WEBVTT - CoreWeave’s $14 Billion Meta Deal, Spotify’s Ek to Leave CEO Role 0:00:02.560 --> 0:00:13.840 Bloomberg Audio Studios, podcasts, radio news. Bloomberg Tech is alive 0:00:13.880 --> 0:00:17.720 from Coast to Code with Caroline Hide in New York 0:00:18.000 --> 0:00:21.160 and vl Loow in San Francisco. 0:00:22.520 --> 0:00:24.239 This is Bloomberg Tech coming up. 0:00:24.280 --> 0:00:24.480 Core. 0:00:24.520 --> 0:00:27.320 We've signed a deal to supply Meta with as much 0:00:27.360 --> 0:00:31.200 as fourteen point two billion dollars worth of computing power. 0:00:31.000 --> 0:00:34.120 Plus a change of leadership at Spotify, a CEO Daniel 0:00:34.280 --> 0:00:37.120 Ck stepping aside after almost two decades. 0:00:36.760 --> 0:00:40.080 And Anthropic releases a new AI model designed to code 0:00:40.120 --> 0:00:43.640 longer and more effectively than prior versions. We speak with Aropis, 0:00:43.720 --> 0:00:45.440 chief product officer, but. 0:00:45.479 --> 0:00:48.159 First ed we are both here in New York this 0:00:48.280 --> 0:00:50.840 week and checking out the markets that are just taking 0:00:50.880 --> 0:00:54.200 a pause amidst some macro focus. Of course, TENSI shutdown 0:00:54.200 --> 0:00:56.800 of the US government that's on every investor's mind. What 0:00:56.840 --> 0:00:58.240 does that mean in terms of data to what does 0:00:58.240 --> 0:00:59.880 it mean in terms of the future of rape policy 0:01:00.080 --> 0:01:03.160 the FED. We're currently completely flat, just pinching into the 0:01:03.200 --> 0:01:04.959 green on the Nastet one hundred, but it's way more 0:01:04.959 --> 0:01:06.959 interesting underneath the hood and you're looking at that. 0:01:07.120 --> 0:01:09.400 Yeah, our top story is Core, Weave and Meta. It's 0:01:09.440 --> 0:01:13.000 the latest Core Weave capacity deal fourteen point two billion dollars. 0:01:13.240 --> 0:01:15.240 The stock has been on a tear since it's listing 0:01:15.280 --> 0:01:17.600 in March right now trading at its highest level in 0:01:17.720 --> 0:01:21.800 around six weeks. Clearly that big game following the Bloomberg report, 0:01:22.400 --> 0:01:24.280 The market's kind of focused on this idea, that is 0:01:24.640 --> 0:01:28.440 evidence core weaves move beyond one single customer, which is Microsoft. 0:01:28.680 --> 0:01:31.679 But actually, as Brodie up Forward often says, carry the 0:01:31.720 --> 0:01:33.280 devil's in the detail on this one. 0:01:33.560 --> 0:01:35.840 Let's get that detail, Dy forward with us. You help 0:01:35.880 --> 0:01:39.680 break the story. Core weave has tripled since it's IPO. 0:01:39.800 --> 0:01:42.720 We're up another fifteen percent. What does it mean to 0:01:42.760 --> 0:01:45.119 be adding Meta to the fold as well as open 0:01:45.120 --> 0:01:48.120 Ai more recently of course in Microsoft O. 0:01:48.080 --> 0:01:50.720 Well, it means that core Weave isn't just a colonial 0:01:50.800 --> 0:01:52.040 state of Microsoft. 0:01:52.120 --> 0:01:52.280 Right. 0:01:52.320 --> 0:01:54.680 That was the concern for so long with these neo 0:01:54.760 --> 0:01:58.680 clouds that gosh, if their biggest customers are companies that 0:01:58.760 --> 0:02:01.840 also effectively compete with them, that seems like a pretty 0:02:02.000 --> 0:02:06.200 indefensible business. But you start getting companies like Meta and 0:02:06.240 --> 0:02:09.280 open Ai who are likely to be longer term buyers 0:02:09.280 --> 0:02:11.840 of this technology, and so I think folks are getting 0:02:11.840 --> 0:02:16.480 more comfort that Corewave and companies like it may have 0:02:16.520 --> 0:02:19.600 a more sustained place in this AI infrastructure build out. 0:02:19.919 --> 0:02:22.880 Fourteen point two billion dollars is a very large number, Brodie, 0:02:22.880 --> 0:02:26.120 But the terms of the contract extent of twenty thirty two. 0:02:26.360 --> 0:02:28.000 You know what I'm going to say, because we discussed 0:02:28.000 --> 0:02:30.680 it this morning. There is skepticism here because there are 0:02:30.680 --> 0:02:34.080 many unknowns. For example, does call We've actually have a 0:02:34.160 --> 0:02:37.240 data center somewhere that's built and operating that they can 0:02:37.280 --> 0:02:39.440 assign Meta's workloads too. 0:02:40.520 --> 0:02:40.720 Well. 0:02:40.760 --> 0:02:43.000 What the CEO told me is that they never sign 0:02:43.040 --> 0:02:45.920 a deal if they don't have the right power and 0:02:46.040 --> 0:02:49.639 data center allocation ready to go. I don't think there's 0:02:49.720 --> 0:02:51.520 a data center that they're just going to flip on 0:02:51.680 --> 0:02:54.240 and say, aren't Meta here you go. But you know, 0:02:54.480 --> 0:02:57.200 this contract goes till twenty thirty one. They're probably going 0:02:57.280 --> 0:02:58.880 to need to buy a bunch of chips, fill it 0:02:58.960 --> 0:03:00.920 up and hand it over to Meta. And what that 0:03:01.000 --> 0:03:03.239 also means is they're going to take out a lot 0:03:03.280 --> 0:03:05.639 more debt. I mean, Corewave is kind of famous at 0:03:05.639 --> 0:03:09.200 this point for levering itself up to a pretty incredible degree, 0:03:09.240 --> 0:03:10.880 and we'll probably see a little more of that. 0:03:11.360 --> 0:03:14.600 The role of debt in financing AI infrastructure. That's becoming 0:03:14.600 --> 0:03:16.960 a story of the year twenty twenty five, Blue Most 0:03:16.960 --> 0:03:18.640 Brady Ford with the core we've metastory. 0:03:18.720 --> 0:03:19.080 Thank you. 0:03:19.560 --> 0:03:23.120 Meanwhile, AI's boom is sending investors searching for trades in 0:03:23.200 --> 0:03:27.160 an overlooked group, suppliers of the gear used to make chips. 0:03:27.360 --> 0:03:30.919 Let's bring in Bloombos Tech equity reporter Ryan blaceelca lots 0:03:30.919 --> 0:03:34.160 of readership on this story and some names that have 0:03:34.320 --> 0:03:37.720 not been at the forefront of what's happened within the semispace. 0:03:38.160 --> 0:03:39.120 Give us your reporting. 0:03:40.120 --> 0:03:42.320 Hey, good morning, Thank you for having me so. Yeah, 0:03:42.320 --> 0:03:44.600 this is a trend that we have been seeing throughout 0:03:44.640 --> 0:03:47.560 twenty twenty five now that some of the more well 0:03:47.600 --> 0:03:50.080 known AI trade seems like maybe those are getting a 0:03:50.080 --> 0:03:53.400 little bit mature. Investors are looking elsewhere. We've seen it 0:03:53.440 --> 0:03:56.360 in storage companies, We've seen it in memory companies, So 0:03:56.400 --> 0:03:59.480 now we're seeing it in the semiconductor capital equipment companies 0:03:59.720 --> 0:04:04.600 named like LAMB Research, Applied Materials, KLA Core, Terradyne. These 0:04:04.680 --> 0:04:07.320 kinds of companies which make the equipment that is used 0:04:07.320 --> 0:04:10.760 for building the semiconductors. Building all this manufacturing out these 0:04:10.760 --> 0:04:14.480 are getting sort of a second derivative move. Given the 0:04:14.600 --> 0:04:17.040 sheer build out we're seeing all the chips that are 0:04:17.080 --> 0:04:20.159 required for AI, obviously it's going to mean higher demand 0:04:20.200 --> 0:04:22.720 for the chip for the machines that make those chips. 0:04:22.839 --> 0:04:26.800 We constantly question valuations, ran should we be worried about them 0:04:26.839 --> 0:04:27.479 for these names? 0:04:28.400 --> 0:04:30.039 Well, I'd say if we were talking about this maybe 0:04:30.040 --> 0:04:31.520 a month or so ago, it would be a lot 0:04:31.640 --> 0:04:33.480 less of a ponent issue. But some of these companies 0:04:33.480 --> 0:04:37.120 have really seen very steep rises. They've really become momentum 0:04:37.160 --> 0:04:39.680 favorites over the past couple of weeks and months, and 0:04:39.720 --> 0:04:41.280 I think right now they are getting to a little 0:04:41.279 --> 0:04:43.599 bit more elevated levels. But certainly they are not sort 0:04:43.600 --> 0:04:46.760 of a nosebleed, real sort of bubble kind of valuation, 0:04:46.839 --> 0:04:48.440 but they certainly come up a lot from where they 0:04:48.440 --> 0:04:49.320 were earlier this year. 0:04:49.440 --> 0:04:51.680 Ryan, if you walk into a fab the clean room 0:04:51.720 --> 0:04:54.800 where semi condut is are manufactured, along the line is 0:04:54.880 --> 0:04:57.440 machines for lots of different companies. At the beginning of 0:04:57.440 --> 0:05:01.000 that line, you're probably likely to have one from LAMB Research. 0:05:01.080 --> 0:05:02.480 The chart that we took from your. 0:05:02.400 --> 0:05:05.479 Story shows LAMB being a real outperformer here. 0:05:05.520 --> 0:05:07.080 Is there anything more to know about that name. 0:05:08.360 --> 0:05:10.599 I think they are one that especially used with Micron, 0:05:10.680 --> 0:05:12.560 so the memory chips that has been another area of 0:05:12.600 --> 0:05:15.039 focus this year, so that high bandwidth memory. I believe 0:05:15.040 --> 0:05:17.560 that LAMB works on machines that are used in those 0:05:17.640 --> 0:05:20.920 I think that's probably certainly a component to their business there. 0:05:21.760 --> 0:05:24.719 Ran, So it's great to catch up with you, Ran, Plaselkah. 0:05:24.760 --> 0:05:25.520 We appreciate it. 0:05:25.520 --> 0:05:29.159 Meanwhile, sticking with chips at Taiwan's premiere, says that trade 0:05:29.200 --> 0:05:31.920 talks with the United States have ended, quote the crucial 0:05:32.080 --> 0:05:35.760 closing stages, indicating the global chiphub is finally nearing a 0:05:35.800 --> 0:05:37.000 deal with the Trump administration. 0:05:37.200 --> 0:05:37.320 Now. 0:05:37.400 --> 0:05:38.760 Investment in the US was. 0:05:38.760 --> 0:05:41.560 Among the issues discussed in Washington in recent days, according 0:05:41.560 --> 0:05:44.360 to a source, as well as lowering the twenty percent 0:05:44.440 --> 0:05:46.520 tariff imposed on the island ed. 0:05:46.800 --> 0:05:48.040 Yeah. Some other news we're tracking. 0:05:48.120 --> 0:05:51.480 Google's agreed to pay twenty four point five billion dollars 0:05:51.640 --> 0:05:55.400 to resolve Donald Trump's claims that being blocked from posting 0:05:55.400 --> 0:05:58.039 on his YouTube channel after the January sixth, twenty twenty 0:05:58.080 --> 0:06:01.919 one riot at the US Capitol amounted to illegal censorship. 0:06:01.960 --> 0:06:04.600 That's according to a court filing which also shows twenty 0:06:04.600 --> 0:06:07.360 two million dollars. We'll go toward construction of a new 0:06:07.400 --> 0:06:10.360 ballroom in the White House, a project near and dear 0:06:10.440 --> 0:06:13.040 to Trump, while the remainder will go to a handful 0:06:13.080 --> 0:06:17.599 of other plaintiffs who joined him in legal action. Okay, 0:06:17.640 --> 0:06:20.400 Coming up, and Thropic has a new AI model that 0:06:20.480 --> 0:06:23.440 codes on its own for up to thirty hours straight. 0:06:23.640 --> 0:06:26.680 No need to feel it soylent, no red bull, and 0:06:26.720 --> 0:06:29.400 Thropic chief product officer talks to us about Claude Sonnet 0:06:29.440 --> 0:06:31.360 four point five A conversations. 0:06:31.400 --> 0:06:32.880 Next, this is Bloomberg Tech. 0:06:46.800 --> 0:06:50.240 Deep Seak update is an experimental AI model, calling it 0:06:50.279 --> 0:06:54.200 a step toward next gen AI. The latest version introduces 0:06:54.200 --> 0:06:58.359 a new technique it calls deep Seek Sparse Attention or DSA, 0:06:58.480 --> 0:07:02.200 a mechanism designed to explore and optimize AI training and 0:07:02.240 --> 0:07:07.120 operation and improve efficiency when processing long tech sequences. 0:07:07.240 --> 0:07:10.600 Correct, let's stick on the models AI startup Anthropic is 0:07:10.600 --> 0:07:13.440 that with a new one Claudes on it four point five. 0:07:13.680 --> 0:07:16.480 The company says it can code longer, more effectively than 0:07:16.480 --> 0:07:20.200 prior versions. Let's get more with Anthropics chief product officer 0:07:20.280 --> 0:07:20.800 Mike Kreger. 0:07:20.960 --> 0:07:22.000 Mike, it's wonderful. 0:07:22.040 --> 0:07:22.920 How to have you on. 0:07:23.160 --> 0:07:26.840 Thirty hours straight is how long then it can code 0:07:26.960 --> 0:07:29.640 on its own? What are the technical feats needed to 0:07:29.680 --> 0:07:32.440 be able to go that long where humans can well 0:07:32.520 --> 0:07:34.880 definitely not survive that unless as a whole load of 0:07:34.880 --> 0:07:35.600 caffeine involved. 0:07:36.360 --> 0:07:36.880 Good morning. 0:07:36.920 --> 0:07:38.760 I think one of the main advancements we made was 0:07:38.800 --> 0:07:41.000 around memory and what we call context management. 0:07:41.000 --> 0:07:41.560 So if you think. 0:07:41.480 --> 0:07:44.320 About how you human works for longer periods of time, 0:07:44.440 --> 0:07:46.400 you're writing things down, You're making sure you can always 0:07:46.440 --> 0:07:48.320 pick up where you're left off if you're coming back 0:07:48.360 --> 0:07:50.920 the next day. So with cloud sign It four point five, 0:07:50.920 --> 0:07:52.960 we did a lot of work on that memory management. 0:07:53.000 --> 0:07:54.600 So the model, if you think about it, sort of 0:07:54.760 --> 0:07:57.119 writes down what it's doing, keeps track of its state, 0:07:57.200 --> 0:07:59.280 and then if it needs to sort of backtrack a 0:07:59.360 --> 0:08:01.160 table to then keep going. And that's how it's able 0:08:01.160 --> 0:08:03.280 to stay coherent for a much longer period of time 0:08:03.320 --> 0:08:04.760 than any other of our models. 0:08:05.240 --> 0:08:09.280 How much you've managed to lower therefore those ideas of 0:08:09.320 --> 0:08:13.600 inaccuracies or more broadly, that well that they're making things up. 0:08:13.680 --> 0:08:16.320 Hallucination has always been the key issue. Heren't been something 0:08:16.320 --> 0:08:18.080 that's limited edgentic AI adoption. 0:08:19.240 --> 0:08:20.800 Yeah, this is the model that we have that is 0:08:20.840 --> 0:08:23.800 also besides being our most powerful, is also are safest 0:08:23.920 --> 0:08:27.080 and most coherent, so as the lowest hallucination rate, and 0:08:27.200 --> 0:08:29.160 is the least susceptible to things like jail breaks. 0:08:29.160 --> 0:08:30.800 And I think that matters a lot. I think I 0:08:30.840 --> 0:08:31.800 tell my product. 0:08:31.520 --> 0:08:33.679 Team all the time, it's no use going for twenty 0:08:33.720 --> 0:08:36.040 thirty hours if you're making mistakes along the way, and 0:08:36.080 --> 0:08:39.080 so having it be both accurate producing good code that's 0:08:39.120 --> 0:08:41.360 the prerequisite, and then you can focus on scaling up 0:08:41.360 --> 0:08:42.160 on the time horizon. 0:08:42.520 --> 0:08:44.560 Mike, can we talk a bit about the audience for 0:08:44.679 --> 0:08:47.160 claudes on at four point five. The real emphasis from 0:08:47.240 --> 0:08:51.360 Entropic from the early days was enterprise customer as opposed 0:08:51.400 --> 0:08:54.240 to a sort of direct consumer. But this field of 0:08:54.320 --> 0:08:58.319 tools for the developer is expanding. It's probably more competitive. 0:08:58.640 --> 0:09:00.760 Who are you hoping uses this? 0:09:01.880 --> 0:09:04.320 It's really we've taken a business focus, but that also 0:09:04.400 --> 0:09:06.680 manifests kind of in the prosumer space, and so we 0:09:06.679 --> 0:09:08.640 have a lot of what we call power users who 0:09:08.880 --> 0:09:11.240 might be developers or might just be early adopters who 0:09:11.280 --> 0:09:13.120 want to bring AI to their work. So one of 0:09:13.200 --> 0:09:15.600 the things that SIGNUT four point five can do, along 0:09:15.600 --> 0:09:19.920 with writing code, is also creating really professional looking word 0:09:19.960 --> 0:09:22.839 and PowerPoint and Excel documents. It actually uses the same 0:09:23.080 --> 0:09:25.360 coding capabilities under the hood, but not to write code, 0:09:25.360 --> 0:09:28.560 but instead to produce documents. And that sort of capability 0:09:28.600 --> 0:09:31.079 means that we're starting to see adoption in the enterprise 0:09:31.120 --> 0:09:32.120 as well. 0:09:32.520 --> 0:09:34.000 Mike, you and I have discussed this in the past. 0:09:34.000 --> 0:09:38.760 We know Mike Krieger as Mike Krieger CTO Instagram, And 0:09:38.800 --> 0:09:40.839 what I'm seeing right now in the field of your 0:09:40.880 --> 0:09:44.679 peers is the reports on open Ai and what Meta 0:09:44.840 --> 0:09:48.880 is doing in social media and video. Is that a 0:09:48.920 --> 0:09:51.400 direction you want the product team to take, claud. 0:09:51.120 --> 0:09:55.280 In, We're focused much more on the productivity use case, 0:09:55.280 --> 0:09:56.960 and so when I think about our roadmap, it's very 0:09:57.040 --> 0:09:59.280 much how do we take workoff people's hands, or how 0:09:59.320 --> 0:10:01.320 do we accelerate folks and make them, you know, the 0:10:01.320 --> 0:10:03.080 work the best that it can be. How do we 0:10:03.160 --> 0:10:05.760 automate your work in the browser? So much more on 0:10:05.760 --> 0:10:07.679 that productivity side of things, and I don't think you'll 0:10:07.679 --> 0:10:09.920 see us play very much in that entertainment space. 0:10:10.200 --> 0:10:15.640 Might productivity perspective. It's come under some concerns recently. Think 0:10:15.640 --> 0:10:18.800 about the MIT report everyone's suddenly talking about this only 0:10:19.160 --> 0:10:22.560 well ninety five percent of the tests were basically failing 0:10:22.640 --> 0:10:25.240 out there in the wild. How are you making sure 0:10:25.280 --> 0:10:28.520 that enterprises adopt your products but actually see the productivity 0:10:28.559 --> 0:10:29.560 gains from them. 0:10:30.040 --> 0:10:32.200 I think this is really important, where if you know, 0:10:32.400 --> 0:10:34.959 AI gets brought into the workplace without the right either 0:10:35.120 --> 0:10:37.320 tools around it or enablement, what you end up with 0:10:37.440 --> 0:10:40.000 is this disillusionment a couple of months later around while 0:10:40.000 --> 0:10:42.160 folks aren't adopting it, or yeah, it helped me a 0:10:42.200 --> 0:10:43.680 little bit, but not enough. And so we have a 0:10:43.679 --> 0:10:45.560 lot of emphasis on let's. 0:10:45.320 --> 0:10:46.880 Make sure the work is actually good. 0:10:46.920 --> 0:10:49.160 You know, you might hear this word online like slop, 0:10:49.200 --> 0:10:51.800 where AI is creating work that actually just is not 0:10:51.920 --> 0:10:53.280 very good. And I think of us, we're trying to 0:10:53.280 --> 0:10:55.640 produce the anti slop work that actually, you know, maybe 0:10:55.720 --> 0:10:57.319 gets you eighty percent of the way there, but it's 0:10:57.360 --> 0:10:59.480 eighty percent that then lets you complete the work in 0:10:59.520 --> 0:11:01.480 a way that you're proud of, rather than you know, 0:11:01.600 --> 0:11:03.400 opes it did something, but now I feel like I 0:11:03.440 --> 0:11:04.880 have to start over because it didn't really help. So 0:11:05.000 --> 0:11:07.560 I think that's the really key piece for enterprise adoption. 0:11:07.960 --> 0:11:11.320 And therefore, does it remove the need for so many 0:11:11.360 --> 0:11:14.880 people or ultimately there's still this argument this can all meant, 0:11:15.240 --> 0:11:16.520 but will it start to replace? 0:11:16.600 --> 0:11:19.440 Mike, we think a lot about what, you know, the 0:11:19.480 --> 0:11:22.960 comparative advantages are of people, you know, as relates to AI. 0:11:23.040 --> 0:11:26.720 There's a lot of still relationship building and trust, critical 0:11:26.760 --> 0:11:30.480 analysis and strategy that really comes on the human side 0:11:30.520 --> 0:11:32.760 of things, and so we really try to design tools 0:11:32.800 --> 0:11:35.560 that as much as possible play up those parts of 0:11:35.559 --> 0:11:38.040 that human AI interaction, knowing that you know there will 0:11:38.040 --> 0:11:40.719 be you know, laborships that are almost inevitable, but if 0:11:40.720 --> 0:11:43.040 we can design our products along the way to maximize 0:11:43.040 --> 0:11:45.480 both people's understanding of AI but also their use in 0:11:45.520 --> 0:11:46.640 a complementary way. 0:11:47.080 --> 0:11:50.400 Mike, open a eyes holding dev day next Monday. It's 0:11:50.440 --> 0:11:55.120 probably on your calendar for peripheral awareness. But I'm very 0:11:55.160 --> 0:11:57.200 conscious that you're kind of speaking to us twenty four 0:11:57.200 --> 0:11:59.640 hours after the news of clots on four point five 0:11:59.720 --> 0:12:02.360 km out. Have you any data on the sort of 0:12:02.360 --> 0:12:07.719 reaction to it, early demand and where that's coming from. 0:12:08.160 --> 0:12:11.040 It's been really interesting how quick people are to adopt 0:12:11.040 --> 0:12:13.440 a new model. So by I think about one pm yesterday, 0:12:13.480 --> 0:12:15.800 we already had more usage of clouds on at four 0:12:15.800 --> 0:12:18.080 point five than all of our other models combined, which 0:12:18.120 --> 0:12:20.280 really speaks to the eagerness of a lot of these 0:12:21.160 --> 0:12:23.600 startups that are building on top of our models, as 0:12:23.640 --> 0:12:26.240 well as early adopters to on day one you saw 0:12:26.559 --> 0:12:29.440 GitHub and Cursor and Windsurf and many of these products 0:12:29.480 --> 0:12:31.880 that build on top of our models want to incorporate 0:12:31.920 --> 0:12:33.520 clouds on at four point five, and so we had 0:12:33.520 --> 0:12:35.680 this really early crossover moment as well. 0:12:35.800 --> 0:12:37.040 That is interesting data. 0:12:38.440 --> 0:12:41.680 Correct me if I'm wrong, But this model is running 0:12:41.960 --> 0:12:45.160 on Project Rainier, right, is that sort of operationally and 0:12:45.240 --> 0:12:49.000 infrastructure wise where the training and now inference of it 0:12:49.040 --> 0:12:49.640 is being done. 0:12:50.640 --> 0:12:53.520 So we do our both training and inference across You know, 0:12:53.559 --> 0:12:56.199 we have partnerships with Google and Amazon, but we have 0:12:56.240 --> 0:13:01.000 a significant part of this model being served now from 0:13:01.040 --> 0:13:03.040 Amazon as well, and we're seeing a lot of growth 0:13:03.080 --> 0:13:04.480 on a tobospedrock as well. 0:13:05.360 --> 0:13:08.160 Just going to that infrastructure layer. You're obviously the product 0:13:08.280 --> 0:13:11.480 visionary here, but you need to have the energy the 0:13:11.520 --> 0:13:14.560 compute to bring your products to life. Many worrying that 0:13:14.559 --> 0:13:17.360 we're in some sort of bubble cycle around AI. How 0:13:17.400 --> 0:13:19.360 do you think about that as you drive this business forward? 0:13:19.440 --> 0:13:22.400 Might I think there's this combined need to both scale 0:13:22.480 --> 0:13:24.959 up for the training side, but also on inference. And 0:13:25.040 --> 0:13:27.800 as we've scaled, especially with our business arrangements and the 0:13:27.800 --> 0:13:29.640 companies building on top of son it, I think we 0:13:29.720 --> 0:13:33.000 now have a sort of forward looking perspective on what 0:13:33.640 --> 0:13:35.440 our inference needs will be, and I think that'll let 0:13:35.480 --> 0:13:37.520 us go out and also secure the kinds of compute 0:13:37.559 --> 0:13:39.600 deals that we need to both feel the training but 0:13:39.679 --> 0:13:43.000 also have that sort of revenue generating inference side as well. 0:13:43.360 --> 0:13:45.760 And we get so focused on the compute needs of 0:13:45.840 --> 0:13:48.320 the United States, but we've been talking a lot about 0:13:48.320 --> 0:13:50.520 that in Europe. How it's scaling in the UK. From 0:13:50.520 --> 0:13:53.600 your adoption, how are you seeing things different globally MIC, 0:13:53.600 --> 0:13:56.520 Because those that are actually deploying what's happening with Sonnet 0:13:56.520 --> 0:13:59.160 four point five in the latest models, we see this 0:13:59.200 --> 0:13:59.520 a lot. 0:13:59.440 --> 0:14:00.920 In terms of our global footprint. 0:14:01.120 --> 0:14:03.640 There's something we started expanding earlier this year, and so 0:14:03.760 --> 0:14:06.400 for our rollout of Trainium too, which is the chip 0:14:06.440 --> 0:14:09.120 that Amazon has built for a WS that we use 0:14:09.320 --> 0:14:12.040 pretty extensively for our cloud models. A lot of that 0:14:12.080 --> 0:14:14.400 deployment is actually international, and when I go to Europe, 0:14:14.400 --> 0:14:17.120 for example, I hear a lot of questions about data 0:14:17.160 --> 0:14:19.400 locality and making sure that inference is happening in local 0:14:19.480 --> 0:14:20.120 data centers. 0:14:20.200 --> 0:14:21.160 And the only way we're going to be. 0:14:21.120 --> 0:14:23.040 Able to do that is to have that international footprint 0:14:23.600 --> 0:14:24.320 of these chips. 0:14:24.320 --> 0:14:25.640 And so you've seen the same in APEC. 0:14:26.880 --> 0:14:27.160 Mike. 0:14:28.040 --> 0:14:30.440 We are going to ask you a question about talent wars, 0:14:30.800 --> 0:14:32.680 but I'm just going to make an appeal to you 0:14:33.520 --> 0:14:36.280 to just be honest with me on this. How big 0:14:36.280 --> 0:14:38.440 a factor it is or isn't for you right now? 0:14:38.480 --> 0:14:41.720 In the product team at Anthropic in particular, I'm looking 0:14:41.720 --> 0:14:44.520 at the pace at which open ai is putting stuff out, 0:14:44.640 --> 0:14:48.960 Meta is putting stuff out. Just through your experience, what's 0:14:49.000 --> 0:14:50.960 the talent situation right now? 0:14:51.920 --> 0:14:54.920 I'm seeing much more of that talent sort of you know, 0:14:55.000 --> 0:14:58.080 back and forth happen within the research site in general, 0:14:58.120 --> 0:15:00.000 a little bit less on the product. I think there's 0:15:00.040 --> 0:15:02.520 some key hires where that's been the case. One thing 0:15:02.560 --> 0:15:06.160 that's been a positive sort of maybe surprise or just 0:15:06.200 --> 0:15:09.440 outcome of how mission oriented a lot of Anthropic, a 0:15:09.480 --> 0:15:11.600 lot of the Entropic team really is has been. It's 0:15:11.640 --> 0:15:13.520 affected us very minimally in terms of that back and 0:15:13.560 --> 0:15:15.400 forth that you're seeing maybe among some of the other 0:15:16.080 --> 0:15:17.120 frontier labs, which. 0:15:17.000 --> 0:15:17.600 Is very encouraging. 0:15:17.680 --> 0:15:19.000 Of course, you have to continue to make sure we 0:15:19.000 --> 0:15:21.440 build a great culture and maintain that mission alignment. 0:15:21.480 --> 0:15:24.480 But so far it's it's been minimally affecting us. 0:15:24.440 --> 0:15:26.120 If we take so on it. Four point five is 0:15:26.160 --> 0:15:29.040 the case study. What were the types of roles that 0:15:29.080 --> 0:15:32.080 you needed to bring in to roll out the release. 0:15:32.960 --> 0:15:35.920 I think people think of, you know, research and model 0:15:35.960 --> 0:15:37.920 science as being fairly cut and dry. I actually think 0:15:37.960 --> 0:15:39.800 that there's a lot of art and taste to it 0:15:39.840 --> 0:15:42.320 as well. You are making a lot of decisions from 0:15:42.320 --> 0:15:45.280 a research engineering perspective around what are the task the 0:15:45.280 --> 0:15:47.360 model needs to improve on, how will it improve on that, 0:15:47.400 --> 0:15:49.640 how will we know that it's improving on that? And 0:15:49.680 --> 0:15:52.560 so a lot of that reinforcement learning post training piece 0:15:53.200 --> 0:15:55.880 is the key shape of what we really thought about. 0:15:55.920 --> 0:15:57.120 And son at four point five. 0:15:57.400 --> 0:15:59.600 My creagain andthropic chief product off. So it was a 0:15:59.640 --> 0:16:01.840 real deep dive into Sonnet four point five. Thank you 0:16:01.880 --> 0:16:04.400 so much for joining us. Back on Bloombo Tech, I've 0:16:04.440 --> 0:16:06.920 coming up on the show Door Dash has a new 0:16:07.080 --> 0:16:08.520 autonomous robot. 0:16:08.640 --> 0:16:11.040 Meet Dot, Hi Dot. 0:16:11.560 --> 0:16:13.920 We'll talk to the VP of door Dash Labs about 0:16:14.160 --> 0:16:16.520 what this robot delivers. 0:16:16.480 --> 0:16:19.760 In terms of its capabilities. That's next. This is Bloomberg tech. 0:16:46.200 --> 0:16:50.000 Delivery company door Dash is unveiling a new autonomous delivery 0:16:50.160 --> 0:16:53.920 robot called Dot. The company says Dot is the first 0:16:53.920 --> 0:16:58.880 commercial delivery bot to seamlessly navigate bike lanes, roads and sidewalks. 0:16:59.280 --> 0:17:03.080 His stirs more actually read vice president of door Dash Labs. 0:17:03.520 --> 0:17:08.320 This robotics experience includes work with zooks and video, but 0:17:08.960 --> 0:17:14.200 the real terms parameters for deployment, which cities, how many 0:17:14.240 --> 0:17:15.720 deliveries when. 0:17:17.359 --> 0:17:20.520 So, we have been focusing on the greater Phoenix area 0:17:20.800 --> 0:17:24.639 and that's our focus for this year. We are hoping 0:17:24.720 --> 0:17:28.040 to address about one point five million customers by the 0:17:28.119 --> 0:17:30.840 end of this year, and then you know, as we progress, 0:17:30.960 --> 0:17:33.600 we'll see what happens. DoD likes to travel though, so 0:17:34.359 --> 0:17:36.639 hopefully we can expand to more cities. 0:17:36.880 --> 0:17:39.439 Team, let's bring up some pictures of Dot again. I 0:17:39.480 --> 0:17:42.600 mean to me assue you that this looks kind of 0:17:42.960 --> 0:17:46.399 the design like a bit like a stroller or a 0:17:46.440 --> 0:17:48.679 pram as we might have said in the United Kingdom. 0:17:49.119 --> 0:17:51.199 Could you just talk us through the design of it 0:17:51.280 --> 0:17:53.760 and how this is what you arrived at. 0:17:55.080 --> 0:17:55.399 Sure. 0:17:55.640 --> 0:17:58.360 So there were three sort of key pillars for us 0:17:58.440 --> 0:18:01.600 when we looked into the line of Dot. The first 0:18:01.600 --> 0:18:03.840 one was product market fit, so we wanted to make 0:18:03.880 --> 0:18:08.800 sure that Dot would have the right cargo capacity and 0:18:09.200 --> 0:18:11.600 the payload that it can carry up to thirty pounds 0:18:12.040 --> 0:18:15.919 fits a vast amount of the door dash deliveries that 0:18:15.960 --> 0:18:19.080 we do today. The other big part was it had 0:18:19.119 --> 0:18:22.880 to be going at speeds that allowed us to do 0:18:23.320 --> 0:18:26.040 a big chunk of deliveries in the three to mile range, 0:18:26.040 --> 0:18:29.919 and that's why the design point was it needs to 0:18:29.960 --> 0:18:33.400 be able to go on bike lanes, on roads, on sidewalks. 0:18:33.920 --> 0:18:36.480 But the other key part is the pickup and drop 0:18:36.520 --> 0:18:39.880 off for deliveries, which makes it very different from right hail. 0:18:40.160 --> 0:18:42.120 And one of the key feedbacks we had gotten from 0:18:42.160 --> 0:18:46.440 merchants was it is absolutely imperative to have the robot 0:18:46.800 --> 0:18:50.720 come as close as possible to the merchant, right next 0:18:50.720 --> 0:18:53.440