 So, why do we have all these AI technologies, but no real AI? We have machines that do things that we used to think only humans could do. But nothing like the flexible general purpose intelligence that each of you uses to do every one of these things for yourselves. So why not? What's the gap? Well, the neural networks and deep learning that's driving today's AI is based on simple equations that were worked out in the 1970s and 80s to describe the most basic animal learning, think Pavlov's dogs or a rat in a maze. They're finding patterns, detecting associations, but they're not understanding. Which means it's at most one small step towards real machine intelligence. Real intelligence goes beyond pattern recognition to modeling the world, explaining and understanding what we see, imagining things we could see but haven't yet, making plans and solving problems to make those things real, and then building new models as we learn more about the world. The goal of my work is to reverse engineer these human abilities to write down their equations and then to use that to make more human like AI. And we're starting with trying to understand how children learn and think because they're the original model builders. And we are still very far from having AI that's as intelligent as even a one and a half year old. But imagine if we could get there. Imagine if we could build AI that grows into intelligence the way a person does, that starts like a baby and learns like a child. It may not be our only route to AI, but it could be our best bet because think about it, a human child is the only learning machine in the known universe that reliably, demonstrably grows into human level intelligence starting from much less. And we know that even small steps towards this goal scaled up could be big. So we're starting with the most basic common sense that's in every child and no AI. What we call intuitive physics like you see here for playing with blocks, stacking up cups, or what we call intuitive psychology that lets even young children, like for example this little guy in the back here, see what somebody's doing and understand what they're doing and why, what their goals are, and how to help them. Even for complex actions like this that you've never seen before. Now watch this kid carefully. These kids I'm showing you are just one and a half. But imagine if we could build AI that had this kind of common sense intelligence, this kind of helpfulness, this kind of helpfulness around the house would be amazing. So now for the technical part. To do this, we've invented new AI programming tools called probabilistic programs, which build on but go way beyond the neural networks that's the big AI breakthrough you've been hearing about for the last few years. And this is what you're going to be hearing about in the next few years if you haven't already. Just as one example, just in the last year, in 2018, we've used these tools to give robots a kind of intuitive physics that lets them do all these things. Stack up blocks and play the game Jenga. Imagine uses for new tools that they've never seen before, even for complex actions. Make sushi or at least the rice for sushi. Pour ice water, even learn to walk. What's next is model learning. So what would it take for a robot or a baby to learn an intuitive physics model? Learning one of these probabilistic programs means that your learning algorithm has to be a program writing program. It's the child is coder. Now this kind of learning is much harder than learning in a neural network, but somehow children do it, so a machine can too. And recently we've made a small step. Right now we can learn programs that capture a simple visual concept like these handwritten characters. Each of you can learn thousands of handwritten characters in many alphabets that you don't already know. And you can learn each one from just a single example. You don't need hundreds or thousands of examples to learn just a single concept like today's machine learning. Now our Bayesian program learning system learns like you do. It uses these probabilistic programs to capture the causal processes that put ink on the page. So action programs for drawing and writing. And then it can run these programs backwards to learn the concept most likely to have produced any one observed character. This lets our system generalize from a single example and even pass a very simple kind of Turing test. So we can show one example of a new concept to both humans and machines and say, imagine new ones. Draw new instances of that concept. Can you tell the difference between the human and machine before I showed you? Try again and here's another chance, think you can tell? My bet is that you couldn't. Now this is just again one small step towards more human like machine learning, but it scales. We can learn programs that capture the concept of a chair, that can answer questions about pictures, and that can even learn to play a new video game a hundred or a thousand times faster than today's deep reinforcement learning and almost as fast as a person. So looking ahead, once we get to this 18 month old stage of intelligence, what's next? Well, stage two is using that common sense to learn the most important thing that every child learns between one and a half and three, language. And then stage three is age three on up to adulthood, using language to learn everything else, to access the full sweep of human knowledge that builds culturally across generations and puts you in a position to contribute new knowledge yourself. So think about it, imagine if we could build AI that learns like that. This would be AI that truly lives in a human world. This would be AI that you could talk to, teach, and trust the way human beings have always done with each other, even people that were just meeting for the first time. This would be AI that could make us actually truly smarter and better off. Thanks.