ROBERT SAXBY: So, we’re going to have a little deep dive into machine learning today I thought we’d take the opportunity to build some models and stuff like that, and no We’re going to have a look at useful applications in machine learning So we’re going to start with some use cases So I couldn’t resist I know that you had an astronaut here today and I thought, I had to take a photo that you should all know from the Kuiper Belt So the little blue dot there, that’s us In 1990, this was Voyager turning back a six billion– yeah, six billion kilometers from Earth and actually looking and saying bye And that’s all that was left of us, this little blue dot So we’re going to zoom in a bit, because my use case is actually done here on Earth This is my favorite view of the Earth Every time I get on in plane and I fly across the Atlantic, I start playing around with that screen once I’ve watched all 20 films I didn’t see And then I start looking at where the plane’s flying And then I’m twiddling the globe around And I always like to get this backward face of the globe, where you can’t see any land on it I think it’s a really nice way of looking at the planet And it gives you a kind of perspective It really kind of messes with your head a little bit You have to think, is it really? So this use case I’m going to tell us about– actually, whenever we tell this use case at Google for a conference, for example, we always mention this island And I didn’t know where it was so I thought I’d take the effort and find out where it was And it happened to be slap bang in the middle of that view that I like to look at so much And I don’t know if you can see this here properly So the island is called Kiribati And this island has villages on it This is nothing to with machine learning yet This it just really interesting So this island has villages on it of London, Paris, Poland, and Banana So like, if you ever look into anything, you can– so, what is the use case? The use case is global fishing Maybe some of you’ve heard about it It’s a really cool use case And normally we show this picture So we show this picture of all these boat moving around, and Africa and Madagascar So I thought Kiribati was there, but it’s not It’s on the other side So that was my little adventure into coming up to this use case So what’s going on here? Basically, whenever a fishing boat passes through a certain zone, they want to know if they’re actually allowed to be fishing there And whenever they catch someone that’s illegally fishing, they say, oh, we were just passing through Yeah, that’s a bit of a problem OK, if their nets were down in the water, but you know, you’ve got some time to get the nets out The say, oh, we were just passing through So the machine learning application that they looked at was actually to take this data, and we’re talking about 40 billion rows of events, and we’ve got 200,000 ships broadcasting AIS all the time That’s the little positioning stuff for boats And what they did was actually work out how these boats are traversing the water, and what patterns correspond to fishing patterns And they applied this, and so they could actually now actually start finding well, fishing companies that were illegally fishing And one of the first ever successful prosecutions was in the waters of Kiribati And they fine that they actually issued was $2.2 million, which was 1% of the country’s global GDP This is something you can do with machine learning I’m guessing most people here are interested in fishing waters and stuff like that So this is like, really take note and you can start applying this as well I’m going to take a few use cases today from retail The reason I’m going to do retail is because I think it’s quite broad and everyone has some association with retail Either we’re working with retail or we’re involved in buying for ourselves So we’ve got a good understanding So Ocado is a British supermarket They are online They’re online only It’s got a ton of active customers Well five tons of them, to be precise And basically what they want to do, they want to process all the messages that are coming at them So they’ve got all these messages coming in on email and social media and things like that And if one of their customers is unhappy about something– maybe the delivery didn’t turn up on time, or the food was damaged, or whatever– they want to capture that message and act So they’re– you know, customer satisfaction is really important to them So they have to process all these messages This message here is like someone, I’m happy You know, my children love the website It’s so easy, and blah, blah That’s great We’ll write back to that, but maybe later The thing that we need to deal with now is maybe another message that’s coming at us, which is, where’s my food? So it’s a way that they can actually take text, unstructured text, and basically process that, apply NLP, so Natural Language Processing We offer an API for this And they can actually then find out what the sentiment of the text is, extract the entities, and they can then go and use this to actually help provide a better customer satisfaction This, and I’m not going to read them all out because it’s going to take a long time But these are some of the areas within retail where you might want to look at machine learning And there’s actually quite a lot
I mean, it ranges from everything from supply chain to pricing to customer insights So we’ve got this whole range, so it’s not just in one area It might be that you’re looking for anomalies It might be you’re trying to forecast a trend, or you’re trying to predict You’re trying to make recommendations So that all these kind of applications that you can use machine learning for to actually improve your business, whether that’s inside to actually become more economic, or outside to improve your relationships with your customers, or generate more revenue, all those important things This is just a number It’s done by a research company But I want this number It’s just to basically think This is not the only important number out there But nine in 10 people in this survey said, if the company was able to predict my intent better and suggest more intuitive products, I would use them above everyone else Now I know there are tons of research firms coming up with tons of statistics like this, but I just want to think about it for a moment I just want to ask you to think about, OK if this is happening, if customers are starting to expect something more from us, can we afford not to do this? I would say that the answer is no Maybe– I mean, I’m sure there’s a lot of discussion to be had around that But if I look at my own life, my personal life, I see more and more how I interact with machine learning and AI Or the result, the product of that, whether it’s, for example, when I connect to Spotify and I actually want to find some new music, or when it’s, as Bill showed us earlier, I’m trying to work on a document like now, and I can just go and find the document It’s becoming more and more something I use And it’s becoming more and more intuitive And I think our customers, they really appreciate advice which actually seems– this is not something that is obstructive This is something that is just actually improving my engagement with you So I think that companies are already acting in this space Companies are actively thinking about this space They have a great chance of actually going forward and being more successful with their customers But what this means is that there’s a lot of work that needs to be done behind the scenes to actually get that information out so we can start doing these things Recommendation engines don’t just pop up out of the ground, you know? You have to think about, what data am I going to use for it? How am I going to collect that data? How am I going to process it? Which model should I choose? And when you’ve done all of those things and you think, well great, I have this fantastic model now It’s able to predict stuff for me, and my customers really like it That was today Tomorrow, there’s more insight, there’s more data Things are changing You need to go and retrain that model, and see if your new model is actually more predictive and more useful than what you had before If you’re not doing that, then you’re static These things don’t just learn by themselves You need to go and actually feed them new data and retrain them So this is a very simple naive look at a neural network Neural networks are immensely popular at the moment for all sorts of things, including things like recommendations There are more classical regression models and [INAUDIBLE] There are more classical things that you could use to do a recommendation, like collaborative filtering for example But people are starting to use– they’re looking at the capacity of neural networks, because neural networks have something that’s really quite unique to them, I would say And that’s the ability to generalize very well If you can’t generalize, what does that mean? It means for example, if I was creating a model which is taking advantage of a whole load of features, and then trying to get a good regression, and let’s say I’m overfitting because I’m just plugging all that stuff in there And I think OK, this is working well But then I go and apply that to a new set of people, where they’re in the same domain, but let’s say, on the edge of it And things are changing My model, if it’s too specific, it might not really recognize their behavior or what’s going on Deep neural networks help with that They’re very good at generalizing So for recommendations, there’s actually a couple ways that you can take like good parts of regression, with a more, let’s say, a more classical set up And then go for a deeper neural network in the same model to actually get a model which has both aspects of memorization, so learning the things, very important, they’ve seen before But at the same time, flexible enough to actually be useful and not just come with the same answer because it’s this So what’s changed in the last few years? Why is this all happening now? Well, machine learning has been around for a long time And a lot of the models that you see, especially used in competitions up until a few years ago, and still when people are running this by themselves in their own machines, are kind limited by the amount of processing power you had If you think about a good model, a good model implies that it’s probably going to have to do quite a lot of work And then on top of that, we’ve got a lot of data
Now if we mash those two things together, we need a lot of those things on the right to start processing it These things on the right here are our TPUs And you don’t have to use TPUs You could use GPUs or anything else But TPUs are actually very useful for us We brought our second generation This is worth standing still, because I think this is one that differentiates us with Google Cloud in the space of machine learning So these things are 180 teraflops per circuit They’re A6, so they’re designed specifically for a certain task In this case actually calculating floating point operations 180 tera, what does that mean? It means, if we put them in a pod, and I’ll show you a photo that in a moment– sorry, Shelby, and she knows why Well, Shelby said to let you all in the joke When I showed her this photo, she said, it looks like something from the 1970s I’m a kind of 1970s guy But I like the ’70s I like a lot of stuff in the ’70s, so this really appeals to me These are all pods of actually of those TPUs Each pod is 11 and 1/2 petaflops What do those numbers mean in real terms? What do they mean for anyone else? They mean, we were translating– for example, we were taking our models like Translate and retraining that in a couple of days And now we can do it in a couple of hours So if you think about these use cases that maybe were involving lots of data and lots of complexity, now we have the ability to actually go and retrain them in relatively short amount of time, and to get a new model out there which could potentially impact our conversation with our users So the message I want to take from this is that machine learning is a really good fit with cloud Cloud offers you this, it will provide you this infrastructure to go and actually take advantage of that So how do we do that? Well, we do it in two ways So on the one hand, on the right hand side here, we have our managed APIs This is a little bit about the democratization of the machine learning So what we’re trying to do in the machine space is get more people involved I just explained in my last talk that I have two children, and whilst I’m not sure if my children are going to grow up and become data scientists– I think one wants to be a firefighter and the other one, we’re going to find out But I know when they grow up and they participate in the world, machine learning is going to be there It’s already here now We’re on the cusp and things are just accelerating I want my kids to know what machine learning is, what it’s about, and more importantly, what people are using it for I don’t want it to be something that they don’t have that connection with the world, that the world just kind of passes them by In our effort to democratize artificial intelligence and machine learning, what we’re trying to do is provide a space for people to interact with On the right hand side, we’re taking models that we’ve actually trained to solve, let’s say, common problems– OCR, taking a translation, going from speech to text We’re opening them up on open APIs that developers can use in their applications, making single rest calls to actually get a result So you saw the video demo earlier, you know, those kind of things So that’s one way of opening up at least the usage and getting more people interacting with the capabilities On the left hand side, we have TensorFlow and Cloud Machine Learning I would say the big differentiator between Google Cloud and other clouds, and I think that other clouds are actually following us, so we definitely have the lead in this space And it was the reason why I joined Google as well I’m a big data fan And if you look at the big data technology, a lot of it, there’s a relationship with Google– either they published white papers, or, in the most recent years, open sourcing a ton of frameworks So why is this important? TensorFlow is an open sourced project for machine learning It has something like 50,000 stars on GitHub So you saw Linux has 20,000 Imagine how many people are involved with this then? We know everyone knows Linux People have an affiliation with Linux So 50,000, that’s 1 and 1/2 times more, are actually they’re using– I should say 2 and 1/2 times– but they’re there, using it, experimenting with it So first of all, there’s a lot of collaboration That innovation that we spoke about earlier, about actually having people contribute to the space and seeing what’s going on and moving the needle, is the same thing what’s happening in machine learning Having it open source means we can all collaborate and work together And we have a managed service So we do that for our big data frameworks as well But we have managed service that you can take your model, come and run at us It kind of puts the kind of, let’s say responsibility, square on our shoulders to do a couple of things– make sure that we’re giving you the best price, make sure that we’re providing the best security, make sure we’re providing the best performance And if not, you’ve written it in open source, so you can go and run it somewhere else So you know your investment’s in good hands Also, if a lot of people are working on this project, it means there are a lot of developers and data scientists and analysts out there in the market that are working with it So when you’re looking to recruit more people to actually accelerate your effort in that space, there are people in the market that understand this technology You don’t have to go and just start from scratch
and train them in something new So open source is a really big part of machine learning as well If you’re not using an open source framework for your machine learning, shame on you You really should start considering it, because it really will put you in a better position for the future This is a quick architecture And we don’t have to go for all the details here, but what I basically say, this is the architecture of the Video API The most difficult part of this, the part that cost the most time to program, was the front end, the stuff that you saw And we do that in AppEngine [CHIRPING] My egg is now hard boiled So we do that in AppEngine All the stuff on the bottom is basically completely serverless So you take your video, you upload it to Cloud Storage We have serverless functions that are listening And if there’s more events happening, we will spin up automatically more functions and then with each video we pick up, we will call the API, process it, extract the metadata We take a metadata, so like, the labels that you saw earlier, which Lee demonstrated We see which index it happens And we will serve that up, in this case with putting it back on Cloud Storage But we could put that in a database or whatever else we want to put it, so we can make it searchable and available for our application So if you think, if you’ve got something around those APIs that we just mentioned, its like speech to text, translation, processing videos, processing images, extracting sentiment from images or recognizing where people are standing on images There are many things you can do And you can just use a service architecture and really get up to– hit the ground running If you’re going to go deeper, and you’re going to start building some of the stuff yourself, there are a few things I want to leave you with to actually think about in that process Make sure the developers that are working for you are giving you good answers on these questions So we want to do machine learning We want to be here at the top doing our machine learning We want to start getting this predictive power making a business impact It is right at the top of this stack Right at the bottom is infrastructure So again, using a framework like TensorFlow that you can run on basically any architecture, that gives you that capability If you’re using another project which doesn’t have those capabilities, and then you want to take something and bring it into production, you’re going to have to have your engineers find a way to actually take that from the research space into the production space That is one of the key motivators behind TensorFlow inside Google itself The Google Brain team, they were working, they were doing research stuff And then when they tried to move it to production, basically there had to be an engineering effort to take those learnings and that model, and now how are we going to get this thing in production? So having infrastructure that we can do that on, and having the applications that actually facilitate that, is a very powerful thing As we move up, we come to the kind of, let’s say the body of water and the data that we need to actually run this on Some of the key challenges there are around collecting data, processing data These are all major efforts that need to be made if you’re actually going to use machine learning So you have to have good answers to these things And then on the top, we can do machine learning If you don’t do it like this, you end up like this And this is basically what’ll happen You’ll be up there and you’ll be like, oh, I’ve got something in production And bomb, it’s gone, and you’ve got a big problem This architecture, and I won’t go into too many details about this at the moment, but I will just say that this is an architecture we use to help customers build things when they’re using TensorFlow And the important part to look at here is that we have frameworks like data flow which is this one here, which is actually unified batch– a streaming framework What that means is, I can take an action I want to perform on my data, like mapping, or a function that I want to perform, a processing action, and I can apply it to batch data So think about the data that you already have that you can use to train your model But I can also apply it to real time data, data coming in on a stream The same model, just making a few changes of how the data’s actually coming in That’s what a unified model allows you to do So if you look at the bottom there, we have date flow coming So when you’re ready to make those predictions, sometimes people forget I’ve just spent all this time taking this data, extracting all the features, and then I trained a model Now we’re going to go make predictions What are you going to make predictions on? Yeah, I’m going to make a prediction on this OK, so what do we have to do with that to actually get it in the same space as our prediction? Yeah, you have to extract those features, and now you see why having a unified model might actually make a lot of sense They’re both serverless as well, and again, as I said, they have open source projects Apache VM, which you can create your stuff online You develop this crazy stuff All the things I said about TensorFlow apply to Apache VM as well– developers working in the community, something which you can go and run on different infrastructure if you want to move, that flexibility The other one is the Cloud ML So here we’re using it to train our data And here we are using it to predict A question that comes up quite often, if I create something in TensorFlow, where am I going to make my predictions? Well, it’s up to you You could take your TensorFlow model
and deploy a compiled version to your mobile phone So you could, for example, you would be doing Translate and like, I’ve seen the example where we actually take– you take your camera, and you hold it up against a sign, for example, in Japanese, and then it gives you the English text But let’s say you’re not online at the moment You still want that functionality to work So you might be saying in your application, if I don’t have an internet connection, use this local model, which is going to be smaller It’s going to be less powerful If I do have a connection, then use this model And then you would make an API call, and then you can come back to Cloud Machine Learning So you have that ability TensorFlow, itself, I’m letting here into some of the things I explain to developers as well, because I think this is where things kind of go wrong if you’re not asking these questions So I just told you about all the good thing about TensorFlow This is the engine of TensorFlow If you want improve the engine, you have people work together to actually collaborate in that space But there’s a lot going on there at the moment It’s pretty solid As you go up the stack, these are the kind of important things to pay attention to, layers If you’re building models there, it’s a great place to learn about machine learning But it’s a place that, when you want to move from one architecture to another, you need to go and think about how is that going to look? How’s that process going to go run? That’s why we created estimators and canned estimators An estimator itself gives you the ability to actually just say, this is my model This is something to think about, doing all the wiring up And we have a custom estimator as well So you can take advantage of something very specific that you want to create And you can take that model and then run it on your local machine But then also say, to Cloud ML for example, go and run this in a cloud and train it on big data So it’s really important– the thing to take home is, really important to challenge people, at what level are they interacting with this stack? Where are we? Where should we be? And right at the top there are the canned estimators And these things these very, these are out of the box They’re not like the machine learning APIs that I showed you You still have to actually tell how your data is going to fit, and you have to extract the features and do more of that kind of stuff But basically, if you’ve got a common problem, like a common regression problem, something like that, there are estimators there that you can use as well And this stuff, Keras Keras is other framework which you can use, which we support as well A lot of data scientists are using Keras as well It also gives you flexibility of back ends So there are things to think about there And there are things that to have that discussion with your lead data scientist, or the product manager, or your data science team, and say, how we are actually going to do this? Or your data engineering team– and just start asking some of these questions It’s not about necessarily actually understanding what’s going on in all those layers But its having good answers like, are you going to build a model? OK, can we train that in different spaces? What’s the effort that you need to make to help me get this to production, because I want this in production because I want to actually start offering new capabilities to my customers This, if you get any time left today, which I’m imagining you won’t, but if you do have any time, downstairs is an instant Insights demo And basically they show Cloud ML making predictions about which advert to show which taxi So imagine taxis driving around town Some are coming from the airport going into town Other ones are going back Some have three people and some have two And this is a running model you can go and look at and basically see the whole flow So we take the message And once we’ve received that message, we make a prediction and we say, oh, look, there’s one person in the taxi They’re heading to the airport Maybe they’d like to have a coffee Or there are three people going into town Maybe they’d like to go to this theater show You can draw your own conclusions what’s a good idea But the thing is, this is all serverless So you can actually go and see how you would set up this production pipeline And the other thing is here as well to see is that we can take data flow, and the same data that’s coming in to make our predictions, we can filter off and have it go into our, for example, bit query, which is our data warehouse, streaming real time with the data that’ already there So when someone phones up and says there’s a problem, you can actually look at the data you’ve been receiving and query not only the data you had yesterday, but the stuff that’s coming in now These kind of capabilities enable you to act in this space They enable you to make things quickly They enable you to get it to production And they enable you to actually start working with it in production when you have problems That’s the story that we have around machine learning This is a photo of a TPU And I’ve got six minutes If there are any questions, I’m happy to take some Silence Come n, someone give me a question Yeah? AUDIENCE: So I have a question [INAUDIBLE] natural language But how come you mentioned that [INAUDIBLE] single, from a particular project [INAUDIBLE],, whether you have equations or [INAUDIBLE] For example, you have [INAUDIBLE]?? ROBERT SAXBY: So I would say there’s a few answers
So the question is– AUDIENCE: Repeat the question ROBERT SAXBY: Yeah, I’m going I’m going to repeat it Yeah, I got that So the question is, when natural language processing, how do we cope with different stacks? Which one should we use? Is one open, the other one closed? How do we navigate that space? So the open APIs that we have, those APIs will give you capability out of the box And so if I was designing an application and I wanted, for example, my application is, let’s say I’m taking– we used an example earlier with that chat So the chat’s coming in and I want to process a chat and actually have a signal real time, about saying if the person is happy or angry so I can, after the chat’s finished, I can maybe steer that chat somewhere else to actually have someone call back or look into what was going wrong, and give a very short turnaround If I was looking at bringing those capabilities to an application, let’s say soon, quickly, I would look at those APIs and see if that fits, because that’s something you can really speed up your development The good thing there, and if you’re going in that space is, choose an API which is on a good standard, a good rest standard And you can actually, it’s a little work to implement And you can actually, let’s say, proxy, so you can put something in front of it and actually work with it And if you want to change that out for some other capability or some other vendor later, the engineering effort is very small If you’re going to go, let’s say more specific, and you’re really trying to target nuances which are very unique to use case, use something like TensorFlow as opposed to, let’s say, one of our competitor’s offerings, which is a little bit more closed, because TensorFlow is an open project If you get into problems, it’s not about one company saying this is our thing It’s something supported by a whole group of people And they will give you help in that space They will give you a chance to take the development further And importantly, a question I got asked recently and I think it’s really worth thinking about as well, was at some point some organizations are going to get asked a question And that question is going to be, to an auditor, explain me how it works And that is going to be an exciting time So if you get that question, it’s going to be really good to have a system that you’re using, which is open and you can use tools to actually look at what’s going in at each level to provide data to the auditor saying look, you might not understand it, but at least I can show what’s going on This is happening there It’s not just some black box and your core business is now tied into this thing that you can just say, well, I ask it something and it gives me an answer There’s a big dialogue around TensorFlow And even the models that we use internally, we have a lot of effort spent For example, for an inception model for images is a trained model in images which we’ve brought out into community so other people can develop their own image So there’s a lot of shared knowledge in this space So I’d really tried to be in that space with the shared knowledge, so you can have access to people but also get access to what’s going on when you need to Does that answer your question? OK, thank you SPEAKER 2: If there are other questions, we have a mic now So everybody going to need a question This is your chance ROBERT SAXBY: There’s time for one last question, if there’s a pressing question SPEAKER 2: Time for one last question ROBERT SAXBY: Brave AUDIENCE: One of the issues with speech recognition, and living in the Netherlands is that Dutch is mostly one of the last languages you support What is the speed you think that something like that will come here? ROBERT SAXBY: So you’re talking like actually processing speech to text? So I’m not going to commit to any like planning on stage I do know from my own personal engagements with company or personal, like in a business context, that we’re really open to have that conversation, because if it’s something that’s going to change what you do, and you’ve got a stake in it, we’re happy to be a partner to explore those possibilities So we’re happy to have that conversation We are obviously working to have that support for all languages It’s just a question of engineering, resources, and linguists and everything else that we need to make that happen So I lot of our models have moved into NLP space at the moment So you’ll see a lot much better But if you have a specific requirement around that, maybe reach out to the field sales that’s on your table if you have someone there Otherwise I will point someone in your direction, and let’s have that conversation That’s something we’re really happy to talk about SPEAKER 2: Thank you Thank you Robert Robert will stick around, too So in the natural way the agenda changed a little bit And at this point, we will be at our round tables, and your table lead will lead the discussion And we take it from there, OK? MARIO: My name is Mario, and I’m going to be your host for the next 45 minutes because I love the interaction you’re currently having
on the roundtable And we’re going to go a little bit deeper We’re going to make from the conversation, we’re going to turn it into a challenge So I hope you’re excited Yeah, a little bit? OK So why are we doing this? We are going to apply a little bit of design thinking This is something that’s very deeply embedded in Google, and we want to share with you how that works But before I got to explain what we’re actually going to do, I want to recap a little bit and give you some context So digital is radically redefining how consumers interact and our businesses are being shaped, or need to transform If you think back the early 2000s, it was through the web Now it’s through this device And in the very near future, it’s going to be in much more ambient ways if you think of all the intelligent devices that are around you, and can help you in any context So for example, Gardner forecast that in just two years time, one fifth of the world’s population will have a virtual personal assistant that they will use in their mobile device And at Google, we see the same sort of transformation going on More than 50% of all the queries happen on this device But what really strikes me is that 20% of these queries are voice based And that leads me to the following thought This is a new area of consumer experience It creates a new standard So think of it as, you’re not only competing against other companies, but you’re actually competing against the best experience a consumer has ever had, regardless of industry So at Google, we have the mission to organize the world’s information and make it relevant and universally accessible And we do that in many ways and shapes So what we’re going to do next is to give you a closer look how we are using that in the Google Home, on your mobile device, and even in the car And more importantly, how you as a business can also benefit from that So with that I’d like to introduce my colleague, Lee Give her a big round of applause And after that I’ll come back to you to explain a little bit more about the challenge LEE VONSTRA: Hi, everyone My name is Lee Vonstra I work as a customer engineer for Google, for Google Cloud Team And I’m here today to talk a little bit about Google Home and Google Assistant, just how that works all under the hood– also chatbots This picture that you see here is actually a picture that was made in 2016 of what’s on the Google I/O This is our CEO, Sundar And he came on stage and he said like, at this time from now on, we will move from a mobile first world to a more AI first world Now what does he mean? Back in the past when we were creating applications, we made it that way, that we’ll make sure that it works on a mobile phone, because if it works on a mobile phone then we can use resize it, do responsive design and so that it runs across any type of device And for a very long time, that was our way of thinking, our design thinking how we would create applications But now we are on a time that we collected so much data and we have so much computing power in the cloud that we finally can change our way of thinking and make a more AI first approach, where we let our applications or interfaces be created by machine learning And we do that with the Google Assistant And the Google Assistant, that is our smart assistant, is implemented in all the latest Google products that you can buy, like think about the Google Watch, think about the Android phone Think all Android phones from the Marshmallow on up The Google Home, the smart speaker, but also in cars we’re try to implement there so that you can, when you’re driving in a car, you can ask like, oh, how is my commute to Paris? How should I drive? Can you turn on the radio, instead of looking through a screen or interacting with a screen So it’s a conversation between you and Google that helps you to get more things done in this world Now just to make a couple of things clear, the Google Home, that is the voice activated speaker and it really is nothing more than a speaker that’s connected to the internet As a microphone, you can talk to it, but the whole magic around it, that is actually to Google Assistant So that is the conversation between Google and you, and that is implemented in the phones, but also in the Google Home Now you will also hear me a couple times referring to actions on Google What is that? Think about it as apps
It’s how you can extend the system So you can create your own conversations, your own chat logics, and implement that on top of the Google Assistant And we can do that with tools like API.AI, which I will talk about later But just to let you know what Google Assistant can do, you can ask the Google Assistant to manage tasks This photo that you see here is actually a question to the assistant, like, yeah, what’s my confirmation number to my flight to Philadelphia? And then the Google Assistant will answer you with the PNR that they got from the flight How does the assistant know that? Well, you’re locked into a Google account, which has connection to your email and all that, and your calendar and so on So the Google Assistant is smart enough to figure that out for you You can ask things like, OK, how will how will my transit be? Is it busy on the road? And the Google Assistant will know that by connecting to all kinds of external services So you can use it to plan the day But you can also use it to ask questions So I could, for example, ask like, who’s the King of the Netherlands? And the assistant will answer, well, that’s Willem-Alexander But then my next question could be like, OK, and what’s the name of his wife? And then, the Assistant will know that the context is Willem-Alexander, so his wife, her name is Maxima So I mean that is something that is very smart You need more than a decision tree to figure that out You need to understand the context You need to understand what was passed in and to keep that context You can use also the Assistant for making memory, so I can ask, like, what is my favorite– my favorite color is blue And if I ask the Google Assistant next week, what is my favorite color again? Then the Assistant will know, well, you told me it was blue So it can remember things And you can also use it to control your home, if you have all types of smart devices like the Philips hue bulbs or the Nest You can say, oh, turn on the lights Turn up the heat Turn the volume on, or play my favorite TV show from YouTube or on Netflix As long as you have smart devices that are connected to the internet, then your Google Assistant can communicate with that Here, I came up with an example how you can implement that logic for Google Home, or for the Google Assistant, within your company I mean, we’re working with companies, and I assume that we all have internal applications running Those are applications that are run behind a firewall So one of such applications could be a meeting room application, to book a meeting room So I could, for example, ask the Assistant hey, OK, Google, let me talk the book a meeting room Book a meeting room in this case is my custom action It is the action that I, as a developer, will create And what you will see is, the Google Assistant will talk back And it will say like, OK, sure here is book a meeting room And then, at that moment, the voice of the assistants will change, because then it will go to a voice that the developer chose for their custom action So you’re coming into the custom action flow now So think about it as an app or a flow And it will start asking with the welcome in text, like, OK, welcome How can I help you? And I say like, OK, yeah, I’d like to book a meeting room for three persons Sure, for when? Tomorrow from 2:00 to 3:00 PM And what happens here is, tomorrow from 2:00 to 3:00 PM– or I want to book a meeting room for three persons Three persons is a parameter that will be sent to an external system So imagine that you have a book a meeting room internal application, which has a REST API I just pass in three persons, and then the API would handle it back It would say, OK, the next request that I should make is a request for the time OK, tomorrow from 2:00 to 3:00 PM And then the system will tell me back like, OK, it went fine The meeting room is booked And then the Assistant can tell me that Or, the meeting room was not booked because maybe it was already booked before And then it can come up with a second suggestion Something like this is not so difficult to create with a tool like API.AI API.AI is like a developer console It’s like a SaaS application, a website you call there And you can create all these intents, so you can write out these conversations But it also contains machine learning so you can train the system So when you ask questions, it becomes smarter over time If you ask the question a little bit wrong, it will understand which question
you get like a similar behavior as if you would have a real Google Home So I could start typing to it and it gives me answers back But I could also use the microphone for my computer and ask questions And then it would read out the answers as if it would be the Google Assistant Once you’re done, when you’re really happy with this conversation, then you can deploy it And kind of like an app store, you need to register the name and then it will be whitelisted in the Google App Store, or in Assistant App Store, kind of That means that there’s approval process Somebody within Google will check your bot, see if it’s valid enough See also if you don’t let the microphone open nonstop, because that is actually one of the guidelines You can never leave the microphone open Out of the box that doesn’t work on a Google Home It always keeps it open for a minute or maybe less, and then automatically shuts off But if you don’t let your users know that the microphone is open, then your [INAUDIBLE] process will fail So make sure that you always ask for more questions So when you create a conversation, you should always answer like, OK, this is the answer Do you like to hear more? OK, and this is the answer And in case you’re really interested in this, in creating Google Assistant conversations, or using API.AI, there is this design document that Google created that can help you with creating conversations like this Definitely a good read I think with that I can hand it over to you, Mario MARIO: Thanks, Lee! LEE VONSTRA: Yes MARIO: Well, thank you, Lee What we try to do is to basically show you the technical toolkit So the next ingredient we want to share with you, and after that you’re really going to get your hands dirty, is a conceptual creative toolkit So if you can switch to the next slide, thank you So this is all about creating the future You now have the opportunity together to create a future together If you go to next slide So at Google, we think of creativity and innovation not as an event We see it as a process, perhaps even more as a mindset And we worked a couple of years ago to get with Stanford’s IDEO and a couple of other experts to develop this method, which is called designed thinking And basically, it consists of three main concepts So I wanted to walk you through, because I think it’s very useful in your exercise So the first one is, if you are starting to build something you don’t start with a revenue model and buying some great technology It always starts with knowing the user Once you know the user, their beliefs, their needs, that’s the perfect starting point to build a magical solution Now the next part you want to do– and basically, again some practical guides You could apply a framework of three pillars of knowing the user The first is analyze So what data do you have available to analyze the user? Second one is observe There are ways you can observe your users And the third one is to ask them There’s no better inputs than asking a user how they feel, what they need, et cetera So the second one, Andrea Cabus already touched upon We don’t think in incremental improvements of 10%, but we want to improve and think about 10x A perfect Google example, I think, is Google Loon In rural areas where there was no Wi-Fi, it was a childish idea to come up with air balloons that could send out Wi-Fi signals But it’s these ideas that in a grown up setting are considered impossible, which are the strongest ideas So ask yourself in the team, what if, because asking the what if question will make it bigger And a final one is be proto driven You want to make sure that you fail fast, because if you incorporate that, you will learn if you succeed, but also if you fail you can incorporate that learning So we see the success not by how high your success rate is, but how often you have tried something OK, go to the next one All right, so coming back, Lee gave you the technical toolkit,
how you can build a conversational customer experience I hope I’ve just gave you a little bit of a toolkit how you can have a creative process to come up So we’re going to put that together So the challenge, we ask you, and you can continue the conversation you had in your roundtable, redesign the customer experience by creating a concept that uses this conversational AI So that’s your objective We’re going to have hand out tables on the round tables, which is a framework you can use to put in your ideas You’re going to be time limited I’m sorry, we’re already far over time But again, you only have 20 minutes Your chance to win a Google Home, I’m afraid I can’t keep up with that promise We don’t have Google Homes, unfortunately But the important thing is– it was not intended to be in there So apology It’s going to be Netherlands, end of year We’re going to score based on business impact So how big is the impact you’re going to come up with the concept? What’s the customer experience like? So who is your customer? What are the integrations you could use? So that could be your own data, could be external data And finally, originality– could you make a flow describing the conversation you would be having with a customer, with one of your partners, employees, et cetera? OK, go to the next slide So as a final example, to give you some food for thought, this is a very cool concept that eBay demoed a couple of weeks ago So I bought a camera six years ago, but with the technological advancement of smartphones I think the camera in my pocket is far more powerful and easy to use So I don’t use my bit camera anymore So what I actually could ask Ebay is, how much is my camera worth? And it would ask me a couple of questions like, OK, tell me the brand, the color, the condition And it would then search in its own great database and could provide me with a very precise estimation of how much my camera is worth I think these are the kinds of concepts you might be thinking about So with that, I wish you all good luck And the 20 minutes starts now AUDIENCE: Here we go [MUSIC PLAYING] MARIO: That’s the end of the challenge, so please finish your discussions It’s a good sign, I think You’re having great conversations So at the end of the challenge, what we’d like to do is to have a couple of teams come onstage to share with us their great concepts they came up So are there any volunteers that think, my concept is the next big thing Awesome What’s your name, sir? EDWIN: Edwin MARIO: Edwin? EDWIN: Yeah MARIO: Hi, Edwin Can I ask you to pitch your concept you and your team came up? EDWIN: I will All right, so the idea started from the company I work for Part of the activity is a retail store around gardening, pets, and baking And what we wanted to do is to create Rop, the personal gardening coach, because what we are, we are a retail store, or retail chain, that provides premium
products with a premium of advice, and by doing so justify a premium price But of course, if we have to wait until the customer shows up in our store with a problem that we can help to solve, then we might miss an opportunity So the idea would be that we create a personal gardening coach, which is a chatbot in an app, so that when a client of ours walks through the garden and spots a bug on some of the plants, that he can open the app, call up Rop, and say hi, I’ve got a problem I’ve spotted a bug And then the chatbot would say, OK, let’s see what we can do Can you send me a picture? Which would mean access to the camera You send a picture and then you start an entire conversation And at the end of the conversation, you would have a solution to the problem It would have the link to a product and it would be able to tie it to an online order or a reference to the nearest store to come and pick up the product that we might provide to them And then it became really spacey because we said, you know, we have added value services What would you think if we’d send over a drone right now to do a scan of your garden as an added value service, so that we can find out how your garden is organized? And what if this bug we spotted might spread out to other areas in your garden? That was one of the fancier ideas around it So what is the business impact? Well, the business impact is that we are there for our clients the moment they spot something So the moment they detect something that we can assist them and find a solution We can tie back to a full experience So it’s an online order or a reference to a nearest store And we would get better insight, because if we spot a bug in a certain area, and we know that that is something that spreads very easily, we would be able, based upon client information, to send out warnings to other clients saying, well, there is something going on in the neighborhood Be aware, and if something goes wrong with the plans, there’s a very big likelihood that it might be linked to this In terms of integrations, well, voice API, image and video API, geolocation– where are you? Dear Mr. Client, but based upon what you have on the phone that’s easy Then of course, the client profile data we have Geolocation, weather data API– maybe there is something wrong in that area because of lack of water or something, a lack of rain like we are experiencing right now And then of course, a bug disease and solution database that we need to tie it all together So you see a lot of integration points And we’re not even talking about the drone So in the customer flow, that’s what I already explained So the customer journey, you walk through the garden You see something and you open the chatbot, and you start having the conversation MARIO: I loved idea one One question Would you also have a female assistant? EDWIN: Of course MARIO: Great I love the idea Any other volunteers? Come up front What’s your name, sir? HEIT: Heit MARIO: Heit HEIT: Heit OK MARIO: Excellent HEIT: Thank you Hello, my name is Heit And we’re in the home care business, I think, or supporting the home care system in the Netherlands [INAUDIBLE] the Netherlands So we looked at another solution for home care nurses So the nurse assistant would be a male, I think, probably I think they’re really independent nurses They’re most of the time at home with patients, so assistant would be appropriate I think Most of the time they need to supply themselves with materials for wounds, for other things for the patient at home So what are things they will need external people to help them? That’s one thing we do with our company, to just help them with IT solutions, to provide them information But I think the customer experience could be for supplying them with medical supplies without any extra time for calling to other companies for delivering them So the business impact would be, they will have much more time with the patients That’s something we always want to try to do with our solutions So let me read something That’s most important business impact, I think, that it will have more time with the patients and less time administrating things So the assistant will start, I think, with personal health records Social media ratings could be an integration for looking up ratings for suppliers Speech API, of course Electronic health records we already have And maybe personal data from patients when they are at home, when do you want something delivered? Well, how could it work? So the assistant would ask, what do you need?
Well, I know it’s Mr. X, because I’m already there So the system knows He has a wound OK, well, what do you want to do, based on the wound? Well, I want some new supplies So it could be a question you can ask to the assistant Well, in connection with the data that we already have, or could be available, then you could say, well, is there already a wound– are the materials ordered? Can we ask you when you are at home? At what time should it be delivered? So it could be in combination type outs based on some questions And eventually they will order the supplies And I think the benefit would be that the supplier would do the rest So the nurse can go to the next patient So that’s part of our brainstorm session solution MARIO: Thank you, Heit So no more visits to the doctor We have room for one final concept Yeah? One more You want to go as well? OK, we’re going to have two Ladies first But you need to be a little bit shorter on time to make sure that we cover both Thanks What’s your name? DIVIA: Divia MARIO: Divia DIVIA: Yes Good afternoon, everyone My name is Divia And the product, or the innovation that we are launching is called as Companion So it’s a companion for the elderly people And we see that there are lots of issues in our society now because all of us are becoming very individualistic There’s a lot of growth in the elderly age group, and they have lots of aging effects– going into depression, feeling loneliness, and lots of old age illness So what does the business impact to our Companion would be that we will increase the quality of life for the elderly We will bring happiness to their life And we will improve their healthy life state, which in turn would help us in society and lower the health cost and the insurance cost And I’ll go to the customer flow first, and then the integration So it’s morning The Google Home talks, hello, Richard So the interaction is starting from the Google Home So it’s not like, I have to speak to my Google Home It’s Google Home speaking to me And through Natural Language Processing, NLPs, we can identify the mood of the person We can identify time of the day We can identify the agenda of the person Looking at the weather, health of the person, we can have a call to action in terms of each one of that To give you an example, if I haven’t spoken since morning, it will call to action to my friends or family that there’s something wrong I’m not in a right mood, or I’m not having good health, so it will call my relatives If the weather is nice, it will call my friends to play bridge with me If it’s raining outside and if I have to go outside, it will advice me to carry my umbrella So this is the customer flow just to give you an example And the integrations, third box was integrations to APIs It could be with voice recognition, to interpret my mood It could be with the health care systems, the weather data, the sentiment analysis, insurance companies, medical records, and the pill box Thank you MARIO: Wow Really amazing The last team that can present with a plant Oh my god Is this your Google Home? What’s your name, sir? PETER: Peter MARIO: Hello, Peter PETER: Hi MARIO: One minute PETER: OK, you want to be short? You asked me to be shorter OK Actually, this is Wilson You know Wilson from the movie “Outcast?” Actually, the first, the lady, ladies first, her idea is brilliant Actually, we had the same idea But we already have the solution here This is the interface for the elderly You can speak to your plant It sees you, it feels you So this is the solution And we just build on the APIs that they were making So, you know, here’s the partnership MARIO: And actually to first team, how to make sure the plant doesn’t die, right? PETER: [INAUDIBLE] MARIO: OK, cool Excellent So I’d like to close off I hope you’ve had fun, great interaction with your table, learned from new people some new ideas And I really hope that you can take away to start tomorrow building the new future using some of the conversational toolkit So with that, I’d like to close off Please give another round of applause for yourself And I’d like to invite Eric Haddad for the closing session
It’s OK? ERIC HADDAD: Yeah, yeah, yeah, OK Thank you Hi, everyone So I understand I am the last speaker, so I will try to be very short So I want just to give you– to not let you leave without having to say you shot a bit We stated of the business that we have hopefully now with Google Cloud So I’m running the business for Google Cloud in Europe, so I’m looking at [INAUDIBLE] like countries in Europe So [INAUDIBLE] or Holland and Belgium and Luxembourg And so you know, I joined Google enterprise– actually it wasn’t Google Enterprise six years ago And I used to work at Microsoft before so I am in this cloud business for, I’ve been in this cloud business for the last 10 years What I can see, and what I can say, is at the beginning, the description was coming from Google to Microsoft, actually, with the [INAUDIBLE] platform and mobility It was the kind of description we could see at the beginning So, was very beginning So we saw kind of linear growth, on the cloud adoption, depending on the country But we could say that Nordic countries were maybe a little bit early adopter– France as well, UK as well And now we start to see quite a big acceleration This description comes from machine learning and data management is where I see the description And it is where I see, when I discuss with customers like you, from all industries across Europe, what we see is a lot of discussion related to machine learning, data management Why? Because there was a big impact with a business There is a big drive already to a business And CIO, CDO, and so on, they look for this kind of impact and they can look at very direct correlation So we see this exertion And what we do as well in order to do all this exertion is to invest So we have– we probably learn about that We have approx 500 new releases of services, I think that for the last 12 months And we continue to have quite an ambitious road map We are missing $28 billion dollars of capex in terms of data centers and infrastructure So over the next few years, so the big, big, big investment from Google in order to be very present, very serious, in this cloud business And we are looking for more investments So we want to be a real partner for enterprise, not only for start up Start up and dot com companies I think, we’re quite recognized for that We want to invest more in order to be sure that for enterprise business, all industries, we have this recognition as a partner So we are investing in your success And I mean that we are not selling product, hopefully I mean, that’s a kind of mindset that I want the team to have with you is more, we are trying to understand your challenges, your priorities And in order to make that, we have to reorgnize our team in terms of verticals, industries, so we are trying to have people with a mind set of vertigo, like banking, like manufacturing, like gaming And to be sure that we understand, for sure, from day zero we cannot understand everything But at least people in the team, we try to learn from your challenge and prioritizing that to be sure they understand and they can give you good answers And so, in order also to be sure that you benefit from best practices, we always are trying also to integrate, to share our best practices that we have in terms of economics in order to give you some solution in terms of AI, in terms of what works and what does not work, in terms of business solution And lastly in terms of something very tangible, we are, as we speak, we are hiring 3,000 people, technical people, from pieces to posters It’s quite big, I can tell you [INAUDIBLE] first time I see it so bigger, having investment So in part of a commitment of hiring So every week we have a dozen of new candidates that we approve And hopefully are we are able to bring these people in front of you, for workshop but also technical support and any support or technical advice that you need
Today we have some [INAUDIBLE] See, this is like old times So I should have removed it Actually, I’m sure you saw this slide, so in a few [INAUDIBLE] So we have two million customers– enterprise customers, B2B customers And what’s new today, quite with some modesty, we are walking with all industries You could say before, some industries, some vertical were a little bit scared of us, with the cloud and maybe with Google So today we work with big bank like BBVA, BCG, ING So we have a real conversation and real partnership, strong partnership with bank We also have strong parnership with a few [INAUDIBLE] as well And as you know, there is a quite very important rule called GDPR in Europe Good to come by May next year So we are planning to be absolutely compliant for that We need to be absolutely comfort with that We’ll be compliant OK? So I can tell you, it’s not– because I got some question already to that Be relaxed It is all life, [CHUCKLES], OK? If we are not GDPR compliant, we stop the business, OK? We be GDPR compliant So we have a plan for that We communicate Maybe we don’t communicate enough, OK? And I would just [INAUDIBLE] to see how we can better communicate poor activity with our customer to be sure, do you understand the plan and roadmap And for sure, I need to mention that we have quite a huge business with [INAUDIBLE] and fashion And so in all Europe, most of the big brand– not all, but many of the big brand in retail and fashion and relative business, we have quite a good partnership There is no good customer discussion without partners And we say the partnering evolution has been somewhere a mirroring of the business, OK? So when the business was still immature, OK, like a niche, the partners were mostly niche partners, OK? Very good partners but niche partners And the more we evaluate and we offer you, we get to a level of maturity of this business and the cloud business, and the level of conversations mature for you And the more we have also other partners So we diversify partners We have also still niche partners– very specialized But we have also more and more large [INAUDIBLE],, like [INAUDIBLE] or [INAUDIBLE] And also software as a service partner kind of application providers that are also completing the ecosystem of partnerships And to be very clear, partnership for us is quite important, is [INAUDIBLE] We can have a direct conversation with you But we have always partners not between us but at least with us, to be sure that we have the best complete solution for you and that the [INAUDIBLE] and the success of the project is guaranteed Talking about partners– I’m not going to talk too much [INAUDIBLE] But I wanted to take the opportunity of this session to give you an example of three or four customer So, OK, very different The first one is obviously the France in space That the leader in satellite imagery, OK? So the problem was to deliver a high quality of image when you have snow, OK? You can imagine that Imagine when you have snow, you obviously also cloud And when I say “cloud,” not the IT cloud I mean the cloud– real cloud in the sky And [INAUDIBLE] it’s quite difficult to identify with what is cloud and what is snow, OK? And so what we did was [INAUDIBLE] them to solve a problem that was they span several years to improve the quality And a few months– a couple of months– I think, three months maximum, we willfully improve a lot the quality of the image So that mean customer satisfaction– that’s quality to machine learning, customer satisfaction, and hopefully more business Ocado– that the pure grocery online The UK– quite a good [INAUDIBLE] in the UK Many today So they have no store You probably heard that to some companies like Amazon But stores, OK? And so this company does not have store They are only online So for them, the key success factor is customer services They need to have a fantastic customer satisfaction and be sure the answer to any problem [INAUDIBLE] So in order to manage and to do a triage of text or messages
or calls from the customers, OK, they’re using the platform– machinery platform– in order to triage And hopefully, they can announce where there was a time accompanying, they can answer very quickly and solve the issue– so business as well Spotify– you know probably this leading companies in music streaming So the [INAUDIBLE] and the problem, they want to be sure that every one of us is able to have all the music you want or we want in your hand So what we do is to make possible to have a maximum of [INAUDIBLE] to the user to be sure they are happy with the music hopefully they get from Spotify Air Liquide the last example, Air Liquide is quite a very old company, probably the oldest one company– very traditional– in France, part of the CAC 40 They are [INAUDIBLE] company And they were in the process of acquisition and growth as well So looking for new talent And they were not able to acquire so much because their IT was quite obsolete and not organized to acquire new companies So what they did was to start the process of [INAUDIBLE] with BCG, the Boston Consulting Group And from this first [INAUDIBLE],, they went to the creation of a night– what they call a night lab– a kind of lab, OK– out of the company to understand and to look at the new processes and new ideas of business and usage they need to have for entering the new world of digital They did it And from that, we enter an additional transformation We help them to manage transformation of whole the users– I mean, 50,000 users And now they just got the acquisition of Airgas in the US– 20,000 users And we’re working with them for the next transformation That the kind of transformation we did So all coming up to you So and that somewhere what we need to have is as a DNA, when we are not [INAUDIBLE],, is for to be sure that you are accessed to when your customer or prospect or partner– you have access to all innovation– innovation from Google by discussing with product manager, discussing with all of people based in Amsterdam or in Europe, and also to our partners or ecosystems So we are also looking always for new partnership in order to bring more value to you or in terms of benefits for your user or for your customers Looking at always on that principle I love is having– since I work for traditional companies, I would say, I love to see that we are not looking contractually, technically, commercially the business with you So what we are trying to do is to always be very open in terms of platform, in terms of conceptualization and be sure that when you hopefully you work with us, and we have a partnership, this partnership is not something like one unique, a mono partnership It has to be a [INAUDIBLE] partnership We understand perfectly well that you should have– and maybe is that an advice from me– you should have a multi-cloud strategy in order to be sure that you can hopefully manage abilities for [INAUDIBLE] and leverage what you need, OK? And so lastly, so something quite new because when I look at the first years I was working at Google, I was worried that we didn’t have professional services So at the beginning, so we are adding some support Now fully we have a professional services organization, that not a professional services organization that is for replacing the partners This is for completing the partners or just to be sure that you feel secure, OK, or you feel that your project will go to success, OK? And that just for that, OK? The purpose of this professional service is that we want to be big enough to be good– not good, great– in terms of impact for you, but not too big to not create some uncomfortable for the partners ecosystem is part of the plan fully of the new Google Cloud that you want to have for you So my conclusion– I think that the first time in advance It’s 1 minute, 20 [CHUCKLES] Only one thing I would love that you keep is if you are in the process of integration with Google Cloud today, one thing you should just open for conversation, for Discovery Workshop, OK? Work with the team I’m sure you know most of the people in the Amsterdam office or the [INAUDIBLE] office Start a discussion for Discovery Workshop, and identify the project where we could help you, OK? Maybe problem that you don’t know how to solve Maybe we have a solution And that free OK, that absolutely free We can invest some time with you in order to be sure that you can get some plan of [INAUDIBLE],, OK?
So no need for a big project We can start small discussion Look at what you need short term And after, we learn by walking through it all So I’m done with my presentation I told you, it’s short 20 seconds left I understand that we have also cocktail Andre probably would present that So we’ll have some glass of wine– hopefully good wine– have good wine? Not sure Yeah [CHUCKLES] Good wine So I’m happy to follow up with you, OK? AUDIENCE: [INAUDIBLE] ERIC HADDAD: Discussion– happy to discuss with you Bye-bye [APPLAUSE] ANDRE HOEKZEMA: Thank you, Eric And thank you, everybody And I hope the wine is good And this brings us to the end of this day And I really hope you enjoyed and got inspired today with the keynotes, with the machine learning, with all the possibilities that we brought I really hope that you got inspired by imagining the future like you did And I hope you got inspired by exploring possibilities with the design thinking and all the things that we discussed I saw some great examples of that And I don’t want to close, really Actually instead, I want to do two proposals– whoops [CHUCKLES] I want to do two proposals One– stay connected So all the folks here on the tables, right? So we have customer engineers with Lee And we have here the business people with Petra Every table has them We have our services organization We have our partner organization We have our marketing organization And by the way, super thank you for Susana and Shelby for pulling this off [APPLAUSE] Thank you I want to propose, stay connected And start a project in a safe environment because we believe at Google to just create a safe environment And experiment stuff Just start stuff Just do maybe some stuff that’s already bubbled up at the tables here Let’s continue Let’s continue the discussion Now we are all connected And let’s keep the conversation going in a safe environment And it’s OK to fail Don’t worry And do it a few times You know how that works, right? And at some point, the future you imagined might be closer, OK? The second proposal is– and Eric already gave a little bit a heads up– we have a boat, right? And the boat is park– you can get on the boat from 5 o’clock So if you walk easy to the boat, and there will be Lollipop girls and boys And they will bring us to the boat And we have an awesome time till 7:00 on the boat We will have nice drinks And I hope the wine is good Really, Eric, I hope so [CHUCKLES] So thank you again AUDIENCE: [INAUDIBLE] [APPLAUSE]