WEBVTT 1 00:00:00.040 --> 00:00:02.560 I want you to imagine a campfire. Like you're sitting 2 00:00:02.600 --> 00:00:05.160 out in the woods, right, Yeah, it's a freezing night. 3 00:00:05.759 --> 00:00:08.160 You desperately need to keep this fire going just to 4 00:00:08.199 --> 00:00:12.000 stay warm. But instead of grabbing a log or some 5 00:00:12.119 --> 00:00:16.320 dry branches, right, you reach into this heavy canvas duffelback 6 00:00:16.800 --> 00:00:19.920 and you pull out a thick stack of CRISP one 7 00:00:19.960 --> 00:00:21.320 hundred dollars bills. 8 00:00:21.120 --> 00:00:23.760 Just straight cash, straight cash. 9 00:00:23.559 --> 00:00:27.160 And you toss them directly into the flames. Wow. They 10 00:00:27.199 --> 00:00:31.039 curl up, turn black and disappear. Wow, But you don't stop. 11 00:00:31.679 --> 00:00:34.960 You keep tossing those stacks in second after second, minute 12 00:00:35.000 --> 00:00:35.759 after minute. 13 00:00:35.840 --> 00:00:38.399 I mean, the physical logistics of that are it's almost 14 00:00:38.399 --> 00:00:41.280 impossible to even visualize, you know, right, to burn through 15 00:00:41.320 --> 00:00:43.960 that much money you're talking about setting fire to like 16 00:00:44.399 --> 00:00:46.920 hundreds of thousands of pounds of pure currency, you'd need 17 00:00:46.960 --> 00:00:48.920 an industrial incinerator running around the. 18 00:00:48.880 --> 00:00:52.399 Clock, exactly. It's a completely absurd mental image. But here's 19 00:00:52.439 --> 00:00:54.840 the thing. You are going to keep tossing those hundred 20 00:00:54.880 --> 00:00:58.000 dollar bills into the fire until you have burned through 21 00:00:58.159 --> 00:01:02.079 exactly twenty point nine billion dollars in a single year, 22 00:01:02.399 --> 00:01:05.239 in a single year, and that twenty point nine billion 23 00:01:05.280 --> 00:01:08.239 dollar cash burn. That is not a hypothetical number. That 24 00:01:08.400 --> 00:01:12.480 is the audited financial reality of open AI for the year. 25 00:01:12.280 --> 00:01:14.680 Twenty twenty five, which is just staggering. 26 00:01:14.879 --> 00:01:17.480 It really is so welcome to Thrilling Threads today, we 27 00:01:17.519 --> 00:01:20.640 are pulling apart a truly fascinating and frankly deeply alarming 28 00:01:20.680 --> 00:01:24.000 piece of source material. We're looking at a comprehensive breakdown 29 00:01:24.040 --> 00:01:25.480 by the entrepreneur Tom Billu. 30 00:01:25.560 --> 00:01:27.680 Yeah, his video on this is intense. 31 00:01:27.840 --> 00:01:30.840 It's titled, uh, Big Tech ran out of ideas and 32 00:01:30.920 --> 00:01:33.000 AI is the cover story. We had to react. 33 00:01:33.680 --> 00:01:36.000 And you know, his analysis is so critical right now 34 00:01:36.040 --> 00:01:39.760 because it really forces us to look past the I 35 00:01:39.760 --> 00:01:41.840 guess you'd call it the dazzling magic tricks we see 36 00:01:41.879 --> 00:01:42.840 on our screens every day. 37 00:01:42.920 --> 00:01:44.599 Oh, for sure, the shiny object syndrome. 38 00:01:44.719 --> 00:01:45.159 Exactly. 39 00:01:45.200 --> 00:01:49.159 We're being asked to stare directly into this terrifying financial 40 00:01:49.200 --> 00:01:52.799 engine that is actually powering the entire artificial intelligence boom. 41 00:01:53.280 --> 00:01:57.760 And he pulls from these incredibly dense financial reports insider 42 00:01:57.799 --> 00:02:01.560 interviews and he builds a very very unsettling picture. 43 00:02:01.879 --> 00:02:04.680 So our mission for today's Thrilling Threads is to untangle 44 00:02:04.719 --> 00:02:08.520 this massive, just glaring contradiction sitting right at the very 45 00:02:08.560 --> 00:02:11.560 heart of the AI industry, right, because on one hand, 46 00:02:11.599 --> 00:02:16.680 you have these mind bending, world changing technological leaps happening 47 00:02:16.800 --> 00:02:18.560 like almost weekly at this point. 48 00:02:18.680 --> 00:02:21.479 Yeah, curing diseases, writing code, driving cars. 49 00:02:21.639 --> 00:02:24.800 Right. But on the other hand, the industry is quietly 50 00:02:24.879 --> 00:02:31.080 building what looks like an apocalyptic, multi trillion dollar financial bubble. 51 00:02:31.159 --> 00:02:33.319 Yeah, it's a bubble that could pop at any minute. 52 00:02:33.680 --> 00:02:36.199 So to set the table for you, the listener, our 53 00:02:36.240 --> 00:02:40.599 source material presents two very distinct viewpoints on this contradiction. 54 00:02:40.960 --> 00:02:44.400 Right, we have ed Zitron anchoring one end of the spectrum. 55 00:02:44.520 --> 00:02:46.280 He's the bear, right, the ultimate bear case. 56 00:02:46.319 --> 00:02:46.560 Yeah. 57 00:02:46.800 --> 00:02:50.560 Zitron is this fierce, just unrelenting critic who looks at 58 00:02:50.560 --> 00:02:54.039 the current AI industry and literally calls it a nothing burger. 59 00:02:54.159 --> 00:02:55.199 A nothing burger. Wow. 60 00:02:55.280 --> 00:02:55.680 Yeah. 61 00:02:56.159 --> 00:02:59.240 His core argument is that artificial intelligence, at least how 62 00:02:59.240 --> 00:03:02.520 it's currently monetized and packaged for enterprise use, is really 63 00:03:02.560 --> 00:03:05.319 just a ten to thirty billion dollar industry. Okay, But 64 00:03:05.400 --> 00:03:07.879 the problem is that it has put on this very elaborate, 65 00:03:08.199 --> 00:03:12.560 very expensive costume, and it's masquerading as a trillion dollar industry. 66 00:03:13.039 --> 00:03:16.000 He believes the fundamentals simply do not exist to support 67 00:03:16.080 --> 00:03:17.439 these massive valuations. 68 00:03:17.719 --> 00:03:20.680 And then providing the counterweight to that is Tom Billy 69 00:03:20.800 --> 00:03:25.000 himself kind of acting as the cautious bull here. Right. 70 00:03:25.319 --> 00:03:28.120 He looks at the exact same data as Zitron, but 71 00:03:28.159 --> 00:03:31.599 he comes away to a totally different conclusion. Like Billy, 72 00:03:31.639 --> 00:03:35.800 you genuinely believes that AI is a revolutionary technology. He 73 00:03:35.879 --> 00:03:40.159 thinks it's already fundamentally changing the fabric of human society. 74 00:03:39.680 --> 00:03:41.680 Which is hard to argue against when you see it 75 00:03:41.719 --> 00:03:42.080 in action. 76 00:03:42.400 --> 00:03:46.360 Right, And this is the massive caveat here. Even as 77 00:03:46.400 --> 00:03:49.199 a believer in the technology, he openly admits that the 78 00:03:49.240 --> 00:03:52.319 financial debt being accumulated to build it is absolutely terrifying. 79 00:03:52.400 --> 00:03:55.479 Yeah, it creates this fascinating tension for our whole discussion today. 80 00:03:55.599 --> 00:03:59.199 We have one prominent voice arguing the technologies business model 81 00:03:59.360 --> 00:04:02.479 is a total illusion, like a scam basically, yeah, and 82 00:04:02.560 --> 00:04:05.120 another saying the tech is an absolute miracle, but the 83 00:04:05.120 --> 00:04:07.680 sheer cost of it might just destroy the broader economy. 84 00:04:07.719 --> 00:04:08.080 Anyway. 85 00:04:08.360 --> 00:04:10.280 It makes me think of a Buddy mine a few 86 00:04:10.360 --> 00:04:13.719 years back. He got into this crazy situation. He went 87 00:04:13.759 --> 00:04:16.720 to a high end dealership and saw this absolute spaceship 88 00:04:16.759 --> 00:04:18.000 of a luxury sports car. 89 00:04:18.160 --> 00:04:19.839 Oh no, I see where this is going. 90 00:04:20.120 --> 00:04:23.240 Right. He was so blinded by the stitched leather seats, 91 00:04:23.319 --> 00:04:26.560 the zero to sixty speed, and just the sheer social 92 00:04:26.600 --> 00:04:30.040 status of driving the thing that he signed the financing 93 00:04:30.079 --> 00:04:31.480 papers without actually reading them. 94 00:04:31.519 --> 00:04:33.399 Did he even look at the maintenance schedule? 95 00:04:33.480 --> 00:04:35.360 He definitely didn't look at the natance schedule. He drives 96 00:04:35.360 --> 00:04:37.399 it off a lot, feeling like, you know, the smartest 97 00:04:37.399 --> 00:04:40.560 guy in town, naturally, But a year later he realizes 98 00:04:41.000 --> 00:04:44.680 the car requires this special type of imported synthetic oil 99 00:04:44.720 --> 00:04:49.199 that costs like one thousand dollars a gallon. Oh my god, 100 00:04:49.560 --> 00:04:52.879 the brake pads need replacing every three thousand miles, and 101 00:04:53.000 --> 00:04:56.759 his loan interest is compounding daily. So the car is real, 102 00:04:56.879 --> 00:04:59.839 the speed is real, the incredible speed is completely real. 103 00:05:00.480 --> 00:05:02.480 But the math dictates that he's going to lose his 104 00:05:02.480 --> 00:05:03.360 house to pay for it. 105 00:05:03.680 --> 00:05:03.920 Right. 106 00:05:04.079 --> 00:05:06.959 The utility of a product doesn't just like rewrite the 107 00:05:07.040 --> 00:05:10.600 laws of mathematics exactly. A miraculous product can still bankrupt 108 00:05:10.680 --> 00:05:13.079 you if the cost to produce it exceeds what anyone 109 00:05:13.120 --> 00:05:15.680 is actually willing to pay. And we are talking about 110 00:05:15.680 --> 00:05:18.959 that exact same dynamic today, just scaled up to a 111 00:05:19.040 --> 00:05:21.720 global multi trillion dollar level. 112 00:05:22.040 --> 00:05:24.959 Okay, let's unpack this. Let's ground that campfire image in 113 00:05:25.040 --> 00:05:28.480 the hard numbers. Our source brings up that audited FTI 114 00:05:28.600 --> 00:05:32.639 report detailing the open AI twenty twenty five cash burn. 115 00:05:33.600 --> 00:05:36.240 I want to really understand the mechanics of that burn rate, 116 00:05:36.279 --> 00:05:38.639 because this isn't just a tech startup spending too much 117 00:05:38.879 --> 00:05:43.040 venture capital on like super Bowl ads or luxurious office campuses, right. 118 00:05:43.079 --> 00:05:43.800 No, not at all. 119 00:05:43.879 --> 00:05:47.040 There is a fundamental flaw in the actual economic model 120 00:05:47.079 --> 00:05:50.800 of the software itself. The core thesis ed Xyitron lays out, 121 00:05:50.839 --> 00:05:54.519 which Billu heavily highlights in the video, hinges on one 122 00:05:54.519 --> 00:05:56.800 incredibly dangerous reality. 123 00:05:56.399 --> 00:05:57.399 For Silicon Valley. 124 00:05:57.480 --> 00:05:58.319 What's the reality? 125 00:05:58.399 --> 00:06:01.759 It boils down to a single sentence. AI costs increase 126 00:06:01.839 --> 00:06:02.560 linearly with. 127 00:06:02.600 --> 00:06:04.360 Revenues, linearly with revenues. 128 00:06:04.519 --> 00:06:07.000 Yes. And to understand why that is a literal death 129 00:06:07.040 --> 00:06:09.040 sentence for a tech company, you have to look at 130 00:06:09.040 --> 00:06:11.720 the golden rule of the digital age, the marginal cost 131 00:06:11.759 --> 00:06:14.519 to software. Precisely, it is the exact opposite of how 132 00:06:14.560 --> 00:06:17.720 AI works. Think about a traditional software product like a 133 00:06:17.759 --> 00:06:22.079 word processor or a photo sharing app. Okay, a company 134 00:06:22.160 --> 00:06:24.879 might spend fifty million dollars paying engineers to write the 135 00:06:24.920 --> 00:06:25.680 original code. 136 00:06:25.720 --> 00:06:27.360 That's a huge upfront cost. 137 00:06:27.519 --> 00:06:31.199 Huge, But once the code is finished, compiling and sending 138 00:06:31.279 --> 00:06:34.240 the second copy of that software to a customer costs 139 00:06:34.279 --> 00:06:39.399 practically zero. Sending the two billionth copy costs practically zero. 140 00:06:39.959 --> 00:06:42.959 You build it once and you sell it infinitely. The 141 00:06:42.959 --> 00:06:46.040 profit margins approached ninety nine percent over time. 142 00:06:45.839 --> 00:06:49.120 And that mechanism is why companies like Microsoft and Apple 143 00:06:49.160 --> 00:06:51.360 became the most valuable entities in human history. 144 00:06:51.439 --> 00:06:55.560 Exactly, they achieved infinite scale and zero marginal cost. 145 00:06:56.120 --> 00:06:58.680 I'm seeing the disconnect here because when I open a 146 00:06:58.720 --> 00:07:01.480 traditional app, my my own phone or my own laptop 147 00:07:01.560 --> 00:07:04.000 is doing the processing work to run it right. The 148 00:07:04.040 --> 00:07:07.720 software company isn't paying for my device's electricity or as processor. 149 00:07:07.839 --> 00:07:08.759 Nope, that's on you. 150 00:07:08.920 --> 00:07:11.279 But when I use a large language model like chad GPT, 151 00:07:11.879 --> 00:07:14.199 my phone isn't doing the heavy lifting at all. 152 00:07:14.240 --> 00:07:15.120 Not even a little bit. 153 00:07:15.120 --> 00:07:18.480 The heavy lifting is being done miles miles away every 154 00:07:18.519 --> 00:07:20.879 single time a user types of prompt and hits enter. 155 00:07:21.240 --> 00:07:24.360 The AI doesn't just like retrieve a pre written file. 156 00:07:24.639 --> 00:07:28.160 It's not a search engine exactly, a massive array of 157 00:07:28.240 --> 00:07:33.079 highly specialized hardware, specifically these incredibly expensive in Vidia GPU 158 00:07:33.160 --> 00:07:36.560 sitting in a giant data center has to physically spin 159 00:07:36.639 --> 00:07:40.600 up and perform billions of mathematical calculations in real time 160 00:07:40.839 --> 00:07:44.160 just to generate that specific unique string of text. Wow, 161 00:07:44.279 --> 00:07:48.800 it requires raw physical computing power. It requires massive amounts 162 00:07:48.800 --> 00:07:51.399 of electricity to run the chips, and then massive amounts 163 00:07:51.439 --> 00:07:53.439 of water and power to cool the chips down so 164 00:07:53.480 --> 00:07:54.839 they literally don't melt. 165 00:07:54.959 --> 00:07:57.560 So every single time someone uses it, they are burning 166 00:07:57.720 --> 00:07:59.079 physical resources. 167 00:07:58.680 --> 00:07:59.879 Every single time. 168 00:08:00.399 --> 00:08:03.000 So as your user base grows, and as those users 169 00:08:03.040 --> 00:08:07.000 start prompting the AI more frequently, your server costs and 170 00:08:07.040 --> 00:08:10.480 your electricity bills go up and lockstep with your revenue. Right, 171 00:08:10.600 --> 00:08:14.279 you never reached that software sweet spot of infinite free replication. 172 00:08:14.560 --> 00:08:17.480 It's like you are essentially running a heavy manufacturing plant, 173 00:08:17.560 --> 00:08:18.959 but your product happens. 174 00:08:18.600 --> 00:08:20.800 To be text, and the source material points out that 175 00:08:20.800 --> 00:08:24.040 it's actually getting worse worse. Well, as these models get 176 00:08:24.040 --> 00:08:26.959 more complex, you know, moving from GBT three to GPT 177 00:08:27.040 --> 00:08:31.160 four and whatever comes next, they require exponentially more compute 178 00:08:31.199 --> 00:08:34.519 to train and to run, so the margins are deteriorating. 179 00:08:34.679 --> 00:08:37.799 The more successful the product becomes in terms of user engagement, 180 00:08:37.879 --> 00:08:39.759 the more money the company actually loses. 181 00:08:40.200 --> 00:08:44.080 But the tech industry's standard defense mechanism to this argument. 182 00:08:44.399 --> 00:08:47.360 I hear it constantly from the engineers. They always say, Oh, 183 00:08:47.480 --> 00:08:50.159 don't worry about the power draw today, the specialist silicon 184 00:08:50.200 --> 00:08:51.159 will save us tomorrow. 185 00:08:51.320 --> 00:08:52.799 The magic chip argument. 186 00:08:53.000 --> 00:08:56.480 Yes, They point to Nvidia's roadmap, things like the upcoming 187 00:08:56.600 --> 00:09:00.559 via Reuben's chips, promising that the next generation of hardware 188 00:09:00.600 --> 00:09:03.919 will be so fast and so efficient that the influence 189 00:09:03.960 --> 00:09:05.679 costs will just drop to zero. 190 00:09:05.960 --> 00:09:09.480 Yeah, but our source pushes back aggressively on that hardware narrative. 191 00:09:10.000 --> 00:09:13.639 Xetron argues there is zero historical or mathmatal proof that 192 00:09:13.679 --> 00:09:16.679 hardware upgrades can magically fix these inverted margins. 193 00:09:16.879 --> 00:09:19.240 Zero proof none, Because while. 194 00:09:19.080 --> 00:09:21.159 It is true that in Nvidia is making chiffs that 195 00:09:21.279 --> 00:09:25.159 are say, twice as efficient, the AI companies are simultaneously 196 00:09:25.159 --> 00:09:26.919 building models that are ten times large. 197 00:09:26.960 --> 00:09:27.759 I see the. 198 00:09:27.679 --> 00:09:32.200 Software's demand for power vastly outpaces the hardware's efficiency gains. 199 00:09:32.600 --> 00:09:35.639 It's like if your car engine gets twenty percent better 200 00:09:35.679 --> 00:09:38.519 gas mileage, but you suddenly decide tow a battleship instead 201 00:09:38.519 --> 00:09:40.600 of a rowboat, you are still going to run out 202 00:09:40.600 --> 00:09:41.000 of fuel. 203 00:09:41.159 --> 00:09:41.360 Right. 204 00:09:41.440 --> 00:09:42.279 The math is not. 205 00:09:42.360 --> 00:09:46.679 Trending toward profitability. It's trending toward a hard physical ceiling. 206 00:09:46.879 --> 00:09:51.240 It's like inventing a futuristic, fully automated bakery. And this 207 00:09:51.320 --> 00:09:54.639 bakery produces the most incredible, life changing bread you have 208 00:09:54.840 --> 00:09:56.159 ever tasted in your life. 209 00:09:56.200 --> 00:09:57.399 Okay, I like this right. 210 00:09:57.759 --> 00:10:00.440 It cures ailments, it gives you energy for day. It's 211 00:10:00.480 --> 00:10:03.960 a literal miracle. But the catch is the oven only 212 00:10:04.080 --> 00:10:08.399 runs if you shovel flawless one carrot diamonds into the furnace. 213 00:10:08.480 --> 00:10:12.159 The fuel cost is just astronomical and inescapable. 214 00:10:12.320 --> 00:10:14.600 Exactly everyone wants the bread. You have millions of people 215 00:10:14.600 --> 00:10:16.759 lining up around the block to buy a loaf, but 216 00:10:16.919 --> 00:10:19.120 every single time you bake one, you have to go 217 00:10:19.159 --> 00:10:20.440 buy another diamond to burn. 218 00:10:20.639 --> 00:10:22.960 You can't just scale up the recipe. 219 00:10:22.600 --> 00:10:25.840 No, you can't scale it up without scaling up your 220 00:10:25.840 --> 00:10:28.559 diamond purchases. So no matter how much you charge for 221 00:10:28.600 --> 00:10:31.759 the bread, unless you start charging millions of dollars a loaf, 222 00:10:31.960 --> 00:10:35.759 which no one can afford, the business model is inherently 223 00:10:35.799 --> 00:10:39.159 doomed because the fuel cost scales directly with the output. 224 00:10:39.440 --> 00:10:42.960 That analogy perfectly captures the panic that's kind of bubbling 225 00:10:43.039 --> 00:10:45.679 under the surface in Silicon Valley right now. Because if 226 00:10:45.720 --> 00:10:48.600 the costs are this astronomically high, if they are burning 227 00:10:48.639 --> 00:10:52.279 through twenty point nine billion dollars just to keep the 228 00:10:52.320 --> 00:10:55.639 diamond furnace running, someone has to be footing that bill. 229 00:10:56.240 --> 00:10:57.879 And retail consumer is paying. 230 00:10:57.639 --> 00:11:00.200 Like twenty bucks a month for a premium chare, but 231 00:11:00.320 --> 00:11:03.360 subscription that barely puts a debt in that number. The 232 00:11:03.440 --> 00:11:06.159 true financial targets for these companies are corporate. 233 00:11:05.919 --> 00:11:09.200 Enterprises, which is where the entire narrative starts to fracture, 234 00:11:09.399 --> 00:11:14.440 right because according to the source material, those massive corporate enterprises, 235 00:11:14.480 --> 00:11:16.480 the Fortune five hundred companies that are supposed to be 236 00:11:16.519 --> 00:11:20.639 injecting hundreds of billions of dollars into this ecosystem, they 237 00:11:20.799 --> 00:11:23.840 are suddenly slamming on the brakes. The video pulls this 238 00:11:23.919 --> 00:11:27.519 incredible clip of Alex Karp. He's the CEO of Palenteer, 239 00:11:28.039 --> 00:11:31.919 and he's being interviewed on CNBC's squawk Box, and he 240 00:11:32.080 --> 00:11:35.799 issues a very very blunt warning about enterprise adoption. 241 00:11:35.960 --> 00:11:38.440 And Karp is uniquely positioned to see this right because 242 00:11:38.519 --> 00:11:42.919 Palenteer builds data integration software for massive institutions. We're talking 243 00:11:42.919 --> 00:11:43.960 from the military to. 244 00:11:44.360 --> 00:11:45.799 Global banks, heaby hitters. 245 00:11:45.879 --> 00:11:49.360 Yeah, and he notes that corporate enterprises are realizing they 246 00:11:49.360 --> 00:11:53.240 are getting absolutely zero measurable value out of AI tokens. 247 00:11:53.720 --> 00:11:57.279 Karp literally labeled the current enterprise AI push as a 248 00:11:57.279 --> 00:11:58.000 waste of time. 249 00:11:58.279 --> 00:11:58.600 Wow. 250 00:11:58.919 --> 00:12:02.080 He said, the basic view among these massive companies is, quote, 251 00:12:02.200 --> 00:12:04.759 I'm going to chilax and waste my time with tokens. 252 00:12:04.840 --> 00:12:06.840 I'm going to get no value and they're going to 253 00:12:06.919 --> 00:12:07.480 get my IP. 254 00:12:07.799 --> 00:12:10.879 That is a brutal public assessment from a major industry insider. 255 00:12:11.200 --> 00:12:13.639 Waste my time with tokens. I really want to break 256 00:12:13.639 --> 00:12:16.519 down why this is happening, Like, why are sophisticated companies 257 00:12:16.559 --> 00:12:19.120 paying huge sums of money to integrate these large language 258 00:12:19.120 --> 00:12:21.639 models and then feeling like they were sold a bill 259 00:12:21.679 --> 00:12:22.080 of goods. 260 00:12:22.320 --> 00:12:25.240 Well, it comes down to the fundamental architecture of the 261 00:12:25.279 --> 00:12:29.759 models themselves. As Ziitron details in the source, large language 262 00:12:29.759 --> 00:12:32.240 models are inherently hallucination. 263 00:12:31.720 --> 00:12:33.360 Prone, right, the hallucinaty. 264 00:12:33.440 --> 00:12:35.440 Yeah. And this isn't just a bug that can be 265 00:12:35.519 --> 00:12:38.080 patched out in the next update. It is a core 266 00:12:38.200 --> 00:12:41.720 feature of how the technology operates. It's a reality that 267 00:12:41.799 --> 00:12:45.240 even companies like open ai have openly acknowledged in their 268 00:12:45.279 --> 00:12:46.399 own technical. 269 00:12:46.000 --> 00:12:49.120 Papers because they aren't actually looking up facts in a database. OK. 270 00:12:49.600 --> 00:12:51.159 I think a lot of people misunderstand this. 271 00:12:51.480 --> 00:12:53.120 They completely misunderstand it. 272 00:12:53.200 --> 00:12:55.799 When you ask a chatbout a question, it's not like 273 00:12:56.240 --> 00:12:58.840 googling the answer and reading a Wikipedia page back to you. 274 00:12:59.320 --> 00:13:03.960 It's just predicting the next most statistically likely word in 275 00:13:04.000 --> 00:13:07.279 a sequence based on all the billions of texts that 276 00:13:07.399 --> 00:13:08.799 read during its training phase. 277 00:13:08.879 --> 00:13:12.000 Exactly, they operate in a probabilistic vector space. They don't 278 00:13:12.000 --> 00:13:15.039 actually know anything, They just know patterns of language. So 279 00:13:15.480 --> 00:13:19.080 if an enterprise asks an AI to summarize a highly 280 00:13:19.159 --> 00:13:23.679 complex proprietary financial audit, the AI might encounter a gap 281 00:13:23.720 --> 00:13:26.879 in its pattern recognition, and instead of saying I don't know, 282 00:13:27.399 --> 00:13:30.799 it's programming basically compels it to generate the most plausible 283 00:13:30.840 --> 00:13:33.799 sounding next word, Oh I am, So it will confidently 284 00:13:33.840 --> 00:13:37.720 invent a completely fictional set of numbers, format them perfectly 285 00:13:37.759 --> 00:13:41.000 in a beautiful table, and present them to the executive 286 00:13:41.120 --> 00:13:42.399 as absolute fact. 287 00:13:42.559 --> 00:13:45.159 And in a corporate environment, that is a tipping time bomb. 288 00:13:45.440 --> 00:13:48.080 You cannot run a global supply chain, or a hospital's 289 00:13:48.159 --> 00:13:51.960 legal department, or an airline's maintenance schedule using a tool 290 00:13:52.000 --> 00:13:54.840 that occasionally just hallucinates a fictional reality. 291 00:13:54.919 --> 00:13:55.919 You'd be insane to try. 292 00:13:56.080 --> 00:13:59.440 Precision is everything. A ninety nine percent accuracy rate sounds 293 00:13:59.440 --> 00:14:01.720 great if you're at a high school essay, but that 294 00:14:01.759 --> 00:14:04.679 one percent failure rate will bankrupt a fortune five hundred 295 00:14:04.720 --> 00:14:06.399 company if they act on bad data. 296 00:14:06.440 --> 00:14:10.000 Exactly, and because of this hallucination problem, zyschron points out 297 00:14:10.000 --> 00:14:13.559 an incredible limitation for the entire AI industry. They cannot 298 00:14:13.600 --> 00:14:16.000 charge their clients based on outcomes or success. 299 00:14:16.159 --> 00:14:18.879 We explain that think about a consulting firm. 300 00:14:19.080 --> 00:14:21.639 A consulting firm can charge a client based on how 301 00:14:21.720 --> 00:14:24.399 much money they save them, right, a percentage of the savings. Sure, 302 00:14:24.879 --> 00:14:27.440 AI companies can't do that because they can't guarantee the 303 00:14:27.480 --> 00:14:30.879 output is even real. So they are forced to charge 304 00:14:30.960 --> 00:14:33.559 based purely on usage, on how many compute tokens you 305 00:14:33.559 --> 00:14:37.320 burn through, regardless of whether the answer actually solved your 306 00:14:37.360 --> 00:14:40.879 problem or you know, created a massive legal liability. 307 00:14:41.240 --> 00:14:44.759 So they actually have a perverse incentive to encourage waste. Exactly, 308 00:14:44.799 --> 00:14:47.759 they want the enterprise employees to spin their wheels, go 309 00:14:47.960 --> 00:14:51.039 back and forth with the chatbot all day long, tweaking prompts, 310 00:14:51.039 --> 00:14:53.519 try and get a usable answer. Because the AI company 311 00:14:53.519 --> 00:14:55.080 gets paid for every single keystroke. 312 00:14:55.200 --> 00:14:55.399 Yep. 313 00:14:55.840 --> 00:14:59.919 But Karp's warning on CNBC went beyond just wasting money. 314 00:15:00.519 --> 00:15:03.159 Let's dig into the second half of his quote. They're 315 00:15:03.159 --> 00:15:06.039 going to get my IP. This introduces what we're calling 316 00:15:06.120 --> 00:15:07.279 the Great IP heist. 317 00:15:07.799 --> 00:15:10.399 This is one of the most shocking allegations discussed in 318 00:15:10.440 --> 00:15:14.240 the source material. The accusation here is that AI companies 319 00:15:14.240 --> 00:15:18.360 are effectively pillaging their own clients intellectual property to launch 320 00:15:18.399 --> 00:15:19.840 direct competitors against them. 321 00:15:20.039 --> 00:15:24.120 It's wild. Our source gives a very specific, devastating example 322 00:15:24.159 --> 00:15:27.840 of this dynamic in action. There's this massive collaborative design 323 00:15:27.879 --> 00:15:30.799 platform called Pigma used by designers all over the world, 324 00:15:30.879 --> 00:15:33.840 very popular. Yeah and Anthropic, which is one of the 325 00:15:33.960 --> 00:15:38.679 leading AI labs, allegedly trained their foundational models by scraping 326 00:15:38.720 --> 00:15:42.320 and absorbing massive amounts of data from Pigma, and then 327 00:15:42.879 --> 00:15:46.240 armed with Pigma's own underlying logic and design patterns and 328 00:15:46.320 --> 00:15:52.519 Tropic turnaround and launched a tool called clotted Design. Dylan Field, 329 00:15:52.600 --> 00:15:56.399 the CEO of Figma, was reportedly blindsided and just absolutely 330 00:15:56.440 --> 00:15:57.360 shocked by this move. 331 00:15:57.480 --> 00:16:00.679 It sets a terrifying precedent for literally any corporate executive. 332 00:16:00.799 --> 00:16:03.679 Imagine you are the CEO of a highly specialized logistics 333 00:16:03.720 --> 00:16:07.480 company rap Your only competitive moat is your proprietary data, 334 00:16:07.759 --> 00:16:11.759 the decades of routing information, supplier negotiations, efficiency metrics you've 335 00:16:11.759 --> 00:16:14.080 built up over years. If you feed all of that 336 00:16:14.120 --> 00:16:16.120 into an off the shelf AI model to try and 337 00:16:16.200 --> 00:16:17.799 like optimize your routes. 338 00:16:17.559 --> 00:16:19.360 You're training the AI on your secret sauce. 339 00:16:19.440 --> 00:16:20.240 You're handing it over. 340 00:16:21.039 --> 00:16:24.960 The AI company can then absorb that knowledge, generalize it, 341 00:16:25.360 --> 00:16:29.960 and sell a cheap automated logistics expert tool to everyone 342 00:16:30.000 --> 00:16:30.360 else in. 343 00:16:30.320 --> 00:16:32.159 The world, including your direct competitors. 344 00:16:32.200 --> 00:16:33.679 Including your direct competitors. 345 00:16:34.120 --> 00:16:37.240 You essentially paid the AI company to automate your own 346 00:16:37.279 --> 00:16:39.600 business model and sell it to the highest bedder. 347 00:16:39.840 --> 00:16:42.720 It's like inviting a famous chef into your family restaurant 348 00:16:42.759 --> 00:16:46.159 to help you organize your kitchen. The chef takes detailed 349 00:16:46.159 --> 00:16:49.360 notes on your grandmother's secret recipe, walks right next door, 350 00:16:49.679 --> 00:16:52.759 and opens a massive chain restaurant selling the exact same 351 00:16:52.840 --> 00:16:54.080 dish for half the price. 352 00:16:54.360 --> 00:16:57.360 That is exactly it. Of course, enterprises are slamming on 353 00:16:57.399 --> 00:16:58.240 the brakes. 354 00:16:58.000 --> 00:17:01.159 Right, But Alex Karp isn't just complaining on television about this. 355 00:17:01.559 --> 00:17:05.680 Palenteer actually proposes a technical solution to this IP theft problem, 356 00:17:05.759 --> 00:17:06.119 don't they? 357 00:17:06.200 --> 00:17:06.599 They do? 358 00:17:06.920 --> 00:17:10.039 Karp argues that the only way enterprises will ever safely 359 00:17:10.039 --> 00:17:12.799 adopt this technology is through what the source describes as 360 00:17:12.799 --> 00:17:16.440 an obfuscation layer or a middle layer. Palenteer calls their 361 00:17:16.519 --> 00:17:18.039 version of this an ontology. 362 00:17:18.319 --> 00:17:21.319 Okay, explain the mechanics of an ontology. How does a 363 00:17:21.319 --> 00:17:23.039 middle layer actually protect the data? 364 00:17:23.359 --> 00:17:26.720 You can visualize it as a secure walled garden. So 365 00:17:27.079 --> 00:17:31.119 enterprises desperately want the analytical reasoning power of an LM. Right, 366 00:17:31.400 --> 00:17:34.480 they have mountains of unstructured data that humans just can't 367 00:17:34.480 --> 00:17:38.000 process fast enough, but they absolutely cannot allow that data 368 00:17:38.039 --> 00:17:41.039 to be sent back to the AI's core training lervers. 369 00:17:41.200 --> 00:17:44.160 Right, keep the chef out of the recipe book exactly. 370 00:17:44.759 --> 00:17:48.440 So the ontology acts as a secure translator and a barrier. 371 00:17:48.839 --> 00:17:52.799 It utilizes technologies like vector databases and something called retrieval 372 00:17:52.839 --> 00:17:54.200 augmented generation or. 373 00:17:54.279 --> 00:17:57.640 RAG a RAG. Okay, So the enterprise keeps its data 374 00:17:57.640 --> 00:17:59.920 completely inside its own locked house exactly. 375 00:18:00.240 --> 00:18:03.079 When an employee asks a question, the ontology searches the 376 00:18:03.119 --> 00:18:08.319 company's internal highly secure database, retrieves the highly specific proprietary facts, 377 00:18:08.599 --> 00:18:10.839 and then hands those facts to the AI model in 378 00:18:10.880 --> 00:18:11.400 a sort. 379 00:18:11.240 --> 00:18:13.920 Of temporary workspace. Let's a sandbox, Yeah, sandbox. 380 00:18:14.240 --> 00:18:17.440 The AI uses its reasoning capability to formulate an answer 381 00:18:17.480 --> 00:18:20.440 based only on those provided facts, and then the workspace 382 00:18:20.480 --> 00:18:24.240 is just wiped clean, deleted. Oh well, the AI model 383 00:18:24.240 --> 00:18:27.519 itself remains frozen. It does not learn from the interaction, 384 00:18:27.720 --> 00:18:29.960 and it cannot send the data back to its creators. 385 00:18:30.559 --> 00:18:34.119 It allows the enterprise to benefit from the intelligence without 386 00:18:34.240 --> 00:18:36.079 surrendering the intellectual property. 387 00:18:36.400 --> 00:18:39.759 As Bill U heavily stresses in the video, The Future 388 00:18:39.759 --> 00:18:43.759 of Corporate AI guarantees that you must somehow turn that 389 00:18:43.920 --> 00:18:47.839 generic AI into something proprietary to you. You have to 390 00:18:48.039 --> 00:18:50.640 because if your company is just using the exact same 391 00:18:50.839 --> 00:18:53.559 off the shelf chat GPT that your competitor across the 392 00:18:53.599 --> 00:18:57.039 street is using, there is zero competitive advantage. If the 393 00:18:57.079 --> 00:18:59.759 AI running your strategy gives the exact same answers as 394 00:18:59.759 --> 00:19:03.160 the A running your competitor strategy, the entire industry just 395 00:19:03.160 --> 00:19:05.599 commoditizes and everybody loses their edge. 396 00:19:05.960 --> 00:19:08.839 Right, and the struggle over application and utility really raises 397 00:19:08.839 --> 00:19:11.480 a glaring question about the current leadership of these major 398 00:19:11.519 --> 00:19:14.680 AI labs. The source points out something very very telling 399 00:19:14.720 --> 00:19:18.640 about figures like open ayes Sam Altman and anthropics Daryo Miday. 400 00:19:19.039 --> 00:19:21.319 This part of the deep Sorry, this part of our 401 00:19:21.359 --> 00:19:24.640