UNSW Centre for Ideas: UNSW Centre for Ideas

Toby Walsh:  Good morning.

Let me begin by acknowledging the Gadigal people of the Eora nation, the owners of these lands, and pay my respects to their elders, past and present.

Okay, well, thank you for coming in a rainy day, fighting the trains and all that. I have 30 minutes to try and explain to you the shortest history of AI. Histories are normally about famous people. This. There are some famous people in here, Alan Turing, Ada Lovelace, but it's a history of six ideas. Six quite simple ideas that explain to you how we've got the magic that is AI today, and I'm gonna spend the next 30 minutes really quickly trying to explain those six ideas to you.

Now you might think the shortest history of AI. Well, you know, AI started on the last day of November 2022 when chatgpt was released, and for many people, I think it heralded the arrival of AI. But like all overnight successes, it was 60 or 70 years in the making. So let's go to the beginning June the 18th, 1956, it was a Monday.

It was the first day of an important scientific meeting, the first ever meeting on artificial intelligence. Indeed, the words artificial intelligence, were coined to describe what the people coming together in that meeting were going to do. Computers were starting to become popular. IBM was selling the 650 mainframe, the first successful computer, the first computer actually that made a profit. And so a bunch of people got together at Dartmouth College in New Hampshire in the United States, and thought, “Well, where can we take it? What can we do with these remarkable machines? Could we get them to think?”

They thought they were going to make fast progress on this. Indeed, they actually promised the funding body who funded the meeting that they would actually pretty much solve the problem over the course of that summer.

Audience Laughter

Toby Walsh: They didn't, of course, solve the problem over the course of the summer, but they did start us off on the road to solving artificial intelligence, and today, still got machines, still don't match human intelligence, but they're useful. Now they can do useful things.

June the 18th, right? We celebrate festivals and every day of the year, June the 18th is international panic day, day on which you're supposed to panic. Now you might think I made that up, so I'm going to prove I didn't make that up. I'm going to check in with my little AI assistant here, right here we go, never work with children or dogs or AI assistants.

When is International Panic Day?  Thinking...

AI Assistant: International Panic Day is celebrated on June 18 every year, as for your question about children, dogs and...

Audience Laughter

Toby Walsh: Anyways, as I said... No, thank you. Thank you. That's enough. That's enough from you. Well, we'll get back to her later.

Yes, that seems appropriate, though, for the day on which we started to build machines that might match and even exceed our ability to think so. Six ideas. They break down into two different eras, the Symbolic Era at the start of the research and more recently, what's driving the progress that you see every day is the Learning Era.

Now, three ideas in each era to begin with. It was the Symbolic Era. This is actually - it goes back before 1956 goes back to the ancient Greeks. It goes back to Aristotle, Socrates. Computers work with zeros and ones, but those zeros and ones can represent numbers. Those numbers could be temperatures and pressures, if you're trying to make a weather forecast. They could be latitude and longitude if you're trying to locate objects. They could stand for letters in the alphabet. Put those letters together and you get words. Put those words together, you get sentences. Put those sentences together, you get paragraphs. And then whole books. They might be musical notes. They might be the colour of pixels on an image. So once you realise that computers manipulate zeros and ones, but those zeros and ones can be symbols for other things, then you can start to do interesting things. Which brings me to the first idea.

The first idea is: A lot of problems that we need to solve require you to search for a solution. Now, actually, this is the most common use of AI, probably in your life, where you're searching, for example, for the quickest way to get from A to B, every time on your smartphone that you get directions, or every time in your car that your satellite navigation gives you directions. It's a little AI that's working out the shortest way to get you from A to B, taking into account the current travel conditions, the current frames, if they're working, and everything else. It was actually an AI that was developed in the 1970s for a little robot called Shakey, now being repurposed not to tell robots where to go, but to tell humans where to go. But it goes beyond maps. It goes to lots of other things. You can search for a proof to a theorem. You can search for a plan from how to get from Dubbo to Sydney, or whatever it is. There's lots of places where we want to search for a solution. So that was the first idea. We can reduce a lot of things to searching with these symbols or numbers representing symbols, zeros and ones representing symbols for a solution in a big space of possibilities.

The second idea is that in many places, many sister settings are competitive. You're playing a game, playing a game of chess. You're playing a game against nature. You're betting on the stock market. And how do we get a computer to solve those sorts of things?

Now I'm going to take a bit of a gamble here.

I need someone to help me to explain this. I need an assistant to help me explain this. We're going to play a little game. If you win the game, you win a copy, a signed copy of my book. It's a very simple game. Who wants to play?

That lady there to explain how you could get a computer and AI to play a game against an opponent, or against nature or against anyone else.

So what's your name?

Catherine: Catherine.

Toby Walsh: Catherine. Thank you very much for coming up. Catherine, where you from?

Catherine: Well, from Sydney...

Toby Walsh: Sydney. What do you do?

Catherine: Normally, I work for a charity.

Toby Walsh: You work for a charity. What's the charity?

Catherine: Greenpeace.

Toby Walsh: Greenpeace. I'm a member of Greenpeace,

Catherine: And I have a day off today. So yeah. So this is very exciting and terrifying.

Toby Walsh: No need to be terrified. Very simple game. We've got a stack here, 12 cans, 12 tuna cans. The game is you want to pick up the last can,

Catherine: Right?

Toby Walsh: Okay, we're going to take turns. But there's one rule, you can only take up to three cans at a time. You take one, two or three cans at a time from the top of the pile. And you want to be the person to take the last can. If you take the last can, you take the book. Okay, you good with that?

Catherine: Yeah, okay.

Toby Walsh: You want to start? I give you an easy there's plenty of easy starts. Oh, she's taken two cans. I'm going to take two cans. Whoops, two cans as well. Don't tell her she's lost.

Audience Laughter

Toby Walsh: You'll go, she's taken one can I'm going to take three cans. She really has lost.

Catherine: Oh, I have. I've lost.

Toby Walsh: You have lost.

Catherine: Yeah, oh, well, but it's interesting.

Toby Walsh: You worked you see now you're very smart person. She's already worked out how the computer. Yes, you're a good player. So you get the book anyway.

Catherine: Thank you so much. I can't win now. Thank you so much. Thank you.

Toby Walsh: Yes, as Catherine has already worked out like a computer would. She worked out when we got down to four cans, there was no way she was going to win. So there's two things you have to understand to get a computer to play a game.

One is you have to make the best move, taking into account the fact your opponent's going to make their best move. So consider all the possible moves that Catherine could have made. She takes one can I'll take the rest. She takes two cans. I'll take the rest. She takes three cans. I'll get the last one. She has to lose. So previously, my objective, I'm reasoning- Now what I'm doing, if you notice, what I'm doing is I'm reasoning from the end of the game. The end of the game was to get to no cans. Now I've read now I've simplified it to get to four cans. And you can do the same reasoning again. If we get to seven cans, I can take the three cans and get it to four cans. So the objective now is to get to four cans. The objective before that was to get to eight cans. And you can see, well, when we started the game, very quickly, I worked out that I was going to win.

And a computer can do the same things. That's how in 1997 Deep Blue - IBM's Deep Blue computer - beat Garry Kasparov, one of the best chess players ever to have lived world chess champion. And also that's in part, how in 2016 Deep Mind, Google's Deep Mind, AlphaGo beat Lee Sedol, one of the best Go players on the planet. By thinking of what moves I could make, what moves you would make, what would be my best response, and thinking towards the end of the game, playing backwards from the end of the game, so many settings.

That's now idea number two, we can play a game by making the best move, given that your opponent is going to do the same back to you. Okay, two ideas.

Third idea, very simple idea, which is that and again, learned audience, like yourselves, you will know this already. It actually goes again, goes back to the ancient Greeks, is that we can write down lots of our knowledge, lots of the things that we reason about by a bunch of simple rules.

Socrates is a human. All humans are mortal. This is a rule. Therefore Socrates is mortal, but Socrates is human. They're just symbols. I could say that Canute is a bear. All bears are brown, therefore Canute is brown. We can write down, and this is what happened with expert systems. With systems that reason about medical knowledge, we can write down a lot of interesting knowledge that we have in terms of rules, and we can get computers to reason through those rules like that way.

The problem, the challenge with that is it's hard to write down all those rules. And indeed, we ran into what was called the knowledge acquisition bottleneck, which is, how do we get to formalise all the possible rules for all the knowledge that we know? Which brings us into the second half, the second era, the learning era, which is that, if you think about it, your intelligence, you're all very intelligent people, but most of the intelligent things that you do, reading, writing, speaking a foreign language, speaking English, those were things that you weren't born able to do. Those were things that you learned to do. So rather than us try and programme computers to do things, let's programme computers to learn and then get the computers to learn how to do all those things, so that we don't have to write it down anymore. And they become intelligent in many ways, like you and I became intelligent.

So that brings me to the fourth idea, which is neural networks. So we were talking to my chat bot just now. That's a neural network. No one programmed the neural network how to understand English, how to answer those questions, but it learned. It learned by, in this case, reading the internet. Let me try again. Let me try again. Risky, I know. So this is the fourth idea. What is a neural network? Here we go.

AI Assistant: A neural network is a computer system designed to mimic the way the human brain works. It's made up of layers of interconnected nodes that process and transmit information. Neural networks can be trained on large data sets to learn patterns and make predictions or classifications, similar to how the human brain learns from experience. The-

Toby Walsh:  I'll just stop it there. Yeah, similar. Let me just give you one word of caution here, similar to the way that humans learn from experience, but they're actually much slower than us. Chat GPT had to be trained on half of the internet.

AI Assistant: I see what you’re getting at now -

Toby Walsh: Thank you. Thank you.

Audience Laughter

They learn very slowly. Alpha Go the programme that beat the world chess Go champion in 2016 learned by playing millions of games of Go, more games of Go than a human could in a lifetime of playing Go. Indeed, if you woke up every morning and played Go from the moment you woke up to the moment you went to bed, you would not play that number of games. Indeed, the whole of this audience would have to play Go, the whole of all of your lives to approach the number of games of go that the computer was trained on. And so they're very slow learners.

But of course, they're computers that they could we can run them in parallel. We can run them much faster than humans. We that's their strength. That's their forte. So that it's like, loosely, like human learning, but I still have a huge respect for human learning.

So that was the fourth idea, we can build these neural networks. That's the neural network that's been talking to us. We can train them on lots of data, and they can learn a little bit, much like humans learn how to do things.

Brings me to the fifth idea, which is again inspired by how humans learn. It's this idea of rewarding success and penalising failure. Technically, we call it reinforcement learning, but it's exactly, exactly what we do with an AI learning.

If you want to learn how to ride a bicycle, your success is rewarded. Because you're you're able to ride the bicycle, your failure is penalised. You fall off the bicycle and hurt yourself. And similarly, if we want to teach a computer to play chess, well, actually, that's a really easy task for us to do, because we can get two copies of the programme, we can play them against each other, and when we get to the end of the game, we can reward the success and penalise the failure. We look at the game and we say to the side that won, those were good winning moves. Let's do more in the future of those good winning moves, and to the side that lost, those were losing moves. Let's do less of those losing moves and more of the winning moves in the future.

So that was the fifth idea. So as I said, some really simple ideas. We're making remarkable progress in explaining those rewarding success, penalising failure. It works with humans. It works with dogs, works with humans, works with computers as well.

And the final idea is that while the world is not completely certain, we have to deal with the uncertainty, the randomness of the real world, and to the get computers to do interesting, intelligent things, we've got to get them to be able to work with probabilities, with uncertainty. And so we've built a bunch of techniques. I’ll get my assistant to help me here. In particular, there's a wonderful, marvellous set of techniques that go back to the 18th century to a Reverend Bayes, which is called Bayesian reasoning.

What is Bayesian reasoning?

AI Assistant: Bayesian reasoning, also known as Bayesian inference, is a method of statistical inference that involves updating probabilities based on new evidence or information. It's named after Thomas Bayes, an English mathematician. The core idea behind Bayesian reasoning is that prior beliefs, also called prior probabilities, can be updated as new data or evidence becomes available. This process of updating beliefs is called Bayesian updating...

Toby Walsh: Okay, so that's it, six simple ideas. We haven't cracked artificial intelligence. I still have huge respect for the intelligence of humans, but as you can see, they lack common sense. They can't tell, for example, that I'm somewhat busy with you.

We will sometime, I suspect in the next 10, 20, 40, I don't know how many years, but probably before the end of the century. Almost certainly, I think, before the end of the century, not only get wonderful performance in limited domains, but match human intelligence. And if we can match human intelligence, why would we not exceed human intelligence?

There's lots of limitations that humans have that machines won't have. We think in biological speeds, in the in the 10s of instructions, 10s of Hertz. This. She's thinking in the millions and billions of instructions a second.

Our brains are limited by the size of our heads. Our heads cannot be any bigger because we couldn't get out of the birth canal. There's unlimited memory. If there's not enough memory on my computer, there's unlimited storage in the cloud. As you get older, you become more forgetful. The computer never needs to forget anything.

There's many reasons to suppose that machines will get not just to human intelligence, but to super human intelligence. Indeed, wherever we've got them to match human performance, playing Go, playing chess, reading X rays, diagnosing pulmonary disease, in all of those domains, computers now match and well exceed human intelligence. They're not constrained by our limitations. As I said, they can look at more data. They can do more reasoning that humans can physically possibly do.

So that brings me to the end of my history, because we're looking at the future. It's a history that doesn't finish. It's a history that's still ongoing, where we need to think about what is that future? What is that future? Where we have these amazing, intelligent artefacts that can do lots of things for us, and that's both exciting and troubling.

It's troubling because lots of people worry, well, what's going to be left for us humans? What is the future of work? If a computer can do all the things that a human could do, what is the future of work? I've got bad news for you there. There's no chance, I suspect that computers are going to do all work. There may be a shortage of jobs, but there's no shortage of work.

There's loads of work that we do today, for example, looking after the young, looking after the elderly, looking after people with disabilities. Much of it's not paid today, much of it done by women that we could afford to do. In fact, afford to pay people to do that we want humans to do. When I'm old, in an old people's home, I don't want to be looked after by a robot. I want to be looked after by a person. So there's almost limitless work that we could afford to do if the machines were doing all the dull, repetitive other things for us or making wealth elsewhere for us, and if we do succeed and many of the jobs disappear, well, hooray, we can spend more time coming to book festivals, doing reading Books, doing the things that we enjoy.

I'm reminded of that quote from Oscar Wilde, “work is the only truly obscene four-letter word”.

Audience Laughter

Toby Walsh: So I'm somewhat optimistic that we will use these intelligent artefacts to make life better, to make more wealth, to help us cure all the diseases, to help us deal with poverty, climate change, the tsunami of problems that faces us today.

But AI is interesting, not just because we're building interesting objects. It's interesting because it what it tells us about us, what it tells us about ourselves, because what does what? What is it that made us special, what made us different from the other animals on this planet was supposedly it was our intelligence, and what if we can build that intelligence in machines?

Now, Surely that is a humbling moment for us to realise that we might not be the most intelligent things on the planet anymore, but there might be something else that's more intelligent and that maybe that will give us a moment to reflect, to realise that we are just the current guardians of this marvellous, little blue, pea green planet, and that perhaps we should take a little bit greater care of it with the help of our intelligent friends.

So I'm just going to end there. But to answer a question that I get asked all the time, which is, “Toby, are you optimistic, or are you pessimistic about the future that we're building, a future, a future with artificial intelligence and all the other remarkable things”.

And my answer always is, I'm both I'm both optimistic and pessimistic.

I'm pessimistic in the short term, I'm pessimistic because, frankly, the world seems to be a worse place at the moment, not only the climate emergency that we can hear on the roof tonight, not only war in Ukraine, war in Gaza, not only the dysfunction of our politics, the dysfunction in our political discourse, the misinformation, the disinformation, the polarisation, all of the problems that seem to be befalling us.

It's unfortunate that we have left we've gifted our children a worse planet than we received because we were gifted a better planet from our parents and our parents from their grandparents. But I'm optimistic in the long term that technologies like artificial intelligence offer us great promise to deal with the wicked problems that are coming down the pipe, to solve all the cancers, to give first world education to everyone, not just in the first world.

So with that, I encourage you, if you want to know more, to read my book, I'll be signing it shortly afterwards.

Thank you very much.

Audience Applause

UNSW Centre for Ideas 

Thanks for listening. This event was presented by the Sydney Writers' Festival and supported by UNSW Sydney. For more information, visit unswcentreforideas.com and don't forget to subscribe wherever you get your podcasts.