This week, Microsoft hosted a vastly different Build, its developers conference and biggest event of the year. Build 2020 had plenty of big news for businesses, developers, and business developers. There was unexpected news and expected news. Even Microsoft haters had plenty to discuss. Cloud and AI announcements abounded for good reason. But the highlight of the event was where these all overlapped: a supercomputer in the cloud.
Microsoft’s $1 billion investment in OpenAI to jointly develop new technologies for Azure is bearing fruit. You should read about all the technical details here: OpenAI’s supercomputer collaboration with Microsoft marks its biggest bet yet on AGI. But I want to discuss Microsoft CTO Kevin Scott’s talk on the second day of Build, which I think largely flew under the radar. That’s where Microsoft brought it all together and explained why you should care about self-supervised learning and a supercomputer in Azure.
Scott’s technical advisor Luis Vargas defined Microsoft’s phrase du jour “AI at Scale” as the trend of increasingly larger AI models and how they’re being used to power a large number of tasks. Watch Vargas explain it all:
Romeo and Juliet, Star Trek, and Star Wars all got shoutouts. What’s not to love? The accuracy and especially the speed of the answers that the system spits out are impressive. I encourage you to pause the video and carefully look at the inputs and outputs. Still, this is a staged demo. Microsoft certainly carefully selected the examples, and this year it could pre-record everything.
Halfway through, Scott brought in OpenAI CEO Sam Altman. That demo was even more mind-blowing. You’ll want to tune in at about 28:30.
A year ago, Amanda Silver, CVP of Microsoft’s developer tools, told me that the company wanted to apply AI “to the entire application developer lifecycle.” The conversation was about Visual Studio IntelliCode, which uses AI to offer intelligent suggestions that improve code quality and productivity. At the time, IntelliCode comprised statement completion, which uses a machine learning model, and style inference, which is more of a heuristic model.
OpenAI showed off Microsoft’s supercomputer not just completing code and offering suggestions, but writing code from English instructions. Yes, this is a demo. No, it’s not terribly practical. I’m frankly more interested in tracking IntelliCode’s evolution because helping developers code is more helpful, at least right now, than trying to code for them. Still, this is incredible to see just one year after IntelliCode hitting general availability.
Machine learning experts have largely focused on relatively small AI models that use labeled examples to learn a single task. You’ve likely already seen these applications: language translation, object recognition, and speech recognition. The AI research community has lately shown that applying self-supervised learning to build a single massive AI model, such as the ones shown above trained on billions of pages of publicly available text, can perform some of those tasks much better. Such larger AI models can learn language, grammar, knowledge, concepts, and context to the point that they can handle multiple tasks, like summarizing text, answering questions, or even apparently — if trained on code — writing code.
Microsoft and OpenAI are talking about this cool technology not simply to show it off, but to tease that it’s eventually coming to Azure customers.
In a Q&A session the same day, Scott Guthrie, Microsoft EVP of cloud and AI, answered a question about what is holding back AI. Here is just the first part of his response:
The more compute you throw at AI — part of the reason why we’re building our AI supercomputer as part of Azure is, we definitely see there’s a set of algorithms that as you throw more compute at it, and you do that compute not just in terms of CPU, but also in particular the network interchange, and the bandwidth between those CPUs is just as critical, because otherwise, then the network becomes the limiter. But you see smarter algorithms getting built. I think Kevin Scott will cover it very well in his talk here at Build. About some of the amazing types of problems that we’re able to solve now, that even two or three years ago looked like science fiction, but they’re now here. In terms of text understanding, machine understanding. I think his talk is happening this morning or maybe it just happened, but I don’t want to steal all his thunder, but there’s some really cool demos.
Guthrie understandably didn’t want to spoil who shot first, not to mention that they have a supercomputer writing code.
I think a supercomputer in Azure makes perfect sense. We’re soon going to see a lot more practical use cases than the demos above. This week’s announcement was just the first step — right now that supercomputer is only for OpenAI to use, but Microsoft is going to make those very large AI models and the infrastructure needed to train them broadly available via Azure. Businesses, data scientists, and developers will take that and build.
ProBeat is a column in which Emil rants about whatever crosses him that week.