 Live from New York, it's theCUBE. Covering AWS Summit New York 2018. Brought to you by Amazon Web Services and its ecosystem partners. Hello everyone, welcome back here live. Cube coverage in New York City for AWS Amazon Web Services Summit 2018. I'm John Furrier with Jeff Rick here at theCUBE. Our next guest is Dr. Matt Wood, General Manager of Artificial Intelligence with Amazon Web Services, Cube alumni. Been so busy for the past year. I haven't been on theCUBE in a year. Thanks for coming back. I appreciate you spending the time. So promotions keep on going on. You got now General Manager of the AI Group, AI Operations, AI Automation, Machine Learning Office. A lot of big category of new things developing in AI. You guys have really taken AI and machine learning to a whole other level. It's one of the key value propositions that you guys now have for not just a large enterprise but down to startups and developers. So congratulations and what's the update? Oh, well, the update is this morning in the keynote, I was lucky enough to introduce some new capabilities across our platform. When it comes to machine learning, our mission is that we want to be able to take machine learning and make it available to all developers. We joke internally that we just want to, we want to make machine learning boring. We want to make it vanilla. It's just, it's another tool in the tool chest of any developer and any data scientist. And we've done that, this idea of taking technology that is traditionally only within reach of a very, very small number of well-funded organizations and making it as broadly distributed as possible. We've done that pretty successfully with compute, storage and databases and analytics and data warehousing. And we want to do the exact same thing for machine learning. And to do that, we had to kind of build an entirely new stack. And we think of that stack in three different tiers. The bottom tier really for academics and researchers and data scientists, we provide a wide range of frameworks, open source programming libraries that developers and data scientists use to build neural networks and intelligent systems. They're things like TensorFlow and Apache MXNet and PyTorch. And they're really, they're very technical, but you can build arbitrarily sophisticated systems. Mostly open source too, right? Mostly open source, that's right. We contribute a lot of our work back to MXNet, but we also contribute to PyTorch and to TensorFlow. And there's big, healthy open source projects growing up around all of these popular frameworks, plus more like Keras and Gluon and Horovod and all these other things. So that's a very, very, it's a key area for researchers and academics. The next level up, we have machine learning platforms. This is for developers and data scientists who have data they see in the cloud, or that they want to move to the cloud quickly, that they want to be able to use for modeling. They want to be able to use it to build custom machine learning models. And so here we try and remove as much of the undifferentiated heavy lifting associated with doing that as possible. And this is really where SageMaker fits in. So SageMaker allows developers to quickly fill, train, optimize and host their machine learning models. And then at the top tier, we have a set of AI services which are for application developers that don't want to get into the weeds. They just want to get up and running really, really quickly. And so today we announced four new services really across those, their middle tier and their top tier. So for SageMaker, we're very pleased to introduce a new streaming data protocol which allows you to take data straight from S3 and pump it straight into your algorithm and straight onto the compute infrastructure. And what that means is you no longer have to copy data from S3 onto your compute infrastructure in order to be able to start training. You just take away that step and just stream it right on there. And it's an approach that we use inside SageMaker for a lot of our built-in algorithms. And it significantly increases the speed of the algorithm and significantly, of course, decreases the cost of running the training because you pay by the seconds. So any second you can save off, it's a cost saving for the customer. And also it helps machine learn more. That's right. Yeah, you can put more data through it, absolutely. So you're no longer constrained by the amount of disk space. You're not even constrained by the amount of memory on the instance. You can just pump terabyte after terabyte after terabyte. And we actually had another thing that I taught about in the keynote this morning, a new customer of ours, Snap, who are routinely training on over 100 terabytes of image data using SageMaker. So the ability to be able to pump in lots of data is one of the keys to building successful machine learning applications. So we brought that capability to everybody that's using TensorFlow. Now you can just have your TensorFlow model, bring it to SageMaker, do a little bit of wiring, click a button, and you'll just start streaming your data to your TensorFlow algorithm. What's the impact of the developer? Time, speed, money? It is the ability to be able to pump more data. It is the decrease in time it takes to start the training. But most importantly, it decreases the training time all up. So you'll see between a 10 and 25% decrease in training time. Some of these you can train more models or you can train more models in the same unit time or you can just decrease the cost. So this is a completely different way of thinking about how to train over large amounts of data. We were doing it internally and now we're making it available for everybody through SageMaker. So that's the first thing. The second thing that we're adding is the ability to be able to batch process in SageMaker. So SageMaker used to be great at real-time predictions. But there's a lot of use cases where you don't want to just make a one-off prediction. You want to predict hundreds or thousands or even millions of things all at once. So let's say you've got all of your sales information at the end of the month. You want to use that to make a forecast for the next month. You don't need to do that in real-time. You need to do it once and then place the order. And so we added batch transforms to SageMaker so you can pull in all of that data, large amounts of data, batch process it within a fully automated environment and then spin down the infrastructure and you're done. So very, very simple API. Anyone that uses a Lambda function can take advantage of this. Again, just dramatically decreasing the overhead and making it so much easier for everybody to take advantage of machine learning. And then at the top layer, we had new capabilities for our AI services. So we announced 12 new language pairs for our translation service and we announced new transcription capability which allows us to take multi-channel audio such as might be recorded here. But more commonly on contact centers, just like you have a left channel and a right channel for stereo, contact centers often record the agent and the customer on the same track. And today you can now pass that through our transcribe service. Long form speech will split it up into the channels automatically, transcribe it, will analyze all the timestamps and create just a single script. And from there, you can see what was being talked about. You can check the topics automatically using Comprehend or you can check the compliance. Did the agents say the words that they have to say for compliance reasons at some point during the conversation? So that's a material new capability for a transcribe. What's the top services being used? Obviously Comprehend, transcribe, and a variety of others. You guys have put a lot of stuff out there. All kinds of stuff. What's the top sellers top usage as a proxy for uptake adoption? Yeah, I think we see a ton of adoption across all of these areas but where a lot of the momentum is growing right now is SageMaker. So if you look at Formula One, they just chose Formula One racing. They just chose AWS and SageMaker as their machine learning platform. But National Football League, Major League Baseball today announced that they're re-upping their relationship and their strategic partnership with AWS for machine learning. So all of these groups are using the data which just streams out of these races or these games. And that can be the video or it can be the telemetry of the cars or the telemetry of the players. And they're pumping that through SageMaker to drive more engaging experiences for their viewers. So guys, hey, streaming this data, let's get this into SageMaker quickly, this cute video. Yeah, just get it all in there, all of it. Well, you know we love data, we'd love to follow up on that. So the question is, is that when will SageMaker overtake Aurora as the fastest growing product in history of Amazon? Because I predicted that re-invent that SageMaker would go on a tear. Is it looking good right now? I mean, Aurora's still on paper. You guys are saying is that the best seller? Yeah, Aurora's still the fastest growing, but SageMaker, give us some indicator. Well, I mean, we don't break out revenue but even the same excitement, I'll say this, the same excitement that I see for SageMaker now and the same opportunity and the same momentum, it really, really reminds me of AWS 10 years ago. It's the same sort of transformative democratizing approach to, which really engages builders. And I see the same level of, And the excitement at levels are super high as well. They're super high in general, there's a little hype out there, but I see the same level of enthusiasm and momentum. People are building with it basically. Absolutely. So what's this toy you have here? I know we don't have a lot of time, but you got a little problem. This is the world's first deep learning enabled wireless video camera. We call it DeepLens. We announced it and launched it at re-invent 2017. Hold it up, hold it up for the camera. I can hold it up to the camera. It's a cute little device. We modeled it after Wally, the Pixar movie. And it is a HD video camera on the front here. And in the base here, we have an incredibly powerful custom piece of machine learning hardware. So this can process over a billion machine learning operations per second. You can take the video in real time. We send it to the, it's got a GPU on board and we'll just start processing the stream in real time. So that's kind of interesting, but the real value of this and why we designed it was we wanted to try and find a way for developers to get literally hands-on with machine learning. So the way that builders are lifelong learners, right? They love to learn, they have an insatiable appetite for new information and new technologies. And the way that they learn that is they experiment. They start working and they kind of spin this flywheel where you try something out, it works, you fiddle with it, it stops working, you learn a little bit more and you want to go round and round and round. That's been tried and tested for developers for decades. The challenge with machine learning is doing that is still very, very difficult. You need labeled data, you need to understand the algorithms, it's just, it's hard to do. But with DeepLens, you can get up and running in 10 minutes. So it's connected back to the cloud, it's connected back to SageMaker. You can deploy a pre-built model down onto the device in 10 minutes to do object detection. We do some wacky visual effects with neural style transfer. We do a hot dog and no hot dog detection, of course. But the real value comes in that you can take any of those models, tear them apart inside SageMaker, start fiddling around with them and then immediately deploy them back down onto the camera. And every developer on their desk has things that they can detect. They have pens and cups and people, whatever it is. So they can very, very quickly spin this flywheel where they're experimenting, changing, succeeding, failing and just going round and round and round. So that's for developers, just target audience for this, right? Yeah, that's right. And what are some of the things that have come out of this? Have you seen any cool evolutionary demos? It has been incredibly gratifying and really humbling to see developers that have no machine learning experience. Take this out of the box and build some really wonderful projects. One really good example is exercise detection. So you know when you're doing a workout? They build a model which detects the exercise you're doing and then detects the reps of the weights that you're lifting. And we saw skeletal mapping. So you could map a person in 3D space using a simple camera. We saw security features where you could put this on your door and then it would send you a text message if it didn't recognize who was in front of the door. And we saw one which was amazing which would read books aloud to kids. So you would hold up the book and it would detect the text, extract the text, send the text to Polly and then speak aloud for the kids. So there's games, there's educational tools, there's little security gizmos. One group even trained a dog detection model which detected individual species, plugged this into an enormous power pack and took it to the local dog park so that they could test it out. So it's all of this from a cold start with no machine learning experience. You having fun? Yes, absolutely. If you can't tell, I'm super passionate about this. One of the great things about machine learning is you don't just get to work in one area. You get to work in Formula One and sports and you get to work in healthcare and you get to work in retail and developer tool. I mean, CTO's going to love this. Chief toy officers. Chief toy officers, I love it. Yes. So I got to ask you, so what's new in your world? GM of AI, Audition Intelligence. What does that mean? Just quickly explain it for our audience. Is that all the software? What specifically are you overseeing? What's your purview within the realm of AWS? Yeah, that's a totally fair question. So my purview is I run the products for deep learning, machine learning and artificial intelligence really across the AWS machine learning team. So I get it, I have a lot of fingers and a lot of pies. I get involved in the new products that we're going to go build out. I get involved in helping grow usage of existing products. I get it to do a lot of invention, spend a ton of time with customers. But overall, work with the rest of the team on setting the technical and product strategy for machine learning at AWS. But what's your top priorities this year? Adoption, uptake, new product introductions. I mean, you guys don't stop introducing. Well, we don't stop. We don't stop. I mean, you keep on introducing more and more things. Any high ground that you want to take, what's the vision? I think the vision is to genuinely to continue to make it as easy as possible for developers to use machine learning. I can't overstate the importance or the challenge. So we're not at the point where you can just pull down some Python code and figure it out. We're not even, we don't have a JVM for machine learning where there's no developer tools or debuggers. There's very few visualizers. So it's still very hard if you kind of think of it in computing terms, we're still working in assembly language and machine learning. And so there's this wealth of opportunity ahead of us. And the responsibility that I feel very strongly is to be able to continually improve on that stack to continually bring new capabilities to more developers. Well, cloud has been disrupting IT operations, AI ops, they're calling it Silicon Valley and the venture circles auto ML as a term that's been kicked around auto automatic machine learning. You got to train the machines with something, data seems to be it. What strikes me about this compared to storage or compared to compute or compared to some of the core Amazon foundational products, those were just better ways to do something that already existed. This is not a better way to do something that already exists. This is a way to get the democratization at the start of the process of the application of machine learning and artificial intelligence to a plethora of applications and use cases. That is fundamentally different and just a step up in terms of the power to the hands of the people. It's something which is very far scenario which is very fast moving and very fast growing. But what's funny is it totally builds on top of the cloud. You really can't do machine learning in any meaningful production way unless you have a way that is cheap and easy to collect large amounts of data in a way which allows you to pull down high performance computation at any scale that you need it. And so through the cloud we've actually laid the foundations for machine learning going forward. And other things too coming. So you guys, sets of services, you guys announced the cloud highlights the power that it brings to these new capabilities. Absolutely, yeah. And we get to build on them at AWS and at Amazon just like our customers do. And so SageMaker runs on EC2. We wouldn't be able to do SageMaker without EC2. And in the fullness of time we see that the usage of machine learning could be as big if not bigger than the whole of the rest of AWS combined. That's our aspiration. Dr. Matt Wood, I wish we had more time to chat. Love talking with you. I'd love to do a whole another segment on what you're doing with customers. I know you guys are very customer focused as Andy always mentions when on the cube you guys listen to customers, want to hear that. Maybe reinvent, we'll circle back. Sounds good. Congratulations on your success. Great to see you. Appreciate it, thanks Tom. Dr. Matt Wood here in the cube. We're streaming all this data out to the Amazon cloud as we host all of our stuff. Of course it's theCUBE bringing you live action here in New York City for cube coverage of AWS Summit 2018 in Manhattan. We'll be back with more after this short break.