WEBVTT 1 00:00:00.080 --> 00:00:03.319 Look, I'd rather live in the United States where we 2 00:00:03.480 --> 00:00:08.119 have control over the technology and information that affects Americans. 3 00:00:08.800 --> 00:00:11.119 And when you have tech like that used to be 4 00:00:11.279 --> 00:00:16.399 fine for basically a century because America made the best 5 00:00:16.399 --> 00:00:19.559 technology in most of the technology. Now we're entering a 6 00:00:19.559 --> 00:00:23.519 world where China makes a lot of technology, and in 7 00:00:23.559 --> 00:00:27.320 some cases may make better technology, but they don't want 8 00:00:27.359 --> 00:00:33.119 to abide by our use or our constitution, our rules 9 00:00:33.119 --> 00:00:37.679 around this. So I would just say, fine, like, you 10 00:00:37.679 --> 00:00:39.520 can do whatever you want in your country, but we 11 00:00:39.640 --> 00:00:41.240 are not going to use your technology. 12 00:00:41.280 --> 00:00:48.679 Then, everybody, welcome to another episode of Eyes on Geopolitics. 13 00:00:48.679 --> 00:00:51.479 I'm here today with Jonathan Hacket, and our guest is 14 00:00:51.560 --> 00:00:56.079 Anthony Vincey. He's the author of The Fourth Intelligence Revolution, 15 00:00:56.240 --> 00:00:58.840 The Future of Espionage, and the Battle of Save America. 16 00:01:00.200 --> 00:01:04.159 He's got a pretty amazing resume. He previously served as 17 00:01:04.159 --> 00:01:07.879 the first CTO and Associate Director of Capabilities at the 18 00:01:07.959 --> 00:01:12.760 National Geospatial Intelligence Agency, following earlier work as an intelligence 19 00:01:12.760 --> 00:01:15.400 officer in Iraq, Africa, and Asia. You could actually watch 20 00:01:15.480 --> 00:01:17.480 an episode with Anthony on the Team House that we 21 00:01:17.560 --> 00:01:22.040 did probably a good six months ago. Now, great episode, 22 00:01:22.079 --> 00:01:23.680 check that out. I'll put a link in the description 23 00:01:23.719 --> 00:01:25.120 you can check it out. I'll also put a link 24 00:01:25.159 --> 00:01:28.319 in the descripture for his book. He is now the 25 00:01:28.439 --> 00:01:31.760 co founder and CEO of Vico, an AI company focused 26 00:01:31.760 --> 00:01:34.439 on decision making and finance and national security, and is 27 00:01:34.439 --> 00:01:38.000 an adjunct Senior Fellow at the Center for a New 28 00:01:38.000 --> 00:01:41.640 American Security and to boot, he holds his PhD in 29 00:01:41.719 --> 00:01:45.319 International relations from the London School of Economics. So I 30 00:01:45.359 --> 00:01:48.519 try to get somebody here that's on the level of 31 00:01:48.680 --> 00:01:51.719 smart as Jonathan Hackett is, because Jonathan's got a pretty 32 00:01:51.719 --> 00:01:54.560 good resume too, and I think I did a pretty 33 00:01:54.560 --> 00:01:56.680 good job with the Anthony. So thanks for joining us today. 34 00:01:56.879 --> 00:01:58.200 Yeah, man, thanks for having me. 35 00:02:00.159 --> 00:02:03.120 This whole talk to be about what's going on with AI. 36 00:02:03.200 --> 00:02:05.000 Me and Jonathan have been talking about this for like 37 00:02:05.040 --> 00:02:06.879 a couple of months, like putting together a show and 38 00:02:06.879 --> 00:02:09.400 trying to find a guest that can, like, you know, 39 00:02:09.439 --> 00:02:11.319 give us the lay of the land of what's going on. 40 00:02:11.520 --> 00:02:13.479 And I've joked with Jonathan. I think I joked with 41 00:02:13.520 --> 00:02:15.719 you want an email too, about saying like this literally 42 00:02:15.759 --> 00:02:18.919 could be a three hour conversation because it's like there's 43 00:02:18.960 --> 00:02:23.879 just so many rabbit holes we can go down full disclosure. 44 00:02:24.000 --> 00:02:27.360 I am more on the bearers side of AIS, particularly 45 00:02:27.439 --> 00:02:30.960 when it comes to like just all all the capex 46 00:02:31.000 --> 00:02:33.280 that's going on, and like, you know, wh where's it 47 00:02:33.319 --> 00:02:36.599 really going? The cow expensive the computer is compared to 48 00:02:36.680 --> 00:02:39.240 like how you know enterprises are using them, like they 49 00:02:39.280 --> 00:02:42.400 doesn't really look like there's a big at least for 50 00:02:42.439 --> 00:02:44.199 the frontier model. So it doesn't look like there's a 51 00:02:44.240 --> 00:02:50.800 big profit center right now. But anyway, let's just get 52 00:02:50.800 --> 00:02:53.280 into it. I mean, I don't know how you guys 53 00:02:53.319 --> 00:02:55.960 want to play it, Jonathan, please, like let's just you know, 54 00:02:57.080 --> 00:03:00.719 improve this basically, do we want it explain like what 55 00:03:00.759 --> 00:03:03.319 the frontier models mean, like should we go more of 56 00:03:03.360 --> 00:03:05.599 a basic route for like the Layman. I'm in Layman 57 00:03:05.639 --> 00:03:07.560 obviously too, but I just read this insensitively. 58 00:03:08.400 --> 00:03:10.680 I think it's helpful to have like a general question 59 00:03:10.759 --> 00:03:14.280 of Okay, why should people care about the crossover question 60 00:03:14.360 --> 00:03:17.560 of China and AI? Because there's the China national security 61 00:03:17.560 --> 00:03:20.000 threat that's existed for a long time, especially international security 62 00:03:20.000 --> 00:03:21.840 strategy all the way back to nineteen eighty seven when 63 00:03:21.840 --> 00:03:25.159 the first one came out. Now AI has entered the chat, 64 00:03:25.639 --> 00:03:28.360 and the question for the American listener is why should 65 00:03:28.360 --> 00:03:29.000 they care about that? 66 00:03:31.000 --> 00:03:32.680 Yeah, I love it. Let's go because we can go 67 00:03:32.719 --> 00:03:35.759 in any direction here. So Anthony, what are you track 68 00:03:35.840 --> 00:03:38.319 in basically right now? With AI? I mean, I know 69 00:03:38.400 --> 00:03:41.879 I sent you over that Financial Times article and that 70 00:03:41.919 --> 00:03:44.400 does seem like the big, the big scary. 71 00:03:45.080 --> 00:03:49.680 Yeah, well look China AI, I think is that's the question. 72 00:03:49.919 --> 00:03:54.560 That's like the generational level question because I think about 73 00:03:54.560 --> 00:03:58.439 it like this. AI is not like software. AI does 74 00:03:58.560 --> 00:04:03.000 what people do right, and it communicates information to you. 75 00:04:03.280 --> 00:04:08.960 It doesn't just show you something, but you interact with it, right, 76 00:04:09.319 --> 00:04:11.960 So anything when you get on claud, you get on chat, 77 00:04:12.000 --> 00:04:16.800 GPT and you ask it a question, it can ask 78 00:04:16.839 --> 00:04:20.399 you questions. You can continue to ask it questions. And 79 00:04:20.480 --> 00:04:24.000 the thing is that it can persuade you to do 80 00:04:24.120 --> 00:04:27.199 things right, like it can lean you in a direction 81 00:04:27.680 --> 00:04:30.399 that's different than anything we've had before. Only people were 82 00:04:30.480 --> 00:04:33.279 able to do that previously. So that's one thing. The 83 00:04:33.319 --> 00:04:36.959 second thing is that AI that's kind of LMS like 84 00:04:37.199 --> 00:04:39.720 AI you talk to. There's another form of AI that 85 00:04:39.759 --> 00:04:45.000 you're using for autonomous vehicles that can drive way more 86 00:04:45.120 --> 00:04:48.959 cars around, or it can fly drones around. That's also 87 00:04:50.079 --> 00:04:54.560 super important. There's a third kind of AI that I 88 00:04:54.600 --> 00:04:57.000 think is you were going to see more and more of. 89 00:04:57.120 --> 00:04:59.079 And it's kind of what I do in my company 90 00:04:59.279 --> 00:05:03.399 at Vico, which is using AI to help run businesses 91 00:05:04.240 --> 00:05:08.240 or even to help run the military. It's making strategic choices. 92 00:05:08.399 --> 00:05:14.360 It's allowing you to play through lots of decision lots 93 00:05:14.399 --> 00:05:17.519 of decisions. Right, those three kinds of AI. That's kind 94 00:05:17.519 --> 00:05:21.560 of how the world works. You have people talking to 95 00:05:21.600 --> 00:05:25.040 people now talking AI. You have AI that's able to 96 00:05:25.120 --> 00:05:28.879 drive cars and other vehicles, and you have a system 97 00:05:28.920 --> 00:05:31.319 that can run companies. Now, these are the things that 98 00:05:31.360 --> 00:05:36.879 we obviously are competing with China around. And the question 99 00:05:36.959 --> 00:05:39.800 is going to be are we going to use Chinese 100 00:05:39.839 --> 00:05:44.079 software or AI systems to do these things? Or we're 101 00:05:44.079 --> 00:05:46.959 going to use American systems? And then what about our 102 00:05:47.839 --> 00:05:52.120 allies or just third party countries like that are neither 103 00:05:52.199 --> 00:05:54.920 aligned with the US nor China kind of sit in 104 00:05:55.000 --> 00:05:58.399 the middle and what are they going to use? And 105 00:05:58.439 --> 00:06:03.519 we've seen this situation before or with Huawei with telecommunications equipment, 106 00:06:04.120 --> 00:06:06.199 where the question became, are you going to have your 107 00:06:06.199 --> 00:06:11.120 telecommunication system set up on Western equipment from ericson or something, 108 00:06:11.319 --> 00:06:12.399 or are you going to have it set up on 109 00:06:12.480 --> 00:06:15.800 Chinese equipment. The downside risk there was that Chinese equipment, 110 00:06:15.879 --> 00:06:19.160 they can spy on it and steal all of your data. 111 00:06:19.240 --> 00:06:23.079 It's completely unsecure. Well, we have the same situation with 112 00:06:23.240 --> 00:06:26.319 China right now, is that if you use a Chinese 113 00:06:26.399 --> 00:06:30.240 AI system, they can spy on it, they can access 114 00:06:30.279 --> 00:06:34.160 your data. But it's even worse now because they could 115 00:06:34.519 --> 00:06:38.560 do two other things. One, in theory, they could use 116 00:06:38.600 --> 00:06:42.399 that AI system to control what you're using it for. 117 00:06:42.560 --> 00:06:45.360 So if you're using a piece of AI to run 118 00:06:45.399 --> 00:06:48.160 an autonomous vehicle, in theory, they might be able to 119 00:06:48.160 --> 00:06:51.079 control that autonomous vehicle. That could be really problematic in 120 00:06:51.160 --> 00:06:55.600 war for example. But the even more nefarious possibility is 121 00:06:55.639 --> 00:06:59.120 that Chinese AI remember the part where you can talk 122 00:06:59.199 --> 00:07:02.639 to it, you use chat GPT where you ask it 123 00:07:02.680 --> 00:07:06.759 a question, say about how to vote in this coming midterm. 124 00:07:07.360 --> 00:07:10.360 If it's Chinese AI system, they could mount an information 125 00:07:10.439 --> 00:07:13.360 attack against you and can actually persuade you to do something. 126 00:07:14.399 --> 00:07:18.240 So those are like really serious consequences for like the 127 00:07:18.360 --> 00:07:24.519 average everyday American citizen, Like do you really want your vote, 128 00:07:24.680 --> 00:07:27.160 for example, in the midterm to be biased in some 129 00:07:27.279 --> 00:07:30.480 way because of your interaction with the Chinese system. Do 130 00:07:30.519 --> 00:07:33.920 you really want this waym car that you're driving in 131 00:07:35.240 --> 00:07:39.279 to potentially be controlled by China and say in a 132 00:07:39.319 --> 00:07:41.439 state of war, to be used against you maybe I 133 00:07:41.480 --> 00:07:44.920 don't know, runs people over or something, or American military 134 00:07:44.959 --> 00:07:48.839 equipment to be harmed by a Chinese AI system. And 135 00:07:48.920 --> 00:07:51.040 worst case scenario, if you're running your business and you're 136 00:07:51.160 --> 00:07:53.519 using AI to do this, or the government is being 137 00:07:53.560 --> 00:07:55.600 run by AI, do you really want it to be 138 00:07:55.639 --> 00:07:58.519 compromised by Chinese AI system? And I think these are 139 00:07:58.600 --> 00:08:00.519 like really serious questions to address. 140 00:08:02.000 --> 00:08:03.759 And I think also too, I was actually reading an 141 00:08:03.800 --> 00:08:06.600 article recently, I think this week about open weight models 142 00:08:06.639 --> 00:08:08.879 that China is kind of pioneering and getting out there 143 00:08:08.920 --> 00:08:10.600 a lot of and people are seeing them their lower 144 00:08:10.600 --> 00:08:12.639 costs and all these there are all kinds of benefits 145 00:08:12.639 --> 00:08:15.759 and advantages on one ledger. And as I'm reading more 146 00:08:15.759 --> 00:08:19.000 about what open weight actually means, at my first thought 147 00:08:19.240 --> 00:08:21.920 before I dug into this was that open weight was 148 00:08:21.959 --> 00:08:24.240 the same as open source, and I've learned that that's 149 00:08:24.279 --> 00:08:26.879 not true. Can you talk a little bit about why 150 00:08:26.959 --> 00:08:28.079 that's problematic? 151 00:08:29.920 --> 00:08:34.840 Yeah, the yeah, open weight means you're getting the open 152 00:08:34.840 --> 00:08:36.799 source means it's a piece of software that you can 153 00:08:36.919 --> 00:08:38.720 use and you can modify, and you can do whatever 154 00:08:38.720 --> 00:08:42.039 you want with and build on. So like Linux, for example, 155 00:08:42.440 --> 00:08:45.240 open weight means that they have provided the model, but 156 00:08:45.360 --> 00:08:48.720 not the underlying data in the model. So it does 157 00:08:48.799 --> 00:08:50.799 mean that you can use it now and you can 158 00:08:51.440 --> 00:08:56.480 build that, you can build an app, say on top 159 00:08:56.519 --> 00:08:59.879 of it, but you don't actually know the data that 160 00:09:00.039 --> 00:09:02.440 and into training it, you don't know exactly how it's 161 00:09:02.480 --> 00:09:05.919 going to respond. So the danger is that it could 162 00:09:06.240 --> 00:09:10.000 respond in a way that is biased. So kind of notoriously, 163 00:09:10.039 --> 00:09:15.039 the Chinese open weight models are biased towards the Chinese 164 00:09:15.080 --> 00:09:18.159 Communist Party. Like if you ask a deep Seek model 165 00:09:18.240 --> 00:09:21.320 or Kimmi or choose your Chinese open weight model a 166 00:09:21.399 --> 00:09:23.919 question about tam and Square, it's going to give you 167 00:09:23.960 --> 00:09:27.360 the party line, right, it's going to say what the 168 00:09:27.399 --> 00:09:32.600 CCP wants you to hear, and that that's problematic on 169 00:09:32.639 --> 00:09:37.159 it on its own. And but really what open weight 170 00:09:37.200 --> 00:09:40.200 and open source mean is that it's free. And so 171 00:09:40.480 --> 00:09:45.679 what China's done is they've made these free models. Deep 172 00:09:45.679 --> 00:09:48.120 Seek and a couple of other companies have gone out 173 00:09:48.159 --> 00:09:51.440 and done this, and so what's happening is they're essentially 174 00:09:51.519 --> 00:09:55.480 giving like this low cost solution that US or other 175 00:09:55.799 --> 00:09:59.879 country companies can build on top of. And so it's 176 00:10:00.200 --> 00:10:03.720 as similar to the Huawei thing again, right, Huahwei was 177 00:10:03.759 --> 00:10:08.480 able to produce routers and other telecommunications equipment at a 178 00:10:08.559 --> 00:10:11.200 fraction of the cost of Western equipment, so people started 179 00:10:11.279 --> 00:10:14.799 using that kind of equipment, and these these systems were 180 00:10:14.960 --> 00:10:18.600 compromised in a way or could be compromised by Chinese 181 00:10:18.639 --> 00:10:21.919 intelligence and so forth. Well, same thing with AI. They're 182 00:10:21.960 --> 00:10:24.960 sort of giving away these open source models that are 183 00:10:25.399 --> 00:10:29.840 effectively free so that people will build their own software 184 00:10:29.879 --> 00:10:32.159 and apps and so forth on top of them. But 185 00:10:32.639 --> 00:10:33.519 they're compromised. 186 00:10:35.279 --> 00:10:36.799 And actually that brings up a good point. I was 187 00:10:37.519 --> 00:10:39.840 working somewhere over the summer where we got fable from 188 00:10:39.879 --> 00:10:43.279 Anthropic for like two days, and then the government shut 189 00:10:43.320 --> 00:10:45.240 it down, and there was this kind of disruptive moment 190 00:10:45.240 --> 00:10:46.879 where the government was trying at the US government was 191 00:10:46.879 --> 00:10:49.360 trying to figure out, how do we regulate this, do 192 00:10:49.399 --> 00:10:51.360 we put export controls on it? How dangerous is it? 193 00:10:51.399 --> 00:10:53.720 And all these questions. But at the same time China 194 00:10:53.799 --> 00:10:55.919 is looking at this in the opposite way. They're trying 195 00:10:55.919 --> 00:10:58.399 to get as much access to people as possible, whereas 196 00:10:58.399 --> 00:11:00.840 the US seems to be trying to be more careful. 197 00:11:00.879 --> 00:11:03.200 What's kind of the right way to balance these two 198 00:11:03.240 --> 00:11:04.679 competing philosophies. 199 00:11:06.600 --> 00:11:10.320 Yeah, so there's a downside to letting AI models out right, 200 00:11:10.399 --> 00:11:15.559 which is that they can be used to do dangerous things. Right, 201 00:11:15.679 --> 00:11:17.759 you could take an AI model and in the case 202 00:11:17.799 --> 00:11:23.320 of Fable and Mythos the Claude models, that they could 203 00:11:23.360 --> 00:11:28.519 be used to hack to perform an offensive cyber attack, 204 00:11:29.679 --> 00:11:31.879 and that could be really dangerous. Imagine you just sort 205 00:11:31.879 --> 00:11:33.799 of let everybody in the world have that, and now 206 00:11:33.799 --> 00:11:37.440 these criminals, criminal hackers were able to use the stuff, 207 00:11:37.440 --> 00:11:40.080 and then they can go and pack into banks or 208 00:11:40.120 --> 00:11:43.080 into healthcare systems and so forth. Could be super dangerous. 209 00:11:45.519 --> 00:11:49.080 So there is a debate just on that, like when 210 00:11:49.159 --> 00:11:51.559 and how should we release these models. And it gets 211 00:11:51.639 --> 00:11:55.639 worse by the way than just hacking. As these models 212 00:11:55.679 --> 00:11:57.840 get more powerful, it's possible you could use them to 213 00:11:57.919 --> 00:12:02.000 create like a bioweapon, right, that a person with little 214 00:12:02.039 --> 00:12:07.440 to no biological science experience could actually create a bioweapon 215 00:12:07.480 --> 00:12:10.799 by following the instructions like you would with you know, 216 00:12:11.039 --> 00:12:14.679 a cooking recipe from an AI system, and that like 217 00:12:14.759 --> 00:12:17.200 that's like terrifying to even think about. So we want 218 00:12:17.240 --> 00:12:22.399 to have some control. China like does have some control 219 00:12:23.279 --> 00:12:27.000 even on the the open weight models, right. The CCP 220 00:12:27.200 --> 00:12:31.080 has established guidelines and controls to those companies like to 221 00:12:31.120 --> 00:12:34.960 Deep Seek, for example, around information that they want certain 222 00:12:34.960 --> 00:12:38.440 information about Tenement Square or the Wigers, which is a 223 00:12:38.480 --> 00:12:42.240 minority group that's that's persecuted in China. They want to 224 00:12:42.240 --> 00:12:45.000 control that information, so they are putting in those controls. 225 00:12:45.720 --> 00:12:49.159 But then these open weight models are being released to 226 00:12:49.200 --> 00:12:53.440 everybody and they don't have there's no way for the 227 00:12:53.519 --> 00:12:56.519 US government to put its own controls, like say, around 228 00:12:56.559 --> 00:13:00.720 cybersecurity issues. Now, this wasn't a problem until recently because 229 00:13:01.240 --> 00:13:05.039 the US models like Claude and Chat, GPT and GROC 230 00:13:05.120 --> 00:13:07.679 and so forth, were by far the most powerful models 231 00:13:07.720 --> 00:13:10.159 and could do the most powerful things. And these open 232 00:13:10.159 --> 00:13:14.480 weight models were, you know, a generation or two generations behind, 233 00:13:15.000 --> 00:13:17.480 and so they weren't as dangerous. But now they're sort 234 00:13:17.480 --> 00:13:19.919 of catching up, and as they catch up, they might 235 00:13:19.960 --> 00:13:23.960 get as dangerous as a Claude model for hacking and 236 00:13:24.080 --> 00:13:26.360 for other purposes, and there's no way for the US 237 00:13:26.480 --> 00:13:29.279 government to regulate it or control it in any way. 238 00:13:29.759 --> 00:13:33.320 And in theory, China might let people, you know, allow 239 00:13:33.799 --> 00:13:36.519 a model out that could be used for cyber hacking. 240 00:13:37.919 --> 00:13:40.879 What do you guys take what's your take on? You know, 241 00:13:40.960 --> 00:13:44.679 Microsoft announced maybe a few weeks ago, also Microsoft being 242 00:13:44.720 --> 00:13:47.080 a big investor into open ai, that they're going to 243 00:13:47.120 --> 00:13:49.600 start using deep seek or at least look into using 244 00:13:49.639 --> 00:13:52.679 deep seek for their some of their enterprise stuff. I 245 00:13:52.679 --> 00:13:57.480 mean that's going to ring alarm bells for everybody, know, Yeah. 246 00:13:57.120 --> 00:14:03.120 I mean, look, my personal take is that just like Huawei, 247 00:14:03.159 --> 00:14:08.480 it should effectively be made illegal or at least very 248 00:14:08.480 --> 00:14:11.480 problematic for US business to use these things because for 249 00:14:11.519 --> 00:14:14.879 all the reasons I mentioned, like these things, they're compromised 250 00:14:15.159 --> 00:14:17.440 in that sense, and there's a danger in that they're 251 00:14:17.480 --> 00:14:22.799 already large American companies using Chinese open weight models. Airbnb 252 00:14:23.080 --> 00:14:27.399 uses them, there's a few other big ones, and then 253 00:14:27.440 --> 00:14:30.720 there's lots of small startups that are using them. Microsoft 254 00:14:30.799 --> 00:14:33.639 saying they're going to use it is not great. I 255 00:14:33.720 --> 00:14:37.000 suspect that really they would be using it for something 256 00:14:37.039 --> 00:14:41.559 that they consider to be low security risk, right, And 257 00:14:41.600 --> 00:14:44.679 that's fine, But then who's making that call about what's 258 00:14:44.720 --> 00:14:50.440 low security risk? Right? And we are in effect creating 259 00:14:50.759 --> 00:14:53.480 the same situation as Huawei, where we're going to kind 260 00:14:53.480 --> 00:14:56.360 of build up their open weight model companies over time 261 00:14:57.240 --> 00:15:00.600 and then we won't be able to compete. 262 00:15:00.639 --> 00:15:02.840 You mentioned something really interesting, you said, like Airbnb and 263 00:15:02.840 --> 00:15:05.879 these smaller companies are using deepseek. You know, I used 264 00:15:05.919 --> 00:15:09.639 to teach at our Human Intelligence Collection course and we 265 00:15:09.679 --> 00:15:12.840 taught about ubiquitous technical surveillance and how to do things 266 00:15:12.960 --> 00:15:16.399 under a uts threat and so on. That was before AI, 267 00:15:16.960 --> 00:15:20.240 And now I'm very curious about how it's not just 268 00:15:20.279 --> 00:15:23.000 a vector. I mean, in ubiquitous technical surveillance, you have 269 00:15:23.080 --> 00:15:26.320 multiple vectors that overlap that make the thing a threat 270 00:15:26.399 --> 00:15:29.399 or a problem for you. And it seems that it 271 00:15:29.440 --> 00:15:32.240 goes beyond the vector problem when you're using AI. Like 272 00:15:32.320 --> 00:15:34.600 let's say you're booking on Airbnb, and there's certain behavioral 273 00:15:34.679 --> 00:15:37.600 patterns that you the user use on Airbnb that are 274 00:15:37.759 --> 00:15:40.519 almost exactly the same behaviors or exactly the same behaviors 275 00:15:40.559 --> 00:15:43.639 when you're undercover under a different persona on some other platform, 276 00:15:43.639 --> 00:15:45.879 and AI is able to see that and correlate you 277 00:15:45.919 --> 00:15:49.799 immediately way faster than a human analyst could. I already mentioned, 278 00:15:49.799 --> 00:15:51.320 you know, what would you say to the average American. 279 00:15:51.320 --> 00:15:53.320 But I'm curious also to the average case officer who 280 00:15:53.360 --> 00:15:54.720 might be listening, because I know there are some who 281 00:15:54.799 --> 00:15:56.960 do listen to the show. What would you kind of 282 00:15:56.960 --> 00:15:57.840 caution them on? 283 00:15:59.120 --> 00:16:01.960 Yeah, I mean, well, look, even for the average American, 284 00:16:02.000 --> 00:16:05.360 by the way, I don't think the average American wants 285 00:16:05.759 --> 00:16:10.759 a Chinese company or a Chinese security service or the 286 00:16:10.840 --> 00:16:14.679 Chinese government to know what they're doing and to be 287 00:16:14.720 --> 00:16:17.240 able to match the records from Airbnb to other things 288 00:16:17.279 --> 00:16:19.960 in their life. We don't want our government doing that, right, 289 00:16:20.080 --> 00:16:24.840 Like your average American like me, is like at least 290 00:16:24.840 --> 00:16:27.440 a little bit libertarian leading in that sense. We don't 291 00:16:27.440 --> 00:16:29.320 want our government to spy on us. Why would we 292 00:16:29.360 --> 00:16:31.159 want the Chinese government to spy on us? 293 00:16:31.240 --> 00:16:31.399 Right? 294 00:16:31.799 --> 00:16:34.840 So I think everybody, But yeah, for case officers. Look, 295 00:16:34.879 --> 00:16:38.519 I was a case officer, you know, and by the 296 00:16:38.559 --> 00:16:43.159 time I stopped doing that, the UTS issue is just starting, right, 297 00:16:43.200 --> 00:16:46.240 and so I only really saw a little bit of it, 298 00:16:46.279 --> 00:16:50.080 and I can't imagine, starting from scratch today what it 299 00:16:50.159 --> 00:16:55.919 might must be like. And the you know, even without AI, 300 00:16:56.279 --> 00:17:00.279 you can match people's patterns right, just literally, like old 301 00:17:00.320 --> 00:17:04.440 school big data analytics, I crunch enough data about you, 302 00:17:04.599 --> 00:17:10.000 I can detect patterns. The AI makes it easier and 303 00:17:10.119 --> 00:17:14.279 more ubiquitous, right, Like I don't even need to you know, 304 00:17:14.559 --> 00:17:16.279 I can. I can do it like that example you 305 00:17:16.359 --> 00:17:20.319 brought up as a great example. Actually, like my habits. 306 00:17:20.960 --> 00:17:23.759 It might not, by the way, even be what I'm 307 00:17:23.799 --> 00:17:26.799 doing on Airbnb. It could even be like the way 308 00:17:26.839 --> 00:17:30.880 I move my mouse or the way I type. Right, 309 00:17:30.960 --> 00:17:35.519 there are signatures that are everybody has. These patterns can 310 00:17:35.559 --> 00:17:39.559 be taken from one place and then and then applied 311 00:17:39.640 --> 00:17:41.920 to another. So as a case officer, you're trying to 312 00:17:41.960 --> 00:17:46.599 act and cover somewhere. And you know, even if you 313 00:17:47.240 --> 00:17:50.319 even if you use a false name for everything you 314 00:17:50.359 --> 00:17:53.359 did in work, and in fact, even if you used 315 00:17:53.359 --> 00:17:56.880 a false name for everything you did in your commercial life, 316 00:17:57.000 --> 00:18:01.000 well your pattern still might be there. Right, Like maybe 317 00:18:01.039 --> 00:18:04.079 you're using your wife's account to log into Airbnb, but 318 00:18:04.119 --> 00:18:08.119 you're the one on the mouse and making the decisions. 319 00:18:08.599 --> 00:18:12.000 It would still be able to recognize these patterns. So 320 00:18:12.200 --> 00:18:17.319 it's like becomes nearly impossible to kind of like keep 321 00:18:17.359 --> 00:18:19.640 yourself out of there now. I actually wrote a whole 322 00:18:19.720 --> 00:18:22.599 chapter about this in my book because like I think, 323 00:18:22.680 --> 00:18:25.000 what where it goes to be a case officer. I 324 00:18:25.039 --> 00:18:28.519 think you just have to sort of assume that anything, 325 00:18:29.079 --> 00:18:32.039 any pattern is or could be tracked, and then you 326 00:18:32.480 --> 00:18:34.440 I think we will enter a world where you just 327 00:18:34.519 --> 00:18:37.039 have to sort of live in pattern and what it 328 00:18:37.119 --> 00:18:39.400 is to kind of be and cover and so forth 329 00:18:39.440 --> 00:18:42.680 completely changes. Like there's no more doc swapping and all 330 00:18:42.720 --> 00:18:45.039 that stuff we saw in spy movies, Like that's over 331 00:18:45.839 --> 00:18:48.000 and like now now you just have to sort of 332 00:18:48.000 --> 00:18:50.839 live your life as that person, like kind of closer 333 00:18:50.880 --> 00:18:52.440 to what it was like to be a source back 334 00:18:52.440 --> 00:18:53.640 in the day. 335 00:18:54.440 --> 00:18:57.319 Hello everyone, if you're hearing this you are not a 336 00:18:57.400 --> 00:19:01.759 member of the Patreon you can get your hands on 337 00:19:02.279 --> 00:19:05.240 ad free episodes that come out early and you can 338 00:19:05.279 --> 00:19:07.519 ask us questions maybe we'll answer them. You can also 339 00:19:07.599 --> 00:19:11.359 watch the team House podcast completely ad free and early, 340 00:19:11.440 --> 00:19:13.559 and you can also watch it live as we interview 341 00:19:13.559 --> 00:19:17.279 our guests, and it's the best way to support the channel. 342 00:19:17.440 --> 00:19:19.599 So if you can, if you're open to it, you 343 00:19:19.599 --> 00:19:22.720 can go to Patreon dot com slash the Teamhouse. That 344 00:19:22.799 --> 00:19:26.759 link is in the description. We appreciate the supporters and 345 00:19:26.880 --> 00:19:29.359 we hope you would consider being one. So it's Patreon 346 00:19:29.440 --> 00:19:31.839 dot com slash the Teamhouse. Thanks guys. 347 00:19:33.759 --> 00:19:36.119 Yeah, Actually, I was reading about when the strikes are 348 00:19:36.160 --> 00:19:38.680 going on February twenty eighth and forward in Iran at 349 00:19:38.680 --> 00:19:41.200 the very beginning of the war, how Israel used some 350 00:19:41.240 --> 00:19:44.039 security cameras and some uts access they had inside of 351 00:19:44.079 --> 00:19:47.200 Iran to better triangulate targets and then predict follow on 352 00:19:47.240 --> 00:19:50.359 targets very rapidly, which is fantastic. But then the question 353 00:19:50.400 --> 00:19:52.480 is that's with the AI that we have right now. 354 00:19:53.079 --> 00:19:55.720 What happens when it's an order of magnitude either more 355 00:19:55.720 --> 00:20:00.079 intelligent or faster or in the wrong hands, you know what, 356 00:20:00.200 --> 00:20:04.359 how do we develop safeguards for that in a military context? 357 00:20:05.440 --> 00:20:08.680 Yeah, Or what happens if it's applied to us here 358 00:20:09.079 --> 00:20:13.200 right where every you know, every factory in America has 359 00:20:14.279 --> 00:20:17.559 some electronic device in it that might be, you know, 360 00:20:17.720 --> 00:20:21.640 could be compromised. And China has an AI system that's 361 00:20:21.720 --> 00:20:25.839 able to sort of assess and analyze all of those factories, 362 00:20:25.880 --> 00:20:29.359 all the people working on them, and then do do 363 00:20:29.559 --> 00:20:32.359 something about it. It could be you're right, it could 364 00:20:32.400 --> 00:20:35.599 be using that for targeting information, right. It could be 365 00:20:36.799 --> 00:20:41.640 they can figure out just by these assessment features, you know, 366 00:20:41.680 --> 00:20:45.640 which ones are building key you know, components in a missile. 367 00:20:46.599 --> 00:20:48.920 It also could be well, why not just target the 368 00:20:48.960 --> 00:20:53.119 people working at the factories non kinetic means, right, Like 369 00:20:53.200 --> 00:20:58.920 why not use an information attack against you know, sometimes 370 00:20:59.000 --> 00:21:01.160 in these factories for exams, you have like one or 371 00:21:01.200 --> 00:21:03.720 two guys who are the key guy, like they know 372 00:21:03.839 --> 00:21:06.519 how to do everything. Why don't I just target that guy? 373 00:21:06.559 --> 00:21:10.359 Why don't I implant some criminal information on that guy's 374 00:21:10.400 --> 00:21:14.359 computer and then put him in prison? And now that 375 00:21:14.400 --> 00:21:18.960 factory is fifty percent as effective because we've lost that guy, right, Like, 376 00:21:19.079 --> 00:21:22.400 there's all sorts of things that can be done if 377 00:21:22.400 --> 00:21:24.519 you you know, if you're sneaky thinker with this stuff. 378 00:21:27.720 --> 00:21:30.039 I question into going back a little bit to uh, 379 00:21:30.400 --> 00:21:34.240 you know, being worried about American companies and American startups 380 00:21:34.319 --> 00:21:38.359 using the Chinese models that are some places I've read 381 00:21:38.359 --> 00:21:40.359 it's like one third cheaper and stuff like that. How 382 00:21:40.400 --> 00:21:42.839 do we stop? And there was like how do we 383 00:21:42.880 --> 00:21:47.319 stop basically it being so expensive for American companies to 384 00:21:47.480 --> 00:21:48.720 use frontier models? 385 00:21:49.960 --> 00:21:56.640 Yeah, yeah, the the first of all I would start 386 00:21:56.680 --> 00:22:00.559 to question some of the like costs different is right, 387 00:22:00.720 --> 00:22:04.960 just like like with anything open source, like the actual 388 00:22:05.079 --> 00:22:08.200 cost to buy it versus like the core cost to implement. 389 00:22:08.240 --> 00:22:11.119 It might be totally different, but let's just assume from it. 390 00:22:11.160 --> 00:22:12.640 And I think I think it is true. It is 391 00:22:12.680 --> 00:22:15.160 going to be it is going to be cheaper. I 392 00:22:15.240 --> 00:22:19.440 think we need basically cheaper models. And I have this 393 00:22:19.519 --> 00:22:24.640 idea which I'll throw out right here, which is the 394 00:22:24.680 --> 00:22:30.279 Trump administration has made for its own reasons, has has 395 00:22:30.519 --> 00:22:34.559 said to Anthropic that there's a problem. They don't want 396 00:22:34.599 --> 00:22:38.319 it to be used in Department of War. We are 397 00:22:38.359 --> 00:22:42.960 where we are, well. I think at some point we 398 00:22:42.960 --> 00:22:46.160 should make a deal with Anthropic, get them back into 399 00:22:46.200 --> 00:22:50.160 the into the Department of War. But I think we 400 00:22:50.160 --> 00:22:52.359 should say in exchange, one of the things that we 401 00:22:52.400 --> 00:22:56.119