The AI Gaming Revolution

[@sclair2854]

No game challenges left? When an AI successfully navigates an entire D&D campaign without obvious meta-gaming I shall be impressed! D:

[@Chocl8215]

How would an AI not meta-game? That's the whole point of the AI.

[@CH-bd6jg]

>How would an AI not meta-game? without obvious metagaming, what +S Clair is asking for, is a D&D ai that passes the turing test, so it can create good in-character reasons for all the metagaming it does. just like how humans metagame, really.

[@sclair2854]

Chocl8215 By the AI playing the role it's given to the maximum of what the role would be, not the min-max of D&D itself. If I say gave the AI a gnome ranger, and let it randomly generate from the flaws and ideals lists, it should not only try to maximize that roles damage, but play that role correctly. It shouldn't for example figure "If I multiclass to wizard, get to level 17, and pick shapechange, I can turn into a dragon and get the highest numbers" "Winning" d&d is about having fun, a AI that could win D&D would like Christian Higginston says, have to pass the Turing Test, then have to understand the game to coordinate with allies, and then have to understand probability well enough to make intelligent dice rolls. It involves very high levels of nuance, that I doubt a weak AI could accomplish. Tbh if an AI could play D&D and be indistinguishable from a human, humans would likely be obsolete already

[@CH-bd6jg]

yes, and that's a D&D implementation of a freaking Turing Test. that's all you're asking for, not an AI that outplays humans. that's not the same thing, and yes, it's a decently unsolved problem so far as I know.

[@sclair2854]

Christian Higginson Completing a campaign is beating a human, it's beating a DM. I'm not saying its reasonable, Im saying that theres a whole genre of games AI's cant even touch yet.

[@cpob2013]

but can it find something to do in no man's sky?

[@smilingaxolotl9576]

FUCKING FIRE

[@gafeht]

asking the real questions here

[@SanctuaryReintegrate]

Savage af

[@tisisonlytemporary]

Quit

[@genericprofile4847]

lmao

[@seahawk124]

I'd love to see an A.I. that can play Monopoly without losing its temper, and end up wanting to kill everybody!

[@AvailableUsernameTed]

I'm for letting AI's handle the real economy.

[@ireallyhatemakingupnamesfo1758]

Oh my god, think about it, we would get to the point where computers can take over the world before then, so once we get to that point and we test it, humanity dies

[@zebpeep1390]

It will be like the Magi system from Evangelion.

[@TheSirGoreaxe]

That is why Skynet is out to kill us all.

[@joshuabianchi673]

that would cause the singularity

[@samk4911]

You can't call it an Eastern Chess if chess was invented in the east.

[@javonyounger5107]

Go is more eastern.

[@TheBelrick]

chess came out of india which is middle. and it was the west that formalized it into the version we know today

[@niveshproag8660]

If the Muslim countries are the middle east, I don't see how India is just middle.

[@TheBelrick]

Nivesh Proag they are only muslim through conquest as islam exists to conquer the world

[@niveshproag8660]

Bel Rick A bit out of subject but you pic is new to me...looks like the stars on the confederate flag, on a rainbow flag. What does that mean, lol. Are you secessionist and pro-homosexuality? An unusual mix. Or did I misunderstand?

[@linkinparkundrground]

It's actually awesome that AI could process 100 million moves per second, and yet Gary Kasparov could still beat it. Shows what the human mind is capable of.

[@HyruleJose]

Sure A.I. can beat us at every game ever but where is that intelligence when they're our in game companions?

[@alecwhatshisname5170]

+Dogmeat

[@scroomshake]

the ai is restricted at that point, not win, but help

[@Trekki200]

+every companion I Skyrim

[@rathcohen]

play Uncharted. tell me how acurate Elaina is.

[@KaltatheNobleMind]

as John was exlaining these AIs require huge resources which is so expensive only institutions can afford them. i dont think the Xbox or playstation or even the highest end gaming rig can put out the calculations like 2-3 stories worth of servers could, doubly so when the machine has to ut resources to other aspects of the game like graphics sound collision detection and physics. though i could see a future where computer manufacturers sell specialty AI cards that have the components and firmware needed to have a basic neural net, the same how graphics and sound cards have the components and firmware to tackle their respective aspects :D and aparently that future is now: www.engadget.com/2016/04/28/movidius-fathom-neural-compute-stick/

[@explosivedude8295]

They can not defeat us in the Game of getting defeated by an AI.

[@explosivedude8295]

Haku infinite Infinite game , Glitch in matrix , Illuminati confirmation , Year Zero , Big Crunch etc.

[@BlueFan99]

end of Universe

[@busTedOaS]

You sure? I can easily write an AI that loses against every other AI.

[@The1SlayerChannel]

What happens when you play it against itself?

[@loayzc10]

Oh shit... now THAT would be an interesting game

[@Ryukachoo]

but can they win the game OF LOVE?

[@GT101Nofear]

Most HUMANS can't even do that one.

[@maniacalo2901]

Very likely, when they learn how

[@raging_dogs7165]

i sure cant win that

[@Rodoadrenalina]

we can totally go extinct by falling in love with evil beauty robots, damn.

[@JRyan-lu5im]

when AIs outperform humans in being good company, they will.

[@s3cr3tpassword]

Heuristics: it isn't always right, but it is almost always right. so......... 60% of the time it works all the time?

[@GotekSC]

It isn't always right, but its always almost right.

[@the-thane]

Closer to the 70-90% range, actually. That is generally considered "Good enough" for these kinds of applications in machine learning. Depending on whether or not speed is an issue, that accuracy changes. Fast heuristics are developed for fast paced games and thus can be seen around 70-80%. Slower paced games might have upwards of 90%.

[@zwerko]

If you were to gamble against an opponent who got it right 60% of the time, given enough time you'd end up broke. So even 60% correctness is much better than a blind chance. And heuristics engines, in dependence of application, can be programmed to aim at 90%+ correctness range with acceptable performance penalties...

[@katzen3314]

60% of the time it works, every time.

[@davecullins1606]

The alien in Alien: Isolation is actually a handicapped AI that for real is trying to find and kill you. It gets clues and has rules so that it often has an idea of which area you're at but it doesn't have enough clues to figure out where you are and so it searches around.

[@Carrera075]

any game ever invented, good luck getting a computer to beat me in MARIO CART

[@hyrekandragon2665]

play mario kart 8 on 200cc and see how the ai beats you at that and come back. Tip you actually have to slow down around turns in 200cc like in real life, or else you will crash.

[@pw7225]

already existent.

[@kennyc002]

TAS tech in Mario Kart 64 can execute wario stadium with 6 perfectly timed short cut jumps from the beginning. So it's possible to program an AI that does that, but do you really want that?

[@IceMetalPunk]

SethBling, of KZbin and Minecraft fame, once designed a neural network to learn how to play MarioKart, and then live-streamed its training process. It's called SethBling's MarI/O Kart.

[@kennyc002]

***** so an AI cannot learn to perform said instructions with the same precision?

[@KaltatheNobleMind]

i think the next "game" Ai should tackle is Rap battle :D i hear a huge hurdle AI scienctists have currently is the issue of semantics and natural language especially with regards to context. what better way to test those things than a contest about who can make the most clever and eleoquent rhyme in real time?:D i bet IBM Watson would be the perfect canidate due to how powerful its natural language processing is :D and the staff at IBM can finally let it use the urban Dictionary for once :P also i just wanna hear a robot finally gloat about its superiority over puny humans in its own words:D

[@TheToric]

That... That's actually a cool idea...

[@KaltatheNobleMind]

TheToric thinking about it further the time limits could help test for reaction/search time and i figure audience reaction can hep test for emotional tone assiting context. and if it's a diss battle we can also test if can understand or "feel" emotional context,like have it recognize its opponent insulted them and react accordingly by making a ryme that is equally insulting to the opponent or make a ryme that boosts its conifdence repairing emotional "damage" also the ryming scheme could test how well the system as a whole can work when given specific rules or parameters such as every piece of output has to follow some ryming scheme:) bet there are other factors i may have missed XD

[@zemerick]

Problem is the judging would be incredibly subjective. It's not like a game where you can tally up points and say exactly and definitively who won. I do wish though they had mentioned a little about actual video games. There are plenty of those that make GO look incredibly simple, where the AI could be pitted against humans. They may not have the long history and super grand champions, but humans are also so much better suited to them, that it would take a huge amount of work to reliably beat a professional gamer.

[@KaltatheNobleMind]

zemerick13 hah got me there on measuring success. i figured they would gauge audience reaction like any other rap battle. if anythign we can write it off as a turing test where if the disses re sick and the burns are big we can classify it as human XD as for Ai doing videogames i think they tackled a few old favorites like space invader and mario. in fact i think they modded a super mario port so the AI worked mario as its own body,essentially making it live the game form its perspective and it aced the entire gameXD and i forget wich AI firm is doing it but they wanna plug an AI into minecraft,not so much to see how it plays but to see how said AI can "survive" and problem solve due to the resource gathering and forging nature of the game. and i think i know why they havent tackled games in seirous competitions:D being already virtual entities they will treat the game as another simulation with every mechanic the avatar posesses be just stimulus if i make sense. they will win every time because they "live" there. guess checkers chess and go were preffered because the mechanics are so abstract and far removed from anything "tactile" so it has to rely on pure strategy.

[@williancruz9657]

but they need to first have a way to know it can do semantics to then test it, right?

[@karvinsunny4199]

so basically humans is trying to defeat itself

[@thejackanapes5866]

How very Human!

[@callumwale6254]

-

[@mitchbankss]

*"without actually playing the game at focus"* that is a big part of your observation that you are missing thats like saying "humans can move themselves at 800 miles an hour" without including the fact that theyre in an airplane that the humans created. dont undermine this break through.

[@irisho5027]

Yes lets ban all cars, calculators, computers, and boats. How dare we invent things that outperform us!

[@WatsitTooyah]

I wouldn't say defeat, maybe augment?

[@AEther0238]

It's kind of cool to think that a 'natural intelligence' could build an artificial intelligence smarter than itself.

[@johnie102]

Well, this already happens when people get children right?

[@AEther0238]

johnie102 Sometimes, yes. But children aren't created with brains. xD

[@johnholmes8315]

Ethan what the fuck are you talking about? Humans literally create other human beings, feed them, teach them, shelter them to become as smart or better than they are. We are the creators of intelligent life, have been for millions of years.

[@AEther0238]

John Holmes Holy fuck, what is wrong with you? Do you honestly not know what I'm talking about? I mean we don't program children with our brains. Not directly anyway. By most birth order psychology, parents often train their first born to be smartest by usually implementing the strictest punishments. But that doesn't mean that the parents are smarter than the first born child. Why am I even bothering? I can already tell this conversation is going to go nowhere fast with an idiot like you involve a lot of swears and insults.

[@johnholmes8315]

Ethan McDonald _we don't program children with our brains_ So what, why are you even picking on body parts here? The human organism is a complex system that involes all organs working together to create and sustain life. Your brain cannot function without rest of the body. It's silly to think you are just a brain, knowing full well that your emotions are being influenced by your other organs.

[@plan8067]

Wait, Deep Thought? The Computer from "The Hitchhiker's Guide to the Galaxy"? I love science

[@metalkez]

now ask an AI to invent a table game to compete with a human... hehe

[@maniacalo2901]

It will probably involve something to do with hyper cubes, tesseracts.

[@elliottmcollins]

AI: "I call it, Calculate The Square Root of this 64-bit integer or die."

[@LetsPlayStarBound]

+Ristar85 Well, no shit.

[@elliottmcollins]

Ristar85 #funnierwhenexplained

[@elliottmcollins]

Ristar85 Matthew H You're both being petulant mean-spirited children about something that doesn't matter. Both of you. Either of you could have had this whole conversation in a friendly way, and your decision to be this mean instead belies a basic flaw in how you have conversations. And it makes the whole internet a worse place, so please, handle your shit and learn to be nice.

[@doncheto4825]

Meanwhile Hearthstone AI..

[@Punkledunk]

The number of times I've seen buffs played before minions is insane.

[@roboterson]

The game has to much RNG so I don't think anyone will be able to get a super high win rate even with a great AI.

[@doncheto4825]

Big game hunter on its own minions is to op for an AI.

[@frostypakki836]

Im playing hearthstone right now lol

[@Shaeress]

MtG and Hearthstone do different things. It's like comparing Football to Twister. They were designed for different settings (one physical, one digital), for different purposes, with different capacities, for different people and in different eras. I love Magic and I've been playing it for over a decade now, but if we look at those things I would say that Hearthstone actually has a better design. Of course, there are reasons for this and the main one is that Hearthstone could never have been a direct competitor to MtG, because HS would've failed that competition miserably; Magic has too great a legacy and history. That's why the design is so brilliant! It uses its digital nature to do things Magic could never do, it saturated the market in a way that Magic never could and they developed a great and intuitive. This is also why all attempts at digitising MtG has kind of failed... at least until Hearthstone created a demand for it. With all this said, I do prefer MtG and its failures in design are understandable; Magic was always the first TCG, establishing it as a thing in the first place, so Blizzard got to look at all the mistakes WotC ever made and learn from them. WotC didn't get to learn from anyone else. It's also 23 years old, long before the Internet was ubiquitous and saturated, and it was released with great uncertainty as to its popularity and aimed at super nerds (tabletop roleplayers), not the huge demographic of general gamers that we have today. MtG wasn't designed to be this huge, it wasn't designed with netdecking and online trading in mind and it wasn't designed to survive this long. That it still exists makes it amazing. Then again, there's also a card game developed (by Richard Garfield et al) with all of this taken into account; Android Netrunner.

[@j-money2337]

But can it win at tic tac toe?

[@maniacalo2901]

Yes

[@j-money2337]

What about two ai's of equal ability playing tic tac toe against each other?

[@denull1]

The game will always end with a draw.

[@FramedHamProductions]

The only winning move is not to play.

[@VienerSchnitzel69]

Always a draw if u know what ur doin

[@Erowens98]

All this perfection, yet Ai's still try to run though walls in the middle of enemy territory while shouting "What" at you

[@quito787]

If it took that much computing power to beat Garry Kasparov, it shows how good he really is at chess!

[@CJCWIS]

So could we get better bots for CS:GO?

[@PleasestopcallingmeDoctorImath]

sure, PODbot 15 years ago. valve are fuckin idiots

[@Shlaboza]

the reason bots in csgo are bad is because valve doesn't want people kicking the anchor because the bot will be better so the bots are intentionally bad

[@saidburhan4212]

+Shlaboza dis da truth ma frend

[@TidusleFlemard]

It's really easy ro make invincible bots, the difficulty is to make bots that are challenging but not unkillable.

[@BIkaloss]

You could very easily make a bot that goes to soundmakers and just instantly headshots you on sight, would that be very fun to play against thou?

[@inademv]

It's all well and good to play games but it gets scary when you task the AI with building a better AI.

[@Ermude10]

6:25 By any definition of "complicated", go has to be much *_less_* complicated than chess (as in simpler). I think the word you meant was "complex".

[@FrostJarl]

"minutes to learn, a lifetime to master" Go is a very simple game with a ridiculous amount of nuance whereas chess is a more complex game with significantly less nuance

[@blacktimhoward4322]

"By any definition..." Well you just shot your argument didn't you? If I define 'more complicated' as 'more possible game states you lose. And if you try to reject my definition, I refer you back to your own statement gg ez

[@Kayakasaurus]

We need to make Data, the android. And then give him a hug, and a removable mustache.

[@AkuNaAme]

Hank, I love you for that Arkham Horror comment. When you said "most complicated game" that was what I immediately thought of, and THEN YOU SAID IT. I am dying of laughter right now

[@Thurgor_Supreme]

"My CPU is a neural-net processor; a learning computer"

[@Kei3th1424]

what?

[@templarkiller2926]

+Khairul Effendi I agree, I have no idea what he is talking about

[@VD913]

Terminator. DUDUN DUN DUDUN

[@veryfastpicket]

It's a quote from the 90's movie Terminator 2: Judgment Day

[@Kei3th1424]

Thank you guys

[@Sandul666]

I remember looking up to gary kasparov as a child... he was such an inspiration and got me into chess. (i was a nerdy elementary kid)

[@moehisho]

Can you actually get better at chess over time?

[@dragon12234]

By learning strategies and tricks, and chess does hone your intelligence.

[@doomgu544]

You can get better at anything with practice

[@vemu333]

Yes. like every other game.

[@ABitOfTheUniverse]

Sure it wasn't Bobby Fischer you idolized as a child? Bobby died in 2008 btw. =( "Searching for Bobby Fischer" came out in 1993. A lot of people saw this movie. Now several more movies have come out about Bobby since his death; 2 in 2009, "Bobby Fischer Live" and "Me and Bobby Fisher", 1 in 2011, "Bobby Fischer Against the World", "Pawn Sacrifice" in 2014 and others are in the works. He lived an interesting life in his later years. I haven't seen any movies with Kasparov, but I'm looking into watching "Game Over: Kasparov and The Machine" that came out in 2003.

[@creativbuildr]

Well, at least there's still Calvinball

[@SgtSupaman]

No AI will ever beat humans in Calvinball!

[@hannahnelson4956]

I made an ai... calvinball time

[@theosnelson1639]

Damn Right there is

[@XiueXelu]

Surprise, the AI are actually on the human's team, meaning the AI are winning.

[@AxionZetaOne]

I was going to post this.

[@Mertaal]

Go is the Japanese name of the game. In Chinese it's called Weiqi.

[@roksraka9241]

In Korean it's called Baduk.

[@Mertaal]

Indeed. It just irks me to refer to it (correctly) as a Chinese game and then call it by its Japanese name.

[@roksraka9241]

Mertaal Yeah, it doesn't make any sense...it's probably called go internationally because it's the easiest one to pronounce :)

[@VashdaCrash]

Likely. How do you pronounce weiqi anyways?

[@amazinglyimaperson]

(fyi it's pronounced kinda like way chi)

[@d3s3rtray]

Its worth mentioning that in the game Lee did win, he did so by playing something called a 'wedge' where he spiked the number of possible moves to be a massive number. Basically he confused the machine long enough to win.

[@EebstertheGreat]

Oh, and the 19x19 dimensions of the Go board _are_ for the grid, not the (meaningless) squares in-between. Calling it 19x19 does not undersell it at all.

[@deanwcampbell]

Now that is "Intelligent Design"! Thanks SciShow!

[@afrojacks4462]

What happens when they try to make an AI with the goal of having that AI build a better AI and so on?

[@jeller9519]

Thats the dark side of AI

[@mso1ps4]

Technological singularity

[@PerfectlyFunctioningAI]

Idk if it was a game A.I like the one in the vid it would hit a wall of all it could possibly learn about the game. But if it was an A.I over something much more complex then it can become spooky.

[@_chirp_6108]

already exists i think

[@rosestube1233]

as long as they don't make AI which can duplicate/produce itself, everything is fine

[@MentalidadeBIM]

Another amazing episode. Congratulations guys.

[@nathanielsoule8067]

Go, Chess, and Checkers are all open-information games (meaning no player has hidden information from another player). Poker is some hidden information, but with only 52 cards for all players to share, the hidden information is still fairly easy to figure out. I would be interested in seeing how an AI plays a game like Magic: the Gathering, or another similar TCG where each player has hidden information that it is almost impossible to guess if you haven't played that person before.

[@SupLuiKir]

Can human AI programmers write an AI program that can write AI programs better than a team of best human AI programmers in the world? Yes. It's a matter of when, not if.

[@snowstorm9818]

that's called the singularity, where a program can change it's own code (or write new code, and use it to replace the old one) to make itself better at changing it's own code.

[@azgarth1]

Human Being's reply is right. Depending on how you look it it, neural networks might fit this bill, too, at least so long as you count systems that program themselves. When it comes to neural networks, you don't set up the weights that actually cause the right/wrong answers to be given. You initialize randomly (to prevent repetitive behaviour), then let it modify its own behavior after every (wrong) prediction.

[@maniacalo2901]

Yeah that will happen

[@venkatchait007]

that is the ultimate goal.

[@JorWat25]

That's known as Superintelligence. en.wikipedia.org/wiki/Superintelligence And it's a serious concern.

[@devilsmessanger]

but what will happen if we develop AI that can develop smarter AI ?

[@DarkPrject]

In principle we already have that. You can use an evolution inspired approach, but that is usually slow and might not go anywhere, depending on what you are aiming for, because if you are on a schedule evolution sucks.

[@DarthBiomech]

Singularity.

[@Followmybliss2013]

Hardware a kill switch unless that's too simplistic means of keeping control.

[@thegreatandpowerfultwily394]

Then humanity either dies out, or thrives. If the new intelligence helps us, then we officially enter a golden age of science. Imagine it. AIs that study, research, and solve any and all real-world problems! It would be able to improve on technology exponentially faster than any human! But that would be if it decided to help us. If it didn't...

[@DustinRodriguez1_0]

Then whoever determines the definition of "smarter" gets the best tool possible for that job.

[@GBart]

He's right, Arkham Horror is the next step above Go

[@GBart]

fuck

[@GBart]

that

[@GBart]

game

[@GarethXL]

lose 10 sanity points please

[@LostMekka]

I've got three corrections: * A Go board has 19x19 positions at which one can play a stone. One position is a cross with 3 possible states, not a square with more states. * A neural network does not have a database. The data that is used for the computation inside the network is stored only in the weighted connections between the layers. (The connections in AlphaGo were trained using a large database of professional Go games, but the program does not use or need it when it plays) * AlphaGo uses two feed forward networks that only process information in one direction: from the first down to the last layer. The only possible feedback loop is, when a move is played and the board position changes, which triggers a new wave of evaluation. (Also, there is a search algorithm involved called monte carlo tree search)

[@theletterwynn]

As a chess player, I remember Kasporov commenting that "computers are idiots who never lose". He has a point, being that computers do not possess a mental capacity of understanding positions or patterns in chess and utilize them in a game like humans do intuitively at the highest levels of play. Thus, computers may play the game, but they don't understand it. Therein lies the difference between man and machine.

[@AzureFides]

One small step for AI, another step for humanity toward being obsolete :P Will it be considered an evolution if humanity creates a manifestation of our own intelligent into another "being"? Or will it be an extinction?

[@DeepVoicedDude]

Yes

[@chrisv4496]

The two aren't mutually exclusive, necessarily. The dinosaurs went extinct... yet most birds alive are descendants of dinosaurs.

[@dilo19000]

not most, all birds are dinosaurs. All Non-Avian dinosaurs are extinct.

[@IceMetalPunk]

You have to be careful using the word "evolution" like that. When most people use it, they generally mean "genetic biological evolution", which this wouldn't be. However, evolution actually applies to virtually everything in the universe (or at least everything with the three common properties of reproduction, mutation, and competition)--so in that regard, computers now are already evolving constantly. Now, I believe what you're asking is could an AI ever be considered an evolved form of human. If you define human in the biological sense, then no, it can't, but if you define a human by some psychological properties, then yes, it could.

[@irisho5027]

We are already made many of our abilities obsolete. Cars can "outrun" us, calculators can perform complex arithmetic in a fraction of a second, and cranes can lift hundreds of tons of stuff with ease.

[@yaa4796]

But can it win Rock, Papers, Scissors?

[@sowthistles]

probably no

[@knights8400]

There is already a computer/robot that does this by scanning your hand in mid throw and then playing whatever will beat it - all in a fraction of a second

[@TheRobster2007]

I have a mesh hammock. I just lie under it and look up. Et voila: Skynet.

[@chairio6212]

thank you. you won the internet. wait, no the world.

[@TheDaxxC]

No, for a game that complex it would need like 30 more "neron" layers!

[@johngreen8465]

THE YEAR IS 20XX

[@juanmi2565]

everyone plays Fox

[@rumfordc]

YES

[@Werwa_]

At AI levels

[@aviinthepast]

Nope, its solved by Rock, Paper, Scissors,

[@huwguyver4208]

The development of ChatGPT got me thinking about the possibility of having realistic free-form conversations with NPCs in video games instead of just picking from pre-programmed lines of script. It could be particularly great for detective games, where the dialog lines can sometimes give away the solution. You could actually have realistic interviews with witnesses and suspects. It's amazing how far AI has come.

[@DeliverToChrist]

AI is so fascinating! Also, thanks Hank for using terms repeatedly in different context so I can better understand exactly what they mean. I'm glad this channel was recommended to me.

[@mvvo7366]

pffft make it play dungeons and dragons

[@NoConsequenc3]

No, that's actually EASY AS FUCK The hard part is making it play D&D in a way that is fun for everyone else

[@markedfang]

+Pseodo Lain Make it the dungeon master, as the dungeon masters goal isn't to necessarily defeat the player.

[@NoConsequenc3]

markedfang The player's goal isn't necessarily to do anything, either. Even if given a quest they can refuse, kill the other players, and then commit suicide. "Winning" or "Losing" in D&D is very much up to house rules

[@killerbee2562]

No gurps, program in _every_ source book and allow the computer to dynamicly create a game world and set of rules it wants to use!

[@MrGamer1992]

+markedfang We kinda have that already (RPG games)

[@abattlescar]

Now get AI to play CAH

[@CJCWIS]

LOL YES

[@mlgesuschrist5518]

wat

[@CJCWIS]

+MLGesus Christ Cards Against Humanity, an adult version of Apples to Apples.

[@Miranox2]

Cards against humanity?

[@milesjsandifer]

how would you differentiate between an intelligence and a random selection of cards, seriously

[@carlosboterosuarez7027]

We cant make AI beat professional Starcraft players!

[@rdizzy1]

......... I think they already have beaten some of them, I read it somewhere a while back.

[@Xynth22]

That wouldn't be hard at all actually. Starcraft is all about apm and good management, both of which would be trivial to a computer which can process and manage things far faster than a human can.

[@AldjinnTV]

Some of them ? The only thing I found (I have conflicting findings on the outcome) is that the AI was playing against an ex-starcraft broodwar pro foreigner. That's not impressive. From what I heard, the AIs were really bad at macroing, which is the most important thing in a starcraft game. Also, AIs don't understand mindgames. So yeah, for now all Starcraft base are belong to us. And it ain't changing anytime soon.

[@kennyc002]

That'd be strange. I mean, the AI should have perfect injects/mule management/chrono boost management, can simultaneously command multiple workers to mine optimally, etc. You'd think the AI would have the least amount of trouble with micro and macro, but would have more problems with the more strategic portion of the game. I'm not talking about like the blizzard AIs tho, that might be the reason. We just haven't really ever built a sc AI like Alpha Go.

[@kylekafka6636]

+NickBaynes While your statements are true, AI would have much more trouble with the imperfect information in Starcraft (all the example board games are games with no hidden information), the diversity of builds possible, and unit positioning (the per-pixel accuracy means a massive grid of positional options, all of which are changing in strategic value every second). Not to mention the game is played at a very fast pace, so at the very least a large amount of computing power would be needed to play fast enough. I'm pretty certain it would be done at some point, but I don't think we're there yet.

[@alfioemanuelef]

+SciHow, at approximately 9:08 you say that "[...] layers send information up and down the stack to all other layers [...]". This seems to be incorrect, as the deep neural networks used in AlphaGo, are 'feed-forward' convolutional networks, which don't have feedback connections, i.e. the information only travels from the lower layers to the upper layers. The opposite would be a recurrent neural network (RNN), but these are used for other purposes and are also much slower to train, so much in fact that a recurrent network of that scale would probably be unfeasible even for Google standards. Very good video overall, it's very informative and factually accurate -- that was just a small detail! Great job :)

[@romkoppel5302]

One of my favourite scishow videos of all time.

[@stocktonjoans]

why the hell did they not name Deep Thought's succesor "Earth"?

[@Philw94]

Actually if you dont mind it is the other way around: The IBM computer came out in 1988, THGTTG aired first in 1978.

[@MuadDib1402]

I wish they had gone with Great Hyperbolic Omni-Cognate Neutron Wrangler.

[@adammartinelli8274]

I wish they had gone with Genetic Lifeform and Disc Operating System.

[@niveshproag8660]

I want a depressed robot as companion.

[@AntiCitizenX]

Still waiting on a decent AI for RTS games. :P

[@pifilixxiv3192]

there is one already dude

[@cridr]

a good example , www.theverge.com/2017/8/11/16137388/dota-2-dendi-open-ai-elon-musk .. also some of the current AIs will work better on faster hardware and that can be scaled a lot. AI domain is very complex, so please do not assume that if we can beat human at all games it will be "intelligent", it will only model very well the winning strategy for games.

[@Kev376]

Googles AI beat starcraft players =\ which is literally a war simulation game (were so screwed)

[@jaysun4069]

No one or no thing will ever beat me at checkers!

[@IceehawkPSN]

saying you're really good at checkers is a friendlier way of saying "I suck at life"

[@maniacalo2901]

You're going to lose

[@JerehmiaBoaz]

I refuse to play checkers also.

[@premchand828]

what about go?

[@jaysun4069]

Prem Chand Never played but im sure i'll destroy anyone in that too

[@tredbobek]

9:20 It's like Legion (/geth) from Mass Effect. Building consensus.

[@videogyar2]

6:52 That's also true for chess.

[@ericktv97]

For a while, I thought Hank was saying Lisa Dole rather than Lee Sedol.

[@twistedbunny527]

Artificial intelligence will never be able to beat us at Calvin-ball.

[@JorneDeSmedt]

So... Is there an AI that can beat humans at Calvinball yet?

[@firedrake110]

If we ever make one, we'd be better off just letting it have the planet from then on

[@volkiodin]

NOPE xkcd.com/1002/

[@charliecrome207]

+firedrake110 GOLDEN SUN IS AN AMAZING GAME

[@Turalcar]

Interestingly, Starcraft is listed before Go and computers have yet to beat that.

[@JorneDeSmedt]

Turalcar Well, computers already beat me at Starcraft, depending on the difficulty level. I just enjoy taking my time way too much...

[@Zakdaman97]

This was an absolutely awesome and intriguing episode of scishow!

Brute force as described at the beginning, with no heuristics, does not "chose the play with the best probability", it chooses the play that will lead to a win with 100% determination. Also, I don't think the layers in the NN communicate backwards. When it comes to game-playing AIs, there's *plenty* of challenges left. Board games are just the beginning.

[@ChristophelusPulps]

How are AI's faring in various complex real-time strategy games? Can they defeat top Starcraft players, for instance?

[@WatzUpzPeepz]

Not well I would imagine. Board games have a set of available moves, large - but not infinite, especially ones available at one time that are both valid and beneficial. While RTS games consist of thousands of different factors on the effectiveness of a single unit alone. That's why we construct strategies to beat others, an AI will need to do the same without any strategies being pre-programmed - soon I would say but not right now AFAIK.

[@ChristophelusPulps]

Just too many variables when you're not playing on a grid.

[@00fabian7]

Not at all. Look at for example Civilization, or Europa Universalis. AI's can only possibly make it difficult for human players to win by 'cheating' (all information readily available that human players don't, relative technological advantages, etc.). Let alone use meta-gaming, dynamically learning how his opponent plays, knowing the contextual (keyword, context is so fucking hard to program) semantics certain plays have and responding to these by using mind games, etc.

[@Animeanem]

Well, its actually even easier, considering that RTS only have an handful of competitive move-set and that 90% of the game is about APS, witch cannot be beaten by human speed. The only reason that you're winning against AI in RTS is that they got a delay before each actions they are allowed to make. The 1-2 seconds they gain at the start of the game keep accumulating and exponentially grow in a 10-15%+ additional resources/troups throughout the match.

[@00fabian7]

Mathieu Audet That holds for regular players, but not top players. If the AI only receives the information a human player gets then it'll become easier for the player to trick the AI (especially if you just train it on certain strategies). It may work the first time (though I'm sceptic), but give it a few games, let the players get a feel for the AI and then that advantage of the AI will be mostly gone. The thing is, for human players it's mostly about APS, whilst for an AI it's more about contextual information gathering and countering the human psyche. The first part comes natural to an AI, the 2nd part to a human. The first part is static and hardly improvable, the second part is dynamic and a pain in the ass to code (if doable).

[@ConradJD777]

It can beat us at Board Games... But will it blend?

[@partingmist8550]

well its not chuck norris so my guess is gonna be yes.

[@Kaish3k]

I normally don't do this, but I feel like I need point out all the errors - let's face it, not many people know AI well enough (it's not easy, right, or everyone would be doing it), and those of us who do know how it works, need to correct the mistakes of others. (I work with bleeding edge deep learning on a daily basis). 0:15 False, AI cannot beat humans at "pretty much any game." (yet) Some of the best DQNs can sort-of-kind-of shoot other players in Doom. That's about as advanced as it gets from a pure modern AI perspective (other than AlphaGo). 8:17 This description of deep neural networks is hopelessly wrong: ----------Neurons are not "computer systems." They exist only as an abstract concept. What really exists are the weights and biases of the connection of a layer and the layer before it. These weights and biases are represented as tensors (similar to matrices from linear algebra, they're just a higher dimensional abstraction of a matrix from multilinear algebra, often 3+ dimensions). Neural networks basically are just large tensor equations (en.wikipedia.org/wiki/Tensor). The input to a neural network is also a collection of numbers, so another tensor. The neural network is essentially the operations between the weights, biases, and input tensors, from layer to layer (the output tensor of layer n-1 is the input of layer n). So obviously "neurons" don't run in parallel (since they don't even exist!). If you calculate the tensor equation on a GPU, then the tensor equation can be calculated in parallel, using the exact same processors that your browser is using to decode this video in real time. GPU = Big, fast calculator. Neural network = Big massive equation. GPU + NN = A fast computation of the NN. There's really nothing special about it. You can do the same thing with a normal processor, it's just MUCH slower for huge networks... ----------Layers in a DNN do not work as described. At all. Period. The process of a neural network "learning" is actually a minimization process. This basically involves enough math to make any undergrad puke (well, I suppose it depends if you like math.. :p). Like I said before, neural networks are large tensor equations, often with 10 million+ variables. Yes. Like the variables you learned in school. Yes. 10 million. You see the problem is you cannot solve an equation this large. Neural networks are iteratively approximated in the best case scenario (which never actually happens). Like I said, it's a lot of math, but if you're interested, you basically use the average partial derivative of the cost function of the output with respect to the weights and biases (which is a high dimensional vector) (this is called a gradient) to effectively crawl down the slope of the cost function and find a minimum in the cost function in some crazy high dimensional space (millions of dimensions). ANYWAYS, this produces an optimized neural network if everything works correctly (which is often not the case). SO WHAT I'M TRYING TO SAY IS YOU DON'T KNOW WHAT THE **CK IT IS LEARNING BECAUSE IT'S JUST WAY TOO COMPLICATED FOR OUR LITTLE HUMAN MINDS TO UNDERSTAND! Yes, it is thought that neural networks learn abstractions in their layers. But this is very theoretical, and should not be considered outside of a pure informational sense of an abstraction, which goes beyond the scope of most people's understanding of math, and even leaks into physics in some regard with respect to what information is fundamentally. Best case, if none of that makes sense. Just think of a neural network as a huge statistical equation that says, given this input X, I ^probably^ think this is the output Y - how it does that is incredibly complicated, as stated above. ----------Neural networks simply DO NOT have databases. I don't know how to elaborate on this. They just don't. It's just not how it works... ----------Layers do not communicate back and forth typically (in normal feedforward deep neural networks), unless you're using recurrent neural networks, which are like a whole 'nother level of mind**ck. Even then, a layer in a RNN typically only "communicates" with its future state (I said it got weird..). ----------AlphaGo used two deep neural networks as part of independent Q-learning systems, being the Deep Quality Network that estimated the Bellman equation used for fitness function in typical Q-learning. DQNs essentially estimate the quality of output actions given an input state (so for example the best quality action is the best action to take given a particular state). All of the above apply to DQNs. One of the DQNs learned to determine who was winning at a particular state, and the other determined the most interesting next moves. Using this information, AlphaGo was then able to dramatically reduce to space complexity and branching factor of the monte carlo search tree (which is used in chess too). Using the information gained between the 2 DQNs and the MCST, AlphaGo selected the single most probable action to result in a future desirable state. ----------Go is NOT the most complicated game. There are MANY MANY MANY MANY more games that basically are provably mathematically infinitely more complex. Those games are games where information is hidden. Go is a game where all the information is available. In games like Doom or really any other 3D games (they're actually 4D because they heavily depend on time too), there are vast amounts of hidden information that is inferred and analyzed at extremely high levels of abstraction. Neural networks cannot currently begin to solve these massively complex games. Within 10-15 years they will though. Sorry to be nit picky, but your description basically entirely missed the mark. It wasn't even metaphorically correct.

[@requireloginisstupid]

The first reported AI controlled game was a Chess program written between 1948 and 1952 by Alan Turing, it lost one game and won another, both in 1952-3.

[@TimVels]

Awesome video! Thanks!

[@kelpia]

"I never asked for this"

[@9Tensai9]

I wonder how those guys felt aftey they lost. I mean, the uktimate champions... defeated.

[@ShaunHusain]

Kasparov was furious, he has since come around some but he was not happy to have been beat by a machine. I believe the go players took it a bit better. When Watson won at Jeopardy the contestants it beat were shocked how it crushed them and quite angry but have also come around. Being beaten when you are the greatest your species has to offer is a bit of a downer but to be fair he's playing a team of engineers with the most advanced technology from all of humanity.

[@aasherahmed4389]

there is always someone or something vetter than you

[@InstxntNoodles]

+Aasher Ahmed unless u are the champion of the world lol

[@aasherahmed4389]

Hamza Kittaneh you never know

[@tyvernoverlord5363]

And then a machine goes and rekts your title of world champion...

[@admiralpercy]

Wait, they did checkers before tic-tac-toe?

[@elliottmcollins]

No, but tic-tac-toe has a brute force solution that can be written on an index card and the only measure of skill is not losing, so it doesn't really bare mentioning.

[@admiralpercy]

Elliott Collins Worth mentioning? Are you saying there WAS one before checkers?

[@elliottmcollins]

Admiral Percy There were lots of games solved before checkers. There were lots of games solved before computers were invented. But none were solved by artificial intelligence until checkers (that I know of).

[@Activsoul]

Yes, there was a computer named "Joshua" that loved playing Tic-Tac-Toe. At least, that's how I remember how the story goes.

[@kohZeei]

"The only way to win at tic-tac-toe is not to play it at all" From the movie War Games.

[@АсенДоцински]

"Go was it, the most complex game ever... although I would like to see it play Arkham Horror"

[@trip1565]

Great episode!

[@Archontasil]

i bet they can't beat dark souls

[@manbob15]

They ARE dark souls...

[@Vyz3r]

Dark Souls? Dark Souls games are easy as fuck. Try Monster Hunter on G rank.

[@lancelindlelee7256]

Dark souls would be easier. If you have data on all animations, you can programs a machine to watch for all animations for the specific enemy then dodge immediately when it sees these animations and attack when this is not present, taking stamina into account. This is true for all fair video games (where it is always possible to react, no need to anticipate)

[@beegieb4210]

I bet they can... eventually. Not with our current algorithms. Dark souls is actually 'easy'... in the sense that there is no 'randomness' in the game's combat that the learning algorithm needs to deal with. Every enemy move is telegraphed, so learning the optimal dodge for every possible attack it sees on the screen is relatively easy. A more difficult challenge is actually navigating the world, and solving the environmental puzzles since they often require long-term dependencies.

[@tdo0msday]

realistically speaking dark souls is only a challenge simply because they can't (or can't be bothered to put in the money and effort to) find enough strong gaming computers to run which is quite a resource demanding game for the simulation tests over and over. do you know how many hundreds of thousands of matches that AlphaGo ran and played with itself? that's the level of complexity the game of Go is at, but for dark souls all it requires is some nice reaction speed (which computers are quite good at given brute forcing is their specialty) and some simple strategies that isn't so hard to figure out. I bet it wouldn't even take more than a few tens or at most several hundreds of playthroughs until the AI is capable of producing a perfect run, depending on how many levels you want it to play or if you want it to get through the entire game, which doesn't prove to be much more complex the longer it goes anyway.

[@clover7359]

For the past 15 minutes, this video has been getting 3 views/second.

[@PleasestopcallingmeDoctorImath]

on average maybe

[@ElectricFan91]

Until AI gets a 100% win rate with dumb people in its team in CSGO, I think humans still come out on top.

[@christopherbazaka1564]

aimbot

[@tropicaltundra6409]

aimbot is not an AI

[@christopherbazaka1564]

cold hunter laser vision

[@christopherbazaka1564]

cold hunter super strength

[@christopherbazaka1564]

cold hunter stinky breath

[@OpZeroFilms]

This video made me miss the old Japanese High school Go club I managed to join years ago. . . A couple of great online friends, all very strong players. . .

[@Masterpouya]

Scishow team is the best ! I'm taking my shot of Scishow before starting to work ! Neurons [OFF] --> ON ! Curiosity [OFF] --> ON ! EPICNESS [OFF] --> ON ! Add some Epic Battle Music and you obtain : Neurons [ON] --> TURBO ! Ready to beat the IA now ! :D

[@Tachyon836]

Can a computer beat me in Melee though?

[@DragonCharlz]

Did you ever play any of the Street Fighter games with Shin Akuma, where he would do super moves if you pressed a button in his face? I'm pretty sure they could make something that just blocks and counters everything you do, but that wouldn't be terribly fun to play again.

[@Tachyon836]

DragonCharlz Then the computer isn't playing by the same rules as I am is it? If you're going to claim that it's superior to me, it has to be playing by the same system I am.

[@DragonCharlz]

Beni Well, you can't pick Shin Akuma in most games, so I guess that's a violation of the rule.

[@Tachyon836]

DragonCharlz That's the idea.

[@DragonCharlz]

Beni Take his AI and put it on regular Akuma. It would still be hard, but that's because, as stated earlier, he reacts to your inputs with super moves or dragon punches if he doesn't have the meter.

[@martinkunev9911]

I am yet to see AI that can beat a good human player in RTS games like StarCraft.

[@patrickwienhoft7987]

RTS games are actually easier in many aspects, as the reaction time of a bot is just much better.

[@martinkunev9911]

Patrick Wienhöft What do you mean by reaction time of a bot being much better? In such games the world is usually very big and diverse. The computer has to make descisions fast while taking into account lack of information and the complexity of the world. In addition, it is often very hard to determine which is the better of two mutually exclusive actions.

[@novareaetem]

+

[@beegieb4210]

+Martin Kunev It's actually not particularly difficult for an AI to learn to assign a value to two mutually exclusive actions. There's already an algorithm that does this... it's called an Actor-Critic Reinforcement Learning Algorithm, a variant of a Policy Gradient Reinforcement Learning Algorithm. All it needs is time to play a few hundreds of thousands of matches, and get the final outcome. It will eventually learn how to assign value to moves given it's current knowledge of the world. Where humans have the advantage is simply this: we don't need to play a game a million times to be able to play it.

[@luisalberto-st8bn]

if the devs actually wanted to they would make an unbeatable AI

[@95frekhaug]

Magnus Carlsen is the current world champion. He is also the highest ranked chess player in history, beating Kasparov. Carlsen's peak/current rating: 2882/2857. Kasparov peak/current rating: 2851/2812.

[@bikernumber7180]

the last game played and first game won by the machine Carlson was 2 years old. at the time of the game kasparov was king.

[@kronusexodues7283]

Saying that Go is the end of the line is just wrong. In chess and go you have time to calculate. Programm it to play stuff like Leauge of Legends where you have to make decisions a lot faster, have to anticipate the behaviour of 9 instead of just one other guy and still have those strategic components.

[@WatsitTooyah]

To expand on his idea on heuristics a little, a program can simply look a predefined number of moves ahead of time, which can be adjusted, which directly affect the run time of the algorithm as well as the "best move" accuracy of the AI.

[@Tutorp]

There's one infamous board game that I'm pretty sure AI isn't quite ready for: Diplomacy. Admittedly, that is not due to the on-board mechanics in itself - those are comparatively simple - but due to the negotiations that form an integral part of the game. In Diplomacy, players must cooperate with and against each other (through secret messages and negotiations, which are not binding), before announcing their moves simultaneously. The diplomacy part of the game requires a fair bit of social engineering, and I don't think that's an AI forte quite yet.

[@johnie102]

I recall reading an article some time ago about researchers succesfully creating an "AI" that could predict when people would break their alliance in a game of Diplomacy.

[@jorgeriveramx]

I see a lot of comments saying things like "try to make an AI to beat me in Dota, Mario Kart, (insert your favorite game here), etc..." people doesn't realise that is much easier to make an AI for those kind of games than Go. Actually AI have to be dumbed down for those games in order to keep them fun for humans. Believe me, if AI can beat us (humans) in Go, there is no other game they can't beat us.

[@TheNewVoxel]

You've never touched a line of code in your life.

[@jorgeriveramx]

Vox​ you have no idea what you are talking about. I've been a software developer for 7yrs. While AI is not what I do at work, it is something that I liked since I was in college and I have experimented with expert systems, artificial neutral networks and evolutive algorithms. You are the one that doesn't have any idea of artificial intelligence. Are you even a programmer? Have you ever wrote a line of code?

[@beegieb4210]

I build AIs for a living, and used to work in a games company. He's actually correct.

[@TheNewVoxel]

I actually wrote my thesis on programming AI for the game GO. Albeit it wasn't anywhere near as successful as what the good folks at Google where able to deliver since I only provided ~10,000 sample games to learn from so it remained largely unheard of. I believe it would be much easier to deliver an AI for this game since there are only three variables that have to be mapped into the programs memory white, black, and empty. Then on top of that they only have to be mapped to some set locations on the board with only 9x9 or 13x13 tiles or whatever sized board you're playing with. While this still leaves us with an insane amount of data to work with like the video said it is at least much easier for the computer to collect and learn from this data from previously played games then it would be to collect and provide sample data for a game like Starcraft or something with many more variables.

[@beegieb4210]

First off, alpha go didn't have the board state given to it. It literally had to learn the board state on its own from a video feed of the board (A CNN). The CNN builds up a representation of the board state in a vector space. That is then what the agent uses as input to decision making. This approach allows you to collapse the state space to something a little more tractable. Then two different neural networks worked together. A 'policy network' had the sole purpose of determining a probability distribution over possible moves to make given the current state representation. While a 'value network' tried to estimate the likelihood a particular state representation would lead to a win. These two networks, and the condensed state space made using a search algorithm tractable. At any given stage of the game, the number of moves you could make was highly reduced by the policy network, this reduced the breadth of a tree-based search method (Monte Carlo Tree Search), the value network decreased the depth of the search tree since you could simply terminate when the value network assigned a high probability to winning, rather than actually rolling out the entire game to observe reality. Alpha go, in a sense still used brute force. They distributed the search algorithm over a server farm, but due to the policy and value networks, they didn't have to explore the entire search space at every turn, only the most probable. This is still not the ideal way to build an AI. That's where algorithms like Policy Gradients and Actor Critic methods come in. They try to learn through single game rollouts (how did this game end after a sequence of moves, and how can I adjust my internal parameters in a way to improve my rewards next time) . This method allows you to build AIs that can determine the next move to take in time proportional to your model's parameters, not proportional to the space of possible configurations.

[@RelatedGiraffe]

God, so many misconceptions being spread in this video. Is it the same in all SciShow videos?

[@jrkirby93]

Yeah, hank clearly doesn't know anything about machine learning.

[@upandatom]

What misconceptions? I'm curious.

[@SnzyBat]

ENLIGHTEN US THEN, AI GURUS. SHOW US THE TRUE PATH AND LEAD US TO SALVATION FROM THIS WRONGED AND FLAWED EXISTENCE. LET US REGOICE IN THE LIGHT OF ACCURACY AND SAVE US FROM HUMAN MISTAKE.

[@jrkirby93]

"Layers send information up and down the stack to all of the other layers." This is completely wrong. In a It really only goes one way, from one layer to the next. The only exception is during training time, the NN is analyzed using backpropagation, you guessed it, backwards. "When all the layers agree on a move..." This shows a complete misunderstanding. None of the layers make a decision on a move to be made. Only the final layer outputs any information used to make that decision, and then it outputs a board with unplayed positions filled in ranked by how preferable they are. Other techniques figure out what move to make. Basically, every single thing he said about neural networks is either completely incorrect, or showing a complete lack of understanding about how neural networks actually function.

[@Regic]

Precisely. Neural network layers doesn't work this way at all. They are levels of abstractions, f.e. first layer finds edges or corners in a picture, second layer finds shapes like squares and rectangles based on the first layer's edge recognition (they don't really recognize "lines" or "squares", but shapes which are sometimes similar to ones that we have names for.), then deeper layers try to find shapes of shapes, (arrangements of different shapes), etc... Information only flows one direction: the 1st layer works based on inputs, 2nd layers works based on the result of the 1st layer, 3rd layer works based on the result 2nd layer, etc... I guess it might learned to find "uncontrolled areas" and similar features (sorry I don't know how to play Go and the publication about AlphaGo doesn't mention these), but these can happen in one layer or more, layers are definitely not assigned to recognize features like those and don't make any decisions on their own. I'm sure you can find a lot of people knowing much more about machine learning than I do, but believe me, anyone who knows neural networks can see the explanation in the video is plain wrong.

[@mr_ekshun]

There is now an AI that has defeated pros at StarCraft 2! It's called AlphaStar. The matches shouldn't be hard to find and are totally worth watching.

[@maggiedunsford8709]

I have a Cognitive Psych test today and AI and heuristics vs algorithms are gonna be on it. Thanks, Hank! Perfect timing with the video. :)

[@merrymachiavelli2041]

While I'm not denying the size of the accomplishment here, in some ways it still feels like this pales in comparison to what the human mind does. Humans can be presented with any game, learn the rules and come up with their own strategies . These machines are much more rigid than that, they are designed to play specific games, with the framework around how decisions should be made pre-determined by human programmers.. You can't just plop the Go AI in front of a chessboard and have it play chess or in front of a deck of cards and have it play go fish. Of course, it's still problem-solving, but it's inflexable in a way that severely limits real-world appications. It will be interesting to see an AI machine that can one day beat a Go grand master without being designed specifically to play Go.

[@moanguspickard249]

EXACTLY. All those AIs think inside the box they were given, there is no thinking outside the box.When they make AI that can learn new games (not pre programed) and then defeat world champions, ill be impressed.

[@avakining]

Mari/O made by Sethbling on KZbin can learn to play most platformers from nothing

[@animowany111]

+Pirateboy04 Not really, with MarI/O the programmer needs to create a game-specific fitness scoring function. For the level the video was on, it was {distance right - time since started playing} (with some coefficients on each term). For a more 'general' AI, look into "Learnfun" on KZbin

[@Sinnessa]

Well, some humans anyway.

[@TheMikernet]

+animowany111 Um, well, yeah...you need to be taught a fitness scoring function to solve a game as well. You think if you throw super mario in front of a caveman he will know what to do? Yeah, didn't think so.

[@IRonIcScopez]

Wait what lmao... "There are more combinations for a GO board than there are Atoms in the universe" That was a joke right....

[@TheJourney667]

No. Not at all.

[@kimberlycuizon2766]

Its true

[@nicholastrombone9899]

No he was speaking the truth

[@Diogofrn]

These idiots claiming it's not a joke still fail to give an explanation on how that's possible. I can also claim that I have 2m schlong without taking my pants off. Saying something is true just because you think so without any thought or sensible explanation is a behaviour that has no place in science... only in religion.

[@steliostoulis1875]

Actually it's Much bigger than the atoms in the universe

[@BrainOnVacation]

"There are more possible board configurations in a game of GO than there are atoms in the universe." Color me fucking intrigued! I feel like this ALONE needs a separate video explaining how the hell does it even make sense.

[@i_boole6639]

There is actually one game that AI have not beat yet. It's called Calvin Ball.

[@Marlas009]

I don't think you can say that's it since GO like Chess or Checkers is a complete information game. A game with incomplete information can be much more challenging. I love that the google deepmind team (the one that did beat Go) are now going to see if the can beat Starcraft2.

[@ZennExile]

The real value of AI is how it will be used to organize massive amounts of data, and to enhance the accuracy of simulation algorithms. You should really do some videos about Quantum Computing and the inevitable Quantum Reality Simulation that will shape the future of all humanity.

[@michaelbuckers]

Neural networks are huge mathematical matrices. You feed in some data on the input, it goes through the matrix and gets multiplied by all sorts of numbers and goes through re-mapping functions, and final result goes to the output. Neural network is tweaked repeatedly until it can produce "right" outputs given specific inputs. Sounds like shitty way to make a program respond to things, especially the ones it was never trained to work with, but it does ultimately gets really good at doing exactly that. A huge mathematical matrix is a very good substitute for a brain, as far as data processing goes.

[@musclebrainsmartypants6275]

Very good episode!

[@LKfan1234]

Go hasn't exactly existed unchanged since it's inception. There have been various updates to tournament rules to handicap the player going first for game balance reasons.

[@EtoThe1toTheV]

Gary Kasparov is no longer the best Chess player in the world, it's the prodigy Magnus Carlsen.

[@gmanbuilder9843]

The names for these computers are badass. I would name my star ship AlphaGo

[@cypressz]

The next game it should try tackling is Magic: The Gathering. That would be massively complex and I'd like to see what this form of problem solving could come up with. Beyond games we could try to get it to solve problems we haven't yet; making Algaefuels economically viable, for instance. Give it materials and then program it to figure out the most efficient way to get fuel out of them.

[@veggiet2009]

So I watched this this morning and it's lunch time and as I take bite it hits me "He forgot about Jeopardy! He forgot about Watson!"

[@DrathVader]

And yet every single one of my companions in Skyrim is dumb as a brick.

[@1234567890rom]

Chess was not solved by brute force in the strictest sense. It could not crunch "the numbers for every possible move" because our estimation for the possible games of chess is shannon's number, which is also greater than the number of atoms in the observable universe.