0016: More on AlphaGo

This is Ted. And this is Randy. Hey, welcome back, Randy. Oh, it's so good to be back.

Glad to be healthy in here. And here we are on Concerning AI, which is a show. It's a show. And what is this?

Episode 16? Episode 16, which is pretty cool. We're having so much fun doing this thing. It is pretty cool.

What is this thing? Why are we doing this? Why are we doing show about Concerning AI? What is 16 in two to the fourth?

Two to the fourth. Two to the fourth. Two to the fourth. So why are we doing this?

Yeah, why are we doing this? Are we concerned? Yeah, it's funny how that goes through cycles for me. At the moment, I'm mostly interested.

Yeah, I am too. I'm having so much fun right now. Yeah, but sometimes I'm panicking, this kind of concern and stuff. And I wrote a really scary short story here.

At least it was scary for me, recently, that was about AI and the choice and stuff that people might have. Can we put that in the show notes? Is it available yet? Are you published?

Oh, I'm not. Is it a secret? It's a little weird. So I don't know that I want to add her.

We'll talk about it offline. If Brandon decides that he's ready, then we'll put in the show notes. And you guys can read it. Yeah, I don't want to offend any of our listeners either, because it's got some references and stuff.

So yeah, so big news. Let's hit the big news again. Yeah, so let's hit the big news. The big news is AlphaGo.

He mind-blowing to me. Honestly, he did not expect it. Yeah, so just as soon. I just don't have this.

If anybody doesn't know. So AlphaGo is a Go playing computer made by DeepMind, which is owned by Google. Yeah, from Cambridge, from Cambridge. And everybody seemed to think that Go was a game that would be at least a decade off that the professional Go players would start losing to computers.

This is a game that computers have been playing against people for a long time. Yeah, I've played a lot of computer players. I've played a lot of computer players. I've beaten a lot of computer players.

Speaking of which, Go is a game we both love. We do. I'll be honest, it's my favorite game. It's my favorite game as well.

I feel like I'm very bad at the game of Go. I'm just good enough at Go to appreciate how bad Go I am. Yeah, great players like Lisa at all. Yeah, yeah.

I can't fully appreciate how great it is, but I can appreciate that there is something great there to appreciate. Yeah, yeah. I was really struck. So let's back up.

So I want to express more appreciation when we get to that. OK, just the back up. So the game of Go is a search space that's really big. Sometimes people say there are as many games of Go as there are atoms in the universe.

But I did a little math. And it turns out that if there was a parallel universe for every atom in our universe, then the game of Go would have as many Go boards as all the atoms in all of those universes. It's just mind-boggling. Yeah, it really is.

It's 10 to 170th or something, which is just ridiculous. So it's such a big number that we're talking to all the elementary particles in the universe. Yeah, it's just a ridiculously huge search space. It's not enough.

No, right. It's not even close. I can't get to you about it because humans are terrible at big numbers like that. Yeah, absolutely.

And so I went off a go, it could really. Intuitively grasp that. Well, that's right. And here's the thing.

Everybody thought that Go was at least another decade off before computers would get any good at it. And then all of a sudden, in January, there came out of paper in nature that AlphaGo had beaten a top European professional last fall, five games to zero. It was shocking. Totally shocking.

I think it was the top European professional. But that's not saying that much because Europeans are so far behind Asia. That's right. Yeah, I mean, it's not saying nothing.

I mean, I'm going to be wrong. But this is literally the first professional to have been ever beaten by a computer. By a computer. Not the top professional.

It was regardless. It was big news. And then AlphaGo, the deep line team set up a match, a five game match, between AlphaGo and LisaGo, LisaGo being one of the top professionals in the world. I think that he would be the consensus best player in the last decade.

I think that that's true. He's not ranked number one right now. But that's why I say he's not the best right now. He's an 18 year old kid that's beaten LisaGo eight on the last 10 times.

So it's, you know, he might be, he's probably on his way down. But he's, you know, he's Michael Jordan year or two after the sixth champion. Right. Exactly.

So he's nothing to sneeze at. And no professional would willingly want to play. And LisaGo thinking that he's an easy win. Right.

Not true for anybody. So they, Google put up a million dollar prize and said a million dollars to the winner. He said, oh, you will get a million dollars if you win. And if we win Google wins, then we will go to the charity.

And LisaGo was pretty confident a couple weeks before, predicting a five zero human win. He picked up the he would win. And it's good. It's good to be confident.

It is good to be confident. Yeah. So it's like he was pretty crashed after that first game. After the first game, he, because he didn't play his very best.

I don't think he did. Yeah, he did. I don't think he did. There were some really interesting things that happened in that first game.

There were moves that the human commentators watching the game said immediately, oh, that's a mistake. And then 20 moves later, they were scrambling to explain how brilliant that move was, because they now saw it. Ah, that's just, and they all started speaking with such respect that they didn't have at the beginning of the match. And they didn't have when these moves were made.

And yeah, so LisaGo lost that first match. And he was very upset. He's a professional. He pulled himself together.

And he was going to learn to do better in the second game. In the second game, he lost again. And he lost the reader out. Yeah.

But between the second and third game, they were headed day off. So he stayed up until six in the morning with his friends studying these first two games. He got a new plan for game three. Then in game three, he lost again.

And then in game four, have you ever read the Hikaru and Nogo series of manga? Yeah. In game four, he played the divine move. Yes.

In game four, I was watching this live when it happened. I've never watched Go Live before. And I was watching this live. And when the move was played, there was a pretty long pause before the move was played, where the commentator was showing all the different moves that could happen.

It was a very complicated position. Didn't mention this move at all. Yeah. And then I was playing the commentator kind of stopped and looked and really was like, I didn't see that coming.

The commentator, by the way, is a nine-dawn professional. This is Michael Redmond. Yeah. He's the top American, North American born, Go professional in the world.

The only one ever to become nine-dawn from North America. He's a Japanese professional. He's a Japanese professional there. So he's nothing to sneeze at either.

No, he's nowhere near least at all. But he's an amazing commentator. He's definitely the best commentator I've ever seen. And it was amazing to see him watch that move and just be in awe and try to understand it.

And then he played out a whole bunch of variations. And he saw that move, that move was brilliant. The divine move could be called the divine move. I saw some commentary later where professionals were calling it.

And they didn't call the divine move. They called it the God-zone move or something like that. Yeah. And it turns out, AlphaGo only gave that move a 1 in 10,000 chance of being played.

And so AlphaGo didn't consider it ramifications of it very far. It didn't keep me right. That may be a gap in the programming. Maybe it needs to be tweaked so that when it comes up with an unusual move against it, it should go deeper and take more time.

It didn't do a very good clock management. It played its next move pretty much the same speed it normally plays. So that might be just a little flaw. But the fact is, it didn't know what to do because it hadn't gone that far down that branch of its search tree.

And Lisa Dolannegan. And so people thought, well, maybe Lisa Dolanne now has learned enough about this computer, learned enough about AlphaGo's style that he gave five, maybe he's got its number. Maybe he's got it. Game five went the longest.

It took the longest to get to a result. And ultimately, Lisa Dolan did resign. So not very much of an move from the end where you would count the board. When you and I play, we always have to count the board because we don't know who the head is behind.

But Lisa don't know. If he's behind by three points, there's no reason to play on when you get to a certain point. Yes. AlphaGo's not going to make it a mistake.

Not in the end game. So it was a fascinating match. The reason why I thought Go was safe for a while is because of how humans play it, like the best humans in the world play. I mean, it seems like it gets right at the core of being human.

It's very intuitive. It's very aesthetic. These moves are beautiful. That's why I play that move is because it's beautiful.

I've heard it described as sort of the intersection between math and poetry in a way. It's nice. It's an intellectual part of understanding why certain moves are good. I mean, the element of it that the computer is going to kick our behinds out every time.

What is the score at the moment? Some of those things. Yeah. Yeah.

I just didn't expect it. It definitely feels a lot closer to general intelligence than chess does to me. Yeah. In chess, it appears that no matter what's going on on the board, humans can look at the board and they will know who's ahead or who's behind.

There are proxy heuristics even for bad players just counting in the board. Yeah. Pieces on the board is not a bad proxy for who's ahead or who's behind. In Go, it's very difficult to know in a lot of board positions.

It's very difficult to count. And this was, I thought in evidence because there were nine Don professionals commentating in one channel and nine Don professionals commentating in a different channel on the same game and they differed about who was ahead in different points, which was fascinating, especially game five, which was very close. And Michael Redmond thought that Lisa had done it had won again and he was wrong. And some other commentators were saying that they thought I'll go with the head.

We're just not that good at counting. We're not that good at knowing what's going to happen. Yeah. So here we have a beautiful game, a game that does not succumb to brute force.

It's not like it's impossible to beat this game based on brute force searches. It's impossible to do what we did to chess. Even if we waited 20 more years of Moore's law or whether the success rate Moore's law is impossible forever to do it. Forever.

Right. And so we've talked on this show less for some reason quantum computers are able to make equal NP. That seems unlikely. Yeah.

So we've talked on the show. We've had a little bit skepticism whether software was getting any better or whether we were seeing gains in artificial intelligence just because of the advances in hardware. And I think this has to be algorithmic. This is a win for the algorithm.

This is, yeah. I mean, I still think, you know, I've been thinking more about the stool with three legs and there's the algorithmic, you know, what are our algorithms? Do we have training data and do we have the computational resources we need to support those algorithms? Definitely the theoretical algorithms leg of the stool here.

Yeah. Yeah. So how did this come about? How did AlphaGo, like, let's talk about that a little bit?

Sure. Well, I think it's interesting because, you know, there was a big some crawling, I think, from Facebook not too long ago about how their Go player was going to do some amazing things and it turned out to be not that good. Yeah. And it's also using, you know, these deep neural networks.

So Facebook Go program is not as good as the two other best Go playing programs in the world before AlphaGo. Yeah. So it could be that in a month or three months it will be, but it's not yet. Well, that, I mean, I think that's one of the things that the humans working on AlphaGo did well is they figured out how to structure the learning problem in a way that the computer could actually make progress playing against itself.

And that's not, that's often the case in, okay, so I would say there are some meta parameters in the learning apparatus that the, that the Facebook people got right. And the, I mean, sorry, that the Google people, the deep mind people got right, but the Facebook people did not. What do you mean by meta parameters? Well, it's easier for me to illustrate it in the realm of like genetic algorithms.

Okay. So genetic algorithms, you have a population of these computer programs that are just basically a string that says what the program is, how it works. And then you have a fitness function and you take that population, you do evolution with it. You say, all right, these ones end up with the best score against my fitness function.

So I'm going to let them breed and I'm going to make some random mutations to that, to that encoding of the program. Right. And that will make a new population, we'll run it against the task and score them all and do the same thing. Well, in that realm, you think, oh, that's very general, that should work easily.

But the magic is in how you actually encode the problem. And a lot of times when you encode a, like, like what a, what a complete machine looks like, a lot of times what you end up with is very quickly you hit a local optima or, yeah, it's, you just don't, you don't make very much progress because the encoding, there's a lot of structure and domain knowledge in code, goes into specifying how that encoding works. And does it help for go have that kind of specific information in it? So an example here is they ended up choosing the value network and the policy network and, you know, how they structured their reinforcement learning.

You know, those are some choices that are not obvious that you would make those choices. So one thing you could do, if you're going to teach a program how to play chess, you might teach it all the known openings and you might teach it what to do about those. Right. And then there are other ways that you could go about teaching a computer, but that's one way you could, you could teach it specific oristics.

These kinds of moves are good. These kinds of moves are how you, you can have a, a good of a chess expert, be one of the teachers to a program. But AlphaGo didn't have it to go expert on the team at all. And what you said there though is, is actually a tree.

Like, like, what do you mean teach it? How would you teach it? I don't know. You could, you could make decision trees for it.

If you see exactly this, then you're going to do this next. Which is part of how humans learn to play chess. If they study openings, they could go spend studying openings. You could, you could build in some decision trees that way where they're very, they're hard coded.

I wouldn't call that so much learning as it is just setting out recipes. Well, okay, so in Go, there are similar things that human players study that could, you could hard code in a program. You could hard code certain things. But that, I mean, if you could actually teach it, those openings and stuff, you would, you would be doing an amazing thing with a computer program.

Right. So, in AlphaGo's case though, they didn't teach, they didn't actually teach it anything. Deep mind, deep mind is this company. Before they did Go, they took on all the old guitar games.

Yeah, with the pixels. With the pixels. So all they gave it was numbers for each of the pixels on the screen. Because I don't know if it did visual recognition that it had to do itself.

Or if they just gave it the numbers. Oh, it's pixel. It doesn't matter. That name would just give it the numbers.

They just gave it the numbers. But they didn't tell it how to play the game or didn't tell it what the game even was. They just knew what the score was at the end of the game. And so if it was playing Pong or whatever, then it just saw the numbers.

And then it figured out itself, even though what it's control was supposed to do. When I move the controller, what pixels on the screen move? It had to figure that out on the phone. And within a matter of hours, it could be any game, any of those are the two games.

They just did one after another after another after another. And then it would learn how to play them all by itself. So, anyway, they didn't teach you how to play any of these things. Now, this is a different kind of game than go because there's no future stake that you have to plan for.

Yeah, but again, there, like how you actually encode the game and how you keep track of what time means, like one step forward, how you slice it and how you feed the information back. All of that, I would say, count as meta parameters. I'm not sure I'm using that term, right? But there are things that it's not clear how to choose those.

Yeah, there's still some like black magic in choosing those things. When a computer can learn those things, that would be really exciting. You can learn the meta parameter part. Yeah.

You can actually learn how to structure it. So, let's talk about that a little bit. So, we've sort of talked about artificial general intelligence. And then we've talked about artificial narrow intelligence.

Is that what we call it? Yeah, yeah, that seems right. And so, people call it weak guy and strong guy. And we maybe two of those, they were sure the same thing.

I kind of think so, yeah. So, every AI that we have today on the planet is a weak AI. It's an artificial narrow intelligence. It does one thing.

It does it well. The Google search algorithm is an example. It's an AI. It does one thing.

Alpha goes and other examples. Self driving cars are another example. But you're sort of pointing ahead to in the direction of a general intelligence. If it can somehow help structure itself, that seems to be a whole different conversation.

Yeah, I don't know how that would work. But it seems a lot closer than it ever has to me right now. Yeah. And why does it seem closer as a result of AlphaGo?

Well, it's not just AlphaGo. It's that we've gotten so much better at identifying where all of our heuristic hacks sort of fit in the general. Okay, it's because we're understanding the general learning case better and better. So we can go back and look at something that we did before that worked in a particular narrow domain and reframe it.

Use different words, sort of the standard machine learning vocabulary that's come up. The standard machine learning models and say how this differs from this other thing that we did over here in this domain. So the fact that we can now compare what we're doing here, I guess what we're doing there, all within a single model makes me think that those meta parameters, those differences between those two cases are things that we're going to be able to learn eventually. Okay, so when I said we didn't learn that much from Deep Blue, which was the chess playing computer, it's because that computer was so, that algorithm was so chess dependent that it wasn't generalizable.

And it feels to me, like, correct me if I'm wrong, but that AlphaGo's methods are generalizable. And DeepMind is very clear. That's what they want to do. The reason that they built AlphaGo at all was not to be a gimmick.

It was because they wanted to learn some things because they want to solve intelligence, that's their vision, to solve intelligence and then use that to solve everything else. That's what they're here on the planet to do as a company. And so AlphaGo doesn't have any go specific go knowledge at all, but it has some methodology. And then what they're going to do now is they're going to figure out how can we transfer that methodology into other problems that are interesting that are not so constrained.

Actually, even though it has a huge search base, it's very constrained. There's a winner and a loser and there are very specific rules. And it's turn based. There's perfect information at every.

Exactly. At every. So that's different than, you know, put AlphaGo in a hospital and have it try to do diagnosis. It's a whole different thing.

But could we use these methods to assist doctors with diagnosing? Yeah, no question. And that is pretty darn exciting. Yeah, it is just no question that that's on the horizon.

Now, we're still early. We're still very, very early. One of the reporters in one of the press conferences, I think it might have been after game four, asked the AlphaGo team, all right, so we saw the commentators say that was a mistake. And then 20 minutes later, say, oh, that was brilliant.

Sorry, I was wrong. And then we saw in game four actual mistakes. Right. Yeah.

And so AlphaGo can play both of those kinds of moves. Move the human state curve mistakes, but then learn that are brilliant and moves their actual mistakes. So what happens when we put it in a medical setting and humans have come to trust the algorithm to do things that look like a mistake to the humans, but we trust it anyway, because we've seen good results in the past. How do we prevent that from being a disaster?

And I thought the answer was brilliant. The answer was AlphaGo is a prototype. We're not at a beta, ready to release next version. We're not even at Alpha.

We're at a prototype. We're very, very early. We really are a long way from feeling anywhere new. Well, first of all, we haven't even done anything in the medical field yet, but even if we were, we're just a long way from that.

That was the answer. Which doesn't answer the question of, all right, when you're at a beta level, how do we have assurances that we haven't been lulled into a false sense of security? I mean, but that's the thing is Alpha Medical or whatever, you know, it's all good make a ton of mistakes and still be way better than the humans are, because humans make human doctors and the medical solution, it makes so many mistakes. Just huge, what they're called, iatrogenic conditions.

Those caused by the mistakes of the medical community. It's staggering how huge that is. So yes, it'll make mistakes, but we can know very clearly whether it makes more mistakes or less mistakes. So last week, Eric and I talked about self-driving cars and how the statistics so far are that they're way better drivers than human beings and they're accidentally way lower.

And we talked about how the social aspect of this is actually probably way more important than the math aspect of it. The math aspect of it might tell us we ought to have self-driving cars for everybody right now, but that when a child is killed by a self-driving car, the social aspect of it will trump the math aspect of it, likely, for a while. Yeah, that seems totally right. If we're used to 20,000 or 30,000 deaths in the US from cars every year, but self-driving cars would result in 10, well, those 10 parents are not going to be happy about that because this fallacy of everybody thinks they're a better driver than the average driver.

I wonder, are doctors more scientists more scientific? Well, do they see the error of their own ways? The doctors are already using technology, this generation are using technology so much more already and are already moving into that sort of supervisory role. So you don't think that social aspect of it is going to be as much of a problem as it might be?

I don't. I think the bigger issue is they are just really slow moving hospitals and health systems and stuff. There's a huge amount of inertia and a lot of regulation and actually moving things is really hard and slow. So I think it will take a while not because technology couldn't do it but because there's just slow to move.

It makes me wonder if we're not going to have medical breakthroughs happening in countries without strong regulatory frameworks. Well, I wouldn't call this a country without strong regulatory frameworks but they are just supposedly, and this source may not be that great. So don't hold me to this. Moving to some human trials in Australia of these chemicals that basically put humans into the hibernation in order to save people who are having acute heart attacks and other things like that in order to give them time to fix things up.

I saw these headphones and they scared me so I didn't read articles. But that's a case of where it might be that it's easier to do that on humans in Australia than it is here. So I think that speaks to what you're hinting at that it may be in developing countries that have a little less of the safety conscious risk-averse nature of the United States, for example. We could even say they have more to gain if their life expectancy is lower in other places.

Maybe those risks seem more appropriate there. Who knows? There's lots of reasons. I mean diversity.

This is a question of diversity. If we've got a 206 countries in the world, some of them want to go slower. Well, that's okay. Some of them will want to go faster.

And that result is likely the technology will not be held back by the social pressures of one or even a huge collection of countries because there's somewhere that technology will get a chance to try it out. Yeah. Yeah. I mean, it's absolutely coming for diagnosis.

No question. There's already way too much information that comes into the chart, all the electronic records. It's a perfect big data problem with all the electronic health records. We have just wonderful, wonderful data available to figure out all kinds of things that people would never notice just through unsupervised learning techniques.

Yeah. I mean, yeah. There's a renaissance just right around the corner here. It's exciting time.

Yeah. Definitely. Definitely in the next decade, we're going to see some big moves there, I think. But one of the things that Eric and I talked about last week and I love your take on this is we've been thinking about, well, how soon is soon?

And this thing called the singularity or the, we're going to get to artificial general intelligence and we're going to have super intelligence. It seems like before too long. Some people say, well, long time from now and they're talking about 30, 40, 50 years. We had an advance with AlphaGo that seems like a decade advance in a course of five months.

Yeah. That may be an accurate assessment, but it seems like that was the consensus that we were a decade away. And there was this breakthrough. You and I have also talked about, we don't know what breakthroughs are needed.

And it's impossible to predict when those breakthroughs are coming. But what if it's 30 years away and we have three decade breakthroughs in a row? Could we be to AGI in five years or 10 years? Well, I think it's clear that we have enough training data.

Yeah. It's really easy to get training data. Yeah. So I don't think that is a limiting factor.

It's not clear to me whether or not we have the computational substrate that we need. Or whether we've got cycles. Yeah. That's not clear to me.

So I think we may be limited by that for another 15 years or so. But then at that point, then at that point, speaking of that point, did you see the news this week about the quantum computing breakthrough? I did not. I don't know if there are five atoms that were able to compute three times five is 15 or something like that.

I can't remember what the problem was trivial. But the fact that five atoms were able to do the computation, I don't understand what that means. It's mind blowing. Quantum computing is, wow, quantum computing could change everything forever.

I mean, that could completely just knock the doors off of everything. If we can use quantum computers to basically do P equals NP, this non-deterministic, the fact, like the search. If you could search every possible path at the same time in these exponential spaces like that. Wow.

So one of the things I'm noticing now is we're sitting and talking about this. At times we've been more concerned. At times we've been more excited. I'm so excited right now.

And I feel like I'm excited because there's some current breakthroughs that were not predicted. And I'm excited about each of those individual breakthroughs. I'm excited about what these things could do for my life right now. I'm not thinking about the existential issues in this month.

I'm not grounded in the existential issues. I'm not thinking about 30 years from now. I'm not thinking about my kids' kids. I'm not thinking about AGI.

I'm thinking about each algorithm and what it might mean for us as a species, for a shorter term. And I wonder if that dichotomy maybe defines this problem. That we get excited about the breakthroughs now, but that doesn't allow us to see what are the possible unintended consequences down the road. Yeah, I'm just not sure what we can do about them.

Because the genie is out of the bottle. I just feel like it's so inevitable. The only thing I can imagine doing is somehow if you're one of the people who could be at the tip of the sphere working on these problems, you would certainly affect the outcome. I thought given my degree and if the situation were right, I might be able to find myself on one of those teams somehow.

And that's terrifying. I don't know that I want to be. Why would you want to be the person who is going to tip the world or one of the people who's going to tip the world? I think that the positive scenarios seem so great that you just have to go after them.

But how do you know? I mean, I make so many mistakes. I don't want it to be me. I don't know.

Yet I know that I'm just a normal human like other people. So I don't want it to be them either. So Demis has a sibilis or whatever the guy's name is. I'll put it in the show notes.

You guys can read it. The guy that started the mind who ended up selling to Google. One of the conditions for selling to Google was that they would create an ethics board so that he wanted some oversight. He wanted somebody thinking about these bigger issues.

I don't know with him or over. I don't know the details. We're probably going to need to look into that. What was he thinking?

We kind of screwed it around some of these issues. Open AI is something we're going to need to dig into to understand what are they doing there. Why do they think that open sourcing is going to make a difference in AI safety? I think we're just hoping.

I know they're just hoping. But we've tried to lay a lot of ground work in this show so far. It might be time to start digging in more specific ways about safety. I think there's a lot of sort of fundamental issues.

We'll never run out of those. We'll never run out of topics to talk about. But I think it might be time to start thinking about what could a person do and when. I don't know exactly how to approach that.

But we might look into this at the sport, for example, and think about what they're thinking about. We might look interesting. Or at some of the, you know, at some of the boss room stuff in detail or some of the user stuff in detail. We might try to understand what they're looking at and what they're thinking about.

Even the framing of the problem itself, I find very difficult because some of these problems just seem so insurmountable. Maybe we just need to do some more on AI safety. We try to do the value loading problem. We've got ourselves sidetracked.

Yeah. I kind of feel like there's some fruit in that direction though. But instead of saying we're going to do the value loading problem, we start with, we're going to, you know, instead of advertising it on the show, we start in our own minds. We're going to try to go towards X and then we go off on the first off ramp that we hit and go deep on that thing.

Because I mean, we could have just taken one of those assumptions that undergirds the value loading problem and then gone off of one assumption that underbout, you know, and that would have been an interesting problem. Okay. Well, let's do that. Let's try to find something interesting there.

We've got a whole bunch of future topics queued up that are all fantastic and exciting. So we'll find something for next week. Yeah. Well, thanks for listening.

Yeah. This was fun. I hope you guys had fun too. Until next time, this is Ted and Brandon signing off.

Bye.