1 00:00:00,040 --> 00:00:03,000 The United States government has officially designated a leading 2 00:00:03,000 --> 00:00:06,720 American artificial intelligence company a supply chain risk to 3 00:00:06,720 --> 00:00:10,040 national security. Which is a label they usually 4 00:00:10,040 --> 00:00:13,400 save for foreign adversaries. We are talking about the exact 5 00:00:13,400 --> 00:00:16,520 same classification used for Huawei and ZTE. 6 00:00:16,800 --> 00:00:19,640 The Department of War has effectively placed a Do not 7 00:00:19,640 --> 00:00:22,360 touch order on Anthropic. That sounds like a lot more than 8 00:00:22,360 --> 00:00:24,080 just a cancelled contract. Oh, it is. 9 00:00:24,120 --> 00:00:27,600 It's an absolute blacklist. It prohibits military 10 00:00:27,600 --> 00:00:31,080 contractors, suppliers, basically any partners from 11 00:00:31,080 --> 00:00:33,120 conducting commercial activity with the company. 12 00:00:33,160 --> 00:00:36,240 If the government can blacklist a domestic company for refusing 13 00:00:36,240 --> 00:00:39,840 to remove safety guardrails from its software, does the private 14 00:00:39,840 --> 00:00:43,160 sector actually retain any control over how its technology 15 00:00:43,160 --> 00:00:46,120 is used in warfare? That is the multibillion dollar 16 00:00:46,120 --> 00:00:48,480 question here. And this whole thing didn't 17 00:00:48,480 --> 00:00:50,000 start in some corporate boardroom. 18 00:00:50,000 --> 00:00:53,760 It started with a specific military operation, specifically 19 00:00:53,760 --> 00:00:57,240 the raid in Caracas to capture Nicolas Maduro, right? 20 00:00:57,520 --> 00:00:59,840 Operation absolute resolve. Exactly. 21 00:01:00,080 --> 00:01:03,520 According to reports, the US military utilized Anthropic's 22 00:01:03,520 --> 00:01:07,080 model CLAWED during the planning and execution of this operation. 23 00:01:07,920 --> 00:01:11,280 Now Anthropic doesn't have a direct contract to sit in the 24 00:01:11,280 --> 00:01:14,800 situation room. The integration happened through 25 00:01:14,800 --> 00:01:17,040 Palantir. We should probably clarify 26 00:01:17,040 --> 00:01:20,000 Palantir's role here, because they aren't just, you know, a 27 00:01:20,000 --> 00:01:22,360 simple database company. No, no. 28 00:01:22,360 --> 00:01:25,080 Think of Palantir as as the operating system for Modern 29 00:01:25,080 --> 00:01:26,920 Warfare. They have this platform called 30 00:01:26,920 --> 00:01:30,880 the Maven Smart system. It ingests satellite imagery, 31 00:01:30,880 --> 00:01:34,800 drone feeds, intercepted communications, basically all 32 00:01:34,800 --> 00:01:37,800 the messy data of war. And it synthesizes it into a 33 00:01:37,800 --> 00:01:40,520 coherent picture for the commanders on the ground. 34 00:01:40,680 --> 00:01:42,040 Right. And they were running clawed 35 00:01:42,040 --> 00:01:44,960 inside that ecosystem to help process all that information. 36 00:01:44,960 --> 00:01:46,960 The mission itself was technically a success. 37 00:01:46,960 --> 00:01:49,400 I mean, the administration got what they wanted, but the 38 00:01:49,400 --> 00:01:51,200 friction actually started afterward. 39 00:01:51,200 --> 00:01:53,680 Yeah, it started because Anthropic asked a question. 40 00:01:54,080 --> 00:01:57,520 Following the operation, company leadership allegedly inquired 41 00:01:57,520 --> 00:02:00,560 whether their technology had been utilized in the raid, and 42 00:02:00,560 --> 00:02:03,360 that inquiry triggered an immediate alarm at the Pentagon. 43 00:02:03,480 --> 00:02:05,880 Why though? I mean, if I sell a product, 44 00:02:06,160 --> 00:02:09,039 asking how it was used seems pretty standard. 45 00:02:09,280 --> 00:02:12,080 Not when you were dealing with classified operations. 46 00:02:12,480 --> 00:02:17,080 But Anthropic felt they had to ask because of constitutional 47 00:02:17,120 --> 00:02:19,200 AI. Let's use plain English for 48 00:02:19,200 --> 00:02:21,400 that. What exactly is Constitutional 49 00:02:21,400 --> 00:02:23,760 AI? Sure, so most AI models are 50 00:02:23,760 --> 00:02:27,040 trained on massive amounts of Internet text and then just fine 51 00:02:27,040 --> 00:02:29,840 tune to be helpful. Anthropic takes a totally 52 00:02:29,840 --> 00:02:32,000 different approach. They give the model a 53 00:02:32,000 --> 00:02:35,200 constitution, a specific set of ethical principles it has to 54 00:02:35,200 --> 00:02:36,880 follow. Like rules against generating 55 00:02:36,880 --> 00:02:38,560 harm or things like that. Exactly. 56 00:02:38,560 --> 00:02:42,160 Rules like do not help kill people, do not violate privacy, 57 00:02:42,160 --> 00:02:45,720 do not create hate speech. The model essentially checks its 58 00:02:45,720 --> 00:02:48,440 own output against this Constitution before it even 59 00:02:48,440 --> 00:02:51,200 responds. So to know if the model actually 60 00:02:51,200 --> 00:02:54,240 followed its own constitution, the company needs to know what 61 00:02:54,240 --> 00:02:56,360 it was doing. Yes, they need a degree of 62 00:02:56,360 --> 00:02:58,960 transparency to ensure their ethical guidelines aren't being 63 00:02:58,960 --> 00:03:01,720 violated, and the Pentagon viewed this request for 64 00:03:01,720 --> 00:03:04,040 transparency as a massive liability. 65 00:03:04,040 --> 00:03:06,280 Because it's an operational security issue for them. 66 00:03:06,560 --> 00:03:08,800 Right. They saw a private vendor in San 67 00:03:08,800 --> 00:03:11,600 Francisco claiming they had the right to audit a military 68 00:03:11,600 --> 00:03:14,120 operation. To the generals, that just looks 69 00:03:14,120 --> 00:03:17,080 like a vendor trying to exercise veto power over national 70 00:03:17,080 --> 00:03:19,800 security decisions. They see a tool they bought and 71 00:03:19,800 --> 00:03:22,960 the tool maker wants to verify if it was used to hurt someone. 72 00:03:23,320 --> 00:03:26,320 And that friction LED directly to the ultimatum. 73 00:03:26,640 --> 00:03:30,480 The Pentagon, led by Secretary Pete Hegseth, issued a directive 74 00:03:30,480 --> 00:03:33,680 to all AI vendors. The rule was very simple. 75 00:03:34,360 --> 00:03:38,000 AI vendors must allow their tools to be used for all lawful 76 00:03:38,000 --> 00:03:40,160 purposes. All lawful purposes. 77 00:03:40,160 --> 00:03:44,160 That phrase does a lot of heavy lifting, but Anthropic refused 78 00:03:44,160 --> 00:03:46,960 to agree to that broad language. They did. 79 00:03:46,960 --> 00:03:50,480 They insisted on maintaining 2 very specific red lines in their 80 00:03:50,480 --> 00:03:53,480 contract, regardless of whether the government called it lawful 81 00:03:53,480 --> 00:03:54,920 or not. What were the 2 lines? 82 00:03:55,080 --> 00:03:58,200 1st, The AI cannot be used for mass domestic surveillance of 83 00:03:58,200 --> 00:04:00,920 American citizens. OK, that seems straightforward. 84 00:04:01,040 --> 00:04:03,080 And the second. Second, it cannot be used for 85 00:04:03,080 --> 00:04:05,280 fully autonomous weapons. Wait, hold on, let me just make 86 00:04:05,280 --> 00:04:07,640 sure I have this straight. What is the specific definition 87 00:04:07,640 --> 00:04:09,680 of an autonomous weapon in this context? 88 00:04:09,880 --> 00:04:13,000 We are talking about systems that select and engage targets 89 00:04:13,000 --> 00:04:16,600 without human intervention. In military terminology, it's 90 00:04:16,600 --> 00:04:18,959 lethal autonomous weapons systems. 91 00:04:19,040 --> 00:04:23,280 Meaning the software processes the sensor data, identifies a 92 00:04:23,280 --> 00:04:26,800 human being as a target, and authorizes the strike. 93 00:04:26,880 --> 00:04:29,680 Yes, all without a person ever pressing a button. 94 00:04:29,840 --> 00:04:32,840 So Anthropic basically said we will not build the software that 95 00:04:32,840 --> 00:04:34,280 decides who dies. Correct. 96 00:04:34,560 --> 00:04:37,240 The government's counter argument is that lawful use 97 00:04:37,240 --> 00:04:40,600 already covers those concerns. They argue that if an action is 98 00:04:40,600 --> 00:04:43,960 legal under the Constitution, a private company has no right to 99 00:04:43,960 --> 00:04:46,720 block the government from using its procured tools to execute 100 00:04:46,720 --> 00:04:49,160 it. But we if the Pentagon says 101 00:04:49,160 --> 00:04:51,960 lawful use is the only standard they will accept, and they 102 00:04:51,960 --> 00:04:54,720 refuse to sign a ban on autonomous weapons, are they 103 00:04:54,720 --> 00:04:58,560 admitting they plan to use them? That is the real Gray area, the 104 00:04:58,560 --> 00:05:01,400 Pentagon spokesperson stated. They have no interest in mass 105 00:05:01,400 --> 00:05:04,400 surveillance or autonomous weapons without human oversight 106 00:05:04,400 --> 00:05:06,320 right now. They actually called those fears 107 00:05:06,320 --> 00:05:08,800 fake. However, they refuse to sign a 108 00:05:08,800 --> 00:05:11,160 contract that explicitly forbade it. 109 00:05:11,400 --> 00:05:14,320 Exactly. Their position is entirely about 110 00:05:14,320 --> 00:05:18,200 operational flexibility. They on a you they cannot have a 111 00:05:18,200 --> 00:05:20,680 Terms of Service agreement overriding a commander's 112 00:05:20,680 --> 00:05:22,960 decision in the field 5 or 10 years from now. 113 00:05:23,440 --> 00:05:26,120 And Anthropic is saying the technology just isn't safe 114 00:05:26,120 --> 00:05:28,480 enough for that. That is their core argument. 115 00:05:28,480 --> 00:05:31,200 They point out that current large language models 116 00:05:31,200 --> 00:05:32,880 hallucinate. They make up facts. 117 00:05:33,160 --> 00:05:35,280 They misinterpret context entirely. 118 00:05:35,280 --> 00:05:37,480 Which is a known issue across the industry. 119 00:05:37,480 --> 00:05:40,560 Yes. So Anthropic believes relying on 120 00:05:40,560 --> 00:05:44,120 them for autonomous killing is reckless and mass surveillance 121 00:05:44,120 --> 00:05:47,560 violates fundamental rights. They view these strictly as 122 00:05:47,560 --> 00:05:50,080 safety issues. While the administration argues 123 00:05:50,080 --> 00:05:53,440 that lawful is the only standard that matters and what corporate 124 00:05:53,440 --> 00:05:55,840 policies cannot constrain the military? 125 00:05:55,840 --> 00:05:58,200 Exactly. OK, let's pause here to reset 126 00:05:58,200 --> 00:06:00,600 the pace. Usually when a company says no 127 00:06:00,600 --> 00:06:02,880 to the government, they just lose the contract. 128 00:06:02,960 --> 00:06:04,120 They don't get the money. Right. 129 00:06:04,120 --> 00:06:06,880 You walk away, you lose the revenue, maybe your stock dips a 130 00:06:06,880 --> 00:06:10,120 bit, but you just continue doing business with everyone else. 131 00:06:10,120 --> 00:06:12,800 But the administration didn't just cancel the contract in this 132 00:06:12,800 --> 00:06:16,040 case, they went much further. They absolutely did. 133 00:06:16,280 --> 00:06:19,640 President Trump ordered all federal agencies to immediately 134 00:06:19,640 --> 00:06:23,200 cease using Anthropics technology, and then Secretary 135 00:06:23,200 --> 00:06:25,840 Hegseth followed up by designating the company a supply 136 00:06:25,840 --> 00:06:29,320 chain risk under 10US Code Section 3252. 137 00:06:29,480 --> 00:06:33,640 I want to focus on that specific statute, NUS Code section 3252. 138 00:06:33,640 --> 00:06:36,200 What does that actually mean? This is a statute designed 139 00:06:36,200 --> 00:06:39,560 specifically to prevent espionage and sabotage by 140 00:06:39,560 --> 00:06:43,080 foreign enemies. It is the exact legal tool used 141 00:06:43,080 --> 00:06:47,000 to ban Huawei and ZTE because of fears they were funneling data 142 00:06:47,000 --> 00:06:48,600 straight to the Chinese Communist Party. 143 00:06:48,800 --> 00:06:51,520 So it implies the company is an active threat to the integrity 144 00:06:51,520 --> 00:06:54,320 of the nation's defense. Yes, and applying it to an 145 00:06:54,320 --> 00:06:56,760 American firm for a contract dispute is entirely 146 00:06:56,760 --> 00:06:58,720 unprecedented. It acts as a corporate death 147 00:06:58,720 --> 00:07:00,200 penalty in the federal sector because. 148 00:07:00,200 --> 00:07:02,640 It's not just the Pentagon that has to top using them now. 149 00:07:02,720 --> 00:07:04,800 Exactly. It effectively forces any 150 00:07:04,800 --> 00:07:06,480 company that wants to work with the military. 151 00:07:06,480 --> 00:07:10,080 Companies like Boeing, Lockheed Martin or Palantir to strip 152 00:07:10,080 --> 00:07:12,560 Anthropic out of their own internal workflows. 153 00:07:12,560 --> 00:07:15,800 So if I'm an engineer at Lockheed Martin and I use Claude 154 00:07:15,840 --> 00:07:19,000 to write code or summarize technical documents for my daily 155 00:07:19,000 --> 00:07:21,320 work. You are now a liability. 156 00:07:21,960 --> 00:07:26,160 Lockheed Martin cannot risk its massive government contracts by 157 00:07:26,160 --> 00:07:29,640 harboring a designated supply chain risk in its software 158 00:07:29,640 --> 00:07:31,760 stack. They have to RIP it completely 159 00:07:31,760 --> 00:07:34,080 out. There's a massive contradiction 160 00:07:34,080 --> 00:07:37,200 in the order itself though. It claims Anthropic is a 161 00:07:37,200 --> 00:07:40,040 security risk, but simultaneously mandates they 162 00:07:40,040 --> 00:07:43,440 continue providing services for a six month transition period. 163 00:07:43,600 --> 00:07:46,760 Which really highlights that this isn't about espionage. 164 00:07:47,080 --> 00:07:50,800 Think about it, if Huawei was actively spying on the Pentagon, 165 00:07:50,800 --> 00:07:54,040 you wouldn't say, OK, keep the routers plugged in for six more 166 00:07:54,040 --> 00:07:55,400 months while we find a replacement. 167 00:07:55,440 --> 00:07:57,080 Right, you would cut the line immediately. 168 00:07:57,080 --> 00:07:59,880 You would sever it that second. So that tells us they need the 169 00:07:59,880 --> 00:08:01,760 tech. They absolutely need the tech, 170 00:08:01,760 --> 00:08:03,520 they just don't want the rules that come with it. 171 00:08:03,720 --> 00:08:06,200 Keeping them on for six months admits that the government is 172 00:08:06,200 --> 00:08:08,000 highly dependent on these models. 173 00:08:08,120 --> 00:08:10,040 So this designation is just a power play. 174 00:08:10,040 --> 00:08:13,360 Completely. It threatens Anthropics entire 175 00:08:13,360 --> 00:08:16,400 enterprise ecosystem and its projected IPO. 176 00:08:16,720 --> 00:08:18,880 It signals to the rest of Silicon Valley that 177 00:08:18,880 --> 00:08:21,560 noncompliance comes with an existential cost. 178 00:08:21,640 --> 00:08:24,480 And almost immediately, another company stepped in to take 179 00:08:24,480 --> 00:08:27,800 advantage. Yes, hours after the ban on 180 00:08:27,800 --> 00:08:32,280 Anthropic was announced, open AICEO, Sam Altman announced a 181 00:08:32,280 --> 00:08:35,360 new agreement. The pivot was instant open AI 182 00:08:35,360 --> 00:08:38,000 stepped right into the vacuum. They announced an agreement to 183 00:08:38,000 --> 00:08:41,480 deploy open AI models on the Department of War's classified 184 00:08:41,480 --> 00:08:43,600 networks. Here's the part that is really 185 00:08:43,600 --> 00:08:46,440 confusing to me. Open AI claims they have the 186 00:08:46,440 --> 00:08:50,040 exact same red lines as Anthropic regarding surveillance 187 00:08:50,040 --> 00:08:53,160 and autonomous weapons. Publicly, yes, their safety 188 00:08:53,160 --> 00:08:55,120 guidelines prohibit the exact same things. 189 00:08:55,120 --> 00:08:59,080 So if they have the same red lines, why did the Pentagon ban 190 00:08:59,080 --> 00:09:02,560 one and sign the other? Why is Anthropic a security risk 191 00:09:02,560 --> 00:09:05,760 and Open AIA trusted partner? It comes down to the mechanism 192 00:09:05,760 --> 00:09:08,240 of enforcement. Anthropic wanted strict 193 00:09:08,240 --> 00:09:11,520 contractual prohibitions. They wanted a written veto in 194 00:09:11,520 --> 00:09:13,680 the deal that they could enforce legally. 195 00:09:14,080 --> 00:09:16,720 Open AI adopted A layered approach instead. 196 00:09:16,720 --> 00:09:19,360 What does a layered approach actually mean in this context? 197 00:09:19,360 --> 00:09:22,840 It means they are relying on a cloud only deployment structure 198 00:09:22,960 --> 00:09:25,320 and the government's own interpretation of the law. 199 00:09:25,560 --> 00:09:28,360 Open AI agreed to the All Lawful Use standard. 200 00:09:28,440 --> 00:09:31,440 So they are effectively betting that the government won't define 201 00:09:31,440 --> 00:09:34,000 mass surveillance or autonomous weapons as lawful. 202 00:09:34,280 --> 00:09:35,760 Correct. Altman stated that the 203 00:09:35,760 --> 00:09:38,720 department agrees domestic surveillance is illegal, so they 204 00:09:38,720 --> 00:09:42,480 didn't need to fight over it. Open AI integrated their safety 205 00:09:42,480 --> 00:09:45,320 stack into the deployment rather than demanding external 206 00:09:45,320 --> 00:09:48,720 oversight of specific missions. So Anthropic wanted the 207 00:09:48,720 --> 00:09:52,000 government to sign a paper saying we promise not to do X, 208 00:09:52,600 --> 00:09:55,320 Open AI said. We know X is illegal, so we 209 00:09:55,320 --> 00:09:57,680 don't need you to sign a paper. Essentially, yes. 210 00:09:58,200 --> 00:10:00,400 They gave the Pentagon the lawful use language they 211 00:10:00,400 --> 00:10:03,200 demanded while assuring the public that their technical 212 00:10:03,200 --> 00:10:05,120 architecture prevents the bad stuff. 213 00:10:05,200 --> 00:10:08,000 It sounds like Open AI gave the Pentagon an optical win. 214 00:10:08,040 --> 00:10:11,280 And a contractual 1 The Pentagon gets to say we don't bow to 215 00:10:11,280 --> 00:10:14,720 terms of service and Open AI gets the massive contract. 216 00:10:14,880 --> 00:10:18,360 During this entire fallout, the administration publicly attacked 217 00:10:18,360 --> 00:10:21,560 Anthropic as a radical left and woke company. 218 00:10:21,960 --> 00:10:24,080 The record really contradicts that characterization. 219 00:10:24,080 --> 00:10:27,240 Though Anthropic is funded by major corporate players like 220 00:10:27,240 --> 00:10:31,640 Amazon and Google, their CEO Dario Amadei has publicly stated 221 00:10:31,640 --> 00:10:35,040 that AI is existentially important for National Defense. 222 00:10:35,040 --> 00:10:37,400 He's not exactly a pacifist. Far from it. 223 00:10:37,760 --> 00:10:41,160 He has explicitly stated he supports helping the US military 224 00:10:41,160 --> 00:10:45,360 defeat autocratic adversaries. He even admitted in interviews 225 00:10:45,520 --> 00:10:48,800 that autonomous weapons might eventually be necessary. 226 00:10:48,800 --> 00:10:51,320 Wait, really? So what is his actual objection? 227 00:10:51,440 --> 00:10:54,440 His objection is entirely that the current technology isn't 228 00:10:54,440 --> 00:10:57,920 safe enough yet. He argues that LLMS hallucinate 229 00:10:57,920 --> 00:11:01,200 and make unpredictable errors and therefore shouldn't be 230 00:11:01,200 --> 00:11:04,160 making keel decisions today. That seems like a technical 231 00:11:04,160 --> 00:11:05,920 argument, not a political one. Exactly. 232 00:11:06,000 --> 00:11:08,880 It suggests the woke label is just political cover. 233 00:11:09,120 --> 00:11:12,200 The real issue here is the sovereign AI doctrine. 234 00:11:12,200 --> 00:11:14,480 Meaning what? The administration is 235 00:11:14,480 --> 00:11:17,720 establishing that the state, not private labs, is the final 236 00:11:17,720 --> 00:11:19,680 arbiter of how technology is deployed. 237 00:11:20,040 --> 00:11:22,960 It is a fundamental shift from a safety first model to a 238 00:11:22,960 --> 00:11:25,440 deployment first mandate. They are worried about falling 239 00:11:25,440 --> 00:11:27,040 behind. They are looking straight at 240 00:11:27,040 --> 00:11:29,040 China. The argument is that if American 241 00:11:29,040 --> 00:11:31,520 companies are permitted to constrain the military with 242 00:11:31,520 --> 00:11:34,680 safety guardrails, the United States will lose the advantage 243 00:11:34,680 --> 00:11:37,920 to an adversary that operates without any such constraints. 244 00:11:38,040 --> 00:11:41,040 They want the tech to be totally subservient to national security 245 00:11:41,040 --> 00:11:44,800 objectives. This supply chain risk, does it 246 00:11:44,800 --> 00:11:48,040 stop at the Pentagon? No, the General Services 247 00:11:48,040 --> 00:11:51,120 Administration has already terminated Anthropics One Gov 248 00:11:51,120 --> 00:11:54,040 deal. That means agencies totally 249 00:11:54,040 --> 00:11:58,120 outside the military, like the Department of Energy or the EPA, 250 00:11:58,160 --> 00:12:01,520 might have to stop using Quad. It it effectively chills the 251 00:12:01,520 --> 00:12:04,480 entire market for them. It forces every company to make 252 00:12:04,480 --> 00:12:06,560 a hard choice. If you want to do business with 253 00:12:06,600 --> 00:12:09,960 the US government in any capacity, you cannot use the 254 00:12:09,960 --> 00:12:12,960 Risk software. It consolidated power strictly 255 00:12:12,960 --> 00:12:15,320 around the vendors who agree to the government's terms. 256 00:12:15,760 --> 00:12:18,440 To wrap this up, Anthropic held the line on their terms of 257 00:12:18,440 --> 00:12:22,320 service and was designated a National Security Risk Open AI 258 00:12:22,320 --> 00:12:24,760 aligned with the government's legal framework and became the 259 00:12:24,760 --> 00:12:26,560 primary partner for the military. 260 00:12:26,760 --> 00:12:29,520 The main take away here is that the Silicon Valley consensus, 261 00:12:29,520 --> 00:12:32,400 this era where companies felt they could dictate ethical terms 262 00:12:32,400 --> 00:12:34,440 to the government, is effectively dead. 263 00:12:34,840 --> 00:12:37,360 The government has demonstrated it is entirely willing to use 264 00:12:37,360 --> 00:12:40,640 the full weight of executive power to ensure AI companies are 265 00:12:40,640 --> 00:12:42,880 subservient to national security objectives. 266 00:12:43,560 --> 00:12:47,480 If you're not subscribed yet, take a second and hit follow on. 267 00:12:47,480 --> 00:12:49,880 Whatever a you're using, it helps us keep making this. 268 00:12:49,880 --> 00:12:50,880 We appreciate you being here.