Kevin Scott is the chief technology officer for Microsoft, a company that built its business on software, but he also appreciates hardware and devices. In fact, he has been building his own machine shop and hardware lab as a personal hobby in his spare time, looking to demonstrate the potential of combining hardware with the latest software and cloud technologies.
One of his upcoming projects, for example, is a vacuum-siphon coffee machine with a vision and speech interface.
“No buttons,” he said. “The user interface will be all AI.”
Sorry, coffee lovers, but it’s not an upcoming Microsoft product.
“For the love of God, don’t tell Satya I’m doing this,” Scott joked, referring to Microsoft CEO Satya Nadella.
However, the project does reflect a broader trend that Scott, Nadella and Microsoft believe will drive new waves of innovation in the years ahead: the confluence of machine learning, artificial intelligence, 5G mobile broadband, cloud technologies and a growing number of sensors and IoT devices around the world.
“You’ve got all of this opportunity for places where you can interface with the physical world,” Scott said. “It’s just where business opportunity is going to live.”
That was one of the takeaways from Scott’s talk at the GeekWire Cloud Summit. The Microsoft CTO was previously vice president of engineering and operations at LinkedIn, led engineering and operations at mobile advertising company AdMob, and worked two stints at Google, starting as a senior engineering manager in 2005. He’s also the host of the podcast Behind the Tech and is involved in diversity initiatives both inside and outside the company.
Listen to our conversation with Microsoft CTO Kevin Scott in the podcast above, watch the video below, or subscribe in your favorite podcast app. Continue reading for edited highlights.
The confluence of trends that he’s watching closely today: There are a bunch of things that are really important right now where you have a convergence of a bunch of technological trends. We have the end of Moore’s Law rapidly approaching. The important thing that’s already gone is Dennard scaling, which says that you can put more transistors onto a chip and actually power them without burning your dyes up. We’re already at the point where the density improvements we get in silicon aren’t as usable as they were when Dennard scaling was actually working.
That’s coming at the same time that we have these AI workloads — even what’s happened over the past six to nine months has just been mind-boggling. We’re in this mode where the quality of the results that you can compute with the most ambitious flavors of these machine learning systems, whether it’s simulation-based reinforcement learning or pure unsupervised learning, is entirely a proportion of how much compute you can throw at them. You have this new workload that is hungry for compute in ways that nothing ever has been before. That is a really interesting set of factors. It has implications for how we build next generation compute and networks and software stacks. What you’re going to be able to do with all of this AI is just mind-blowing in ways that are hard to predict.
Then you also have this other really interesting thing where IoT and sensing are accelerating because actually Moore’s Law is still working there. Most of these low cost micro-controllers and microprocessors are still on older process technologies. We’ve got another couple or three doublings of price-performance coming, and so more and more of the world is going to get instrumented and you’re going to be able to take these data-enriched AI powered feedback loops that have built the consumer internet, for instance, and that’s going to migrate into all sorts of things like industrial controls and home automation.
Unsupervised learning and reinforcement learning probably sounds far off in the distance and some of the stuff about IoT probably sounds like it’s far off in the distance. The insight that I have is, don’t think of it as far off in the distance. Especially on the machine learning front, that needs to be a part of every company and every developer’s toolkit. Progress is happening so quickly that if you wait too long, you’re going to be in this position where you’re scrambling to catch up.
The state of machine learning and artificial intelligence: The thing that’s really changed across the entire AI community over the past six months is there have been a couple of astounding breakthroughs in unsupervised learning for natural language processing, natural language understanding. We have figured out a new way to build models that build some sort of conceptual notion of what natural language looks like, and from that, you can build a bunch of very specific applications that are higher performing and it’s much lower cost to adapt the general model to a use case and the use cases are all over the place, from question answering to document summarization, to search, to literally there are a dozen or more things that folks are looking at right now and so we are just starting to see these things sort of trickle out into the product portfolio.
Practical applications of artificial intelligence: AI is not a product. It is a technology that you use to build products. We have a different and better way now to build the features of a whole bunch of products and so the way it will show up is not in some sort of Big Bang thing. It’s going to be that a bunch of the features of a whole lot of products are going to get better in material ways and some of it is, you’re just not going to be aware what’s going on. Content moderation on the Xbox platform, for instance, has got a whole bunch of this tech in it and it’s being used to try to prevent bullying and bad behavior on the platform and to protect kids, for instance. It’s just much easier to scale those systems when they’re powered by AI and enhanced by human beings.
The intersection of software and hardware: The interface between software and the physical world is extremely interesting. The real world is not governed by rules that are as logical and straightforward as the rules that we have for software systems. That’s not to diminish how complex software actually is, but machine learning gives us a bunch of powerful new tools and techniques to deal with that complicated, messy, interface. You can’t expect an average tinkerer to sit around and understand these nonlinear partial differential equations that characterize a physical system, whereas you may be able to give them a set of tools like a machine learning system that can give them enough control over that physical thing that they can do something interesting with it. You’ve got all of this opportunity for places where you can interface with the physical world. I think it’s just where business opportunity is going to live.
The potential of mobile broadband: What 5G does is it makes the edge platforms even more interesting. You get lower latency and higher throughput. You still will have to have different solutions for rural areas because the 5G technologies require higher densities of towers than we even have right now with 4G, but there are solutions. 5G presents this opportunity for us to really think about the architecture of these edge applications in different ways. For instance, one of the things that we see already emerging is these computer vision models. On a constrained compute edge device, you’ll be able to run one level of inference. You can do generic event detection, for instance, on the edge and then refer those interesting events back up to the cloud to do a heavyweight inference to actually classify an object or whatnot. 5G just lets you draw those lines differently, and you can build much more interesting applications.
The state of privacy in tech: I was lucky enough to be an engineer and I think many engineers are very aware of what’s going on inside of all of these systems and have very strong opinions about privacy and security in the first place. I’ve always thought that it’s extraordinarily important at AdMob and I was an early Google employee. We took all of these things deadly, deadly seriously, just in terms of access controls and whatnot. But if anything, it gets more and more important every day. AdMob was 2007-2010. Nine years later, the world of data that is residing in clouds is many orders of magnitude larger, and presents this very interesting challenge for folks. The digital world has become more complex. The footprint you’re leaving as you use digital services has grown much larger as we spend more time online, and so I think the onus is on all technology companies, large and small, that are handling people’s data. You’ve got some regulatory things that you have to comply with, like GDPR. I think more of those regulations are coming and I think that’s a good thing for consumers.We as an industry need to go above and beyond because we have early line of sight on what the trends are with privacy and security as well.