Unlocking AI Innovation: A Conversation with Eric Boyd and Asha Sharma | Studio06 (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:10) SETH JUAREZ: Hello, my friends. (00:00:12) Welcome back. (00:00:12) This is the Build Live stage. (00:00:15) My name is Seth, and no pressure for (00:00:17) me because this is my boss. (00:00:20) Then my boss' boss. How's it going, my friends? (00:00:26) ERIC BOYD: Very well. How are you? (00:00:27) SETH JUAREZ: Eric and Asha, (00:00:28) why don't you introduce yourself? (00:00:29) We'll start with you and then Asha. (00:00:30) ERIC BOYD: I'm Eric Boyd. (00:00:31) I lead the AI platform team. (00:00:33) ASHA SHARMA: Asha, I lead product (00:00:34) for the AI platform team. (00:00:36) SETH JUAREZ: Let's start with (00:00:37) what is the now (00:00:38) of AI because there's a lot of stuff going on, (00:00:40) and maybe you can help us rationalize it. (00:00:42) Want to get your thoughts and (00:00:43) then Asha on product, how that looks. (00:00:45) ERIC BOYD: There's so much (00:00:46) going on in AI. (00:00:47) I think some of the most exciting things to (00:00:49) think about are what are our customers doing? (00:00:52) If I start with Mercedes Benz, (00:00:55) everyone's been in a car and has (00:00:57) tried to talk to that car voice interface, (00:00:59) and it doesn't understand (00:01:00) you and can't say anything back. (00:01:01) Mercedes Benz integrated with GPT-4 for (00:01:04) their MBUX and so now you can (00:01:06) have a much more natural conversation in the car, (00:01:08) being able to get directions and get things done. (00:01:11) They even integrated GPT-4 vision into the dash cam, (00:01:14) so now they can see (00:01:15) things going on around the car and near you. (00:01:18) It's so exciting, things like that. (00:01:20) You can go from car manufacturers (00:01:23) all the way to developers with the unity platform, (00:01:26) which has now made a bot that can help (00:01:28) developers figure out all the things (00:01:30) that you're doing on Unity. (00:01:31) Unity is a platform for making 3D games. (00:01:33) Helping you figure out all the APIs (00:01:35) and how everything works, (00:01:37) all within the unity environment and not having to leave. (00:01:39) When you look at those two extremes of just (00:01:42) how much these applications are just all over the place, (00:01:45) it just shows you how quickly companies are (00:01:47) adopting AI across the spectrum. (00:01:50) We're just building tools as (00:01:52) quickly as we can to try and serve those needs. (00:01:55) Companies need things like search to be able to (00:01:58) get their data in and store it. (00:02:00) Azure AI Search, which is (00:02:02) now 12 times as much storage for the same price, (00:02:05) really all integrated in. (00:02:07) The Azure AI Studio, making it easy. (00:02:09) There's just so much going on in AI (00:02:11) just really bringing all those applications together. (00:02:13) It's really fun and exciting to see. (00:02:15) ASHA SHARMA: It almost feels (00:02:16) like applications today are (00:02:17) feeling increasingly broken when (00:02:19) they don't have AI as part of it. (00:02:20) We're starting to see (00:02:22) organizations go from experimentation (00:02:24) now to scale and (00:02:26) even think about customizing the experiences. (00:02:28) A lot of the tools that we're working (00:02:30) on are going towards, how do they fine tune? (00:02:33) How do they leverage data to (00:02:35) ground their application and things like that. (00:02:37) SETH JUAREZ: What are (00:02:37) some cool things that people are (00:02:39) doing now that you're getting a sense for? (00:02:42) Because we talk about AI all the time, (00:02:44) and everyone's like, AI, AI. (00:02:46) But what are some things that (00:02:47) people are doing to sort of soften (00:02:49) the edge around software and hardware and people? (00:02:53) ERIC BOYD: You just see it in (00:02:54) all the different use cases. (00:02:55) I talked about a couple where we (00:02:57) see what Mercedes Benz and Unity is doing. (00:03:00) But there are so many other examples where they're (00:03:04) trying to build solutions to (00:03:05) make it easier to interact with people. (00:03:07) I love the example Satya gave in (00:03:09) his keynote this morning at the Khan Academy. (00:03:11) The ability to have an individualized education tutor (00:03:15) really connecting directly with each student is (00:03:19) really powerful and the types of (00:03:21) things that they are the applications you (00:03:23) couldn't even have really dreamed of (00:03:25) a few years ago until AI really got to this point. (00:03:28) SETH JUAREZ: What are your thoughts? (00:03:28) ASHA SHARMA: Yeah, I think multimodal (00:03:30) is becoming a big aspect of that. (00:03:32) Now with the power of GPT-4o, (00:03:34) which is GA today on Azure OpenAI and Azure AI Studio, (00:03:38) I think there's going to be a new world (00:03:40) of possibilities to actually (00:03:41) interact with your AI (00:03:43) much more fluently than ever before. (00:03:46) SETH JUAREZ: I love that because they (00:03:48) let me play with these models sometimes, (00:03:50) and they're actually really fun (00:03:51) to imagine the things that you can do. (00:03:54) We started talking about GPT-4 Omni, (00:03:57) but there's a lot of other models (00:03:58) that we're looking at right now. (00:04:00) Eric, why don't we start with some (00:04:01) of the cool things that are out today? (00:04:03) Then Asha maybe some ways (00:04:04) that we're helping people use those things. (00:04:06) ERIC BOYD: Some of the other things (00:04:08) that we're excited to see today, (00:04:10) we announced the Phi series of models, Phi-3. (00:04:13) We've got Phi-3-mini, (00:04:14) small, and medium, (00:04:15) which are 3, 7, (00:04:17) and 14 billion parameter models. (00:04:19) The thing about these Phi series models (00:04:21) is we took a different approach to training them, (00:04:24) where we really tried to think of (00:04:26) the data that they train on as a curriculum, (00:04:29) like you would give to a student where you're trying to (00:04:31) teach the model step by step, (00:04:33) how do we make it learn everything about the world. (00:04:37) In doing that, we got a model that (00:04:39) is much smaller and more powerful. (00:04:41) All the Phi-3 models, (00:04:43) they basically punch a full weight class (00:04:45) ahead for their size, (00:04:46) which is really interesting as (00:04:48) customers are looking for what's (00:04:50) the most power that I can get from (00:04:51) a model at the right price point. (00:04:54) Looking to select all sort (00:04:55) of along that price performance curve, (00:04:57) what are the places (00:04:58) where the model is going to work really well? (00:05:00) It's exciting to see what we've done with that. (00:05:02) We also added, Asha is talking about multimodal. (00:05:06) Phi-3 for Vision is also available today. (00:05:08) Now you can in these smaller model still get (00:05:11) those same rich vision capabilities that (00:05:13) you're used to seeing with GPT-4V and GPTo. (00:05:16) Really excited to see that. (00:05:18) ASHA SHARMA: We also expanded (00:05:19) our partnership with Hugging Face, (00:05:20) so there's a lot of open source models. (00:05:23) We are also doing a lot around the tool chain. (00:05:25) I think that's what's most special. (00:05:27) Eventually, we have to teach people to get (00:05:29) the right model for the right job at the right cost. (00:05:32) If they can do that, then it's about (00:05:33) the tool chain to build those applications. (00:05:35) We announced the general availability of Azure AI Studio, (00:05:39) but we're also announcing a lot around responsible AI to (00:05:42) keep those applications safe and secure end to end, (00:05:45) so content filters, prompt shields, (00:05:48) all of those things are going to become (00:05:49) even more important as we start to scale. (00:05:51) As we look at this from a perspective of developers, (00:05:57) I might be like, oh, I got to use AI. (00:05:59) I don't know what to do. (00:06:00) Because I've been to conferences around the world, (00:06:02) and people are literally being told by their management, (00:06:06) you need to use AI, (00:06:07) and they're like, what? (00:06:09) What is some of the stuff that we provide to (00:06:12) help developers get started? (00:06:14) We'll start with you then Asha. (00:06:15) ERIC BOYD: There are (00:06:15) all these different tools that (00:06:17) we make to make things easier. (00:06:20) You can go to Azure AI Studio, (00:06:22) which is the integrated place to go (00:06:24) and build your generative AI applications. (00:06:26) The most common pattern that we're (00:06:29) seeing is this retrieval augmented generation, (00:06:32) the rag pattern of how do I take (00:06:34) some data and put it into a search engine, (00:06:37) usually one that has vector search. (00:06:38) That's why Azure AI Search is (00:06:41) super helpful for building those applications. (00:06:43) It's a really rich search engine that (00:06:45) brings all the power of vector search with (00:06:48) all the things that we've learned over decades of doing (00:06:50) keywords based search into a single platform, (00:06:53) and then has a semantic ranker on top of (00:06:55) them to really retrieve the most relevant results. (00:06:58) Then you feed that to your model that then (00:07:01) the model now has all of (00:07:02) the information for your application, (00:07:04) just all that built into it. (00:07:06) That's the pattern that (00:07:07) developers are really going after. (00:07:09) How do we make it easy for them to build that, (00:07:12) to put all those pieces together, (00:07:14) and then to start experimenting (00:07:15) and iterating and learning, (00:07:17) how do I tweak this to get the best quality out of it. (00:07:20) ASHA SHARMA: It seems like some of (00:07:21) the more sophisticated developers, (00:07:22) once they do that, (00:07:23) they're starting to turn to tools (00:07:25) to actually help them solve complex tasks, (00:07:27) so Assistants, Assistants API, (00:07:30) which we have for Azure OpenAI. (00:07:32) That's becoming increasingly more popular. (00:07:33) I saw Coca Cola is doing (00:07:35) something really cool where they have (00:07:36) Assistant API templates for every single department, (00:07:40) so it makes it much easier than (00:07:41) for a business decision maker to actually (00:07:43) get started with the tools in a really non scary way. (00:07:48) SETH JUAREZ: The cool end. (00:07:49) Did you want to say something? (00:07:49) ERIC BOYD: You can say, (00:07:50) you even mentioned prompt shields, (00:07:51) which we also didn't talk about is like. (00:07:53) The whole capability of building (00:07:55) these tools responsibly and safely. (00:07:57) We provide the Azure Content Safety system, (00:08:01) which is built into the Azure OpenAI service, (00:08:03) and you can then add on to any (00:08:04) other model that you want. (00:08:06) But it gives you the full set (00:08:07) of safety tools that you might want, (00:08:09) things about how to give different levels of control (00:08:12) around violence or sexual content (00:08:15) or how do you want to manage that, (00:08:16) all the way to reducing hallucinations, (00:08:19) to managing prompt shields, (00:08:21) which is really controlling (00:08:22) jail break and things like that. (00:08:23) All those are just making the developer's life (00:08:25) easier as they go and build these applications. (00:08:27) SETH JUAREZ: I want to lean in here just (00:08:29) a little bit before we get to some of (00:08:30) the other questions about safety and responsibility. (00:08:33) Can you tell me how Microsoft is taking (00:08:35) this particular task seriously (00:08:38) in helping developers do it as well, Asha. (00:08:40) ASHA SHARMA: Well, we've been (00:08:41) working on it (00:08:41) for the last eight years. (00:08:43) We think about this as (00:08:45) cutting edge technology needs cutting (00:08:47) edge responsibility and we're still pretty early in it. (00:08:50) But after eight years, (00:08:51) we have 90 tools, hundreds of features. (00:08:54) We do a lot of the work ourselves. (00:08:57) Before we release a model, (00:08:58) there's a lot of work that we do to (00:09:00) provide the Azure promise, (00:09:01) but then we also provide tools so (00:09:03) developers can actually customize their need. (00:09:05) Custom Filters is a big feature that we released today, (00:09:09) and that's allowing you to then set (00:09:11) up different tools for moderation, (00:09:13) for things that maybe we haven't even thought about yet. (00:09:15) ERIC BOYD: That's right. (00:09:16) SETH JUAREZ: That's really cool. (00:09:17) The thing that I (00:09:18) think now I want to dive into, (00:09:20) if it's possible, is the notion of, (00:09:22) what are customers doing with? (00:09:24) I know we spent a lot of time talking (00:09:26) about here's all this good stuff, (00:09:27) goodness that we have, but what are customers doing? (00:09:29) We'll start with you, Eric, and then (00:09:31) maybe Asha tell us about that. (00:09:32) ERIC BOYD: We just keep telling all (00:09:33) of the customer stories of (00:09:34) all the amazing things that they're able to (00:09:36) go and deliver on top of these tools, (00:09:38) and it's been in every industry (00:09:41) that we've really worked with. (00:09:43) You see obvious cases in customer support (00:09:45) where people are trying to build customer support bots (00:09:48) to help answer users questions or (00:09:50) even for their own agents to (00:09:52) go and help them develop that. (00:09:54) But I think the interesting things (00:09:56) are the new applications, (00:09:57) the things that we just couldn't imagine (00:09:59) being lit up previously and now seeing (00:10:02) how those are coming to light (00:10:03) with the way the customers are (00:10:04) pulling all these different tools (00:10:06) together in incredible applications. (00:10:08) Some of the most interesting things are this new space, (00:10:12) of having multiple agents talk to (00:10:14) each other and sort of string (00:10:16) different conversations together. (00:10:17) I think that's a really interesting space (00:10:19) where we're starting to see (00:10:21) the standard computer science abstractions (00:10:24) and separation of control, (00:10:25) that's starting to materialize into (00:10:27) these much richer applications. (00:10:30) ASHA SHARMA: We now have more than (00:10:32) 53,000 customers that are on Azure AI. (00:10:35) One of the things that we're thinking deeply about is, (00:10:37) how do we start to do safety and security by default. (00:10:40) Again, removing as much work (00:10:42) as possible from our customers. (00:10:43) The other interesting thing is that (00:10:45) Microsoft runs on Azure AI. (00:10:47) All of the copilots, (00:10:49) all the millions of users, (00:10:50) we're getting a first class ride at learning (00:10:53) and using the technology for our customers internally, (00:10:56) and then we put it in the platform and we (00:10:57) extend it to our third party customers. (00:10:59) SETH JUAREZ: Microsoft has always been (00:11:01) customer number zero? (00:11:02) ERIC BOYD: Literally from the start (00:11:04) when we first started (00:11:06) building out our capabilities (00:11:08) with the Azure OpenAI service. (00:11:10) We first launched Bing Chat, (00:11:12) and then we launched M365 Copilot and GitHub Copilot, (00:11:15) even 18 months before that. (00:11:17) We've built all of that as Asha said, (00:11:19) on the same platform (00:11:20) that we make available to our customers. (00:11:22) They know we've literally (00:11:24) bet all of Microsoft business on it, (00:11:25) so they can trust it with their business as well. (00:11:27) SETH JUAREZ: That's awesome, because, like (00:11:29) I said, I I use this stuff, (00:11:31) and I know what it does. (00:11:33) Every once in a while, it's (00:11:35) pretty surprising what you can do. (00:11:38) As people are getting immersed in the rag pattern (00:11:41) and doing that development (00:11:43) and they're getting out to work. (00:11:44) What are some things that people should look (00:11:45) forward to or how should they (00:11:47) orient their thinking when it comes to (00:11:49) experiences that they may be able to develop? (00:11:52) ASHA SHARMA: One of the big things is, (00:11:54) I feel like we have to just be thoughtful to (00:11:56) not do AI for AI's sake. (00:11:58) It has to create real values. (00:11:59) It has to improve the way you live, (00:12:00) improve the way you work. (00:12:02) There was a study that showed that skilled workers that (00:12:04) use GenAI are seeing (00:12:05) a 40 percent improvement in the way that they work. (00:12:08) We are constantly thinking about (00:12:10) that from a product perspective, which is, (00:12:11) how do we build tools to help (00:12:13) developers solve real problems. (00:12:15) I would say, don't lose sight of that. (00:12:17) The second thing to look forward to is, (00:12:19) and we showed this in one of our sessions. (00:12:21) We're moving much more code first, (00:12:23) so we want to meet developers where they are. (00:12:25) More and more, we're seeing (00:12:26) all applications becoming AI applications, (00:12:28) and we have the best IDs in the world. (00:12:30) We want to make that really (00:12:32) simple to wire up your applications, (00:12:34) change out your models, (00:12:35) and do so without any thrash. (00:12:37) ERIC BOYD: I would add, too. (00:12:39) Asha touched on multimodality earlier. (00:12:41) This is something that we've been (00:12:42) seeing coming for a while. (00:12:44) But really, with GPT-4.0, (00:12:46) it really brought home just how (00:12:48) powerful this could be to have (00:12:50) that much fluency with voice (00:12:52) and talking to a model and having it talk back. (00:12:56) There's these inflection points on (00:12:58) all technologies where suddenly you get to (00:13:00) a point where it's good enough that it (00:13:03) opens up scenarios that didn't use to previously work. (00:13:05) In the early 2000s, (00:13:08) speech recognition had to be word at a time, (00:13:11) and now it's much more natural for me to talk to (00:13:13) my phone than to swipe text messages. (00:13:16) I think this level of voice fluency (00:13:19) is going to change the way that we interact with (00:13:22) machines where we'll just (00:13:24) talk much more naturally in our voice as opposed to (00:13:26) doctoring things to machine (00:13:28) speak and making it work that way. (00:13:30) I think that combined with vision, (00:13:32) is going to be really (00:13:33) interesting the types of things that we can do. (00:13:34) ASHA SHARMA: Now I was just going (00:13:35) to say, I agree. (00:13:36) In many ways, we've taught ourselves to (00:13:37) prompt engineer queries for search. (00:13:40) Your demo this morning was pretty (00:13:42) amazing because it was just you (00:13:44) expressing yourself to an application that now becomes (00:13:48) much smarter over time and can help you in (00:13:50) any mode that you want in any language that you want. (00:13:52) SETH JUAREZ: For those wondering, (00:13:53) that actually I coded that up. (00:13:54) That's a real thing. (00:13:55) ASHA SHARMA: I noticed (00:13:56) SETH JUAREZ: It's pretty amazing because, (00:13:58) as I was interacting with it, (00:14:00) I forgot the interface. (00:14:04) It wasn't a keyboard thing anymore. (00:14:07) Where should people go to get (00:14:08) started with this stuff, Asha? (00:14:11) ASHA SHARMA: Azure Studio is (00:14:12) the best place. (00:14:13) We have a playground, so you can just start to (00:14:14) dabble with the models and (00:14:16) then you can consume the models. (00:14:17) That's a great place to get started. (00:14:19) SETH JUAREZ: Any final thoughts from you, (00:14:20) Eric and then Asha. (00:14:22) ERIC BOYD: There's so much happening (00:14:23) in AI these days, (00:14:24) that the main thing for developers (00:14:26) is they just have to get their hands wet, (00:14:28) like just dig in (00:14:29) and start building something and just start (00:14:31) to see what you can do with it (00:14:33) because until you have that experience, (00:14:35) until you're really seeing how things (00:14:36) compose and how you can use it. (00:14:39) You're just not going to have that full understanding. (00:14:41) That's the main thing I would encourage people to do is, (00:14:43) go try these tools, (00:14:45) go build applications, and you'll (00:14:47) be surprised at the power (00:14:48) that you can really create with things. (00:14:50) SETH JUAREZ: Asha for your last word. (00:14:51) ASHA SHARMA: He took it. (00:14:52) I feel like AI just (00:14:53) say get second. That's okay. (00:14:55) No. I think that we just need (00:14:58) to imagine the world as (00:15:00) every application becoming an AI application. (00:15:03) I would say, think at a much bigger scale (00:15:05) than what we have before. (00:15:06) SETH JUAREZ: Well, this has been amazing, (00:15:07) and thank you so much for letting me work (00:15:09) with you on these cool models. (00:15:11) It's really exciting. Thank you so much, (00:15:12) and we'll see you after this.