20. Yeah, I think everybody. And welcome to the 2018 Michaelmas term strategy lecture organised by the Computer Science Department in Oxford. This series of distinguished lectures is named after Christopher Street XI, the first professor of computation at Oxford University. The straight she lectures are generously supported by Oxford Asset Management. I'm Peter Jevons.
I'm acting head of the Department of Computer Science and it's my job to I pleasure to welcome Speaker today Professor Rodney Brooks. Mr. Brooks is the Panasonic professor of robotics emeritus at M.I.T., where he was for ten years, the director of the MIT Artificial Intelligence Lab, and then the Computer Science and Artificial Intelligence Love. After I merged with C.S.
He was born in Adelaide, Australia and received his bachelor's and master's degree in mathematics before joining Stanford University for his Ph.D. He was a research scientist at Carnegie Mellon University and at MIT on the faculty at Stanford before joining the MIT faculty in 1984. Facebook's research has been in computer vision, robotics, artificial life, artificial intelligence. He develops behaviour based robotics, which has been deployed in robots on the surface of Mars.
In thousands of bomb disposal robots in Afghanistan and Iraq inside Fukushima Daiichi after in the intimate aftermath of the 2011 tsunami in thousands of factories around the world. And in tens of millions of homes in the form of robot vacuum cleaners. He was co-founder, chairman and CTO of iRobot and of Rethink Robotics. He's won numerous scientific awards.
Too many to mention, and also starred as himself in the 1997 movie Fast, Cheap and Out of Control that was named after one of his scientific papers. But these days, he describes himself as trying to inject realism into public discourse about the AI boom. We may be getting some of that this afternoon and also working on the question of how order can arise from disorder. So there's much more I could say, but I think it's better to let Ronnie speak for himself.
So I ask you to welcome my speaker, Professor Rodney Brooks. We see people talk about artificial intelligence that's been around for a long time. Recently, there's been a movement that says it's called artificial general intelligence, which tries to distinguish itself from artificial intelligence to say they want to build a complete, intelligent entity.
And I'm going to show you that I think that's a little bit of over marketing, because that's been the goal of the original AI people from from day one. And then there's been the idea of artificial superintelligence, which is rather poorly defined. But is the idea that with soon going to have artificial intelligence, which is much better and smarter than people, and so somehow it's going to kill us all. Nick Bostrom, is Nick here, by the way? Nick's not here.
Okay. Well. Nick Bostrom, you know, did a survey a little while ago when we're going to get artificial general intelligence. And then there was another sort of extra survey around that. And the claim is that the median estimates from people in the field is that we'll get artificial general intelligence in 2040 and an artificial superintelligence in 2060. And once we get artificial superintelligence, everything's going to go crazy. I think both those numbers are very optimistic.
And Nick has had his book on superintelligence, which the press has picked up. And I think, you know, because he's associated with Oxford and it says, you know, there's a real chance that superintelligence is going to kill us all. The press really picked up on that. But if you go to Nick's Web page, these are some of his featured papers. Where are they? Why? I hope the search for extraterrestrial terrestrial life finds nothing.
Particularly he even mentions Mars. I hope there's no artificial, you know, extraterrestrial life there because it might come and kill us all. And the vulnerable world hypothesis is about all the ways that things could kill us all. And the how unlikely the doomsday. Doomsday catastrophe is about how experiments with the physics super colliders could kill us all. It's really and so on. Even the a typology of potential harms is about knowledge.
That is true, but that probably we shouldn't know because it might kill us all. But the only thing missing is, is talking about AI that, you know, in a way that it might kill us all and then damping down AI and that will kill us all. But that's maybe a bit self-referential. So I feel sorry for Nick actually, about what it must be like to be him, because he's really afraid of a lot of stuff, but it hasn't stopped other people.
Martin Rees is a dear friend of mine, the former astronomer, or maybe still the Astronomer Royal has talked about how I may, you know, end up killing us all. Stephen Hawking is worried about it. Oh, now finally we've got an MIT professor, but he must know about AI right now. He's a professor of physics, and he's never been he's never given a talk in AI or computer science at MIT. Well, then there's Elon Musk. He must know about AI, and he thinks it is going to kill us all.
It's going to be an apocalypse now. He thought that I and robotics was going to help build cars, but it turned out that was a little beyond that. And Stuart, I haven't mentioned you and I'm sorry I left you out. So Alan Turing either. I will. I will. Yeah. But here's the here's the point. Maybe maybe I'm just a grumpy old guy. And, you know, maybe they were right and I'm, you know, get off my lawn. So I decided, well, okay, let's let's rethink this, you know, am I being unfair?
And I go back to Marvin Minsky's, 1961 paper, which, for those who haven't seen it, is an amazingly good paper, well over 100 references where he talks about how to build artificial intelligence and he breaks it down in five sections. The search pattern recognition, which most closely corresponds, I think today to two deep learning sort of things. And then there's learning, planning, induction, and he he frames each of those as methods of controlling search and that's his view.
So I thought, well, you know, Marvin did that for for artificial intelligence. If I want to, you know, really not be just a grumpy old guy, maybe I should rethink things and think about steps towards superintelligence. What would it take to get there? And so that's what this talk is about. What would it take to get to superintelligence? And I'm going to go through these sections, starting with a brief history of AI.
I'm sure Alan Turing, too. Well, very well known papers, uncountable numbers in 1936 and his 1950 computing machine intelligence, which is where he would get the Imitation Game and the Turing Test from. But he had another paper in 1948, which was not published until 1970. His boss, whose name was Sir Charles Darwin, was grandson of another Charles Darwin wouldn't let him publish it,
and so that was only published posthumously. And in that he starts off and you can see echoes of this paper in his 1950 paper. The possible ways in which machinery might be made to show intelligent behaviour are discussed. The analogy with the human brain is used as a guiding principle, so I think he was perhaps the first person to talk about. Computing machinery, and he was talking about computing machinery, being able to emulate human intelligence.
He specifically talks about discrete controlling machinery by which he means essentially a digital computer. If you read the you know, this was 1948, the words went around and he says brains very nearly fall into this class. And so he talks about in this paper ways that machines could learn. And then in 1950, he published he did publish this paper computing machinery and intelligence. I proposed to consider the question, can machines think?
And this is where he started talking about what he called The Imitation Game, which started out as Could I, man. And could you tell whether it was a man or a woman? If you were just getting questions back and forth to them, if they were trying to fool you. And then he goes on to use that to say, Well, what if it was a machine that could fool you about whether it was a person or not? Then surely it could be said to be thinking is essentially his argument.
I think I think, Aaron, you may disagree with me a little bit on this, but I come back to this a little later. I go back and forth on what he meant. He does say there's no convincing arguments of a positive nature to support his views, which is an honest statement from him. And he also talks about ESP as being being a proven thing and that ESP should be perhaps considered. So telepathy, etc. So it's a little strange by today's standards.
But he says that, you know, if we look at The Imitation Game and a computer can fool a human observer 70% of the time. That's a good substitute for can a machine think? And he says, I believe that in about 50 years time, we be possible to program computers with a storage capacity of about ten to the ninth to make them play The Imitation Game and win 70% of the time.
Where does he get that ten to the ninth from? Well, he he says that the Encyclopaedia Britannica 11th edition has two by ten to the ninth bits in it. So somehow that turns into ten to the ninth probably being enough. And but 10th of life is the program space because in the computers fast enough, he thinks because then he says he can produce about a thousand digits or bits. He's referring to bits there, a program a day and that's before assemblers, remember.
So it was really arduous to write code. And so he says 60 workers working steadily through 50 years might accomplish this job. And if you multiply it out and make them work 333 days a year each, you get exactly ten to the ninth bits. So somehow the two by ten, the ninth bits in the 11th edition of Encyclopaedia Britannica becomes a program of length ten to the ninth bits.
By the way, I'm not sure anyone's ever checked this, but I looked around my my living room and that was, you know, 30 year old robot. But over in that bottom right hand corner. What's that? Oh, it's Encyclopaedia Britannica. Britannica, the 11th edition. I happened to have it, so I open to a random page counted, and it comes out to about two by the end of the night. So he was accurate. Now, people remember the term artificial intelligence coming from the 1956 A.I. Workshop at Dartmouth.
And this is the proposal was only about 12 pages long and the first few pages were written by John McCarthy. And he just goes right in and starts using the words artificial intelligence. And this is, as far as we can tell, the first use of that term.
But he doesn't explain it. We propose a two month ten man study of artificial intelligence and the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine could be made simulated. So I think despite what the artificial general intelligence people say that, oh, it's a new thing to go after all aspects of intelligence.
Certainly John McCarthy, Marvin Minsky, etc. thought they were going after all aspects of intelligence back in the fifties. John does say, well, maybe the speeds as as the Turing. Turing said we may need more memory, but not more speed. John says more speed is probably necessary, but the major obstacle was just knowing how to do the programming.
And later, you know, into the sixties, seventies, you saw John and Marvin and other people thinking that we were going to have human level intelligence within, you know, a decade sort of thing. So that idea has been around for a long time. Now I'm going to give four approaches to artificial intelligence that have come about. And these are cartoon total cartoons. They're very cartoonish. What I'm going to describe here there just to give a general flavour to it, please hold back.
You know, it's a cartoon. It's not a. It's not your book. You know, it's it's just a cartoon. So the first one is the symbolic approach, which definitely John McCarthy was, was was it was a leader. And here basically you have symbols and these are these what they call bolded words. And those symbols are sort of indivisible things. But you can talk about relationships between them and make and you can talk about rules of inference and how the rules of inference work.
And. So you can have some thing resembling intelligence discussing objects in the world as symbols. The problem has been how you ground those symbols in actuality, in real perception. And that will come in a minute when we get to deep learning and other things. But these symbols. To the machine that's looking at them don't have the meaning that we associate with and read. You know, we read a lot in the cap. But to these symbolic systems, you know, it could as well being zero five, three seven.
They're all instances of g0083. So I just did substitutions for symbols there. Oh. In those relationships, instance of that has a lot of meaning to us, but they are symbolic sorts of things and there are strict rules of inference, etc. And so, you know, really it's something more like this is how things are speaking. There's a great advantage to these symbols though.
You can compose symbols and compose things from different subsystems by using the symbols, and that isn't true for some of the other approaches. Second approach I want to talk about is neural networks cartoon again. And it's, you know, 2.0, 2.1. It's been rediscovered and rebuilt many times. Marvin Minsky, in his 1954 thesis, was doing neural networks at Princeton.
But then, you know, it came again, late fifties, came again, sixties, etc., etc. And just for those who who don't know, again, a cartoon neural networks, roughly the idea is you have these things which are in some really abstracted way, modelled after neurones, starting with a 1943 paper by McCulloch and Pitts. They're not really like real neurones in any way, but the press likes to pick up on, Oh, they're modelled after the brain. It's a very far thing from the brain.
The idea is you have inputs which come into the neurones on the left and a feedforward network where the neurones do something. I'll tell you what that is in a minute and push it onto the ones further to the right. And then and the output your these the ones on the right are some sort of classes. And and the one that is most activated and we'll talk about that in a second is the the winning class of what the inputs are looking at.
And each of those neurones is typically a linear weighted sum of the inputs. The inputs are between zero and one X, one through X, and the weight them and the weights are what I learned the numbers, just a bunch of numbers. This multiplied in sum to get a number that could be quite big, negative or positive. So you put it through a function which squashes it down to zero one again, and that's the individual neurone.
So you have inputs from some sort of sensory apparatus and it goes through these hidden layers, feeds forward. And then on the right, maybe, you know, one of the if the if the one labelled cat has the highest output score, it says this cat in the image, the dog in the image is a car in the image, but it's feedforward. And we know that the brain doesn't work like that at all in in the human visual system. And we'll come back to this in a bit.
There's always things going the other direction from the output towards the input in I think it's V v2 one part of the visual cortex, there's ten times as many connections going backward as there are forward. These are all feedforward and that has some implications. And then the big revolution about nine years ago was something called deep learning, where we went from two or three layers to 12 layers, and now you can find hundreds of layers and some versions of it.
And this has been the revolution in A.I., which I think has got everyone thinking important again, just like it was with expert systems decades ago. And it is what is driving everything about AI today. It's exploitation of this particular algorithm. And it was nine years ago that that this was was done. I think people tend to think, oh, every three weeks as a new revolutionary concept coming out in AI, but it's not true. It's it's fairly slow. The revolutionary concepts, this was a big one.
And one of the things that happened was that previously with these, you know, just two or three led networks, there'd been some algorithms developed by people that's a person, you know, a person programming, not analysing the picture of the car, but writing the operators that look at the images, maybe find circles, maybe find straight lines, etc. And they're the inputs, the network out of which comes classification in deep learning. People got rid of that handcrafted input processing.
And that, by the way, is why speech understanding has gotten so good. You know, five years ago, speech understanding systems were still pretty bad. But Alexa, the Amazon Echo, if any of you used that, it's pretty damn good transliterated speech to text, not talking about understanding, but getting what the words correspond. And we had accents don't matter too much. Our noise in the room doesn't matter too much.
And that's because the deep learning got rid of the human built early operators on the speech signal and learnt much better ones. People have overgeneralise that to say, well, we should never have the human doing any part of the job. A lot of learning. People say that, and I think that's a mistake in many ways. And the world came to know about this in a in a 2014 article in The New York Times by John Markoff, where this image was showed.
And I Google a couple of networks. Google labelled this a group of young people playing a game of Frisbee, and I think most A.I. researchers were surprised that was pretty damn good and much better than. Were you surprised, Stuart? Yeah. Yeah, I was surprised. It was surprising. This was a of an eye opener that they could do this and. Uh, so that, that, that really got the world talking about, I think around 2014. In neural networks, I'm going to include something called reinforcement learning.
I'm going to just lump that together. This was Donald Mickey, who was a colleague of Alan Turing at Bletchley Park. He did this in 1961 at the University of Edinburgh. I think it was a professor of surgery, so he wasn't allowed to use a computer. So he built the set of matchboxes with with coloured beads in them and and had reinforcement learning. Learn to play the game. Tic tac toe. Traditional robotics is the third approach I want to talk about.
I think this, you know, got started with the. Uh oh. Just lost his name. Someone at MIT. Who? Larry Roberts. Yes. Thank you. Larry Roberts. Who showed how to take images of Polyhedra and get the lines out of now. This time, taking an image was a was a job. You took a picture with a camera. You got film, and then you scanned the film. Know bit by bit mechanically to get the image. And he showed that you could get polyhedra, simple polyhedra from actual images.
And then I went off and saw, okay, what about line drawings? Can we figure out the three dimensional structure from that? And that worked out fairly well with shadows in there, etc. And it was then used for robotics experiments as the cube stacker at MIT, which looked at a stack of cubes and tried to model them, and the computer copied the A.I. program, copying the stack of blocks. Unfortunately, this is the only photo I can find of it, which doesn't really show you much much.
Swri Stanford Research Institute. Now, of course, these are right next door to Stanford. In 1970, they built the robot shakey and it lived in a world of polyhedra and with each side of each polyhedron painted in a matte colour, and it built 3D models of the Polyhedron planned how to get around. This is Hans. This is a photo I took in 1979. That's Hans Moravec up there filming his robot, the cop out.
And it's only ever outdoor run because it needed the whole mainframe and no one was going to let him use the computer. But this was the last day before the lab closed down to move the campus. So he got the computer to run all day, this one outdoor run. And you can see he built Polyhedra for the obstacles. But it took 15 minutes of mainframe processing to to to process the images, to plan a motion of one metre.
And the shadows moved a lot in that 15 minutes, and it really screwed up the algorithms. But traditional robotics then was in a sense, the world built a complete world model and then plan what to do. This is the robot Freddie at University of Edinburgh, with a fairly simple polyhedron modelled by the way, that hand is about a metre wide and a metre high. That is not a miniature camera there next to it. That's about a 30 centimetre long camera.
You know, we all today think cameras are tiny and plentiful and cheap. They were not for most of the history of I. And then there's behaviour based robotics which I attribute originally to Grey Walter, who was at Bristol at the American at Bristol. This is a 1950 paper he wrote on his tortoise's tortoise as on the bottom left there. This particular one had a circuit with two vacuum tubes or valves. They're the you know, the those that's a vacuum tube there.
There's a filament which heats up, spits up electrons and anode there, which collects them. And these plates in between can modulate it. And that's how that's how we that's how people used to do electronics. It's only got two modules, but it could demonstrate all sorts of behaviours. And this is from that Scientific American four page article. The next year he published an another one on a machine that could learn there.
It had grown to seven vacuum tubes or valves for learning and it was very much like Pavlov. The 3000 cycles there is about a whistle that it could hear and and do Pavlov type experiments. So there that was behaviour generated in a sort of holistic way.
I took that and changed that a little bit for digital, where I connected sensors to actuate is instead of building one central model, maybe had lots of partial models all happening at once and each layer generating different behaviours and having the actuate it sort of figure out what to do with the conflicting demands and put them together in a into networks. This is a robot called Genghis. On the left are these little finite state machines which make up the layers.
And there's actually six copies of many of them for each leg, one for each leg, 57 finite state machines and 12 layers there on the left. And I got these robots to walk around and do interesting things. But they they did have implications. Sojourner, which went to Mars, landed in 1997 for the primary mission of seven Sol, seven days. And the secondary mission for a further 21 Sols was was operating from the ground.
But at the Sol 28 the behaviour system was turned on and there's this was taken on Sol 72. There's Sojourner off in the distance from the lander, wandering around exploring Mars by itself. And then, as Pete mentioned, this has been the basis of the Roomba. 20 million of those in homes, a lot of robots in Afghanistan and Iraq dealing with roadside bombs in Fukushima. We got there a week after it happened, helped shut down.
And really it was essential for the shutdown of the reactors that were still operating during the cold shutdown. And then the top middle layer, you see one of the larger robots, a 200 kilowatt robot with a suction tube. That's a really big Roomba that was used to clean up radioactive stuff. And when I was last in Fukushima in 2015, these robots were still operating and many more. And the cleanup is not going to be finished till 2050, by the way.
And then more recently have put robots in factories using this behaviour based system. Around the year 2000, Damian Esler and Bruce Blumberg took my finite state machines, which sort of mix logic and behaviour together and split them into something separate called behaviour trees. And those behaviour trees have become quite popular at MIT, at Rethink Robotics, my company.
When someone shows the robot what to do, it automatically builds a behaviour tree on the left there and you can go and edit it. But more importantly, perhaps or more interestingly, about two thirds of all video games are now programmed with behaviour trees. This is one of the frameworks, unity, there's a whole bunch of them. And so people all around the world are programming that AI little characters in video games using behaviour trees.
So it's, you know, if you count the, the actual number of instances, you know, every one of the horde of those little creatures in the video game that are coming to attack you is, is running a separate behaviour tree or running an instance of a behaviour tree. The raw numbers is most AI systems in the world. Sorry, are behaviour based. I have forgiven Stuart for in his book saying there's been no known application of behaviour based robots.
Anyway, with first edition, this is my summary of the four approaches to AI. Symbolic is very deliberative. Traditional robotics is very deliberative. Behaviour with behaviour trees can be reactive and deliberative and neural networks. I see a cat, it doesn't really know what it's doing. It labels and here I get a really scientific study that's a joke of what are the strengths of the different approaches for composition?
Symbolic is the best for composition in on a scale of 1 to 3, it's symbolic. It's best composition because those symbols let you patch different things together via the symbols. But symbolic is pretty bad at being grounded, whereas the neural systems really ground in sensor data to the symbols that are the labels.
So they're good at grounding. And then, you know, just for fun, I again very scientifically compared this to a child and I added cognition, remember, on a scale of 1 to 3 and these are the numbers that I that I came up with. I think we're very, very far away from what even a child can do. And I'll talk about that in some detail a bit later. Now, I think a lot of people have been very bad at predicting the future of AI.
And I. I had a thinking technology review last year about the seven ways I see people getting their predictions wrong, including treating a sufficiently advanced stuff as magic and thereby attributing any quality you want for the purpose of an argument about what it's capable of. I'm just going to talk about two of these cases performance versus competence and suitcase words.
So performance versus competence. When we see a person perform some tasks at some level, we know we have a good intuition for what else they know around that performance in that general area. So if we see a person describe this image as a group of young people playing a game of Frisbee, we'd expect to be able to, you know, ask the person, what's the shape of a Frisbee?
We'd expect them to know that. We'd expect them to say no. Whether a person can eat a Frisbee, can a three month old person play Frisbee, how young are they? You know, kind a person. But the labelling systems know nothing about this. They not have any general competence around these symbols, which they label the images with. They have a performance, but not a competence. In the same way that deep blue Kasparov at chess, but Deep Blue couldn't be a coach in chess in any shape or form.
Whereas a human who be Kasparov could probably even teach me to play a little bit better chess than not really lousy chess that I play. And it would be weird if a person could label those images and didn't have these more general competence. So I think people hear about the performance of a system. It beat the world champion and go, Wow, it must be really intelligent. It must be able to do just about anything. But no, it's very narrow. And why are these labelling?
You know, why have these, you know, neural nets that label images based on probabilities when it comes down to it and you know, yeah, that's a 90% chance, that's a person, a 60% chance, it's a person. But we would never make the mistake. Oh, it's 20% chance. That piece of tree is a person. But they get put together and come up with something. His his adversarial attacks. On on on on. Deep learning has become very popular.
This is one of the earliest ones. His a deep learning network. It says with 9090 9%, that's a guitar, 100% that's a penguin. And then a program goes and plays around with images and tries to create images and does hill climbing and a genetic space. And it comes up with that. There's this image here, literally this image that says at 100%, that's a guitar and that's a penguin. And also it's found, you know, a set of pixels that provoke the early stages.
And that's sort of weird. And and so you get these weird classifications. My favourite there is the school bus. You can see the school buses, shortness of it. For an American, you know, that's what school buses sort of look like with a yellow. But it's clear that there's no real spatial understanding that convolutional neural networks actually do lack coherence of spatial input. And I'll come back to this later, because it actually has significant. Problems for us. Suitcase words.
This is a term that Marvin Minsky came up with where he pointed out that many words we use have many, many meanings. Now what happens is an AI researcher comes up with something that can do a little bit of classification or a little bit of recognising or a little bit of reading. The Institutional Press Office says, We got to we got to get a press release about this. And before you know it, it gets turned into, you know, a local local professor has built up an AI program that can hallucinate.
For instance, we've seen hallucinations o. AI systems going on loosely. This is really wild. But these are suitcase words. Take. Take the word learn. For instance, learn to ask means many, many different things. Learning to play tennis is related to, but rather different from learning to ride a bicycle. It's certainly a very different sort of process than learning ancient Latin.
And even though they're both in both blackboards, learning Latin again is a very different process from learning algebra. I'm good at algebra. I was really lousy at Latin and I'm really lousy chess. Learning to play chess is a different sort of skill again, learning to play music again. These are all very different skills, but we use that same suitcase word and in my last company I'd have the VCs call me up and say, So the carrier is is doing just come in.
In the sounds in the Bay Area, the carriers is learning about robots. The robots have learning. They've got to have learning. They'll be beaten. And that learning is it also shows my love for VCs and learning learning away around a new city is a very different sort of process. So these are suitcase words. And when we hear, you know, the machine can do a little aspect of it, it often gets generalised that big aspect and it's a much bigger set of skills on a code, no cheering.
And this is getting back, Aaron, to what he really meant in The Imitation Game. Can you tell whether it's a computer or a person answering your questions? And I sometimes thought that he was using it as a rhetorical device to say, well, if you can't tell, then you can't say anything more about thinking because the machine, you know, can fool you. So you can't say a machine can't think because you just don't know. And I don't know what the step from man versus woman.
To Machine versus persons doesn't quite fit in here, but with what it was doing it's a little strange. Then can it be said to be thinking? And so some people and I at various times, various versions of me over the decades have thought that it was just a rhetorical device, a thought experiment to show in principle, we can't rule out that a machine can think. But then as I reread his papers, now I see that he talks about how to construct a program for such a machine.
He thinks 3000 person years that 50 years or 60 years or 50 programmers was 50 years of 60 programmers was way too onerous. So he suggests having the program learn like a like a child and then play the game. And so at some level, he does seem to treat it as a real test. And so it's become, you know, a benchmark for A.I., supposedly, but it's something that can easily be hacked. So all the Turing Test competitions are won by stupid programs which are not intelligent at all.
So I suggest that we get rid of Turing's imitation game and think about better tests, get machines to do real tasks in the world. And I'll just give you a couple of examples. If we could do these and we we do not have a clue how to do these. I mean, you could you could give, you know, the best AI company in the world, you know, 5000 people devoted to these tasks.
And they wouldn't get very far in the next five years. Certainly if they were really lucky, maybe they'd get somewhere in the next ten years. One is an elder care worker, which is an embodied task. So a living care provider for a for an elderly person over decades. And we will need these, by the way, as as the baby boomers get older.
And it's got to understand human relationships, expectations in a household provide physical help the person, including manipulate the whole body as they get weaker, understand their degrading language. You know, Alexa can only understand pretty good language, but as a person gets a little hard of hearing and can't get nouns so well, they start pointing, they start nodding, they start grunting, understand, provide for human needs, etc., etc. a whole bunch of things it has to do.
A service logistics planner would be a disembodied system design implement perhaps a new dialysis ward in an existing space from scratch. This is just an example. Now people can do this. So if we're going to have superintelligent superintelligence better be able to do all these things, because otherwise it isn't superintelligence.
And to do that, it's got to do all sorts of geometric reasoning, quantitative geometric reason, quantitative physical simulation, understand human needs and fears, understand how family members will feel and act in dialysis for all sorts of stuff. So the elder care worker, these are happy pictures that you get from Google. When you Google elder care. The reality is not this happy but, you know, give physical help to people.
If you if they at this stage it's really not so happy because they normally wear a diaper. And, you know, figuring out how to help people at that stage is going to be very difficult for a robotic elder care worker, but help them with all sorts of physical tasks. The service logistics planner I base on the idea of an army colonel. When the US military goes into some new area, Army colonel is given tasks, set up a hospital, set this up, set up schooling, set up all sorts of stuff.
Well, set up a dialysis ward amount and Colonel will be expected to do it. And they have to figure out, you know, how the patients are going to sit or lie, how many places you need in dialysis ward, what the flow of people is going to be through it, what the the nurses or other attendants need to do with the people, what sort of information they need to give, what sort of feedback, what the layout of the space for dialysis needs to be, what for waiting room?
How people get from public transportation in the city or cars in into an out of whatever facilities whole bunch of problems which we cannot have an automatic system do at the moment. And if you say, oh, deep learning will do it, well, you've got to have an awful lot of dialysis machine examples. That's going to be pretty hard data to get in in the in the quantities that you need. So I think these sorts of challenges are the sorts of trouble.
If we if we can't do these sorts of things, we're not getting towards human level intelligence. So what's hard the day I'm going to my blog, I've got just seven random ones. I'm going to briefly talk about four things that are hard today that we that we have no idea of. But that superintelligence proponents sort of assume that we're pretty good at first. Is real perception as distinct from labelling those images? Here's some other examples. No. One, these are adversely generated.
No one would get that third one from the left and top row and say that's an armadillo. But this particular network does well. That's sort of fun and games, but it does have some real implications. So. This is from a Senate hearing in the US Senate from two or three months ago where there's a stop sign that an automatic driving vision system labelled as a 45 mile an hour sign. Now it's got four pieces of tape attached to it. It was produced. Purposely to fold it.
And if you look around the S and the T with those two pieces, it's sort of like a four and over the O and P, maybe it's sort of like a five. You can sort of see maybe how it got, but how a stop sign is red. Why didn't it get? Why didn't it get that? That couldn't possibly be a speed limit sign. It's read. Well, it turns out the colour is not colour. We think colour is just from the colour of the pixels and.
When you use deep learning and you don't use a human designed input system, it actually doesn't get colour constancy. This is from Ted Adelson. And they might be. What do we see that anyone? Want to look for in check? Yeah. How do you know it's a checkerboard? Because it's black and white. Oh, it's black and white. See those two squares there? They're the same colour. And there I've expanded the pixels. Now, how could they be the same colour? Look, black and white.
Remember all those things going downward instead of upward? It turns out we compensate for their shadow. There's a shadow there. We compensate for that. Right. And that's how we know that it's black and white, which it really is. But if you just look at the pixel value and don't label the individual parts, then you get it wrong. And what does this. And what colour are the strawberries? Red. Yeah, they're red.
So of the three quarters of a million pixel, there are only 122 pixels where the red component is bigger than both the green and the blue. They're the ones with the the biggest red component where they're both bigger than blue and green, and they're the three pixels which just have the biggest red in them.
They're not very red. And when you get rid of the strawberries, the any any if when the strawberries are there, you sort of see a little red in the grey right next to it that's flowing over from you seeing the strawberries. And when you get rid of it, it becomes less red. Some some may or may not see that, but that's where you know, you're using are these are strawberries. They must be red. Your reconstructing everything. We do that all the time.
And colour constancy is just one of many, many visual tricks that we use which make good sense. If you can't label the colours and teach the deep learning that what it doesn't learn them. So the Deep Learning Network never realises that stop signs are red because in pixels space a lot of them aren't red. And so it's not a feature and we don't have colour constancy.
So, um, realism in real problems is, is, is, is different from what I call tech rows, you know, with the myriads of them in the Silicon Valley, you know, I'm going to apply machine learning to X, I'm going to apply machine learning to I just need data. But they don't necessarily know what the data they need to get the robustness if we're going to have an elder care worker better work in human homes at a human home.
Right. This is from Aaron editing his thesis. This is what his home looks like, his kitchen. That's the sort of world we have to deal in. But we're pretty good, you know? Anyone know what that is? Yeah. It's a tow container. And how do we know that? Well, maybe that's the. The salt shaker is priming. That's the salt. But we sort of know that that's. But if you just saw it by itself or you saw that by itself, it's a rice cooker.
You know, you can get that because, you know, it's a kitchen. I'm going to do a little experiment now. Does anyone know what steampunk is? Sure. Does anyone not know what steampunk is? Oh, good. You're my test subjects. This is theme. Okay, that is a style. Steampunk is more steampunk. Here is this these are usually make up is there's more steampunk. Okay, now we do the test. You've had three examples. Is that steampunk? There are two maker faire. No, that's not steampunk.
She has goggles. Not that steampunk wear is goggles. The others had goggles while they around his neck. Okay. They thought we had goggles. Steampunk. No steampunk. Yeah. I think that's debateable. I think it's sort of lousy steampunk now. Steampunk? What about that? Yeah. It's got nothing to do with goggles. You know, this is the examples. It wasn't. We had robot arms, but you, you know, some of you, they were able to learn that category from three examples.
And you got it pretty, pretty well. When we use a deep learning, we show it millions of examples, hundreds of thousands of times each very different. So we don't have real perception. We don't have real manipulation. This is the air lab at Stanford in 1978. And there you see on the right of the image, the blue arm. We also had a gold arm which is off of the camera. I don't have a picture of it. How do I know that gold arm is there? Because that's me.
So here's the gold arm in the lobby of the computer science department at Stanford today. And you might remember these arms, Steve. And you see the gripper. There is a parallel jaw, gripper and worm screws where the two fingers come together in parallel, back and forth. That's 1978. Here's what my company was selling for grippers in Nice in 2018. Same thing. Not much improvement in the hands. We actually sold kits for parallel grippers and kits for suction grippers.
Here's the shunt catalogues thousands of pages of different size of parallel jaw grippers. And that's what people use for robot today. We are lousy at manipulation, but you know, humans can do all sorts of manipulation and superintelligence should be able to do them too. Superintelligence should be out of play or run. You know, if superintelligence is going to, you know, take over and kill all the people, it better be able to.
Well, it doesn't have to cook, but it better be able to do the all the manufacturing that people currently do with their hands. That's Julia hit. Hit again, doesn't have to be cooking, but this just shows the dexterity of humans. This is a sushi chef where they're using the force and changing the angle of the knife so they don't pull apart the pieces of the fish. And we just we can just figure out how to do these things with our hands.
We cannot get Steve, can we can we get a robot to do any of these? No, he says. Steve says, no, it must be true. We can't we can't get a robot. And all our manufacturing relies on these sorts of dexterity with sloppy floppy objects, etc. You know, some people are experimenting with manipulation for people and, you know, putting clothes on people and stuff, but we can't. This is from my home state of South Australia and if Nick was here I should tell him, you don't have to worry.
We can't give the robots the knives. They can't cut up rib cages yet. We're a long way from that. We can't do most of the manipulation that people can do the very simple pick and place, even after working on it for 40 years. So the thing read a book. Why read a book? A lot of human knowledge is in books. If we're going to have superintelligence, maybe it should read just like we read to get a lot of knowledge. And you know, every so often there's a scare a bit humans at reading.
Good headline or maybe not. This was this is where a little while ago people thought, oh, suddenly this AI system can read a book when examined, closed, not so well. And here's some examples of it. So these are the Winograd schemas from NYU. Alice tried frantically to stop her daughter from chatting at the party, leaving us to wonder why she was behaving so strangely. Who's who? She. Well, Alice. Yeah, but if we change one word, the dog was barking.
The delivery truck zoomed by the school bus because it was going so fast. What was going fast in the truck? Yeah. The delivery trucks went by the school bus because it was going so slow. Now it's the school bus. Sam pulled up a chair to the piano, but it was broken. So we had the stand and said the chair was broken. But if instead he's seeing it was the piano that was broken.
Now, to answer all those questions, you're doing all sorts of different, different sorts of simulations for each question in your head about what's going on. And we can't do those sorts of sort of common sense gross simulations at the moment. And writers have written books. Just assume we know a lot of stuff. You know, you all know that Prince William is taller than his son, Prince George.
But you don't know and you know that you don't know whether that's going to be true in 20 years, which is an interesting thing. You just know that I didn't even have to get trained on that. You just know that, you know, down the bottom, dolphins eat raw fish. So to some humans, dolphins don't usually cook food. Most humans do. We just know this stuff. And it's background and it's assumed in all writing that we know the stuff.
Our air systems have very little common sense. John McCarthy wrote his first paper on this topic in 1958. We're still away from it. And Tampa just announced a $2 billion research program on common sense. And, you know, there's a couple of sorts of it. Here's one of the pieces. Know in the main objects, agents, places. They're saying this is what we want you to do. Do all these things for objects, agents and places. One of those months on the right then not months for the research.
They're the age of babies when they can do them. These are all lives spell qui Susan Carey. So it's a really it's a multibillion dollar project on getting the common sense of an 18 month old. That's what the funding agency believes is still hot. Not quite at superintelligence yet. And I'm going to skip over writing or debugging a program. But some people think that, you know, we're going to have artificial intelligence going to rewrite its code from scratch.
Basically, my argument here is that people keep saying this, you know, we cannot we cannot have. And then they say, what's superintelligence going to rewrite the rules of physics, too? By the way, even Max TEGMARK may not believe that he's a cosmologist. But, you know, we we don't have anything that can understand this program. And I was going to go through and explain how I just glance at it.
And I know all sorts of stuff about this program. And none of those deductions, anything that we can have any sort of system do at the moment. So what do we work on now? I'm running out of time. I think we can't work on those. You know, the health care worker directly, the older care worker. We can't work on the systems logistics planet directly. It's too hard. But these are sorts of goals that we could work towards.
And if we're making progress towards any of these goals, our AI systems would have more common sense and be more robust. A two year old has colour constancy and what colour categories can map form to function can figure out that they can sit on this. This can act as a chair even though it's not a chair. So and then form and function. They have object classes and can categorise objects that look totally new at the pixel level. Unlike the deep learning they can do one shot subclass learning.
You take, you take a kid to the zoo and they see a giraffe for the first time in their life. You don't say, Oh, by the way, that's an animal. They know it's an animal. At age two and after they've seen a zoo for a few seconds, you go home and they open a book. They know that's a giraffe. One shot learning. Four year old can talk and listen when they come in about turn taking understanding cues.
They know when they're in a conversation with someone and one of the participants changes they if multiple participants in the conversation they know how to get attention and direct their remarks to someone, they can tune in a lot of conversations. They know when someone is suddenly speaking, we speaking differently from normal. So they know a lot of stuff. And currently Alexa knows none of this stuff, for instance. And Alexa is a fantastic push forward, by the way.
I'm not belittling it. It's a lot more. Six year old can estimate from vision how to pick up many objects. Whether it's going to be one hand, wants to handle what's going to wrap their arms around it, use the whole body, they pre shape their hands for a grasp they can apply and control force appropriate for a task they can use. Chopsticks. A six year old who's grown up in a chopstick house can use chopsticks as a complex thing.
They can do all sorts of tasks. None of our industrial robots can do any of these. They can even pick up. This is a step towards elder care. They can pick up cats and dogs and pet them and they can wipe their own bums, which is important for elder care, too. An eight year old can articulate their beliefs, desires and intentions, and they understand that different people, other people have different.
Beliefs, desires and intentions. Some of this might be a nine year old, but eight year olds are fit the pattern better. I apologise and they can deduce many of these things by observing others. And we don't have systems that can do that today. But if we made progress on any of these four things, it would help our systems operate in the world where things go wrong. Oh, the superintelligence destroying us. Thanks, Nick. But here's where it goes wrong.
It goes wrong in the datasets getting putting bias into systems. And some of the headlines get a little hysterical, but there's real issues there. Watson Health You know IBM tried to push Watson too fast the in the health care and it was a bit of a big failure but not why they are pushing it into into digital marketing now and that's a really. Oh well, anyway, forget it. But what if we're completely off track?
You know, we think building intelligent machines is like building copies of ourselves ultimately. Okay. Well, maybe we can do that. What if we see these two dolphins, A and B, and then we look more closely and we notice that B is a robot. Are we going to conclude that a bunch of A's got together and built the B? Did the dolphins build the robot dolphin? We don't think they're capable of it. Okay. Are we capable of it? Are we capable of building human level intelligence?
We'd like to think we are. We're pretty arrogant about it. But maybe we're not. Maybe we were just not good enough. Just like we think the dolphins are just not good enough. Maybe the aliens up there are looking down on us. Look at those little humans trying to kind of make copies of themselves. But we're just not smart enough. But maybe we are smarter, but we're going about it the wrong way.
And I want to use flight as an example here. You know, people always say, well, you know, we don't we don't we don't make things fly by copying birds. And Lord Kelvin was wrong just a few years before heavier than air flight, although it's not clear exactly what he was saying. But. No. The Wright brothers. One out. You know, earlier people had said, well, we need flapping, flapping wings and that wasn't helpful.
And then Lilienthal maybe mispronouncing his name, really understood that it was a static wing in airflow that was important for gliding. He he did over 2000 glides before he died doing what he loved gliding. And he wrote a book saying in the German title, says, It's a human flight inspired by birds. And Wilbur Wright was certainly a knew about that book read that book and he observed that birds use a change of shape of their wings in order to roll left and right.
And that was they did have better engines. But the key innovation for for Wilbur and Orville Wright was realising that control of the wing shape in the airflow was important. Up until that point, people hadn't been thinking about control. That turned out to be the the critical thing. We use our computers. We've had, you know, Moore's Law just delivering us better and better cocaine for a long time and computation and certain metaphors.
But maybe we're not thinking about things the right way. Maybe. And maybe it's you know, maybe it's going to be 150 years before someone figures that out. Some things take a long time. We do not know how long it will take. But I want to end with something that Alan Turing said in 1950. We can only see a short distance ahead, but we can see plenty that needs to be done in AI and there's plenty for us all to do. So thank you. And I'm sorry I went a bit long. Thanks.