 So, I'd like to tell you a bit about the progress in training AI scientists so that perhaps one day in the future, Nobel Prize in Science could be awarded to a computer. In order to convey how this works, I'll have to take you a little bit deeper into the technical details of my field, which is nano-photonics. I often joke with my grad students that the goal of this is to replace graduate students with AI. Somehow, they never laugh when I tell this joke. Ideally, you would like to train AI scientists in a lab, but such a real-world environment is very messy and difficult for two AI systems to navigate in. One of the reasons why computers are so successful in finance is because you can close the entire finance loop within a computer. Fortunately, there are a few fields of science where you can simulate experiments nearly exactly on a computer, and that makes those fields particularly suitable for training AI scientists. My field, nano-photonics, is one such field. Nano-photonics is a subfield of nanotechnology that deals with light. In terms of applications, these are lasers, LEDs, sawers, cells, optical fibers, things like that. We create artificially-created materials, which are nanostructures that lengths are substantially smaller than the wavelength of light, and their nanostructure determines their optical properties. In particular, by tailoring the nanostructure, we can tailor the laws of physics, at least as far as light is concerned, almost at will. Let me illustrate this on an example of a recent project of ours. This recent project dealt with this one amazing light bulb, which not only lit up the world, but it was actually also the best lighting device which we had for more than 100 years. Now, we have to unfortunately unbend on it because it's very inefficient. It sends most light into infrared, and only as more portion goes into visible. So we were curious to see if we can use nano-photonics techniques to enhance the efficiency of this ultimately challenging system. And in order to do that, we need to recycle infrared light back onto the filament. So we did detailed numerical optimization of this system, and then we fabricated it. And as you see, numerical experiments indeed reproduce real experiments basically point by point. And this way, we demonstrate the three times more efficient incadescent light bulb. Our nano-photonics process typically starts by first calculating something called bands of our nano-photonics system. Bands are very important to us. They tell us how the system will respond when illuminated with light of a given frequency. For example, they tell us whether it will reflect or transmit light depending on the incident wavelength or the incident angle, as you have seen on the previous slide. And we can easily calculate bands of many very different nano-photonics systems. And we can use this plethora of data in order to train our AI systems. And this ease of generating a lot of data is one of the reasons why nano-photonics is a suitable field for training AI scientists, because when we show all this data to AI, it can then develop intuition how the systems behave. So you can show it, for example, a nano-photonics system it has never seen, in this case, to a neural network, and it gives you an excellent prediction as to how the bands will look like. So you can indeed argue that it developed a physical intuition of source for behavior of these systems. And that this can be an excellent match indeed. You can see here by comparing red dots are the neural network prediction and the rest are the exact bands. But perhaps even more interesting would be if we could give some desired bands to neural network. And the neural network could tell us how the physical system has to look like in order to have this physical property. Such task requires creativity, which is something that computers are supposedly not good at. And recently AI systems started appearing which are creative. So here you train these two neural networks together and this bottom one learns a function which can transform any arbitrary random noise into images of a given desired particular class, in this case of human faces. And you can use this for many different applications. So for example, the dogs that you see on the screen do not exist in real life. They've all been imagined by a neural network of the type I've showed you simply starting from noise. People have also used similar AI systems in order to create new music, create new art paintings and actually sell them for a lot of money. We have used a similar AI system to imagine new nanophotonic systems for us against starting from noise. And now this AI system can generate as many new nanophotonic systems of certain desired properties. Now, job of a scientist is also sometimes to discover the underlying physical laws governing behavior of a system. Here you see an example of wave behavior. Neural networks can excellently predict behavior of these waves, but they can actually also discover the underlying physical law which governs the behavior of this system. I would like to end on a provocative note and that is that science is maybe the most important field that AI has a chance of cracking in the next 20 to 40 years because if we crack science, that could help us solve some of the most important problems of the humanity including cancer, climate change and many others. Thank you very much for your attention.