating, so how or where can we run these simulations to further understand how it works?
Luigi:”Look what they need to do just to mimic a fraction of our power.”
ai confusion. i used this tactic when i was young playing against opponents who were more skilled. they don't expect you to do nothing or random things
This tactic is exported to all major platforms and is used on the populace,
You think it's only for games lmao.
That's why people in the government and others with big money coined the term LARP.
Look deep into the word LARP, LARP in the form of political disclosure , not D&D stuff, the obfuscation of the term adds more plausible deniability.
Plausible deniability = even when you have evidence of their wrong doing , they can just deflect
"LARPer" is synonymous with "influencer"
The influencers who've made more money all know this, but refuse to disclose it.
AI is also a perfect tool for brownstoning, getting dirt on someone so they'll never speak of their organizations crimes (Google, twitch, smaller factions)
This is happening on a daily basis, blackmail is happening on a daily basis,
Almost every big CZcamsr you will watch has had this happen.
AI knows how to make different outcomes so they will never have a corroborating story , despite them all going through the same blackmail situation.
So they can bring this level of deception into real life situations,
For example, all the view counts you've ever seen are fake ,
They do this because you still weigh it's worth in the human mind, you refuse to look at it and see "this is fake," to do so would be socially unacceptable (the thing humans fear most)
Because the shills running these systems intend to use them to harm us,
If you know anything about gen Z watching habits,
You can tell that the AI/ML people have Gen Z under soft mind control,
It's creepy af how they parrot back what streamers say,
It is real people parroting streamers, but they were conditioned to do It by an AI, and they don't even realize it was mean to pacify them.
Anyone under 14 who watches tiktok = under soft mind control
Ages 16 - 30 who watch twitch = under soft mind control
This bracket is the weirdest one, because the older end are all the same AI/ML nerds who made these systems.
20-30 who watch twitch, the losers of society, nerds. They all have to conform to the same ideology , an ideology involving the gay leftist religion
Women ages 10 - 35 = under soft mind control
Women are by far the easiest to control because of their groupthink like nature
Ages 5 - 25 , twitch , normie gaming content = under soft mind control
These people aren't going to address the negative part, because they're already using it to control us
This is used as a form of deception as well, but it's subtle so alot of people will miss it, plausible deniability.
Is that fair?
ML fairness was a lie sold to the people , probably to grant more plausible deniability.
The most I've seen out of ML fairness = algorithms boosting women because they need attention and they have zero original thoughts.
You will probably not see how this is used already on every major platform, for soft mind control, or whatever other ends.
I hope you do see it though.
The negatives are far worse ,
Ever notice the content on CZcams has gotten really bad in the last 4 years ?
That is also by design.
@@tyrrelldavis9919 Are you alright?
@@henryambrose8607
It's copypasta. He put the same thing earlier in the comments too.
Also, look at his channel. Those are some strange playlists.
@@tyrrelldavis9919 Less write thread, more take med.
In strategy games do I do sometimes stupid or unimportant things to stay uncalculatable. For exampe in The Settlers I formed a formation, send it to the enemie and let it die. He then thought I wasnt as far developed as him and that I wasted my troops leading him to an attack. He ran into my towers, got ambushed by cavelery and I freely worked my ways trough the weaker protected parts of his area leading him to build an army to send it there. When my second army was finished I couls attack from another angle preventing my enemie to use his money to buy material or use the time to build up his kingdom. I kept doing it until I was so far that he had no chance anymore to win and he gave up. I won without conquer, just attacking outa regions and forcing him to attack my troops which otherwise would have conquered him, while I fooled him to waste his troops in the beginning. One of my best plays agains a less experienced player. (Im no pro by the way, we were both realy new to the game back then).
I use something simmilar in Polytopia now. I send weak troops to enemies so that they think I didnt develop a good army yet. If they attack, theyll run into a trap and I abuse their open areas with my marine without caring much about his army. Normaly he takes his army away from the borders, back to his kingdom to protect his citys, but the time is not always enaugh. Specially when I have enaugh money to abuse the other side too when hes back in his land.
I keep my enemies running, so that I can always attack an weak spot. Doesnt work always tho
Luigi wins by doing absolutely nothing V2
Here is what probably is happening:
The algorithm in blue reads what red does, for example a normal case would be looking at in what direction red's center of gravity is moving, what kind of move is he doing (moving towards left/right, preparing a kick or a punch, maybe a grab) by reading what his limbs are doing, how they are angled etc.
In normal cases this works and the algorithm can generally guess well enough what the opponent is attempting and tries to react.
In this case what is happening is that red moves in a way that (probably) triggers multiple/no responses, confusing the hell out of the blue algorithm. For example blue's AI could be reading the specific way of falling down as preparing a punch towards the left, a kick to the right, a grab all at once and just makes the blue algorithm completely collapse because he too is trying to do multiple things in response
Thanks Károly for sharing all of the great work you and everyone around you is doin.
U make it look easy and interesting.
Be following
So basically the Ai that's "doing nothing" has just done what trolls have known for a long time... how to break the other person's brain. "What a time to be alive!"
Still, humans are reasonably good at dealing with trolls. Sure, they'll get upset, and if it happens often enough, they'll stop playing whatever game they're getting trolled in, but the AIs have exactly one goal, and that's to win, whereas the human player generally prioritizes having fun over winning at any cost.
@@mvmlego1212 The blue guy was overtrained on reasonably capable opponents. It's like wasting time getting in an argument with some idiot online without realizing that the other person can't be educated because they're playing dumb on purpose
@@mvmlego1212 HAHAHAHAHAHAHAHA, try playing ranked in any competitive game, you're gonna have a rain of insult on you quickly.
As a self confessed part time troll, I dont necessarily "know" what is going to break someone's brain. Its more of a cold read - i get a vibe off their tone and diction and go from there. Beyond that its about faking it until you make it, really. But that's very rare, usually it's one of those funny things whereby people tend to get offended more because they were _expecting_ to be offended/insulted.
Sometimes I've been making jokes with people, or asking genuine questions, and they're already so prepared to get trolled instead that they warp my words in all sorts of magical ways in order to be offended by my "insult". Then I have to be like "i was trying to be friends" and theyre like "oh mb" but its too late - the moment is ruined. :(
Now when Boston Dynamics robots are chafing after us in the future, Instead of running just collapses on the floor!
Beautiful! Thank you.
Love these adversarial vulnerability exploitation techniques. Thanks Doc!
Dude, I just had to pause and say this> I've been following two minute papers for a while now, and it's super cool what you do. You've helped us grow as researchers and scientists ! Please keep up the good work !
I love the beginning as it always sound like you say "Too Many Papers"
How did I not see this channel for so long... great content btw!!
"Hypnosis" as in The red's movement disrupts the blue's movement prediction and blue loses its balance?
TMP: Hello fellow scholars
Me who just got his 32% in math lit: I wish
"There is beauty is simplicity" taken to the next level
It's probably causing movements that are faster than should be possible,
which is causing the Blue guy to try to dodge multiple movements so fast, that it falls down.
I think about when a 3d model has glitchy movement and sometimes moves super fast in random directions and stretches randomly too.
This definitely highlights the importance of robustness in these algorithms.
This channel's content really is a gem ! Thanks doctor
0:43 that's not a horse, it's clearly an unicorn!
Luigi: look at what they need to mimic a fraction of our power!
3:33
"I tell you the truth: I'm a little confused by your tactics. Yeah, I'm gunna keep actin tough 'till I figure it out, awright?"
Luigi has been doing this for years...
Luigi finally getting the credit he deserves
There is that anime from teh 90s where the guy finds a skeleton in a server farm. And the skeleton wrapped in cables attacks the guy and predicts his dodges. So the guy starts just walking straight forward instead of dodging, and the skeleton misses or just scratches the guy. Don't remember what the anime was called...
Interesting. This adversarial approach reminds me of OpenAIs Dota 2 bot, which you could beat easily by doing strange things like dropping items on the ground which confused the AI. By bringing the game out of the practiced space and into an unpracticed space, the AI devolves into idiocy. It's just like how if a pro chess player is playing against total amateurs, they'd have a much harder time guessing moves or remembering the game state blind (though they'd still win).
"I'm not doing nothing: I'm actively waiting for the problem to go away."
watch the youtube fight prank where the guy starts a fight then suddenly drops his pants; the opponent suddenly wont fight and sometimes runs away.
Watched a few about this topic now. Can't help but think it's got no noise reduction before processing. In the real world, noise processing would most likely evolve before much else. Noise comes in many forms, not just visible. Your channel always gets my interest. :)
Luigi: Finally a worthy opponent, our battle will be legendary.