 So it's a thrill to be here. I been worked. So as Jim said, the first part of my career actually was working on windows. And so it's thrilled to be here. Keynoting at an open source leadership summit representing Microsoft of all things. It's fantastic times that we're in. Now the talk this morning that I've got for you is to talk a little bit about how open source and cloud have converged to really drive this. The surge of AI and ML that we see and I'll start by talking about some of the examples that we've already seen in Azure with some of our customers applying AI and ML and shows you just kind of the difference that I can make to businesses. The first case here is Rolls Royce. You might be familiar with Rolls Royce from the cars they make, but they also make aircraft engines. And if you take a think about what an aircraft engine is for a second. It's a giant device. It's got sensors all over it pumping huge amounts of data. And before the advent of cloud and AI capabilities is really hard to make use of that data. And the kinds of use you can make out of that data include things like fuel efficiency and predictive maintenance. And so Rolls Royce has built a platform where that data goes up into the cloud. They apply AI and ML to it and they can predict the failure of an aircraft engine. If an aircraft engine fails and parts need to arrive at the airport where it's failed. That costs an airline about a million dollars a day. So just by able to predict when an engine is going to fail and have the maintenance happen before that happens or at least have the parts in place. So that when it fails it can be serviced more quickly, can save a huge amount of money. And they also use AI and ML to track the efficiency of the engine under various conditions of humidity and altitude. And by using that they can optimize the fuel efficiency of an airline flight and guide the pilot to be able to save fuel. And airline spends about 40 percent of its income on fuel. So just a few percentage points of additional fuel efficiency can be millions of dollars a year for an airline. The second case is actually a Microsoft case. This is us exploring the possibilities of AI and ML. And this was inspired this project here called diagnostic ex inspired by a project at Stanford called checks net where we've taken a data set and a curated data set of radiological images chest images and have applied AI and ML to identify pneumonia to be able to train it to take a look at a new image and determine if there's a diagnosis likely of pneumonia. And why is pneumonia such a focus well around the world pneumonia is the number one killer for children under five years of age about 16% of the deaths of children under five years of age are caused by pneumonia and us alone 50,000 kids die each year from pneumonia. And two thirds of the world's population, even despite having access to imaging devices, do not have access to radiologists. So if an ML can provide a diagnosis for a doctor physician that can potentially save a child's life. So by taking a look at this data set, we've been able to train it to identify pneumonia. And so this is just another example of application of AI and ML. The first case here on the right. I was over there on the right side and this is the predictive maintenance use case. The health use case that I just talked about is more one of image recognition, but you can see other use cases here and this shows you that a ML developers are targeting a variety of use cases. This is from a survey from slash data of a bunch of ML developers and what kind of algorithms, what kind of use cases they're targeting. You can see everything from analysis and prediction of customer behavior, which is number one applies to every single industry all the way down to the ones that I've talked about. Natural language processing, speech recognition, artificial intelligence and games being one. So this AI and ML phenomenon has really already permeated just about every industry. In fact, if you take a look, there's a font problem on this slide, but if you take a look at the most promising jobs of 2018. So LinkedIn, a company that I happen to like is publishes every year the most in demand skills. And for the last since 2014, the number one skill has been statistical analysis and data mining surpassed only this year by cloud, but this shows you that this trend has been going on for some time. Now, what's been powering this trend, like I said, cloud has a big part of it. Open source has a big part of it. Hardware has a big part of it. Data has a big part of it. And so let's break that down a little bit. If you take a look at what cloud represents, it's on-demand compute. And if you've got on-demand compute, what that promotes is experimentation. And AI and ML is really requires, it's an art at this point, and it really requires a lot of experimentation. If you've got a fixed amount of hardware, if it's very hard to get access to hardware resources, that's going to inhibit the ability to experiment. When you've got cloud on-demand, instant scale, you can go experiment where you run, train your models, try different algorithms, try different parameters, and really tune something to be able to solve real-world problems. AI and ML is fueled by data. And in the past, data was relatively expensive. So most companies were very careful about what types of data they kept. Unfortunately, if they didn't keep long enough periods of data, or if they threw away the data that might be the features that were most important for solving a particular problem, then they couldn't have an efficient solution. But with the advent of cloud storage, extremely cheap, now large data sets can be collected, and insights can be gained at a later point in time. Even if you've got data and you might not be sure about what kind of insights you can gain from it, you just store it, and later you might come up with some way to make use of it. And then finally, we've seen a huge surge in hardware capability. And hardware capability has really brought into the realm of possibility solutions for many types of problems. If you take a look at most AI, ML training, it's done on GPUs. And this graph here from NVIDIA shows you the improvement in GPU capability. It's really around 2010, over on the left side of the graph, that GPUs became powerful enough to make deep neural networks of any complexity practical. And you can see that we've gone up in roughly 10x in performance capability of GPUs today versus then. And that opens the door to solving many more types of problems in a much shorter amount of time than was possible before. If you take a look at cloud and you take a look at processing massive data sets, you need a great network to shuttle data back and forth between the data sources and between the GPUs, and even in between GPUs for a distributed neural network, for example. So the network is extremely important, and we've seen huge innovations in networking in the data center. The graph right here shows you for a data center, the amount of bandwidth out of the whole data center's bandwidth that an individual server has over time. So if you take a look, for example, in 2017, 2016, 40 gigabit NICs were added, became standard on data center networks, and you see a jump in the amount of bandwidth that can come from an individual server has as part of that overall data center network. And what this means ultimately is you've got a 40 gigabit NIC. You can see that the amount of bandwidth that's got is around 6 gigabits. That shows you that the oversubscription across the whole network continues to drop despite the fact that the individual bandwidth off a server continues to grow. In other words, the network is becoming flatter and flatter and bandwidth is becoming effectively free. And so this is what's going to promote the processing of massive data sets. And of course, open source has a big role to play in this. If you take a look at the kinds of technologies that are involved with data processing in ML, it starts with data storage platforms like SQL servers. You see MySQL and Postgres, the rise of those over the last 10 years, open source based platforms. Cassandra there, no SQL, huge rise in that. That whole ecosystem, no SQL, has been entirely open source driven. The analytics and forecasting tools, I've just got a couple of examples. There's Spark, open source, R is open source. And if you take a look at the development languages and tools, there I've got Xamarin and I've got Node and I've got Python. And those are all largely open source driven by open source. A couple of examples specific to what we see happening on trends in AI and ML. The use of things of frameworks like CNTK, which has come out of Microsoft. This is a convolutional newer network, CNN optimized platform for deep neural networks. It's completely open source. Next to that is TensorFlow, which has come out of Google, open source. MXNet, which is a framework that's a collaboration between Facebook, Microsoft and AWS, showing a community of companies working on an open source platform, this Apache MXNet. And then finally, Onyx. Onyx is an attempt to define an interoperability standard across these different frameworks. So one can export a model that can be consumed by another. And so you can use the right tool at the right place. For example, if you want to use a dynamic graph model like PyTorch, you can export the model that you create from that and then plummet into TensorFlow and get the optimizations of TensorFlow. And that's made possible by Onyx, which is, again, a collaboration across Facebook, Microsoft and AWS. Now, what has really driven open source in AI and ML? I found this paper from 2007. Remember when Dick Cheney was still president? 2007. And this paper is in a machine learning journal, Journal of Machine Learning Research, a collection of roughly 20 different authors published this paper, The Need for Open Source Software and Machine Learning. And this is screenshots from the paper. You can see the top seven reasons why they believed, why they argued that the community, the scientific community, machine learning community should adopt open source. And some of these are just general applicably to any open source, any argument for open sourcing software. The first two are specific to the research community for AI and ML. Their observation was that a lot of academic papers would be published. They would have results. But because the tools that were used to produce those results were not generally available and not open source, that it was very difficult to go and validate those results and reproduce them. And then second, if they're not open source, it was impossible for the scientific community to go look and see maybe if there's a problem, a bug, a bias somewhere in those tools that were throwing off the results. So while the tool was producing some results that you could say are valid, there was actually some flaw in them. With open source, the community could go look at it and figure out and contribute in enhancements, improvements, and find problems. So already back then in 2007, really before the GPUs really caused this surge in deep neural network, and we already see the scientific community say, hey, we need to do open source with AI and ML for all these reasons. If you take a look at the kinds of tools that people are using now in the community, this is the Kegel 2017 state of data science and machine learning survey. These are the top five tools used by data scientists. All five of these are open source. Well, Python's open source, R's open source. If you take a look at SQL, there's SQL server here, MySQL and Postgres. So largely open source. Jupyter Notebook's open source and TensorFlow, open source. And if you take a look at the libraries they're using, here's the ranking of most popular libraries. These came from dataincubator.com where they went and analyzed GitHub activity, Stack Overflow activity, and Google search results to kind of determine what are the hot libraries right now. And you can see they're all open source libraries. The whole machine learning industry is really built on top of open source. And if you take a look at what languages they're using, besides things like R, which is used for professional machine learning and AI, this survey here, also from slash data, breaks down the preferences for different languages and frameworks across data scientists based on their educational background, what they've come from, if they've been self-taught, if they've gone to university. And the thing that's striking about this is that consistently across all of them, that if they develop their own algorithms, they do so in Java, Python and C. And all of these three are open source languages. And then finally, when it comes to machine learning and AI, it's a community. And it's a very complex subject. It's quickly evolving. And the developers in that space oftentimes are stock or looking for guidance or looking for inspiration for how to solve a problem. The places they turn to for information are shown here, also slash data results, surveyed from them. Where do you go find information about or guidance about the kinds of projects that you're working on? You see number one is GitHub. Why is it GitHub? Because they can go and find projects there that they can take, learn from, reuse, fork, deploy. So nothing teaches more than the actual code. And that's where GitHub comes into play. You can see number two is websites and community forums. So again, open source is about, not just about open source and the code, but actually creating community around that code to support that code. And you see that in AI and ML with the communities that are built around top of the code inside of GitHub. And I talked about before things like MXNet and CNTK. All of that is in GitHub, which has become the repository for all things open source. Now going back to that project that I talked about at the beginning from Microsoft, where we took the NIH radiological imaging library and have trained it with machine learning models. I wanted to show you a breakdown of that architecture to show you how we've made heavy use of open source throughout that whole architecture. So you got the data sources on the left. There's about 112,000 images, by the way. 14 different pathological labels, including pneumonia, across 30,000 patients. That's ingested into cloud storage. We then launch what's called a deep learning virtual machine. This is a virtual machine you can launch in Azure that has built into it roughly 30, 40 different languages, visualization tools, frameworks. The vast majority of which are open source where you can quickly get up and running with your project without having to go and reinstall things. The model that we use here for the training is called DenseNet 121. This is actually a feed forward convolutional neural network that was published last year in an academic journal. And going back to what the scientific community was asking for back in 2007, the entire thing is in GitHub. So anybody can come and make use of this algorithm to apply it to problems like this. And so that's what we did. And you can see that the frameworks we used on top of that algorithm, PyTorch and Keras, both open source. Then we trained it on Azure with Azure Machine Learning and Visual Studio for Tools for AI and Visual Studio Code is, of course, open source. And then when we deploy the models, we deploy in a target a bunch of different run times. And you can see Onyx is one, TensorFlow is another, Core ML is another, all of which are open source. And then finally what the doctor would consume or what the analyst would consume here when they go and get a radiological image, process it through this model, is images over there on the right side that you saw before. So open source infused throughout this whole pipeline, and this project right here is available also in open source. The completely open thing, the whole thing, from start to finish, you can go access. Today we just published the whole thing in GitHub. It's at least some time today. It might have already happened. But the whole thing is going up into GitHub. So if you take a look at what's been happening over the last 10 years, really this rise of AI and ML, why are we hearing about it so much? It's the confluence of a few things. Cloud computing, infinite storage, low cost, on-demand computing, the growth of GPUs and the capabilities of GPUs combined with the tools and run times and languages that people do this kind of technology with, which are all open source and contributed to by the open source community, which has really created this massive wave that everybody's writing now to go solve problems like pneumonia and increasing the efficiency of airplane engines. So with that, I want to thank you very much again for being here. And this AI and ML phenomenon, we just seen the beginning of it. Of course, everybody's talking about how this thing is going to permeate everything we do. Again, driven largely by open source in the communities that we've all created. So thank you very much.