Hi everyone, it's Josh and for this week's select, I've chosen our twenty eighteen episode Some Games You would Surely Lose to a Computer. It's a philosophical discussion about AI that's disguised as an episode on computer games. Honestly, we didn't plan it to be like that. It just turned out that way. We're pretty happy that it did. And in light of the recent advances with machine learning like chat GPT, a few of the things we say seemed
naively quaint now. Plus it has a doll up of our tech stuff colleague Jonathan stricklanet and so that's a bonus. I hope you enjoy.
Welcome to Stuff you should know, a production of iHeartRadio.
Hey, and welcome to the podcast. I'm Josh Clark. There's Charles w Chuck Bryant, there's Jerry over there. I'm just gonna come out and tell everybody making fun of me for some weird reasons, vaguely weird ways. But I'm all right, So check out his story for you. Okay, I'm going to take us back to the seventeen seventies and they'll swing in town of Vienna, not Virginia, not Vanna Georgia, which you know, that's how they pronounce it, right, Yanna, Vienna sausages, right, Vienna Austria.
Have you ever been there?
Vienna, Austria. No, been to Brussels. That was pretty close.
Vienna's lovely, I'm sure.
I think it's a lot like Brussels.
Very clean, lovely town. I just remember it being very clean.
Yeah, very clean, gorgeous architecture, weird little angled side streets. They're very narrow, very pretty town. So we're in Vienna and there is a dude skulking about going to the royal Palace in Vienna. His name is Wolfgang von Kempelin, and he's he's an inventor, he's an engineer. He's a pretty sharp dude. And he's got with him what would come to be known as the Turk, but he called it the mechanical Turk or the automaton chess player, and
that's what it was. It was a wooden figure that moved mechanically, seated at a cabinet, and on top of the cabinet was a chessboard. And when he brought it out to show to the royal court, he would it was cool kind of but nothing they hadn't seen before, because automata was kind of a hip thing by then.
Yeah, people loved building these engineering these automata machines to do various things. And people are just knocked out by the fact that, you know, you hide these gears and levers behind wood or a cloth, and it looks as though there's a real well not real, but you know.
What I mean that it's like a real machine.
Yeah, but not. They weren't fooled anything like is that a real man? It was, but it was for their time. It was so advanced looking that it's like us seeing x Makina in the movie theater.
Sure does that make sense? Yeah, no, it does make sense. But imagine seeing like X makinan being like I've seen this before. This isn't anything special, okay.
Yeah, And this thing, to be clear, looked like a is it Zoltar or Zultan from Big Zoltar Zultan?
I don't know. It's one of those two.
One of those two. Like this, this guy's wearing a turban and it's in a glass case like a bust like, you know, like a chest up thing.
Yeah, he's seated at this cabinet, so there's no need for legs or anything like that.
Yeah.
But the thing, this is what was amazing about the Turk. He could play chess, and he could play chess really well. So yeah, he was like an automaton and he moved all herky jerky or whatever. But he could play you in chess, which was a huge, huge advance at the time. Like, this is something that wouldn't come up again until the nineteen nineties, more than two hundred years later. This thing, this automaton could play a human being in chess and beat them.
Well. Yeah, and it looked like when the game started, it would look down at the chessboard and like cock his head, like what should my first move be?
Right?
And if people I love this part. If people tried to cheat. Apparently Napoleon tried to cheat this thing because this guy he debuted at the VI and East Court. But then it, you know, it went on a world tour.
Yeah, and he was even it was taken over by a successor to the guy who toured with it.
Even further, people went nuts for this stuff.
They did. They loved it because they were like, this is crazy, I can't believe what I'm seeing. Most people, though, were not taken in by it, they're like, there's some trick here. Sure, but von Kempelin and the guy who came after him, I don't remember his name, they would demonstrate you could open this happened and you could see all the workings of the mechanical turk inside.
Right, So what I was saying is if this thing's since a cheater like Napoleon supposedly did it, would you know Napoleon would move a piece out of turner illegally or something. This dude, the turk Turk one eighty two would pick up the chess piece, move it back as if to say like, no, no, Napoleon, let' see what
you're doing. And then if the person attempted to move it again, I don't know how many times on you two or three times, eventually it would just go ah and wipe his hand across the board and knock off all the pieces. Game over, right, Which is pretty great. Yeah, it's a nice little feature.
Yeah it is. But it even showed even more that this thing was thinking for itself. Yeah, that's the key here, right. Sure, chess had been for a very long time viewed as only something that a human would be capable of, because it took a human intellect, and there was actually a guy in English engineer, I think he was a mechanical engineer. His name was Robert Willis. He said that chess was in the province of intellect alone. So the idea that
there was this automaton playing chess blew people away. But again people figured out, like, okay, there's something going on here. We think that von Kempelin is controlling this thing remotely somehow,
maybe using magnets or whatever. Other people hit upon the idea that there was a small person inside the cabinet who would hide when the cab when the workings were shown, when the cabinet was open to show the workings, and then when the cabinet was closed again and the mechanical turk started playing, the person that crawled back out and
was actually controlling it. This seems to be the case that there was a person controlling it, but the idea that it was it was a machine that could think and beat humans in chess had like kind of unsettling implications.
Yeah, this author Philip thick Ness, great name, British author for sure, Philip Thickness. Yeah, he said, and you know, people like you said, all those more complicated explanations. In this article you sent, Astuteley points out that he followed Occam's razor and basically said, he's got a little kid in there. He's got a little a little Bobby Fisher in there that's really good at chess, and that's what's
going on. And other people speculated that other, you know, little people might be in there, just adults who would fit in there. But then you know, there's the explanation that he would open it up and shine a candle around and say, you know, nothing to see here everyone, So what should we reveal the real deal?
Sure, I think I did already.
Well I don't think you spelled it out as.
Oh, we'll spell it out.
There was a little person in there, Yeah, not just one little person, but they would travel around and recruit people. I guess people would get tired of being in there.
Or they'd forget about them and they'd starve and have to replace them.
But it really was a trick. There was a little person in there. They did the same thing as like the magic acts, you know, when they saw a person in half. It's that the lady just gets into a tiny little ball right in one section of that box.
But my thing is this, like, this is not a satisfying explanation to me, Chuck.
I think it's great.
How did the person keep up with the board above?
Well, I mean some I don't know if they ever proved exactly how it was going.
That's what I'm saying.
Oh, okay, whether or not I think they that the zoltar or I'm sorry, the turk was just hollowed out and you you would just put your arms through the arm hole turk, you would become the turk and the churk would fuse.
That's what some people thought. I think that's what Egar Allen Poe thought too.
He Kidney right.
Other people thought that there was the the the little person was underneath the in the cabin operating the trick with levers and stuff like that.
Well, there could have been a mirror or something, you know, I guess, like a telescopic mirror.
That's what's getting me is how would they keep up with the game?
Right?
You could keep track of the game, but how could you see where the other person moved? You would know where you moved, but you wouldn't be able to see where the other person moved. That's what I don't get.
Just mirrors smoking mirrors.
Maybe so, But the point is is it was a fake. It was a fraud, but it raise some really big questions about the idea of a machine beating a person at something like chess.
Yeah, and it really peaked the mind of one Charles Babbage was he was a kid or young at least at the time when he saw the Turk in person, and a few years afterward he began work on something called the Difference Engine, which was a machine that he designed to calculate mathematics automatically. So some point to this is kind of maybe the beginnings of humans trying to create AI.
Well yeah, with Babbage's differential machine or difference machine.
Yeah, difference engine. But at the very least, what this is is the first that I know of example of man versus machine, even though it was really man versus man because it was a man in the machine.
Right. It was a fraud, Yeah, but it sparked that idea, It definitely did. And that's something that like chess in particular, has always been like this idea of like, if you can teach a machine to play chess, you have really achieved a milestone. And there's been you know, plenty of programs, most notably Deep Blue, which we'll talk about. But there's there's been this idea that like part of AI is chess,
teaching it to play chess. But they, the people who develop AI, never set out to make a chess playing AI, just to make a machine that can play chess. That's not the point. Chess has always been this way to demons straight the progress of artificial intelligence.
Yeah, because it's a complex game that you can't just program it, like it almost has to learn.
Well, it depends on how you come at it at first, right, So initially they did try to program it. Okay, there's this from basically nineteen fifty to the about the mid like about say nineteen fifty to twenty ten, sixty years right, That is how they approached AI and chess is you figured out how to break chess down and explain it to a computer. Now, what if you could, ideally you would have this computer or this AI, this artificial intelligence be able to think about the outcome of every possible
or every possible outcome of a move before making it. Right, that's just not possible. Still today we don't have computers that can do that. Right, So what you have to do is figure out how to create shortcuts for the machine, give it best practices, that kind of thing. Yeah, And that was actually laid out in nineteen fifty by a guy named Claude Shannon who's a father of information theory, and he wrote a paper with a pretty on the
nose title called Programming a Computer for Playing Chess. And you have to say it like that when you say the name.
Yeah, it's got a question mark at the end, right, But.
He laid out two big things. One is creating a function of the different moves, and then another one is called a mini max. And if those were the two things that Shannon laid out, and they established about fifty or sixty years of development in teaching an AI to play chess.
Yeah. So this evaluation function is just sort of the base, the very basis of it all kind of where it starts, which is you a kind of give a number to create a numerical evaluation based on the state of the board at that moment, right, and assign a real number evaluation to it. So the highest number that you would shoot for is obviously getting checkmate, getting a king and checkmate.
Right, right, So what you've just done now is by assigning a number to a state like the pieces on a board. What you what you've done is to say, like shoot for this number. Right, the higher the number, like you're going to give this aither rule. Now, the higher the number, the more desirable that this move that could lead to that higher number. Function. Evaluation function is what you want to do.
Right, like capture the night or capture the queen. Capture the queen would have a higher evaluation number.
Right exactly. So that's the function. Then there's another one called the mini max. Yeah, this is pretty great where you want to minimize the maximum And this is another shortcut that they taught computers.
The maximum loss that is right, Yeah.
So what they what they taught computers to do is so you no computer can look through an entire game every possible outcome, but there are computers that can look pretty far down the line at every possible outcome. And what you can say is, Okay, you want to find the evaluation function that is the worst case scenario, the maximum loss, and then find the move that will minimize the possibility for that outcome.
Yeah bye, And this is you're only limited by your programming power. But by looking not only at the state of the board right now, but if I make this move and I move the PAWD to this spot, what are the next like three moves possibly that could happen as a result of this move. And you're only limited, like I said, by programming power. So obviously, the more juice you have, the more moves ahead that you can look exactly.
And then they just shy away from ones with the higher function number exactly or lower function number, depending on how you've programmed it. But they're making these decisions based on these rules. And then there's other things you can do, like little shortcuts to say, if a decision tree leads to the other players King being in checkmate, don't even think about that move any further, don't evaluate any longer, just abandon it because we would never want to make
that move. So there's all these shortcuts you can do. And that's what they did to teach computers. That's what Deep Blue did when it beat Gary Kasparov in nineteen ninety seven. It was this huge, massive computer that knew a lot of chess, a lot about chess. It had a lot of rules, a lot of incredibly intricate programming that was extremely sharp, and it actually won. It became the first computer to beat an actual human chess grand master in like regulation match play.
Yeah. I mean, and I don't think Kasparov gets enough credit for being willing to do this, because it was a big deal for him to lose. It was in this community and the AI community. It sent shock waves, and everyone that was alive remembers, even if you didn't know anything about either, one remembers Deep Blue being all over the news. It was a really big deal. And Kasparov put his name on the line and lost.
Yeah, And I was wondering, Chuck, how how like you would get somebody to do that.
I'm sure the Mountain of catch.
I guess that would probably be part of it. But I mean, shit, I don't know.
I bet, I bet that's out there. We just I just didn't look it up.
So that's possible. It's also possible that they said, look, man, like this is chess we're talking about or whatever. But really, what you're doing is helping advance artificial intelligence.
Right, because we're not really trying ultimately to win chess games. We're trying to cure cancer.
I mean, yeah, we're going to take your title because we're going to beat you, or our machine's going to beat you. But even still, you're going to be helping with cancer. Think of the cancer Casparov. That's probably what they said.
Should we take a break?
Yeah?
Uh wait, well should we tease our special guests first?
Is he okay? I can smell him.
I don't think we even said we're gonna have a special guest later in the episode, Mister Jonathan Strickland of Tech Stuff.
Nice.
It's just been a long time since, like years since we had Stricken.
The last time we had Stricken was like two thousand and nine with the Necronomicon episode.
What is what? Where's he been besides sitting in between us every day?
It's been a strickling drought, is what it's been.
Yeah, so Strickland's coming later, but we're going to come back after this and talk a little bit more about a man versus machine.
Okay, dude, So what we just described was how AI was to play things like chess or to think like you take something, you figure out how to break it down into little rules and things that a computer can think of, right, and then follow these kind of rules to make the best decision. That's how it used to be. The way that it's done now that everybody's doing now is where you are creating a machine that teaches itself.
Yeah, that's the jam.
That was the breakthrough. You may have noticed back in about ty thirteen twenty fourteen, all of a sudden, things like Siri and Alexa got way better at what they are doing. They got way less confused. Really, your navigation app got a lot better. And the reason why is because this type, this new type of AI, this new type of machine learning that can teach itself and learn on its own, just hit the scene and they just
started exploding. And one of the things that they were first trained on was games.
Yeah, and it makes sense. And if you thought chess was complicated and difficult, when it comes to these new AIS that they're teaching to teach themselves game strategy. They said, we might as well dive in to the Chinese strategic game Go because it has been called the most complex game ever devised by humans. Yeah, and this was actually that was actually a quote from Demi Hasabi, a neuroscientist and the founder of deep Mind, which was deep Mind.
They were purchased by Google or were they always part of Google.
I don't know if they were a spun off branch or where they were purchased, but it's one of Google's AI outfits.
Well, they're one of the teams, yeah, that are designing these new programs. And to give you an idea of how complex Go is, it deals with a board with different stones and there are ten how do you even say that?
Ten to one hundred and seventieth power, So.
That means one hundred and seventy zeros and take that number and that's the number of possible configurations of a go board.
Right, So, like you say, chess is very complex and complicated, and it's very difficult to master Go. And I've never played Go, of you no, So it's supposedly it's easy to learn.
Right, but very complicated in its simplicity.
Right, right, exactly, it's extremely difficult to master. And there was a guy in the late nineties and I'm guessing that he was saying this after Deep Blue beat Caspar ofv It was an astrophysicist from Princeton. He said that it would probably be one hundred years before a computer beats a human a go. To give you an idea of just how complex GO is that deeply would just be caspar OFV. And this guy's saying it'll still be one hundred years before anyone gets beat at GO by a computer.
And he was someone who knew about this stuff, who was an astrophysicist. He was just some schmoe at home and drunken as reclining.
Is making asinine predictions.
So and again we've said this before, but I want to reiterate the people that I think Alpha Go is the name of this program. The people that created this at Deep Mind, they wanted to stress that this is a problem solving program. We're just teaching it this game at first, just to make it learn and to see
if it can get good at what it does. But they said it is built with the idea that any task that has a lot of data that is unstructured and you want to find patterns in the data and then decide what to do right, And that's kind of like what we were talking about. It It crunches down all these possible options aka data to decide what move should I make right? And you could apply that Ideally, they're going to apply this to Alzheimer's and cancer and all sorts of things.
Right, it's general purpose thinking, right, Yeah, and thinking on the fly too, and face with novel stuff. So one of the reasons why it's good to use games like chess or Go or whatever, those are called perfect information games where both players or anybody watching has all the information that's available on it. There are definite rules or structure.
It's a good proving ground. But as we'll see, AI makers are getting further and further away from those structure games as their AI becomes more and more sophisticated, because the structure and the limitations aren't necessarily needed anymore, because these things are starting to be able to think on their own in a very generalized and even creative way.
Yeah, it's really really interesting. Yeah, the way that they're like you said earlier before the break, that we don't have computers that can run all the possibilities. So what they teach in the case of Alpha Go, this program teaches itself by playing itself in these games and Go specifically, and the more it plays itself, the more it learns, and the more ability it has during a game to choose a move by narrowing down possibilities. So instead of like,
well there are twenty million different variations here. By playing itself, it's able to say, well, in this scenario, they're really only fifty different moves that I could or should make, right, or that's kind of a simplified way to say it.
But right, No, but it's true. But that's exactly right. And what they're doing is basically the same thing that a human does. It's going back to its memory banks, Yeah, exact experience, huh, and saying well, I've been faced with something like this before, and this is what I used and it was successful. Forty out of fifty times. I'll do this one. This is a pretty reasonable move. Yeah, that is what humans do.
Yeah. Not only I mean, boy, we screwed up the chess episode, but I get the idea that when you're a chess master, you don't just think what do the numbers say and what does the book say?
Right?
But man, I did this move that one time and it didn't go as the book said.
Right.
So that's now factored into my thinking.
Right, except imagine being able to learn from scratch and get to that point in eight days or eight hours. Yeah, So that go team the alpha go the first the first iteration of alpha go, I think they started working on it in twenty fourteen and in twenty sixteen. At the end of twenty sixteen, they unleashed it secretly onto an Alpha Go website and it started just wiping the floor with everybody. Yeah, everybody's like, this thing's pretty good. Oh,
it's Alpha Go. That was the end of twenty sixteen.
Okay, so chess had already come and gone. Like, oh, by this point, you can download a program that's like Deep Blue, right.
That was That's a great point. Yeah, like today the stuff you played chess with on your laptop is even more advanced than Deep Blue was in the nineties, and it's just on your laptop. But this is so, this is Go. This is the end of twenty sixteen. The end of twenty seventeen, Alpha Go was replaced with Alpha Go zero. It learned what Alpha Go had taken two years or three years to learn in forty days by teaching itself, and.
It beat the master. Yeah, and finally in May of twenty seventeen, Alpha Go took on Key g the highest ranked Go player in the world. Don't know if he or she still is.
No. Lisa A. Doll is the current or was until Alpha Alpha Go beat him?
Oh man? Yeah, did they get knocked off and Alpha Go is the champion? Yeah? Like that's that's not fair.
I if it's match play and the player, the human player is accepted a challenge from the computer, I don't see why it wouldn't be the world champion.
Or do they just now say on websites like human champion? Maybe in italics? What's like a sneer?
Right? Maybe? Yeah?
Interesting?
What do they call that? Wetwear? Like your brain, your neurons and all that. What instead of hardware? It's wetwear?
Oh? I don't know about that.
I think that's the term for it.
What does that mean? Though?
It means like you you have a substrate, right your intelligence? Your intellect is based on your neurons and they're firing all that stuff and it's wet and squishy and meat. Then there's hardware that you can do the same thing on, you can build intelligence on, but it's hardware, it's not wetwear.
Oh.
Interesting, so that's probably it. It's the wetwear champion versus the hardware champion. But wetwear is italicized with the sneer.
So where things really got interesting because you were talking earlier about what is it with the chest and go? What do they called? What kind of games?
Perfect information games?
Right? Then you think And my first thought when you said that was well, yeah, and then there's there's games like poker like Texas Hold Them where there are a set of rules. But poker is not about the set
of rules. It is about sitting down in front of whatever five or six people and lying, bluffing and getting away with it in your game face being bluff Like there's so many human emotions and contextual clues and micro expressions and all these things, like surely you could never ever teach a machine to win at Texas Holding Poker.
Yeah, it'll be one hundred years at least before that happens, I predict.
No, they did it, and more than one team has done it.
Yeah. I read there was one from Carnegie Mellon called Liberatus AI Go melon Heads, Yeah, go the Thornton Melons.
Yeah, I mean that's was The University Alberta has one called deep Stack.
That was the one I read about. And it actually here's the thing, like if you read the release on it, you're like, you don't know how this thing works?
Do you really?
Yeah? And I'm pretty sure they don't fully get it, because that's one of the problems. I actually talk about this in the Existential Risks series.
That's scary that is to be released right.
That there is a type of machine learning where the machine teaches itself but we don't really understand how it's teaching. Probably the scariest one, right, or what it's learning, but that's the most prevalent one. That's what a lot of this is is like these machines. It's like here's chess, go figure it out and they go, okay, got it. How'd you do that? Wouldn't you like to know?
So that's the scariest presentation you will see on AI is when someone says, well, how does all this work?
And they go, but we just know it can be to human at poker. But the thing about deep Stack at the University of Alberta is that it learned somehow some sort of intuition. Yeah, because that's what's required is not just the perfect information where you have all the
information on the board. It's with poker, you don't know what the other person's cards are, and you don't know if they're lying or bluffing or what they're doing so that's an imperfect information game, so that would require intuition, and apparently not one, but two different research groups taught AI to into it.
Yeah, Carnegie Mellon came out in January of twenty seventeen with its Liberatus AI and they said they spent twenty days playing one hundred and twenty thousand hands of Texas hold them with four professional poker players and one and smoked them. Basically got up to They weren't playing with real money, obviously, but they they that would have been great.
They were playing with skittles like me as a kid.
Funded their next project, Liberatus was up by one point seven million, and one of the quotes from one of the poker players that he made to Wired magazine said, felt like I was playing against someone who was cheating, Like it could see my cards. I'm not accusing it of cheating. It was just that good.
Right.
So that's a really interesting thing, man, that they could teach self teach a program, or a program could teach itself intuition. Right, it's creepy. I thought this part was interesting, the Atari stuff. This gets pretty fun. Google deep mind. Let it's AI wreak havoc on Atari forty nine different Atari twenty six hundred games. See, they could figure out how to win, and apparently the most difficult one was Miss pac Man, which is a tough game. Still man,
misspac Man. They nailed it. It's still one of the great games.
But their their game, or their Q deep Q network algorithm beat it.
Yeah.
I think it got the highest score nine nine points, and no human or machine has ever achieved that high score from what I.
Understand, amazing. And the way this one does it, the hybrid reward architecture that it uses, is really interesting. It says here, it generates a top agent that's like a senior manager, and then all these other one hundred and fifty individual agents. So it's almost like they've devised this artificial structural hierarchy of these little worker agents that go out and collect I guess data and then move it up the chain to this top agent.
Right, and then this thing says, Okay, you know, I think that you're probably right what these agents are probably doing. And I don't know this is exactly true, but there's there are models out there like this where the agent says this is you have a ninety percent chance of success at getting this pellet. If we take this action, somebody else says, you've got an eighty two percent chance
of av this ghost if we go this way. And then the top agent, the senior manager, can put all this stuff together and say, well, if I listen to this guy and this guy not only while I evade this ghost, I'll go get this pellet. And it's based on what confidence level that the lower agents have in success in recommending these moves. And then the top agent weighs these things.
Wow, they should give him a little cap.
But all this is happening like that. Oh yeah, you know what I'm saying. This isn't like well, hold on, hold on, everybody, what is Harvey? What do you have to say? Well, let's get some Chinese in here and hash it out, and everybody sits there in order some Chinese food, and then you wait for it to come, and then you pick up the meeting from that point on. And then finally Harvey gives his idea, but he forgot what he was talking about, so he just sits down and eats his a roll.
Well, here's a pretty frightening survey. There was a survey of more than three hundred and fifty AI researchers, and they have the following things to say, And these are the pros that are doing this for a living. They predicted that within ten years AI will drive better than we do. By twenty forty nine they will be able to write a best selling novel. AI will generate this and by twenty fifty three be better at performing surgery than humans are.
You know. So again, one of the things that about the field of artificial intelligence.
You know a lot about now famous.
It is famous for making huge predictions that did not pan out. Sure, but you've also seen it's also famous for beating predictions that you know have been levied against it. But there is something in there, Chuck, that stands out to me, and that's the idea of an AI writing a novel. Like for a very long time I thought, well, yeah, okay, you can teach a robot arm to like put a car part or something somewhere if you wanted to just follow these mechanical things. Or it can use in auition,
or it can use logic in reason. But to create that's different, right. That was like the new frontier. It used to be chess and then then was go. The next frontier is creativity and they're starting to bang on that door big time. There's a game designing AI called Angelina out of the University of Foulmouth, which I always want to say Foulmouth, Yeah, but we'll just call it
Foulmouth like it's supposed to. And Angelina actually comes up with ideas for new games, not like a different level or something like you should put a purple loincloth on that player, you know, that'll look kind of cool like new games, but whacked out games that humans would never
think of. One example I saw is in a dungeon Battle Royale game, a player controls like ten players at once, and some you have to sacrifice to be killed to save the others, like the stuff that human wouldn't necessarily think of. This AI is coming up with.
Well, I mean, when you think of creatively, especially something like writing a novel or a film, if there are only seven stories, I mean, and that's sort of the thinking that they're basically every every dramatic story is a variation of one of seven things.
Yeah. So I mean you can look at like AI is scary, and in some ways it very much is and can be, but there's also like, definitely a level of excitement of the whole thing, and the idea that there are artificial minds that are coming online or that have come online now, that are out there that are they'll they'll just naturally by definition, see things differently than
we do. Yeah, and the idea that they can come up with stuff that we've never even thought of there is just gonna knock our socks off, hopefully in good ways. That's a really cool thing. And so maybe there's just seven as far as humans know, but there's an unlimited amount. Is if you put computer minds to thinking about these kind of things, that's the premise of it.
Right, so the robot would be like you never thought of boy meets girl meets well trilobite. But see even that's a variation of right.
Just imagine something that we've never even thought of.
Well, do you know how they should do this? If they do do that is uh? Is not is just release a book and not tell anyone that it was written by an AI program because if they do that then it's going to be so under scrutiny. Oh yeah, they should secretly release this book and then after it's a New York Times bestseller, say meet the Whopper, the author of this.
You know his interests are roller skating, playing Tic tac toe, and global thermal nuclear war.
All right, should we take a break and get strickland in here.
Yeah, we're going to end the Strickland drought because it is about to rain strickland in this piece.
You gross.
Okay, we're back and get this. The scent of strick has permeated our place. That's a beautiful scent.
It smells like a soldering gun and a circuit board and feel a lavender in a protein bar.
That's fair. I was gonna say Draco noir that would have been a lie.
Is that how you said? I always called a drakar dracr.
That's that's fair.
I always pronounced it Benetton colors. That was what I wore.
Oh is that what you wore?
Yeah? During my what I call the Year of Cologne, I had a couple of seven.
Uh.
This is scintillating. Why why am I here?
So we know that you already know because we talked via email about this, but we'll tell everybody nobody else. We have brought you in here because you are the master of tech and we were talking tech today, which we've talked about without you before. But frankly, Chuck and I and Jerry huddled and we said, this is not quite as good with that strict so let's try something different, gotcha.
And we're talking about games and machine versus man and that that whole evolution and how that's gone super crazy over the last few years.
Games without frontiers, as Peter Gabriel would say, yeah, or without fear.
And we've talked, I mean, we've talked a lot about the evolution of machine learning and how now it's starting to take off like a rocket because they can teach themselves, right. But one thing we haven't really talked about are solved games. I mean, we talked about chess. Yeah, we talked about go right, would those constitute solve games?
Not really?
So a solved game is the concept where if you were to assume perfect play on either sides of the game, you would always know how it was going to end.
Which we always assume perfect play, right, Yeah, it's kind of our.
Bags, right stuff.
You should know motto so perfect play just meaning that no one ever makes a mistake, so very much the way I do my work right stuff, you should know exactly So if you were to take a game like Tic Tac Toe and you assume perfect play on both sides, it is always going to end in.
A draw, which is what was in war games.
Yes, right, the only way to win is not to play right. Yes, so a game with like a game like Connect four, whoever goes first is always going to win, assuming perfect play both sides.
Yes, what, I don't think I've played Connect for it. That's where you drop ever a long time. That's wh where you drop the little tokens.
Yeah, like it kind of like Checkers.
You did an interstitial playing Connect four. Remember I was.
Faking it though, and you had perfect play, so I knew it was useless.
I was going to say that I'm so humiliated by all the Connect four games that I've lost starting even.
Yeah, but I mean perfect play. That's something that that obviously only the best players typically achieve with significantly complex games. Obviously the simpler the game, the easier it is to play perfectly. Right tic Tac toe, if you know, once you've mastered the basics of Tic Tac Toe and the other person has, you're never really going to win unless someone has just made a silly mistake because they weren't paying attention.
They put a star instead of an xer right.
Which doesn't count automatically disqualified you. One thing I've found that's very enjoyable is playing with little kids who haven't figured out that tic tac toe is very easy.
To yes and smash their face on the board and rub it in.
Yeah.
I mean the same reason why I like to join in on little league games because I can really whale that ball out of the park. Yeah, I really missed me feel like a man.
That's the most tech stuffy thing you've ever said. You really whaled that ball out of the park.
Well, to be fair, I did just do a textuff episode about the technology behind baseball bats, so it's so fresh on mine.
Nice.
Yeah one.
Actually, it's a lot of fun. So there have been a lot of games that have been solved, but Checkers will one that was recently solved back in recently by the early nineties when it was played against a computer called Chinook and chi in Ok.
Yeah, like the Helicopter or the wins that blow through Alberta exactly.
And so there are certain games that are more easily solved than others. You do it through an algorithm, but other games like chess, are more complicated because you can In chess you have multiple moves that you can do where you can you can move a piece back the way you win, right, it's not you're not committed to going a specific direction with certain pieces, like with a night you know you can you could go right back to where you started on the next move if you
wanted to, and that creates more complexity. So the more complex the game, the more difficult it is to solve. And some games are not solvable simply because you'll never know what the full state of the game is from any given moment. Did you have a chance to talk about the difference between perfect knowledge and imperfect knowledge in a game.
Yeah, yeah, we talked about that some. Yeah.
So computers, obviously they do really well if they understand the exact state of the game all the way through, if they have perfect knowledge.
All of the informations there on the board.
Right and all players can see all information at all times. But games like poker, which you guys talked about, obviously you have imperfect information. You only know part of the state of the game. That's why those games have been more difficult, more challenging for computers to get better than humans until relatively recently, and there have been two major
ways of doing that. You either throw more processing power at it, like you get a supercomputer, or you create neural networks, artificial neural networks, and you start teaching computers to quote unquote learn the way people do.
So we talked about that. Yeah, and one of the things that we talked about was how there's this idea that the programmers, especially say the people who are making programs that are playing poker and or getting good at poker, aren't exactly sure how the machines are learning to play poker or what they're learning. They're just getting better at poker. Yeah, do they know how they're learning poker? They just know that they're learning poker and that they're good at it. Now, like,
where's the intuition? How is that being learned?
An excellent question. The way it typically has learned, especially with artificial neural networks, is that you set up the computer to play millions of hands of poker that are randomly assigned, so it's truly as random as computers can get. That's a whole philosophical discussion that I don't think we're ready to go into right now. But you have games come up where the computer is playing itself millions upon millions of times and learning every single time how the
statistics play out, how different betting strategies play out. It's sort of partitioning its own mind to play against itself. And through that process, it's as if you, as a human player, were playing thousands of games with your friends and you start to figure out, Oh, when I have these particular cards and they're in my hand, and let's say we're playing Texas, hold them and the community cards are are these? Then I know that generally speaking, maybe
three times out of ten I end up winning. Maybe I shouldn't bet. Well, the computer is doing that, but on a scale that far dwarfs what any human can do, and in a fraction in the amount of time, and so it's sort of well, it's intuition in the sense of it's just done it so much.
Right, But does that mean it's completely ignoring micro expressions and facial cues, so that doesn't even come into play.
You should say, Strickland just nodded yet, Yeah I was.
I was waiting for John Well. I still nod when I do a solo show. And I do a lot of expressive dance.
What do you think, Jonathan, I don't know, Jonathan.
It gets lonely in here, guys. No, but yes, what you're saying all the tells, right, Yeah, tells that you would use as a human player. The computer does not pick up.
A speypically taking data.
Yes, typically, what it would do is it would study the outcomes of the games from a purely statistical expression, so that most of these poker games tend to be computer based poker games. So it's not that it's playing like it's not like there's a computer that says, pushed ten more chips into the table. You know, I tick right exactly a little it's a little winky face emoticon,
like I don't have good cards. No, it's it's all usually over, sort of like internet poker, which a lot of the people who play professional poker cut their teeth on, especially you know, in the the more recent generations of professional poker players.
Kids today, Yeah, they don't know what it's like being a smoky saloon like Moneymaker.
When Moneymaker rose to the top a few years ago, more than like a decade ago, now, he had come from the world of internet poker, and so he was using those same sort of skills in a real world setting. But obviously there are subtle things that we humans do in our expressions their computers do not pick up on it. In fact, that leads us sort of into the realm of games where computers don't do as well as humans.
Yeah, is that list you sent a joke or is it real? No, that's real.
It does seem like it's weird, like one of the games on there is pictionary, for example, rag or tag. Yeah, but these are some of these are They sound silly, But when you start to think about them in terms of computation and robotics, you start to realize how incredibly complex it is from a technical perspective, but incredibly easy it is for your average human being. Okay, so with humans a game of tag, once you know the basics, it's it's all an instinct.
You know what to do.
You run after the person you tried to catch up with them and tag them. But you also know push them ind as you can. Well, if you're Josh, you push them as hard as you can. But most of us we tag and we're not trying to cause harm. Robots, however, robots not so good on you said, I'm just saying
Isaac Asimov Isaac Asimov's rules of robotics. Aside, robots are not very good at judging how hard they have to hit something in order to make contact, right, They're not as good at even your bipedal robots that walk around like people, even the ones that can run and do flips and stuff.
Have you seen that one the other day that the footage of that thing running and jumping, it's really impressive and super creepy.
Yeah, but even so, that's that's a clip of the best of If you ever if you ever see the clips where they show all the times the robots fallen over.
Yeah, we're pouring hot coffee in someone's head.
Yes, but they always play those clip shows, toy.
Yes, this is true. So DARPA had its big robotics challenge a few years ago where they had bipedal robots tried to go through a scenario that was simulating Fukushima
nuclear disaster. So the interesting thing was the robot had to complete a series of tasks that would have been mundane to humans, things like open up a door and walk through it and pick up a power tool and use it against a wall, and you can watch the footage of some of these robots doing things like being unable to open the door because they can't tell if they need to pull or push or they open the door, but then immediately fall over the threshold of the door.
And when you see that, you realize, as advanced as robotics is, as advanced as machine learning has become, and as incredible as our technology has progressed, there are still things that are fundamentally simple to your average human right that are incredibly complicated from a technical standpoint, Like a.
Six year old can play jinga better than a robot.
Right right, right, Okay, But the thing is we're talking robots here, and as we go more and more and more online and our world becomes more and more web based rather than reality based, doesn't the the fact that a robot can't walk through a door matter less and less, and the idea that that machines are learning intellect and the robotivity, and you just blew my mind that that's becoming more and more vital and important and something we should be paying attention.
It absolutely is something we should pay attention to. I mean, we have robotic stock traders, the trading on thousands of trades per second, right fast, so fast that we have had stock market booms and crashes that last less than a second long due to that.
So that the robot army that will ultimately defeat us is not something from the terminator.
It's invisible, right, it's online, it will be online, it's.
It's it's what's determining our retirement.
Right, Yeah, the global economy or our municipal water supply or whatever.
Yeah.
No, There's the fascinating thing to me about this is not just that we're training machine intelligence to learn and to perform at a level better than humans, but that we're putting a lot of trust in those devices and things that have real incredible impact on our lives, significant enough impact where if things were to go south, it would be really bad for us, and not in that Terminator respect. Terminator is a terrifying dystopian science fiction story.
But then when you realize what could really happen behind the scenes, you think, oh, robots don't have to do any physical harm to us to really mess things up. So there are certainly some cases for us to be very vigilant in the way we deploy this artificial ants right from the outset exactly, and too depends not necessarily. I think I think it's I don't think it's too late, but I think it's getting to that point of no return very very quickly.
By December this year. Yeah.
Well, if you're if you're someone like if you're someone like Elon Musk, you'd say, if we don't do something now where we're totally going to plummet off the edge of the cliff.
But now is a window that is rapidly closing.
Yes, yeah, yeah, the now is the now is a time where we've got a deadline. We don't know exactly when that deadline is going to be up, but we know that it's not getting further out it. We're just getting closer to that deadline. So, and a lot of this is covered in deep conversations and the artificial intelligence and machine learning fields that has been going on for ages, to the point where you even have bodies like the European Union that have debated on concepts like granting personhood
to artificial intelligence. So this is a really fascinating and deep subject that and the games thing is a great entry point into have that conversation. Uh, you know. I'm lucky if I can win a game of chess against another human being.
Oh yeah, right, so I can't even describe chess.
Did I did? My big thing is I do that night thing. I call it the night shuffle. I just move them back and forth. I just cast.
If I can castle, then I'm so happy.
And that's the third tech stuffiest thing. They come in threes.
Well, Strick, thank you for stopping by.
Should stick around for listener mail.
I think you should too. I love to and throw out any funny comments that you have.
I'll throw out comments and then Jerry can decide which ones are funny.
Okay, all right, fair enough, all right, So if you want to know more about AI, go listen to tech Stuff. Strict does this every week what days.
Monday, Tuesday, Wednesday, Thursday and Friday.
Wow, that's amazing, buddy. And wherever you find your podcast yep, okay, And you've been doing it for years, so if you love this, there's a whole big backlog, nine hundred plus episode.
You're celebrating your ten year as well, right yep, by.
Sure am, I'll be We'll be turning ten and tex stuff on June eleventh.
Congratulations, well, since I said happy anniversary, it means it's time for listener mail.
Guys, I'm gonna call this Matt Groening and cultural relativism about that?
Nice?
Hey, guys, love your podcast so much. The massive archive makes for endless learning and entertainment. My favorite part is you were such rad guys, including Strickland, and I could totally imagine how did they know? I can totally imagine myself getting a beer with you two, but without Strickland, your Simpsons episodes were absolutely perfect. I still live in Portland and drove on Flanders and Lovejoy Streets a lot.
Wait, is this Matt Groening? Okay?
Matt Granning Drew Bart in the sidewalk cement behind Lincoln High School in downtown Portland. You can google that. I would like to offer one interesting observation though, I've noticed that on several episodes you guys have said that you are cultural relativists. Is that pronounce right? Yeah? Yeah? But then in nearly every episode I hear you pass moral judgments on all the messed up stuff that people do, whether it's racism, preak shows, or crematoriums bearing bodies on
the sly. You guys are never shy to condemn something that deserves to be condemned. Reminds me of something I read from Yale's sociologist Philip Gorsky, who points out that our own relativism is rarely as radical as our theory requires. We can't be complete relativists in our daily lives. He then gives the example of how academic social scientists, where diehard relativists, get furious and moralistic at the data fudging
of other researchers. Anyway, love the show, guys, love tech stuff especially, and will forever be indebted to you for your hilarity and knowledgeability. Cheers Jesse Lusco.
Ps go tech stuff.
That's sweet.
Love that. Yeah, thanks a lot, Jesse. There was an actual episode, and I don't remember which one it was, where we a and in our cultural relativism, do you remember, because we used to just be like, no judgment, no judgment, right, we just can't judge, you know, And then finally we were like, you know what, No, that's not true. We changed our philosophy to include the idea that there are moral absolutes that are universal, although sometimes we are just
judge even beyond that. Look at us, Yeah, Well, if you want to get in touch with us, you can send us an email to Stuff podcast at HowStuffWorks dot com. You can send John an email to.
Tech stuff at HowStuffWorks dot com.
Nice, and then hang out with us at our home on the web. Stuff youshould know dot com and just go.
To tech Stuff. Just search it in Google. I come up all the time.
Fair enough.
Stuff you Should Know is a production of iHeartRadio. For more podcasts my heart Radio, visit the iHeartRadio app, Apple Podcasts, or wherever you listen to your favorite shows.