Nick, Peter and Fraser discuss what AlphaGo's triumph over Lee Sedol might mean for analysis and decision making. Artificial intelligence has been making waves lately, and we dive headfirst into the nitty-gritty of its recent triumphs, particularly focusing on AlphaGo's epic showdown against the world champion Go player, Lee Sedol. It's wild to think that a program developed by Google DeepMind could take on the best of the best and emerge victorious in a five-game match, boasting a score of 4-1. We kick things off by exploring the buzz around AlphaGo and the way folks started attributing human-like traits and emotions to this algorithmic powerhouse. Seriously, some commentators were acting like AlphaGo had its own ego and could feel the pressure of the game! But let’s get real—AlphaGo isn’t aware of its moves or outcomes; it just crunches numbers and maximizes its programming to play the best game it can. This leads us into a broader discussion about the human tendency to find patterns and ascribe meaning to behavior, even when there’s none to be found. It’s a quirky quirk of human nature that we can’t help but relate to even the most advanced AI. As we unpack this, we also touch on how the decision-making processes of machines like AlphaGo mirror, to some extent, our own human decision-making, which is all about weighing options and predicting outcomes. It’s like having a mirror held up to our own brains—fascinating, isn’t it?

Takeaways:

• Artificial intelligence like AlphaGo isn't self-aware but just optimizes for winning moves.
• Humans often project emotions onto AI, but it just crunches numbers without feelings or awareness.
• Understanding AI decision-making can improve our own skills in evaluating future outcomes and preferences.
• The leap from chess to Go in AI development showcases the complexities of programming intelligence.
• AlphaGo's victory illustrates that AI can learn and adapt, but real-world applications are much messier.
• In the future, AI could take over tedious tasks, freeing humans for more creative and complex work.

Hello and welcome to the Cognitive Engineering podcast where we take a look at interesting topics and what we think they tell us about analysis and decision making. I'm here with Peter Coghill and Nick Hare of Aleph Insights and today we're taking a look at artificial intelligence and recent developments there.

Specifically, we'll be discussing what happened recently with AlphaGo, which was a computer program developed by Google DeepMind in London. And AlphaGo beat the world GO champion Lee Sedol in a game of Go and it beat him 4, 1. So Peter, thinking about AlphaGo, any initial thoughts on this?

Something I found very interesting about the whole tournament was the way that the commentators and spectators were personifying AlphaGo and ascribing to it human like values and human like emotions. There was talk of it making mistakes, it realizing it had made mistakes and recovering from them, it making boring or imaginative moves.

But really this isn't really how it works at all. It doesn't have any sort of human like values. AlphaGo is not aware that it's playing Go. It's not aware that it's trying to win Go.

All it's trying to do is maximize a certain function within its core, core programming.

And I think this relates to an interesting thing that humans do a lot which is sort of look for patterns and find patterns in seemingly random data that isn't there. And through some sort of, some sort of evolutionary process, this is advantageous.

Finding threats or finding opportunities and particularly in social animals, finding things that look like other humans that you can potentially interact with. I thought that was an interesting sort of feature of the general discussion around.

I think that's sort of natural, that that's what we as humans do is trying to sort of, in trying to understand something sort of behavior, as you say. So yeah, that, that makes sense.

Yeah, I don't, I don't think it's, I don't think in this case it's, it's that erroneous to think in terms of how, how, how the, we might explain the behavior of AlphaGo in vaguely human terms because it still basically does two things which are fundamental to human decision making. There's two elements to what it does.

One is it, it sort of predicts how the move it's potentially going to make will affect the state of the board in future. And it can't, because of computational limitations, predict that perfectly.

It can only look, you know, a certain amount ahead, you know, quite an impressive amount ahead.

But the other thing it does, and I think, I think, you know, this, this is it evaluates the board and that's so it's a combination of putting some sort of value on the state of the board and then trying to look for the move that will, that will push the state of the board into something which is a higher value. Right.

We can think about that in terms of a set of beliefs which is if you like the model of the GO universe, if I do this then potentially it will lead to this outcome and a set of preferences. So its preferences are very simple.

It wants to win, some states of the board constitute winning and it's trying to push the board towards states of the board that effectively mean that it's won. So that's very like what humans do.

We have a set of preferences, there are things, outcomes we're trying to achieve and we have a set of beliefs about the world which are, you know, if I do this then then a certain type of outcome will occur. And if you want to get better at decision making, you can get better at doing those two things.

So first of all you can get better at evaluating the world or you can get better at predicting how it's going to behave. And one of the, it's interesting that we have a combination of those two things.

So there are some things I, I was thinking particularly of aphidophobia, which is fear of snakes where people just people who have that just don't want to be near snakes, they're kind of anti snake.

It's not because necessarily, well they might know that some snakes are poisonous and they might be scared of that, but they're hardwired to run away from snakes. They want to be in a snake free environment.

That's kind of different to somebody who avoids snakes because they know that they might get bitten and the snake might be poisonous.

So one is a kind of preference, a preference not to be near snakes and another is a belief about snakes being harmful and you can get better at doing those two things and become a better decision maker.

Well, hold on.

So this makes me think of a few things but first of all in terms of you were talking there about decision making and I think it's intriguing the fact that the AlphaGo 14 to 1. Right. So as a layperson I would have thought that would have been 5 nil, let's say.

And so that makes me think one of the implications, what you're saying is just should we hand over all decision making to computers? Right. And you know, and so in 15 years time we'll all be driving around in Apple cars and no one's gonna be having any accidents.

So I've got a general question about that, and it makes me think actually of a Gary Larson cartoon from years ago where you had a quiz show and you had God competing against, I don't know, Bert from Alberta or whatever. And you can see the scores and it says God 5, Larry 0. Okay.

And actually, Gary Larson said he was going to put one on there, but then he thought he better not because he'd get in trouble with fundamentalists.

But anyway, that's a roundabout way of saying what are the implications of this in terms of decision making, in terms of, look, yeah, should we just hand it all over to computers and, you know, how did it. How was it that. That Lee managed to win one of the games? Who'd like to pick that up, either? Peter? Nick?

Well, it's pretty.

Pretty.

It's astonishingly unlikely that we have developed a machine that has approximately equivalent go playing ability to a human.

There's much more likely possible scenarios where either one is massively overmatching the other than they are close enough to get at least one to four result. So it seems very unlikely.

And I think this is.

And what Peter's touching on there is the fact that we are sometimes accustomed to thinking of intelligence as something you either have or you don't. You know, that we humans are intelligent, we've got that, and we create Alphago.

You know, it kind of seems maybe natural that Alphago would be about as good as the best human. They're both intelligent. Alphago in a very narrow way, and Lee Sedol in a much, you know, broader way.

But actually, intelligence, we should think of it as being on a big scale, you know, from, from ants up to humans and then beyond, you know, and, and to who knows where, and we don't know where.

And, you know, as artificial intelligence improves, the idea is that, you know, we will see super intelligences that are above us on the line, and we don't really know how they're going to behave because, you know, if we knew how they were going to behave, we would. We would be as intelligent as them.

That's an idea called Vingian uncertainty, which is, you know, it means that that's the behavior of superintelligence is what will be something that's slightly mysterious to us. We don't know why they're doing things that they're doing.

But the point is that that is a big scale from, you know, from ants up to super intelligences. And we don't know where the upper limit of that is.

So it is very unlikely, sort of starting from scratch, that a new intelligence like AlphaGo should come in at sort of more or less exactly where the human is. It's really astounding.

It's not quite the same as with chess, for example, where we saw, you know, chess Mach evolved gradually, and then one day they. They beat humans. You know, they kind of got better and better and better, and one day they beat humans.

This is a sort of a new approach to trying to develop a Go intelligence, and it happens to be about as good as the world champion. I think that's amazing.

I just want to interrupt you there. Sorry. So that being the case, is it. What's happening is on which side is that possible?

Is it possible because of the nature of Go and the nature of chess, or is it on because of what's happening with computer programming? Because also actually, and I should have done this at the beginning really, but I'm not that familiar with go.

I mean, I know what it is and how it works, but I don't know how familiar either of you two are with Go. But can you tell me a bit about the game and how it is possible for this program to do this?

Well, I've played Go a bit. It's much more. It's sort of much more complex than chess in purely sort of search in terms of the number of states of the board.

I mean, the set of rules aren't that complex.

But the point is there's so many more potential moves than there is in chess that, you know, the sort of depth of search you need to do to try and play well is much, much higher. That's kind of actually totally irrelevant, though, because it could be anything.

I mean, we call it Go and we depict it, you know, in terms of stones on a board. But as far as the machine is concerned, it's just a state.

You know, the board is in a state, and we can label those states, we could give them all numbers, and each of those states can lead to other states. The fact it's easy for us to make it a visual thing, and Go players, I think, use quite a lot of sort of visual intuition when they're making moves.

But AlphaGo doesn't have a model of the world that looks like a Go board. You know, it's just a tree of potential moves that lead to other states. Likewise with chess machines.

Okay, so coming back to my question, then, perhaps, Peter, if you can pick up on that. So why is it that. That's just interesting to me.

Why is it with chess, it took a while for an evolution of a program that could beat a player, whereas in Go it's happened almost instantly.

I think that's probably the historical factors.

Of.

Computer science and software design and things. I think we've come on a long way. There's an exponential progression of productivity and complexity of computer science that can be achieved.

And I think we've just come a long way and I think it was a new challenge. People have been working on the Go program for as long as they've been working on chess.

It's just been harder because of the vast number of possible states. It's much greater than chess.

I think chess, the chess problem fell earlier because it's a simpler problem and it's in perhaps recent years when all the people who were doing the chess thing now moved over to Go and the increase in capacity of computing and people realized that now was the time that it could happen. So I think it's just the, it comes down to simply the number of possible states that the Go board represents.

So what does it all mean? What's the conclusion? What do we, what do we come away from this thinking? What do we, you know, especially thinking about.

I know you touched upon it earlier, Nick, but in terms of analysis and decision making, I'm wondering, does anyone look at this because of the results of this? What changes for us in real life or what changes for the way that an organization might make decisions or analyze things or even we as individuals?

Yeah. What are your thoughts on that?

I think you mentioned earlier, you know, things like driving. Driving is a good example of a real world problem which is fairly, it's fairly narrow as real world problems go.

I know it's a complex task and involves lots of different variables, but it's not that hard to specify what the objective of a driving machine is, you know. So we have a fairly good idea about the problem.

Now the interesting thing about AlphaGo is the way that it is actually a learning, a sort of general learning architecture.

It's, it doesn't, it's not as I understand it and I obviously we don't really know the details about exactly how it works, but it is basically a machine for learning stuff. And if it wasn't Go, you could probably feed something else into AlphaGo and get it to learn to be good at that.

But one of the crucial issues about Go is that there's no uncertainty. It doesn't have to learn the rules of go. I mean, it's sort of, it sort of does. I mean, it isn't told the rules of Go.

But the point is that the, you know, there is a direct link between an action that it takes and the way that the state of the board changes with the real world.

There's always going to be much more uncertainty and a machine that was designed to learn how the real world behaved would have to spend a lot longer doing things to the world.

You know, it would almost, you know, one of the things to imagine is that it would sit there effectively gathering information for quite a long time before it, you know, actually did anything because it would have to learn how the world works. And I think AlphaGo gives us a glimpse of how something like that might work.

You know, where you can feed it a problem and it will learn how that, how that problem space behaves and then work out the best decisions. The problem with the real world is just a lot, you know, there's a lot more variables, there's a lot more stuff, a lot more information.

So I think the probably the biggest constraint really is the computational one at the moment.

Peter, you wanted to come in there.

Yeah, I think that's it. I mean chess was easier than Go because there's fewer states, but I think the real world just has many more possible states as Nick described.

But I think the other thing is that on the parameters of what it means to win in Go is quite easy to define. You have more space on the board, you control more space on the board than your opponent.

So the objective is quite easy to define in computable mathematical terms.

Whereas when you say I need to drive this automated car from one place to another, what does that actually mean aside from the getting from one point to another? What does it mean not to hit stuff? What does it mean not to run people over?

Okay, so we need to wrap up. But if we just squeeze this in then. So what makes humans better at doing this stuff than computers?

Is it because we are truly self determining and a computer doesn't really know what it wants to do other than what it's prescribed to do? What, is there a gap there that will never be filled? Is there always going to be this gap?

No.

Yeah, I think, I think we will surpass human intelligence probably within our lifetimes. But the.

So what would be the what's. Okay, so that's interesting. So what will be the consequences of that for people, for us, for the planet?

I think it means that we can hand off a lot of tedious yet intelligence requiring tasks such as sort of low level design tasks, information processing tasks, estate agency, that can all be handed over to intelligent machines to free up human workers to do more interesting, more creative things. Playing to our strengths.

Okay, we need to wrap up there. So the message from Aleph Insights is, watch out, estate agents. Within our lifetime, you will be replaced by robots.

That's what we seem to be driving at here. Okay, so thank you very much for listening.

Once again, this is the cognitive engineering podcast you've been listening to me, Fraser McGrewer, with Nick Hare and Peter Coghill at Aleph Insights. Don't forget to check out our blog. There's always new postings going on there about all sorts of things that interest us.

But in the meantime, thank you very much for listening, and until next time,.