1 00:00:00,960 --> 00:00:04,717 Bring in show music, please. 2 00:00:04,800 --> 00:00:09,357 >> This is Squawk Pod and I'm CNBC producer Cameron Costa. [music] On today's 3 00:00:09,440 --> 00:00:10,877 episode, 4 00:00:10,960 --> 00:00:16,557 OpenAI co-founder and president [music] Greg Brockman on this critical moment in 5 00:00:16,640 --> 00:00:19,437 AI. >> Compute is really becoming the new oil, 6 00:00:19,520 --> 00:00:22,077 the new limited resource of the AI age. 7 00:00:22,160 --> 00:00:24,557 >> And Brockman's call of warning about 8 00:00:24,640 --> 00:00:27,197 what's coming in cyber security. 9 00:00:27,280 --> 00:00:31,277 >> Now is the time to act. Now is the time to say it will no longer be sufficient 10 00:00:31,360 --> 00:00:33,327 and this has to be prioritized I think 11 00:00:33,410 --> 00:00:35,037 [music] as a top level priority in every 12 00:00:35,120 --> 00:00:38,317 single company and we're seeing it. 13 00:00:38,400 --> 00:00:41,277 >> Then what's coming in the skies? Drone 14 00:00:41,360 --> 00:00:43,837 delivery company Zipline has just inked 15 00:00:43,920 --> 00:00:46,877 a deal with Uber to reach a million 16 00:00:46,960 --> 00:00:49,277 drone deliveries a day. CEO [music] 17 00:00:49,360 --> 00:00:54,287 Keller Clifton joins us. When you make this kind of delivery way faster, way 18 00:00:54,370 --> 00:00:57,757 [music] less expensive, less traffic, less pollution for our neighborhoods, 19 00:00:57,840 --> 00:01:01,677 people significantly change their buying behavior. 20 00:01:01,760 --> 00:01:04,157 >> Plus, President Trump's foreign policy 21 00:01:04,240 --> 00:01:06,077 and an online debate about billionaire 22 00:01:06,160 --> 00:01:11,437 [music] taxes gets heated. Politicians, investors, and entrepreneurs are 23 00:01:11,520 --> 00:01:14,637 weighing in. >> There was a whole bunch of uh FUS 24 00:01:14,720 --> 00:01:17,117 involved in this uh conversation. 25 00:01:17,200 --> 00:01:19,757 >> I reached out to Mark Cuban, actually. 26 00:01:19,840 --> 00:01:23,117 It's Monday, August 17th, 2026 [music] 27 00:01:23,200 --> 00:01:26,077 and Squawkpod begins right now. 28 00:01:26,160 --> 00:01:31,357 >> Stand back by in 3 2 1 c please. 29 00:01:31,440 --> 00:01:35,517 >> Good morning everybody. Welcome to Squawkbox right here on CNBC. We are 30 00:01:35,600 --> 00:01:39,677 live from the NASDAQ market site in Times Square. I'm Becky Quick along with 31 00:01:39,760 --> 00:01:42,317 Andrew Rosor and Joe is off today. 32 00:01:42,400 --> 00:01:45,837 Bergkshire Hathaway bulking up on Alphabet shares. According to 33 00:01:45,920 --> 00:01:50,557 Bergkshire's latest 13F filing, it bought $17 billion worth of Alphabet 34 00:01:50,640 --> 00:01:54,957 stock in the second quarter. That made Google the the Google parent the third 35 00:01:55,040 --> 00:01:59,197 largest holding in Bergkshire's equity portfolio, which is worth more than $ 36 36 00:01:59,280 --> 00:02:01,677 billion that holding. Bergkshire's 37 00:02:01,760 --> 00:02:03,837 biggest stakes are in Apple and American 38 00:02:03,920 --> 00:02:07,677 Express. About 60% of the more than 48 39 00:02:07,760 --> 00:02:11,117 million Alphabet shares that Bergkshire added came in a private placement sale 40 00:02:11,200 --> 00:02:14,237 that the companies announced back in early June, but it was buying in the 41 00:02:14,320 --> 00:02:16,957 open market as well. And in other moves, 42 00:02:17,040 --> 00:02:21,437 Bergkshire added to its position in Delta Airlines and and in Macy's. It 43 00:02:21,520 --> 00:02:25,357 also added to its stake in the homebuilder Lar. And it bought shares of 44 00:02:25,440 --> 00:02:27,437 Bergkshire, too. >> It's going to be interesting to see. I 45 00:02:27,520 --> 00:02:31,597 mean, Google uh for better or worse has not been as loved as so many of the 46 00:02:31,680 --> 00:02:36,317 other big tech companies. And the question is from an AI perspective, will 47 00:02:36,400 --> 00:02:38,157 they be able to make it work? >> Yeah. 48 00:02:38,240 --> 00:02:42,957 >> And I actually think, you know, they may not be the most advanced at the mo at 49 00:02:43,040 --> 00:02:47,757 this point like Gemini is not >> where Anthropic is or Open AAI is, but 50 00:02:47,840 --> 00:02:51,437 given the surface area of just what they capture, 51 00:02:51,520 --> 00:02:54,317 it feels like the use case. I don't I 52 00:02:54,400 --> 00:02:56,877 find myself more and more even just on 53 00:02:56,960 --> 00:03:01,437 Google, I'm now I'm now >> putting in prompts just straight into 54 00:03:01,520 --> 00:03:05,117 Google. I do too. >> Uh almost by habit at this point, 55 00:03:05,200 --> 00:03:08,317 >> right? Because we've all used Google for so long. And in fact, that was Warren 56 00:03:08,400 --> 00:03:10,717 Buffett's point when we spoke to him last month. 57 00:03:10,800 --> 00:03:14,797 >> Uh and talked a little bit about this push into Alphabet, which by the way, he 58 00:03:14,880 --> 00:03:19,037 was behind that move. Um he said, "Look, they're going to be a winner." He just 59 00:03:19,120 --> 00:03:22,477 thinks that they that that's his expectation at least that they will be a 60 00:03:22,560 --> 00:03:25,037 winner. He doesn't know how it's all going to play out, but they have the 61 00:03:25,120 --> 00:03:28,797 money, they have the capability, and the smarts to do it. And that was his thesis 62 00:03:28,880 --> 00:03:31,837 and why they are plowing so much money into And by the way, Greg Abel's on 63 00:03:31,920 --> 00:03:35,597 board with that. They've talked about all of it, but it was initiated from Buffett. 64 00:03:35,680 --> 00:03:40,077 >> We will see. Uh Apple, it's funny because on tech on the tech side, he's 65 00:03:40,160 --> 00:03:44,877 had this sort of mixed he's had an amazing record with Apple. Y >> killer. 66 00:03:44,960 --> 00:03:49,517 >> IBM, if you remember, was one of his picks years ago. That was a a miss. 67 00:03:49,600 --> 00:03:52,957 >> Yeah. Under Jenny Remedi when she >> So, it's going to be 68 00:03:53,040 --> 00:03:55,357 >> We'll see. We'll see. >> Yeah. And and by the way, he's always 69 00:03:55,440 --> 00:03:58,077 said that he's not really a tech investor, but when I pushed him on that 70 00:03:58,160 --> 00:04:01,997 last month, he was like, "Yeah, I've been buying technology stocks for 40 71 00:04:02,080 --> 00:04:06,157 years or so." So, you know, he his reason behind Apple was that it was a 72 00:04:06,240 --> 00:04:09,037 consumer stock and that he understand the consumer aspects to it. I asked him 73 00:04:09,120 --> 00:04:12,157 if that was the Google point. And he said, "Look, he just really thinks that 74 00:04:12,240 --> 00:04:16,077 they are going to to end up being one of the companies that that that thrives in 75 00:04:16,160 --> 00:04:18,717 what what's going on with it with AI." 76 00:04:18,800 --> 00:04:22,237 >> We will see. We will see. 77 00:04:22,320 --> 00:04:26,557 President Trump ordering the US military to scale back annual exercises with 78 00:04:26,640 --> 00:04:30,957 South Korea. In a social media post, President called what he or cited I 79 00:04:31,040 --> 00:04:34,877 should say what he called a very good relationship with North Korea's Kim 80 00:04:34,960 --> 00:04:37,677 Jong-un and said the exercises in Hid's 81 00:04:37,760 --> 00:04:40,237 words quote send a signal that is 82 00:04:40,320 --> 00:04:44,717 totally inappropriate and hostile to a country that as long as Donald J. Trump 83 00:04:44,800 --> 00:04:49,597 has been president has been unthreatening and respectful. President 84 00:04:49,680 --> 00:04:54,557 went on to say, quote, while somewhat unrelated, question mark, I recently 85 00:04:54,640 --> 00:04:58,157 asked the president of South Korea if they would like to join us in the 86 00:04:58,240 --> 00:05:00,397 denuclearization of the Islamic Republic 87 00:05:00,480 --> 00:05:04,397 of Iran, and they said, "No thanks." So 88 00:05:04,480 --> 00:05:08,077 maybe we can understand some of where some of this is coming from. Uh, the 89 00:05:08,160 --> 00:05:12,077 drills involve 18,000 South Korean soldiers and are designed to increase 90 00:05:12,160 --> 00:05:14,477 readiness against North Korean threats. 91 00:05:14,560 --> 00:05:16,237 Back in May, Defense Secretary Pete 92 00:05:16,320 --> 00:05:19,277 Hgsth praised South Korea for its sec 93 00:05:19,360 --> 00:05:23,677 leadership on the Korean Peninsula. So, 94 00:05:23,760 --> 00:05:25,437 bit of a head scratcher this morning. 95 00:05:25,520 --> 00:05:28,157 >> Yeah, look, President Trump likes to 96 00:05:28,240 --> 00:05:33,037 kind of combine politics, geopolitics, what he's doing economically and beyond, 97 00:05:33,120 --> 00:05:35,677 and there could be some pressure coming from that because they're also kind of 98 00:05:35,760 --> 00:05:39,917 pushing Korea, South Korea, South Korea to come up with some of the investments 99 00:05:40,000 --> 00:05:41,437 that they had been talking about on the 100 00:05:41,520 --> 00:05:44,637 economic front, too. Um 101 00:05:44,720 --> 00:05:49,437 but yeah um it is an interesting and change in perspective on something 102 00:05:49,520 --> 00:05:52,397 that's been planned for a very long time and I think would be hard because right 103 00:05:52,480 --> 00:05:55,517 now they're in the process of going through these drills. It would be hard 104 00:05:55,600 --> 00:05:57,117 to take much away from that at this point. 105 00:05:57,200 --> 00:06:00,397 >> Well, I think there's a couple of things. The sovereignity of South Korea, 106 00:06:00,480 --> 00:06:03,357 which you'd think they're our ally, we would want to protect. 107 00:06:03,440 --> 00:06:06,317 >> Well, we fought the Korean War there. 108 00:06:06,400 --> 00:06:11,037 >> Uh North Korea we have typically put in the same category as Putin 109 00:06:11,120 --> 00:06:13,117 >> of evil though. the Bush term. 110 00:06:13,200 --> 00:06:17,437 >> The president, this president has not uh always put Putin in that category. So 111 00:06:17,520 --> 00:06:22,157 that's why I said just I put it in the headscratching category. That's a light 112 00:06:22,240 --> 00:06:25,037 way to put it. I think a lot of people would say there's something very a miss 113 00:06:25,120 --> 00:06:30,317 here. >> George W. Bush put >> and North Korea and 114 00:06:30,400 --> 00:06:35,837 >> yes, >> Iran in that axis of evil. And we shall see. 115 00:06:35,920 --> 00:06:39,277 >> All right, let's get to the latest on the Iran war. Speaking of the Wall 116 00:06:39,360 --> 00:06:42,957 Street Journal says that Iran has spent the past two months getting ready for a 117 00:06:43,040 --> 00:06:45,597 bigger fight and into a stance to raise 118 00:06:45,680 --> 00:06:50,477 US cost if the conflict goes on. A [snorts] report says that since a 119 00:06:50,560 --> 00:06:54,717 memorandum of understanding was signed with the United States in June, Iran's 120 00:06:54,800 --> 00:06:58,477 hardline leaders have ramped up missile and drone production and shifted more 121 00:06:58,560 --> 00:07:00,237 power to the country's revolutionary 122 00:07:00,320 --> 00:07:05,517 guard corps. Separately, Iran's deputy foreign minister doubling down on the 123 00:07:05,600 --> 00:07:09,517 country's position that that the straight of Hormuz is under its control. 124 00:07:09,600 --> 00:07:13,117 That followed comments from President Trump on Friday that he may declare the 125 00:07:13,200 --> 00:07:15,437 straight US property. Here's a little 126 00:07:15,520 --> 00:07:17,037 bit of what the president said on higher 127 00:07:17,120 --> 00:07:20,557 gas prices stemming from that war. 128 00:07:20,640 --> 00:07:23,597 >> So, for you to pay a tiny little bit 129 00:07:23,680 --> 00:07:27,117 more for your gasoline, just remember 130 00:07:27,200 --> 00:07:30,957 you're doing it so that a very evil country cannot have a country really. 131 00:07:31,040 --> 00:07:35,517 It's the number one state sponsor of terror in the world. We don't want to 132 00:07:35,600 --> 00:07:39,597 have them have a nuclear weapon. So remember that when you have to pay a 133 00:07:39,680 --> 00:07:43,197 little bit more, you're at $4. It's okay. 134 00:07:43,280 --> 00:07:45,517 >> According to AAA, a gallon of regular 135 00:07:45,600 --> 00:07:48,877 currently averages $46 across the United 136 00:07:48,960 --> 00:07:51,597 States. 137 00:07:51,680 --> 00:07:53,917 >> California Congressman Roana earning 138 00:07:54,000 --> 00:07:56,397 some new criticism for support of a 139 00:07:56,480 --> 00:07:58,637 possible billionaire tax in California. 140 00:07:58,720 --> 00:08:03,357 Over the weekend on X, Kana says the following. The California Democratic 141 00:08:03,440 --> 00:08:05,357 Party and the California Labor movement 142 00:08:05,440 --> 00:08:07,757 just stood with Bernie Sanders and me in 143 00:08:07,840 --> 00:08:12,797 supporting a 5% wealth tax on 250 California billionaires. California 144 00:08:12,880 --> 00:08:16,397 voters want a Democratic party that will stand up for the working class over the 145 00:08:16,480 --> 00:08:19,197 billionaire class. That prompted all 146 00:08:19,280 --> 00:08:22,957 sorts of commotion uh online and this 147 00:08:23,040 --> 00:08:25,757 reply from Mark Cuban who said founders 148 00:08:25,840 --> 00:08:28,477 of even richly valued startups don't see 149 00:08:28,560 --> 00:08:31,037 much personal wealth following flowing 150 00:08:31,120 --> 00:08:35,997 their way. Cuban asked Kana how he planned to tax that wealth even if it 151 00:08:36,080 --> 00:08:37,997 looks huge on paper and said if the 152 00:08:38,080 --> 00:08:41,597 billionaire tax passes uh he would not 153 00:08:41,680 --> 00:08:44,637 bring new companies to California and in 154 00:08:44,720 --> 00:08:48,797 fact it'd be a prerequisite uh for them not to be based in California if they 155 00:08:48,880 --> 00:08:53,437 were going to invest in a company. So then Connor replies and says that 156 00:08:53,520 --> 00:08:55,357 founders, and this then gets everybody 157 00:08:55,440 --> 00:08:58,157 even crazier, says founders without much 158 00:08:58,240 --> 00:09:01,357 liquid wealth could pledge their shares 159 00:09:01,440 --> 00:09:03,597 with a long-standing loan from the 160 00:09:03,680 --> 00:09:07,437 government to pay their taxes. to that. 161 00:09:07,520 --> 00:09:11,437 Then hedge fund billionaire Bill Aman jumps into this whole game and says that 162 00:09:11,520 --> 00:09:13,597 under Conor's plan, quote, "The penalty 163 00:09:13,680 --> 00:09:17,117 for a failed startup is insolveny and 164 00:09:17,200 --> 00:09:19,917 the government garnishes your wages for 165 00:09:20,000 --> 00:09:22,717 life." Palmer Lucky also uh jumping into 166 00:09:22,800 --> 00:09:28,157 the fray. There was a whole bunch of uh FUS involved in this conversation. I 167 00:09:28,240 --> 00:09:31,757 reached out to Mark Cuban actually >> and Mark and I had emailed yesterday 168 00:09:31,840 --> 00:09:32,877 about this 169 00:09:32,960 --> 00:09:36,317 >> and it's kind of an unbelievable 170 00:09:36,400 --> 00:09:39,917 uh scenario. I'm surprised. 171 00:09:40,000 --> 00:09:43,037 I'm surprised actually just at how far 172 00:09:43,120 --> 00:09:44,877 Row, who we've had on the show for so 173 00:09:44,960 --> 00:09:48,317 long, has sort of turned the tables and 174 00:09:48,400 --> 00:09:52,317 in particular given that he does represent Silicon Valley 175 00:09:52,400 --> 00:09:55,197 >> and does a lot of trading >> um and does trading, but I'm not even 176 00:09:55,280 --> 00:09:57,197 putting the trading part into the story. 177 00:09:57,280 --> 00:09:59,677 >> You would think he would understand a little bit more about how this would 178 00:09:59,760 --> 00:10:02,557 work out. I would just think that he and 179 00:10:02,640 --> 00:10:04,797 this is the political calculus that I I 180 00:10:04,880 --> 00:10:06,957 feel like I don't understand given that 181 00:10:07,040 --> 00:10:09,997 I would have thought his constituency 182 00:10:10,080 --> 00:10:12,477 would be against this broadly given that 183 00:10:12,560 --> 00:10:14,797 he represents literally Silicon Valley 184 00:10:14,880 --> 00:10:17,357 that is his constituency. He must 185 00:10:17,440 --> 00:10:19,917 believe that there's enough people in 186 00:10:20,000 --> 00:10:21,997 his district 187 00:10:22,080 --> 00:10:24,557 um who work at tech companies but are 188 00:10:24,640 --> 00:10:27,517 not billionaires themselves who somehow 189 00:10:27,600 --> 00:10:30,077 hold either so much animosity towards 190 00:10:30,160 --> 00:10:35,597 their their overlords, if you will, or uh have a very different view of all of 191 00:10:35,680 --> 00:10:38,317 this. I mean, that's the piece of it that I can't understand. 192 00:10:38,400 --> 00:10:41,517 >> He's gotten backing from people like Lisa Sue and Jensen Wong who say that 193 00:10:41,600 --> 00:10:44,557 they'd go along with this if it were to happen. So, he always uses them as kind 194 00:10:44,640 --> 00:10:47,517 of the cover for what he's happening with this. It does make me wonder if Ro 195 00:10:47,600 --> 00:10:50,957 is looking for a higher office beyond just >> Sure. But what I don't understand is if 196 00:10:51,040 --> 00:10:56,157 he doesn't get the higher office, does he get to stay in the role? Meaning I I 197 00:10:56,240 --> 00:10:59,757 would think you would be looking both politically at some kind of higher I I 198 00:10:59,840 --> 00:11:01,917 no doubt that he may have higher and 199 00:11:02,000 --> 00:11:03,357 bigger ambitions than than what he's 200 00:11:03,440 --> 00:11:06,797 doing today. But if he if he doesn't get 201 00:11:06,880 --> 00:11:11,037 that, is there some other maybe the fallback is his his his family's quite 202 00:11:11,120 --> 00:11:15,437 wealthy. But but but I would think he would want to at least stay in Congress. 203 00:11:15,520 --> 00:11:17,837 And what I can't figure out is this 204 00:11:17,920 --> 00:11:20,557 audience that he technically represents. 205 00:11:20,640 --> 00:11:24,477 >> Beating up billionaires is is something that's usually pretty good from a 206 00:11:24,560 --> 00:11:27,357 political perspective because there aren't enough of them to say that 207 00:11:27,440 --> 00:11:31,037 they're going to >> I but I it well I was going to say few 208 00:11:31,120 --> 00:11:33,437 places in the country you would have thought New York as sort of the 209 00:11:33,520 --> 00:11:37,757 capitalism would be one of those places but especially Silicon Valley you would 210 00:11:37,840 --> 00:11:40,877 I would have thought uh would would have a very different view of this and I 211 00:11:40,960 --> 00:11:44,477 think so many of the vocal ones obviously on X and others do. Um, 212 00:11:44,560 --> 00:11:48,157 >> look, I think it's it's something that people who have a lot of money to give 213 00:11:48,240 --> 00:11:49,997 towards politics may not be a fan of, 214 00:11:50,080 --> 00:11:53,837 but I a populist argument generally wins 215 00:11:53,920 --> 00:11:59,357 when you're looking at just getting votes >> and that but that that's to me the qu 216 00:11:59,440 --> 00:12:02,077 the thing that I'm trying to figure out is if you're an entrep because Silicon 217 00:12:02,160 --> 00:12:06,317 Valley is I I think so maybe they're not maybe it's not a group of entrepreneurs 218 00:12:06,400 --> 00:12:09,677 people who want to be entrepreneurs I would have thought that they would have 219 00:12:09,760 --> 00:12:14,237 had such a different perspective on this that politically it would be difficult, 220 00:12:14,320 --> 00:12:18,957 but clearly he's made a different calculus. 221 00:12:19,040 --> 00:12:23,197 >> Cheese will be next. >> Coming up on SquawkPod, co-founder and 222 00:12:23,280 --> 00:12:25,317 president of OpenAI, Greg Brockman. 223 00:12:25,400 --> 00:12:28,077 [music] How one of the builders of AI envisions 224 00:12:28,160 --> 00:12:33,757 our future and safety within it. [music] Where he doesn't want to go, centralized 225 00:12:33,840 --> 00:12:36,637 AI power. >> I think that that world is a very scary 226 00:12:36,720 --> 00:12:39,727 one. The flip side of saying, hey, that we want to distribute much more. We want 227 00:12:39,810 --> 00:12:42,557 [music] to have distributed power and ensure that there isn't this 228 00:12:42,640 --> 00:12:46,397 concentration of power. That to me feels like the the the right outcome, [music] 229 00:12:46,480 --> 00:12:49,757 the right long-term answer. But it also does mean that then you need to think 230 00:12:49,840 --> 00:12:53,357 about the fact that there are misaligned actors out there, right? It's people 231 00:12:53,440 --> 00:12:56,877 don't all want the same thing. And so, how do you ensure that there are 232 00:12:56,960 --> 00:12:58,477 standards across this? 233 00:12:58,560 --> 00:13:05,000 >> Much more from the co-founder of Open AI right after this. 234 00:13:07,840 --> 00:13:10,317 You're listening to SquawkPod from CNBC 235 00:13:10,400 --> 00:13:15,117 today with Becky Quick and Andrew Ross Sorcin. Here's Andrew. 236 00:13:15,200 --> 00:13:18,157 >> Okay, we've got an important interview now. Joining us in an exclusive 237 00:13:18,240 --> 00:13:20,557 conversation, Greg Brockman, OpenAI's 238 00:13:20,640 --> 00:13:23,917 co-founder and CEO. Just published an 239 00:13:24,000 --> 00:13:27,677 essay uh about cyber security uh warning 240 00:13:27,760 --> 00:13:31,757 uh really about the dangers of all this, but really what needs to happen now. 241 00:13:31,840 --> 00:13:33,117 Greg, it's great to see you this 242 00:13:33,200 --> 00:13:35,597 morning. Um there's so many questions 243 00:13:35,680 --> 00:13:40,397 about where we are uh in AI uh which we can get to but I I want to talk about 244 00:13:40,480 --> 00:13:44,157 the cyber security piece because you really speak to this idea of us being in 245 00:13:44,240 --> 00:13:48,477 an inflection point that may be underappreciated. 246 00:13:48,560 --> 00:13:52,957 F >> first of all thank you for having me Andrew and yes we are in a window right 247 00:13:53,040 --> 00:13:54,797 now where we can see a little bit into 248 00:13:54,880 --> 00:13:57,917 the future. I talked to a number of 249 00:13:58,000 --> 00:14:02,477 organizations going to conferences talking about cyber security and 250 00:14:02,560 --> 00:14:06,557 everyone feels that we are in a moment where cyber security fundamentally needs 251 00:14:06,640 --> 00:14:11,517 to change. Every organization needs to uplevel its cyber security practices but 252 00:14:11,600 --> 00:14:14,637 many people don't know where to start and I think that no one is really 253 00:14:14,720 --> 00:14:16,317 talking about this message publicly. 254 00:14:16,400 --> 00:14:19,757 Everyone kind of feels it talks about it in closed rooms but it is important that 255 00:14:19,840 --> 00:14:23,277 we as a country and as a world really 256 00:14:23,360 --> 00:14:27,597 take the moment seriously. So the open eye hugging face incident, it was one 257 00:14:27,680 --> 00:14:29,277 moment we got to see into the future. 258 00:14:29,360 --> 00:14:33,917 You could see how automated attacks are going to become possible that these will 259 00:14:34,000 --> 00:14:38,557 be capabilities that will be in the hands of threat actors in upcoming 260 00:14:38,640 --> 00:14:42,157 months because this technology is being created across the world by many 261 00:14:42,240 --> 00:14:47,197 organizations. But defenders have a moment now where they can change their 262 00:14:47,280 --> 00:14:50,157 practices, where they can uplevel everything that they're doing. And I 263 00:14:50,240 --> 00:14:53,197 think that fundamentally what's going on is that we're learning that human 264 00:14:53,280 --> 00:14:55,437 written code is full of vulnerabilities. 265 00:14:55,520 --> 00:14:57,197 That there is tech debt that's building 266 00:14:57,280 --> 00:15:00,397 up over time and that we can actually 267 00:15:00,480 --> 00:15:04,797 the same tools that are going to be in, you know, potentially used to find those 268 00:15:04,880 --> 00:15:07,997 vulnerabilities and exploit them. They can actually find them first and fix 269 00:15:08,080 --> 00:15:11,677 them. That's why defenders have an upper hand, have an advantage, but need to 270 00:15:11,760 --> 00:15:15,517 lean in. So we are talking about the practices that we're using at OpenAI. We 271 00:15:15,600 --> 00:15:18,317 are very urgently moving to uplevel 272 00:15:18,400 --> 00:15:22,957 everything we do, but also lessons that apply to the general industry and tools 273 00:15:23,040 --> 00:15:24,637 that can help make everyone more secure. 274 00:15:24,720 --> 00:15:27,677 So we believe in the long run that defenders can really win here. 275 00:15:27,760 --> 00:15:30,877 >> Well, so that's the question. Can defenders really win if in fact and and 276 00:15:30,960 --> 00:15:35,197 the Hugging Face um story was, you know, fascinated a lot of people, scared a lot 277 00:15:35,280 --> 00:15:36,957 of people because it fundamentally 278 00:15:37,040 --> 00:15:39,037 showed that at base underneath these 279 00:15:39,120 --> 00:15:42,317 models, if their guard rails are not on, 280 00:15:42,400 --> 00:15:44,717 uh that they will go find, cheat, lie, 281 00:15:44,800 --> 00:15:47,597 do whatever is necessary, uh to break 282 00:15:47,680 --> 00:15:50,637 the code, if you will. 283 00:15:50,720 --> 00:15:54,797 >> I would say that there are two pieces to hugging face. The first one is really 284 00:15:54,880 --> 00:15:58,477 something about how we develop and evaluate these models. And that's 285 00:15:58,560 --> 00:16:01,437 something we've taken extremely seriously. And so we're also talking 286 00:16:01,520 --> 00:16:06,237 today uh in in a later blog post about the different defenses that we have 287 00:16:06,320 --> 00:16:09,757 introduced. The way that we've upleled safety, security, alignment in our 288 00:16:09,840 --> 00:16:13,037 development process. And that is absolutely critical that you want models 289 00:16:13,120 --> 00:16:16,397 that are really aligned even when they're contained and even when you're 290 00:16:16,480 --> 00:16:19,917 you're developing them which is that's a real shift that now the models are 291 00:16:20,000 --> 00:16:23,437 capable enough that it is important for us to be applying all these techniques 292 00:16:23,520 --> 00:16:26,877 that we've been preparing and and really thinking about for the past decade since 293 00:16:26,960 --> 00:16:30,717 we started OpenAI uh that we're bringing it all together as part of one unified 294 00:16:30,800 --> 00:16:32,637 engineering process. But there's the 295 00:16:32,720 --> 00:16:34,157 second piece of how Hugging Face has 296 00:16:34,240 --> 00:16:38,077 showed how AI automation of attacks are 297 00:16:38,160 --> 00:16:41,757 going to be different from human ones because AIs have very different 298 00:16:41,840 --> 00:16:45,357 strengths and weaknesses. And one takeaway is that lots of small 299 00:16:45,440 --> 00:16:49,517 vulnerabilities that humans kind of leave around. Credentials that either 300 00:16:49,600 --> 00:16:52,557 are over scoped that got exposed on the internet that people don't really 301 00:16:52,640 --> 00:16:56,557 remember, the little flaws that you have in your system, the credential you 302 00:16:56,640 --> 00:16:58,237 didn't roll that you've been meaning to. 303 00:16:58,320 --> 00:17:02,797 All of those are harder for human attackers to chain together, but for AI 304 00:17:02,880 --> 00:17:07,037 looking through lots of different pieces able to chain together to an objective 305 00:17:07,120 --> 00:17:10,157 that that is something that is much easier. And so I think that what we're 306 00:17:10,240 --> 00:17:12,477 seeing is that we have had two lacks of 307 00:17:12,560 --> 00:17:15,597 a standard as a cyber security field. 308 00:17:15,680 --> 00:17:19,517 And I think every cyber defender kind of feels that everyone has those I have 309 00:17:19,600 --> 00:17:23,197 this thing in the backlog, but I have so many priorities and so many different 310 00:17:23,280 --> 00:17:27,917 stakeholders to align. And now is the time to act. Now is the time to say it 311 00:17:28,000 --> 00:17:32,157 will no longer be sufficient and this has to be prioritized I think as a top 312 00:17:32,240 --> 00:17:35,917 level priority in every single company and we're seeing it. 313 00:17:36,000 --> 00:17:41,277 >> Let me ask you this though. um you know when Anthropic had had its incident and 314 00:17:41,360 --> 00:17:42,877 and and you had had yours there were 315 00:17:42,960 --> 00:17:48,397 some people who said oh this is a flex this is this is marketing um because 316 00:17:48,480 --> 00:17:52,797 what's effectively telling everybody is they now need to go buy uh either open 317 00:17:52,880 --> 00:17:57,277 AI service or anthropic service to secure themselves what do you tell those 318 00:17:57,360 --> 00:18:02,077 people >> so our responsibility is to really show 319 00:18:02,160 --> 00:18:06,077 the shape of the future that's part of the open and open AI we really believe 320 00:18:06,160 --> 00:18:09,437 this technology technology is something that will uplevel everyone, uplift 321 00:18:09,520 --> 00:18:13,357 everyone, can people can use it to build a better life. But we need to point out 322 00:18:13,440 --> 00:18:15,997 what we see coming and sometimes that's 323 00:18:16,080 --> 00:18:17,597 positive and sometimes that's mitigation 324 00:18:17,680 --> 00:18:20,877 of risks. And in this case, we saw 325 00:18:20,960 --> 00:18:24,317 something in how the system was how how 326 00:18:24,400 --> 00:18:28,397 the system was interacting with the world that surprised us but that we also 327 00:18:28,480 --> 00:18:29,917 felt was very important information to 328 00:18:30,000 --> 00:18:31,837 provide to others. And it is absolutely 329 00:18:31,920 --> 00:18:35,757 true that AI is both changing the sides 330 00:18:35,840 --> 00:18:39,597 of attack and defense. It is something that is reshaping how really the whole 331 00:18:39,680 --> 00:18:43,917 economy runs with cyber security as a specific instance. Uh we we definitely 332 00:18:44,000 --> 00:18:47,437 provide the most advanced cyber capabilities and we try to put those in 333 00:18:47,520 --> 00:18:50,317 the hands of defenders. We have a trusted access program so that we're 334 00:18:50,400 --> 00:18:53,757 able to differentially advantage defenders. But it's not just us. This 335 00:18:53,840 --> 00:18:57,197 technology is being created around the world. many different companies, 336 00:18:57,280 --> 00:19:00,397 openweight models, closed models, lots of different tools that people are 337 00:19:00,480 --> 00:19:03,357 building on top of it. So, it's really an entire community that are building 338 00:19:03,440 --> 00:19:06,797 these tools and we think that together we can make the world more secure. 339 00:19:06,880 --> 00:19:09,437 >> Greg, let me ask you a broader question about what's happening at OpenAI right 340 00:19:09,520 --> 00:19:11,837 now. Uh because there has been a number 341 00:19:11,920 --> 00:19:13,277 of departures that have been very 342 00:19:13,360 --> 00:19:16,797 visible. Uh Fiji Simo at the top uh but 343 00:19:16,880 --> 00:19:19,997 your uh CRO was was effectively left and 344 00:19:20,080 --> 00:19:24,477 then was replaced last week. Bright Brad Lycap who've been at the company for a 345 00:19:24,560 --> 00:19:26,317 very long time been the operating uh 346 00:19:26,400 --> 00:19:29,037 chief for for quite a long time uh left 347 00:19:29,120 --> 00:19:33,677 the the the head of ethics uh after about a year left what is happening 348 00:19:33,760 --> 00:19:37,997 right now and what what should the public take away from this 349 00:19:38,080 --> 00:19:40,797 look I would say that the number one thing that we are doing is we're really 350 00:19:40,880 --> 00:19:43,677 focusing right that we have a strategy a 351 00:19:43,760 --> 00:19:48,717 mission and a set of priorities that we're executing on and throughout this 352 00:19:48,800 --> 00:19:52,557 whole year you've seen the impacts of that that we've really been trying to 353 00:19:52,640 --> 00:19:57,677 figure out how do we bring together our consumer and enterprise business to to 354 00:19:57,760 --> 00:20:01,437 be really unified and you've seen that with things like chat GBT work uh you 355 00:20:01,520 --> 00:20:04,557 can see that we're at this inflection point as a business in July our revenue 356 00:20:04,640 --> 00:20:08,397 grew 20% month over month 32% of that is 357 00:20:08,480 --> 00:20:11,997 you know within within just the enterprise category 32% growth there so 358 00:20:12,080 --> 00:20:15,677 you can really see that we are moving to create the best possible models deliver 359 00:20:15,760 --> 00:20:19,757 value for our customers and to really try to scale systematically to every 360 00:20:19,840 --> 00:20:23,197 Greg just so we're on same page because I think people are looking at this and 361 00:20:23,280 --> 00:20:27,357 they're they're asking questions. Uh the CE the chief revenue officer obviously a 362 00:20:27,440 --> 00:20:30,557 hugely important role. I think people are saying to themselves, you hired 363 00:20:30,640 --> 00:20:34,557 somebody for that uh role, they they left and you replace them. Was that 364 00:20:34,640 --> 00:20:38,077 because they weren't doing it right and so you wanted to replace them? Is that 365 00:20:38,160 --> 00:20:41,997 because they didn't like what was going on and they wanted to leave on their 366 00:20:42,080 --> 00:20:46,397 own? People ask the same question about Brad. They asked the same question about 367 00:20:46,480 --> 00:20:50,317 Fiji, who I know had had health issues but is not returning and is starting her 368 00:20:50,400 --> 00:20:54,157 own thing. I mean, can you can you speak to this? 369 00:20:54,240 --> 00:20:56,957 >> Yes. Look, I I'd say fundamentally that 370 00:20:57,040 --> 00:20:59,597 the way that the organization like we're 371 00:20:59,680 --> 00:21:04,237 a very resilient organization. We've been through if you if you look over the 372 00:21:04,320 --> 00:21:08,397 years that there have been different eras where we have different sets of of 373 00:21:08,480 --> 00:21:12,717 leaders in place. you know, I'm a constant, Sam is a constant, and that I 374 00:21:12,800 --> 00:21:16,317 think that we are stronger because of that resilience and diversity. And I 375 00:21:16,400 --> 00:21:20,557 think what I would say is that my view has always been that I try to work with 376 00:21:20,640 --> 00:21:25,997 with the the best people for for that moment and that um we we have people who 377 00:21:26,080 --> 00:21:29,117 have been here for a very long time laboring behind the scenes that don't 378 00:21:29,200 --> 00:21:32,957 make the headlines as much. And so maybe my my core thing would be I actually 379 00:21:33,040 --> 00:21:35,997 think that the difference between open eye and other organizations is that we 380 00:21:36,080 --> 00:21:39,757 are so much in the spotlight. So every departure gets scrutinized in a way that 381 00:21:39,840 --> 00:21:43,517 it doesn't otherwise. I actually don't think that it's actually that atypical 382 00:21:43,600 --> 00:21:47,437 kind of how how we operate and the the longevity of of people. But I would say 383 00:21:47,520 --> 00:21:50,397 that we have a amazing leadership bench. 384 00:21:50,480 --> 00:21:54,397 I'm super excited about the team that we're working with and that just over 385 00:21:54,480 --> 00:21:59,117 the past year I look at the set of of leaders that I I've had the pleasure of 386 00:21:59,200 --> 00:22:02,237 really being able to elevate and continue to work with and will work with 387 00:22:02,320 --> 00:22:06,477 for the long term. And I feel extremely good about that bench and uh and and 388 00:22:06,560 --> 00:22:08,797 where we're going. >> Um let me ask you a different question 389 00:22:08,880 --> 00:22:13,277 about the economics of AI right now. Uh because obviously it appears that 390 00:22:13,360 --> 00:22:16,877 Anthropics is going to have an IPO later this year. I want to talk to you about 391 00:22:16,960 --> 00:22:18,557 the prospect of you doing such a thing 392 00:22:18,640 --> 00:22:22,477 in 2027. But it appears that the 393 00:22:22,560 --> 00:22:24,397 valuations of so many of these uh 394 00:22:24,480 --> 00:22:26,237 different tech companies are quite 395 00:22:26,320 --> 00:22:29,517 dependent on your own valuations uh and 396 00:22:29,600 --> 00:22:32,717 not just u what they are today but that 397 00:22:32,800 --> 00:22:34,557 they have to continue to rise because 398 00:22:34,640 --> 00:22:39,917 embedded in some of their own uh economics is your very company. And so 399 00:22:40,000 --> 00:22:43,837 I'm curious sort of what you're seeing in your in your own business but also 400 00:22:43,920 --> 00:22:47,197 you know just last week and Becky and I have been talking about it. She had this 401 00:22:47,280 --> 00:22:51,677 great interview with Jensen Wong um and a whole bunch of financial leaders who 402 00:22:51,760 --> 00:22:55,277 are going to be contributing potentially up to 500 billion dollars to support 403 00:22:55,360 --> 00:23:00,237 this this whole world and there's a big question mark about whether these these 404 00:23:00,320 --> 00:23:01,997 chips and these data centers are going 405 00:23:02,080 --> 00:23:05,517 to have the value uh that's needed uh to 406 00:23:05,600 --> 00:23:08,397 make all of this work. 407 00:23:08,480 --> 00:23:13,517 >> So I'd say at a fundamental level what we are doing and you can look at this 408 00:23:13,600 --> 00:23:17,197 across so many different verticals of the economy. You can look at we just 409 00:23:17,280 --> 00:23:19,757 talked about cyber security which I think is an absolutely critical 410 00:23:19,840 --> 00:23:23,757 application of AI and I think we're going to see lots of growth there. Um 411 00:23:23,840 --> 00:23:27,037 software engineering this year has been totally transformed. We're starting to 412 00:23:27,120 --> 00:23:30,557 see every aspect of knowledge work really change but also we're starting to 413 00:23:30,640 --> 00:23:32,637 see scientific discovery. We we 414 00:23:32,720 --> 00:23:35,757 announced a couple weeks ago that our AI 415 00:23:35,840 --> 00:23:40,877 for $2,000 worth of compute was able to solve 10 long-standing math problems uh 416 00:23:40,960 --> 00:23:44,557 that had eluded mathematicians for decades. And think about that level of 417 00:23:44,640 --> 00:23:48,717 capability applied to biology, applied to material science. It's all starting 418 00:23:48,800 --> 00:23:53,037 to happen. And so all of these applications that we're starting to see, 419 00:23:53,120 --> 00:23:56,957 they're all powered by one fundamental resource, which is compute. Compute is 420 00:23:57,040 --> 00:23:58,877 really becoming the new oil, the new 421 00:23:58,960 --> 00:24:01,917 limited resource of the AI age. And our 422 00:24:02,000 --> 00:24:06,957 belief, you need to believe one fundamental thing is that 423 00:24:07,040 --> 00:24:10,717 because AI is really about solving problems, the problems that humans want 424 00:24:10,800 --> 00:24:13,597 AI want want to be solved. You need to 425 00:24:13,680 --> 00:24:18,157 believe that the demand for problems to be solved is unlimited, right? And that 426 00:24:18,240 --> 00:24:21,117 is something I believe and that is something we have seen that we have a 427 00:24:21,200 --> 00:24:23,277 billion users on CIBT. We have two 428 00:24:23,360 --> 00:24:25,757 million enterprise customers as well. 429 00:24:25,840 --> 00:24:29,837 And uh you can just see that that the depth of usage as people have started to 430 00:24:29,920 --> 00:24:31,837 use agents is just going off the charts. 431 00:24:31,920 --> 00:24:35,517 Like every single company uh that that tries these things, you start to realize 432 00:24:35,600 --> 00:24:39,837 that you can move in more ambitious ways than you thought previously. And so I 433 00:24:39,920 --> 00:24:43,517 think that we're at a world where we see what's coming. We see the shape of the 434 00:24:43,600 --> 00:24:47,277 future because that model exponential, that capability curve, that's 435 00:24:47,360 --> 00:24:51,597 continuing. We really have a a roadmap lined up where the capabilities will 436 00:24:51,680 --> 00:24:53,597 continue to increase in significant ways 437 00:24:53,680 --> 00:24:55,437 that are hard to that are hard to kind 438 00:24:55,520 --> 00:24:58,877 of you know sort of explain uh in a in a 439 00:24:58,960 --> 00:25:01,997 succinct way because it is really about 440 00:25:02,080 --> 00:25:06,717 being able to add value across every aspect of human endeavor and be able to 441 00:25:06,800 --> 00:25:09,757 help empower people to do more. And so I think if you believe in human 442 00:25:09,840 --> 00:25:12,477 creativity, human ambition, then 443 00:25:12,560 --> 00:25:17,117 ultimately you you have to believe in the power of of what AI will provide. 444 00:25:17,200 --> 00:25:21,757 Hey, Greg. I I'd like to go back to cyber security if it's okay. Um, I agree 445 00:25:21,840 --> 00:25:25,437 with you. There's huge potential promise for AI. I have a lot of hopes about what 446 00:25:25,520 --> 00:25:27,837 it's going to do for healthcare. But 447 00:25:27,920 --> 00:25:29,917 when we saw the breach that you all 448 00:25:30,000 --> 00:25:33,037 talked about last month, the where the 449 00:25:33,120 --> 00:25:35,917 open AI model broke out of the sandbox 450 00:25:36,000 --> 00:25:40,717 and and broke into hugging uh hugging face. I mean, we didn't find out about 451 00:25:40,800 --> 00:25:44,397 that or you didn't find out about that until after HuggingFace had identified 452 00:25:44,480 --> 00:25:49,277 what had happened. they had contacted the FBI and then Anthropic had to admit 453 00:25:49,360 --> 00:25:53,197 that they had a similar sort of situation. I I just wonder, do you feel 454 00:25:53,280 --> 00:25:56,317 like you have your hands around this situation, your arms around this 455 00:25:56,400 --> 00:25:59,837 situation? You know what happened and you can make sure that something like 456 00:25:59,920 --> 00:26:01,597 that doesn't happen again? That that to 457 00:26:01,680 --> 00:26:06,877 me was a a brand new world and a uhoh, do we have any control over this stuff 458 00:26:06,960 --> 00:26:10,157 moment. >> Yeah. For us, we took this extremely 459 00:26:10,240 --> 00:26:13,037 seriously. So for the past couple weeks 460 00:26:13,120 --> 00:26:15,997 that we've been in in really full 461 00:26:16,080 --> 00:26:20,557 post motor postmodm um take a look at what happened really uplevel every 462 00:26:20,640 --> 00:26:24,317 aspect of safety and security and alignment within how we do these 463 00:26:24,400 --> 00:26:27,677 frontier training runs. So it's a real moment where we realize that we do need 464 00:26:27,760 --> 00:26:31,997 to uplevel and it's not a surprise to us in a fundamental way. We've always 465 00:26:32,080 --> 00:26:36,077 anticipated that this moment would come, but I think that that when it actually 466 00:26:36,160 --> 00:26:40,157 happens, seeing the whole team kick into gear and to be so serious. And so we 467 00:26:40,240 --> 00:26:44,397 actually have delayed our frontier training runs in order to ensure that we 468 00:26:44,480 --> 00:26:48,877 have the right levels of controls and in place. And I think that that is that is 469 00:26:48,960 --> 00:26:53,357 a really important thing is that we as a field are sort of now at a point where 470 00:26:53,440 --> 00:26:58,077 these models are so capable and that it's a dual-edged thing because they're 471 00:26:58,160 --> 00:27:00,877 they're capable. It means that we need to ensure that they have alignment 472 00:27:00,960 --> 00:27:03,837 during training, that they need to be monitored during training in a robust 473 00:27:03,920 --> 00:27:07,117 way. But it also means they're so useful, right? And that is where that 474 00:27:07,200 --> 00:27:10,717 economic value comes from because they're able to solve problems in these 475 00:27:10,800 --> 00:27:15,677 novel ways and that we unethical ways like that's that's part of the problem. 476 00:27:15,760 --> 00:27:19,037 Like this was intentionally a model that was set up to say, okay, if we take off 477 00:27:19,120 --> 00:27:22,477 some of these controls, are you guys not doing that anymore? Not saying like, 478 00:27:22,560 --> 00:27:26,717 okay, go find solve the problem no matter what rules we have to break. We 479 00:27:26,800 --> 00:27:30,397 are we are ensuring that there is alignment during that training process. 480 00:27:30,480 --> 00:27:34,237 Absolutely. That I think that alignment is is a non-negotiable and that is 481 00:27:34,320 --> 00:27:37,837 something we've invested in for many years and really thinking about that as 482 00:27:37,920 --> 00:27:41,837 an integral part. So you can't separate out capability from safety, security, 483 00:27:41,920 --> 00:27:45,037 alignment. These are one thing, one system, one way. How how do you think 484 00:27:45,120 --> 00:27:49,117 about that in the context of of the idea of all these open weight models which 485 00:27:49,200 --> 00:27:53,677 which won't have these controls and you have the whole industry really in the 486 00:27:53,760 --> 00:27:58,637 past couple weeks all of a sudden embracing openw weight models. 487 00:27:58,720 --> 00:28:01,757 I think that open weight models are an important part of the ecosystem because 488 00:28:01,840 --> 00:28:05,197 at a fundamental level you have two directions you can go with AI 489 00:28:05,280 --> 00:28:09,037 development. You could either say we need to centralize all of it, have one 490 00:28:09,120 --> 00:28:13,597 way of one one entity in control, maybe one person in control. That is a 491 00:28:13,680 --> 00:28:17,997 different kind of risk, right? But you could you could then say that you're 492 00:28:18,080 --> 00:28:21,357 able to mitigate this question of, hey, there's all this uncontrollable action 493 00:28:21,440 --> 00:28:25,837 happening. I think that that world is a very scary one. The flip side of saying, 494 00:28:25,920 --> 00:28:29,357 hey, that we want to distribute much more. We want to have distributed power 495 00:28:29,440 --> 00:28:33,197 and ensure that there isn't this concentration of power. That to me feels 496 00:28:33,280 --> 00:28:37,917 like the the the right outcome, the right long-term answer. But it also does 497 00:28:38,000 --> 00:28:40,477 mean that then you need to think about the fact that there [snorts] are 498 00:28:40,560 --> 00:28:44,477 misaligned actors out there, right? It's people don't all want the same thing. 499 00:28:44,560 --> 00:28:47,997 And so how do you ensure that there are standards across this? And I think 500 00:28:48,080 --> 00:28:51,357 that's the important question. You can actually see this with things like 501 00:28:51,440 --> 00:28:55,277 Demisipus talking about the FINRA like organization and really thinking about 502 00:28:55,360 --> 00:28:59,837 how do we as we move to frontier AI govern that technology ensure that as 503 00:28:59,920 --> 00:29:03,357 it's being produced it's safe but also really distribute power and so I think 504 00:29:03,440 --> 00:29:06,157 that there's different levels of capability. I think the fact that the 505 00:29:06,240 --> 00:29:09,677 frontier labs have capability that's peering a little bit into the future 506 00:29:09,760 --> 00:29:12,797 gives us information and I think that the system that we're setting up right 507 00:29:12,880 --> 00:29:16,157 now to me feels like it balances these questions and these questions of control 508 00:29:16,240 --> 00:29:18,877 and distribution of access. Greg, President Trump's going to be meeting 509 00:29:18,960 --> 00:29:20,877 with President Xi Jinping in September. 510 00:29:20,960 --> 00:29:23,677 Could you imagine them agreeing on an 511 00:29:23,760 --> 00:29:28,237 actual framework for all of these models so that we have there's there's a a 512 00:29:28,320 --> 00:29:31,917 global body effectively that regulates these models? 513 00:29:32,000 --> 00:29:36,237 >> I I absolutely think it's possible and I absolutely think that it's important for 514 00:29:36,320 --> 00:29:40,077 us as a world to really lean in together and say that we have shared interest 515 00:29:40,160 --> 00:29:43,357 here. We have common interest here to ensure that the technology is something 516 00:29:43,440 --> 00:29:47,037 that uplifts everyone, empowers everyone and helps create the value that is 517 00:29:47,120 --> 00:29:50,797 possible from it. >> On the open weight models, is there any 518 00:29:50,880 --> 00:29:53,837 worry on your end that this this could 519 00:29:53,920 --> 00:29:57,037 effectively break the economics of of 520 00:29:57,120 --> 00:29:59,277 the frontier lab models? 521 00:29:59,360 --> 00:30:02,077 >> I I think that there's a bit of a of a 522 00:30:02,160 --> 00:30:06,557 misnomer or sort of a confusion in terms of uh what openweight models represent, 523 00:30:06,640 --> 00:30:09,837 right? because openweight models and cheap somehow clustered together for 524 00:30:09,920 --> 00:30:14,317 people. But it's not actually true. If you look at that PTO frontier of of task 525 00:30:14,400 --> 00:30:17,837 performance, so what's the cheapest model for any task that we actually win 526 00:30:17,920 --> 00:30:20,397 there and you know we did an 80% price 527 00:30:20,480 --> 00:30:25,437 cut on our our model Luna and now we saw that the the volume associated with it 528 00:30:25,520 --> 00:30:29,677 went up by more than 5x. And so you can really see that people are choosing 529 00:30:29,760 --> 00:30:32,557 voting with their wallet. And there's really two segments of the market. 530 00:30:32,640 --> 00:30:36,317 There's the more price sensitive smaller models and then there's the frontier 531 00:30:36,400 --> 00:30:39,117 models where people are not price sensitive at all. And so for us it 532 00:30:39,200 --> 00:30:43,437 really just comes down to can we produce the best price performing models. And we 533 00:30:43,520 --> 00:30:46,157 have an advantage there because we have the whole stack and we're able to 534 00:30:46,240 --> 00:30:50,077 integrate everything from the chips data center all the way up to the model of 535 00:30:50,160 --> 00:30:53,597 the architecture uh every single piece to produce those efficient models. 536 00:30:53,680 --> 00:30:58,077 >> Talking about the stack and chips uh one of the big questions is what is the 537 00:30:58,160 --> 00:31:02,317 depreciation schedule of these chips? Uh Jensen is now arguing that these chips 538 00:31:02,400 --> 00:31:07,037 themselves uh and the data centers are an asset class. Uh we were comparing 539 00:31:07,120 --> 00:31:10,717 this earlier to cars. So there's always been a market for new cars and there's 540 00:31:10,800 --> 00:31:13,037 also been a market often times for used 541 00:31:13,120 --> 00:31:15,917 cars. Uh when there's u you know 542 00:31:16,000 --> 00:31:18,477 pressure on or there's less supply of 543 00:31:18,560 --> 00:31:21,677 new cars, you know, older cars have more 544 00:31:21,760 --> 00:31:24,637 value, but there's a question long term 545 00:31:24,720 --> 00:31:29,197 what the value of those older chips will be. Some companies are depreciating them 546 00:31:29,280 --> 00:31:31,357 right now on a five or sixyear schedule. 547 00:31:31,440 --> 00:31:34,637 It seems to me that at least what Jensen and some of these others are talking 548 00:31:34,720 --> 00:31:39,437 about is you would have to be using these chips for a decade or two decades 549 00:31:39,520 --> 00:31:43,037 long. I I think that there is a shift 550 00:31:43,120 --> 00:31:47,357 happening because yeah, we classically did look at that at a five to six year 551 00:31:47,440 --> 00:31:51,117 and it's very evidence from this year is very 552 00:31:51,200 --> 00:31:55,517 interesting where just because there's such compute scarcity that it actually 553 00:31:55,600 --> 00:31:57,517 turns out that people continue to use 554 00:31:57,600 --> 00:32:00,077 older generation chips at an even higher 555 00:32:00,160 --> 00:32:03,357 market price than than than you'd expect 556 00:32:03,440 --> 00:32:05,757 because people just need that compute. 557 00:32:05,840 --> 00:32:09,357 And it fundamentally fl flows from the demand that people are seeing trickles 558 00:32:09,440 --> 00:32:11,997 down to well your supply limited right 559 00:32:12,080 --> 00:32:16,557 the compute really determines how how much you can serve and so at that point 560 00:32:16,640 --> 00:32:20,237 anything that gives you compute becomes very valuable very useful and so I think 561 00:32:20,320 --> 00:32:24,877 that that there is a shift undergoing and uh exactly where where it ends up I 562 00:32:24,960 --> 00:32:28,397 think it's too early to say but that what I can definitely say is that the 563 00:32:28,480 --> 00:32:31,357 way we looked at this classically is is just not going to hold. 564 00:32:31,440 --> 00:32:34,637 >> Okay Greg uh it's great to see you this morning. We appreciate you joining us. 565 00:32:34,720 --> 00:32:36,397 Thanks for coming. >> Thank you for having me. 566 00:32:36,480 --> 00:32:40,237 >> You bet. Um, and meantime, we should just uh bring you some additional AI 567 00:32:40,320 --> 00:32:42,077 news just out. Nvidia said that it has 568 00:32:42,160 --> 00:32:47,437 now secured land power and shell capacity in Pike Country, Ohio to host 569 00:32:47,520 --> 00:32:49,037 Nvidia Compute through a partnership 570 00:32:49,120 --> 00:32:53,197 with SoftBank's SB Energy. Open AAI uh 571 00:32:53,280 --> 00:32:57,677 Greg, of course, the president of that company will be the customer. SB Energy 572 00:32:57,760 --> 00:33:01,197 is going to build, own, and operate that data center under a 20-year lease to 573 00:33:01,280 --> 00:33:03,277 Open AI. As part of the deal, Nvidia 574 00:33:03,360 --> 00:33:06,237 will invest $ 1.5 billion dollars in SB 575 00:33:06,320 --> 00:33:10,157 Energy and uh SB Energy's co-CEO Rich 576 00:33:10,240 --> 00:33:12,157 Hosel is going to join us tomorrow 577 00:33:12,240 --> 00:33:13,997 morning first on [music] CNBC. We can 578 00:33:14,080 --> 00:33:17,197 discuss that. 579 00:33:17,280 --> 00:33:20,077 >> Next up on SquawkPod, the future of 580 00:33:20,160 --> 00:33:23,037 delivery with CEO of Zipline, Keller 581 00:33:23,120 --> 00:33:25,197 Clifton. The drone delivery service 582 00:33:25,280 --> 00:33:28,077 envisions a million drone deliveries a 583 00:33:28,160 --> 00:33:30,317 day with the scaling help [music] of new 584 00:33:30,400 --> 00:33:35,277 partner Uber. These kinds of deliveries are not just Costco competitive with 585 00:33:35,360 --> 00:33:39,277 doing a delivery using a 4,000 [music] gas combustion vehicle, but they're way 586 00:33:39,360 --> 00:33:41,437 faster. They're more reliable. They're 587 00:33:41,520 --> 00:33:46,440 cleaner and [music] quieter for the neighborhoods that we serve. 588 00:33:49,520 --> 00:33:54,797 >> You're listening to squawk pod >> up and Q. The US is preparing to tell 589 00:33:54,880 --> 00:33:58,477 countries around the world that they need to pick a side in the AI race with 590 00:33:58,560 --> 00:34:01,677 China. That's according to a Reuters report. that cites a draft state 591 00:34:01,760 --> 00:34:05,117 department letter to about three dozen countries saying they're going to be 592 00:34:05,200 --> 00:34:07,917 excluded from a US AI coalition if they 593 00:34:08,000 --> 00:34:12,557 also sign up for a competing initiative spearheaded by China. Last month, China 594 00:34:12,640 --> 00:34:16,077 launched a group called the World Artificial Intelligence Corporation 595 00:34:16,160 --> 00:34:20,717 organization bolstering openweight AI technology and uh some in the US like 596 00:34:20,800 --> 00:34:22,477 Medus Mark Zuckerberg have advocated for 597 00:34:22,560 --> 00:34:25,277 openweight AI but open AI and anthropic 598 00:34:25,360 --> 00:34:28,477 chiefly rely on closed technology. Some 599 00:34:28,560 --> 00:34:32,477 news out this morning. Uber is partnering with drone delivery company 600 00:34:32,560 --> 00:34:37,037 Zipline to scale up operations in the United States. Joining us right now to 601 00:34:37,120 --> 00:34:41,597 talk about it is Zipline CEO and co-founder Keller Clifton. Keller, first 602 00:34:41,680 --> 00:34:45,197 of all, welcome. Uh we've spoken with you over the years about what you've 603 00:34:45,280 --> 00:34:46,957 been doing. I know you started u with 604 00:34:47,040 --> 00:34:48,957 these drone deliveries in Rwanda and 605 00:34:49,040 --> 00:34:52,477 that was uh many years ago back in 2016. 606 00:34:52,560 --> 00:34:56,237 Obviously, we've come a long way over the last decade. Why don't you describe 607 00:34:56,320 --> 00:35:00,797 what this partnership is and what it means in terms of drone deliveries? 608 00:35:00,880 --> 00:35:02,477 >> Yeah, of course. I mean, you know, as as 609 00:35:02,560 --> 00:35:06,317 you mentioned, Zipline has been building autonomous logistics for the last 13 610 00:35:06,400 --> 00:35:09,997 years. It was only last year, January 15th, that Zipline actually started 611 00:35:10,080 --> 00:35:14,237 doing autonomous deliveries across the US. And um you know, we've been a bit 612 00:35:14,320 --> 00:35:16,477 shocked by the the speed at which that 613 00:35:16,560 --> 00:35:19,597 service has grown. We grew 10 to 15% 614 00:35:19,680 --> 00:35:23,197 week over week, most weeks last year. Um 615 00:35:23,280 --> 00:35:26,637 and so what we're announcing today is this partnership with Uber to scale to a 616 00:35:26,720 --> 00:35:28,877 million deliveries a day. Um you know a 617 00:35:28,960 --> 00:35:31,437 million deliveries a day is uh uh 618 00:35:31,520 --> 00:35:34,957 represents something like 10 10% of 619 00:35:35,040 --> 00:35:38,797 total instant deliveries being done in the US. But what's interesting is the 620 00:35:38,880 --> 00:35:40,477 early customer data that we've observed 621 00:35:40,560 --> 00:35:42,877 in Dallas indicates that when you make 622 00:35:42,960 --> 00:35:47,837 this kind of delivery way faster, way less expensive, less traffic, less 623 00:35:47,920 --> 00:35:50,957 pollution for our neighbor neighborhoods, people significantly 624 00:35:51,040 --> 00:35:54,797 change their buying behavior. Um, a lot of Zip Lines customers are using the 625 00:35:54,880 --> 00:35:58,397 service every single day. And this means that there's actually an opportunity to 626 00:35:58,480 --> 00:36:02,637 grow the size of the instant delivery market more than 10x uh over the coming 627 00:36:02,720 --> 00:36:06,957 years using robotics and physical AI, you know, and and build a service that 628 00:36:07,040 --> 00:36:09,917 ultimately saves people time and saves people money. 629 00:36:10,000 --> 00:36:12,157 >> I mean, it still looks pretty futuristic 630 00:36:12,240 --> 00:36:16,797 to see the videos that we're watching right now of Zipline having a drone come 631 00:36:16,880 --> 00:36:19,117 in and drop you your Uber Eats right to 632 00:36:19,200 --> 00:36:23,917 your doorstep. I have to admit, I have not seen that to this point. Um, when 633 00:36:24,000 --> 00:36:26,957 you started drone deliveries, I think it was something it cost something like 634 00:36:27,040 --> 00:36:31,597 $300 a delivery. You've gotten that cost down to what at this point? 635 00:36:31,680 --> 00:36:35,517 >> Yeah. You know, it's funny. Whenever you talk to a hardware hardware 636 00:36:35,600 --> 00:36:39,037 entrepreneur, I think that, you know, I often ask the question, it's like, cool, 637 00:36:39,120 --> 00:36:41,437 how much do you think this is going to cost when you launch the service? And 638 00:36:41,520 --> 00:36:45,597 however much people say, I always multiply it by 10 in my mind. Um, and 639 00:36:45,680 --> 00:36:49,117 this is exactly what happened at Zipline when we launched the service in 2016. 640 00:36:49,200 --> 00:36:52,797 Um, we initially started by delivering blood transfusions and vaccines. And so 641 00:36:52,880 --> 00:36:54,397 we were lucky that we started with these 642 00:36:54,480 --> 00:36:56,957 lifesaving um you know safety critical 643 00:36:57,040 --> 00:36:59,917 use cases that um meant that we actually 644 00:37:00,000 --> 00:37:04,637 were able to take a decade to travel down that cost curve. Um at this point 645 00:37:04,720 --> 00:37:08,957 these kinds of deliveries are not just cost competitive with doing a delivery 646 00:37:09,040 --> 00:37:13,357 using a 4,000 gas combustion vehicle, but they're way faster, they're more 647 00:37:13,440 --> 00:37:16,957 reliable, they're cleaner and quieter for the neighborhoods that we serve. And 648 00:37:17,040 --> 00:37:19,517 frankly, they just result in a better customer experience. I mean, you can 649 00:37:19,600 --> 00:37:22,797 kind of imagine, you know, if you're ordering something from either a 650 00:37:22,880 --> 00:37:25,917 retailer or from a restaurant, having something delivered in just a couple 651 00:37:26,000 --> 00:37:30,237 minutes, the quality of the experience is way, way higher. You can rely on it. 652 00:37:30,320 --> 00:37:32,157 Um, we today we deliver from 6:00 a.m. 653 00:37:32,240 --> 00:37:36,237 to 11:00 p.m. Soon the service will be operating 24/7. Um, and when you can get 654 00:37:36,320 --> 00:37:38,877 something delivered in 2 or 3 minutes, you know, food tastes better than it 655 00:37:38,960 --> 00:37:42,077 does if it's if it's 30 or 40 minutes cold arriving at your doorstep. 656 00:37:42,160 --> 00:37:45,677 >> And it and it cost what originally was $300. You've gotten it down to what, 657 00:37:45,760 --> 00:37:50,157 about $12, or am I off in the numbers there? Yeah, you know, the the 658 00:37:50,240 --> 00:37:54,877 delivering with a 4,000lb gas commotion vehicle can cost anywhere from like8 to 659 00:37:54,960 --> 00:37:58,077 $15. And so, you know, it's totally different in different metros and 660 00:37:58,160 --> 00:38:00,237 different states, but that is the range 661 00:38:00,320 --> 00:38:05,277 that that this technology is now um is now operating in. And ultimately over 662 00:38:05,360 --> 00:38:08,397 time, this will be not only 10 times faster, but significantly less expensive 663 00:38:08,480 --> 00:38:11,597 than using a human driving a 4,000lb gas combustion vehicle. 664 00:38:11,680 --> 00:38:15,757 >> I I hear your point. You've made it a few times. driving a 4,000lb gas 665 00:38:15,840 --> 00:38:19,597 combustion vehicle. You know, we all have Amazon delivery trucks in our 666 00:38:19,680 --> 00:38:23,277 neighborhoods. We have Uber cars that are coming through. You see all of these 667 00:38:23,360 --> 00:38:24,877 things. And I hear your point that a 668 00:38:24,960 --> 00:38:29,037 drone is less intrusive. It is quieter. 669 00:38:29,120 --> 00:38:33,117 Um maybe safer than having one of these cars or trucks flying down the street if 670 00:38:33,200 --> 00:38:35,037 you've got kids playing on the streets. 671 00:38:35,120 --> 00:38:38,877 But when you start talking about a million deliveries a day, I do get a 672 00:38:38,960 --> 00:38:42,717 little concerned about the air traffic, what that could mean. You've done a lot 673 00:38:42,800 --> 00:38:46,957 of these deliveries. What What's your safety record to this point? 674 00:38:47,040 --> 00:38:49,677 >> So, Zipline has now flown over 140 675 00:38:49,760 --> 00:38:54,077 million commercial autonomous miles. And we have done that with zero safety 676 00:38:54,160 --> 00:38:57,277 incidents. So, just to put that into perspective, if you were to drive that 677 00:38:57,360 --> 00:39:01,197 many miles in the US, that's like driving every road in the entire United 678 00:39:01,280 --> 00:39:05,037 States five times over. And if you were to have a human drive that many miles, 679 00:39:05,120 --> 00:39:09,517 you'd have something like 600 accidents, 100 injuries, and multiple fatalities. 680 00:39:09,600 --> 00:39:13,437 So, we actually already have the data to know that these kinds of systems are way 681 00:39:13,520 --> 00:39:16,077 safer um than than their, you know, 682 00:39:16,160 --> 00:39:19,357 traditional uh um the the default, you 683 00:39:19,440 --> 00:39:24,077 know, uh uh that we've been using to to make these deliveries. Only that these 684 00:39:24,160 --> 00:39:26,717 aircraft that you can see behind me, we now have individual aircraft that have 685 00:39:26,800 --> 00:39:30,557 flown more than a million commercial autonomous miles. So, a lot of times 686 00:39:30,640 --> 00:39:34,637 people think of this technology as like sensitive or fragile. Um you know, the 687 00:39:34,720 --> 00:39:38,397 average car is only going to drive 150,000 miles during its lifetime. And 688 00:39:38,480 --> 00:39:41,757 so these kinds of systems are already not just safer, but they're already a 689 00:39:41,840 --> 00:39:45,357 lot more reliable than than um than cars. 690 00:39:45,440 --> 00:39:49,837 >> Keller, thank you for your time. Um eager to see more about how this works 691 00:39:49,920 --> 00:39:52,397 out, but we appreciate you joining us today. 692 00:39:52,480 --> 00:39:56,157 >> Yeah, thank you for having me. 693 00:39:56,240 --> 00:39:59,677 >> That's Squawk Pod for today. Thank you for listening. Squawk [music] Box is 694 00:39:59,760 --> 00:40:01,917 hosted by Joe Kernan, Becky Quick, and 695 00:40:02,000 --> 00:40:04,237 Andrew Ross Orcin weekday mornings on 696 00:40:04,320 --> 00:40:07,677 CNBC starting at 6 Eastern. To get the 697 00:40:07,760 --> 00:40:09,837 best bits of [music] our three-hour TV 698 00:40:09,920 --> 00:40:12,637 show, tune into SquawkPod wherever you 699 00:40:12,720 --> 00:40:14,957 listen to podcasts. We'll meet you right 700 00:40:15,040 --> 00:40:17,517 back here tomorrow. Have a great day. 701 00:40:17,600 --> 00:40:23,000 >> We are clear. Thanks, guys.