WEBVTT - DeepSeek's AI Brings Tech Rout 0:00:02.520 --> 0:00:09.600 Bloomberg Audio Studios, podcasts, radio news from the heart of 0:00:09.680 --> 0:00:15.400 where innovation, money and power collide in Silicon Valley and beyond. 0:00:15.800 --> 0:00:21.079 This is Bloomberg Technology with Caroline Hyde and Ed Ludlow. 0:00:35.000 --> 0:00:37.040 Live from New York. I'm Caroline Hyde. 0:00:37.159 --> 0:00:40.959 And I'm Mike Sheppard in San Francisco. This is Bloomberg Technology, and. 0:00:40.920 --> 0:00:44.280 We have special coverage of the Chinese AI competitor Deep 0:00:44.320 --> 0:00:47.880 Seek and its impact on technanues the broader markets at large. 0:00:47.920 --> 0:00:50.360 Let's look at the nastat, currently having its worst day 0:00:50.400 --> 0:00:52.800 in more than a month, as we question the need 0:00:53.040 --> 0:00:56.080 for compute power to get us ever more efficient, ever 0:00:56.200 --> 0:00:59.040 more specific generative AI. 0:00:59.480 --> 0:01:00.720 We're down too point six percent. 0:01:00.760 --> 0:01:02.560 Move on and look at the individual names, because this 0:01:02.600 --> 0:01:05.360 is when we question in video the compute power, the 0:01:05.440 --> 0:01:08.720 necessity forever more efficient chips, and the cost of them. 0:01:08.959 --> 0:01:09.959 In Vidia has its. 0:01:09.959 --> 0:01:14.080 Worst route in history in terms of market capitalization ever 0:01:14.160 --> 0:01:16.280 on the market. We have never seen almost half a 0:01:16.319 --> 0:01:19.440 trillion dollars wiped out from one single name. This is 0:01:19.560 --> 0:01:23.240 global ASML chip equipment maker in Europe down almost six percent, 0:01:23.480 --> 0:01:26.319 But Hong Kong stocks do well as we start to 0:01:26.319 --> 0:01:30.360 see AI as a real champion for China. Let's talk 0:01:30.360 --> 0:01:32.720 about it all with Bloomberg Intelligence senior analyst man Leap 0:01:32.760 --> 0:01:35.880 Singh who joins us some more. How does this change 0:01:35.880 --> 0:01:39.119 the game for llm's in the US, for chip equipment 0:01:39.440 --> 0:01:40.040 in the US. 0:01:40.400 --> 0:01:42.360 Yeah, Look, I mean we look at it from a 0:01:42.400 --> 0:01:45.920 stack perspective, and with generative AI, so far, all the 0:01:46.080 --> 0:01:50.320 value was captured by the chip makers, the semicap equipment guys, 0:01:50.360 --> 0:01:54.120 and the LLM companies like open Ai and Tropic, and 0:01:54.200 --> 0:01:57.560 the software guys were really missing out because everyone was like, 0:01:57.600 --> 0:01:59.960 their margins will be compressed, they have to pay more 0:02:00.080 --> 0:02:03.360 for chips, pay for this LLM layer. Well, guess what 0:02:03.680 --> 0:02:06.800 if the price can come down and what deep seek 0:02:06.840 --> 0:02:10.200 has shown us is the price per token, price per 0:02:10.280 --> 0:02:14.960 query can come down drastically, then suddenly the I think 0:02:14.960 --> 0:02:17.440 the power goes back to the software guys because now 0:02:17.480 --> 0:02:20.800 they can incorporate AI in everything they want, in their 0:02:20.840 --> 0:02:26.040 app and their website, whatever functionality they have, they can 0:02:26.120 --> 0:02:29.720 add a layer of AI without having to pay too much, 0:02:29.880 --> 0:02:32.560 you know, and hurting their growth margins. And so I 0:02:32.600 --> 0:02:36.520 think this really is a big development. And every cloud company, 0:02:36.520 --> 0:02:40.600 whether it's Alphabet or Microsoft, they'll be looking to leverage 0:02:40.639 --> 0:02:43.800 their compute more efficiently because they will benefit from the 0:02:43.800 --> 0:02:46.640 cloud revenue on the infrastructure side, but also on the 0:02:46.680 --> 0:02:49.080 application side, which is why I think Meta raise their 0:02:49.120 --> 0:02:50.480 capex on Friday. 0:02:51.639 --> 0:02:55.160 Mandep How does this undercut in Video in particular in 0:02:55.200 --> 0:03:00.080 its argument for advances like the Blackwell product that it 0:02:59.960 --> 0:03:02.960 is pushing out now and what may come after it? 0:03:03.000 --> 0:03:06.640 Is it becoming much harder for Jensen want to sell 0:03:06.680 --> 0:03:08.919 that to Microsoft and it's other customers. 0:03:08.919 --> 0:03:11.680 Now all of a sudden, look, I mean the last 0:03:11.680 --> 0:03:14.520 three months we were debating about the scaling laws and 0:03:14.600 --> 0:03:17.600 how big of a cluster we need. We're talking about 0:03:17.600 --> 0:03:21.480 one hundred kgpus in a cluster and possibly one million 0:03:22.160 --> 0:03:25.200 GPUs in a cluster. I think this puts into question 0:03:25.280 --> 0:03:28.360 mark how big of a cluster you need actually for 0:03:28.440 --> 0:03:31.919 training your LLM And look, all these LLM companies have 0:03:32.000 --> 0:03:35.360 to keep training their algorithms. You know, this is not 0:03:35.440 --> 0:03:38.360 a one and done thing, so you need compute capacity. 0:03:38.480 --> 0:03:41.160 But at the same time, and Video was accruing all 0:03:41.200 --> 0:03:45.120 the benefits in terms of pricing because of the advances 0:03:45.120 --> 0:03:47.640 in their chips. So if you can do more with 0:03:47.760 --> 0:03:50.440 even the you know, the older version of their chips 0:03:50.680 --> 0:03:53.760 and you don't need a cluster of one million GPUs 0:03:54.120 --> 0:03:57.320 connected together, then probably I think you can do more 0:03:57.440 --> 0:04:00.280 in terms of the distributed infrastructure. And I think that's 0:04:00.320 --> 0:04:02.880 what a lot of people are trying to figure out today. 0:04:03.920 --> 0:04:06.720 The guys Man Deep saying thank you. Let's bring in 0:04:06.800 --> 0:04:10.440 Tony Wong, portfolio manager of the Science and Technology Fund 0:04:10.480 --> 0:04:13.480 at t row Price and a key and video shareholder 0:04:13.520 --> 0:04:14.400 for more on this. 0:04:15.000 --> 0:04:15.320 Tony. 0:04:15.400 --> 0:04:18.520 We've got a big week and month of earnings coming up. 0:04:18.720 --> 0:04:21.000 How does this change the way you were looking at 0:04:21.000 --> 0:04:23.480 companies as they prepare to deliver their results. 0:04:25.279 --> 0:04:27.919 Yeah, I think it's a great question. And if you 0:04:27.920 --> 0:04:31.000 look at the kind of trends and technology generally, when 0:04:31.040 --> 0:04:35.039 things become more efficient, increases demand and there's more ubiquity. 0:04:35.080 --> 0:04:38.280 And so it wasn't that long ago, you know, a 0:04:38.320 --> 0:04:41.200 few months that we're all debating that, like the cost 0:04:41.360 --> 0:04:43.960 was too great to deliver and so where is the ROI? 0:04:44.000 --> 0:04:47.640 And so I actually feel like this is tam expansionary 0:04:47.760 --> 0:04:50.640 for the market long term. I think that Man Deep, 0:04:50.720 --> 0:04:54.560 you know, obviously has highlighted what the market is kind 0:04:54.560 --> 0:04:55.320 of talking about today. 0:04:55.360 --> 0:04:56.919 I think that's a very valid concern. 0:04:57.040 --> 0:05:01.560 But Javon's paradox essentially is that when things become more efficient, 0:05:01.800 --> 0:05:04.599 consumption goes up, demand goes up, and so I think 0:05:04.600 --> 0:05:08.560 that's that's necessary, that's healthy. I do think that there's 0:05:08.600 --> 0:05:13.760 a little bit of sensational reporting perhaps on Twitter and 0:05:13.839 --> 0:05:16.839 just in terms of how deep seek only took you know, 0:05:16.880 --> 0:05:18.159 six million dollars to train. 0:05:18.240 --> 0:05:20.559 I think that's a that's a that's a nice marketing number. 0:05:20.560 --> 0:05:22.320 But if you dig in that, you know, that doesn't 0:05:22.320 --> 0:05:25.159 include R and D costs, experimental runs, and the market. 0:05:25.520 --> 0:05:28.840 The model is a lot smaller than what chatchbt is 0:05:28.880 --> 0:05:31.320 and it's only tech based, so there's a lot of 0:05:31.360 --> 0:05:34.920 reasons to think that this is not exactly you know, 0:05:35.040 --> 0:05:38.680 you know, a kind of direct comparison. They do have 0:05:38.720 --> 0:05:43.240 a lot of great breakthroughs in the efficiency of the model, obviously, 0:05:43.360 --> 0:05:45.680 but I think it's good for the industry, and it's 0:05:45.680 --> 0:05:49.000 open source and the players will you know, the community 0:05:49.000 --> 0:05:53.120 will adopt the innovations and then further kind of pursue 0:05:53.160 --> 0:05:55.200 like kind of AI in a bigger way. 0:05:55.240 --> 0:05:58.880 I think to that point, let's just reconfirm to our 0:05:58.920 --> 0:06:02.719 audience what is that deep seek has done. Because Bloomberg 0:06:02.800 --> 0:06:06.360 Intelligence Bloomberg News has been reporting on deep Seek four months. 0:06:06.360 --> 0:06:09.880 Bloomberg Intelligence ranked it as the seventh most powerful LLM 0:06:09.960 --> 0:06:12.120 all the way back in June of last year. But 0:06:12.240 --> 0:06:14.440 now it gives us the fact that it's latest. Our 0:06:14.560 --> 0:06:18.280 one model is as sophisticated. Money would say on the 0:06:18.279 --> 0:06:21.680 internet and other experts out there is as sophisticated as 0:06:21.680 --> 0:06:24.800 that of anthropics latest model, or of open A's open 0:06:24.800 --> 0:06:27.200 AI's latest model. But they've done it at a fraction 0:06:27.240 --> 0:06:29.760 of the cost. But six million is what is reported 0:06:29.800 --> 0:06:32.200 to have been cost. We have been able to delve 0:06:32.240 --> 0:06:36.120 into the veracity of that number, Tony. When it comes 0:06:36.160 --> 0:06:41.240 to the rush to the bottom ever more cheap of 0:06:41.240 --> 0:06:45.800 applications of generative AI, do we therefore see Microsoft have 0:06:45.920 --> 0:06:47.400 to cut its prices. 0:06:47.480 --> 0:06:48.360 Do we see. 0:06:48.320 --> 0:06:52.000 Ultimately revenue and profitability of these LM providers having to 0:06:52.040 --> 0:06:52.920 become less? 0:06:53.560 --> 0:06:54.920 Yeah, I think that's a good question. 0:06:55.080 --> 0:06:58.280 I would say that, you know, for why we don't 0:06:58.320 --> 0:07:00.800 know why open ai in it Thropic are going to 0:07:00.839 --> 0:07:03.039 release next and a lot of times, like you know, 0:07:03.080 --> 0:07:06.200 there are so many innovations and marketing is moving very quickly, 0:07:06.360 --> 0:07:08.920 so you can very well that they've got something that's 0:07:09.000 --> 0:07:10.200 really exciting to be launched. 0:07:10.680 --> 0:07:11.600 So that's number one. 0:07:11.600 --> 0:07:13.520 I think like number two is that a lot of times, 0:07:13.680 --> 0:07:16.120 you know, this is a fast follower bress model that 0:07:16.400 --> 0:07:17.920 has a ceiling in terms of what it can do 0:07:18.080 --> 0:07:20.600 versus larger models, and it's you know, open ai is 0:07:20.640 --> 0:07:24.120 going to be multimodal with their model, and you know, 0:07:24.160 --> 0:07:24.560 I think. 0:07:24.440 --> 0:07:26.559 There's a lot of innovation that will happen. 0:07:26.640 --> 0:07:28.280 And then I think that like at the end of 0:07:28.320 --> 0:07:30.280 the day, like you know, there's a lot of kind 0:07:30.280 --> 0:07:35.400 of ecosystem building, enterprise aspects that that also factor into it. 0:07:35.480 --> 0:07:37.680 So I think that the market for oms can be 0:07:37.720 --> 0:07:40.000 more diverse than people are giving credit for. 0:07:40.720 --> 0:07:43.120 Is this a buying opportunity Tony to stack up on 0:07:43.240 --> 0:07:43.640 end video? 0:07:43.680 --> 0:07:44.800 Even more, I. 0:07:44.800 --> 0:07:48.280 Think long term, you know, you think about the technology trends, 0:07:48.720 --> 0:07:52.480 and I don't think that just because there's one model 0:07:52.520 --> 0:07:54.600 that is more efficient that people are going to say, Okay, 0:07:54.600 --> 0:07:57.040 we're just going to stop here. You know, at the 0:07:57.120 --> 0:08:00.440 end of the day, everybody's racing towards hopefully AGI and 0:08:00.840 --> 0:08:03.440 you know, looking to continue to invest. 0:08:03.440 --> 0:08:05.040 And I think that you saw Mark. 0:08:04.920 --> 0:08:08.400 Zuckerber, even though he knew about Deep Seek, you know, 0:08:08.440 --> 0:08:11.680 since it was released over Christmas, you saw that he 0:08:11.720 --> 0:08:14.440 actually increased CAPEC spend, and so, you know, I think 0:08:14.480 --> 0:08:16.760 that there's it's not just the model, it's also the 0:08:16.800 --> 0:08:20.480 compute as well, and so I think that as there's 0:08:20.520 --> 0:08:22.440 more demand of the app layer, that's actually good for 0:08:23.200 --> 0:08:26.080 filling up the compute capacity that we're building. 0:08:26.800 --> 0:08:30.560 Tony, you talked about Mark Zuckerberg and his announcement last 0:08:30.600 --> 0:08:35.160 week on capex. Is there any thought to the way 0:08:35.400 --> 0:08:38.760 these companies should revisit those spending plans? 0:08:38.800 --> 0:08:39.920 Are they going too big? 0:08:40.040 --> 0:08:43.240 Is it too big and maybe not efficient enough? 0:08:44.000 --> 0:08:44.200 Yeah? 0:08:44.240 --> 0:08:45.479 Well, naturally, absolutely. 0:08:45.520 --> 0:08:47.840 I think that everybody's going to the engineering teams and 0:08:47.880 --> 0:08:49.840 saying like, hey, what can we do? This is open 0:08:49.840 --> 0:08:52.440 source model, let's take it apart, let's figure out how 0:08:52.480 --> 0:08:55.080 we can incorporate the best and you know, as cost 0:08:55.160 --> 0:08:57.760 him down, Like, I think that actually increases the number 0:08:57.840 --> 0:08:59.920 of use cases and the people say like, oh wow, 0:09:00.240 --> 0:09:04.120 I wasn't able to afford perhaps the latest like equipment 0:09:04.200 --> 0:09:06.719 because of the cost of doing so, we're not sure 0:09:06.720 --> 0:09:07.839 about the ROI and so. 0:09:08.080 --> 0:09:10.600 This I do think that it's it's absolutely natural. 0:09:10.640 --> 0:09:13.320 But I don't think that companies are going to say like, oh, like, 0:09:13.600 --> 0:09:14.560 let's just stop here. 0:09:15.480 --> 0:09:18.480 And because there are so many like potential benefits, the 0:09:19.800 --> 0:09:21.439 kind of goal at the end of the day is 0:09:21.800 --> 0:09:22.440 pretty grand. 0:09:23.640 --> 0:09:23.920 Tony. 0:09:23.920 --> 0:09:26.120 I wanted to follow with a question about the potential 0:09:26.320 --> 0:09:31.440 policy responses from Washington. Given the hawkish mood towards China 0:09:31.559 --> 0:09:35.559 with the new administration. Do you see any risk of 0:09:35.640 --> 0:09:39.800 a response that would pose a you know, a challenge 0:09:39.840 --> 0:09:41.960 to the companies that you own and over and are 0:09:42.040 --> 0:09:42.840 taking a look at. 0:09:43.520 --> 0:09:45.840 Yeah, well, I think that definitely there might be some 0:09:45.880 --> 0:09:48.800 geopolitical aspects of what's going on this week. You know, 0:09:48.840 --> 0:09:53.200 I think President Trump announced the Stargate project and so 0:09:53.400 --> 0:09:56.040 this was perhaps you know, deep seek is a response 0:09:56.080 --> 0:09:58.040 to that, and that you know, the question is like 0:09:58.400 --> 0:10:03.040 can you build AI more cheaply and openly versus like 0:10:03.080 --> 0:10:05.920 building a bunch of AI infrastructure, And I think that, 0:10:06.440 --> 0:10:08.920 you know, my view is it probably you need both, 0:10:09.400 --> 0:10:13.520 and you know, each country will continue to race towards 0:10:14.160 --> 0:10:17.200 you know, AI capabilities, and I think there could be 0:10:17.240 --> 0:10:21.240 some export restrictions more more along the way, possibly, but 0:10:21.320 --> 0:10:26.319 it does show that you know, perhaps like restricting isn't 0:10:26.400 --> 0:10:32.319 particularly helpful because then you actually bore innovation out of scarcity. 0:10:32.720 --> 0:10:34.800 I think that what we've seen is that there is 0:10:34.800 --> 0:10:37.200 a lot of innovation we had and you know, I think, 0:10:37.400 --> 0:10:40.599 you know, competition is good, and you know, at the 0:10:40.679 --> 0:10:42.280 end of the day, like you want an open market. 0:10:42.920 --> 0:10:47.280 Who will win from your portfolio from that innovation or 0:10:47.320 --> 0:10:50.959 potentially from that shift from hardware more to the software side. 0:10:51.160 --> 0:10:52.760 Yeah, I think it's a it's a really exciting time 0:10:52.800 --> 0:10:55.360 because I do think that at the app layer this 0:10:55.600 --> 0:10:58.360 improves kind of the r oy obviously, and you know, 0:10:58.559 --> 0:11:01.079 anybody that has like a really taking customer base and 0:11:01.679 --> 0:11:04.959 large kind of kind apps that they have a lot 0:11:05.000 --> 0:11:06.560 of data from, I think that's good. 0:11:06.840 --> 0:11:08.280 You know, I do like you know, the. 0:11:08.440 --> 0:11:11.880 AI infrastructure still, I think that there's you know, still 0:11:12.000 --> 0:11:15.520 steady kind of consistent demand there and you know, history 0:11:15.559 --> 0:11:19.640 of technology would say that people take that innovation that 0:11:19.679 --> 0:11:23.160 improved performance and they put into good use for developing 0:11:23.559 --> 0:11:25.640 the next next here technologies. 0:11:26.280 --> 0:11:29.240 Turning on Portfolio Manager at t ro Price, one of 0:11:29.360 --> 0:11:32.679 the most significant in video shareholders, we so appreciate it. 0:11:32.720 --> 0:11:35.280 Coming up will be joined by former White House Chief 0:11:35.280 --> 0:11:40.360 Information Officer Theresa Paytent to discuss deepseeks, cybersecurity implications, and more. 0:11:40.880 --> 0:11:43.439 But there are rather earnings upon us keep us a 0:11:43.480 --> 0:11:45.520 float with AT and T. We'll bring you what AT 0:11:45.640 --> 0:11:47.800 and T is currently doing on the market. We're higher 0:11:48.080 --> 0:11:51.040 in that particular. Stop fourth quarter results came in better 0:11:51.120 --> 0:11:54.040 than had been expected, driven by seasonal promotions for AT 0:11:54.160 --> 0:11:56.280 and T and bundled product offerings. 0:11:56.600 --> 0:12:03.200 This is bloot technology. 0:12:09.200 --> 0:12:12.440 There's no question that this could be a potential game changer. 0:12:12.640 --> 0:12:14.880 It's a game changer for the Max seven stocks. 0:12:15.120 --> 0:12:17.880 Tech has to be okay simply because it's arch a 0:12:17.920 --> 0:12:20.520 large weight in the market, and if Tech is not okay, 0:12:20.600 --> 0:12:21.720 the entire market goes. 0:12:21.840 --> 0:12:24.120 We're all tied to these seven stocks, and in particular, 0:12:24.120 --> 0:12:25.280 we're all tied to video. 0:12:25.480 --> 0:12:27.920 It looks like the story is China is not as 0:12:27.960 --> 0:12:29.320 far back as what people thought. 0:12:29.360 --> 0:12:30.800 They are much closer now. 0:12:30.880 --> 0:12:33.840 China looks like a very viable competitor, and a competitor 0:12:33.840 --> 0:12:35.560 that perhaps might catch the eyre of. 0:12:35.559 --> 0:12:36.520 The Trump administration. 0:12:36.720 --> 0:12:37.400 This has got to. 0:12:37.360 --> 0:12:42.360 Be a core concern of not just the Trump administration, 0:12:42.400 --> 0:12:45.160 but all the tech universe that has moved into the 0:12:45.200 --> 0:12:45.600 West Way. 0:12:45.760 --> 0:12:49.600 The US China trade wars and the restrictions on chips 0:12:49.600 --> 0:12:52.719 and things like that could get even more heated. 0:12:52.520 --> 0:12:54.640 Which will bring more volatility to this market. 0:12:56.240 --> 0:12:58.960 That was what some of Bloomberg Television's guests had to 0:12:59.000 --> 0:13:03.040 say about the deep impact earlier. Today, this is Nvidia 0:13:03.160 --> 0:13:07.240 is hitting session lows now. Deepsek reminds us that technology 0:13:07.320 --> 0:13:10.480 lies at the heart of geopolitical tensions between the US 0:13:10.520 --> 0:13:14.840 and China, with disputes over AI TikTok and cybersecurity and 0:13:14.880 --> 0:13:17.160 for more, and that we're joined by Teresa Payton. She 0:13:17.280 --> 0:13:19.760 is the CEO of fort E Lee's Solutions and the 0:13:19.760 --> 0:13:23.280 former White House Chief Performation Officer during the George W. 0:13:23.400 --> 0:13:27.480 Bush administration. Teresa, thank you for joining us. Obviously this 0:13:27.640 --> 0:13:31.720 is going to cause a significant impact m Washington. How 0:13:31.760 --> 0:13:34.680 do you expect policy makers there we have taken a 0:13:34.880 --> 0:13:37.400 hawkish approach towards China, will. 0:13:37.200 --> 0:13:41.520 Respond, Yeah, I mean this tech route and massive market 0:13:41.559 --> 0:13:44.560 sell off, it's a wake up call. It is for me, 0:13:45.120 --> 0:13:48.280 and it underscores the urgent need for robust policies. We 0:13:48.400 --> 0:13:52.440 have got to safeguard the United States leadership in AI. 0:13:52.600 --> 0:13:55.960 I certainly don't want China setting the gold standard for 0:13:56.000 --> 0:14:01.000 the rest of the world for privacy, safety, secure resiliency 0:14:01.559 --> 0:14:03.600 and ethics around AI. 0:14:04.240 --> 0:14:05.360 I believe this is. 0:14:05.360 --> 0:14:10.160 Why you saw sort of the Technology Advisors Elon Musk included, 0:14:11.200 --> 0:14:17.480 and President Trump last week speaking very boldly about AI 0:14:17.559 --> 0:14:21.360 and the United States leadership position that is necessary. I mean, 0:14:21.720 --> 0:14:24.680 we didn't even have a full week before the Trump 0:14:24.680 --> 0:14:28.040 administration passed an executive order on AI, so this is 0:14:28.160 --> 0:14:28.840 very serious. 0:14:29.880 --> 0:14:33.320 They've set AI as a priority across Washington and including 0:14:33.400 --> 0:14:37.480 from the White House itself, and yet previous measures, including 0:14:37.600 --> 0:14:41.880 export controls by the Biden administration, don't appear to have worked. 0:14:41.920 --> 0:14:43.440 So what is the. 0:14:43.320 --> 0:14:46.600 Formula to achieve what Donald Trump is set as a 0:14:46.600 --> 0:14:49.160 policy goal, that is US leadership in AI. 0:14:50.320 --> 0:14:51.280 No, you're right about that. 0:14:51.320 --> 0:14:53.920 This sort of this open source method in which deep 0:14:53.960 --> 0:14:58.400 seek was released, it's now the top downloaded free app 0:14:58.880 --> 0:15:02.560 at the Apple Store definitely got around sort of the 0:15:02.600 --> 0:15:04.680 different protocols that were put in place by the US 0:15:04.760 --> 0:15:08.080 government over the last couple of years. And so it 0:15:08.240 --> 0:15:12.600 shows that although regulatory frameworks are incredibly helpful to set 0:15:12.960 --> 0:15:16.640 the right standard in the right tone, that sometimes regulatory 0:15:16.640 --> 0:15:19.280 frameworks don't service and they get in the way of innovation. 0:15:20.280 --> 0:15:23.200 So we're going to have to figure out how do 0:15:23.280 --> 0:15:26.320 we go back to the drawing board and look at 0:15:26.360 --> 0:15:30.520 how we engineer AI. In the United States today and 0:15:31.720 --> 0:15:35.880 ask ourselves what can we be doing differently? And it's 0:15:35.920 --> 0:15:39.440 a race right now, It's a long race. Right now 0:15:39.600 --> 0:15:42.560 we kind of lost one of the heats, so but 0:15:42.640 --> 0:15:45.360 you know, there's time to make up the difference. 0:15:45.680 --> 0:15:48.960 With your security expertise. We bring you this latest in 0:15:49.040 --> 0:15:52.280 terms of breaking news that deep seek says it's subject 0:15:52.320 --> 0:15:54.840 to a large scale malicious attack. If you've tried to 0:15:54.840 --> 0:15:59.120 download it, you can't. Currently it seems to be offline 0:15:59.160 --> 0:16:01.960 in many ways. How do you think this will be 0:16:02.040 --> 0:16:05.800 able to be responded to? Is it right that we 0:16:05.840 --> 0:16:07.600 can't access it here in the United States? 0:16:08.680 --> 0:16:10.800 Yeah, I think this is a real challenge. I saw 0:16:10.800 --> 0:16:13.840 that they were this is late breaking news that they 0:16:13.880 --> 0:16:19.920 were having some resiliency and recoverability issues, ostensibly because there 0:16:20.000 --> 0:16:24.000 was such heavy volume in traffic and interest. But if 0:16:24.000 --> 0:16:26.920 they are under some type of a cyber attack, I 0:16:26.960 --> 0:16:30.080 think that should give pause to everybody who's thinking about 0:16:30.120 --> 0:16:34.320 testing out the app, perhaps on company data, perhaps using 0:16:34.360 --> 0:16:40.040 your own personal data. This is an untested app. I 0:16:40.160 --> 0:16:45.360 would caution people not to put too much proprietary company 0:16:45.400 --> 0:16:49.600 information into it until there's been an opportunity to actually 0:16:49.680 --> 0:16:52.880 do something called ethical hacking or red teaming, pen testing 0:16:53.440 --> 0:16:56.160 of the actual app, and learning more about how is 0:16:56.200 --> 0:16:59.240 your data treated, where is it stored, how do the 0:16:59.280 --> 0:17:02.680 algorithms work. It is open source, but things do still 0:17:02.720 --> 0:17:05.040 need to be put through their paces and theresa. 0:17:05.320 --> 0:17:07.920 This comes at a time when an app that everyone 0:17:08.000 --> 0:17:11.600 in the United States from a bipartisan perspective thought was 0:17:11.600 --> 0:17:15.160 a national security threat that of TikTok doesn't get banned 0:17:15.200 --> 0:17:18.480 for US users. We see this ongoing post to drive 0:17:18.560 --> 0:17:21.240 in AI but actually perhaps pull away from some of 0:17:21.280 --> 0:17:23.560 the commitments when it comes to ethics, when it comes 0:17:23.600 --> 0:17:24.720 to safety. 0:17:25.320 --> 0:17:27.560 Is that something that you're reading between the lines right now? 0:17:28.600 --> 0:17:34.120 Yeah, I'm definitely watching this very closely and my interest 0:17:34.200 --> 0:17:39.120 is peaked here for starters. Where is Siphius on this right? 0:17:39.200 --> 0:17:42.880 So we're we've been upset about TikTok. We've demanded TikTok 0:17:42.960 --> 0:17:45.280 do a lot of architecture changes. They did a lot 0:17:45.280 --> 0:17:49.080 of architecture changes with Oracle. The United States still said 0:17:49.080 --> 0:17:52.760 that there was some concerns there. Where does Cyphias stand 0:17:52.920 --> 0:17:55.200 on this particular app. How do they feel about it? 0:17:55.240 --> 0:18:00.160 For both US citizen data as well as US corporation data. 0:18:00.720 --> 0:18:04.359 The origin story is China. It is headquartered in China, 0:18:05.240 --> 0:18:08.720 and so again a lot of questions remain. Are we 0:18:08.840 --> 0:18:11.480 going to allow this app to be downloaded in the 0:18:11.560 --> 0:18:15.159 United States? And is it allowed to be downloaded just 0:18:15.280 --> 0:18:19.639 because Siphius hasn't had the time to analyze it and 0:18:19.680 --> 0:18:20.760 give a ruling on it. 0:18:22.880 --> 0:18:25.719 Teresa Payden, CEO of for de Lee's Solutions and the 0:18:25.720 --> 0:18:36.360 former CIO in the Bush White House, thanks for joining us. 0:18:38.800 --> 0:18:41.800 Let's not bring in Bloomberg's Jackie Dabolos to discuss the 0:18:41.840 --> 0:18:45.840 impact of Deep Seek on Open AI and Stargate. Jackie, 0:18:46.000 --> 0:18:49.400 we worked on the coverage of this Stargate project unveiling 0:18:49.560 --> 0:18:52.359 last week at the White House. This is not the 0:18:52.480 --> 0:18:55.960 kind of moment that President Donald Trump and his administration 0:18:56.520 --> 0:19:00.520 wanted to start this week, so soon after that big announcement, 0:19:00.760 --> 0:19:01.800 where do they go from here? 0:19:03.400 --> 0:19:07.240 Well, from here, we really have to assess how powerful 0:19:07.480 --> 0:19:11.119 is Deep Seek? What makes these models so threatening to 0:19:11.160 --> 0:19:14.000 the ones that we are building here in the United States. 0:19:14.440 --> 0:19:16.200 One of the things that we can look to is. 0:19:16.119 --> 0:19:20.600 The fact that this are one model family that is 0:19:21.080 --> 0:19:22.160 really making a splash. 0:19:22.200 --> 0:19:23.240 They're open weights. 0:19:23.359 --> 0:19:26.160 What makes us so surprising is the fact that many 0:19:26.240 --> 0:19:29.600 open weight models have actually kind of been lagging their 0:19:29.640 --> 0:19:34.560 closed source peers like open AI, like and anthropics, and 0:19:34.600 --> 0:19:37.879 so what the United States government really now has to 0:19:37.920 --> 0:19:40.960 assess is what are we missing here? How are they 0:19:41.080 --> 0:19:45.040 able to get past these export controls? 0:19:45.560 --> 0:19:46.760 It might not even need them. 0:19:46.760 --> 0:19:50.600 It seems like these models are really efficiently run because 0:19:50.600 --> 0:19:53.480 they're far more computational in nature than what we might 0:19:53.520 --> 0:19:56.120 see from an open AI. So this kind of turns 0:19:56.200 --> 0:19:59.679 the whole premise that chips were really going to be 0:20:00.119 --> 0:20:03.200 the key here to getting ahead in the AI race. 0:20:03.440 --> 0:20:05.119 This really kind of flips. 0:20:04.720 --> 0:20:09.280 That necessity is a mother of all invention. They had 0:20:09.320 --> 0:20:11.800 to do it. They had lack of access to the 0:20:11.800 --> 0:20:14.919 most sophisticated chips coming from Nvidia. And we understand from 0:20:14.960 --> 0:20:20.360 Dooming Intelligence that this is about novel mixture of experts architecture, right. 0:20:20.480 --> 0:20:22.080 This is how they lower the cost, This is how 0:20:22.080 --> 0:20:25.320 they bring computational power is a different form of running 0:20:25.359 --> 0:20:27.560 the models, but others have been bringing it on too. 0:20:27.600 --> 0:20:28.680 Can you tell us a little. 0:20:28.520 --> 0:20:31.200 Bit about how they're managing can compete and when out 0:20:31.240 --> 0:20:33.840 when it comes to mass challenges, when it comes to 0:20:33.920 --> 0:20:37.960 reasoning versus the lightest out of Claude or indeed open Ai. 0:20:39.119 --> 0:20:40.480 You're absolutely rad, Caroline. 0:20:40.480 --> 0:20:43.520 It's no surprise that the computational framework here really has. 0:20:43.400 --> 0:20:44.160 Gone a long way. 0:20:44.240 --> 0:20:47.919 It's a startup that has its roots in a quantitative 0:20:47.960 --> 0:20:50.320 hedge fund, so this is really their bread and butter, 0:20:50.320 --> 0:20:52.120 at least it used to be. But when we look 0:20:52.119 --> 0:20:55.960 at what the numbers actually show their performance across several benchmarks, 0:20:55.960 --> 0:20:58.720 the ones that stand out to me are in mas encoding. 0:20:59.200 --> 0:21:02.560 These are the especially when you think about how widely 0:21:02.640 --> 0:21:04.679 used it is in the coding community. The fact that 0:21:04.720 --> 0:21:09.840 it outperformed Claude and open Ai on some benchmarks. Deep 0:21:09.880 --> 0:21:13.639 Seek here really is a player that is contending with 0:21:13.760 --> 0:21:14.840 these incumbents. 0:21:15.320 --> 0:21:17.080 But you know kind of past that. 0:21:17.240 --> 0:21:20.560 You're right, there is a difference here when it comes 0:21:20.600 --> 0:21:23.920