WEBVTT 00:00:05.240 --> 00:00:07.260 [music] 00:00:10.480 --> 00:00:12.500 [music] 00:00:24.000 --> 00:00:27.557 >> Please welcome Swyx, Sara Goa, A Lad 00:00:27.640 --> 00:00:30.917 Gill, and Chairman and Chief Executive 00:00:31.000 --> 00:00:35.800 Officer of Microsoft, Satya Nadella. 00:00:39.300 --> 00:00:41.320 >> [cheering] 00:00:41.840 --> 00:00:46.397 >> Hello. Satya? >> Uh I'm so excited to be here. Welcome to 00:00:46.480 --> 00:00:51.117 a crossover episode of No Priors and Lain Space with Satya Nadella. Um 00:00:51.200 --> 00:00:53.317 congratulations on an amazing build. 00:00:53.400 --> 00:00:56.557 >> No, thank you so much and it's great to be with both of you. I listen to both of 00:00:56.640 --> 00:00:59.357 you or both the podcasts all the time. 00:00:59.440 --> 00:01:01.157 It's great to be on it. >> Thank you so much. 00:01:01.240 --> 00:01:04.797 >> Sara has been talking about um these amazing uh announcements from across the 00:01:04.880 --> 00:01:09.277 Microsoft estate all morning for I think 3 hours. What is the uh 00:01:09.360 --> 00:01:11.597 what's the most important reflection or takeaway you have? 00:01:11.680 --> 00:01:13.957 >> I I'd say there are 00:01:14.040 --> 00:01:14.557 uh 00:01:14.640 --> 00:01:18.917 perhaps the the biggest one for me is 00:01:19.000 --> 00:01:21.917 let's sort of conceptualize this more as 00:01:22.000 --> 00:01:27.317 an ecosystem play as opposed to a single model or even a single platform, right? 00:01:27.400 --> 00:01:30.357 I mean yeah, whatever I at least for me having 00:01:30.440 --> 00:01:35.157 grown up at Microsoft, having seen whatever four major platform shifts, 00:01:35.240 --> 00:01:38.037 uh I sort of fall into that uh 00:01:38.120 --> 00:01:41.317 uh camp where a platform is defined by 00:01:41.400 --> 00:01:45.437 fundamentally its ability to create more value above the platform versus what's 00:01:45.520 --> 00:01:47.797 captured in the platform. And so if you 00:01:47.880 --> 00:01:50.237 you view what's happening right now, I 00:01:50.320 --> 00:01:55.557 think this morning's keynote was how can any company 00:01:55.640 --> 00:01:57.077 whether it's an AI-native company a 00:01:57.160 --> 00:01:59.557 traditional enterprise company, 00:01:59.640 --> 00:02:03.557 participate as a first-class participant 00:02:03.640 --> 00:02:07.397 where they can point to AI they created. 00:02:07.480 --> 00:02:11.157 Right? It's not that they don't use other people's AI. Of course, they will. 00:02:11.240 --> 00:02:15.557 But to me, what's the path? What's the recipe? How do I do it? What does the 00:02:15.640 --> 00:02:19.557 stack look like? What does the tooling look like? What is valuable? How do you 00:02:19.640 --> 00:02:23.877 do that? That's it. That's sort of our job to do. 00:02:23.960 --> 00:02:29.357 Yeah. >> Ecosystem strategy is uh very complicated, right? Because you end up 00:02:29.440 --> 00:02:32.677 building certain components, partnering for certain components, supporting them. 00:02:32.760 --> 00:02:36.517 You just announced this big suite of models. Like, tell us a little bit about 00:02:36.600 --> 00:02:39.317 the uh training strategy for Microsoft. 00:02:39.400 --> 00:02:42.117 >> Yeah. So So, the thing that we wanted to do with 00:02:42.200 --> 00:02:45.437 the MAI models was to build, and as 00:02:45.520 --> 00:02:50.357 Mustafa talked about, first of all, a great lineage, right? Starting with 00:02:50.440 --> 00:02:52.037 pre-training 00:02:52.120 --> 00:02:55.197 uh with very good data quality, 00:02:55.280 --> 00:02:57.757 uh doing all the ablations, making sure 00:02:57.840 --> 00:03:02.397 because in in some sense, it's become even harder to build a clean lineage 00:03:02.480 --> 00:03:05.597 model. Yes, because there's so much stuff out there 00:03:05.680 --> 00:03:10.317 uh that you truly need to ablate out to be able to have a fantastic pre-trained 00:03:10.400 --> 00:03:12.917 model. In fact, that's one of the challenges of a lot of the open weight 00:03:13.000 --> 00:03:15.277 models is they look great on one 00:03:15.360 --> 00:03:20.037 benchmark or two, but they're not great on practice. So, that's why, in fact, 00:03:20.120 --> 00:03:21.997 even in the RFDs are 00:03:22.080 --> 00:03:25.197 are pretty gone really excited about 00:03:25.280 --> 00:03:27.077 these MAI models because how the heck 00:03:27.160 --> 00:03:31.037 can a small 5B model hill climb? 00:03:31.120 --> 00:03:33.277 Uh and it goes back a little bit to what 00:03:33.360 --> 00:03:35.957 I think is ultimately the key thing to 00:03:36.040 --> 00:03:38.597 do, which is try to pursue finding that 00:03:38.680 --> 00:03:41.437 cognitive core. Uh so, to me, starting 00:03:41.520 --> 00:03:44.957 with a clean lineage, then creating that 00:03:45.040 --> 00:03:48.437 ability for companies to be able to use 00:03:48.520 --> 00:03:51.357 this, right? Not just as a generalist, 00:03:51.440 --> 00:03:56.677 but to create their own specialist by building this hill climbing scaffold 00:03:56.760 --> 00:04:00.437 around it, right? So, it's not just the model, but you have a hill climb 00:04:00.520 --> 00:04:05.077 scaffold around it, then you will start building your RLE. You 00:04:05.160 --> 00:04:08.357 will start collecting the traces. Most importantly, you'll have private emails 00:04:08.440 --> 00:04:12.037 because we know all the emails out there are good, interesting, but they're not 00:04:12.120 --> 00:04:16.237 really that critical at this point because they all can be maxed. And so, 00:04:16.320 --> 00:04:17.837 the point is each company will have its 00:04:17.920 --> 00:04:20.397 own private email. And so, 00:04:20.480 --> 00:04:23.677 that end-to-end platform story 00:04:23.760 --> 00:04:28.477 around our models is sort of uh what I think is interesting. And then the one 00:04:28.560 --> 00:04:30.997 other thing, Sarah, is since you brought that up is 00:04:31.080 --> 00:04:34.757 I do feel there's a new frontier. Like people talk about the frontier and are 00:04:34.840 --> 00:04:36.637 you operating at the frontier? 00:04:36.720 --> 00:04:41.317 Um interestingly enough, if you add a little temporality to it, 00:04:41.400 --> 00:04:45.637 you can use, let's say, in in in fact, that the Land O'Lakes demo we showed was 00:04:45.720 --> 00:04:49.637 pretty cool. We used whatever GPT-55, right? Then you collected a bunch of 00:04:49.720 --> 00:04:52.237 traces, and then you took a 5B reasoning 00:04:52.320 --> 00:04:56.917 model and achieved higher uh so, that is another aspect of what it 00:04:57.000 --> 00:05:00.357 means to appear you know, operate at the frontier. 00:05:00.440 --> 00:05:04.557 >> Yeah. I I think uh I first of all have to congratulate you on basically 00:05:04.640 --> 00:05:08.597 building a frontier neural lab inside of Microsoft in 2 years. Um I'm wondering, 00:05:08.680 --> 00:05:12.237 you know, you have all this AI strategy that you're rolling out. I'm wondering, 00:05:12.320 --> 00:05:15.597 what do you know now that you wish you would tell yourself 2 years ago where or 00:05:15.680 --> 00:05:19.917 2 or 3 years ago? 3 years for the Jensen partnership, 2 years for uh MAI. 00:05:20.000 --> 00:05:22.397 >> Yeah, I mean, I think the the thing with 00:05:22.480 --> 00:05:26.197 that I reflect quite a bit, right? Which is sort of obviously I got into all this 00:05:26.280 --> 00:05:30.317 when I got excited by the the scaling laws paper and, you know, when, you 00:05:30.400 --> 00:05:33.357 know, even the OpenAI partnership came about when those folks said, "Hey, we're 00:05:33.440 --> 00:05:37.357 going to really throw a lot of computer transformers." Uh and they've held, 00:05:37.440 --> 00:05:39.237 right? The thing that I always look back 00:05:39.320 --> 00:05:41.597 and say, "Wow, these things um 00:05:41.680 --> 00:05:44.397 do have capability that they're climbing 00:05:44.480 --> 00:05:48.677 up with I mean, this, you know, this crude way of saying it is intelligence 00:05:48.760 --> 00:05:51.077 is log of compute kind of works. 00:05:51.160 --> 00:05:54.837 Now, what I think we underestimated 00:05:54.920 --> 00:05:58.477 perhaps is the real world complexity 00:05:58.560 --> 00:06:01.477 of deploying these so that they actually 00:06:01.560 --> 00:06:06.117 deliver the value in the real world, right? So the outcomes as measured by 00:06:06.200 --> 00:06:08.037 any benchmark 00:06:08.120 --> 00:06:11.397 is interesting, important, but the true 00:06:11.480 --> 00:06:16.957 eval is when people out there are able to do unique things that they only can 00:06:17.040 --> 00:06:20.597 value. And it's very measurable. 00:06:20.680 --> 00:06:25.277 Right? That I wish we had sort of even like had more in our consciousness, 00:06:25.360 --> 00:06:28.877 right? Which is as an industry because right now I think 00:06:28.960 --> 00:06:30.477 when people say, "Wow, I don't want a 00:06:30.560 --> 00:06:34.197 token max." It's an artifact of us not 00:06:34.280 --> 00:06:37.037 having thought ourselves as an industry 00:06:37.120 --> 00:06:39.477 that we are using tokens to create value 00:06:39.560 --> 00:06:44.237 every step of the way. So I think that's kind of what I wish we had gotten there, 00:06:44.320 --> 00:06:46.797 but I'm glad we are here. >> What are some other use cases that 00:06:46.880 --> 00:06:49.957 you've seen that have created the most value for your customers? Because I know 00:06:50.040 --> 00:06:53.197 that people talk a lot about code and I think it's pretty clear that that's 00:06:53.280 --> 00:06:56.317 something that's having very large-scale impact. Are there other areas that you 00:06:56.400 --> 00:06:58.397 find in common that your customers are really benefiting? 00:06:58.480 --> 00:07:02.877 >> Yeah, I think yeah, to your point, obviously coding is now got but it's 00:07:02.960 --> 00:07:06.557 interesting by the way Elijah to even talk about the coding, right? Which is 00:07:06.640 --> 00:07:10.677 coding is worked so well that we now have to rebuild the IDE, right? I mean 00:07:10.760 --> 00:07:14.917 it's kind of nuts to see what we saw large is like, "Oh my god, I have these 00:07:15.000 --> 00:07:17.397 100 agent sessions. I the cognitive load 00:07:17.480 --> 00:07:20.477 it transfers back to me as a human is so 00:07:20.560 --> 00:07:24.557 excessive that now I need a new UI. Uh 00:07:24.640 --> 00:07:27.797 oh, by the way, I like the the chat as 00:07:27.880 --> 00:07:31.597 the only artifact is also impossible. So that's why we need a canvas. So it's 00:07:31.680 --> 00:07:35.877 kind of interesting for all the things about where is software needed or where 00:07:35.960 --> 00:07:38.277 is UI needed? Uh you kind of need that 00:07:38.360 --> 00:07:43.597 even for code, right? In a fully agentic world. But that said, one of the things 00:07:43.680 --> 00:07:46.477 that we are starting to see, we started seeing with co-work, but even some of 00:07:46.560 --> 00:07:50.597 the work we showed with auto autopilot, 00:07:50.680 --> 00:07:52.717 right, on what you see with claws, is a 00:07:52.800 --> 00:07:54.037 good one, because if you sort of think 00:07:54.120 --> 00:07:58.037 about a lot of human capital is doing 00:07:58.120 --> 00:08:01.037 the glue work, right? If you now can 00:08:01.120 --> 00:08:05.317 augment that with tokens {slash} agents 00:08:05.400 --> 00:08:08.077 that are long-running, durable, 00:08:08.160 --> 00:08:11.837 right, then your ability to scale even 00:08:11.920 --> 00:08:14.397 what is still judgment and glue work 00:08:14.480 --> 00:08:17.157 gets amplified like coding does. 00:08:17.240 --> 00:08:21.037 So you can like I'm positive that 6 months from now we'll all be saying, 00:08:21.120 --> 00:08:24.317 "Oh, wow, like all through night the night there 00:08:24.400 --> 00:08:26.517 was a bunch of stuff that all these 00:08:26.600 --> 00:08:31.637 autopilots that I have working on my behalf with my delegated authority, so 00:08:31.720 --> 00:08:36.117 to speak, right? I can sort of given even my identity, did a bunch of work. 00:08:36.200 --> 00:08:40.677 Then of course I'll need my new ADE to say, "What did you do?" Like that. My 00:08:40.760 --> 00:08:44.477 Did I do this work and so on. So I think that that's where compressing of 00:08:44.560 --> 00:08:48.837 workflows, completing of tasks, that's where I 00:08:48.920 --> 00:08:50.357 think a lot of the value gets created. 00:08:50.440 --> 00:08:53.597 >> you raise a really interesting point, which is there's the actual agent is 00:08:53.680 --> 00:08:56.997 doing the code, and then there's a harness around it. And that's the 00:08:57.080 --> 00:08:59.637 environment, that's the context, that's everything you're setting up as a 00:08:59.720 --> 00:09:02.957 developer around actually a coding agent. What is the harness for the 00:09:03.040 --> 00:09:07.197 enterprise? Is there an equivalent concept for broader productivity work, 00:09:07.280 --> 00:09:09.237 or how do you think about that concept sort of generally? 00:09:09.320 --> 00:09:11.517 >> So so some sense you kind of want the 00:09:11.600 --> 00:09:15.957 harness to define the models, the the 00:09:16.040 --> 00:09:19.757 data, and the tools, and so that you have a 00:09:19.840 --> 00:09:24.197 loop across those three. And so what we are trying to first of all make sure is 00:09:24.280 --> 00:09:27.997 each of our products that we build, right, whether it's GitHub Copilot or 00:09:28.080 --> 00:09:32.397 the security copilot the stuff we showed with M-Dash, or even the discovery for 00:09:32.480 --> 00:09:37.317 science, it doesn't matter. All of them are multimodal harnesses 00:09:37.400 --> 00:09:40.437 um tools access so that you can do this 00:09:40.520 --> 00:09:42.997 progressive uh disclosure of tools even 00:09:43.080 --> 00:09:45.157 so that they're token efficient. Uh and 00:09:45.240 --> 00:09:47.917 then you're feeding it with very rich 00:09:48.000 --> 00:09:50.557 context because that's sort of the other 00:09:50.640 --> 00:09:52.637 hard lesson we've learned in the last 2 00:09:52.720 --> 00:09:55.997 years is oh my god the amount of work 00:09:56.080 --> 00:10:00.677 you need to do to prep the context layer 00:10:00.760 --> 00:10:04.157 uh such that your plan can execute in 00:10:04.240 --> 00:10:08.437 the most efficient way is where the magic is. So we have in our case we have 00:10:08.520 --> 00:10:12.037 the get up harness which essentially we're using across all our products. 00:10:12.120 --> 00:10:15.797 It's available in foundry and we're open like you can use your llama harness 00:10:15.880 --> 00:10:18.757 whatever or you can use the um 00:10:18.840 --> 00:10:21.077 you know any open harness or any harness 00:10:21.160 --> 00:10:25.917 of yours and train with your tools and multiple models and your context. And so 00:10:26.000 --> 00:10:30.517 that's the pitch because right now a lot of dialogue is um hey if I train the 00:10:30.600 --> 00:10:32.397 harness plus tools and the model 00:10:32.480 --> 00:10:35.077 together you get evals. And what we're 00:10:35.160 --> 00:10:37.837 proving out is and the best example of 00:10:37.920 --> 00:10:40.357 that is what we did with M dash right 00:10:40.440 --> 00:10:42.717 because when it launched uh 00:10:42.800 --> 00:10:46.157 it found bugs or vulnerabilities that 00:10:46.240 --> 00:10:48.197 were not found by mythos. 00:10:48.280 --> 00:10:51.597 Uh and so there is existence proof I 00:10:51.680 --> 00:10:55.677 would claim that you can have a multimodal harness 00:10:55.760 --> 00:10:58.317 uh that can in fact be more uh 00:10:58.400 --> 00:11:01.157 performant in the real world. 00:11:01.240 --> 00:11:04.997 >> So premise behind the uh training at the 00:11:05.080 --> 00:11:09.757 independent frontier labs is really you know we're going to have these models 00:11:09.840 --> 00:11:13.837 and we'll have an API business and we'll support enterprise and startups but 00:11:13.920 --> 00:11:18.237 a first party product be it productivity or code or search drives the majority of 00:11:18.320 --> 00:11:21.517 revenue. That's a different value equation than you're describing. I think 00:11:21.600 --> 00:11:26.077 with the Microsoft ecosystem uh if if that's the case tell me if it's the case 00:11:26.160 --> 00:11:29.077 cuz obviously you have first party products and you have enablement 00:11:29.160 --> 00:11:32.917 products. Um what is the role of the develop like what's going to be hard and 00:11:33.000 --> 00:11:36.237 the set of skills and the value capture the developer has in that world? 00:11:36.320 --> 00:11:39.557 >> Yeah, so I think that there's always going to be the case that 00:11:39.640 --> 00:11:44.517 someone who's super successful and as a platform builder can also have 00:11:44.600 --> 00:11:46.277 first-party products. 00:11:46.360 --> 00:11:48.877 It was true with Windows, it was true uh 00:11:48.960 --> 00:11:51.757 with uh the the SaaS side and the cloud 00:11:51.840 --> 00:11:53.797 side as well with us and others and so 00:11:53.880 --> 00:11:56.917 on. But, the thing that is is it should 00:11:57.000 --> 00:11:59.837 not be a limiter to other people 00:11:59.920 --> 00:12:01.757 achieving that same success, right? That 00:12:01.840 --> 00:12:04.077 I think is the core difference, which is 00:12:04.160 --> 00:12:07.197 the the network effects this time around 00:12:07.280 --> 00:12:09.517 around intelligence are such because 00:12:09.600 --> 00:12:12.917 they learn from data and not really lots 00:12:13.000 --> 00:12:15.917 of data. It's just the few samples that 00:12:16.000 --> 00:12:17.877 you have to see to understand what's 00:12:17.960 --> 00:12:20.917 novel about something. So, that's why 00:12:21.000 --> 00:12:24.077 the game becomes how to protect, so 00:12:24.160 --> 00:12:26.637 that's why I would say every company 00:12:26.720 --> 00:12:31.117 having private evals maybe the biggest IP. Right? I think about it. Like, 00:12:31.200 --> 00:12:35.157 what's that private eval that you can then use even a frontier model to hill 00:12:35.240 --> 00:12:38.077 climb on and not leak the traces maybe 00:12:38.160 --> 00:12:40.277 one of the biggest drivers 00:12:40.360 --> 00:12:42.237 uh of IP. Like, so in other words, 00:12:42.320 --> 00:12:45.477 another acid test is you have an eval 00:12:45.560 --> 00:12:49.157 that's private. You're using a a model 00:12:49.240 --> 00:12:53.517 A. Can you switch it to model B and you 00:12:53.600 --> 00:12:58.437 know, climb up? If you can, then you're in control. If you can't, you're not in 00:12:58.520 --> 00:13:02.277 control. And that's where even the harness becomes super important, right? 00:13:02.360 --> 00:13:04.557 So, therefore, having an open harness, 00:13:04.640 --> 00:13:07.237 letting all models come in, having your 00:13:07.320 --> 00:13:10.877 evals, your contacts, your tools help 00:13:10.960 --> 00:13:13.597 you hill climb, I think is the skills 00:13:13.680 --> 00:13:18.917 that an AI native startup needs, a SaaS company needs, or every enterprise 00:13:19.000 --> 00:13:23.757 needs. >> Yeah, I think in a very real way, you are Microsoft historically as an 00:13:23.840 --> 00:13:27.517 operating systems company and then becoming a cloud company. Maybe like the 00:13:27.600 --> 00:13:31.037 third act is that you're harness or eval's 00:13:31.120 --> 00:13:34.877 company. Whatever whatever the the sort of conglomerate of concepts that you 00:13:34.960 --> 00:13:39.637 want to put together. Um I I I think like enabling every company to have like 00:13:39.720 --> 00:13:43.197 frontier intelligence or what what I forget the the exact term that you used. 00:13:43.280 --> 00:13:47.917 Um is the is the mission, right? That is that is the platform promise that you 00:13:48.000 --> 00:13:51.437 build with us, you will get your intelligence for your data. 00:13:51.520 --> 00:13:53.997 >> That's it. That to me that is the like 00:13:54.080 --> 00:13:56.917 if there was one tagline for this entire 00:13:57.000 --> 00:13:58.797 developer conference is 00:13:58.880 --> 00:14:02.357 can everybody operate at the frontier 00:14:02.440 --> 00:14:04.517 with their frontier intelligence, right? 00:14:04.600 --> 00:14:07.437 To me that is so important because otherwise 00:14:07.520 --> 00:14:11.837 I I don't know how you achieve stable equilibrium, right? Which is how do I 00:14:11.920 --> 00:14:14.157 then go and say, "Wow, my company is 00:14:14.240 --> 00:14:16.717 going to have a terminal value because I 00:14:16.800 --> 00:14:20.317 now know how to continuously compound on 00:14:20.400 --> 00:14:23.997 top of what's a platform that gets better, right?" So when like Windows 00:14:24.080 --> 00:14:25.917 obviously came out, 00:14:26.000 --> 00:14:29.197 Adobe built, Autodesk built, 00:14:29.280 --> 00:14:34.317 or even like take but Jensen said, "We built DX." And he built, you know, CUDA 00:14:34.400 --> 00:14:37.677 on top of it. Right? I mean, I always say to Jensen, 00:14:37.760 --> 00:14:40.557 "God, I got the short end of that." But I wish 00:14:40.640 --> 00:14:42.237 we had recognized it. But nevertheless, 00:14:42.320 --> 00:14:45.557 but that idea that you can build a 00:14:45.640 --> 00:14:50.117 platform layer that someone else can then extend out 00:14:50.200 --> 00:14:53.397 and build their own intelligence layer in this case, I think is everything, 00:14:53.480 --> 00:14:56.917 right? Without it, why have a developer conference? I can just come and have you 00:14:57.000 --> 00:15:01.557 all sort of just worship at the altar of one model. But that's not a developer 00:15:01.640 --> 00:15:04.437 conference. >> Uh backstage we you had a discussion 00:15:04.520 --> 00:15:08.757 about what is IP or what is the the value in a company. It used to be the 00:15:08.840 --> 00:15:11.197 length of human experience at a company. 00:15:11.280 --> 00:15:15.157 And now it's this other thing, which is the eval's, the 00:15:15.240 --> 00:15:18.477 experience in sort of applying agents to the company. He I just want you to like 00:15:18.560 --> 00:15:21.157 flesh that out a little bit more cuz >> Yeah, it's It's way to frame it, right? 00:15:21.240 --> 00:15:24.717 Because you have at the end of the day every company is going to have both the 00:15:24.800 --> 00:15:26.797 human capital that is still going to be 00:15:26.880 --> 00:15:29.717 super valuable because humans 00:15:29.800 --> 00:15:33.677 and their ability to find the gaps that 00:15:33.760 --> 00:15:37.917 exist at all times is going to be the way we all will create value, right? I 00:15:38.000 --> 00:15:40.157 mean so I'm definitely in the camp that 00:15:40.240 --> 00:15:42.717 this is going to be about expressing new 00:15:42.800 --> 00:15:48.157 forms of human agency and ambition even as token capital goes up, right? So 00:15:48.240 --> 00:15:51.717 let's say any corporation has lots of tokens and 00:15:51.800 --> 00:15:54.157 lot of human capital. The question is 00:15:54.240 --> 00:15:59.237 how do you compound the two? So if you have a like if you take in teams, I have 00:15:59.320 --> 00:16:02.997 a bunch of agents doing work and a bunch of humans doing work and the traces 00:16:03.080 --> 00:16:04.477 between those 00:16:04.560 --> 00:16:07.557 that is really important context of the 00:16:07.640 --> 00:16:10.397 how that enterprise is creating value. 00:16:10.480 --> 00:16:13.077 Then that goes back to train not a 00:16:13.160 --> 00:16:15.837 generalist model, but to train the train 00:16:15.920 --> 00:16:20.637 the company veteran agent. Right, that is super valuable again, 00:16:20.720 --> 00:16:23.757 right? Which is when a company goes says it should in fact go on to the balance 00:16:23.840 --> 00:16:25.437 sheet is how I think about it, right? 00:16:25.520 --> 00:16:29.797 That's what in fact they may be like human capital was never possible to go 00:16:29.880 --> 00:16:33.277 put on a balance sheet because you didn't know how to capture 00:16:33.360 --> 00:16:37.317 the tacit knowledge, whereas now I think you can with the agents that have 00:16:37.400 --> 00:16:41.517 learned through the through time, through all the traces. 00:16:41.600 --> 00:16:43.877 So that's what at least we think will happen. 00:16:43.960 --> 00:16:47.997 >> I think the SEC is going to have to have accounting standards for token 00:16:48.080 --> 00:16:52.957 expertise. >> Uh you're talking about the equilibrium 00:16:53.040 --> 00:16:56.157 state and a stable equilibrium where companies 00:16:56.240 --> 00:17:00.237 have this compounding value and can see terminal value for themselves. Another 00:17:00.320 --> 00:17:02.717 challenge to you know, the considered 00:17:02.800 --> 00:17:04.877 equilibrium of okay, there are 00:17:04.960 --> 00:17:07.597 applications and workflows that are sort 00:17:07.680 --> 00:17:12.317 of common to a vertical or a horizontal and this was like the generation of SaaS 00:17:12.400 --> 00:17:16.157 companies and you know, Microsoft has lots of SaaS properties as well. And 00:17:16.240 --> 00:17:18.477 then there are things that are very specific to every enterprise that 00:17:18.560 --> 00:17:20.397 they're differentiated against. 00:17:20.480 --> 00:17:23.437 Um I'm sure you have heard much and participated much of the debate about 00:17:23.520 --> 00:17:28.397 the end of software because all these workflows are are cheap to generate now. 00:17:28.480 --> 00:17:32.557 Um do you think the equilibrium looks different between what agents get built 00:17:32.640 --> 00:17:35.797 in enterprises versus in their vendors in the future? 00:17:35.880 --> 00:17:41.117 >> Yeah, so I think what's happening there is see we we had a particular way we 00:17:41.200 --> 00:17:45.837 captured uh I would say workflow 00:17:45.920 --> 00:17:50.717 in apps, right? Because we built a a data model, right? We schematized some 00:17:50.800 --> 00:17:52.517 part of some business process. 00:17:52.600 --> 00:17:55.197 >> Mhm. >> We then built a bunch of business logic. 00:17:55.280 --> 00:17:57.637 >> Yep. >> And then we put a bunch of UI on top of 00:17:57.720 --> 00:18:00.877 it, right? So that's kind of what every SaaS company >> configuration, right? 00:18:00.960 --> 00:18:02.597 >> 20 20 years that was >> Right. 00:18:02.680 --> 00:18:07.037 >> And that was it. So interestingly enough, now you kind of get to 00:18:07.120 --> 00:18:11.917 re-litigate that vertical stacking, right? So I still think, for example, 00:18:12.000 --> 00:18:15.997 that data model that you built underneath every SaaS application is 00:18:16.080 --> 00:18:20.317 super good, right? It like why reinvent it? Like I I my general ledger better be 00:18:20.400 --> 00:18:23.837 a general ledger. I don't need new schema creation. 00:18:23.920 --> 00:18:26.997 >> Yeah. >> Uh in fact, that entity relationship 00:18:27.080 --> 00:18:30.837 uh is actually pretty good robust thing that I want to feed. 00:18:30.920 --> 00:18:31.877 >> And you want it to be stable. 00:18:31.960 --> 00:18:34.797 >> That's right. >> Yeah. >> Then same thing with business logic, 00:18:34.880 --> 00:18:37.117 right? If If you look at uh we have this 00:18:37.200 --> 00:18:38.957 product called Power BI, right? It is 00:18:39.040 --> 00:18:42.597 like dashboards galore people created. 00:18:42.680 --> 00:18:44.917 The beauty underneath that dashboard is 00:18:45.000 --> 00:18:47.317 a very rich semantic model, right? 00:18:47.400 --> 00:18:49.757 Someone took the pain to create a 00:18:49.840 --> 00:18:53.037 dashboard and do all the measures. 00:18:53.120 --> 00:18:55.437 And you want that that's business logic, 00:18:55.520 --> 00:18:57.637 right? I want that to be available to 00:18:57.720 --> 00:19:01.077 me. So I think the challenge of the SaaS 00:19:01.160 --> 00:19:04.197 business model is we packaged one way. 00:19:04.280 --> 00:19:09.157 We now have to learn how to unbundle these things and re-bundle in new ways 00:19:09.240 --> 00:19:10.997 and discover new business models, right? 00:19:11.080 --> 00:19:15.357 I mean if you look at it, te- what's happening today with Microsoft 365 is a 00:19:15.440 --> 00:19:17.997 great example. Right? We have this thing called Work 00:19:18.080 --> 00:19:22.717 IQ. In fact, like what we are realizing is, oh my god, like you know, if you 00:19:22.800 --> 00:19:26.717 look at In fact, there's a historical parallel to, right? We sold first 00:19:26.800 --> 00:19:30.597 Exchange and SharePoint and uh you know, before Teams, we had a 00:19:30.680 --> 00:19:33.917 thing called Link Server and what have you. And we thought, oh, that's all 00:19:34.000 --> 00:19:38.117 going to move to the cloud, but little did we realize that oh the number of 00:19:38.200 --> 00:19:42.677 people who will use servers in the cloud is 10x 100x, right? Because people were 00:19:42.760 --> 00:19:45.677 not buying servers, they were just buying a subscription. 00:19:45.760 --> 00:19:47.277 The same thing is now happening with 00:19:47.360 --> 00:19:50.477 M365 because with Work IQ, 00:19:50.560 --> 00:19:55.357 we have exposed what is perhaps the most important database in a company that 00:19:55.440 --> 00:20:00.317 never got used as a database because it is only captive to our apps, right? It 00:20:00.400 --> 00:20:04.157 is all email operated on it, Teams operated on it, Word, Excel, PowerPoint, 00:20:04.240 --> 00:20:06.637 SharePoint. But now, like this is one of 00:20:06.720 --> 00:20:11.277 the coolest things I get to do with Work IQ. I go to a GitHub repo and I say, 00:20:11.360 --> 00:20:13.477 "Hey, I attended a bunch of design 00:20:13.560 --> 00:20:16.397 meetings last week related to this repo. 00:20:16.480 --> 00:20:19.957 Can you capture all that and tell me what changes I should make?" I mean, 00:20:20.040 --> 00:20:24.197 think about that. Right? It literally can go look at all those transcripts, 00:20:24.280 --> 00:20:26.357 come back with a plan to change a code 00:20:26.440 --> 00:20:28.597 base. Right? Previously, you could never 00:20:28.680 --> 00:20:31.637 have thought of using M365 for something 00:20:31.720 --> 00:20:33.957 like that. So, the value creation 00:20:34.040 --> 00:20:36.877 opportunity now in the agent world is in 00:20:36.960 --> 00:20:41.237 fact 10x more. But it does require us to have, for example, there's going to be 00:20:41.320 --> 00:20:43.637 usage around M365, right? Which is going 00:20:43.720 --> 00:20:46.477 to be perhaps more than even the end 00:20:46.560 --> 00:20:48.437 users. And we have to even re-architect. 00:20:48.520 --> 00:20:51.197 Like in fact, like what I used to serve 00:20:51.280 --> 00:20:55.277 an inbox or a mailbox cannot be used to 00:20:55.360 --> 00:20:59.357 serve an agent. Uh and so, that's sort of what we're doing. 00:20:59.440 --> 00:21:03.157 >> I don't believe in like permanent business models for any of these 00:21:03.240 --> 00:21:06.877 domains, but in the near term, do you have a prediction between 00:21:06.960 --> 00:21:11.837 uh you know outcomes based pricing token based pricing 00:21:11.920 --> 00:21:15.597 enterprise bundles >> Yeah the way I think about this is 00:21:15.680 --> 00:21:19.157 always we have like let's even take the per user pricing. 00:21:19.240 --> 00:21:23.397 >> Mhm. >> The per user pricing is really an 00:21:23.480 --> 00:21:27.597 artifact of someone creating a budget needing certainty. 00:21:27.680 --> 00:21:31.437 Right because it's the most important thing like somebody wants a budget. 00:21:31.520 --> 00:21:33.957 >> Mhm. >> They need a per user and and per user is 00:21:34.040 --> 00:21:37.517 just a set of entitlements to usage. 00:21:37.600 --> 00:21:38.997 Right that's kind of what it is and so 00:21:39.080 --> 00:21:41.677 the way is if the first bundling will be 00:21:41.760 --> 00:21:46.477 take some usage bundle it into per user stacks and you know then sell 00:21:46.560 --> 00:21:50.157 subscriptions. So subscriptions I think are going to be there per user is going 00:21:50.240 --> 00:21:53.917 to be there. Then the next big thing will be consumption. So people will say 00:21:54.000 --> 00:21:57.517 I want consumption. And it's also possible that people will say I don't 00:21:57.600 --> 00:22:00.397 even want to pay for any of the subscriptions or the consumptions 00:22:00.480 --> 00:22:04.557 outcome. But remember most people love outcomes until they 00:22:04.640 --> 00:22:07.317 have an outcome. Because once you have an outcome it's like giving away 00:22:07.400 --> 00:22:11.717 royalty. Right I mean like I've talked to customers who love you know outcome 00:22:11.800 --> 00:22:16.117 based pricing and I say I'm all in until they oh my god like what are you talking 00:22:16.200 --> 00:22:19.597 about you're sharing in my outcome no no no I want you to go back to per user 00:22:19.680 --> 00:22:23.277 pricing and I want you to consumption price. Right so I think that debate will 00:22:23.360 --> 00:22:26.757 go on and all all the all of these business 00:22:26.840 --> 00:22:28.637 models have a particular time and a 00:22:28.720 --> 00:22:32.197 place versus one to rule them all. And 00:22:32.280 --> 00:22:36.557 if anything if you're a SAS vendor or you're a platform vendor having that 00:22:36.640 --> 00:22:40.557 flexibility and quite frankly we face this with GitHub. Right we just recently 00:22:40.640 --> 00:22:43.157 announced a per user pricing on GitHub. 00:22:43.240 --> 00:22:45.037 Because little you know 00:22:45.120 --> 00:22:48.517 GitHub co-pilot was constructed at a per 00:22:48.600 --> 00:22:52.037 user level before we understood even 00:22:52.120 --> 00:22:54.877 the intensity of usage of agents. Right 00:22:54.960 --> 00:22:59.157 it was an interactive way for a developer to use code complete maybe 00:22:59.240 --> 00:23:01.637 task. It was not like oh I launched 00:23:01.720 --> 00:23:04.517 10,000 you know, agents that are going 00:23:04.600 --> 00:23:09.637 on all day. Right? So, that is what the adjustment is about. So, now that we 00:23:09.720 --> 00:23:13.317 really want there will always be a per user, but there will have to be a 00:23:13.400 --> 00:23:16.117 consumption meter. >> How do you think about the durability of 00:23:16.200 --> 00:23:19.477 SaaS more generally? One thing I've observed is in a lot of enterprises 00:23:19.560 --> 00:23:23.677 internally, there'll be teams that almost have agent euphoria. They're so 00:23:23.760 --> 00:23:26.357 excited about the explosion of things they can build that they're trying to 00:23:26.440 --> 00:23:29.037 rebuild a lot of applications or going to other SaaS vendors and saying we're 00:23:29.120 --> 00:23:32.517 not going to work with you anymore or we're considering an internal project. 00:23:32.600 --> 00:23:35.197 And it seems like in 6 to 9 months maybe some of those people will come back and 00:23:35.280 --> 00:23:38.197 say actually we we can't rebuild everything. How do you think about 00:23:38.280 --> 00:23:40.397 what's durable in this world and what isn't? 00:23:40.480 --> 00:23:42.717 >> I I think I think we have to go through 00:23:42.800 --> 00:23:46.637 one full budget cycle on this to really 00:23:46.720 --> 00:23:52.397 see the uh uh the sort of the emergence of the equilibrium. Because at the end of the 00:23:52.480 --> 00:23:56.317 day there's marginal cost to even generating 00:23:56.400 --> 00:23:57.797 the app. Right? So, in fact, there can 00:23:57.880 --> 00:23:59.997 be even a a simple way to say it like if 00:24:00.080 --> 00:24:03.357 you should always acquire something if 00:24:03.440 --> 00:24:07.117 the marginal cost of building and maintaining 00:24:07.200 --> 00:24:09.957 uh something on your own is higher. 00:24:10.040 --> 00:24:13.957 Uh right? That should be like it's a quantifiable right? A quantifiable 00:24:14.040 --> 00:24:18.037 thing. And the maintenance part is important. Right? Even like you got to 00:24:18.120 --> 00:24:22.437 remember like hey, you know, all the security stuff that now AI will find you 00:24:22.520 --> 00:24:26.277 better fix them too fast. Uh of course, there's a coding agent to help you with, 00:24:26.360 --> 00:24:30.197 but then that burns tokens. Right? So, whose responsibility is it? It's kind of 00:24:30.280 --> 00:24:34.677 like a a cycle that you've got to think through. And I think we have gone 00:24:34.760 --> 00:24:38.477 through the excitement that I can generate a lot of software. I think the 00:24:38.560 --> 00:24:43.077 next thing would be what software do I really want to generate? What software 00:24:43.160 --> 00:24:45.397 do I want to use from others? How do I 00:24:45.480 --> 00:24:50.797 compose these two into some agentic workflow that I have agency over? Right? 00:24:50.880 --> 00:24:55.757 Because I think there'll be very little tolerance for anybody who's inflexible 00:24:55.840 --> 00:24:59.837 uh at the vendor level. Uh but at the same time, I think that 00:24:59.920 --> 00:25:04.797 anyone who has got that flexibility shows up, delivers the value, will be 00:25:04.880 --> 00:25:08.597 back at again, right? We're selling software, but with just different 00:25:08.680 --> 00:25:13.037 business models, in fact. >> Uh speaking about building software, um 00:25:13.120 --> 00:25:15.997 one of my favorite moments from I think a previous build, maybe 1 or 2 years 00:25:16.080 --> 00:25:19.077 ago, was they had a they they there was a section of you building your own 00:25:19.160 --> 00:25:21.597 software. I'm curious if you're building anything now. 00:25:21.680 --> 00:25:26.317 >> Yeah, so I I think the you know, first of all, let's face it, right? Building 00:25:26.400 --> 00:25:31.317 software has made it possible for even the incompetence of a CEO of a company like 00:25:31.400 --> 00:25:32.157 ours, [laughter] 00:25:32.240 --> 00:25:34.517 uh you can build. So, thank God. But 00:25:34.600 --> 00:25:39.237 that said, I I I I do feel that 00:25:39.320 --> 00:25:41.757 you know, something like um 00:25:41.840 --> 00:25:44.317 GitHub Copilot to me, and especially the 00:25:44.400 --> 00:25:47.397 new sessions app or the new app, has 00:25:47.480 --> 00:25:51.037 just made it so much more possible 00:25:51.120 --> 00:25:54.397 for you to have agency over artifacts 00:25:54.480 --> 00:25:56.157 that you felt you couldn't touch before, 00:25:56.240 --> 00:26:01.437 right? So, to for me as a CEO even, to go to a codebase, uh to be able to learn 00:26:01.520 --> 00:26:05.037 about it, like I remember joining Microsoft long back, you know, first, 00:26:05.120 --> 00:26:07.517 and then you say, "Well, everybody had to go in and look at, you know, 00:26:07.600 --> 00:26:12.517 whatever, Cutler's Mallet or what have you, to learn how to do good C C++ 00:26:12.600 --> 00:26:13.877 code." Um 00:26:13.960 --> 00:26:16.437 so, now that ability to be more full 00:26:16.520 --> 00:26:20.517 stack up and down is so good. But that 00:26:20.600 --> 00:26:24.157 doesn't mean every one of us should be doing the same thing. The question is, 00:26:24.240 --> 00:26:28.437 how do you then have the ability to inspect things, learn things, see 00:26:28.520 --> 00:26:31.197 things, um I think it's just so much more. And 00:26:31.280 --> 00:26:33.357 so, to me, what I'm building a lot of is 00:26:33.440 --> 00:26:36.357 these long-running foundry agents, 00:26:36.440 --> 00:26:38.637 uh right? So, there's autopilots. So, 00:26:38.720 --> 00:26:41.037 the easiest thing is, to me, I think I 00:26:41.120 --> 00:26:44.037 just built one uh even last week, where 00:26:44.120 --> 00:26:47.757 the idea was, hey, can I have an agent 00:26:47.840 --> 00:26:50.557 that is continuously monitoring, 00:26:50.640 --> 00:26:53.157 essentially, my own chief of staff 00:26:53.240 --> 00:26:55.117 autopilot, right? We're going to have 00:26:55.200 --> 00:26:59.117 that obviously in Scout. That's what 00:26:59.200 --> 00:27:01.677 we showed, but it is so easy and trivial 00:27:01.760 --> 00:27:04.077 to build. I took it work IQ. I said take 00:27:04.160 --> 00:27:06.437 work IQ, go 00:27:06.520 --> 00:27:09.037 and build a foundry long-running agent, 00:27:09.120 --> 00:27:11.837 store all the memory in 00:27:11.920 --> 00:27:16.237 using Raven, right? Basically at my back end as a service. And lo and behold, it 00:27:16.320 --> 00:27:18.437 built it and not only built it, I could 00:27:18.520 --> 00:27:23.797 say publish to Teams and it published the damn thing to Teams. So, the ability 00:27:23.880 --> 00:27:27.397 to have you know some end-to-end project like 00:27:27.480 --> 00:27:29.517 this complete is just pretty miraculous. 00:27:29.600 --> 00:27:32.477 >> How do you think that impacts the different types of engineering roles 00:27:32.560 --> 00:27:36.277 that exist in the future because right now I think there's you know a dozen 00:27:36.360 --> 00:27:40.597 different types of engineers that you can be from QA, front end, etc. You 00:27:40.680 --> 00:27:43.917 know, there's a big swath. I've heard some people argue that in four or five 00:27:44.000 --> 00:27:47.237 years we'll basically end up with four engineering roles. It'll be people who 00:27:47.320 --> 00:27:51.597 are managing agents. It'll be forward deployment engineers or FTEs. It'll be 00:27:51.680 --> 00:27:55.437 security engineers and then people working on large-scale infrastructure 00:27:55.520 --> 00:27:58.877 for a small number of services and then everything else just collapses into the 00:27:58.960 --> 00:28:01.597 agentic world. >> Do you think that's a correct view of the world? 00:28:01.680 --> 00:28:05.757 >> Yeah, I mean I think I think we'll have to experiment our way through it. But 00:28:05.840 --> 00:28:08.357 what you said is what there are some 00:28:08.440 --> 00:28:10.757 very at scale things. At LinkedIn, they 00:28:10.840 --> 00:28:14.437 did structurally change 00:28:14.520 --> 00:28:17.917 and it you know basically built up a new discipline called full stack builder, 00:28:18.000 --> 00:28:21.957 right? So, they went and said, "Hey, let's bring people from design and 00:28:22.040 --> 00:28:26.557 product management, front end engineering, all put them together, 00:28:26.640 --> 00:28:30.157 but also have an edge, right? It's not like the design person still doesn't 00:28:30.240 --> 00:28:33.157 have the design edge or the front end person doesn't have the front end edge, 00:28:33.240 --> 00:28:36.077 but you can give yourself bigger scope 00:28:36.160 --> 00:28:37.997 in role so that you're not confined to 00:28:38.080 --> 00:28:40.677 one role. And then 00:28:40.760 --> 00:28:45.077 equally infrastructure has become very critical, right? So, in other words, 00:28:45.160 --> 00:28:47.677 like I mean our R&D team, I mean one thing we've 00:28:47.760 --> 00:28:51.117 is even for the Excel team, for example, 00:28:51.200 --> 00:28:55.077 building the RLE in which a reward can 00:28:55.160 --> 00:29:00.037 be learned is actually one of the hardest sort of infrastructure problems. 00:29:00.120 --> 00:29:03.637 And so you kind of need even new talent, right? Distributed systems people even 00:29:03.720 --> 00:29:06.957 in what was considered an end-user app team 00:29:07.040 --> 00:29:10.477 because it's a different skill set. So yes, infrastructure science is the other 00:29:10.560 --> 00:29:14.077 one, obviously. So I think we'll see how these evolve, 00:29:14.160 --> 00:29:17.037 right? Where's the reals, I mean, always the world will have a bunch of 00:29:17.120 --> 00:29:18.877 specialists. 00:29:18.960 --> 00:29:23.677 You know, I think the generalist role is 00:29:23.760 --> 00:29:25.637 going to be the most exciting, right? 00:29:25.720 --> 00:29:29.477 Because the leverage of a generalist 00:29:29.560 --> 00:29:33.317 is where we're going to see the maximum returns, right? When when you said, 00:29:33.400 --> 00:29:38.877 "Hey, I'm coding. I'm now a general like what I've basically translated knowledge 00:29:38.960 --> 00:29:40.037 work, 00:29:40.120 --> 00:29:42.197 right, which I did where I created a 00:29:42.280 --> 00:29:46.637 Word document or a spreadsheet or even 00:29:46.720 --> 00:29:48.557 and now I can build an app, right? It's 00:29:48.640 --> 00:29:51.117 in the same sentence, 00:29:51.200 --> 00:29:56.117 right? That idea that, "Oh wow, my generalist skills have gotten a higher 00:29:56.200 --> 00:29:59.077 leverage." I think is what we're going to see across the board. 00:29:59.160 --> 00:30:03.597 >> Music to the ears of CEOs and VCs that are like a little dangerous and a lot of 00:30:03.680 --> 00:30:04.997 fun. >> Golden age for idea people. 00:30:05.080 --> 00:30:09.157 >> Idea people with a lot of agency. If you take that idea of personal agency and 00:30:09.240 --> 00:30:13.717 you just zoom it out to the organizational context, um 00:30:13.800 --> 00:30:17.917 uh my partner Mike Vernal, who actually started his career at Microsoft, just 00:30:18.000 --> 00:30:21.157 wrote an essay where one of the big takeaways is it's an age where you can 00:30:21.240 --> 00:30:25.597 be much more ambitious and you need to be given the pace of the environment and 00:30:25.680 --> 00:30:28.157 how quickly actually users and companies 00:30:28.240 --> 00:30:30.357 are open to adopting new technologies. 00:30:30.440 --> 00:30:34.957 Um how do you think about I feel silly asking this of somebody running a, you 00:30:35.040 --> 00:30:37.957 know, trillion-dollar-plus company already, but how do you think about how 00:30:38.040 --> 00:30:40.557 Microsoft can be more ambitious now? 00:30:40.640 --> 00:30:44.637 >> It's a great question. Um I think um 00:30:44.720 --> 00:30:46.997 I think the the thing in these type of 00:30:47.080 --> 00:30:48.837 transitions 00:30:48.920 --> 00:30:53.357 is to have a conceptual model 00:30:53.440 --> 00:30:55.677 of how 00:30:55.760 --> 00:30:58.637 work can change 00:30:58.720 --> 00:31:01.877 to go after outcomes that you could 00:31:01.960 --> 00:31:06.637 hardly imagine previously, right? In fact, Kevin Scott has this nice line, 00:31:06.720 --> 00:31:10.837 right? Which is um when you can make the impossible like 00:31:10.920 --> 00:31:13.317 when you're making hard things easier 00:31:13.400 --> 00:31:18.197 that's sort of one point of leverage but true ambition is about making the 00:31:18.280 --> 00:31:21.437 impossible possible. So 00:31:21.520 --> 00:31:24.437 now the thing that is missing a little 00:31:24.520 --> 00:31:26.597 bit in all of our organizations is what 00:31:26.680 --> 00:31:30.117 is that new conceptual model of what can 00:31:30.200 --> 00:31:33.717 we build? What was impossible and what can we build? 00:31:33.800 --> 00:31:37.357 And I'll give you one example of this, right? Which is I take great inspiration 00:31:37.440 --> 00:31:41.597 from sort of the people who were managing the Azure net network. 00:31:41.680 --> 00:31:44.717 And they came to the This was from even last year. 00:31:44.800 --> 00:31:48.437 You know, we were scaling. You saw that I I talked about sort of how we built in 00:31:48.520 --> 00:31:53.717 the last 15 months more Azure capacity than we built in the first 15 years. I 00:31:53.800 --> 00:31:56.157 mean, it's crazy. Right? It's pretty wild. And it's the 00:31:56.240 --> 00:31:59.557 same team. So they saw that and they said, "Bob, this just ain't going to 00:31:59.640 --> 00:32:02.877 work if we don't reconceptualize our work." 00:32:02.960 --> 00:32:05.597 So they built Essentially, they said, 00:32:05.680 --> 00:32:07.997 "Our job is not to do Azure networking. 00:32:08.080 --> 00:32:10.477 Our job is to build the agentic system 00:32:10.560 --> 00:32:12.557 does that does Azure networking." Right? 00:32:12.640 --> 00:32:15.637 These are the folks managing the 500 00:32:15.720 --> 00:32:19.317 plus fiber operators managing the van, 00:32:19.400 --> 00:32:23.477 right? All over. And fiber operations ultimately is a physical operation. 00:32:23.560 --> 00:32:27.357 Things get cut. Things get you know, have to be repaired. You know, we have 00:32:27.440 --> 00:32:29.437 fancy words called devops and so on. 00:32:29.520 --> 00:32:33.397 Basically, emails are coming in and you got to go respond to them, take care of 00:32:33.480 --> 00:32:35.557 it. So they built this agentic system. 00:32:35.640 --> 00:32:39.037 They even have a character for it. It's called Miles and it sort of does all 00:32:39.120 --> 00:32:42.397 this stuff. Right? They started sort of 00:32:42.480 --> 00:32:46.237 screaming for more tokens and so on. And so they were saying, "Look, we don't 00:32:46.320 --> 00:32:47.917 need head count. We need tokens in order 00:32:48.000 --> 00:32:50.837 to be able to manage our operation." 00:32:50.920 --> 00:32:53.557 That reconceptualization 00:32:53.640 --> 00:32:55.997 of what their work is, right? They They 00:32:56.080 --> 00:33:00.757 basically took their work and made it meta. Right? That meta work is now their 00:33:00.840 --> 00:33:03.957 new work. >> Mhm. >> Right? In the '80s, if somebody had come 00:33:04.040 --> 00:33:07.677 to us and said, "4 billion people are going to get up in the morning and start 00:33:07.760 --> 00:33:12.877 typing." My model would have been, "We need 4 billion typists." But we're not 00:33:12.960 --> 00:33:16.757 doing typing. We're doing knowledge work. So that to me, I think, is it, 00:33:16.840 --> 00:33:20.637 right? Which is whether it's Microsoft or whether it's any organization is to 00:33:20.720 --> 00:33:23.557 give ourselves permission to do new 00:33:23.640 --> 00:33:27.637 types of meta cognition, meta work using 00:33:27.720 --> 00:33:31.757 these new tools to change the outputs that matter, 00:33:31.840 --> 00:33:35.637 uh and then really make the impossible possible. So completing that dot or that 00:33:35.720 --> 00:33:40.117 that connective tissue across those, I think, is where a lot of the enterprise 00:33:40.200 --> 00:33:43.637 value will get created. >> Can we talk about the data centers? 00:33:43.720 --> 00:33:48.677 >> Yeah, please ask. >> Oh, okay. Well, we This leads nicely into the data 00:33:48.760 --> 00:33:52.157 center buildout. I always think just I'm just impressed with the sheer scale of 00:33:52.240 --> 00:33:56.237 the buildout from Microsoft but also everyone else that this is redefining 00:33:56.320 --> 00:33:58.637 what it means to be a hyperscaler. And I 00:33:58.720 --> 00:34:03.597 just feel like that that that is unprecedented scale on finances, 00:34:03.680 --> 00:34:06.957 on the way you run the company, but also the communities that are that are 00:34:07.040 --> 00:34:09.637 impacted. Um then just talk a bit more about what 00:34:09.720 --> 00:34:11.317 you're seeing on the ground. Like when you visit your 00:34:11.400 --> 00:34:15.957 >> Yeah, I think there are two aspects of it. Obviously, the the buildout is 00:34:16.040 --> 00:34:20.797 extraordinary. You know, nothing like this has happened 00:34:20.880 --> 00:34:24.717 and it's great to be a one of the participants in it. Uh but you brought 00:34:24.800 --> 00:34:29.797 up the other part, right? I think at this point it's clear that unless we as 00:34:29.880 --> 00:34:32.197 an industry uh 00:34:32.280 --> 00:34:36.437 are very principled about ensuring that 00:34:36.520 --> 00:34:38.357 the benefits of all the stuff we're 00:34:38.440 --> 00:34:41.997 talking about are felt in real ways 00:34:42.080 --> 00:34:43.797 at the community level, right? Because 00:34:43.880 --> 00:34:47.477 this is not just a a campaign. 00:34:47.560 --> 00:34:52.157 Right, it has to be real where people are saying, "Look, this is not changing 00:34:52.240 --> 00:34:56.917 the prices on energy for me." In fact, if anything, it's bringing down prices 00:34:57.000 --> 00:35:00.237 because long-term there's going to be a better grid. There's going to be more 00:35:00.320 --> 00:35:05.077 energy. Water consumption is in fact not sort of in fact water is being 00:35:05.160 --> 00:35:06.997 replenished, right? You got to really, 00:35:07.080 --> 00:35:09.797 you know, educate folks on truly what's 00:35:09.880 --> 00:35:14.397 happening the the close-loop systems we're building. We have to invest in the 00:35:14.480 --> 00:35:18.277 training, the jobs, the tax base. In fact, the least talked about stuff is 00:35:18.360 --> 00:35:20.757 the amount of jobs that get created 00:35:20.840 --> 00:35:23.677 during construction, after construction, 00:35:23.760 --> 00:35:27.837 what's the tax base that's there in the community. And and all this has to be 00:35:27.920 --> 00:35:31.277 real. And and if that is the case, then we 00:35:31.360 --> 00:35:34.877 will have permission. If it is not, we won't have permission. It's as simple as 00:35:34.960 --> 00:35:38.117 that, right? Which is we we I think we have to take it as an 00:35:38.200 --> 00:35:41.997 industry pretty seriously. I think it's good for communities to be 00:35:42.080 --> 00:35:44.877 skeptical, ask the hard questions for us 00:35:44.960 --> 00:35:47.837 to do the hard work, earn that. 00:35:47.920 --> 00:35:50.797 But at the end of the day, if this if we 00:35:50.880 --> 00:35:55.077 can really be the I've always felt like in human history, 00:35:55.160 --> 00:35:57.517 if you use a lot of energy, but also 00:35:57.600 --> 00:36:00.357 create a lot of value for society, 00:36:00.440 --> 00:36:04.757 the story has been fantastic. If you don't do that, it's not been that great. 00:36:04.840 --> 00:36:09.237 And this time around I'm a firm believer that ultimately, if you do have a token 00:36:09.320 --> 00:36:14.077 economy that drives productivity, that drives economic growth, that drives 00:36:14.160 --> 00:36:17.637 broad-spread, you know, participation, 00:36:17.720 --> 00:36:22.117 better health outcomes, then I think we will be in a great 00:36:22.200 --> 00:36:25.957 place. And that's at least what we all have to be focused on. 00:36:26.040 --> 00:36:29.437 >> Yeah, it makes me think actually that with all these initiatives that you're 00:36:29.520 --> 00:36:33.997 doing, might be easier to see ROI in the communities first before in an enterprise. 00:36:34.080 --> 00:36:36.157 >> Yeah. >> [laughter] >> I I mean, I think both sides. In fact, 00:36:36.240 --> 00:36:40.477 it comes back together. It has to be the people in the communities are going to 00:36:40.560 --> 00:36:45.557 be employed, are going to be participants uh in the real economy, right? That's I 00:36:45.640 --> 00:36:47.357 think the question is like if we if the 00:36:47.440 --> 00:36:50.637 broad economy is doing well and the 00:36:50.720 --> 00:36:53.197 communities are doing well, the dots get 00:36:53.280 --> 00:36:58.077 connected. It's sort of the market forces are such that we will connect the 00:36:58.160 --> 00:37:00.757 dots. And that I think is it. Like you ought to be able to see the evidence. 00:37:00.840 --> 00:37:05.917 You can't be about any one company. Uh but it has to be broad economic growth 00:37:06.000 --> 00:37:08.597 and broad you know, community permission. 00:37:08.680 --> 00:37:12.357 >> Yeah. What's the biggest mistake you've made about currently 00:37:12.440 --> 00:37:17.477 or whatever you most updated your personal models on regarding societal impact of AI? 00:37:17.560 --> 00:37:21.357 >> So, you're saying what's the the >> What have you updated most on in terms 00:37:21.440 --> 00:37:24.877 of societal impact of AI? >> I think the 00:37:24.960 --> 00:37:28.877 um the the most uh 00:37:28.960 --> 00:37:33.197 critical thing is the first question we even started with, which is we need to 00:37:33.280 --> 00:37:35.037 tell 00:37:35.120 --> 00:37:38.117 the story and make it real 00:37:38.200 --> 00:37:42.757 that everybody has a real shot to 00:37:42.840 --> 00:37:46.117 participate as a first-class participant 00:37:46.200 --> 00:37:48.917 in this new economy. Right? That's kind 00:37:49.000 --> 00:37:52.437 of I think in the next 12 months, 18 00:37:52.520 --> 00:37:55.197 months, we need a way for people to say, 00:37:55.280 --> 00:37:57.397 "Oh, wow, I get it." Right? There's 00:37:57.480 --> 00:38:00.197 going to be tremendous capability, 00:38:00.280 --> 00:38:05.397 tremendous amount of infrastructure, but I can see what is going to happen 00:38:05.480 --> 00:38:06.957 whether it's the benefits like health 00:38:07.040 --> 00:38:09.717 outcomes or my ability to create a 00:38:09.800 --> 00:38:13.677 startup or my ability to run my local 00:38:13.760 --> 00:38:16.957 sort of uh store more efficiently, it's 00:38:17.040 --> 00:38:19.877 just happening and I see that uh benefit 00:38:19.960 --> 00:38:22.597 myself. Right? That to me 00:38:22.680 --> 00:38:26.917 you know, earning that permission in a path-dependent way, we can't wait. See, 00:38:27.000 --> 00:38:28.877 the one thing you learn I've now learned 00:38:28.960 --> 00:38:31.477 is I think the world is going to be way 00:38:31.560 --> 00:38:35.157 skeptical of tech and tech companies 00:38:35.240 --> 00:38:39.557 that say, "Trust us. We've got it. The future is going to be 00:38:39.640 --> 00:38:40.797 glorious." 00:38:40.880 --> 00:38:45.437 Uh you kind of have to deliver tangible 00:38:45.520 --> 00:38:50.517 benefits. Um And but frankly, politicians winning 00:38:50.600 --> 00:38:54.837 elections uh because they have advocated for that. 00:38:54.920 --> 00:38:59.917 That will be at least my adjustment because without it um 00:39:00.000 --> 00:39:02.997 thinking that somehow because it's too important this time around. It's too 00:39:03.080 --> 00:39:05.957 much of the economy for it not to be the case. 00:39:06.040 --> 00:39:08.597 >> So, one very simple framework I have 00:39:08.680 --> 00:39:11.117 for, you know, what are what is going to 00:39:11.200 --> 00:39:14.397 be the broad benefit of AI um 00:39:14.480 --> 00:39:18.717 beyond the communities just working in technology uh our sort of wealth 00:39:18.800 --> 00:39:22.117 creation is going to happen in a ton of different companies, startups and large 00:39:22.200 --> 00:39:25.797 companies. Then you have healthcare. Uh you you had amazing demos today. There 00:39:25.880 --> 00:39:30.317 are companies like Open Evidence. I think that is happening. Um education 00:39:30.400 --> 00:39:34.237 seems like another one that's an obvious good where we haven't seen as much 00:39:34.320 --> 00:39:37.757 impact as I'd expect. Do you have a hypothesis on why that might be or if 00:39:37.840 --> 00:39:39.277 it'll come? >> Yeah, I mean I think this is where 00:39:39.360 --> 00:39:42.197 again, how we think about education, how 00:39:42.280 --> 00:39:46.837 you know, recently I met with uh the founders of Alpha School and learned a 00:39:46.920 --> 00:39:51.117 lot about what they were going and going about it. It is fascinating to listen 00:39:51.200 --> 00:39:55.717 uh to how to even rethink uh what is education really look like because I 00:39:55.800 --> 00:39:58.917 think it's actually very important. Uh and I'm not saying anything 00:39:59.000 --> 00:40:01.877 traditionally being done is less important, right? I was even looking at 00:40:01.960 --> 00:40:04.557 the It's fascinating to see I forget to 00:40:04.640 --> 00:40:07.437 which Stanford class it was uh the the 00:40:07.520 --> 00:40:10.637 Asian guidelines for CS something. Uh 00:40:10.720 --> 00:40:13.397 because you still need people to learn. 00:40:13.480 --> 00:40:16.517 Uh like it is an interesting AI class that they were making sure people were 00:40:16.600 --> 00:40:20.317 learning how to apply softmax appropriately versus saying, "Hey, fix 00:40:20.400 --> 00:40:24.357 my training run." Uh so, I think learning concepts is 00:40:24.440 --> 00:40:26.677 important. It's going to be a critical. 00:40:26.760 --> 00:40:29.837 But, the way we create the incentives, 00:40:29.920 --> 00:40:34.077 what are the credentials, how we value those credentials, what is the 00:40:34.160 --> 00:40:35.837 employment opportunity for those 00:40:35.920 --> 00:40:37.717 credentials? So, I think that there is a 00:40:37.800 --> 00:40:41.197 complete change that has to happen uh 00:40:41.280 --> 00:40:44.237 given the way to get to information, way 00:40:44.320 --> 00:40:48.397 to educate yourself, way to continuously keep yourself updated has changed so 00:40:48.480 --> 00:40:51.237 much. So, I think interestingly enough 00:40:51.320 --> 00:40:54.917 maybe the next big startup and success 00:40:55.000 --> 00:40:59.677 story could be someone who builds a new university 00:40:59.760 --> 00:41:01.637 um or a new um 00:41:01.720 --> 00:41:04.677 pedagogy even of how to get someone to 00:41:04.760 --> 00:41:09.357 go through a curriculum and find economic opportunity uh that's highly 00:41:09.440 --> 00:41:12.077 valuable. >> Well, that has felt uh perhaps 00:41:12.160 --> 00:41:14.957 impossible for a long time, but it's a great note to end on and something that 00:41:15.040 --> 00:41:17.037 might be possible. Yeah. Thank you, Sayan. 00:41:17.120 --> 00:41:20.427 >> Thank you so much. Thank you. I appreciate it. Thank you all. 00:41:20.510 --> 00:41:22.530 >> [music] 00:41:25.750 --> 00:41:27.770 [music]