1 00:00:00,080 --> 00:00:04,080 GPT 5.2 didn't just process data this week, it actually 2 00:00:04,080 --> 00:00:08,880 conjectured a brand new formula for single minus gluon tree 3 00:00:09,000 --> 00:00:11,840 amplitudes. Which, if you aren't a 4 00:00:11,840 --> 00:00:14,560 theoretical physicist, is a problem that human 5 00:00:14,560 --> 00:00:17,600 mathematicians generally describe as incredibly long and 6 00:00:17,600 --> 00:00:19,600 painful. Yeah, we are talking about what 7 00:00:19,600 --> 00:00:24,280 is typically 1/4 page of really dense messy algebra. 8 00:00:24,800 --> 00:00:28,080 And the AI identified a hidden pattern that simplified all of 9 00:00:28,080 --> 00:00:31,040 that into a single, elegant product structure. 10 00:00:31,120 --> 00:00:33,960 And that is massively significant because this goes 11 00:00:33,960 --> 00:00:37,520 way beyond a calculator doing arithmetic faster than a human 12 00:00:37,760 --> 00:00:39,360 right. We are looking at a system 13 00:00:39,440 --> 00:00:42,920 analyzing a mess of complexity and spotting A symmetry that 14 00:00:42,920 --> 00:00:45,040 human experts had missed for decades, right? 15 00:00:45,040 --> 00:00:47,920 It is the difference between simply computing a result and 16 00:00:47,920 --> 00:00:49,800 actually understanding the architecture of the physics 17 00:00:49,800 --> 00:00:51,480 problem itself. It proves we are building 18 00:00:51,480 --> 00:00:54,320 genuine collaborators now. But while the science side is 19 00:00:54,320 --> 00:00:57,360 having this massive Eureka moment, the way these AI 20 00:00:57,360 --> 00:00:59,600 companies operate is shifting just as drastically. 21 00:00:59,680 --> 00:01:02,480 We are moving entirely away from the era of chat bots where you 22 00:01:02,480 --> 00:01:03,840 type a question and get an answer. 23 00:01:04,000 --> 00:01:06,960 We are entering the era of agents and Co workers that 24 00:01:06,960 --> 00:01:10,280 actually execute tasks. That distinction really matters. 25 00:01:10,520 --> 00:01:15,000 A chat bot talks, an agent acts, and right now the industry has 26 00:01:15,000 --> 00:01:18,200 split into two very distinct battles to make that happen. 27 00:01:18,640 --> 00:01:21,000 There is one battle for raw speed and hardware 28 00:01:21,000 --> 00:01:24,760 infrastructure, and another for deep autonomous execution. 29 00:01:24,920 --> 00:01:27,120 Exactly. Which brings us to the core 30 00:01:27,120 --> 00:01:28,880 question we are going to be exploring today. 31 00:01:29,440 --> 00:01:32,560 In a week where AI is proving theorems and learning to drive, 32 00:01:32,560 --> 00:01:35,720 our computers, are the safety guardrails crumbling under the 33 00:01:35,720 --> 00:01:38,120 pressure to win massive government and enterprise 34 00:01:38,120 --> 00:01:40,440 contracts? It is the defining tension of 35 00:01:40,480 --> 00:01:43,400 everything happening right now. And the autonomous side of that 36 00:01:43,400 --> 00:01:45,760 battle just got a huge injection of talent. 37 00:01:46,080 --> 00:01:49,600 As of this morning, February 25th, 2026, Anthropic has 38 00:01:49,600 --> 00:01:52,680 officially acquired Vercept. This is a major signal flare. 39 00:01:53,040 --> 00:01:56,040 Vercept was a Seattle based startup and they were founded by 40 00:01:56,040 --> 00:01:59,120 some real heavy hitters from the Allen Institute for AI. 41 00:01:59,720 --> 00:02:02,240 They had built this desktop agent called V. 42 00:02:02,440 --> 00:02:06,640 I've actually seen demos of V It is a creepy and impressive and 43 00:02:06,640 --> 00:02:08,479 equal measure. It doesn't just read the 44 00:02:08,479 --> 00:02:11,240 underlying text code, it physically sees the screen 45 00:02:11,240 --> 00:02:13,360 elements. Right, the visual component is 46 00:02:13,360 --> 00:02:15,640 the main hurdle everyone has been trying to clear. 47 00:02:15,920 --> 00:02:18,320 Yeah, for a long time. If you wanted an AI system to 48 00:02:18,320 --> 00:02:21,200 use a computer, you had to hook it up to an API. 49 00:02:21,320 --> 00:02:24,400 That is essentially a special backdoor that lets the software 50 00:02:24,640 --> 00:02:28,120 talk directly to other software. But the real world is incredibly 51 00:02:28,120 --> 00:02:30,200 messy. Most software out there doesn't 52 00:02:30,200 --> 00:02:32,040 have clean API's. Exactly. 53 00:02:32,040 --> 00:02:35,560 So VAI interacts directly with the graphical user interface. 54 00:02:35,840 --> 00:02:37,560 The actual pixels on the monitor. 55 00:02:37,880 --> 00:02:39,680 It looks at the screen just like you do. 56 00:02:40,400 --> 00:02:44,120 It identifies that a specific Gray rectangle is a submit 57 00:02:44,120 --> 00:02:47,520 button and a specific white box is a text field. 58 00:02:47,840 --> 00:02:51,400 Which sounds so simple to us. It does, but for a machine it is 59 00:02:51,400 --> 00:02:54,360 incredibly difficult. It knows the grounding problem. 60 00:02:54,360 --> 00:02:57,440 You have to perfectly map pixel coordinates to semantic actions. 61 00:02:57,720 --> 00:03:00,560 And this acquisition perfectly explains the sudden jump in 62 00:03:00,560 --> 00:03:03,720 Anthropics performance numbers. We saw the new benchmarks for 63 00:03:03,720 --> 00:03:07,000 Claude Sonnet 4.6 this week, specifically looking at AUS 64 00:03:07,000 --> 00:03:08,040 World. Yes. 65 00:03:08,560 --> 00:03:11,640 For context, the Alice World benchmark is the standard test 66 00:03:11,640 --> 00:03:14,840 for measuring how well an AI can navigate an operating system. 67 00:03:15,480 --> 00:03:18,960 The test asks things like can it open a spreadsheet, copy a 68 00:03:18,960 --> 00:03:22,360 specific cell, open a web browser, paste that date into a 69 00:03:22,360 --> 00:03:24,800 form and hit enter. Real computer use. 70 00:03:25,000 --> 00:03:28,880 Yes, and in late 2024, the best models in the world were scoring 71 00:03:28,880 --> 00:03:32,320 under 15% on that benchmark. Which is functionally useless 72 00:03:32,320 --> 00:03:34,600 for a user. You would spend more time 73 00:03:34,600 --> 00:03:37,440 correcting its mistakes than the time it theoretically saves you. 74 00:03:37,760 --> 00:03:42,560 A 15% success rate is basically a toy, but Sonnet 4.6, which is 75 00:03:42,560 --> 00:03:45,400 clearly integrating this new Recept tech, is now hitting 76 00:03:45,480 --> 00:03:49,040 72.5%. That crosses a massive threshold 77 00:03:49,160 --> 00:03:51,920 at over 72%. You can actually walk away from 78 00:03:51,920 --> 00:03:54,480 your desk and let the agent run. You can trust it with a multi 79 00:03:54,480 --> 00:03:57,040 step workflow. That is entirely why Antropic 80 00:03:57,040 --> 00:03:58,600 bought them. They aren't interested in 81 00:03:58,600 --> 00:04:01,200 keeping the Vibe brand alive. They are integrating the team of 82 00:04:01,200 --> 00:04:04,600 about 20 engineers to make Claude fully capable of direct 83 00:04:04,600 --> 00:04:06,920 computer use. They want an AI that writes the 84 00:04:06,920 --> 00:04:09,800 e-mail for you, opens your client, attaches the PDF and 85 00:04:09,800 --> 00:04:12,880 click send. So Anthropic is heavily betting 86 00:04:12,880 --> 00:04:16,959 on the smart autonomous agent that navigates a messy desktop 87 00:04:16,959 --> 00:04:19,959 environment. Open AI, however, seems to be 88 00:04:19,959 --> 00:04:21,760 betting on something else entirely. 89 00:04:21,760 --> 00:04:26,280 This week they released GPT 5.3 codecs Spark. 90 00:04:26,440 --> 00:04:28,000 Spark is the operative word there. 91 00:04:28,160 --> 00:04:31,560 This new model is entirely obsessed with latency. 92 00:04:31,600 --> 00:04:34,440 And they are achieving this incredible speed through 93 00:04:34,440 --> 00:04:37,440 specialized hardware, right? This goes beyond standard 94 00:04:37,440 --> 00:04:39,840 software optimization. It is absolutely a hardware 95 00:04:39,840 --> 00:04:41,960 play. Spark is running on Cerebra's 96 00:04:41,960 --> 00:04:44,560 Wafer Scale Engine 3. I really need you to visualize 97 00:04:44,560 --> 00:04:47,120 this for a second because I saw a photograph of this chip 98 00:04:47,120 --> 00:04:48,960 recently. Standard computer chips are 99 00:04:48,960 --> 00:04:50,600 small. They are roughly the size of a 100 00:04:50,600 --> 00:04:53,080 postage stamp. This Cerebra's chip is the size 101 00:04:53,080 --> 00:04:55,120 of a dinner plate. Yeah, it's the size of a giant 102 00:04:55,120 --> 00:04:58,160 pancake. It is an entire uncut silicon 103 00:04:58,160 --> 00:05:01,200 wafer, usually in standard chip manufacturing where you take a 104 00:05:01,200 --> 00:05:04,360 silicon wafer and cut it into hundreds of tiny individual 105 00:05:04,360 --> 00:05:06,480 chips. So reverse just keeps the whole 106 00:05:06,480 --> 00:05:08,600 thing intact as one massive processor. 107 00:05:08,600 --> 00:05:11,400 Why does keeping it intact matter so much for speed? 108 00:05:11,520 --> 00:05:14,400 Because in a traditional setup, you have your memory sitting on 109 00:05:14,400 --> 00:05:17,240 one physical stick and the processor sitting on another. 110 00:05:18,040 --> 00:05:21,560 Data has to physically travel back and forth through wires to 111 00:05:21,560 --> 00:05:24,360 compute anything. Even if we get the speed of 112 00:05:24,360 --> 00:05:27,760 light, that travel takes time and consumes a lot of energy. 113 00:05:27,920 --> 00:05:31,760 Right on the wafer scale engine, the memory and the compute cores 114 00:05:31,760 --> 00:05:35,000 are right next to each other on the exact same piece of silicon. 115 00:05:36,120 --> 00:05:39,040 The data transfer delay is effectively eliminated. 116 00:05:39,040 --> 00:05:41,600 It is nearly instant. And the practical result of that 117 00:05:41,600 --> 00:05:44,520 architecture is over 1000 tokens per second. 118 00:05:44,800 --> 00:05:47,200 Well over 1000. Wait, hold on, let's back up a 119 00:05:47,200 --> 00:05:49,040 second. To put that in perspective, a 120 00:05:49,040 --> 00:05:53,040 human being reads roughly 5 words a second, generating 1000 121 00:05:53,040 --> 00:05:55,320 tokens. A second means a massive wall of 122 00:05:55,320 --> 00:05:57,320 text appears instantly on your screen. 123 00:05:57,480 --> 00:05:59,760 You can't even begin to read it as it generates. 124 00:05:59,760 --> 00:06:02,200 You absolutely cannot. But for writing code, which is 125 00:06:02,200 --> 00:06:05,640 exactly what Spark is designed for, it fundamentally changes 126 00:06:05,640 --> 00:06:08,480 the texture of the work. It is essentially 15 times 127 00:06:08,480 --> 00:06:10,480 faster at coding than their standard model. 128 00:06:10,600 --> 00:06:13,000 They are marketing this as conversational coding. 129 00:06:13,320 --> 00:06:15,760 Think about how you use a standard chat bot right now. 130 00:06:16,240 --> 00:06:18,920 You type a prompt. You wait maybe 10 seconds. 131 00:06:19,280 --> 00:06:21,680 The code appears block by block. You read it. 132 00:06:22,320 --> 00:06:25,400 It feels very much like sending a letter and waiting for a reply 133 00:06:25,400 --> 00:06:27,440 in the mail. A turn based interaction. 134 00:06:27,560 --> 00:06:30,640 Right, with Spark the code generates so fast you can 135 00:06:30,640 --> 00:06:32,880 interrupt at mid thought. You can see it going down the 136 00:06:32,880 --> 00:06:36,280 wrong logical path in line three of a function and just stop it 137 00:06:36,280 --> 00:06:37,840 instantly. You correct it on the fly. 138 00:06:37,840 --> 00:06:41,120 So it feels more like jamming with a musician in a studio. 139 00:06:41,360 --> 00:06:44,800 That is the perfect analogy. It creates a tight real time 140 00:06:44,800 --> 00:06:47,440 feedback loop. But there is a serious trade off 141 00:06:47,440 --> 00:06:49,120 here. You do not get that kind of 142 00:06:49,280 --> 00:06:51,560 speed for free. The model is less rigorous. 143 00:06:51,560 --> 00:06:54,480 Exactly. Spark is completely latency 144 00:06:54,480 --> 00:06:57,240 first. If you look at the SWE Bench Pro 145 00:06:57,240 --> 00:06:59,880 scores, which is the premier software engineering benchmark, 146 00:07:00,080 --> 00:07:03,840 Spark scores significantly lower than the full GPT 5.3 codecs 147 00:07:03,840 --> 00:07:06,400 model. And perhaps more worryingly, 148 00:07:06,760 --> 00:07:10,400 Open AI explicitly states it is not rated for high capability 149 00:07:10,400 --> 00:07:12,880 cybersecurity work. So we have a purposeful division 150 00:07:12,880 --> 00:07:15,240 now. If you want the AI to discover 151 00:07:15,240 --> 00:07:19,160 the gluon tree amplitude formula, you wait patiently for 152 00:07:19,160 --> 00:07:22,920 the slow deep thinking model. If you want to hack together a 153 00:07:22,920 --> 00:07:26,640 website infrastructure in 10 minutes, you use Spark for fast 154 00:07:26,640 --> 00:07:29,080 execution. It is a deliberate split between 155 00:07:29,080 --> 00:07:32,280 deep reasoning and velocity. Speaking of deep reasoning and 156 00:07:32,280 --> 00:07:34,800 that physics discovery we mentioned at the start, we sort 157 00:07:34,800 --> 00:07:37,160 of glossed over where that actually took place. 158 00:07:37,520 --> 00:07:40,640 That discovery did not happen in a standard chat window on a 159 00:07:40,640 --> 00:07:42,800 browser. It happened inside Prism. 160 00:07:42,800 --> 00:07:45,080 Open AI Prism. Yeah, this is their brand new 161 00:07:45,080 --> 00:07:48,720 workspace designed specifically for scientists, and it is 162 00:07:48,720 --> 00:07:51,560 fascinating because it attacks the actual workflow of doing 163 00:07:51,560 --> 00:07:54,080 science. It is a fully Latex native 164 00:07:54,080 --> 00:07:56,880 writing environment. Latex is the complex typesetting 165 00:07:56,880 --> 00:08:00,160 system that pretty much every physicist and mathematician uses 166 00:08:00,160 --> 00:08:03,800 to write their formal papers. Prism integrates the AI directly 167 00:08:03,800 --> 00:08:06,560 into that source mode. Plenty of standard text editors 168 00:08:06,560 --> 00:08:10,440 have AI plugins these days, but Prism is completely different 169 00:08:10,440 --> 00:08:12,920 because of the context window and the deep integration. 170 00:08:13,160 --> 00:08:16,480 Prism reads the entire project structure simultaneously. 171 00:08:16,760 --> 00:08:20,560 It sees your equations, your citations, your raw empirical 172 00:08:20,560 --> 00:08:22,720 data files, and all of your figures. 173 00:08:22,720 --> 00:08:26,080 So if I go in and tweak a variable in my core equation on 174 00:08:26,080 --> 00:08:30,160 page 2, Prism automatically knows to update the resulting 175 00:08:30,160 --> 00:08:34,200 graph in figure 3 on page 10. Yes, it validates if your 176 00:08:34,200 --> 00:08:37,120 empirical results actually match your theoretical model. 177 00:08:37,360 --> 00:08:40,120 It can check all your citations against the actual text of the 178 00:08:40,120 --> 00:08:42,559 reference papers to ensure you are quoting them correctly. 179 00:08:43,039 --> 00:08:45,880 It is doing the tedious grunt work of consistency that usually 180 00:08:45,880 --> 00:08:48,640 drives researchers crazy. And I asked about the business 181 00:08:48,640 --> 00:08:51,680 model earlier because I noticed they are offering this entirely 182 00:08:51,680 --> 00:08:54,520 for free for personal accounts. That is the classic Silicon 183 00:08:54,520 --> 00:08:56,880 Valley play. They want to become the default 184 00:08:56,880 --> 00:08:58,800 infrastructure for scientific discovery. 185 00:08:59,440 --> 00:09:02,000 Right now, a scientist might use Overly for collaborative 186 00:09:02,000 --> 00:09:06,040 writing, Zotero to manage their citations, and Python for data 187 00:09:06,040 --> 00:09:08,600 analysis. Prism is designed to replace all 188 00:09:08,600 --> 00:09:11,760 of those fragmented tools. They want the next Nobel Prize 189 00:09:11,760 --> 00:09:14,800 winning discovery to happen natively inside an open AI 190 00:09:14,840 --> 00:09:16,920 interface. Discovery is an incredibly 191 00:09:16,920 --> 00:09:19,640 valuable commodity. If you own the tool where the 192 00:09:19,640 --> 00:09:22,600 science happens, you get to see where human knowledge is going 193 00:09:22,600 --> 00:09:25,600 before anyone else does. Which brings us directly to the 194 00:09:25,600 --> 00:09:29,720 money, because whether it is via navigating a messy desktop or 195 00:09:29,720 --> 00:09:33,160 Prism writing a theoretical physics paper, this technology 196 00:09:33,160 --> 00:09:37,200 has to be monetized and simply selling an API key to developers 197 00:09:37,200 --> 00:09:39,840 isn't cutting it for these massive valuations anymore. 198 00:09:39,960 --> 00:09:43,000 We are seeing the rise of the true AI Co worker and the 199 00:09:43,000 --> 00:09:45,680 massive consulting army is required to install them. 200 00:09:45,680 --> 00:09:48,480 Open AI refers to these as frontier alliances. 201 00:09:49,000 --> 00:09:50,760 They have realized a harsh truth. 202 00:09:51,400 --> 00:09:55,160 You cannot just hand a Fortune 500 company access to a super 203 00:09:55,160 --> 00:09:57,920 intelligent model and expect them to magically become more 204 00:09:57,920 --> 00:09:59,800 productive. The companies literally do not 205 00:09:59,800 --> 00:10:01,800 know how to use it. They have no idea how to wire it 206 00:10:01,800 --> 00:10:05,000 into their legacy systems, so Open AI has officially partnered 207 00:10:05,000 --> 00:10:08,520 with McKenzie, BCG, Accenture, and Cap Gemini. 208 00:10:08,520 --> 00:10:11,040 These are the massive consulting firms you traditionally hire 209 00:10:11,040 --> 00:10:14,080 when you want to fire half your staff and completely restructure 210 00:10:14,080 --> 00:10:16,840 the remaining half. Or, phrase more charitably, 211 00:10:17,200 --> 00:10:20,040 they're the people you hire when you need to redesign your 212 00:10:20,040 --> 00:10:21,840 organization's central nerve, the system. 213 00:10:22,680 --> 00:10:25,280 These consulting firms are wiring the frontier platform 214 00:10:25,280 --> 00:10:28,840 directly into massive corporate data warehouses and customer 215 00:10:28,840 --> 00:10:32,320 relationship management systems. They are actively redesigning 216 00:10:32,320 --> 00:10:35,120 organizational workflows to accommodate autonomous agents. 217 00:10:35,120 --> 00:10:37,080 It is essentially organizational surgery. 218 00:10:37,160 --> 00:10:38,840 And it is very expensive surgery. 219 00:10:39,120 --> 00:10:41,800 Entropic is playing this enterprise game just as hard 220 00:10:41,800 --> 00:10:44,360 right now. They are currently hitting a $14 221 00:10:44,360 --> 00:10:48,840 billion revenue run rate, and to fuel that massive infrastructure 222 00:10:48,840 --> 00:10:51,880 and enterprise expansion, they just closed their Series G 223 00:10:51,880 --> 00:10:53,600 funding round. The number on that round was 224 00:10:53,600 --> 00:10:57,240 staggering, $30 billion. $30 billion in cash. 225 00:10:57,800 --> 00:11:01,720 That completely values Anthropic at $380 billion. 226 00:11:01,760 --> 00:11:04,800 That is an immense, almost incomprehensible amount of 227 00:11:04,800 --> 00:11:07,760 capital. But here is where the tension we 228 00:11:07,760 --> 00:11:09,520 talked about earlier really surfaces. 229 00:11:09,720 --> 00:11:13,200 We have $30 billion funding rounds, we have heavy military 230 00:11:13,200 --> 00:11:15,560 interest, and we have these massive consulting armies 231 00:11:15,560 --> 00:11:18,040 deploying agents. What happens to the original 232 00:11:18,040 --> 00:11:20,280 safety mission? Enthropic was founded 233 00:11:20,280 --> 00:11:23,520 specifically by people leaving Open AI to be the definitive 234 00:11:23,520 --> 00:11:26,000 safety company. That core mission is severely 235 00:11:26,000 --> 00:11:27,680 colliding with reality right now. 236 00:11:27,960 --> 00:11:31,760 Just yesterday, on February 24th, Anthropic released version 237 00:11:31,760 --> 00:11:35,160 3 Point O of their Responsible Scaling Policy, or RSPI. 238 00:11:35,320 --> 00:11:36,800 Read through that document last night. 239 00:11:36,800 --> 00:11:39,560 There is a very specific, very controversial change in the 240 00:11:39,560 --> 00:11:41,400 language. In the previous versions of the 241 00:11:41,400 --> 00:11:44,120 RSP, Anthropic had a hard written commitment. 242 00:11:44,760 --> 00:11:46,760 They stated they would unilaterally pause all 243 00:11:46,760 --> 00:11:49,400 development if they couldn't meet certain strict safety 244 00:11:49,400 --> 00:11:51,520 measures. If a model was deemed too 245 00:11:51,520 --> 00:11:54,200 dangerous or autonomous to contain, they would stop 246 00:11:54,200 --> 00:11:55,400 training. Period. 247 00:11:55,440 --> 00:11:57,360 And in version 3 point O that is gone. 248 00:11:57,360 --> 00:11:59,440 That unilateral commitment is entirely gone. 249 00:11:59,640 --> 00:12:01,520 They've replaced it with a clear bifurcation. 250 00:12:01,680 --> 00:12:04,640 They now differentiate between industry recommendations and 251 00:12:04,640 --> 00:12:07,320 company plans. So they're publicly recommending 252 00:12:07,320 --> 00:12:10,400 that the entire AI industry should pause if things get 253 00:12:10,400 --> 00:12:13,720 dangerous, but they are no longer promising that they will 254 00:12:13,720 --> 00:12:15,720 pause if their competitors keep going. 255 00:12:15,760 --> 00:12:18,800 They are framing it as a classic collective action problem. 256 00:12:19,480 --> 00:12:22,440 Their argument, which is highly rational from a pure business 257 00:12:22,440 --> 00:12:26,400 perspective, is that if Anthropic pauses to build 258 00:12:26,400 --> 00:12:30,840 perfect safety guards, but open AI or a massive state backed lab 259 00:12:30,840 --> 00:12:34,600 in China keeps scaling and throffic just loses market 260 00:12:34,600 --> 00:12:36,480 share. They lose their influence over 261 00:12:36,480 --> 00:12:37,680 the industry. Exactly. 262 00:12:37,920 --> 00:12:41,960 They effectively cede the future to actors who might be far less 263 00:12:41,960 --> 00:12:43,360 concerned with safety than they are. 264 00:12:43,600 --> 00:12:45,400 It is the ultimate prisoner's dilemma. 265 00:12:45,400 --> 00:12:48,440 If I play nice and you decide to play rough, I lose everything, 266 00:12:48,440 --> 00:12:50,120 so I am forced to keep playing rough. 267 00:12:50,560 --> 00:12:52,800 But there's another massive pressure point here that we 268 00:12:52,800 --> 00:12:55,360 absolutely have to discuss the Pentagon. 269 00:12:55,520 --> 00:12:58,360 The US Department of Defense has been getting incredibly loud 270 00:12:58,360 --> 00:13:00,240 about AI integration over the last year. 271 00:13:00,320 --> 00:13:03,000 They have. The Pentagon explicitly 272 00:13:03,000 --> 00:13:06,200 communicated that a strict refusal by an AI company to work 273 00:13:06,200 --> 00:13:09,640 on national security tasks was viewed as a supply chain risk. 274 00:13:09,920 --> 00:13:14,200 Supply chain risk, that is highly specific bureaucrat speak 275 00:13:14,200 --> 00:13:16,720 for we are not going to buy anything from you. 276 00:13:16,920 --> 00:13:19,480 Exactly. It is all about reliability. 277 00:13:19,680 --> 00:13:22,400 If the United States government is going to heavily integrate 278 00:13:22,400 --> 00:13:25,520 your AI agents into their defense system, they need 279 00:13:25,520 --> 00:13:28,120 absolute certainty that you aren't going to suddenly turn 280 00:13:28,120 --> 00:13:31,320 the servers off because of an internal moral qualm about how 281 00:13:31,320 --> 00:13:33,960 the tech is being used. So this policy shifted. 282 00:13:33,960 --> 00:13:36,320 Anthropic isn't just about preserving enterprise market 283 00:13:36,320 --> 00:13:39,400 share, it is about making themselves a viable long term 284 00:13:39,400 --> 00:13:42,800 government contractor. This new policy is a calculated 285 00:13:42,800 --> 00:13:46,400 pivot to survive in a world where the US government is 286 00:13:46,400 --> 00:13:48,960 suddenly the biggest, most important customer in the room. 287 00:13:49,200 --> 00:13:52,560 The entire concept of safety is being redefined in real time. 288 00:13:53,160 --> 00:13:56,120 It is no longer about slowing down to ensure alignment, it is 289 00:13:56,120 --> 00:13:58,760 about staying in the absolute lead so you have the power to 290 00:13:58,760 --> 00:14:00,720 set the rules. And while they are fiercely 291 00:14:00,720 --> 00:14:03,240 fighting for those lucrative government contracts, they're 292 00:14:03,240 --> 00:14:06,080 also simultaneously writing massive checks to make their 293 00:14:06,080 --> 00:14:08,360 foundational legal problems disappear. 294 00:14:08,840 --> 00:14:12,200 We really have to talk about the data that powers all of this. 295 00:14:12,360 --> 00:14:15,120 The cost of content, we finally have a concrete price tag on it. 296 00:14:15,360 --> 00:14:19,120 The Barts V Anthropic Settlement. $1.5 billion. 297 00:14:19,120 --> 00:14:22,240 That is a historic number for a copyright lawsuit. 298 00:14:22,440 --> 00:14:25,280 It completely sets the precedent for the entire industry. 299 00:14:25,720 --> 00:14:28,440 For a very long time, the standard defense from these AI 300 00:14:28,440 --> 00:14:32,080 labs was fair use. The argument was always that an 301 00:14:32,200 --> 00:14:34,760 AI learns exactly like a human learns. 302 00:14:35,080 --> 00:14:38,520 It reads publicly available information and internalizes the 303 00:14:38,520 --> 00:14:41,280 concepts. But the settlement strongly 304 00:14:41,280 --> 00:14:44,720 suggests that the actual cost of doing business moving forward 305 00:14:44,920 --> 00:14:47,880 involves paying out billion dollar class action settlements 306 00:14:48,160 --> 00:14:51,680 to the authors and creators. And the details of this specific 307 00:14:51,680 --> 00:14:54,000 case are crucial. This wasn't just a broad 308 00:14:54,000 --> 00:14:56,640 philosophical debate about whether an AI reading a 309 00:14:56,640 --> 00:15:00,320 purchased book is fair use. It was specifically focused on 310 00:15:00,320 --> 00:15:02,760 the book's three data set. Right, the shadow libraries. 311 00:15:02,760 --> 00:15:05,680 This was a massive data set that was largely composed of directly 312 00:15:05,680 --> 00:15:09,280 pirated books, so the legal battle shifted away from a high 313 00:15:09,280 --> 00:15:12,320 level debate about copyright theory to a very specific, 314 00:15:12,320 --> 00:15:15,320 undeniable accusation. You downloaded this material 315 00:15:15,320 --> 00:15:18,280 from a known pirate site and your engineers knew exactly what 316 00:15:18,280 --> 00:15:21,440 they were doing. A $1.5 billion penalty 317 00:15:21,440 --> 00:15:25,000 absolutely stings, but for a company that just raised $30 318 00:15:25,000 --> 00:15:28,240 billion in cash a few days ago, it is highly affordable. 319 00:15:28,240 --> 00:15:30,960 It is literally just a line item on a spreadsheet for them. 320 00:15:31,640 --> 00:15:33,920 And that is the dark irony of this entire settlement. 321 00:15:34,320 --> 00:15:39,000 A $1.5 billion penalty instantly destroys any new startup. 322 00:15:39,040 --> 00:15:41,960 It completely bankrupts the university research lab trying 323 00:15:41,960 --> 00:15:46,560 to build an open source model. But for Anthropic or Open AI or 324 00:15:46,560 --> 00:15:49,080 Google, it is just the toll they have to pay to get on the 325 00:15:49,080 --> 00:15:50,720 highway. It effectively entrenches the 326 00:15:50,720 --> 00:15:52,560 giants. They're the only entities on 327 00:15:52,560 --> 00:15:55,560 Earth who can actually afford to retroactively pay for the data 328 00:15:55,560 --> 00:15:57,000 they already scraped from the Internet. 329 00:15:57,000 --> 00:15:58,840 It completely clears the competitive field. 330 00:15:59,040 --> 00:16:02,200 So tying all of this together, we have AI models right now that 331 00:16:02,200 --> 00:16:04,960 are capable of genuine world changing brilliance. 332 00:16:05,200 --> 00:16:07,600 They are discovering hidden physics formulas, you're coding 333 00:16:07,600 --> 00:16:10,320 at the speed of thought, and we have autonomous agents that can 334 00:16:10,320 --> 00:16:13,840 finally navigate the messy pixel based reality of our everyday 335 00:16:13,840 --> 00:16:16,720 computers. But to actually get those agents 336 00:16:16,760 --> 00:16:19,720 out of the lab and into the real world, these companies are 337 00:16:19,720 --> 00:16:22,200 making a very specific set of compromises. 338 00:16:22,520 --> 00:16:25,440 They're wiring themselves deep into the corporate structure 339 00:16:25,440 --> 00:16:27,120 through massive consulting firms. 340 00:16:27,480 --> 00:16:31,160 They're softening their founding safety pledges to seamlessly 341 00:16:31,160 --> 00:16:34,480 aligned with military interests. And they are paying billion 342 00:16:34,480 --> 00:16:38,480 dollar fines to retroactively legalize the aggressive data 343 00:16:38,480 --> 00:16:41,200 gathering that made the model smart in the first place. 344 00:16:41,320 --> 00:16:44,440 We spent years intensely debating whether artificial 345 00:16:44,440 --> 00:16:47,200 intelligence would be safe or whether it would be perfectly 346 00:16:47,200 --> 00:16:50,280 aligned with human values. But looking at the reality of 347 00:16:50,280 --> 00:16:53,360 2026, safety isn't a philosophical stance anymore. 348 00:16:53,720 --> 00:16:56,440 It is simply a clause in a massive enterprise Oregon 349 00:16:56,440 --> 00:16:59,360 government contract. And it is being defined entirely 350 00:16:59,360 --> 00:17:01,760 by whoever is signing the biggest check, whether that 351 00:17:01,760 --> 00:17:04,319 happens to be the Pentagon or the venture capitalists. 352 00:17:04,400 --> 00:17:06,760 If you're not subscribed yet, take a second and hit follow on 353 00:17:06,760 --> 00:17:09,000 whatever app you're using. It helps us keep making this. 354 00:17:09,000 --> 00:17:10,240 We appreciate you being here.