WEBVTT - Dell CTO on Artificial Intelligence Infrastructure 0:00:02.520 --> 0:00:06.880 Bloomberg Audio Studios, podcasts, radio news. 0:00:07.200 --> 0:00:11.080 This is Bloomberg Business Week with Carol Messer and Tim 0:00:11.119 --> 0:00:13.399 Stenebek on Bloomberg Radio. 0:00:14.400 --> 0:00:19.480 These are not upie. We've got to have some music. 0:00:20.040 --> 0:00:23.960 Front, all right. We are talking new frontiers. We're talking 0:00:24.000 --> 0:00:25.759 actually we talk about it. I feel like every day 0:00:26.079 --> 0:00:30.200 anything and everything to do with artificial intelligence, specifically jen AI, 0:00:31.280 --> 0:00:34.760 LM's large language models, the components that play into it, 0:00:34.800 --> 0:00:37.280 the power that's going to be needed to power the 0:00:37.360 --> 0:00:40.919 data centers, and who is doing all of the build out. 0:00:41.440 --> 0:00:44.800 Having said that, we've been watching Dell Technology stock rallied 0:00:44.840 --> 0:00:46.519 in today's session, up about eight percent. 0:00:47.200 --> 0:00:47.760 Some news. 0:00:47.840 --> 0:00:50.240 Loop Capital, which has a buy on the stock, raised 0:00:50.240 --> 0:00:52.560 the price target from one twenty five to one five, 0:00:53.000 --> 0:00:56.280 boosted some investor confidence. Keep in mind, Dell does report 0:00:56.320 --> 0:01:00.880 earnings on May thirtieth, so that's tomorrow after the And 0:01:00.920 --> 0:01:04.080 then yesterday some news of Dell expanding its AI factory 0:01:04.080 --> 0:01:07.480 with Nvidia to include new server, edge workstation solutions and 0:01:07.800 --> 0:01:10.920 services advancements that speed AI adoption and innovation. There's a 0:01:10.959 --> 0:01:13.560 lot going on. Michael Dell just about a week and 0:01:13.600 --> 0:01:16.720 a half ago talking about aipcs being pretty standard in 0:01:16.800 --> 0:01:19.880 twenty twenty five. So we have a great guest to 0:01:19.920 --> 0:01:21.959 talk about a lot of what is going on here. 0:01:22.400 --> 0:01:25.280 He participated in a panel with me put on by 0:01:25.319 --> 0:01:29.080 our Bloomberg Intelligence team looking at generative AI, specifically on 0:01:29.120 --> 0:01:33.520 the potential build out by companies of AI data capabilities 0:01:33.520 --> 0:01:36.160 on site or on premise with us as John Rose. 0:01:36.200 --> 0:01:39.280 He's global Chief Technology Officer at Dell Technology. He's joining 0:01:39.319 --> 0:01:41.520 us from New Hampshire. John, So, nice to have you here. 0:01:41.560 --> 0:01:42.039 How are you. 0:01:43.120 --> 0:01:45.000 I'm doing great? How are you all doing well? 0:01:45.040 --> 0:01:48.760 Doing well? There is a lot coming at us right now. 0:01:49.240 --> 0:01:51.720 Talk to us about what you are seeing. We talked 0:01:51.720 --> 0:01:56.200 about at that Bloomberg Intelligence event about what companies were doing, 0:01:56.240 --> 0:01:58.520 what their demands would be, you know, to bring things 0:01:58.640 --> 0:02:02.280 on site on PREMI what's been some of the conversations 0:02:02.280 --> 0:02:04.480 that you are having with clients around this as of late. 0:02:05.720 --> 0:02:08.840 Yeah, I mean, let me rewind a little bit. Last year, 0:02:08.880 --> 0:02:11.960 the first year of the AHI era. You know, I 0:02:11.960 --> 0:02:14.720 think most enterprises were trying to figure out what should 0:02:14.760 --> 0:02:18.040 they do, and largely most of the large enterprises did 0:02:18.040 --> 0:02:19.040 a lot of experimentation. 0:02:19.120 --> 0:02:22.160 This year is different. This year, many of those. 0:02:21.919 --> 0:02:25.080 Initial experiments which we're really trying to figure out where 0:02:25.080 --> 0:02:27.960 to apply this technology for the best return. You know, 0:02:28.120 --> 0:02:30.840 if you're an enterprise, you could apply aid anything, but 0:02:30.919 --> 0:02:32.920 if you apply it to your supply chain or your 0:02:32.919 --> 0:02:35.240 product development cycle or something that really moves the needle 0:02:35.320 --> 0:02:38.000 from an economic perspective, that'll have a bigger impact. So 0:02:38.040 --> 0:02:41.440 I think today most enterprises are triangulating on where to 0:02:41.480 --> 0:02:43.920 apply it, which then gets them to the discussion of 0:02:43.960 --> 0:02:46.960 how to apply it. And that's where you know, the 0:02:47.000 --> 0:02:49.040 panel we add when we were chatting kind of went, 0:02:49.200 --> 0:02:53.000 which is this is not a workload that's you know, 0:02:53.080 --> 0:02:55.120 sitting on the side used by three people. This is 0:02:55.160 --> 0:02:57.280 the center of your enterprise. It's going to run non 0:02:57.320 --> 0:02:59.240 stop seven by twenty four. It's going to power your 0:02:59.280 --> 0:03:03.640 sales processing tompower your customer satisfaction. And so choosing the 0:03:03.720 --> 0:03:06.400 right place to run it, whether that you know that 0:03:06.520 --> 0:03:09.000 gives you the best economic outcome, the best control, doesn't 0:03:09.040 --> 0:03:11.800 get you into compliance and regulatory challenges, is really the 0:03:11.880 --> 0:03:14.120 dialogue that's happening now. And so a lot of the 0:03:14.120 --> 0:03:16.240 work we're doing that you mentioned on the intro around 0:03:16.280 --> 0:03:19.360 with our ecosyst around Nvidia and Meta and everybody else, 0:03:19.480 --> 0:03:22.920 is how do we reduce the complexity to make that decision. 0:03:22.960 --> 0:03:24.680 How do we make it easier for people to. 0:03:24.720 --> 0:03:27.519 Get started, to not have to do everything, and to 0:03:27.600 --> 0:03:29.840 really get the platforms in place where they need them. 0:03:29.880 --> 0:03:32.000 And our opinion, one of the best places to do 0:03:32.080 --> 0:03:35.200 some of this stuff is clearly in their own owned infrastructure. 0:03:35.240 --> 0:03:37.760 One it tends to be a Capex model, and two 0:03:37.960 --> 0:03:40.320 it's in your control and it's much more predictable. 0:03:40.400 --> 0:03:40.640 John. 0:03:40.680 --> 0:03:43.240 For years we've heard about the cloud being what's nimble 0:03:43.400 --> 0:03:46.040 and the cloud being the place that we can do 0:03:46.120 --> 0:03:50.800 this stuff and secure, secure, do this stuff for less expensive. 0:03:52.240 --> 0:03:56.280 I'm wondering the shift on premise is not a shift. 0:03:56.280 --> 0:03:58.760 It's been around for but since before the cloud. But 0:03:58.760 --> 0:04:00.800 why are we seeing and top thing so much about 0:04:00.800 --> 0:04:01.640 the shift right now? 0:04:02.440 --> 0:04:03.080 Yeah, this is. 0:04:03.000 --> 0:04:05.960 Remember the cloud here was all about taking your existing workloads, 0:04:05.960 --> 0:04:09.440 your web servers, your email, your office productivity and maybe 0:04:09.440 --> 0:04:10.920 trying to figure out a different way to operate it. 0:04:10.920 --> 0:04:13.040 If by the cloud wasn't just public cloud, the cloud 0:04:13.080 --> 0:04:16.680 model pervade everything and so you know, the shift autonomous, 0:04:16.800 --> 0:04:20.479 automated elastic infrastructure happened with a set of applications that 0:04:20.480 --> 0:04:23.200 we understood well. The things we're building now look nothing 0:04:23.320 --> 0:04:26.080 like that. A large scale generative AI system for an 0:04:26.160 --> 0:04:30.160 enterprise is arguably the most demanding and complex workload you 0:04:30.200 --> 0:04:33.440 will create in your lifetime, and so you know, we're 0:04:34.000 --> 0:04:36.920 you could argue that, you know, an OPX driven as 0:04:36.960 --> 0:04:39.280 a service model that gives you lots of agility but 0:04:39.400 --> 0:04:42.800 you pay by the drip is not a very good outcome. 0:04:43.000 --> 0:04:45.640 If the thing you're running runs seven x twenty four 0:04:45.680 --> 0:04:48.919 at enormous performance levels, you know you don't want that 0:04:49.000 --> 0:04:51.839 meter running. You want that meter to be predictable, and 0:04:51.920 --> 0:04:54.480 so it's just a different class of workload. By the way, 0:04:54.520 --> 0:04:56.400 we're big proponents in multi cloud. A lot of the time, 0:04:56.440 --> 0:04:59.440 the best place to develop your AI system is in 0:04:59.440 --> 0:05:01.039 one of the cloud providers because they have a great 0:05:01.080 --> 0:05:04.240 tool chain. The best place to test it might be there, 0:05:04.240 --> 0:05:06.200 Maybe the best place to train your models might be 0:05:06.200 --> 0:05:08.000 there because you only need the infrastructure for a short 0:05:08.040 --> 0:05:10.440 period of time. But the minute it becomes inference that 0:05:10.440 --> 0:05:12.960 you're putting it into production and it's using your data, 0:05:13.000 --> 0:05:14.880 which by the way, most of that is on prem 0:05:14.920 --> 0:05:18.320 even today, then it starts to become a very different discussion, 0:05:18.320 --> 0:05:21.719 which brings kind of this modern on prem architecture that 0:05:21.760 --> 0:05:24.520 we talk about with the AI factory into play as 0:05:24.600 --> 0:05:26.720 probably one of the more logical places to start. 0:05:26.800 --> 0:05:29.280 Well, what is exactly the concept of an AI factory. 0:05:30.279 --> 0:05:34.200 Yeah, here's the important thing for people to realize. AI 0:05:34.320 --> 0:05:36.520 is a new workload and by the way, it actually 0:05:36.560 --> 0:05:39.640 needs a new class of infrastructure. The type of compute 0:05:39.720 --> 0:05:43.400 is not CPUs, it's GPUs. The type of data is 0:05:43.480 --> 0:05:46.920 not traditional databases, it's vectorized data that lives in large 0:05:47.000 --> 0:05:49.920 language models. The kind of tools you use are different, 0:05:49.920 --> 0:05:52.599 and so what we talked about at Dell Technologies World 0:05:52.680 --> 0:05:55.160 last year last week was was not a new set 0:05:55.200 --> 0:05:57.960 of products exclusively. It was you really probably need to 0:05:58.000 --> 0:06:01.320 have a separate type of infrastructure where your AI than 0:06:01.360 --> 0:06:03.560 you do for the traditional things that you do, those 0:06:03.640 --> 0:06:06.000 workloads that went through the kind of cloud migration we 0:06:06.000 --> 0:06:08.400 talked about, and what the AI factory is is it's 0:06:08.400 --> 0:06:11.160 an articulation of what that infrastructure looks like. That it 0:06:11.200 --> 0:06:14.359 is accelerated compute, that it's a different kind of data architecture, 0:06:14.360 --> 0:06:17.640 it's a better and different type of networking architecture, and 0:06:17.680 --> 0:06:21.120 that it probably lives in a different footprint because it 0:06:21.120 --> 0:06:24.640 itself has different requirements. Than your legacy applications and the 0:06:24.680 --> 0:06:27.000 other applications you run in either a public cloud or 0:06:27.000 --> 0:06:27.760 a private environment. 0:06:27.839 --> 0:06:29.000 And so the II. 0:06:28.800 --> 0:06:32.600 Factory is how do you create a methodology and organize 0:06:32.640 --> 0:06:35.600 all the technology to put that in play, whether it's 0:06:35.640 --> 0:06:38.040 at a rack level or even an entire data center, 0:06:38.320 --> 0:06:41.040 that builds you the optimal infrastructure to run these new 0:06:41.040 --> 0:06:43.120 workloads that you're going to need. So think of it 0:06:43.160 --> 0:06:44.840 as just a paradigm shift. We're going to have to 0:06:44.880 --> 0:06:47.800 build new infrastructure for this new workload. It might be 0:06:47.839 --> 0:06:50.320 redesigning or optimizing what we have, and it might be 0:06:50.360 --> 0:06:51.840 in fact being your data center build out. 0:06:52.080 --> 0:06:54.040 John. It's really interesting and I think we might have 0:06:54.080 --> 0:06:56.640 talked about this at the Bloomberg Intelligence event, but about 0:06:56.640 --> 0:06:59.400 you know, technology companies, they compete, they work with each other, 0:06:59.480 --> 0:07:02.880 and it was an interesting at Dell Tech World the 0:07:02.960 --> 0:07:07.480 keynote stage to see Jensen Wog and videos CEO up there, 0:07:08.400 --> 0:07:11.280 and so there's really you know, you can see the partnership. 0:07:11.360 --> 0:07:13.920 You can see there's clearly a show support from in 0:07:14.040 --> 0:07:16.800 video to you guys at Dell. What is the nature 0:07:16.840 --> 0:07:19.560 of this partnership, especially when they make their own AI 0:07:19.840 --> 0:07:23.840 server racks, which makes you competitors to some degree. 0:07:24.920 --> 0:07:28.400 Yeah, Remember, Dell is a unique company. We are obviously 0:07:28.560 --> 0:07:32.120 very large, and you could argue we're the largest technology 0:07:32.120 --> 0:07:34.640 integrator in the world. Now it means something to different people, 0:07:34.640 --> 0:07:37.440 but basically, like I don't build my own CPUs or GPUs, 0:07:37.480 --> 0:07:40.280 but what I do is I organize that technology in 0:07:40.320 --> 0:07:44.520 the consumable units of it that my customers across the 0:07:44.560 --> 0:07:47.320 world can consume. Now, when you look at what in 0:07:47.400 --> 0:07:50.120 Video is doing, they clearly are the provider of some 0:07:50.160 --> 0:07:51.880 of the better GPUs in the world. There are other 0:07:51.960 --> 0:07:54.920 choices that we also work with. They also have organized 0:07:54.960 --> 0:07:56.880 their stack in a way that makes it very easy 0:07:56.880 --> 0:07:58.640 to consume, and I think that did a great job there. 0:07:58.680 --> 0:08:00.800 And so they have an early lead and it's importantly 0:08:00.960 --> 0:08:04.120 that they're they're they're making it much more consumable and 0:08:04.160 --> 0:08:08.160 they're keeping the innovation cycle up. However, their ability to 0:08:08.200 --> 0:08:10.640 engage with the large enterprise across the world, they don't 0:08:10.640 --> 0:08:12.880 have the We own much larger salesforce, We have a 0:08:12.960 --> 0:08:16.040 global services capability, we have the largest supply chain that's 0:08:16.040 --> 0:08:19.240 secure in the world and technology and so and by 0:08:19.280 --> 0:08:22.040 the way, you also don't just need the GPUs, you 0:08:22.080 --> 0:08:24.200 need the advanced storage services. You need to integrate it 0:08:24.200 --> 0:08:26.000 with your existing infrastructure. You need to talk to your 0:08:26.040 --> 0:08:28.000 existing data, which, by the way, rides pretty much on 0:08:28.080 --> 0:08:31.840 Dell technology storage systems. And so the nature of the 0:08:31.840 --> 0:08:34.640 relationship is, look, hey, you're trying to build an aifactory. 0:08:34.640 --> 0:08:37.120 There are some leading edge parts that absolutely have to 0:08:37.160 --> 0:08:39.320 be produced and on way to. One way to actually 0:08:39.320 --> 0:08:41.600 make sure that they happen correctly is to integrate them 0:08:41.640 --> 0:08:45.000 into a system, an early system like what Nvidia does. However, 0:08:45.080 --> 0:08:48.000 those parts are decomposable and then they can reassemble into 0:08:48.120 --> 0:08:51.040 other form factors that companies like Dell can take to 0:08:51.160 --> 0:08:54.240 a much more scalable market. The announcements around edge and 0:08:54.280 --> 0:08:57.280 other areas. We have the ability to reach more customers 0:08:57.320 --> 0:08:59.480 than anybody in the world. We need partners to help 0:08:59.559 --> 0:09:01.760 us build technology to bring to them. 0:09:01.880 --> 0:09:04.199 Is it a deeper relationship? Just got about forty seconds 0:09:04.240 --> 0:09:07.160 and we'll come back and continue. But is your partnership 0:09:07.200 --> 0:09:09.000 with Nvidia a little bit different? Is it a deeper 0:09:09.040 --> 0:09:12.239 one because you do have partnerships with other server manufacturers 0:09:12.240 --> 0:09:14.640 as well in others? Is it a deeper relationship or 0:09:14.640 --> 0:09:15.400 how would you call it? 0:09:15.480 --> 0:09:17.599 Yeah, we worked with you know We work with a 0:09:17.679 --> 0:09:19.240 lot of companies, and what I will tell you is 0:09:19.520 --> 0:09:20.560 the first one. 0:09:20.600 --> 0:09:21.520 It was the deepest. 0:09:21.679 --> 0:09:24.280 Last year we announced Project Helix, which was the first 0:09:24.280 --> 0:09:27.280 time anybody articulated putting all the parts together into something 0:09:27.280 --> 0:09:31.400 people could consume. So it has a significant first mover advantage. However, 0:09:31.559 --> 0:09:32.880 you know, like you said, we work with a lot 0:09:32.960 --> 0:09:34.440 of companies and we have a lot of partners. 0:09:34.559 --> 0:09:36.839 You want to continue with our guest, John Roses with US. 0:09:37.080 --> 0:09:39.559 Roses with us. He's a global chief technology officer at 0:09:39.559 --> 0:09:42.920 Dell Technology, still with us from New Hampshire. Hey, John, 0:09:42.920 --> 0:09:44.960 you know we were talking Tim and I in the 0:09:44.960 --> 0:09:48.400 break and just you know, curious about as people go 0:09:48.559 --> 0:09:50.880 right t him to build out their data centers, you 0:09:50.920 --> 0:09:53.320 do wonder how busy it kind of gets for Dell. 0:09:53.640 --> 0:09:57.280 Yeah, I'm wondering how big John, the AI server opportunity 0:09:57.640 --> 0:09:59.880 is for Dell in twenty twenty five, and then how 0:10:00.040 --> 0:10:02.160 big it could be over the next few years. What 0:10:02.200 --> 0:10:04.599 are you folks talking about internally and externally? 0:10:05.480 --> 0:10:07.760 Well, so I'm a CTO, so I'm not going to 0:10:07.840 --> 0:10:08.480 talk about. 0:10:08.480 --> 0:10:10.400 Particular year, But are you really busy? 0:10:10.520 --> 0:10:10.800 John? 0:10:11.840 --> 0:10:14.800 I am extraordinarily busy, But let me paint a picture 0:10:14.840 --> 0:10:17.320 that's a little longer term, you know, and that is, look, 0:10:17.320 --> 0:10:20.040 we are, as I mentioned, we're in now year two 0:10:20.120 --> 0:10:22.840 of the AI cycle, the modern AI cycle, and year 0:10:22.880 --> 0:10:26.120 one was all about surprise, get organized. Year two is 0:10:26.160 --> 0:10:29.920 about the first kind of enterprise deployments and kind of 0:10:29.960 --> 0:10:33.400 the prototypes. What that means is that the enterprise buildout 0:10:33.440 --> 0:10:36.640 hasn't actually begun. And if we look at the size 0:10:36.640 --> 0:10:38.679 of the A market today and what's going on, it's 0:10:38.679 --> 0:10:41.200 a pretty interesting market. What's happening is we're building the 0:10:41.200 --> 0:10:44.480 foundational technologies, we're training the large language models that this 0:10:44.520 --> 0:10:47.800 is a very robust ecosystem right now, we are developing 0:10:47.800 --> 0:10:50.400 the tool chains and as you can see from you know, 0:10:50.440 --> 0:10:52.480 just the state of the industry, it's a pretty exciting 0:10:52.520 --> 0:10:56.080 and a pretty significant shift that is in front of 0:10:56.480 --> 0:10:59.840 the enterprise cycle. We're conservatively, you know, most most people 0:11:00.240 --> 0:11:03.839 over the long term, the AI cycle is about rebalancing 0:11:04.080 --> 0:11:06.520 a sizable portion of the work in the world into 0:11:06.520 --> 0:11:09.400 the machine layer, and so as that occurs, you know, 0:11:09.720 --> 0:11:11.800 it represents you know, sometimes we use the phrase we're 0:11:11.800 --> 0:11:13.680 in the training era. Now we're about to enter the 0:11:13.720 --> 0:11:16.640 infront zer for enterprise, which is AI gets put into production. 0:11:16.760 --> 0:11:18.840 When it does, you know, you can. 0:11:18.720 --> 0:11:21.480 Calculate imagine a world where you know a third of 0:11:21.480 --> 0:11:24.480 the work is now happening in a machine layer. It's 0:11:24.480 --> 0:11:26.920 being done by a machine an AI system. What does 0:11:26.960 --> 0:11:29.720 that look like? How big is that? Well, you can't 0:11:29.720 --> 0:11:32.319 calculate it accurately, but you know it's a gigantic number. 0:11:32.360 --> 0:11:34.680 And it's as big as the Internet build out, it's 0:11:34.720 --> 0:11:37.640 as big as the Industrial revolution. In many of the 0:11:37.679 --> 0:11:40.360 discussions that we have, the timing on it, it's an 0:11:40.360 --> 0:11:43.160 extended cycle. This'll be a twenty year cycle, but we're 0:11:43.160 --> 0:11:45.960 about to enter that phase. And what it tells us 0:11:46.040 --> 0:11:48.839 is it's a significant amount of replumbing of enterprise. It's 0:11:48.840 --> 0:11:52.160 building out AI factories, it's rethinking your data strategy, it's 0:11:52.200 --> 0:11:55.080 rethinking your footprint in the multi cloud, and all of 0:11:55.080 --> 0:11:57.360 that tends to give it our breath and depth in 0:11:57.400 --> 0:12:00.600 our ecosystem, drag us into an awful lot of customer 0:12:00.800 --> 0:12:04.120 conversations in a pretty active world, even in advance of 0:12:04.160 --> 0:12:06.760 the significant buildouts occurring on the infront side. 0:12:06.920 --> 0:12:09.360 So okay, so if we're only in year two when 0:12:09.400 --> 0:12:14.880 it comes to AI, where are corporates in their AI journey? 0:12:14.880 --> 0:12:17.080 How do they adopt their technology. We know Dell won 0:12:17.160 --> 0:12:19.679 a meaningful chunk of Tesla's AI roll out, for example, 0:12:19.720 --> 0:12:21.960 But where are corporates in their AI journey? 0:12:22.960 --> 0:12:25.199 Yeah, the corporate As I mentioned before, last year, it 0:12:25.280 --> 0:12:27.280 was all about getting your feet on the ground, understanding 0:12:27.280 --> 0:12:29.600 the technology of earning what a large language model, earning 0:12:29.640 --> 0:12:32.720 what RAG was, and this year it's all about finding 0:12:32.760 --> 0:12:36.120 those first projects. The first projects are largely under development 0:12:36.160 --> 0:12:38.880 in most large enterprises, and some of them are emerging 0:12:38.920 --> 0:12:41.960 as chat thoughts and other services that we're starting to see. 0:12:42.559 --> 0:12:45.360 And there are definitely early examples of technology that have 0:12:45.400 --> 0:12:49.400 been deployed as new offerings. But the full pivot where 0:12:49.400 --> 0:12:52.640 a company now declares that I am in the center 0:12:52.679 --> 0:12:55.640 of my business, you know, building my product, selling my product, 0:12:55.800 --> 0:13:00.000 servicing my product, engaging with my customers based on primary 0:13:00.080 --> 0:13:03.040 early in AI architecture, that is still work to be done. 0:13:03.400 --> 0:13:04.800 And so so I think, you know, at. 0:13:04.720 --> 0:13:07.880 This point, we're still right now in most large enterprises 0:13:07.920 --> 0:13:11.480 in the first proof of concepts, the first prototypes. But 0:13:11.559 --> 0:13:13.240 one of the things that's different about AI is it 0:13:13.240 --> 0:13:15.000 does not take three years to build one of those. 0:13:15.360 --> 0:13:17.640 You can go from idea based on the tool chains 0:13:17.640 --> 0:13:20.960 available to having something in production that you can start 0:13:21.000 --> 0:13:22.760 to really do, and we're doing that inside of Dell 0:13:22.880 --> 0:13:25.480 make your developers more productive in a matter of months. 0:13:25.559 --> 0:13:27.880 And so the velocity of this cycle is something that 0:13:27.880 --> 0:13:31.520 we've never seen before, which means the gap between getting 0:13:31.520 --> 0:13:33.640 your feet on the ground and being in production and 0:13:33.640 --> 0:13:35.920 transforming your enterprise is not a ten year cycle. 0:13:36.000 --> 0:13:38.120 Hey, John, I keep hearing you know you've said it, 0:13:38.200 --> 0:13:40.280 and I've had other guests, but when they talk about 0:13:40.280 --> 0:13:43.480 AI in inference, right, am I saying it correctly? 0:13:43.920 --> 0:13:44.839 Yeah? Right? 0:13:45.400 --> 0:13:48.160 How is that different from AI training? And is that 0:13:48.320 --> 0:13:50.680 kind of a new concept or is that just something 0:13:50.720 --> 0:13:53.760 more complicated when it comes to generative AI. 0:13:54.160 --> 0:13:56.640 No, they're part of the same cycle that The idea 0:13:56.720 --> 0:13:59.680 behind AI is like you're trying to have a machine 0:14:00.480 --> 0:14:04.160 do some kind of cognitive work, answer a service, call, 0:14:04.320 --> 0:14:07.480 sell something, build a writ code. In order to do that, 0:14:07.559 --> 0:14:10.360 the first step is that you must have that machine 0:14:10.520 --> 0:14:13.960 have some access to the knowledge necessary to do that, 0:14:14.000 --> 0:14:16.680 which is what training is about. The difference in large 0:14:16.720 --> 0:14:19.920 language models is that we've developed techniques that allow us 0:14:19.960 --> 0:14:22.880 to instead of trying to as human beings decide how 0:14:22.880 --> 0:14:24.760 to code, we've learned that if. 0:14:24.640 --> 0:14:27.320 You just expose these new. 0:14:27.200 --> 0:14:31.000 Techniques, these new technologies, large language models, to a gigantic 0:14:31.120 --> 0:14:34.840 set of coding, just examples of coding, they will classify them, 0:14:34.920 --> 0:14:37.760 organize them, and create a neural network. And interestingly enough, 0:14:37.800 --> 0:14:41.320 they will then be able to replicate that intelligence, that behavior. 0:14:42.520 --> 0:14:45.320 And so the training phase is about taking gobs of data. 0:14:45.400 --> 0:14:49.600 In the current phase, it's the entire Internet and run 0:14:49.640 --> 0:14:53.360 it into these models that create systems that can understand 0:14:53.520 --> 0:14:56.560 or communicate human language, that can code. And all that 0:14:56.640 --> 0:15:00.520 is is them deriving from a gigantic data set the 0:15:00.600 --> 0:15:03.120 knowledge that's contained within it to create a set of skills. 0:15:03.280 --> 0:15:06.760 That's training. Inference is totally different. Inference is once you 0:15:06.840 --> 0:15:08.680 have that model, now you want to do. 0:15:08.680 --> 0:15:09.320 Something with it. 0:15:09.440 --> 0:15:12.400 You have a thing that can code great well inferences 0:15:12.400 --> 0:15:14.960 when you tell it to code a program to actually 0:15:15.000 --> 0:15:17.560 produce source code that does something. So they're just two 0:15:17.560 --> 0:15:19.680 habs at the same coin. One is the learning to 0:15:19.680 --> 0:15:21.720 create the capability. The other is the act of using 0:15:21.760 --> 0:15:22.920 the capability and production. 0:15:23.520 --> 0:15:25.160 Hey listen, something I got to ask you. It's a 0:15:25.200 --> 0:15:27.600 story that's on the Bloomberg and this has to do 0:15:27.640 --> 0:15:30.040 with Nobel Laureate he's also an economics professor. We're talking 0:15:30.040 --> 0:15:33.560 about Paul Romer, and he was talking with our team 0:15:33.640 --> 0:15:37.640 here and he said, runaway confidence and artificial intelligence risks 0:15:37.640 --> 0:15:39.800 repeating the mistakes of the crypto hype bubble of only 0:15:39.840 --> 0:15:42.440 two years ago. He says, right now, there's way too 0:15:42.520 --> 0:15:45.920 much confidence about the future trajectory of AI. When people 0:15:45.960 --> 0:15:47.920 project this phot I think they're at risk of making 0:15:47.960 --> 0:15:51.200 a very serious mistake. Do you think that there's too 0:15:51.320 --> 0:15:55.240 much confidence about AI? Do you think there's too much euphoria? 0:15:55.480 --> 0:15:58.240 Two answers to that in the general population. I think 0:15:58.280 --> 0:16:01.520 it's a very confusing space because we have basically the 0:16:01.600 --> 0:16:06.760 human race is maybe susceptible to science fiction, and we 0:16:06.840 --> 0:16:09.440 think that what we're producing is artificial general intelligence, and 0:16:09.480 --> 0:16:11.800 we're producing these these sentient beings. 0:16:11.840 --> 0:16:14.800 These are not sentient beings. These are technologies that yah. 0:16:15.720 --> 0:16:18.400 Yeah, maybe sometimes we all believe there's a path to AGI. 0:16:18.480 --> 0:16:21.720 It's just not in the near term. That whole dialogue, 0:16:21.720 --> 0:16:24.840 and let's call it the consumer general market, which is 0:16:24.880 --> 0:16:28.000 primarily where most of the big public AI players play, 0:16:28.360 --> 0:16:31.080 is a risk because people think these have personalities. They 0:16:31.080 --> 0:16:33.840 don't understand the technology. When you go to the enterprise world, 0:16:33.960 --> 0:16:37.080 it is far more conservative. There is no enterprise in 0:16:37.120 --> 0:16:39.360 the world that's trying to build the terminator or a 0:16:39.440 --> 0:16:41.760 sentient being. What we are doing is we're applying the 0:16:41.800 --> 0:16:44.920 techniques that we're pioneered in the public AI world in 0:16:45.000 --> 0:16:49.000 large language model space. Two very specific problems within a 0:16:49.040 --> 0:16:53.480 corporation that could benefit from it, writing code, finding your customers, 0:16:53.880 --> 0:16:55.640 engaging to solve problems. 0:16:55.960 --> 0:16:57.400 Those are the things that are mattering. 0:16:57.400 --> 0:17:00.080 So enterprise is kind of boring to be perfectly on 0:17:00.320 --> 0:17:02.400 versus what is in the part of the possible in 0:17:02.440 --> 0:17:05.960 the public world. However, as you know, our industrial complex 0:17:06.000 --> 0:17:08.040 is quite large and the impact is much bigger. 0:17:09.080 --> 0:17:11.840 Great stuff already looking forward the next time we get 0:17:11.840 --> 0:17:13.720 to catch up John, Thank you so much. B. While 0:17:13.760 --> 0:17:17.159 John Rose, he's global chief technology officer Dell Technology, is 0:17:17.240 --> 0:17:20.280 joining us from New Hampshire. You are listening and watching 0:17:20.280 --> 0:17:23.679 Bloomberg Business Week Carol Masser along with Tim Stanovic, and 0:17:23.720 --> 0:17:25.320 this is Bloomberg. 0:17:29.720 --> 0:17:33.280 Some disturbing information that we found our disturbing This next 0:17:33.280 --> 0:17:36.000 interview report by the global media company with a health 0:17:36.000 --> 0:17:39.439 and biotech focus. Marianne Liebert found that quote in nearly 0:17:39.600 --> 0:17:42.879 three quarters of the cases where a disease afflicts primarily 0:17:42.960 --> 0:17:47.320 one gender. The funding pattern favors males in that either 0:17:47.359 --> 0:17:50.000 the disease affects more women and is underfunded with respect 0:17:50.040 --> 0:17:53.880 to burden, or the disease affects more men and it's overfunded. 0:17:53.960 --> 0:17:56.800 So double whammy when it comes to the gender health gap. 0:17:57.000 --> 0:17:59.200 Yeah. Meantime, in a recent report for McKinsey, it found 0:17:59.200 --> 0:18:02.400 that reducing the time women spend in poor health by 0:18:02.440 --> 0:18:05.240 twenty five percent could be worth one trillion dollars, in 0:18:05.280 --> 0:18:08.920 large part because health disparities disproportionately hit women during their 0:18:08.960 --> 0:18:12.320 working years. So let's get to it. There are huge 0:18:12.359 --> 0:18:15.640 disparities between men and women's health that definitely tim need 0:18:15.680 --> 0:18:16.360 to be addressed. 0:18:16.520 --> 0:18:18.560 Back with us to talk about that gender health gap 0:18:18.640 --> 0:18:21.840 is Elizabeth Staddinger, managing board member at the sixty five 0:18:21.880 --> 0:18:25.320 billion dollar medical tech company Semen's Health and Ears, publicly 0:18:25.320 --> 0:18:27.960 traded one. I should note she joins us from Germany. Elizabeth, 0:18:27.960 --> 0:18:29.800 good to have you back with us. Thanks for staying 0:18:29.840 --> 0:18:32.440 up a little bit later once again to join us 0:18:32.560 --> 0:18:35.040 from Germany. Talk a little bit about what you're doing 0:18:35.040 --> 0:18:37.320 over at Semens Health and Ears to address this gender 0:18:37.359 --> 0:18:37.879 health gap. 0:18:40.320 --> 0:18:44.280 As you already mentioned, in the opening, there is significant disparities. 0:18:44.359 --> 0:18:47.119 And if you just to make this tangible, if you 0:18:47.280 --> 0:18:51.680 think of the typical Hollywood heart attack where somebody kind 0:18:51.720 --> 0:18:55.119 of reaches to his breast, has this burning sensation in 0:18:55.160 --> 0:18:57.720 the arm, and everybody will start rushing and saying, oh 0:18:57.720 --> 0:19:00.639 my god, something really serious is going on. Let's rush 0:19:00.720 --> 0:19:05.879 that patient to the ED and take care of him. Unfortunately, 0:19:06.080 --> 0:19:08.960 when you're a woman, most likely the symptoms you will 0:19:09.000 --> 0:19:13.440 be feeling are different. And that matters because the likelihood 0:19:13.480 --> 0:19:16.560