WEBVTT 00:00:00.000 --> 00:00:04.370 Welcome to the deep dive. We've got a... A massive 00:00:04.370 --> 00:00:06.750 stack of your sources today covering the whole 00:00:06.750 --> 00:00:10.169 sprawling, fast -moving world of generative AI. 00:00:10.470 --> 00:00:12.970 Yeah, we're looking at a really extensive, totally 00:00:12.970 --> 00:00:16.190 up -to -date Wikipedia article for 2026. Right. 00:00:16.190 --> 00:00:17.649 And I mean, if you're listening to this, you've 00:00:17.649 --> 00:00:19.829 probably typed a prompt into a box recently and 00:00:19.829 --> 00:00:22.030 just watched a magic trick happen. Oh, absolutely. 00:00:22.149 --> 00:00:25.190 Everyone has. You ask for like a photorealistic 00:00:25.190 --> 00:00:28.309 image of a cat riding a skateboard on Mars, and 00:00:28.309 --> 00:00:30.829 poof, it just appears on your screen. And usually 00:00:30.829 --> 00:00:32.909 when we encounter a technological shift like 00:00:32.880 --> 00:00:35.380 we just sort of clap at the magic. Exactly. We 00:00:35.380 --> 00:00:37.420 tend to focus entirely on the rabbit and we just 00:00:37.420 --> 00:00:40.140 completely ignore the trapdoor. Yeah. But if 00:00:40.140 --> 00:00:41.700 you really want to understand the trick, you 00:00:41.700 --> 00:00:43.539 have to look under the stage, you know, you have 00:00:43.539 --> 00:00:46.100 to look at the pulleys, the hidden compartments, 00:00:46.219 --> 00:00:48.420 and just the really messy reality of how the 00:00:48.420 --> 00:00:50.679 illusion actually works. Which is exactly what 00:00:50.679 --> 00:00:52.679 we're doing with you today. Right. Because the 00:00:52.679 --> 00:00:55.140 goal here is not to just rattle off a list of 00:00:55.140 --> 00:00:58.640 new chatbots or software updates. We want to 00:00:58.640 --> 00:01:00.899 cut through the incredible amount of noise out 00:01:00.899 --> 00:01:02.740 there. We're going to explore the underlying 00:01:02.740 --> 00:01:06.340 mechanics. the sweeping global hype cycle, the 00:01:06.340 --> 00:01:09.859 really complicated legal battles, and the surprising 00:01:09.859 --> 00:01:12.379 real -world consequences we are seeing unfold 00:01:12.379 --> 00:01:15.659 right now. OK, let's untack this. To understand 00:01:15.659 --> 00:01:17.599 what generative AI is doing today, we really 00:01:17.599 --> 00:01:19.379 have to look at where it came from. Because I 00:01:19.379 --> 00:01:21.000 think a lot of people assume this technology 00:01:21.000 --> 00:01:24.680 was born in a Silicon Valley garage maybe a decade 00:01:24.680 --> 00:01:27.180 ago. Oh, no. The timeline goes back way further 00:01:27.180 --> 00:01:29.200 than that, over a century, actually. A century? 00:01:29.400 --> 00:01:32.379 Wait, really? Yeah, it really does. The foundational 00:01:32.379 --> 00:01:35.840 concept of algorithmically generating media dates 00:01:35.840 --> 00:01:39.760 all the way back to 1906. Wow. There was this 00:01:39.760 --> 00:01:42.500 Russian mathematician named Andrei Markov and 00:01:42.500 --> 00:01:45.959 he was studying Alexander Prushkin's poem Eugenianogen. 00:01:46.560 --> 00:01:48.780 And obviously he didn't have a computer in 1906. 00:01:49.120 --> 00:01:51.840 Alright, no computers. He sat there and manually 00:01:51.840 --> 00:01:54.680 analyzed the text. He was counting the patterns 00:01:54.680 --> 00:01:57.780 of vowels and consonants to calculate the mathematical 00:01:57.780 --> 00:02:00.040 probability of one letter following another. 00:02:00.200 --> 00:02:02.099 Which is just incredible to think about. I mean 00:02:02.099 --> 00:02:04.480 he was essentially doing probabilistic text generation 00:02:04.480 --> 00:02:07.219 entirely by hand. Like if you have a consonant 00:02:07.420 --> 00:02:10.919 What are the exact odds a vowel comes next? That 00:02:10.919 --> 00:02:13.620 core idea, just predicting what comes next based 00:02:13.620 --> 00:02:16.939 on established patterns, is the absolute bedrock 00:02:16.939 --> 00:02:19.240 of the systems we use today. It's essentially 00:02:19.240 --> 00:02:22.039 the great grandfather of your phone's autocomplete 00:02:22.039 --> 00:02:24.740 feature. He created what mathematicians still 00:02:24.740 --> 00:02:27.370 call a Markov chain. And it didn't just stay 00:02:27.370 --> 00:02:30.110 as math on paper, right? No, it evolved. If we 00:02:30.110 --> 00:02:33.370 jump forward, by the 1970s, an artist named Harold 00:02:33.370 --> 00:02:36.169 Cohen built a computer program called ARIN, and 00:02:36.169 --> 00:02:38.509 ARIN could actually autonomously create physical 00:02:38.509 --> 00:02:41.449 paintings. Which is wild for the 70s. Yeah. And 00:02:41.449 --> 00:02:44.250 then by the 1980s and 90s, the military and aerospace 00:02:44.250 --> 00:02:46.169 sectors were using something called generative 00:02:46.169 --> 00:02:48.569 planning. They used it to map out crisis action 00:02:48.569 --> 00:02:50.750 plans or even prototype autonomous spacecraft. 00:02:51.009 --> 00:02:53.409 But the real turning point, the big shift, that 00:02:53.409 --> 00:02:55.949 happens in the 20 pens with the rise of learning, 00:02:56.229 --> 00:02:59.169 because for a really long time, artificial neural 00:02:59.169 --> 00:03:01.710 networks were primarily discriminative. Yes. 00:03:01.849 --> 00:03:04.430 They categorize things. Right. That means they 00:03:04.430 --> 00:03:07.590 were used for solding data. You feed a computer 00:03:07.590 --> 00:03:10.310 a million pictures, and it learns to discriminate 00:03:10.310 --> 00:03:13.349 between them. It can confidently tell you, yes, 00:03:13.389 --> 00:03:15.669 this is a dog, or no, this is a blueberry muffin. 00:03:15.960 --> 00:03:18.759 which was a big deal at the time, but training 00:03:18.759 --> 00:03:23.000 a model to actually generate a brand new, completely 00:03:23.000 --> 00:03:25.979 original picture of a dog from scratch, that 00:03:25.979 --> 00:03:28.669 was an entirely different. much harder problem. 00:03:28.990 --> 00:03:31.610 Right, until 2014. That's when we see the introduction 00:03:31.610 --> 00:03:34.830 of JANs, Generative Adversarial Networks. Yeah, 00:03:34.830 --> 00:03:37.449 that was a brilliant conceptual leap in how we 00:03:37.449 --> 00:03:39.629 train artificial intelligence. The way I always 00:03:39.629 --> 00:03:43.310 like to visualize a JAN is to imagine like a 00:03:43.310 --> 00:03:45.750 master art forger and a brilliant art detective 00:03:45.750 --> 00:03:47.949 locked in a room together. Oh, that's a great 00:03:47.949 --> 00:03:49.810 way to look at it. Right. So the forger's job 00:03:49.810 --> 00:03:53.009 is to paint fake masterpieces and the detective's 00:03:53.009 --> 00:03:55.729 job is to spot the fakes. And in the beginning, 00:03:55.810 --> 00:03:57.849 you know, the forger is terrible. Just awful 00:03:57.849 --> 00:04:00.169 at it. Yeah, and the detective easily flags the 00:04:00.169 --> 00:04:02.689 fakes. But they do this millions of times. The 00:04:02.689 --> 00:04:05.250 forger learns from every single mistake, constantly 00:04:05.250 --> 00:04:07.710 improving, until eventually the forger's work 00:04:07.710 --> 00:04:09.990 is completely indistinguishable from the real 00:04:09.990 --> 00:04:12.370 thing. And that is essentially what those two 00:04:12.370 --> 00:04:14.110 neural networks are doing inside the computer. 00:04:14.590 --> 00:04:16.649 What's fascinating here is how that adversarial 00:04:16.649 --> 00:04:20.490 setup forced the technology to improve exponentially. 00:04:21.129 --> 00:04:24.170 It pushed the boundaries. But JANs were mostly 00:04:24.170 --> 00:04:26.990 a revelation for generating images, right? Right. 00:04:27.129 --> 00:04:29.250 The real earthquake for language processing happened 00:04:29.250 --> 00:04:32.769 a few years later. In 2017, researchers invented 00:04:32.769 --> 00:04:34.850 the transformer architecture. This is the big 00:04:34.850 --> 00:04:38.509 one. It is. To understand why this was revolutionary, 00:04:39.009 --> 00:04:43.069 you have to look at how AI read text before 2017. 00:04:43.870 --> 00:04:46.970 Older models processed text sequentially. just 00:04:46.970 --> 00:04:49.750 word by word from left to right. Which is incredibly 00:04:49.750 --> 00:04:52.149 slow. So slow. And more importantly, it means 00:04:52.149 --> 00:04:54.529 the model struggles with memory. Like if you 00:04:54.529 --> 00:04:57.029 feed it a really long paragraph by the time it 00:04:57.029 --> 00:04:59.230 gets to the last sentence, it has mathematically 00:04:59.230 --> 00:05:01.769 forgotten the context of the very first sentence. 00:05:01.970 --> 00:05:04.670 Exactly. And the transformer changed that completely. 00:05:05.170 --> 00:05:08.170 It processes entire sequences of tokens, which 00:05:08.170 --> 00:05:10.490 are essentially just chunks of words or syllables 00:05:10.490 --> 00:05:13.550 in parallel. Right. It uses a mechanism called 00:05:13.550 --> 00:05:16.500 self -attention. This allows the model to look 00:05:16.500 --> 00:05:19.079 at a whole sentence at once and determine the 00:05:19.079 --> 00:05:22.360 contextual importance of every single word relative 00:05:22.360 --> 00:05:24.620 to every other word. So it actually understands 00:05:24.620 --> 00:05:27.439 context. Yes. For example, it instantly understands 00:05:27.439 --> 00:05:29.300 the difference between a riverbank and a bank 00:05:29.300 --> 00:05:31.920 account based on the surrounding words. Oh, wow. 00:05:32.019 --> 00:05:34.420 And this meant researchers could scale up training 00:05:34.420 --> 00:05:38.000 exponentially. It moved AI from being a neat 00:05:38.000 --> 00:05:41.420 parlor trick to actually possessing a profound, 00:05:41.639 --> 00:05:44.079 almost eerie level of linguistic understanding. 00:05:44.250 --> 00:05:46.170 And because of that transformer architecture, 00:05:46.350 --> 00:05:48.410 the timeline just suddenly accelerates. I mean, 00:05:48.529 --> 00:05:51.269 we went from spending 110 years in the theoretical 00:05:51.269 --> 00:05:54.290 lab to a complete cultural explosion in about 00:05:54.290 --> 00:05:57.129 36 months. It was incredibly fast. In 2020, an 00:05:57.129 --> 00:06:00.209 anonymous MIT researcher releases 15 .ai, which 00:06:00.209 --> 00:06:03.290 was an early, highly viral audio deep fake tool. 00:06:04.069 --> 00:06:06.769 Then in 2021, Delay showed the general public 00:06:06.769 --> 00:06:09.310 what text to image generation actually looked 00:06:09.310 --> 00:06:11.949 like. And then 2022 was just the watershed year. 00:06:12.029 --> 00:06:14.079 You had mid -journey, stable diffusion. and of 00:06:14.079 --> 00:06:16.980 course, chat GPT. And what the sources make very 00:06:16.980 --> 00:06:20.259 clear is that the hardware running all this software 00:06:20.259 --> 00:06:23.339 is an entire spectrum unto itself. Right, it's 00:06:23.339 --> 00:06:25.860 not all just supercomputers? No. On one extreme, 00:06:26.060 --> 00:06:28.500 yes, you have the massive foundation models like 00:06:28.500 --> 00:06:32.509 GPT -4. These require vast arrays of advanced 00:06:32.509 --> 00:06:36.110 graphics processing units, or GPUs, running in 00:06:36.110 --> 00:06:38.689 enormous, really energy -hungry data centers. 00:06:38.870 --> 00:06:40.370 But then on the other end of the spectrum, you 00:06:40.370 --> 00:06:42.670 have these smaller, highly efficient models. 00:06:43.129 --> 00:06:46.509 The sources mention a 7 billion parameter version 00:06:46.509 --> 00:06:49.569 of LAMA. Which is remarkably capable. Yeah, and 00:06:49.569 --> 00:06:51.550 just to clarify for everyone, parameters are 00:06:51.550 --> 00:06:54.009 essentially the artificial synapses in the network, 00:06:54.129 --> 00:06:55.870 the connections that help it make decisions. 00:06:56.430 --> 00:06:58.930 7 billion sounds like a lot. But the biggest 00:06:58.930 --> 00:07:01.670 models have well over a trillion. But because 00:07:01.670 --> 00:07:04.769 it's smaller, this version of Llama can run locally 00:07:04.769 --> 00:07:08.430 on a $40 Raspberry Pi computer. People are even 00:07:08.430 --> 00:07:10.829 running models like stable diffusion entirely 00:07:10.829 --> 00:07:13.110 on older iPhones. And the reason a user would 00:07:13.110 --> 00:07:15.430 want to run AI locally on their own device isn't 00:07:15.430 --> 00:07:18.769 just to avoid subscription fees. It's fundamentally 00:07:18.769 --> 00:07:21.050 about privacy. Yeah, that makes sense. If you're 00:07:21.050 --> 00:07:23.600 processing sensitive corporate data, or if you 00:07:23.600 --> 00:07:25.779 simply want to avoid corporate censorship and 00:07:25.779 --> 00:07:28.279 rate limits, running a local model gives you 00:07:28.279 --> 00:07:30.819 complete control. And that hardware reality ties 00:07:30.819 --> 00:07:33.279 directly into the geopolitical landscape, too. 00:07:33.620 --> 00:07:36.720 Because those advanced data center chips, specifically 00:07:36.720 --> 00:07:40.480 NVIDIA's H100s, are so critical to building the 00:07:40.480 --> 00:07:42.699 biggest models, the U .S. government actually 00:07:42.699 --> 00:07:46.699 imposed strict export controls in 2022 to restrict 00:07:46.699 --> 00:07:49.319 them from being sold to China. Right. But the 00:07:49.319 --> 00:07:51.759 sources detail how Chinese companies immediately 00:07:51.759 --> 00:07:54.389 develop work Yeah, the Byron chip, right? Exactly, 00:07:54.509 --> 00:07:58.350 the Byron BR -104 chip. The U .S. sanctions specifically 00:07:58.350 --> 00:08:00.790 targeted a metric called interconnect speed, 00:08:01.250 --> 00:08:03.149 which is basically how fast multiple chips can 00:08:03.149 --> 00:08:06.029 talk to each other to share a workload. So the 00:08:06.029 --> 00:08:08.790 Chinese manufacturer simply designed the BR -104 00:08:08.790 --> 00:08:11.230 to physically throttle its own interconnect speed. 00:08:11.670 --> 00:08:13.829 By doing that, they technically complied with 00:08:13.829 --> 00:08:16.649 the exact letter of the U .S. sanctions, legally 00:08:16.649 --> 00:08:18.829 allowing them to be manufactured, while still 00:08:18.829 --> 00:08:21.329 maintaining high enough processing power to fuel 00:08:21.329 --> 00:08:24.639 their domestic AI ambitions. That is wild. And 00:08:24.639 --> 00:08:26.660 culturally, the reaction to all this technology 00:08:26.660 --> 00:08:28.939 has been entirely fractured, depending on where 00:08:28.939 --> 00:08:31.100 you look. Oh, completely. Like, in the West, 00:08:31.279 --> 00:08:33.259 there's been a significant amount of hand wringing 00:08:33.259 --> 00:08:35.940 and anxiety about the future. But in the Asia 00:08:35.940 --> 00:08:38.460 -Pacific region, the sentiment is overwhelmingly 00:08:38.460 --> 00:08:42.009 optimistic. Very much so. A 2024 survey showed 00:08:42.009 --> 00:08:45.009 68 % of people in the Asia -Pacific believed 00:08:45.009 --> 00:08:48.129 AI was having a positive impact compared to only 00:08:48.129 --> 00:08:53.649 57 % globally. And in China, as of 2023, 83 % 00:08:53.649 --> 00:08:56.149 of respondents were actively using generative 00:08:56.149 --> 00:08:59.129 AI. They also lead the world in generative AI 00:08:59.129 --> 00:09:02.970 patents by a staggering margin. Right. Yet, despite 00:09:02.970 --> 00:09:05.750 that enthusiasm on the consumer side, the enterprise 00:09:05.750 --> 00:09:08.970 reality in 2025 has been a bit of a reality check. 00:09:09.880 --> 00:09:12.039 Analysts at Gertner noted that the business world 00:09:12.039 --> 00:09:14.779 has entered what they call the trough of disillusionment. 00:09:14.879 --> 00:09:17.919 The trough of disillusionment. Yeah. Many major 00:09:17.919 --> 00:09:20.340 companies are quietly abandoning their generative 00:09:20.340 --> 00:09:23.059 AI pilot projects entirely. They're reporting 00:09:23.059 --> 00:09:26.080 poor data quality, integration nightmares, and 00:09:26.080 --> 00:09:28.379 a distinct lack of return on their really massive 00:09:28.379 --> 00:09:30.259 investments. Wait, I have to push back on that 00:09:30.259 --> 00:09:32.929 a little. Sure. If the underlying technology 00:09:32.929 --> 00:09:35.769 is as revolutionary as the transformer architecture 00:09:35.769 --> 00:09:39.129 suggests, and if over 80 % of Chinese respondents 00:09:39.129 --> 00:09:42.590 are finding ways to use it daily, why are Western 00:09:42.590 --> 00:09:44.870 corporations suddenly bailing on their pilot 00:09:44.870 --> 00:09:48.169 projects? Like, is the technology itself failing 00:09:48.169 --> 00:09:50.870 to deliver, or are businesses just deploying 00:09:50.870 --> 00:09:53.690 it incorrectly? It really comes down to the difference 00:09:53.690 --> 00:09:55.669 between having an engine and having a working 00:09:55.669 --> 00:09:59.350 car. The transformer model is a remarkably powerful 00:09:59.350 --> 00:10:01.879 engine. But having a powerful engine sitting 00:10:01.879 --> 00:10:03.620 on your driveway does not mean you can drive 00:10:03.620 --> 00:10:06.320 to the grocery store. Right. Integrating generative 00:10:06.320 --> 00:10:09.139 AI into a business requires clean, organized 00:10:09.139 --> 00:10:11.980 data. But real -world corporate data is usually 00:10:11.980 --> 00:10:14.360 siloed, it's messy, it's full of errors. Oh, 00:10:14.379 --> 00:10:17.559 I'm sure. Integrating AI into that mess is a 00:10:17.559 --> 00:10:19.720 staggering engineering challenge. businesses 00:10:19.720 --> 00:10:21.759 are discovering that if their internal databases 00:10:21.759 --> 00:10:24.600 are full of garbage, the AI doesn't magically 00:10:24.600 --> 00:10:27.340 fix it. It just synthesizes that garbage faster 00:10:27.340 --> 00:10:29.519 and more confidently. That makes perfect sense. 00:10:29.700 --> 00:10:31.820 And honestly, that integration friction in the 00:10:31.820 --> 00:10:35.220 enterprise world is quiet compared to what happens 00:10:35.220 --> 00:10:37.840 when this technology collides with the legal 00:10:37.840 --> 00:10:40.779 system. Oh, yeah. Because when you drop a revolutionary 00:10:40.779 --> 00:10:43.620 data synthesis machine into a society that is 00:10:43.620 --> 00:10:45.580 fundamentally built on intellectual property, 00:10:46.139 --> 00:10:48.539 Things start to break. They break very publicly. 00:10:49.240 --> 00:10:52.620 We are currently watching a huge wave of high 00:10:52.620 --> 00:10:55.240 -stakes lawsuits unfold. The New York Times, 00:10:55.639 --> 00:10:58.820 Getty Images, the Authors Guild, they are all 00:10:58.820 --> 00:11:01.539 suing tech giants like Microsoft and OpenAI. 00:11:01.779 --> 00:11:04.840 And the defense there is fair use. Exactly. The 00:11:04.840 --> 00:11:07.899 core defense from the AI developers is that training 00:11:07.899 --> 00:11:10.500 their models on publicly available copyrighted 00:11:10.500 --> 00:11:13.500 data falls under fair use. They argue that the 00:11:13.500 --> 00:11:16.720 AI is learning concepts, not copying text, making 00:11:16.720 --> 00:11:19.460 it a highly transformative process. But the copyright 00:11:19.460 --> 00:11:20.960 holders are countering with a very different 00:11:20.960 --> 00:11:23.279 perspective. They're essentially arguing, you 00:11:23.279 --> 00:11:25.620 ingested our entire life's work without permission 00:11:25.620 --> 00:11:28.799 or compensation to build a machine that now directly 00:11:28.799 --> 00:11:30.879 competes with us in the marketplace. Yes. And 00:11:30.879 --> 00:11:33.539 beyond just the training data, there is the deeply 00:11:33.539 --> 00:11:36.919 confusing flip side. Who actually owns the output? 00:11:37.139 --> 00:11:39.360 Right. The outputs are a huge legal gray area. 00:11:39.539 --> 00:11:42.200 Initially, the U .S. Copyright Office ruled that 00:11:42.200 --> 00:11:46.080 works created entirely by AI lack human authorship 00:11:46.080 --> 00:11:48.639 and therefore cannot be copyrighted at all. They 00:11:48.639 --> 00:11:51.980 even referenced that famous 2018 legal case where 00:11:51.980 --> 00:11:53.840 a monkey took a selfie with a photographer's 00:11:53.840 --> 00:11:56.740 camera and the court ruled that non -humans cannot 00:11:56.740 --> 00:11:59.220 hold copyrights. But the law is really struggling 00:11:59.220 --> 00:12:02.080 to keep up and it's shifting constantly. By early 00:12:02.080 --> 00:12:05.100 2025, the Copyright Office had to release new, 00:12:05.480 --> 00:12:07.879 highly nuanced guidance. Which changed everything. 00:12:07.899 --> 00:12:10.889 It did. They stated that if a human user exerts 00:12:10.889 --> 00:12:13.009 a significant amount of control over the selection 00:12:13.009 --> 00:12:15.590 and placement of the creative elements, the resulting 00:12:15.590 --> 00:12:18.029 output might actually be copyrightable on a strict 00:12:18.029 --> 00:12:20.789 case -by -case basis. Which leads to one of the 00:12:20.789 --> 00:12:23.230 most surreal details on this entire deep dive. 00:12:23.559 --> 00:12:26.899 the 2025 American Cheese update. Yes. The U .S. 00:12:27.159 --> 00:12:29.179 Copyright Office officially registered a visual 00:12:29.179 --> 00:12:32.039 artwork titled, A Single Piece of American Cheese. 00:12:32.600 --> 00:12:35.320 It became the very first visual artwork composed 00:12:35.320 --> 00:12:37.840 entirely of AI -generated materials to actually 00:12:37.840 --> 00:12:40.559 receive a copyright. And it only happened because 00:12:40.559 --> 00:12:42.980 the user extensively documented their process. 00:12:43.360 --> 00:12:46.139 Right. They showed that they spent hours tweaking 00:12:46.139 --> 00:12:48.980 specific prompts, adjusting sliders, and dictating 00:12:48.980 --> 00:12:53.059 the precise placement of the pixels. If I casually 00:12:53.059 --> 00:12:56.700 type draw a dog into mid -journey, I don't own 00:12:56.700 --> 00:12:59.059 the result. Right. But if I spend three hours 00:12:59.059 --> 00:13:01.399 micromanaging the generation of a picture of 00:13:01.399 --> 00:13:03.759 American cheese, suddenly I am the legal author. 00:13:04.080 --> 00:13:06.759 Like where exactly is the legal line between 00:13:06.759 --> 00:13:09.820 using a tool and being an author? If we connect 00:13:09.820 --> 00:13:12.059 this to the bigger picture, you can see how our 00:13:12.059 --> 00:13:13.940 legal frameworks are currently doing complete 00:13:13.940 --> 00:13:17.019 gymnastics. We are trying to apply 20th century 00:13:17.019 --> 00:13:19.580 copyright law. which is built around printing 00:13:19.580 --> 00:13:22.659 press, cameras, and human illustrators, to 21st 00:13:22.659 --> 00:13:24.919 century statistical probability models. It just 00:13:24.919 --> 00:13:27.980 doesn't fit. It's a profound mismatch. And we 00:13:27.980 --> 00:13:30.399 are seeing that exact same tension play out in 00:13:30.399 --> 00:13:32.500 the labor market. Yeah, the labor impact data 00:13:32.500 --> 00:13:34.759 is incredibly contradictory. On one side, you 00:13:34.759 --> 00:13:38.220 have very real localized pain. Reports from China 00:13:38.220 --> 00:13:41.659 in 2023 showed that 70 % of video game illustrators 00:13:41.659 --> 00:13:44.639 lost their jobs to image generation AI. 70%. 00:13:44.639 --> 00:13:47.820 That's massive. It is. And we all remember the 00:13:47.820 --> 00:13:50.320 2023 Hollywood strikes, where the Writers Guild 00:13:50.320 --> 00:13:52.600 and the Screen Actors Guild essentially went 00:13:52.600 --> 00:13:56.120 to war over AI protections, calling it an existential 00:13:56.120 --> 00:13:58.700 threat to creative professions. Yet, if you zoom 00:13:58.700 --> 00:14:01.379 out and look at the macroeconomic data, the narrative 00:14:01.379 --> 00:14:04.659 shifts entirely. A 2025 study concluded that 00:14:04.659 --> 00:14:07.100 the overall U .S. labor market had experienced 00:14:07.100 --> 00:14:10.000 absolutely no discernible disruption from generative 00:14:10.000 --> 00:14:12.639 AI. None at all. None. Another study followed 00:14:12.639 --> 00:14:14.720 Danish workers who were given access to chat 00:14:14.720 --> 00:14:18.159 bots. It found they saved about 2 .8 % of their 00:14:18.159 --> 00:14:20.860 time on average, with no significant change in 00:14:20.860 --> 00:14:22.980 their earnings or hours worked. So it's not replacing 00:14:22.980 --> 00:14:25.519 them. Right. The discrepancy exists because, 00:14:25.519 --> 00:14:28.600 right now, AI is generally replacing specific 00:14:28.600 --> 00:14:32.519 tasks rather than entire James. Its disruption 00:14:32.519 --> 00:14:35.019 is heavily isolated to highly specific micro 00:14:35.019 --> 00:14:37.539 -industries like concept illustration. The macro 00:14:37.539 --> 00:14:39.759 -level job apocalypse just hasn't materialized 00:14:39.759 --> 00:14:43.220 yet. OK, but if the legal and labor worlds are 00:14:43.220 --> 00:14:45.399 struggling to handle who owns the content and 00:14:45.399 --> 00:14:47.860 who makes the content, the internet itself is 00:14:47.860 --> 00:14:50.059 struggling to survive the sheer volume of what 00:14:50.059 --> 00:14:52.879 is being produced. Here's where it gets really 00:14:52.879 --> 00:14:55.940 interesting. And honestly, a little dark. We 00:14:55.940 --> 00:14:57.820 need to talk about SLOP. SLOP is the perfect 00:14:57.820 --> 00:15:00.440 term for it. It is the Generative AI Equivalent 00:15:00.440 --> 00:15:04.340 of Spam. yeah it refers to the shoddy unwanted 00:15:04.340 --> 00:15:07.460 AI -generated text and images that are currently 00:15:07.460 --> 00:15:10.059 flooding search engine results, social media 00:15:10.059 --> 00:15:12.679 feeds, and even published scientific papers. 00:15:12.879 --> 00:15:15.519 The scale of the slop is just staggering. A research 00:15:15.519 --> 00:15:18.019 paper from Amazon Web Services found that over 00:15:18.019 --> 00:15:21.320 57 % of sentences on a vast sample of the web 00:15:21.320 --> 00:15:23.860 had been machine translated. And usually poorly 00:15:23.860 --> 00:15:26.360 translated at that. Right. It has gotten so pervasive 00:15:26.360 --> 00:15:29.519 that in 2024, a highly respected open source 00:15:29.519 --> 00:15:32.639 database called WordFreak had to shut down entirely. 00:15:32.809 --> 00:15:35.289 That was a huge loss for researchers. It was. 00:15:35.730 --> 00:15:38.309 This database calculated word frequencies across 00:15:38.309 --> 00:15:40.590 the internet to help researchers understand human 00:15:40.590 --> 00:15:43.169 language. But the creator shut it down because 00:15:43.169 --> 00:15:45.710 generative AI had permanently polluted the data. 00:15:46.230 --> 00:15:48.870 Words like delve were suddenly everywhere because 00:15:48.870 --> 00:15:51.830 chat GPT overuses them, which completely skewed 00:15:51.830 --> 00:15:53.950 our measurement of authentic human language. 00:15:54.070 --> 00:15:56.409 And that data pollution leads to a severe technical 00:15:56.409 --> 00:15:59.129 vulnerability known as model collapse. Model 00:15:59.129 --> 00:16:02.500 collapse. Yeah. When an AI model is trained on 00:16:02.500 --> 00:16:05.419 authentic human data, it works wonderfully. But 00:16:05.419 --> 00:16:08.539 as the internet sills with AI slop, the next 00:16:08.539 --> 00:16:11.799 generation of AI models ends up unknowingly training 00:16:11.799 --> 00:16:14.399 on the output of previous AI models. Like an 00:16:14.399 --> 00:16:17.639 echo chamber. Exactly. It's like taking a photocopy 00:16:17.639 --> 00:16:20.679 of a photocopy of a photocopy. The model gradually 00:16:20.679 --> 00:16:23.559 loses the quirks, the nuances, and the specific 00:16:23.559 --> 00:16:26.700 outliers of human behavior. If you repeat that 00:16:26.700 --> 00:16:29.620 cycle recursively, the model degrades until it 00:16:29.620 --> 00:16:32.139 collapses into generic garbled mush. Are we just 00:16:32.139 --> 00:16:34.299 building a digital Auroboros, you know, a snake 00:16:34.299 --> 00:16:36.779 eating its own tail where AI churns out slop, 00:16:36.940 --> 00:16:39.100 trains on its own slop, and burns down the planet's 00:16:39.100 --> 00:16:40.919 energy grid just to give us a lower quality version 00:16:40.919 --> 00:16:43.200 of the Internet? This raises an important question, 00:16:43.259 --> 00:16:45.850 especially regarding the energy grid. Because 00:16:45.850 --> 00:16:48.049 the environmental toll of keeping these massive 00:16:48.049 --> 00:16:51.590 systems running is immense. The research institute 00:16:51.590 --> 00:16:55.610 EPOC AI noted that a single chat GPT query uses 00:16:55.610 --> 00:16:59.169 about 0 .3 watt hours. Which doesn't sound like 00:16:59.169 --> 00:17:01.570 a lot. It might not sound like much in isolation, 00:17:01.909 --> 00:17:04.769 but it is roughly 10 times the energy required 00:17:04.769 --> 00:17:07.289 for a standard Google search. When you multiply 00:17:07.289 --> 00:17:09.789 that by hundreds of millions of users daily, 00:17:10.329 --> 00:17:12.710 the scale of energy consumption becomes a serious 00:17:12.710 --> 00:17:14.680 liability. And the projections in the sources 00:17:14.680 --> 00:17:18.119 are sobering. It's estimated that by 2035 the 00:17:18.119 --> 00:17:20.960 global carbon footprint of generative AI could 00:17:20.960 --> 00:17:24.720 reach up to 245 million tons of CO2 annually. 00:17:25.160 --> 00:17:27.619 Yeah. To put that abstract number into perspective, 00:17:28.059 --> 00:17:30.579 that rivals the entire carbon footprint of the 00:17:30.579 --> 00:17:32.640 United States beef industry. Which is exactly 00:17:32.640 --> 00:17:35.180 why researchers are urgently pushing for mitigation 00:17:35.180 --> 00:17:37.799 strategies. They're advocating for strict audits 00:17:37.799 --> 00:17:39.940 and environmental impacts and exploring ways 00:17:39.940 --> 00:17:42.380 to use synthetic data safely without triggering 00:17:42.380 --> 00:17:43.960 that model collapse we talked about. There's 00:17:43.960 --> 00:17:45.920 also a major push for detection tools, right? 00:17:46.099 --> 00:17:49.299 Yes. Google, for instance, deployed a tool called 00:17:49.299 --> 00:17:53.980 Synthed in 2025. It subtly watermarks AI -generated 00:17:53.980 --> 00:17:57.039 text, images, and video directly at the source, 00:17:57.440 --> 00:18:00.059 aiming to help platforms identify and label synthetic 00:18:00.059 --> 00:18:02.759 content before it spreads. And reliable detection 00:18:02.759 --> 00:18:05.319 is critical because the cybercrime and disinformation 00:18:05.319 --> 00:18:08.019 applications are escalating rapidly. We aren't 00:18:08.019 --> 00:18:10.420 just talking about teenagers using chatbots to 00:18:10.420 --> 00:18:12.559 write high school essays anymore. Not at all. 00:18:13.019 --> 00:18:15.000 Cybercriminals are building their own dedicated 00:18:15.000 --> 00:18:18.140 malicious models. Tools with names like Worm 00:18:18.140 --> 00:18:21.619 GPT and Fraud GPT exist specifically to automate 00:18:21.619 --> 00:18:24.000 phishing attacks and social engineering scams 00:18:24.000 --> 00:18:26.099 at an industrial scale. And at the state level, 00:18:26.200 --> 00:18:28.039 the technology is being leveraged for information 00:18:28.039 --> 00:18:30.839 laundering. Generative AI allows state -sponsored 00:18:30.839 --> 00:18:33.160 actors to disguise the origin of propaganda by 00:18:33.160 --> 00:18:35.240 rapidly rewriting it and distributing it across 00:18:35.240 --> 00:18:37.539 thousands of fake social media accounts. The 00:18:37.539 --> 00:18:39.720 sources actually highlight two very specific 00:18:39.720 --> 00:18:43.140 examples of this. In 2025, a research group called 00:18:43.140 --> 00:18:45.859 the American Sunlight Project reported on a network 00:18:45.869 --> 00:18:48.690 known as Pravda. This network was publishing 00:18:48.690 --> 00:18:51.230 up to 10 ,000 articles a day to push Russian 00:18:51.230 --> 00:18:53.990 narratives, specifically aiming to get that text 00:18:53.990 --> 00:18:56.750 absorbed into the training data of large language 00:18:56.750 --> 00:19:00.670 models. Right. Similarly, in 2025, it was reported 00:19:00.670 --> 00:19:03.349 that Israel signed a $6 million contract with 00:19:03.349 --> 00:19:07.210 a U .S. firm called Clock Tower X. The stated 00:19:07.210 --> 00:19:09.809 goal was to flood social platforms with specific 00:19:09.809 --> 00:19:12.170 information to intentionally influence models 00:19:12.170 --> 00:19:15.049 like chat GPT through a process called a rag 00:19:15.079 --> 00:19:17.960 To understand RAG poisoning, we have to look 00:19:17.960 --> 00:19:20.980 at the mechanics. RAG stands for Retrieval Augmented 00:19:20.980 --> 00:19:23.960 Generation. OK. Normally, an AI only knows the 00:19:23.960 --> 00:19:26.440 data it was initially trained on. But our key 00:19:26.440 --> 00:19:28.559 allows the chatbot to search the live internet 00:19:28.559 --> 00:19:30.880 to answer your questions in real time. So it 00:19:30.880 --> 00:19:33.559 pulls in fresh info. Exactly. But if a state 00:19:33.559 --> 00:19:35.720 actor successfully flugs the internet with thousands 00:19:35.720 --> 00:19:38.359 of fabricated articles, the AI chatbot searches 00:19:38.359 --> 00:19:41.000 the web, unknowingly retrieves those fake articles, 00:19:41.539 --> 00:19:43.759 and summarizes them as objective truth for the 00:19:43.759 --> 00:19:45.819 user. It highlights how the new battleground 00:19:45.819 --> 00:19:48.400 is not just about generating text. It's about 00:19:48.400 --> 00:19:51.220 manipulating the underlying databases of knowledge 00:19:51.220 --> 00:19:54.339 that the world relies on for answers. OK, let's 00:19:54.339 --> 00:19:56.819 take a breath. We have covered a vast amount 00:19:56.819 --> 00:19:59.019 of ground today. We traced the mathematical roots 00:19:59.019 --> 00:20:01.599 all the way back to Andrei Markov calculating 00:20:01.599 --> 00:20:05.000 vowel probabilities in 1906. We explored the 00:20:05.000 --> 00:20:07.880 2017 transformer breakthrough that allowed AI 00:20:07.880 --> 00:20:10.779 to understand context by reading words in parallel. 00:20:11.079 --> 00:20:13.619 We really went through it all. We navigated the 00:20:13.619 --> 00:20:17.519 incredibly complex 2026 landscape where a picture 00:20:17.519 --> 00:20:20.920 of American cheese can secure a copyright. Concept 00:20:20.920 --> 00:20:23.700 artists face significant job losses and the internet 00:20:23.700 --> 00:20:26.519 is slowly filling with automated slop while consuming 00:20:26.519 --> 00:20:28.900 the energy equivalent of the U .S. beef industry. 00:20:29.390 --> 00:20:32.210 So what does this all mean? It means we are navigating 00:20:32.210 --> 00:20:35.130 a profound societal transition. The technology 00:20:35.130 --> 00:20:38.289 is undeniably capable, but our legal, social, 00:20:38.349 --> 00:20:40.529 and technical infrastructures simply were not 00:20:40.529 --> 00:20:42.529 built to handle a machine that can synthesize 00:20:42.529 --> 00:20:45.970 human output at this scale. The trough of disillusionment 00:20:45.970 --> 00:20:47.950 the corporate world is experiencing right now 00:20:47.950 --> 00:20:50.589 is really just society trying to build the guardrails 00:20:50.589 --> 00:20:53.549 to safely contain and utilize this engine. I 00:20:53.549 --> 00:20:55.609 keep coming back to that model collapse problem, 00:20:55.930 --> 00:20:59.329 the snake eating its own tail. If AI models actively 00:20:59.329 --> 00:21:01.549 degrade when they train on their own synthetic 00:21:01.549 --> 00:21:04.789 slop, and if 57 % of the web is already machine 00:21:04.789 --> 00:21:07.410 translated, what happens when the AI companies 00:21:07.410 --> 00:21:09.849 simply run out of fresh human -created data? 00:21:09.970 --> 00:21:12.250 It fundamentally changes the value of human creation. 00:21:12.470 --> 00:21:17.250 Exactly. Will pre -2022 verifiably human -made 00:21:17.250 --> 00:21:20.750 text, art, and data suddenly become the most 00:21:20.750 --> 00:21:24.289 valuable scarce commodity on earth? Will we one 00:21:24.289 --> 00:21:25.970 day reach a point where we have to pay a premium 00:21:25.970 --> 00:21:27.710 subscription just to read an article or look 00:21:27.710 --> 00:21:30.190 at an illustration that we know for an absolute 00:21:30.190 --> 00:21:32.869 fact a human being actually spent the time to 00:21:32.869 --> 00:21:34.869 create? It's entirely possible. When you look 00:21:34.869 --> 00:21:37.430 past the magic trick and past the trap doors, 00:21:37.829 --> 00:21:40.529 maybe the real revelation is that the messy slow 00:21:40.529 --> 00:21:43.049 human mind was the most valuable component all 00:21:43.049 --> 00:21:43.329 along.