View Recording: Best of Microsoft Build: What the Announcements Mean for Your Business

Brian Haydin All right. Hey, welcome in everybody. Last week, Sunir and I were standing in front of this sign in Fort Mason in San Francisco for Microsoft Build 2026. Over the two days, I think there were 25 or 30 different announcements. There’s a concert. 0:0:28.237 –> 0:0:47.997 Brian Haydin and a whole bunch of other exciting things. One thing that I thought was really interesting about Microsoft Build this year is that, and I’ve done this Best of Build for five or six years now, is that I’ve always been able to go back to the book of news, and they completely abandoned the book of news this year. And I was telling poor Suneer here, like, hey, 0:0:48.277 –> 0:1:11.277 Brian Haydin when the keynote launches, book of news is going to drop, we’re going to go in. And I’m like, what is this? There’s like this CLI thing that like, you know, I can’t use as my cheat sheet. I actually have to pay attention and listen to the keynote. So we sat through the keynote, took a bunch of pictures, and then spent two days talking with the product teams. And this is our opportunity to, you know, filter out what was noise. 0:1:11.557 –> 0:1:27.837 Brian Haydin what the good signals were, what the bad signals are, and what kind of piqued our interest. Suneer, introduce yourself real quick. We have a jam-packed amount of things to go through today, so we’re going to go fast and furious, and I’m going to let you introduce yourself. 0:1:28.397 –> 0:1:38.717 Suneer Mehmood Absolutely. Thank you, Brian. Hey everyone, I’m Suneer. I work as a lead architect for CONSULTANCY and looking forward to, you know, talking through the best of builds with you guys. 0:1:40.637 –> 0:2:1.997 Brian Haydin All right, so I’m going to start out with a little bit of a confession. Walking into Build, I had posted a newsletter post, you know, on my blog on LinkedIn with my predictions of things. And as I was getting on the plane to San Francisco, Nvidia announced their RTX Spark. We’re going to talk a little bit more about that. 0:2:2.77 –> 0:2:24.37 Brian Haydin But, you know, I had this thesis, one of the things that was going to be talked about is that Microsoft was building itself out to be the marketplace for AI and not necessarily a competitor. And boy, was I wrong on that one. I just want to admit it. The Microsoft AI models that came out, 0:2:25.117 –> 0:2:44.557 Brian Haydin you know, demonstrated they’re willing to lean in to be that competitor. And there’s a lot of good reasons about it. But I think if I take a step back, you know, I wasn’t really completely off base with this. We saw this theme of like the control plane really starting to take shape. We’ve been using this terminology now for about a year, but 0:2:44.957 –> 0:2:57.797 Brian Haydin I saw this start to take shape this year. And so we’ll see what that actually means. Hopefully we’ll get a good understanding of it in the webinar. Anything that like you got wrong walking into it, Suneer? 0:2:59.437 –> 0:3:17.877 Suneer Mehmood Well, like, you know, I went with an open mind and this was my first build. So, you know, I was like, everything was information from a fire hose to me. So I didn’t have any, you know, pre-concepts, you know, as I was walking into the session. So. 0:3:18.437 –> 0:3:37.757 Brian Haydin Yeah, so let me go ahead and just kind of give you a quick agenda of what we’re going to do here today. You know, we’re going to walk through, you know, some of our initial impressions, and then, you know, we’re going to just basically walk through security and governance, tooling, agent runtime, you know, the compute fabric. I know that 0:3:18.477 –> 0:3:18.957 Suneer Mehmood So. 0:3:38.317 –> 0:3:59.77 Brian Haydin Seniors got a lot of content. I’ve took a lot of good pictures, but high level, this is how we’re going to walk through things. And before we get started, you know, I do want to say that like agents don’t have like just app-shaped traffic anymore, right? We have a different call pattern that is starting to… 0:3:59.157 –> 0:4:3.837 Brian Haydin Bring in a lot of different data, anything. 0:4:4.846 –> 0:4:7.846 Brian Haydin Real quick, Sunir, on this slide. 0:4:6.966 –> 0:4:29.526 Suneer Mehmood Yeah, I mean, these agents, like, you know, they loop, store, retrieve, reason, act, and learn, right? So different call patterns mean different databases, a different warehouse, different context services. So that single fax explains, you know, every data announcement of the build, right? Like, hold that thought and we’ll cash in on. 0:4:29.846 –> 0:4:32.86 Suneer Mehmood That in the next 15 minutes, I guess so. 0:4:32.646 –> 0:4:47.326 Brian Haydin All right, so before we get into the deep dives, I think it’s really, we’re not going to get through every one of our, every one of the announcements. I think it’s really interesting just to take your top five. Why don’t you go first, Suneer? What did you, what really resonated with you coming out of Build? 0:4:47.686 –> 0:5:5.846 Suneer Mehmood All, the first one of my favorite was Microsoft IQ, right? So, context became a product service you can point at point agents at, right? And then we’ll talk more about it, like, you know, two was Horizon DB. It’s A Postgres service rebuilt for agents. 0:5:6.326 –> 0:5:11.526 Suneer Mehmood So it’s transactional plus vector plus search in one engine. 0:5:12.326 –> 0:5:33.46 Suneer Mehmood And 3rd was card speed, as they call it, right? So that’s, so it’s a fabric warehouse powered with NVIDIA GPUs, right? So you’re never gonna have the depth of compute power in fabric anymore. And then other thing that I really didn’t expect was like Power BI in Edge Integra, you know, describing a report and and. 0:5:33.206 –> 0:5:50.6 Suneer Mehmood Seeing A semantic model being built and a Power BI dashboard being built was awesome to see, right? And Cosmos DBs, some of the Cosmos DBs with native AI functionalities was also one of the things that I really liked. You know, these are my top five. 0:5:50.566 –> 0:6:9.246 Brian Haydin I remember that demo with the Power BI. It was really cool. And I was surprised too. The data platform is one of those areas where it’s got some cool AI features to it, but from a development standpoint, it seemed to be kind of lagging. And maybe we’re going to see a little bit of momentum. And I know you got 0:5:53.166 –> 0:5:53.446 Suneer Mehmood Yeah. 0:6:9.366 –> 0:6:27.686 Brian Haydin you got quite a few slides on that concept. So I’ll take over, I’ll go with my top five. You know, I would say like Foundry’s operating layer, you know, actually going GA. I’ve been doing a lot of talking about tracing and evals and, you know, having just one, you know, open telemetry pipeline. 0:6:28.166 –> 0:6:51.286 Brian Haydin And, you know, so it actually started to see this start to take shape. You know, ACS and Assert, there’s like now open, deterministic trust act for agents. You know, I think it got the least amount of coverage, but maybe deserves quite a bit more. Agent 365 was an interesting story kind of coming into it. 0:6:51.926 –> 0:7:11.366 Brian Haydin Beginning of the month was when we first started to be able to use it. But now we’re starting to use, you know, watch that through, you know, MXC and like how we can actually deploy some of these agents in a, you know, in a safety, a safer container. Copilot Studio agents coming under the Foundry control plane. 0:7:12.166 –> 0:7:33.126 Brian Haydin You know, again, something that wasn’t really talked about, I had to like really like get into the product teams to, you know, to, you know, layer into what that actually means. I’ve got a little bit of a hallway conversation. I’m not sure if I want to do it on a recorded line, but some interesting things that are coming through. 0:7:33.766 –> 0:7:55.366 Brian Haydin And then, you know, actually being able to see the agent framework in GA and what that actually means, you know, for organizations, how they’re going to start using it. You know, this is a, I would be remiss if I didn’t kind of lay into it. Not something that got a lot of like keynote attention, but definitely walking the floor, talking to the product teams, they were leaning into this real heavy. 0:7:56.246 –> 0:7:58.966 Brian Haydin So, you know… 0:7:59.846 –> 0:8:19.966 Brian Haydin Those are the 10 picks that we’ve got between the two of us. But I think if we kind of break this down into, you know, three major themes, takeaways that, you know, that I had, you know, the first one, I would say, you know, Suneer, you put the Horizon DP as your number, what was that, #2? 0:8:20.806 –> 0:8:21.446 Suneer Mehmood I guess, yeah. 0:8:21.926 –> 0:8:33.286 Brian Haydin I had it as kind of a footnote, and you know, I remember talking, I remember talking to, we were talking to the group at the same time. What make your case? 0:8:34.686 –> 0:8:55.366 Suneer Mehmood Yeah, so it’s an agentic app data tier, right? So it’s durable state plus vector plus search, all collapsing into one, right? So on the front face of it, you are actually getting your native Postgres, right, which is open source, and you get all of those functionalities. 0:8:55.526 –> 0:9:15.846 Suneer Mehmood plus all of the shared storage architecture for resiliency, for your DR, and extremely capable storage functionalities with, you know, Postgres being on the front. So it is going to be definitely needed in the AI era. 0:9:15.926 –> 0:9:16.726 Suneer Mehmood In my opinion. 0:9:17.446 –> 0:9:38.286 Brian Haydin Yeah, you know, I just feel like for me, it was, there’s tools out there that did it, you know, a lot of what they do. And it underwhelmed me a little bit, but I think there’s an important story with Postgres coming through and having that native boundary, you know, is, I think, going to be kind of important. 0:9:33.366 –> 0:9:33.766 Suneer Mehmood Yeah. 0:9:38.686 –> 0:9:57.846 Brian Haydin We also got this context, you know, became the infrastructure, and we’re gonna like really look into the IQ layers that surface, you know, that surface through the through the keynotes, and then these agents, you know, kind of like got identity and initiative. Any other like thoughts? 0:9:57.966 –> 0:9:59.486 Brian Haydin You know, from your perspective, Suneer. 0:9:58.486 –> 0:9:58.886 Suneer Mehmood Yeah. 0:9:59.926 –> 0:10:6.886 Suneer Mehmood Yeah, like, you know, for you, um, ACS and SERT weren’t even in my notes, so why is it #2 for you? 0:10:7.366 –> 0:10:26.326 Brian Haydin Yeah, so, all right, so ACS is the agent control specification. And this again, like comes with, you know, comes a little bit less from the keynotes and more from like the hallway chatter. And it, for me, it’s like, I’ve been talking so much about the enterprise adoption of 0:10:27.366 –> 0:10:47.446 Brian Haydin you know of these agents and then having the control patterns that go along with it, not just, you know, Agent 365, but you know, really being able to control these agents and the life cycles as well. You know, I think it’s super important, you know, something that’s been resonating with the conferences that I’ve been going to this year. 0:10:48.166 –> 0:11:8.646 Brian Haydin you know, and the people that I’ve been speaking about. So, but hey, you know, everybody’s going to have their opinions. That’s, you know, that’s why we bring a couple of people, you know, to Build is to, you know, take a look at different things. I do want to like just, you know, be a little bit honest about what we’re not going to talk about today. And there’s some really, really cool stuff on this slide. 0:10:56.246 –> 0:10:56.646 Suneer Mehmood So. 0:11:1.766 –> 0:11:2.246 Suneer Mehmood Seven. 0:11:9.446 –> 0:11:28.326 Brian Haydin Every one of these is real. And in fact, actually, if you just look at the raw community, like, who’s talking about what? The GitHub Copilot app, I mean, that was probably the biggest, like, there’s still people are still talking about it, trying to figure out what that means. The Copilot booths were just jam-packed the entire time. 0:11:29.686 –> 0:11:48.726 Brian Haydin If we do a part 2 of this, I think that would get a really significant amount of attention. But hey, I can only go through so much and I know we’re going to run out of time. So these are the things that I think in like 15 seconds or less that I’ll go ahead and talk about. So the GitHub Copilot app, 0:11:49.46 –> 0:12:9.686 Brian Haydin Having like this native, really easy to use, built for developers, you know, development desktop experience. There were things like, you know, concentration mode and all fantastic. All this stuff is great. The Windows Dev experience, working native with like WSOC containers. 0:12:10.86 –> 0:12:30.646 Brian Haydin Also fantastic, this Microsoft Discovery, the R&D platform, man, they’ve been talking about this at every build for the last two or three years. And it’s been really fantastic to see. It’s going GA, but man, I’ve seen it on the keynote like three years in a row. It just doesn’t excite me anymore. The NVIDIA Unified Stack, M-Dash, 0:12:31.526 –> 0:12:44.486 Brian Haydin My, my, my, I can’t, I’ve never figured out how to pronounce my, you keep asking about this, you know, why and what, and you know. 0:12:36.166 –> 0:12:36.886 Suneer Mehmood Marana. 0:12:38.166 –> 0:12:38.646 Suneer Mehmood Yeah. 0:12:46.166 –> 0:12:51.446 Brian Haydin We’re not going to talk about it, but you feel strongly about it, so I’ll give you 15 seconds to make a case. 0:12:52.86 –> 0:13:15.446 Suneer Mehmood Well, like quantum computing is the future. So I believe like Marana is worth the talk. I’ve been trying to read a little bit about quantum computing and why, you know, what are we going to get that, you know, long retention of information that quantum computing would have, you know, with Marana too, I think, you know, retention of information in 0:13:16.566 –> 0:13:36.86 Suneer Mehmood These chips are averaging to almost like 20 seconds, which will, you know, exponentially improve the probability of, you know, new inventions and and and stuff, right? So, so I’m really looking forward to that marijuana too, and they’re saying they know they can go production in 20. 0:13:36.166 –> 0:13:39.526 Suneer Mehmood Twenty-nine, so, so that’s really exciting. 0:13:38.166 –> 0:13:56.966 Brian Haydin Yeah, yeah, that’s a long time away, man. It’s an exciting announcement. I think it’s fantastic news. And it’s really going to change the compute paradigms that we actually use. But it’s so far out in the future, man. I just don’t, that’s kind of why I put it on the list. 0:13:41.46 –> 0:13:41.446 Suneer Mehmood Yeah. 0:13:42.726 –> 0:13:43.206 Suneer Mehmood Yeah. 0:13:53.606 –> 0:13:53.886 Suneer Mehmood Yeah. 0:13:58.166 –> 0:14:15.766 Brian Haydin All right, so let’s talk a little bit about theme one. So the control plane, right? I would say like the sharpest criticism that’s been floating around the developer forums after Build, you know, would come down to like the sentence, this governance layer is nowhere near the capability layer. 0:14:16.806 –> 0:14:37.206 Brian Haydin You know, and so that gap is what enterprises I think are like trying to trying to reconcile with right now. So we’re going to try to look, we’re going to look a little bit into what Microsoft did this year to close that gap. And, you know, I would say that one thing that sticks out to me is that it no longer matters where your agent was born. 0:14:37.606 –> 0:14:55.846 Brian Haydin whether it was in Copilot Studio, whether it’s, you know, something that you developed in Pro Code, maybe even like Google or AWS, you know, Microsoft has the management plane that is going to help you pull all of this together. So founders tracing, the evaluations, 0:14:56.166 –> 0:15:16.86 Brian Haydin I think I’ve done a talk about this. One of my coworkers, Mac, and I went down to Chicago and kind of demonstrated this out inside of the inside of Foundry. It’s fantastic. It’s got one, you know, hotel pipeline. You’ve got evals linked back to the exact traces and how you can produce them. 0:15:16.646 –> 0:15:36.886 Brian Haydin This is really turning the agents into something that enterprises can trust. And that’s like the key word for me is that they can actually trust. So once you have like these, you know, hosted agents that, you know, can get into like per session sandboxes, 0:15:37.846 –> 0:15:57.326 Brian Haydin You can start them up, you know, pretty much instantly. All these things, you know, all the maturity processes that are going around this is what really excites me and I think is what’s going to excite everybody else. So I’m seeing this like talk that I do, the agent ops talk, you know, actually show up in each one of these tools. And the other part that I’ve 0:15:57.606 –> 0:16:6.446 Brian Haydin you know, I just love is that it’s open, man. They’re bringing in all the different ecosystems. What about the data side for you, you know, Suneer? 0:16:7.966 –> 0:16:13.206 Suneer Mehmood Yeah, so again, like, you know, what excites me is like, um… 0:16:14.686 –> 0:16:33.206 Suneer Mehmood You know, at least from a data side, I can see three advancements, like we already mentioned about Horizon DB, we had mentioned about, like, you know, we are gonna mention about Cosmos DB AI, getting AI native, and then the fabric aspect of it, right? So, several advancements there, so… 0:16:26.326 –> 0:16:26.726 Brian Haydin Yeah. 0:16:34.86 –> 0:16:54.6 Suneer Mehmood You know, I can I can definitely see that, you know, this is all gonna work in the cohesive pattern and and and data, you know, every one of those traces and evals is, you know, queryable telemetry, right? Like, you know, from from the back end of Cosmos or Postgres, so… 0:16:54.286 –> 0:16:59.206 Suneer Mehmood The agent fleet just become a data set now, right? So that’s pretty much it. So. 0:16:58.766 –> 0:17:16.886 Brian Haydin Yeah, yeah, exactly. So, you know, I guardrails that I think maybe this is a little bit of a sleeper, you know, from the conference. Like I said, it wasn’t something that was featured, you know, really prominently, but this agent control specification, it’s, you know, 0:17:18.166 –> 0:17:36.726 Brian Haydin Best way for me to talk about it would be like an open standard that you that you can use, giving any runtime A deterministic allow or deny type of decision, at you know, five different checkpoints that you might have inside of the life cycle of an agent, so… 0:17:37.46 –> 0:17:55.686 Brian Haydin at input, at LLM call, maybe at state management, tool execution, and output, right? So being able to like have these, you know, baked into it is giving you like a real structural enforcement to the agents. And so. 0:17:56.406 –> 0:18:14.966 Brian Haydin That helps with this like whole probabilistic versus deterministic like fight that we have on a day-to-day basis. And this is going to be work, this is, you know, baked in, so it works across Foundry, your agent framework, even supported through LangChain. So I’ve been talking… 0:18:15.6 –> 0:18:36.766 Brian Haydin talking about agent ops as a formal like kind of discipline for now about six months. But I’ve been asked by customers, I’ve been talking about how do we make these agents enterprise worthy for a couple of years. And I think this is kind of like the seat belt that’s going to help us, you know, stay safe in our cars as we get on the highway. 0:18:37.686 –> 0:18:56.566 Brian Haydin So that for me is, you know, one that was big. Agent 365, this went GA May 1st. You know, we’ve got a registry, we’ve got access controls, we can do fleet dashboards. I can take a look at how much this is costing, this agent is costing me. 0:18:56.646 –> 0:19:18.486 Brian Haydin per month. Every agent that you’ve got is now governed under an, you know, a entra identity. And I think, you know, what some of you might have heard me say in some of my advisory sessions or on some of the other like talks that I’ve done is that like the lines between Copilot Studio and Foundry were starting to blur. 0:19:18.886 –> 0:19:37.846 Brian Haydin already. And now I’m starting to see this as like a unified platform. And it’s not, you know, when I was talking to, when I was talking to the Foundry group about the control plane and merging or unifying the Copilot Studio and 0:19:37.926 –> 0:19:57.286 Brian Haydin and Foundry, it’s not just the Microsoft ecosystem either, it’s the Google and you know AWS and you know the other agent platforms that you can bring in there as well. This is such an important concept and I still don’t understand the licensing of it. Nobody’s gonna like, you know, I still don’t quite figure what. 0:19:57.326 –> 0:20:13.846 Brian Haydin this is going. But this is going to help you manage shadow IT agents that are being built and deployed within your ecosystem. And it’s going to make your security teams feel a lot better about what’s going on. Before I move on, anything you want to say about the last slide or the last two slides? 0:20:15.46 –> 0:20:23.846 Suneer Mehmood Yeah, like what a cross-platform agent registry means for data governance, right? So… 0:20:26.246 –> 0:20:45.206 Brian Haydin We’ve got, like, you know, you’ve got permissions that are being set up right between, like, you know, you got this agent, they have, they have their set of permissions that, you know, that you can control, and it works the same whether you’re in Teams, whether you’re in Fabric. 0:20:26.286 –> 0:20:27.126 Suneer Mehmood What? 0:20:45.566 –> 0:21:4.566 Brian Haydin it’s going to respect those boundaries as well, right? And I think that’s, you’ve got a big story about the, you know, about the IQ, and I think that’ll kind of piece it all together for the audience. And then, you know, I think, you know, the other thing too is with the Microsoft agent. 0:20:48.486 –> 0:20:48.886 Suneer Mehmood Oh, yeah. 0:20:54.886 –> 0:20:55.366 Suneer Mehmood Mhm. 0:21:4.646 –> 0:21:25.446 Brian Haydin you know, framework kitting, you know, general availability, you’ve got this orchestration layer that, you know, stays, you know, where it’s been, deterministic, and all the participants, all the different models are still kind of plug and play, right? We don’t have any dependencies on, you know, on a specific. 0:21:26.46 –> 0:21:47.446 Brian Haydin you know, we’re not vendor locked, I guess would be the way the best way for me to talk about this is just say that we can we can mix and you know mix and match a little bit as well. So we’re we’ve got plenty of got plenty of slides to go through. Why don’t we start to position a little bit more in the context becoming the infrastructure. 0:21:48.86 –> 0:21:51.206 Brian Haydin We’re talking about IQ here, I think, if I remember correctly. 0:21:48.166 –> 0:21:48.566 Suneer Mehmood Sure. 0:21:51.686 –> 0:22:12.206 Suneer Mehmood Yep, yep. So I mentioned before, right, that the agents don’t have app share traffic. They loop, store, retrieve, reason, act, learn. This is where, you know, I cache that in, right? Everything Microsoft rebuilt in the data tier traces back to that loop. So let’s go ahead and start. 0:22:14.86 –> 0:22:14.646 Suneer Mehmood All right. 0:22:14.566 –> 0:22:16.166 Brian Haydin Microsoft IQ. 0:22:16.406 –> 0:22:35.366 Suneer Mehmood Microsoft IQ, right? It’s A context layer. It does not replace Copilot and does not have a UI of its own, right? It stitches 4 sources. One, Web IQ. This is a brand new AI native web Browning API, fresh. 0:22:35.446 –> 0:22:56.806 Suneer Mehmood Ranked have from official sources, right? Second is fabric IQ, and I’ve been talking about that in, you know, earlier demonstrations that we had as well, right? So, you see a structured business data as ontology built by extending the Power BI semantic models and your other storage solutions within the fabric that. 0:22:57.206 –> 0:23:2.6 Suneer Mehmood We call us lake house and warehouses and has live telemetry data, right? 0:23:3.166 –> 0:23:23.446 Suneer Mehmood And then the work IQ with M365 signals, your mail, calendars, Teams, SharePoint, organization graphs, just again, permission aware, right? And then Foundry IQ, the knowledge retrieval layer for agents, you build in Foundry, right? And it is the year as of June. 0:23:24.6 –> 0:23:47.206 Suneer Mehmood And one thing I really liked was the demonstration by Microsoft, right? So Microsoft showed this live with a power utility bright line scenario, right, where an operator needs to assess a live grid incident and write a response brief. And they did it with one agent, in fact, you know, built in Microsoft Foundry, wired to a Foundry IQ knowledge. 0:23:46.646 –> 0:23:47.126 Brian Haydin Mhm. 0:23:47.526 –> 0:24:8.606 Suneer Mehmood base that packages, you know, with the utilities, documents, operational data, people into context, and then publish to Microsoft 365 for the whole team. So there were three questions that they asked. What are the current electricity prices in SFO? Right? An ungrounded model cannot answer that. The agent now reaches Web IQ. 0:24:8.806 –> 0:24:28.886 Suneer Mehmood you know, and searches the web, you know, and grounds that data with what is available with, you know, the current mark, you know, out there in the public internet, right? Second was like, which of our stations are at risk? So now, it’s not public information, it needs internal context. So it queries fabric IQ, and then the… 0:24:29.46 –> 0:24:47.686 Suneer Mehmood you know, the entire Bright Lines grid is a fabric ontology right now, I mean graph model, right? So it retrieved that internal data. And 3rd was what are the steps to respond to a substation trip? So that policy and people activated work IQ. 0:24:47.926 –> 0:25:5.366 Suneer Mehmood and read the response playbook straight out of SharePoint and that to the latest, right? The live document the team actually maintains. It’s not a stale upload, right? So one situation, three questions, one connected answer. The outside world, your operations, and your people all stitched together. 0:25:6.566 –> 0:25:28.646 Suneer Mehmood It ran as a long running agent, so you can see every step along the way with Foundry routines to run on a schedule. So I don’t think the model was the hard part, reaching the right context at the right moment. It is. I mean, that seems to be the, you know, the key here, right? It’s not about picking the models, getting the data estate clean from a data standpoint, right? 0:25:29.46 –> 0:25:39.846 Suneer Mehmood governed and reachable, right? So there has to be, it has to be clean, governed, and reachable, right? So that’s a lot of IQs, but I hope that helps connect the dots. 0:25:40.726 –> 0:26:1.766 Brian Haydin Yeah, I’m going to throw out here just a quick housekeeping item. You’re going to see on some of these slides, we’ve been annotating them with what’s in GA, what’s in preview. Microsoft Build is a lot of times when they release things that are in private preview before anybody can actually touch them. But some of these things are actually, you know, coming out like right there on the spot. 0:25:40.926 –> 0:25:41.126 Suneer Mehmood But… 0:26:2.486 –> 0:26:24.646 Brian Haydin I also echo you, Suneer. I saw this live and in person yesterday talking with one of my customers where they were in their web browser and opened up the Web IQ, right? You know, so they opened up Copilot and the web. They had CRM open and were asking questions about, you know, like that Web IQ would, you know, ask. 0:26:15.526 –> 0:26:15.926 Suneer Mehmood Yeah. 0:26:25.6 –> 0:26:44.46 Brian Haydin But it also was able to transition really quickly into like the work IQ and pull out like the CRM record that like, you know, there was a list of a bunch of records and pulled these things out. It was so cool to see all these like context layers like really working together. Fantastic. And 0:26:29.846 –> 0:26:30.326 Suneer Mehmood Done. 0:26:44.966 –> 0:26:55.446 Brian Haydin You know, maybe the official branding Microsoft IQ would get rid of some of these other IQs, but fantastic. All right, this is you, this is your slide, man. Make your case stronger. 0:26:53.926 –> 0:26:54.726 Suneer Mehmood AA. 0:26:55.966 –> 0:27:16.6 Suneer Mehmood Horizon DB, right? So Postgres just got an AI native sibling, right? So I’m sure some of you use Postgres already together, I mean, right now, right? So Horizon DB hit public preview at Build. The specs, Satya quoted on stage, built ground up. 0:27:16.246 –> 0:27:35.966 Suneer Mehmood for high availability, scale out, zone redundant with automated failover, 128 terabytes storage per cluster, 15 read replicas, and 3x throughput versus any self-managed PostgreSQL in there as per their internal testing. 0:27:36.86 –> 0:27:36.486 Suneer Mehmood Right. 0:27:38.166 –> 0:27:57.366 Suneer Mehmood So it also has got in-database model invocation through Azure AI extension, inference via SQL, no external orchestration layer, plus advanced vector indexing and semantic search. So it collapses the transaction database plus vector store. 0:27:58.406 –> 0:28:21.686 Suneer Mehmood plus search index, you know, which we all use for RAG applications, right? Retrieval augmented generation applications today. So it is Postgres compatible. So for our Azure DB for PostgreSQL clients, there is a familiar path, but it’s preview and only in five regions at launch. I just want to call that out as well. Right now it’s in Australia, East, Central US. 0:28:3.766 –> 0:28:4.966 Brian Haydin Yeah, exactly. 0:28:22.6 –> 0:28:31.686 Suneer Mehmood Sweden Central and West US 2 and 3, but more are coming. So pilot it before putting that in your roadmap. It’s really promising though. 0:28:32.486 –> 0:28:51.846 Brian Haydin Yeah, I would say pilot it. A couple things that jumped out for me. One is I remember talking with the Postgres team last year. They had a booth and they had a managed service that was out there. And when we were talking to the Horizon DB team, I challenged them. I said, what’s different about this, right? 0:28:34.486 –> 0:28:34.806 Suneer Mehmood Yeah. 0:28:52.246 –> 0:29:13.526 Brian Haydin And so A, you’re not going to have to go outside of the ecosystem and use a third party managed system, which I think is really cool. They’ve got a unified like architecture that is distributed, you know, you know, in up to 15 different zones. So I think it’s, you know, it’s a really fantastic thing. 0:29:13.846 –> 0:29:36.406 Brian Haydin One thing to note, especially around that pilot, is that when we asked them, like, what about the pricing of this, right? They do have pricing that’s published, but they were really quick to acknowledge that they don’t know what this is going to cost to run yet. So as you’re piloting this, like, just expect over the next year that that pricing is probably going to fluctuate, you know, quite a bit as you sort of figure out what this looks like. 0:29:19.846 –> 0:29:20.326 Suneer Mehmood Mhm. 0:29:28.246 –> 0:29:28.646 Suneer Mehmood Yeah. 0:29:36.806 –> 0:29:37.446 Suneer Mehmood Exactly. 0:29:37.46 –> 0:29:39.846 Brian Haydin But all right, so next topic, God speed. 0:29:40.566 –> 0:29:58.286 Suneer Mehmood All right, God speed. So, fabric data warehouse, right? You know, some of us are already in the path of fabric because of its, you know, concept of one leg bringing everything in one under one governance and security architecture, and and you know the AI native capabilities, right? So… 0:29:58.406 –> 0:30:18.86 Suneer Mehmood those who are already in the path or contemplating that path. Now, Fabrics Data Warehouse can now publish eligible queries onto NVIDIA GPUs inside the execution engine. So there is no rewrites, no cluster sizing or resizing. The optimizer decides when the GPU is a better plan. 0:30:18.486 –> 0:30:38.246 Suneer Mehmood Right? So Satya said, a 7x on stage in the keynote, the official fabric blog is precise. And it says up to 7x faster than the three leading cloud data warehouse vendors at 64 user concurrency. So that’s the picture that I have on the right side, right? 0:30:39.46 –> 0:30:57.286 Suneer Mehmood As per the internal benchmarking, I know it’s very small, and you know, because I had to work with the real estate that I had in the slide, so the first time, I mean, on the graph on the extreme left that you see is the three axis, like you know, on a single person concurrency, you know, so when it switches from fabric to… 0:30:57.486 –> 0:31:18.566 Suneer Mehmood NVIDIA GPUs, you are getting, you are making your queries 3x faster. And if it is like 16 user concurrency, it’s 6x faster, it is 64 user concurrency, it’s 7x faster, right? So the research is the real credibility here, right? It grew out of the tensor query processor. 0:31:18.966 –> 0:31:38.806 Suneer Mehmood productized as card speed and won the best industry paper at Sigma 2026. The first fully managed warehouse to offer GPU acceleration fabric, right? So it’s entering early access preview in July 2026. It’s not GA today. So have that as a directional. 0:31:39.286 –> 0:31:58.646 Suneer Mehmood You know, in your roadmap and and pilot it, you know, and when it is available, you know, after you pilot, switch it on in production, you know, once you’re once you know how it works, and and why should we care about this, right? Like, you know, for people who are slightly distant from the… 0:31:59.166 –> 0:32:18.566 Suneer Mehmood constant reporting warehousing gimmicks, right? So heavy reporting or aggregation is exactly where warehouses get slow and expensive. So moving that to a GPU transparently changes the cost and performance math without rewriting any of your layers in the Medallion architecture. And mostly the… 0:32:18.646 –> 0:32:36.966 Suneer Mehmood reports pull from a layer called gold, where you have curated slash summarized data, right, meant for your reports. But you know, you do not have to redo any of that. I’d still get independent benchmarks before quoting the speed ups. So that’s my personal take on that. 0:32:37.606 –> 0:32:57.126 Brian Haydin Yeah, definitely. I think, you know, looking at this graph, what really jumps out to me is that, you know, without the GPU accelerated query performance, you have exponential growth as you start to grow these, you know, the concurrent users. But it looks like a real linear growth pattern with the GPU. 0:32:57.286 –> 0:32:58.86 Suneer Mehmood Exactly. 0:32:57.966 –> 0:33:17.366 Brian Haydin which to me, like, you know, sky’s the limit then, right? So might actually save you some serious money when we start talking about, you know, the difference between an F64 and an F128, right? That’s some big things. I was not as impressed. 0:33:2.6 –> 0:33:2.766 Suneer Mehmood Exactly. 0:33:11.566 –> 0:33:12.606 Suneer Mehmood Mhm, mhm. 0:33:18.166 –> 0:33:25.686 Brian Haydin with Ray Finn and man, you really leaned into this pretty hard. I remember sitting in a couple of sessions with you. 0:33:26.486 –> 0:33:44.886 Suneer Mehmood Yeah, so one thing that really excites me about Rayfin is now we have been perceiving Fabric as your one-stop shop in the future for your data, right? And we have been perceiving it as your data platform, right? That’s the concept that Microsoft had. 0:33:46.6 –> 0:34:3.366 Suneer Mehmood three years ago when they actually introduced Fabric. You do not have to spin up different services like your Azure Data Factories or your storage solutions like your BLOB or your relational databases and whatnot. You have Fabric as your ecosystem. Now they are extending beyond that. For the record, Rafin is an SDK. 0:34:3.846 –> 0:34:23.126 Suneer Mehmood An open source SDK living on GitHub as Microsoft slash Rayfin is the engine underneath Fabric’s new app building experience, not a code button. So Fabric is now going from a data platform concept to even an app building concept. 0:34:23.206 –> 0:34:42.326 Suneer Mehmood concept, right? So you have one command, NPX riff and up, and that provisions a complete enterprise app backend on fabric, a SQL database, Microsoft Entra ID authentication, a GraphQL API, a OneLake connected static hosting, and such, right? 0:34:42.806 –> 0:34:53.366 Suneer Mehmood You can literally start from a template. It could be a blank template, or right now the templates that they have is like one is a blank template, they have a to-do list app template, then they have a data app template. 0:34:54.566 –> 0:35:12.646 Suneer Mehmood So that the data app template, for example, wires straight into your semantic models. So it ships the scaffolding and agents.markdown guide, MCP servers, arrayfin.yaml config, sample code, et cetera. So the unlock from a BI folks. 0:35:13.926 –> 0:35:14.486 Suneer Mehmood You know what? 0:35:15.766 –> 0:35:34.166 Suneer Mehmood You can actually turn a fabric semantic model into a working app, right? Or an agent without hand building the API or plumbing. And because the app runs on fabric, its data lands in OneLake, and it is immediately available in Power BI, notebooks, and data agents. 0:35:35.206 –> 0:35:54.86 Suneer Mehmood There’s also Replit partnership, by the way. So the developers can build in a familiar AI-first environment and deploy it into a governed fabric tent. Again, from an architectural standpoint, it’s new, piloted, you know, before planning it in your production. 0:35:54.806 –> 0:36:10.886 Suneer Mehmood Two questions I would ask is how much of the flexibility do you lose? How much of A flexibility do you lose of, you know, when you go by the templates happy path, right? And how stable is an SDK, which is very new? So. 0:36:12.6 –> 0:36:30.886 Brian Haydin Well, you and I closed out Build at a little workshop, little round table session, talking about Rayfan. And the thing that jumped out to me was that the governance wasn’t quite there. And I could see this turning into like the access databases of the future, you know. 0:36:18.486 –> 0:36:18.886 Suneer Mehmood Yeah. 0:36:25.606 –> 0:36:25.846 Suneer Mehmood Zach. 0:36:30.806 –> 0:36:31.526 Suneer Mehmood Exactly. 0:36:31.686 –> 0:36:42.246 Brian Haydin And that to me is I want to see that problem solved before I’m going to give it my endorsement. But it’s definitely something cool. What about Power BI? This was cool. This was cool. 0:36:40.286 –> 0:36:59.206 Suneer Mehmood Power BI. Yeah, this was really cool, right? So 2 Power BI announcements that compress the path from raw data to a polish report. You know, both are in preview, by the way. Agent skills for Power BI, which is a screenshot that I have. So if you can read on the black. 0:36:59.366 –> 0:37:21.126 Suneer Mehmood screen over there, right? It’s like describe, you can describe the report that you want in plain language and hand it a screenshot. Well, you know, that’s creating a semantic model, by the way. The screenshot that I have is creating a semantic model. But, you know, you can literally describe the report that you want in a plain language, provide it a screenshot, and an AI agent builds the end-to-end solution. 0:37:21.606 –> 0:37:41.926 Suneer Mehmood right, a semantic model from your fabric data, report pages, visuals, with design best practices. So it works on, excuse me, existing reports as well. So cleaning up a scrappy report is no longer a rebuild. It builds on the Power BI modeling MCP that shipped. 0:37:42.6 –> 0:37:43.46 Suneer Mehmood Last November. 0:37:44.246 –> 0:38:4.286 Suneer Mehmood So Fabric apps on semantic models. So AI agents build and deploy Fabric native web apps directly on your semantic models, right? So Fabric as the back end powered by open source Rayfin SDK from the previous slide which we spoke about, right? So my read is like, yeah, like, you know, Fabric is this. 0:38:0.366 –> 0:38:0.726 Brian Haydin Yeah. 0:38:5.526 –> 0:38:23.206 Suneer Mehmood Build announcement on fabric was like really huge with these two, right? Like, plus I would add NVIDIA GPU capabilities as well, right? So, the governance angle is what makes sense, you know, more than a demo standpoint, right? So… 0:38:23.326 –> 0:38:42.566 Suneer Mehmood for years, let business users build apps, meant shadow data and ongoing logic, right? So building on semantic model means the metrics and the security travel with the app, right? So the catch is that, yeah, both are preview and an agent builds your model. 0:38:36.766 –> 0:38:39.486 Brian Haydin It’s important, yeah, that’s important. 0:38:43.126 –> 0:39:2.166 Suneer Mehmood is only as good as your data estate underneath it. So it’s really important that you keep your data accurate, precise, right? So garbage and still gets your polished report out of garbage, which loops right back to the Microsoft IQ standpoint, you know, that I had about clean and governed estate. So. 0:39:1.526 –> 0:39:12.6 Brian Haydin Yeah, exactly. All right, what’s Cosmos DB? You brought this up earlier. What’s good? What’s bad? What did you hear? 0:39:12.726 –> 0:39:30.726 Suneer Mehmood Yeah, so let me make the AI native concrete with two toolkits Microsoft shipped to preview at Build, right? Both built on Cosmos DB for NoSQL as a single store for documents, vectors and full text with vector, full text and hybrid search are built in, right? So 0:39:31.366 –> 0:39:52.206 Suneer Mehmood There is no need of a second vector database, no sync coordinated. First is like the agent memory toolkit, a Python SDK that gives an agent a real memory. It stores raw conversation, you know, and it turns into derived memories. It summarizes extracted facts and cross-thread user profiles, so the agent holds a… 0:39:52.286 –> 0:40:11.286 Suneer Mehmood coherent thread now and recalls the stable user context weeks later, right? So the part I like is that it can process that memory automatically in the background using cosmos, change feed, and durable functions. So new turns get summarized and fax extracted without you having to orchestrate it. 0:40:11.606 –> 0:40:29.926 Suneer Mehmood And it even reconciles contradictory facts with an audit trail. That was like really awesome to see, right? So, and memory is usually the things teams bolt on with a vector store plus a document store plus a custom logic. So here you have one toolkit on the database you already run, right? 0:40:17.686 –> 0:40:18.246 Brian Haydin Yeah. 0:40:30.886 –> 0:40:50.526 Suneer Mehmood And the second will agent retrieval toolkit, a reference architecture for multi-step RAG. One shot RAG retrieves ones and answers ones. This one retrieves, drafts a preliminary answer, spots the gap, generates follow-up sub-questions, retrieves more evidence, and synthesizes A grounded final answer. 0:40:51.366 –> 0:41:6.806 Suneer Mehmood With vector plus full text search, so that’s well, that was cool to see, like, you know, Cosmos DB, you know, having these two new toolkits added, right? So, yeah, that’s pretty much my take on that. 0:41:9.126 –> 0:41:18.966 Brian Haydin All right, we talked a little bit about this already, right? Describing the report, the agent goes out and build it. Anything you want to add here? This is… 0:41:14.526 –> 0:41:15.446 Suneer Mehmood Yep, yep. 0:41:17.846 –> 0:41:18.446 Suneer Mehmood No, I. 0:41:19.766 –> 0:41:38.566 Suneer Mehmood Yeah, like, you know, I think, you know, in my preparation, I had two versions of it. So pretty much like, it’s a beauty of like, you know, your natural language development of your report, right? Like it creates a semantic model and even the visualization, which is like really cool. You don’t have to spend a lot of time on building your semantic layers and dashboards. So that’s what my… 0:41:38.646 –> 0:41:40.646 Suneer Mehmood Problem I, which we already saw, so… 0:41:39.926 –> 0:41:48.166 Brian Haydin So now this is in preview. Have you been able to play with this in Fabric yet or is it something you got you got to turn it’s not rolled out in our tenant right? 0:41:44.646 –> 0:41:46.806 Suneer Mehmood Not yet, not yet, like you know. 0:41:48.646 –> 0:41:56.166 Suneer Mehmood Yeah, I don’t think it’s rolled out in our tenants, so I’m really looking forward to trying that, like, you know, in the coming days, actually. 0:41:56.966 –> 0:42:20.646 Brian Haydin Yeah, what about optimization? I know that there were a lot of improvements about how like less experienced people can can can you know, become kind of optimization experts, right? What do you like, what were you seeing in terms of like Microsoft Research and the R&D tools? 0:42:13.286 –> 0:42:13.846 Suneer Mehmood Yeah. 0:42:21.606 –> 0:42:40.766 Brian Haydin that, you know, I know we’ve seen like the R&D, the researching tools, you know, kind of like over the last couple of builds, but I think this was under described, under talked about, and kind of like a bonus, like founded on the floor kind of thing. 0:42:40.846 –> 0:42:41.486 Brian Haydin Thoughts? 0:42:42.526 –> 0:43:2.766 Suneer Mehmood Yeah, like, you know, OptiGuide, right? That was a floor find right out of Microsoft Research. So it brings supply chain and scheduling optimization to people who are not operations research PhDs. So schedule the line, minimize over time, hit the shift date, and it translate that into a constrained optimization run. 0:43:2.846 –> 0:43:10.206 Suneer Mehmood If you run a plan, the preview signup is open and the bar to enter is low actually. So the whole… 0:43:11.406 –> 0:43:13.726 Suneer Mehmood Thing is, like, go sign up, right? So… 0:43:14.526 –> 0:43:34.846 Brian Haydin Yeah, yeah, I like this. There were some really cool demos in the researcher lab that I was able to find specifically around this. And I don’t know if some of the people that were there with me, some of our customers from Build were watching this as well. It was pretty cool. All right, we’ve got… 0:43:36.46 –> 0:43:41.566 Brian Haydin We’ve got some some other topics here to just kind of quickly through through Ray, some scorecards. 0:43:38.446 –> 0:43:39.646 Suneer Mehmood Scorecards. 0:43:41.966 –> 0:43:58.366 Suneer Mehmood Yeah, so this has a scores cut for the data tier. Build on what is GA, you know, so fabric mirroring and Medallion, Cosmos GA features, Microsoft IQ, and work IQ, I mean, work IQ APIs is GA, you know, this week. 0:44:0.206 –> 0:44:6.926 Suneer Mehmood Always pilot, don’t roadmap, you know, Horizon DB is in preview, card speed in early access in July. 0:44:7.886 –> 0:44:26.966 Suneer Mehmood Power BI agent skills in preview, the operating rule context is the new bottleneck and the data is a state is a moat, I would say, right? A proof of concept is a concept car. Don’t drive it on the highway, right? So Brian, the agents looks like they grew up. 0:44:19.566 –> 0:44:19.966 Brian Haydin Yeah. 0:44:27.726 –> 0:44:28.206 Suneer Mehmood Right. 0:44:28.286 –> 0:44:49.566 Brian Haydin That is very true. So, you know, kind of walking into the next theme, you know, agents, they did get this, they got this identity and importantly, they got initiatives too, right? Like autopilot is something that I thought was really interesting. I had to scratch my head a little bit when he brought it up. 0:44:50.446 –> 0:45:12.366 Brian Haydin And, you know, one mental model for, you know, all this like branding sprawl that’s happening, the three different planes. Plane one, we’ve got IQs. You know, that’s the context of what the agent knows. Plane 2 is the co-pilots. That’s the interaction the human, you know, asks, the AI answers, the episodic, you know, and reactive type of thing. And then plane three, 0:45:12.766 –> 0:45:34.526 Brian Haydin We’ve got, you know, this new one, the autopilots, you know, finally a co-pilot that’s granted an identity, it’s got persistence, it can take initiative, right? They talked about Scout as like the first autopilot, and that’s going to run under its own, you know, its own identity, right? Not a shared service account, but its own identity. 0:45:23.886 –> 0:45:24.286 Suneer Mehmood Yeah. 0:45:35.246 –> 0:45:54.206 Brian Haydin And that’s the kind of like, that’s the kind of like model that we need that’s going to get past the security team when you review it. So I think it’s really cool. You know, I think we’re going to see some more about it. And I, you know, I do want to talk a little bit about Scout. I thought it was really, really interesting. 0:45:55.246 –> 0:46:13.166 Brian Haydin Scout, you know, for those of you that didn’t see the keynote, it acts in the background. It has the ability to read your inbox and draft any kind of replies for you, organize your files, maybe do some meeting prep if that’s what you need. And it works all the way across the IQ system, right? So Teams and Outlook, 0:46:13.566 –> 0:46:31.806 Brian Haydin OneDrive, SharePoint. But the most important thing is that when it does take actions, and that’s the point of this is take actions, they’re all going to be attributable to like, you know, the audit trail that runs under your Entra ID as well as the Scout, you know, Entra ID. So 0:46:32.366 –> 0:46:54.286 Brian Haydin You know, there was, you know, a little bit of leak memo controversy happening around this. I’m not exactly sure. You know, I want to dive into it a little bit. But what I’m saying is like, let’s see how the like preview, you know, terms and how the telemetry disclosures kind of evolve. 0:46:54.686 –> 0:47:14.926 Brian Haydin before you put this in front of your workforce. I haven’t been able to turn it on in the concurrency tenant. I’m not sure that I really want to. You know, when you go through the setup instructions, it requires you to turn OpenClaw on for the organization in order to get this work. And I’m not sure that a lot of organizations are going to be quite there. 0:46:57.406 –> 0:46:58.46 Suneer Mehmood Mm-hmm. 0:47:15.246 –> 0:47:34.766 Brian Haydin But this is definitely frontier preview. Be careful. You know, make sure that you are, that you’re rolling this out in a safe and monitored way. My posture really for this in the next 90 days is to watch it, but don’t buy. 0:47:33.606 –> 0:47:33.806 Suneer Mehmood Mm. 0:47:35.86 –> 0:47:54.446 Brian Haydin I’m going to buy it in my own personal tenant just to play around, but you know, I’ll let you know how that goes. Which I think kind of leads us a little bit into like the trust perimeter, right? The 4th, you know, deep dive. This is where things are starting to reach into 0:47:54.606 –> 0:48:13.406 Brian Haydin you personally into the laptop. So I think that Nadella’s line was, Windows is the best place to run agents, you know, but the strategy, you know, the strategy now isn’t all cloud or all local, it’s something in the routing in between. 0:48:13.486 –> 0:48:36.526 Brian Haydin right? We’ve got, we have to look at the routine, we have to look at sensitivity, we have to look at latency bound work and how that stays local at a, you know, taking into account like the token economy that’s, you know, starting to get a lot of traction right now. So we’re looking for ways to like get near 0 token cost, but still keep that heavy, that heavy reasoning. 0:48:36.926 –> 0:49:0.446 Brian Haydin And that’s probably going to stay in the cloud, because that’s where those compute resources are going to need. Nobody’s going to be walking around with a dolly, you know, in order to carry these, you know, huge like stacks of GPUs to process it. So, you know, if that kind of like, you know, segue a little bit into this like open claw, you know, story. I thought that was. 0:48:46.926 –> 0:48:47.286 Suneer Mehmood Yeah. 0:49:1.166 –> 0:49:21.166 Brian Haydin That was really cool. And it brings together both the OpenClaw and Windows, which I have not had a chance to play with yet, but also like this idea of the Microsoft execution containers. So they work kind of hand in hand. And I see the value of like the Microsoft execution containers. 0:49:21.726 –> 0:49:40.206 Brian Haydin But giving it, like making it real with people with OpenClaw, I think is like the two pieces that brought that together. So what I’m trying to show here is like the whole hybrid thesis in one, you know, single frame. OpenClaw as an open source runtime, it’s got some dangers, you know, running on your Windows machine. 0:49:40.766 –> 0:50:2.446 Brian Haydin maybe running an anthropic model with, you know, OS enforced containment. That’s the permissions, you know, that we need to take this stuff to the next level. And being able to do some of this in like a local inference is super important as well. And the Microsoft execution containers, they’re going to help enforce that boundary. 0:50:2.726 –> 0:50:23.326 Brian Haydin And they’re gonna do that at the OS level, right? So you kind of only have to describe this execution container once, and Windows gonna hold that like boundary for everywhere that this agent’s gonna run. Bring in to, you know, bring in like Defender, Entra, Intune, Purview, all this stuff like extends to these local agents as well. 0:50:23.686 –> 0:50:44.686 Brian Haydin So you’re going to get that like, you know, layer of governance and security across the organization. You know, I would say like this idea of building local agents was kind of a blind spot. People were going to be able to do these open clause, clods, you know, things like that. And I think that blind spot just got kind of closed up a little bit. 0:50:45.6 –> 0:50:49.46 Brian Haydin What about from the data side? You know, do you have any thoughts on that? 0:50:49.646 –> 0:51:9.6 Suneer Mehmood Yeah, from the data side, local inference means unmetered tokens, right? So an A on one plan ships 14 billion parameter reasoning inbox, tool calling, and the data never leaves the device, right? So that’s, yeah. 0:51:7.726 –> 0:51:8.126 Brian Haydin Yeah. 0:51:9.126 –> 0:51:11.166 Suneer Mehmood That’s the upside, so… 0:51:10.926 –> 0:51:28.766 Brian Haydin Yeah, that AI on model is something that I am looking forward to playing with in quite a bit. 14 billion parameters is enough to run on my workstation, even without the RTX Spark. So, you know, I might be able to see some good results out of that. 0:51:25.166 –> 0:51:25.486 Suneer Mehmood Yeah. 0:51:29.486 –> 0:51:48.846 Brian Haydin Project Solera, I had a really nice like close up of that guy’s badge when he was standing up on the stage. And, you know, I’d say like, here’s just, you know, somewhat of a respectful disagreement. Microsoft, they did ask the right question, like, what’s the next form factor for an agent first world? 0:51:49.406 –> 0:52:8.286 Brian Haydin Honestly, I don’t think it’s going to be like, you know, a laptop, you know, you know, in the future. But their answer was, you know, a wearable badge, right? And they had some other reference designs up that I took some pictures when we were walking around the booth. And I just don’t see it, man. 0:52:6.286 –> 0:52:6.686 Suneer Mehmood Yeah. 0:52:9.86 –> 0:52:27.886 Brian Haydin I don’t see these devices. Like, you know, I saw that that I saw that badge up on the stage and, you know, took that picture of it. And I thought to myself, I have something very similar to that. It’s called a smartphone. It looks about the same size, the same weight. What am I getting out of this, right? So I just wasn’t a big fan. 0:52:9.126 –> 0:52:9.766 Suneer Mehmood Okay. 0:52:22.46 –> 0:52:22.366 Suneer Mehmood Yes. 0:52:25.166 –> 0:52:25.806 Suneer Mehmood True. 0:52:28.206 –> 0:52:49.326 Brian Haydin We’re going to see where this goes, but I do think that we need to start thinking about where we meet these agents, where we interact with them. And I don’t think it’s going to be on a laptop or necessarily my phone. I don’t always think it’s going to be a meta glasses or an AI lapel that somebody wears. So it’s great to hear people talking about these ideas. 0:52:39.806 –> 0:52:40.286 Suneer Mehmood Mhm. 0:52:49.646 –> 0:52:51.806 Brian Haydin But let’s find something that actually works, right? 0:52:52.46 –> 0:52:55.486 Suneer Mehmood Absolutely looking forward to that. However, that evolves. 0:52:56.46 –> 0:53:14.766 Brian Haydin Yeah, and that talks down, I, you know, kind of led into this last little part here, which was the Surface RTX. I mean, this is a huge story for me, you know, in a lot of ways. But it’s not a thing yet, right? We’re still, we’re going to wait to see it. 0:53:13.646 –> 0:53:13.846 Suneer Mehmood Yeah. 0:53:15.46 –> 0:53:37.246 Brian Haydin I had took some pictures. Those are the actual pictures from the floor. You couldn’t touch them. They were under these like super secure like glass boxes. Actually, the battery ran out of the laptop and they had to lift the box. And I was like, can I hold it? And she was like, no. You know, armed guards started walking closer to me when I got closer. But they’re real devices. And I was talking about this at 0:53:22.726 –> 0:53:22.926 Suneer Mehmood Yeah. 0:53:29.966 –> 0:53:30.366 Suneer Mehmood Okay. 0:53:37.606 –> 0:53:57.566 Brian Haydin a meetup last night. If you think about it, they’re promising this is going to be available in the fall. So I would expect that you’re going to see these devices, that these are actually working prototypes of the real thing. You know, so I’m not actually surprised about it. What struck me the most is that the dev box on the left side, it was so 0:53:57.646 –> 0:54:16.926 Brian Haydin then. I mean, it really wasn’t that big of a box. And it’s the power that it has, right? A petaflop, you know, on your desk, you know, in your backpack, walking around, local AI compute. It’s got 128 gigs of unified memory. That’s going to compete with the Mac, with the Mac minis and 0:53:58.446 –> 0:53:58.926 Suneer Mehmood Mhm. 0:54:0.446 –> 0:54:1.646 Suneer Mehmood Yeah, I saw that too. 0:54:3.606 –> 0:54:4.166 Suneer Mehmood Mhm. 0:54:15.886 –> 0:54:16.366 Suneer Mehmood Mhm. 0:54:17.246 –> 0:54:36.606 Brian Haydin you’ll be going to be able to do some damage with it. Is it big enough to run 120 billion parameter model right now? I don’t know. We’re going to see that. You know, that’s akin to, I mean, you know, even the sparks that are out there now, you know, Lance with your Sparky, if you’re listening, 0:54:18.926 –> 0:54:19.646 Suneer Mehmood Absolutely. 0:54:26.206 –> 0:54:26.526 Suneer Mehmood Yeah. 0:54:37.406 –> 0:54:58.566 Brian Haydin You know, you even said last night it struggles with some of that, like, you know, really, really large model consumption. But so it’s not shipping till later this year. It’s not real in my mind. But it is, you know, it is exciting to see this on a Windows device because to get this kind of performance, you almost had to go to 0:54:58.686 –> 0:55:11.246 Brian Haydin you had to go to Apple. And that’s just not something that I like to do myself. I know a lot of other people do. But that’s kind of my take on that. What did you think of the RTX? 0:55:0.246 –> 0:55:0.726 Suneer Mehmood Mmh. 0:55:11.806 –> 0:55:20.406 Suneer Mehmood Really promising. I am, you know, I’m personally I’m planning to get one later once they have it. Yeah. 0:55:17.846 –> 0:55:42.606 Brian Haydin Oh yeah, I’m getting, absolutely. I’ve put myself in, you know, put myself, actually I use like 3 different email addresses just to make sure I got in. We’ll see, we’ll see how they ration that out. All right, so couple things, we are pretty much at time here. So two things that I kind of put on the back burner specifically that 0:55:27.966 –> 0:55:28.366 Suneer Mehmood Right. 0:55:42.926 –> 0:55:52.686 Brian Haydin I was hoping we could get a little bit of time to talk about, we got like 2 minutes before we got to shut her down. So, Suneer, what did, did you take a look at the formulator? 0:55:53.6 –> 0:56:7.566 Suneer Mehmood Yeah, I think, yeah, like you know, the Microsoft Research, open source, agent driven data that connects to OneLake, Custo and Postgres. So, it’s all on GitHub today, so I would, I would highly encourage, go play with it. 0:56:8.366 –> 0:56:28.526 Brian Haydin Yeah, I was walking through a demo around this, and it’s a really, really cool way to be able to spin this stuff up, really, you know, get the orchestration, the ETL orchestration all pulled together, and then start to like ask questions about your data. You know, very, very cool, you know, and they’ve got to get… 0:56:19.806 –> 0:56:20.286 Suneer Mehmood Mhm. 0:56:23.726 –> 0:56:24.206 Suneer Mehmood Yeah. 0:56:28.646 –> 0:56:48.766 Brian Haydin good Git repo. I follow the Microsoft Research Group, you know, on GitHub and what they’re doing. I’m going to start following them a lot more because there were some other things in that lab that were just, you know, you know, blowing my mind what those guys are working on. It’s really fantastic. And shout out to that group too because… 0:56:43.86 –> 0:56:43.366 Suneer Mehmood Yeah. 0:56:45.206 –> 0:56:45.406 Suneer Mehmood Yes. 0:56:49.46 –> 0:57:7.166 Brian Haydin I was talking about my 10-year-old son and lemonade stands and what he wants to invest in. And, you know, they gave me a lot of really good pointers and then gave me, you know, some 3D stuff to take home to the kids. So shout out to you guys. You were awesome. Trellis, you know, this was like, 0:57:8.46 –> 0:57:23.246 Brian Haydin Trellis 2 was also in this research lab, you know, and it was, you know, image to 3D, you know, kind of rendering and building out like 3D models that you could, that you could actually use. It was super fun. Did you do anything with it, Suneer? 0:57:24.206 –> 0:57:29.166 Suneer Mehmood No, that I think that that’s a part I missed, so looking forward to your thoughts on that. 0:57:27.726 –> 0:57:48.286 Brian Haydin Yeah, yeah. So, yeah, it was, so 2 aspects that was kind of cool. And that’s too bad that you missed it because one is like this Trellis, you know, aspect of it. So what I was able to do was use the new Microsoft AI models to prompt and get images generated, like 2D images. 0:57:48.606 –> 0:57:49.206 Suneer Mehmood Mhm. 0:57:48.766 –> 0:58:8.126 Brian Haydin and then use Trellis to actually map it into a 3D model, right? And then it could print the 3D model. So what we did was I made this like, it’s a fish. Everybody knows I like outdoor stuff, so of course it was a fish. But I made this fish and the model, like the image turned out great. 0:57:54.126 –> 0:57:54.286 Suneer Mehmood Mm. 0:58:8.366 –> 0:58:31.486 Brian Haydin And then the model actually turned out really nice and smooth. And they’re like, we’re going to go ahead and print this because most of these models that people are trying to build, like, you know, through trellis, they get kind of sketch a little bit. So I got this postcard that I went and sent the next morning to my house and then got them to print out that fish, you know, in the 3D printer that I could take home. Super fun, super cool. 0:58:12.846 –> 0:58:12.926 Suneer Mehmood Huh? 0:58:31.886 –> 0:58:51.406 Brian Haydin And that’s also open source. You can go and find that on GitHub as well. So that being said, we are at 1159. There was a Paige, you put a link in the chat. If you want to go in a little bit deeper on any of these things that we talked about today, 0:58:37.246 –> 0:58:37.886 Suneer Mehmood Austin. 0:58:51.726 –> 0:59:9.406 Brian Haydin Fill out that survey, let us know what you liked, what you didn’t like, what we should have covered, what we shouldn’t have covered. And maybe if we get enough people that are interested, we’ll do a part 2 of this. So with that being said, thanks so much for joining us today. And we’ll talk soon. 0:59:10.366 –> 0:59:11.726 Suneer Mehmood Thank you so much, everyone.