Episode 34: Alpha Go and Creativity
- Links to this episode: Spotify / Apple Podcasts
- This transcript was generated with AI using PodcastTranscriptor.
- Unofficial AI-generated transcripts. These may contain mistakes. Please check against the actual podcast.
- Speakers are denoted as color names.
Transcript
[00:00:11] Blue: Welcome to the theory of anything podcast. Hey guys, how are you doing today doing great Bruce? How are you good? Well, mostly I’m I am actually doing well. You probably know I have a kidney stone. So I’m not necessarily doing great How are you Tracy I’m well good today We’re going to do Alpha go the movie which is a movie that is free on YouTube. Just go Google Alpha go the movie and Don’t listen to this episode until you’ve gone and you’ve watched that movie This this episode is going to contain spoilers now You may think no wait isn’t Alpha go the movie a documentary How can a documentary have something that can be spoiled? I’m telling you go watch the movie And you don’t want to listen to this episode first because it’s going to contain spoilers I thought the movie was just excellent it was tense it was Exciting I had been asked to watch it for one of my classes. I think it was my deep learning class They required it as part of the curriculum, you know required in the sense that nothing in college is truly required ever But it was something that we were supposed to do so I went and watched it And I was so excited by the end the movie and you know what I knew what was going to happen I was already familiar with the story So even though so it had already been spoiled for me to some degree and I was so excited about it Took my wife and whichever kids I could get to watch it with me and I played it on the in the theater downstairs and My wife was like so tense all during the movie.
[00:01:51] Blue: It’s just an excellent dramatic Story that
[00:01:55] Red: happens
[00:01:55] Blue: to actually be true. Sorry go ahead Tracy.
[00:01:58] Red: Oh, yeah, I was just saying it’s very suspenseful
[00:02:00] Blue: Yes, it is And we’re gonna spoil it all so go watch the movie first and then come back to this episode And we’re gonna do kind of an analysis of the movie and we’ve already kind of spoiled parts of it like if you watched our Reinforcement learning episode we talked quite a bit about Alpha go the movie in there And I probably should have warned people then that there was spoilers in it. Anyhow ladies Let’s just start at the the beginning here with the story. So David Silver is a famous guy in machine learning in reinforcement learning and he has a series of Lectures that are available on YouTube, which are excellent where he teaches reinforcement learning So if you watched our reinforcement learning episode and it Made you curious you can learn about the full theory of reinforcement learning at a college level From David Silver probably one of the world experts in reinforcement learning. He’s a very good teacher I say the only real downside to the lectures is that I had a hard time understanding him sometimes He has an accent and the sound isn’t the greatest, but he’s someone who really knows this material well He is in charge of Alpha go. So the setup for this is is that go is It’s sometimes called Chinese chess, which is a total misinterpretation But it’s it’s kind of their strategy game that they like similar to how You know we in America might enjoy chess, but the games are nothing similar at all It’s a much harder game to write a program for to play and the reason why is because the way they usually do
[00:03:42] Blue: Game -playing algorithms is with the mini max algorithm, which basically just tries making a move Tries making a move for its opponent tries making its own move It just tries as many moves as it can out into advance as far as it can quickly that becomes an exponential Nightmare, you know if you can get seven moves out, you’re doing great with go the branching factor the number of possible moves is so large You just can’t really use the mini max algorithm effectively and then to make matters worse It’s not in chess You can kind of come up with really simple algorithms that can tell you Oh, you know your seven moves out if you do this move some moves from now You’ll be in a better position and it can determine quote better position Based on some simple algorithm that says well, you know You haven’t lost your queen and you’ve got this many points for the pieces on your board And they can come up with very simple algorithms that tell you if the board position seven moves out is good or not There’s nothing equivalent for that for go Go masters use their intuition to be able to tell if their board position From this move is good or not and so how do you get a computer to do intuition? So this is kind of the setup that there are go playing algorithms when David Silver steps onto the scene and they’re really bad They can only play at an amateur level
[00:05:01] Blue: Professionals think it’s a joke that it’s kind of a common joke amongst professional go players about how bad Go playing Computer programs are okay, so this is kind of the background They don’t I don’t know if they fully explain that in the movie or not But that is that is the background for which David Silver then decides I’m gonna make this go playing algorithm that can actually compete at a professional level and gets a team together as You know for at his university or whatever whoever is funding it and that is what AlphaGo is there They’re setting out to try to Beat professional level players if possible they want to beat the world champion who’s least of all They don’t know what it’s gonna take to do it They’re just gonna try to use the theory of reinforcement learning to feel like that and deep reinforcement learning I’ll explain what that is in just a second
[00:05:51] Blue: Our technologies that in principle might be possible to build a professional level go playing algorithm out of But no one’s in the world ever done it before they don’t know if it can really be done or not And then here’s a quote from David Silver and we’ve often talked about the fact that so much of what goes into machine learning is Human knowledge, you know machine learning does create knowledge, but it’s not very much It’s mostly knowledge coming from humans and David Silver admits this This is a quote from the movie He says everything that AlphaGo does it does because a human has either created the data that it learns from created the Learning algorithm that learns from the data created the search algorithm all of these things have come from humans So really this is a human endeavor and this is actually an important point because everybody is kind of cheering on the human and Not the computer because we relate to Lisa doll being human being human But he’s kind of making the point Both sides are in a sense human right that you’ve got this team of programmers that are all human And they want to be able to make an algorithm that has never existed before That can do something that’s never been done before This is a very human thing for them as well, right that they they want to be successful So for this movie this made this somewhat exciting for me You can’t help but cheer for Lisa doll the human player But you’re not really against Team AlphaGo because they’re a bunch of humans that you really kind of are rooting for too At least that was for me. What did you guys think?
[00:07:23] Green: I um, you know, everybody is a very sympathetic character you end up rooting for for everybody and it’s it’s hard when people are getting beat that That you know that they’ve dedicated their entire life to being masters And now they’re getting beat by a by a computer that they don’t even understand. So it it’s I found everybody to be a sympathetic character.
[00:07:49] Blue: Yeah with this in mind as the background team alpha go contacts I don’t I I can’t even pronounce his name, but it’s contacts a go player who’s A professional level but not a strong professional level player He’s kind of somewhere in the middle somewhere They they contact him and they say we’re working on a program to play go We need someone who’s good at go to help us We’d like to pay you to come and come in and Be part of our team and and he’s thinking Okay, this is dumb. You know computers can’t play go. I get I don’t I’ll show up I’ll see what they’re into. I don’t know what they need for me You know and and he’s not really sold at all on what’s going on when he first shows up What happens when he plays his first game against him? Do you guys remember?
[00:08:44] Unknown: I
[00:08:44] Red: just watched it this morning I think he loses he
[00:08:49] Blue: does He loses his first game and he’s so embarrassed that he he walks out He walks out and the team is the out team alpha goes thinking he might not come back And he walks wonders off and he’s thinking to himself and he’s just humiliated because to lose to a computer in go Is like the most humiliating thing that could ever happen to you as a go player because it’s so well known that Go programs just don’t play well And he’s walking along being all humiliated and it finally strikes him Oh my gosh, they’ve got something It just beat me And he runs back and he goes I’ve got to be a part of this It’s
[00:09:37] Green: what he can put his own his own ego aside to realize what a phenomenon
[00:09:41] Red: like through it is.
[00:09:42] Green: Yes
[00:09:44] Blue: He’s like they’ve made a go playing program So he runs back and he wants to participate and he gets super excited And he’s one of the main narrators of the documentary through the rest of the movie And he’s he’s interesting because he’s on team alpha go and you can tell he is He really wants to see alpha go win from this point forward But he’s coming at it from the standpoint of the person that helped train alpha go, you know that Um, you know, this is his baby to some degree. He’s not at least at all’s level and he knows he’s not right so Alpha go is way going to exceed his skill level and he knows this But he kind of just understands what it’s like to play alpha go and to be shocked When you suddenly realize this is no normal or go playing program Right and so you get a lot of kind of the sense of what it’s like to be a go player through his eyes throughout the movie so They train up and they they do a challenge to lisa doll who’s the world champion lisa doll
[00:10:49] Green: Excepts and they’re gonna do five games
[00:10:52] Blue: Lisa doll initially He takes the stance Oh, i’m gonna beat it all five games and he seems pretty confident because of course, he knows alpha go programs their bath But at least I don’t know about you guys But it seemed to me like maybe he was just a little bit nervous like you can see him kind of asking questions Oh, I saw it playing your go consultant and and has it gotten better since then They’re like, well, we can’t tell you how much better it is But you can kind of tell that he knows he’s Jumping into the unknown here a little bit and I thought that was interesting too a very very human sort of thing He starts off very confident But he’s not quite sure what he’s dealing with right and he’s never played this this program before He’s seen games with this other go player that who’s not on his level He thinks he can beat it based on seeing those games that it’s playing but Of course, they’re improving the algorithm like right up until the day of the game So what he’s actually going to be playing is not the same algorithm that he’s actually seen It’s going to be something better by the time it gets there So I thought that was interesting too. There’s sort of this tension about how well can the algorithm play and You get a good feel from the from team alpha go How nervous they are that they’re taking their little program that’s only played a mid -level You know go player up to this point and it’s suddenly going to be playing the world champion and It might embarrass them, right? Yeah I
[00:12:26] Green: think you can feel like this Definite like they they’re afraid that he’s that their program is going to get beat,
[00:12:33] Red: you know, three moves in four moves in
[00:12:35] Blue: Yeah Yeah, and they’re also they’re also worried about the game hallucinating. So that’s what they call it They’re worried that the game will at times It will think it’s winning when it’s not It’s like hallucinating a victory that’s not there The game has a history of doing that where You know some percentage of the time it suddenly its algorithm says, oh you have You know a 90 chance of winning with this move and it’s it’s just a bad move, right? It’s it’s completely off with its calculations And when that happens the game doesn’t play like a human and Still put up a decent fight. It just starts to do stupid things embarrassingly stupid things So the biggest fear is that it will hallucinate in the middle of a game And it will start to do really embarrassing moves and when it loses it won’t lose You know gracefully or put up a good fight. It’ll basically just Start doing incredibly stupid things and the team will be embarrassed So there’s this real threat that the team might get embarrassed by lisa doll. Who’s the world champion go player So you kind of feel for them too At tracy, I know you you were wanted to ask about hallucination. I tried to explain it there Do you have any other questions about what it means to hallucinate?
[00:14:03] Red: Well, no I was just confused by the term because I think most people think that you literally imagining Something that’s not there and in this case I have for computer I guess I’m just thinking that Is hallucinate the right word? It’s kind of weird because I think the computer’s I don’t know. Is it imagining versus it’s trying to project or predict Versus just imagining something that’s not there
[00:14:27] Blue: Yeah, you know That’s the word that I believe the team just used or maybe they used the word diluted. I think maybe they used both but If you really think about it, this is just humans making up a word for something, right? I mean it’s They call it diluted or hallucinate But it’s not because it’s somewhat analogous to when a human is diluted or hallucinates But it’s not really the same thing. It’s just an analogy to it
[00:14:53] Green: so I was I will admit I was a little confused on how that would impact The programs about like what did they think that the program would do once it started hallucinating That would be different from what they had Been training it to do
[00:15:10] Blue: So if it gets into one of these states, so to understand why this would happen I Think back to our reinforcement learning episode. Okay Now even if you only just got the gist of the theory of reinforcement learning Remember that there was a world space, right? There’s this solution space where you have to Have every have to have a table a q -table with every single possible Combination that’s possible in this world space. So what’s the world space for for alpha go? It’s um It’s every possible configuration of the board. Okay, so imagine trying to make an array and The array the first element in the array represents a board with no pieces on it And then the second element in the array represents a board with only white having one piece You know in the bottom right corner or something like that Okay, and then you’d have to just have one something in that array for every single possible board position That could possibly exist According to the rules. So an array that large Would be Too large for any computer. There’s no computer on the earth that could store that much memory It’s it’s just too large, right? And this is one of the problems with the theory of reinforcement learning Is that you even for really small The world space is that it can actually handle It can only handle small ones because this q -table grows out of control really fast, right? It’s exponential growth So and then even if you could fit it all into memory, you’ve still got a second problem.
[00:16:45] Blue: Okay The way it works think about how when the little robot went up into the corner Slowly the numbers move out from the goal And the reward moves forward and it Calculates a value for each space until it finds the spaces that are next to the goal And then those get a higher value State. Okay. Remember how that works in from the episode of reinforcement learning?
[00:17:10] Green: Yes,
[00:17:11] Blue: yes Okay So if you have that many spaces How many games would you have to play? For it to actually learn that this board this specific board configuration Is this close to the goal you would have to play probably Trillions to the trillions games Right, I mean like it would be an enormous amount of games You would have to play and this is the second problem with reinforcement learning Is that even even if you can fit and I actually tried this I did a lunar lander reinforcement learning algorithm And I tried to just make a state space That just used a regular q -table and it that it crashed that lander like you wouldn’t believe I mean it was The problem was is is that it I just couldn’t get it to train enough To actually fill up the table so that most the table was always zero and it just didn’t know what to do. Okay so, um Let’s use this as kind of a starting point imagine that Uh, they actually implemented this with the q -table which they didn’t I’ll explain what they did instead But it’s a similar problem Imagine that you got the game into a state that it has never seen before So it it it’s q -table it gives it a value of zero It’s gonna basically make a random move at that point. Okay, so humans never make random moves If it starts making random moves It’s gonna look stupid. It’s gonna look like it’s a little child playing the world champion, right? So it’s gonna be embarrassing for Um the programmers does this kind of make sense?
[00:18:50] Red: Yes. It does Okay
[00:18:52] Blue: so Since you can’t possibly fit the world space world space for this, um For go into a computer and if you did it wouldn’t work anyhow because you’d be able to play enough games What do they do instead?
[00:19:04] Blue: What they do is they they keep in mind that um neural networks They do a very good job of mimicking any function So think of it like this like we somehow recognize faces we go on facebook We see a picture to recognize a person’s face So we know there’s a function that humans use to recognize faces We just don’t know what it is and I talked about this in the machine learning episodes where Visual recognition is something we don’t really understand and so we didn’t know how to program it And so it turned out to be easier just to let machine learning learn how to do it Well, the way they do that is with neural nets neural nets will come up with some sort of function That works at least well enough right and it doesn’t necessarily do it the way a human does And that’s that’s why they have these things called adversarial examples where we see I have a picture of a dog And it recognizes a dog and then you go change a couple pixels and it still looks exactly like a dog And human can’t tell the difference and now the computer thinks it’s a giraffe Right, I mean like there’s those funny things you can do because of the way it comes up with its functions Okay But that’s what neural nets do neural nets will take any function and the whole world is made up of functions That’s part of the theory of computational theory that everything can be simulated as a function So we’re basically just asking the neural net through this it’s training algorithm through using gradient descend as it’s heuristic We’re asking it to try variants of weights and select the best ones Which is what we talked about when we talked about knowledge creation and machine learning until it finds some sort of local minima That works pretty well.
[00:20:42] Blue: You can do that for a q -table. You can say, okay We can’t fit the true q -table and solution space in there So we’re going to use a neural net That takes inputs like a q -table would and gives outputs like a q -table would But it’s just the neural net doing it instead, which is going to be much smaller And neural nets are good at generalization So the neural net will figure out on its own That this board position is somewhat analogous to that board position And it will end up with some sort of knowledge about what to do for every single board position Even though you never actually reached some of the board positions that it has to play Okay But if the neural nets hasn’t been trained well It will start to act a lot like a q -table that’s you reached a state where it just doesn’t know what to do if it if you hit some Combination and it Thinks it’s analogous to something that it isn’t or it just has never really seen anything like that before It’ll just start to do random moves. Basically. It’s the same problem, but on a much smaller scale Okay And I think part of the reason why alpha go Had this particular problem is that the original alpha go used human training examples.
[00:21:52] Blue: It would train off of human games The later alpha go didn’t it just played itself So it didn’t have to use any input of human games at all for its training That version where it just played itself It would just create its own data by playing games with itself and play it billions and billions of times That version didn’t seem to have anywhere near the level of the hallucination problems that the earlier Problem did when it was still trying to use these human games, which means that I mean how many human games have been played in the history of the world, right? I mean Alpha go the final version that doesn’t use human human games It plays more games than it probably been played in the history of the world by humans Okay When you’re trying to train off of the human knowledge, you’re leaving all sorts of gaps in its knowledge as to what to do Okay, but it kind of a shortcut initially that you can show what a good game looks like and it can use that for Generating board positions, but uh, it’s gonna be a weakness also So that’s really kind of what happened there from the technical standpoint And they hadn’t yet figured out how to get it to stop doing that and so there was always this threat That the game might just embarrass them Did I answer your question cameo?
[00:23:02] Green: It does it it does. That was a great explanation.
[00:23:05] Blue: Okay So we get into the game and Do you guys remember what happened? Is one of you want to kind of Summarize what happens in the the first game there?
[00:23:16] Red: Oh, uh, well It was just very tense. Um, you know we If there’s no obvious face to an opponent, which is terrifying and it started out really to me it seemed hesitant and It was anxiety written because the um alpha go took so long to make its first move There was a lot of doubt and like is this really gonna work How’s this gonna
[00:23:40] Blue: play? So what you just mentioned is really interesting When humans play there’s only certain opening moves you can make so humans make an immediate open move Right, they don’t stop and think about it for 10 minutes.
[00:23:54] Red: Right. Yeah
[00:23:55] Blue: So alpha go stopped and thought about it for 10 minutes About which move it wanted to make and and it’s it for an opening move that seems like wow It’s is it just confused But it turns out it stopped and thought about each move for 10 minutes pretty much the whole way through You know, it just always just stopped and thought for some period of time and and it was very Unhuman -like in that regard. So it kind of threw The human commentators and Lisa doll off because they weren’t quite sure what to make of that
[00:24:28] Green: Well, I think everybody is kind of thinking that the machine’s not able to make the next move, right? Like they’re They’re not sure if because why would a machine need to think right, you know, and and even that uh that phraseology Like what what is the machine actually doing during these things? Is it running multiple variations of the move?
[00:24:54] Blue: Yeah, it’s it’s running. It’s probably the way alpha go is programmed. It does use a minimax algorithm. So it actually runs through Every possible move several moves out It just does that regardless of where it is. It doesn’t care if it’s the first move or You know the 10th move or the 100th move it always would just try to run multiple different moves and Try to think outward what its best move is
[00:25:17] Green: Yeah, but they they they do it all of the humans are like, why is it taking it so long? It
[00:25:23] Blue: kind of makes sense though a human player would have a heuristic mind They would think well, there’s only certain good opening moves. I’m just going to pick one of them Whereas the computer doesn’t have a library of opening moves They probably could have given it a library of open moves to make it more human -like But they didn’t instead the computer had to go run a simulation Of a little game in its head to figure out what a good open move would be does that make sense? Yeah, it does. Yeah, so The first big surprise was what they call move 37. Oh another thing to tracy’s point Alpha go is played by one of the programmers. So the programmer has a computer screen up He enters the move that Lisa doll makes And then he waits for alpha go to come back and when alpha go makes the move he then Does that move on the board according to what alpha go told him to do? And lisa doll Keeps trying to read the programmers face. Yeah. Yeah Disturbing Is he’s trying to figure out what the programmers thinking and then he suddenly realizes the programmer doesn’t know what’s going on I’m not playing the programmer. I’m playing the computer Well,
[00:26:38] Green: and it’s especially funny because the programmer doesn’t actually know how to play go at all. He doesn’t. Yeah He doesn’t understand the game at all
[00:26:47] Blue: so A real player would read the tension in their in their opponent’s eyes and would Determine what’s this person thinking and what they’re doing? And lisa doll suddenly finds that that part of his skill set is useless Because the person he’s playing against sitting across from him on the table It doesn’t know anything is totally can’t give anything away So I thought that was an interesting thing there so Move 37 happened while lisa doll was on a break So lisa doll went up and got went to go out and take a break And the programmer who I believe they said his name is aja So here’s what the go consultant said he says so aja Who’s the human playing the alpha go sees alpha go plays move 37 and aja puts the stone on the board When I see this move for me, it’s a big shock. Normally humans. We never played this one because it’s bad. It’s just bad And here’s what the commentator said about move 37. Oh, it’s It’s totally an unthinkable move. Yeah That’s a very that’s a very surprising move And then like one of the programmers said I thought it was a mistake and he laughed So move 37 gets made and the initial impression of He’s knowledgeable commentators and the go consultant who’s a knowledgeable mid -level go player himself Is uh, oh, maybe that’s a mistake lisa doll comes in And the go consultant’s waiting for lisa doll He wants to see what lisa doll is going to think of that move because it was made while he was out on break And lisa doll comes in and looks at the board and here’s what lisa doll Said about it.
[00:28:31] Blue: He said I thought alpha go was based on probability calculation And that it was merely a machine But when I saw this move I changed my mind Surely alpha go is creative. This move was really creative and beautiful. So lisa doll sees That this move is not a mistake that alpha go has made a move That humans wouldn’t normally make human players do not make a move like this They’re built up heuristics and intuitions and such you don’t make a move like this But when when lisa doll saw the move While everybody else was thinking a mistake. He suddenly realized Whoa, that was a smart move Even though it was a one that was a very inhuman sort of move It’s interesting to hear the commentators at this point. Let me see that got David silver He says the professional commentators Almost unanimously said that not a single human player would have chosen move 37 So I actually had a poke around an alpha go. So alpha go has you can like ask it um questions like it’s got Analysis of its own moves and things like that so you can like pull up the data and look at it So as I had a poke around an alpha go to see what alpha go thought An alpha go actually agreed with the assessment alpha go said that there was a 1 in 10 000 probability that move 37 Would have been played by a human player So it knew this was an extremely unlikely move It went beyond its human guide and it came up with something new and creative and different Is what david silver said and then one of the narrators in the show.
[00:30:03] Blue: He says I’m very much watching the game through the commentators. That’s the way it works So when they’re confused, I’m certainly confused at the same time I’m latching on to the fact that they are confused, right This is this that is an interesting moment when everyone is confused who is not confused, right? Besides the machine so alpha go with move 37 It made a creative move that Was outside of human knowledge This is one of the things that was so interesting about it And yet the world master even if most the commentators couldn’t tell immediately goes Wow, that’s an amazingly beautiful move and starts to realize. Uh, oh I’m I’m not playing a normal computer Go playing program here. So that was one of the most exciting moments that was early on in the game where It it makes this move kind of makes people start to realize This is no regular go program and lisa doll starts to get kind of worried. Uh, oh What’s what’s going on here? You know, this this is something that can think differently than a human It’s not like a human master, but it’s a master all its own He goes on to lose the first game Of course, it’s humiliating that he lost to this to this game to this computer But he’s he’s got a great attitude about it And he’s like i’m gonna come back tomorrow and i’m gonna win the next game and and he’s kind of even excited Maybe that he’s found an opponent that’s worthy of playing, right?
[00:31:35] Blue: It also leads to what when alpha go has a victory the media attention starts to get a lot more serious because now it’s actually beat the uh World go champion once so that makes it way more interesting of an event So the media scrutiny and attention starts to really heat up at this point in the story so Now I don’t know. I can’t remember the exact order of games That that takes place. Um, I believe it was Alpha go. No, I do remember alpha go goes on to beat lisa doll three times in a row And at this point lisa doll is actually lost, right? He’s still got two more games to play But alpha go is going to be the overall victor at this point because it’s already won three out of five games Right lisa doll has got a great attitude about this though. He’s he’s his tone has changed though He’s gone from you know, I’m gonna beat it every game to I’m gonna beat it at least once And this is what all the humans are cheering for now Is there’s so much tension over can lisa doll beat alpha go at least once The excitement is around that now is that alpha go is no longer thought of as oh this little go program It’s now the world champion effectively, right? And
[00:32:48] Blue: and it’s like can we get a human to beat this this unstoppable machine Go playing machine And a lot of the tension in the movie kind of shifts towards that at this point where lisa doll is trying so hard to Beat alpha go at least once And this is where one of the so move 37 is one of the big spoilers This next one the god move Is probably the biggest spoiler. So that’s why hopefully you’ve watched the movie first Lisa doll comes back and beats alpha go on the fourth game Everybody’s celebrating and they’re so excited and it’s like yes, we humans. We’re not out of the race, you know There’s a great deal of tension around that. But what’s really interesting is how lisa doll beat alpha go Um in game four they they called the move that he used where you beat alpha go They called the god move. I think it’s also called move 78 or something like that.
[00:33:40] Blue: I can’t remember exactly what it was called Lisa doll As he’s playing alpha go he starts to realize That it’s teaching him ways to play go that he’d never thought of before because he’s only ever played humans before And that he’s starting to think differently about the game of go because alpha go Treats the game of go differently than a human would And in in game five we can explain better what I mean by that because it becomes super apparent How alpha go plays differently than a human during game five But lisa doll is already starting to pick up on it and it’s starting to create he’s starting to think of new sorts of strategies that he’s never thought of before because of him playing against alpha go And this is what so move 37 in particular taught him something about the game He didn’t know and it caused him to then come up with the god move so that he was able to beat alpha go This is what I think is interesting is how um alpha go actually created a whole new play style That lisa doll then started to pick up on and started to use against it And I think that’s what part of what makes game four so exciting is when he kind of figures out I get now how I can beat the machine. So here is what um Here’s the quote it says the lessons that alpha go is teaching us are going to influence how go is played for the next thousand years At the very end the go consultant.
[00:35:06] Blue: He says Where we look back and say Yeah, that was just like move 37 something beautiful occurred there at least in a broad sense move 37 begat move 78 Begat a new attitude lisa doll a new way of seeing the game He improved through this machine. His humaneness was expanded by playing this inanimate creation And and that’s actually true from what I understand World champion go players Have whole new play styles because of the advent of alpha go that it it introduced new play styles into the game That nobody had ever thought of before and it changed the way humans play go also so then probably so after lisa doll beats alpha go in game four and Everybody’s really excited about this victory for the humans. Honestly Lisa doll might as well have won the entire tournament the level of excitement of him beating alpha go once Was just so through the roof Everyone was so excited to see a human stumped the machine for a change Um, the fact that he had actually already lost three games Almost didn’t matter because at this point everyone knew alpha go was nearly unstoppable And that’s kind of weird like the end of dog suddenly won and he was no underdog really.
[00:36:19] Red: Yes
[00:36:20] Blue: He’s the world champion that we’re talking about right? They go into game five and lisa doll is Feeling somewhat confident that maybe he can beat alpha go again and get at least two victories against it and this this leads to Probably the most interesting aspect of the documentary even even though the god move is super interesting And move 37 is super interesting. This last part is Almost it’s almost like a comedy. It’s funny Um, so what happens is that early on in game five? um, oh by the way, that was One of the things that happened is when lisa doll made the god move it made alpha go become a little confused Just like we were talking about the hallucination where it deludes itself. I don’t think it necessarily played bad but um, it it could not Quite figure out what to do from that point forward and it was kind of obvious that it couldn’t He had he had figured out a super creative move That that it hadn’t foreseen strangely as that may sound That allowed him to go into a board position Where it just couldn’t figure out how to recover it continued to play from that point forward But it was just it was kind of obvious it was over just just like when it made move 37 it became kind of obvious Oh once they started to realize what it had done they started to realize. Oh the game’s over alpha alpha go with that move It had won the game. It’s just a matter of time now And apparently that happens and go so in game five in the final game Uh, it make alpha go makes a move and similarly move 37.
[00:37:58] Blue: Everybody thinks it’s a mistake Only this one was a lot weirder. This one was Move 37 at least lisa doll could see it was a beautiful move, right the world champion in this case Everybody thought it was a mistake So here’s here’s what the commentators are saying Uh in the documentary they say is it fair to say that alpha go made a mistake We might have another victory today for lisa doll We’re you weren’t crazy about the timing on this move and the other guy says, yeah I’m sort of thinking that maybe alpha go hasn’t recovered from game four yet. Yeah, they kind of chuckle And then it’s one of them says I think it could be a kind of misreading Says there’s no reason for why keep playing that move. These are like the The commentators who are watching the game. You’ve got these people who are commenting on the game And in the background the programmers have seen this move and one of the programmer says still diluted another one says We don’t know and then one says That is looking good for lisa doll and they’re like, yeah, it looks like lisa doll is going to win this game And the programmer says are we seeing another short circuit? So they check what alpha alpha go thinks and one of the programmers reports Alpha go is saying it’s 91 percent certain now it’s going to win and the other programmer says Yeah, because it’s incorrect again So the um and the consultant said it’s a bad move. Oh, maybe they all
[00:39:17] Green: think that the machine’s just messing it
[00:39:19] Blue: up That’s right. It says, oh, maybe alpha goes weakness comes back. It’s a bad move says the alpha go consultant that trained it It continues after this move after it makes this bad move and everybody thinks that’s it lisa dolls won the game It continues to make bad moves Throughout the rest of the game So it really looks like it’s just hallucinating and it’s just doing these crazy moves Right as the game moves on The commentators suddenly change and they go wait, I think alpha goes winning And it turns out that all these crazy moves that it was making We’re actually really intelligent moves that no human could recognize as a good move the reason why And one of the things that the programmer says is the whole game. We thought the alpha go was wrong about the board position So it analyzes this board position. They thought it was hallucinating. It’s oh, it thinks it’s got a 91 percent chance of winning But it’s it’s hallucinating says we were super worried that oh, it’s going to play garbage It’s going to to be like lose in a very embarrassing way and this continued the whole game as it turns out None of us know go well enough to actually judge what alpha go is doing And the narrator says we all we all say some of alpha goes moves are so weird and strange and maybe mistakes But after a game is finished.
[00:40:41] Blue: We have to doubt ourselves our judgment What alpha go was doing is it it’s it’s true play style came out during game five and the easiest way to explain it is that Throughout the history of the world People had been using number of points as a proxy for chance of winning So human players consistently And this makes sense, right? I mean like this. There’s no big mystery here Consistently tried to go for as many points as they could possibly get On the grounds that that would increase their their margin for the win, right and increase their chances of winning Alpha go realized that wasn’t The way it worked that you won by winning by one point So it played it entirely differently It would almost throw moves away if it had to once it once it had consolidated what it was going to win with It’s board position it was going to win with It would simply defend that board position and it wouldn’t even try to take new territory after that And it would make moves that seemed like it was just throwing stuff away It was just waiting biting its time so that it could now win the game
[00:41:52] Blue: And that was why it seemed like it was making all these stupid moves, but it actually knew it had won the game If it already figured out it had won the game And it was just waiting for lee to come along and rest everybody else all the humans to come along and realize Oh, it’s one and until near the end of the game None of the humans recognized that is that it had already won the game way back with its first crazy move If you think about it from a just the theory of reinforcement learning standpoint What reinforcement learning does is it tries to learn This board position is closer to the goal of winning than this this other board position Right, that’s what it’s trying to learn. Okay That isn’t the same thing as getting the most points. It just isn’t you can see how being defensive might be a better strategy In some circumstances than being offensive, but human players never realized that before they had never understood That you only have to win by a point So sometimes it’s better to be defensive than it is to be offensive. Well alpha go ahead figured this out Okay, that was part of what its algorithm figures out as it tries to work out the probability of winning from each board position and When it had realized this it that’s how come it created this whole new play style Where humans had just never realized.
[00:43:09] Blue: Oh, we’ve been using points as a proxy for chances of winning When really we should be worrying about chances of winning directly because that’s a better thing for us to be worrying about This is what led alpha go to introduce a whole new play style Was this realization that humans were making a mistake Every thousands of years of go players were making the same mistake and alpha go wasn’t and this is why it invented this new play style That’s why it seemed so inhumanly like it was making mistakes, but it was actually winning the game this is one of the most interesting parts of The overall game is just the fact that it had discovered this whole new play style Um through the way the machine learning algorithms work And it had come up with this idea. I’m just going to try to win by one point I’m not I’m not going to try to win by as many points as I can Once I’m convinced I can win by one point. That’s it. I’m going to just make sure I win by one point and I’m done And that’s why alpha goes play style seemed so random at times Uh to a human who only is looking at how to you know, shouldn’t it be taking more work position? Shouldn’t we try to score more points and it didn’t care that this is one of the most exciting points to the of the story to me is when they started to realize that That every human on the planet did not understand this aspect of go that alpha go understood Maybe I can make it just an aside here for a second. We talked about in one of the episodes
[00:44:39] Blue: The pseudo -deutsch theory of knowledge this idea that uh, no No ai algorithm. No algorithm that has ever been invented by humans has ever created knowledge before Based on typically there’s different ways that this is explained. It’s explained as Well, because the human actually input all the knowledge And we know the human doesn’t input most of the knowledge. So that’s not entirely false Based also on this idea that the knowledge somehow exists in the data Which to me sounds kind of inductive, but anyhow that this is typically how it’s said Is that the knowledge is in the data and it’s just sort of reorganizing the knowledge into a useful format But that the knowledge all came from the program Now I’m not saying alpha go refutes that because to be honest This is an irrefutable theory and which is the problem the theory can’t be falsified in any way Which is why it’s not a good explanation But I think the alpha go playing Lisa doll Certainly stretch shows how far you have to stretch this theory to take it seriously To where not a single human understood what alpha go learned by playing itself or By using this algorithm not a single human had ever actually realized.
[00:45:57] Blue: Oh, you shouldn’t be using Points as a proxy for chances of winning You should be paying attention to winning by at least one point The fact that alpha go really and truly did come up with a brand new creative move move 37 You know that it created a new play style that All of these things that something so creative that the world champion was the only one who could recognize the beauty of the move Right it required a world champion to understand how beautiful the move was because it was so creative And it was something so far outside Anything any human had done it wasn’t part of the data that was fed into it from human players because human players don’t make this move You could still of course say oh well, but the knowledge actually came from The algorithm and humans inserted that you can always say that that’s the problem with the pseudodorch theory of knowledge Is that it can be used for any circumstance like this? It can apply to any circumstance it could never ever be falsified or refuted But this should show just how far we’re stretching to make the claim that uh machine learning creates no knowledge I would also point out That we could use the pseudodorch theory of knowledge to claim that humans don’t create knowledge we could say oh all knowledge actually comes from from uh biological evolution And all the knowledge that humans display actually comes from that knowledge created in their genes And I mean of course, this is a ridiculous theory.
[00:47:23] Blue: I mean, of course, it’s a ridiculous theory and then uh humans just take observations and You know using induction, of course Take observations and the knowledge is already in the observations and they they use the these algorithms that that are built into their head by Uh the knowledge that came from the genes And so all the knowledge actually comes from the genes and humans generating no knowledge at all And the pseudodorch theory of knowledge could be used for to prove this also Which is the problem that it can it over explains it explains every single possible outcome There’s no outcome that exists that it can’t explain That’s not what we want in a good theory what we want in a good theory is one that can be falsified That doesn’t explain every single possible outcome I do think though that alphago is one of the strongest challenges where intuitively You look at what happened go watch the movie And then really try to try to hold on to the the pseudodorch theory of knowledge And I think you’ll find that it’s starting to crack around the edges for you right is It just doesn’t make sense that there’s no knowledge creation that took place here Because the knowledge that it’s showing none of the humans knew about it.
[00:48:32] Blue: It just it didn’t exist in any of the human heads It was something there was great knowledge that came from the humans But it went beyond that it went beyond that and generated some of its own knowledge That allowed it to be the world champion that allowed it to be a new kind of Playstyle that no human had seen before this is how this this documentary ties into some of our past podcasts and kind of the importance of really understanding That machine learning does have a form of knowledge creation That’s involved with it that ai is creating knowledge in a sort of narrow setting now It doesn’t create knowledge beyond It’s narrow setting alphago isn’t going to suddenly learn You know how to make pancakes tomorrow, right? It’s it’s stuck to Whatever its specific knowledge space is that it is trained to do It’s narrow ai. It’s no different than any other narrow ai in in that regard But within that narrow domain Machine learning comes up with things that no human being has seen before And that no new human being knows how to do and that’s really the point that I kind of wanted to make to particularly the deutching community that often has embraced the pseudo -deutch theory of knowledge That this is something that needs a stronger look. This is something where we really need preparing is looking at it saying Okay, what’s really going on with machine learning? What is it doing? You know that it’s able to come up with New things like this totally creative new things like this Um that no human new prior to this point. All right off my show box on that one
[00:50:06] Green: And so I’m I’m curious because you said that this is a simplistic ai What you know, I think it’s a pretty big moment to see This machine have um an intuitive leap Where it does something that nobody had taught it And it figures out something we’re not capable of figuring out or or I couldn’t imagine Our intuition didn’t hadn’t come up with Yes, what does it mean and or maybe it doesn’t mean anything but you know looking at general artificial intelligence Do we see this as as a leap forward on our understanding of Machine learning or our understanding of our ability to create an artificial general
[00:50:57] Blue: intelligence? All right, very good question There are In some ways and we’ll do a podcast on this at some point In some ways what you’re asking is what is it that we’re missing about general intelligence? And we don’t know what we’re missing right? I mean like it’s a mystery What it is that we just we’re not even asking the right questions at this point When we try to come up with an intelligent machine to go play alpha go We do it in a certain way.
[00:51:26] Blue: We we figure out, okay We’re going to have this state space and we’re going to have it, you know we’re going to use reinforcement learning And there’s not even really an attempt to make it generalized to do anything Some people will tell you reinforcement learning is a general learning algorithm That’s only true if you don’t take into consideration That a human has to go and put the state space the world space You have to go teach it what the world space is for each problem that you want it to solve It never learns to solve a different problem except for the one it’s been programmed to solve The rest of the algorithm generalizes right is think of it like a module you have to plug in the The reinforcement learning algorithm will learn any World space it’s given but you have to first give it to it You have to first say here’s what how to represent You know every single board position for alpha go here is how you understand what a reward is And it has to have some sort of input from the real world That’s a reward that has to be explained to it has to be programmed not explain the program And then from that point forward it will then play billions of games with itself More games than any humans ever played and it will it will figure out Which board positions are the better ones?
[00:52:43] Blue: Once it knows which board positions are the better ones then you think about it All it has to do is it has to try each move and say is this move going to give me a better board position Than this move or this move or this move The knowledge gets caught up in its board evaluation algorithm. That’s really where it’s its true knowledge exists Okay, it generates this board evaluation algorithm It’s the board evaluation algorithm That is the function that no human has ever seen before that it has come up with And its board evaluation algorithm was so good that looking ahead only one move It could play at professional levels if you let it look ahead more than one move than it was even better And presumably to play at least at all you’re not looking ahead one move, right? You’re looking ahead as many moves as you can And then think about the fact that looking ahead trying out moves that itself is a form of knowledge creation Because you’re trying out each of the moves you figuring out which move is the best one And then you’re using the board evaluation algorithm to tell you as proxy for that’s the best move So the combination of those two algorithms Generate knowledge now. Here’s the interesting thing think about like
[00:53:49] Blue: IBM playing Casper off when big blue beat Casper off Big blue did not create a new play style big blue did not create A brand new sorts of moves that no one had ever seen before people said wow that was a creative move Now there’s a good reason why it didn’t it’s because all the knowledge was caught up in Trying to look ahead as many moves as it could which is just too limited Okay to create a whole new creative play style Alpha go learned its creative play style through its board evaluation algorithm in big blue deep blue. I mean sorry The board evaluation algorithm was done by a human. It was just human knowledge inserted into the computer program with Alpha go that board evaluation algorithm was created by The learning algorithm the reinforcement learning algorithm combined with the deep learning algorithm And so the net result was that it had come up with a new way to think about the game And a new way to understand the game and new creative sorts of moves Okay, that’s because the main ax algorithms just too weak. Yes, it creates knowledge But the knowledge it creates is very local. It just simply says given this board position What’s my next best move looking out seven moves or however moves it can it can figure out It’s more complicated than that. It tries to call out moves that aren’t good And it tries to look ahead 13 moves maybe or 20 moves on the few most interesting possibilities or something like that. Okay with Alpha go it actually plays billions of games And then it figures out based on those games Which board positions are the most likely to lead to a win?
[00:55:27] Blue: Okay, that’s the way reinforcement learning works That was why alpha go Came up with new creative play styles and beautiful new moves and things like that. Whereas deep blue Never did it just simply played chess well. That is the difference between AI regular AI good old fashioned AI and machine learning Where machine learning actually comes up with creative new algorithms functions I should say functions and algorithms can be the same thing It comes up with creative new functions that no human has ever seen before Whereas with AI it’s just simply trying to find the best move out of the possible moves at this position. Does that make sense?
[00:56:06] Green: Yeah, yeah, that totally makes sense But what does it mean for the future of of
[00:56:12] Blue: agi? so I think every approach to AI that we’re currently doing is the long approach for agi. They’re all interesting in their own right Okay, I mean the fact that we can make up These automated algorithms that can generate new new sorts of creativity right new sorts of creative moves things like that Within narrow domains. That’s not a bad thing. Like when we actually create agis, they’re going to be a lot like us They’re going to be people We aren’t going to want to enslave them. I mean that would be bad To enslave a person to then be a chess player only or something like that, right? If you want to make a really good chess player You what you really want is you want a narrow ai to do it and that’s true of you know running Manufacturing plans. I mean most of what we’re going to want to do through automation Is going to be narrow ai forever We’re never going to use agi to automate things That wouldn’t make sense because it’s going to be a creative individual. So it both fields Need to be researched for different purposes Both fields are types of knowledge creation. So there is an overlap between them They’re both they’re both different subsets of a single thing which is knowledge creation But they’re very different sorts of knowledge creation. Now. Why are they different? That’s what we’re trying to figure out, right? One of the things that they that they talk about is there’s a gentleman He uh researches what he’s what is called the problem of open -endedness to get his name here Oh, yes, Kenneth O.
[00:57:42] Blue: Stanley And you can go look him up on youtube and he’ll talk about the problem of open -endedness Now if you think about knowledge creation Everything that humans have created that creates knowledge. So basically ai algorithms and machine learning algorithms They’re all narrow Knowledge creation they create knowledge in some little tiny domain that we understood well enough to to explain And then it it goes and it tries variance and it discover states by doing that That’s different though than the other two kinds of knowledge creation that David Deutsch mostly talks about which is Neo Darwinian evolution And human knowledge creation now both of those are in some sense open -ended think about like You know Darwinian evolution and how it creates all sorts of different species Right, it’s kind of oh, there’s not it’s not limited. It’s not Going to try to create the best mouse and that’s it right It will discover all sorts of new species and new creative ways to live in niches that weren’t inhabited before And that’s what neo Darwinian evolution does it’s it’s got this open -endedness to it And we don’t know how to program that never mind agi yet. That’s Related to a different problem We can’t even program Darwinian evolution as an algorithm now. This is uh, This is what um Leslie valiant, I think I mentioned him in past episodes This is one of his areas of research.
[00:59:15] Blue: He’s a very famous guy in artificial intelligence and machine learning and he’s written a number of books and he points out that We think we understand Darwinian evolution and we don’t we don’t really know what the algorithm is that creates the sort of Facts that you see in Darwinian evolution We try to we we have something called genetic programming Which uh is based on uh Darwin’s theory of evolution. It does Cross over it does mutation It does mating it creates these population pools of replicators It does everything that we think we understand about Darwinian evolution And the end result is narrow ai just like every other type of ai we’ve built It’s even though it’s got in theory because it uses a programming language in theory It’s got an open -ended problem search space. It never Uses it it it always just explores a little tiny part of that space that’s super narrow And it produces results nearly identical to You know any other form of narrow ai It’s it’s just not a super creative algorithm And it’s certainly not open -ended in fact it converges It can it’s the way they set it up Usually they’re trying to have it solve a single problem. So it kind of makes sense that it converges It’s part of it’s just how it’s designed But then you could say oh well then don’t have it converge to one thing let it just try to discover You know random solutions. I can say that but I don’t know what that means Right, I don’t even understand what I just said well enough to go make an algorithm out And this is one of the big secrets and I brought this up in our computational
[01:00:53] Blue: theory episodes Is that human beings understand things through algorithms? Okay, and David Doricius puts the mornings on that. He said well, maybe we shouldn’t you know He’s it’s been his experience that it’s a mistake to try to jump to algorithms immediately that we should start by explaining things without algorithms Well, when you think about Darwinian evolution, we’ve gotten quite far with an algorithm with a an explanation of Darwinian evolution that Can’t be fully algorithmatized yet But in so far as we can’t algorithmatize it It shows that we don’t understand parts of it and maybe we think we do I’ve had people where I’ve talked to them and I’ve said Yeah, we don’t fully understand Darwinian evolution. It’s oh, yeah, we do But I’ve had people argue with me over this but the simple truth is we don’t If we actually understood it completely We could make open -ended evolution of virtual species and they believe me they’ve tried to do this Right, they’ve had a certain level of success But it always seems to kind of top off at some point and it just stops creating new things at some point It’s because we’re missing something There’s something about the theory of evolution that we don’t understand and we don’t even understand What is that we don’t understand? So it’s hard to even ask the right question And that’s what kind of Stanley is trying to study is with the the problem of open -endedness one of the things that he’s researched is Trying to come up with creative search. So
[01:02:18] Blue: instead of trying to search for just a solution to the problem So I could let’s say you have a a virtual robot that you’re trying to teach to walk Instead of just trying to directly give it rewards through reinforcement learning for successfully walking He basically just tells it any new state you haven’t reached before you get a reward for and it will Learn to walk faster than if you directly try to teach it to walk by doing that And so that’s like his approach to try to understand The problem of open -endedness and he’s got a number number of other really interesting Experiments that he’s trying to do using software to explore the problem of open -endedness So that’s one problem that we know we don’t understand correctly Valience pointed out that we don’t even understand a tractable version of Darwinian evolution that the versions that we program Would never tractably run Even if they could solve the problem of open -endedness to to be able to create species In a few billion years like the world was able to do so So we’re missing something there and there’s some interesting ideas there that are worthy of some research there was that paper that Thatchapal in one of our episodes talked about where Someone tried to take Leslie valiance algorithmic evolutionary algorithm that he had proposed that didn’t really work Very well and he got it to solve the bitwise problem, which is something that evolutionary algorithms can’t solve By introducing the idea of ecology Into the mix. So maybe that’s one of the things we’re missing, right?
[01:03:55] Blue: Maybe we’re so narrowly Focusing on variation and selection that we’re missing the fact that there’s these aspects of ecology That are also knowledge created that have their own variation and selection That allow it to be able to solve problems that it can’t currently solve that might lead us to an understanding of open -endedness Beyond that Beyond that there’s an additional problem for agi Which is we talk about like science’s explanations, right? And this is something we’ve talked about with our our epistemology and we say, oh, yeah, science is about explaining things This is one of the things that like the instrumentalists have missed which is why they misunderstand science Um that science is really about finding the best explanation It’s about trying to creatively come up with a conjecture that and then criticize that conjecture and then the one that survives That’s the best one. That’s the the one that has the most verisimilitude. It’s the the truest one Um, that all makes sense But can you explain to me algorithmically what an explanation is? Because I don’t think anybody can right now. I sure can’t right it’s Give me a minute We have a we have a um Upcoming episode where I’m going to talk about causal inference One of the things that I found exciting about causal inference Is that it tried to create a mathematical computational, um graph graph model of what an explanation is It’s too primitive though, right?
[01:05:21] Blue: It’s I don’t see how it’s going to successfully create the kind of explanations that exist in science It’s these really primitive explanations There’s a something called explanation based learning which does the same thing and it uses logic To be able to create its explanations, uh propositional logic or first order logic And that kind of makes sense popper His epistemology if you go read his book, he’s got several things that we call books most of them are just collections of Things that he wrote sometime. He has one actual book called the logic of scientific discovery He his epistemology is based entirely on propositional and first order logic And so it kind of makes sense that an explanation. He models explanations as logic statements Now David Dwayche has criticized that a little and said that that’s only an approximation of an explanation Well, what does that mean? Like if if logic statements are only an approximation of an explanation, then what is an explanation? Right, it’s how would you model that inside of a computer? Well, nobody knows Right, I mean it’s we’ve got a couple different paths explanation based learning causal inference where we’re we’re trying to come up with something Right now they seem kind of primitive Is that possibly one of the things that we’re missing for agi? Maybe right? I don’t know another one is How do we make conjectures? What’s the conjecture engine popper basically treats conjectures as a mystery? He basically says you just come creatively come up with your best conjecture and then from there you criticize it His epistemology Doesn’t explain how to come up with the conjecture. It explains what’s to do what to do once you have one Well, how do you actually come up with one?
[01:07:00] Blue: Well, if I knew that I probably would have solved the problem of agi That’s something missing and again What does it even mean to creatively come up with the conjecture? Right, could could you give me an algorithm that explains what what those words mean? Right, they’re they’re actually very vague and you may not think of them as vague But the moment you try to put them into an algorithm, you’ll start to realize just how vague they actually are And this is why I really favor computational theory approach to things There is value in creating theories that we don’t know how to turn into an algorithm And uh, that that usually is the first step you have to first create the theory at the level of linguistics But there’s these vaguenesses that exist in such a theory That you may not even be aware of there until you try to put them into an algorithm And the moment you try to put them into an algorithm you start to realize Oh, wow, there are like massive gaps in my knowledge that I didn’t even realize I have Until I just tried to get specific and put it into an algorithm and and I think that all of these are What we’re missing, right?
[01:08:04] Blue: We need to Explore out a lot of these ideas better and then we need to make an a failed attempt To put them into algorithms and then figure out what is it that I’m missing by failing This is something popper brings up over and over again It says the way you actually solve a problem is by trying to solve it and failing and that helps you understand the problem And once you understand the problem really well, that’s when you actually are in a position You try to actually solve it for real You so in other words You have to go try to fail to solve a problem to educate yourself on the problem to the point Where you have any chance of solving the problem that makes any sense
[01:08:40] Green: Not only does it make sense, but I like we could almost have an episode just on that. That’s a place. I’ve been Kind of thinking deeply about, um, you know, just in my job You see a lot of organizations trying to become More lean or more agile, but ultimately the goal isn’t to isn’t to be agile It’s to figure out how to effectively fail with within Organizational constraints, you know and and traditionally businesses don’t like the concept of failure and so part of what I think is you see kind of happening in in a lot of especially in software development is How can we organizationally get to a place where we are more comfortable with the concept of failing because it’s The only actual way we can learn So I’m just very interested in that in that particular concept right now That is something that follows directly from poppers epistemology What you really want is you want to be able to fail and not cause problems because then you can learn faster There are some people who consider themselves to be critical rationalists Who really just don’t get this fact and they’ve got these concepts of You shouldn’t you know overreach Because then you’ll make lots of mistakes and that’s failure and failure is bad.
[01:09:58] Blue: It’s like wow You call yourself a critical rationalist. How have you so severely misunderstood the epistemology? What you really want is Sometimes failure is bad, right? That’s why people hate it But what you really want is you want a situation where failure is not so bad and that you can figure out how to Let the failures through not Cause your whole organization to fall apart over it. You know, someone is really good at this is amazon Amazon has an enormous number of failures
[01:10:26] Green: Sure, and and they’re very very good at failing learning from the failure And then moving on and using that failure as a jumping point toward a Toward a a future success Yes
[01:10:41] Blue: And if you just if you just think about it from just a variation in selection, which that’s knowledge creation, right? standpoint if you go out and you try, you know, 20 different ideas And 19 of them fail But one of them’s a massive hit It pays for the failures Right. I mean like you you’re going to find massive hits By trying things and failing and then by chance you get one that’s really good See
[01:11:10] Red: I don’t even if it seems like failure is actually just conformational.
[01:11:14] Blue: That’s it. We’re probably coming up on time here So any other final thoughts on alpha go the movie?
[01:11:20] Green: Well, hopefully nobody ruined it for themselves but even if they hadn’t watched it I still recommend going and watching it because um, I think the impact of seeing like The way lisa doll is is responding to this and the way that everybody’s constantly kind of surprised by what’s going on is hard to Convey here and and I just it’s just a really enjoyable Show it just is really enjoyable
[01:11:47] Blue: I agree me me courting them just doesn’t do it justice You have to kind of live it and that’s what the movie lets you do you have to really see Wow, something something really creative is going on here Right that the algorithm is doing something creative and it’s coming up with things that the humans just don’t get Only within its own little narrow area. We’re still talking narrow AI for sure, right? This is not a conscious agi we’re talking about at all, but uh um, there’s something just interesting in and of itself About machine learning machine learning is an interesting topic all its own Regardless of whether it’s a path to agi or not which it’s not not the current form anyhow Yeah, definitely would encourage people to take a look at alpha go the movie and Experience this for themselves really be there and watch What’s happening and how it unfolds and how it feels to the people that are involved um all right Well, thank you guys. This has been uh entertaining to talk about this movie with you It’s one of my favorite movies. Honestly. Um, I recommend it to anybody So thanks everybody. Thank you.
[01:12:55] Red: Thanks everyone
[01:12:59] Blue: After I finished recording this episode on alpha go David doge posted a video on his youtube channel called poppers problem oriented epistemology with david doge and elie tear And in the video he actually talks with elie about alpha go briefly Eli makes the claim that alpha go creates knowledge and david doge admits that maybe it does Although he still seems somewhat skeptical of that possibility But he he mentions I think this is correct that if it does create knowledge It creates it in the same way that biological evolution does not through explanations I don’t think there’s any doubt about that at all the type of knowledge that is created by a program like alpha go Is not explanatory knowledge. So I was really glad to see david doge finally address alpha go Directly like this and I was glad to see that he’s at least open to the possibility that alpha go does create knowledge This is a good example of what I was trying to get at within the podcast That alpha go does present a problem for the pseudo doge theory of knowledge and that was really my point here Since the pseudo doge theory of knowledge is irrefutable. It could be applied to alpha go very easily It could be applied to literally anything very easily But I think emotionally it becomes far more difficult to apply it to something like alpha go Where it’s clearly come up with this entirely new creative play style That’s never been seen before the history of the world in any case.
[01:14:13] Blue: This is the point I was really trying to make I really wasn’t trying to go beyond that point But I was glad to see that david doge could see that there was a problem here and is starting to open his mind to the possibility That machine learning algorithms do create knowledge The theory of anything podcast could use your help We have a small but loyal audience and we’d like to get the word out about the podcast to others So others can enjoy it as well to the best of our knowledge We’re the only podcast that covers all four strands of david doge’s philosophy as well as other interesting subjects If you’re enjoying this podcast, please give us a five star rating on apple podcast This can usually be done right inside your podcast player Or you can google the theory of anything podcast apple or something like that Some players have their own rating system and giving us a five star rating on any rating system would be helpful If you enjoy a particular episode, please consider tweeting about us or linking to us on facebook or other social media to help get the word out If you are interested in financially supporting the podcast We have two ways to do that the first is via our podcast host site anchor Just go to anchor.fm slash four dash strands f o u r dash s t r a n d s There’s a support button available that allows you to do reoccurring donations If you want to make a one -time donation go to our blog, which is four strands org There is a donation button there that uses paypal. Thank you You
Links to this episode: Spotify / Apple Podcasts
Generated with AI using PodcastTranscriptor. Unofficial AI-generated transcripts. These may contain mistakes; please verify against the actual podcast.