WEBVTT 00:00:00.000 --> 00:00:04.700 Imagine finding out that the flawless legal brief 00:00:04.700 --> 00:00:07.360 defending you in federal court was just, well, 00:00:07.839 --> 00:00:10.240 entirely hallucinated by a machine. Yeah, that's 00:00:10.240 --> 00:00:13.220 a terrifying thought. Right. Or consider this. 00:00:13.820 --> 00:00:16.579 Every single time you ask your AI assistant to 00:00:16.579 --> 00:00:20.059 just draft a few routine emails, that software 00:00:20.059 --> 00:00:23.300 is literally consuming half a liter of fresh 00:00:23.300 --> 00:00:25.850 water. just to cool the servers down. It's just 00:00:25.850 --> 00:00:28.190 for a few prompts. Exactly. I mean, today is 00:00:28.190 --> 00:00:31.550 Thursday, March 26, 2026, and we are basically 00:00:31.550 --> 00:00:34.090 living in a reality that would have sounded like 00:00:34.090 --> 00:00:36.609 pure science fiction, like just a few short years 00:00:36.609 --> 00:00:39.670 ago. Oh, totally. If you think back to late 2022, 00:00:40.070 --> 00:00:43.189 chat GPT was just this... It was a viral novelty. 00:00:43.469 --> 00:00:45.509 Yeah, quirky little toy. Right, exactly. You'd 00:00:45.509 --> 00:00:47.409 log on, type a prompt, and it would spit out 00:00:47.409 --> 00:00:50.270 a slightly clunky poem about your dog or something. 00:00:50.329 --> 00:00:52.509 Yeah, as a parlor trick. I mean, a really sophisticated 00:00:52.509 --> 00:00:54.590 parlor trick, sure. But still, just something 00:00:54.590 --> 00:00:56.289 you played around with in a browser window to 00:00:56.289 --> 00:00:57.929 see what it could do. To test the limits, yeah. 00:00:58.250 --> 00:01:01.049 But that novelty era. It is completely gone. 00:01:01.210 --> 00:01:04.849 I mean, today, ChatGBT is a deeply embedded digital 00:01:04.849 --> 00:01:07.250 worker. It operates across our web browsers, 00:01:07.750 --> 00:01:10.090 it analyzes our private health records, and it 00:01:10.090 --> 00:01:12.950 literally, literally has direct access to our 00:01:12.950 --> 00:01:15.310 wallets. Which is a staggering transition when 00:01:15.310 --> 00:01:17.189 you really look at the timeline. The speed of 00:01:17.189 --> 00:01:19.959 it is just insane. And that is exactly why we 00:01:19.959 --> 00:01:22.159 are doing this today. For this deep dive, we 00:01:22.159 --> 00:01:24.939 are looking at a massive, incredibly comprehensive 00:01:24.939 --> 00:01:28.180 Wikipedia breakdown of Chad GPT's evolution. 00:01:28.620 --> 00:01:30.700 And it is so up to date, too. Yeah, it covers 00:01:30.700 --> 00:01:32.439 everything. We're looking at the foundational 00:01:32.439 --> 00:01:35.299 features, the underlying models, the major controversies 00:01:35.299 --> 00:01:37.620 spanning right from that initial launch, all 00:01:37.620 --> 00:01:40.980 the way up to the brand new GPT 5 .4 model that 00:01:40.980 --> 00:01:42.640 just dropped, what, a few weeks ago? Yeah, just 00:01:42.640 --> 00:01:45.060 a few weeks ago. And what makes the specific 00:01:45.060 --> 00:01:47.579 source document so valuable for you, the listener? 00:01:47.819 --> 00:01:50.760 is that it completely strips away all that Silicon 00:01:50.760 --> 00:01:53.859 Valley marketing hype. Right, no PR spin. Exactly. 00:01:54.120 --> 00:01:56.840 It just presents the raw chronological timeline 00:01:56.840 --> 00:02:00.540 of how this system grew, how the underlying architecture 00:02:00.540 --> 00:02:03.540 fundamentally changed, and how we as a society 00:02:03.540 --> 00:02:07.099 have been, well... forced to adapt to a technology 00:02:07.099 --> 00:02:09.659 that scales way faster than our own loss. Way 00:02:09.659 --> 00:02:12.680 faster. So our mission today is to map out that 00:02:12.680 --> 00:02:15.300 explosive transformation for you. We really want 00:02:15.300 --> 00:02:17.699 to understand how chat GPT went from just a simple 00:02:17.699 --> 00:02:20.740 text generator to an autonomous digital agent. 00:02:20.969 --> 00:02:23.210 And of course, to uncover the real world ripple 00:02:23.210 --> 00:02:26.550 effects this is having on our society, our politics, 00:02:26.650 --> 00:02:28.990 and just our day -to -day lives. It impacts almost 00:02:28.990 --> 00:02:31.110 everything now. It really does. Okay, let's untack 00:02:31.110 --> 00:02:33.129 this. And I think we should start with the most 00:02:33.129 --> 00:02:35.210 fundamental shift in its identity over the last 00:02:35.210 --> 00:02:37.409 year or so. The fact that it doesn't just, you 00:02:37.409 --> 00:02:39.289 know, talk to you anymore. Right, it takes action. 00:02:39.430 --> 00:02:41.389 It actually does things on your behalf. And this 00:02:41.389 --> 00:02:45.289 is really the crux of the 2025 shift. Open AI 00:02:45.289 --> 00:02:47.409 moved super aggressively away from the concept 00:02:47.409 --> 00:02:49.650 of a passive conversationalist, right? And they 00:02:49.650 --> 00:02:51.759 moved toward what the industry calls agentic 00:02:51.759 --> 00:02:54.240 capabilities. Agentic, like acting as an agent. 00:02:54.560 --> 00:02:57.300 Exactly. So in January of 2025, they released 00:02:57.300 --> 00:03:00.219 this feature called Operator. And Operator wasn't 00:03:00.219 --> 00:03:02.639 just generating text for you to copy and paste. 00:03:02.879 --> 00:03:06.259 It was actively navigating web browsers. Which 00:03:06.259 --> 00:03:08.500 blew my mind when it came out. Oh, it was wild. 00:03:08.539 --> 00:03:10.419 It could fill out online forms, it could place 00:03:10.419 --> 00:03:12.979 orders, it could schedule appointments, all by 00:03:12.979 --> 00:03:15.719 controlling a software environment inside a virtual 00:03:15.719 --> 00:03:17.979 machine. OK, but I remember reading about that 00:03:17.979 --> 00:03:20.210 and thinking, how does that actually work in 00:03:20.210 --> 00:03:21.990 practice? Because when I look at a website, I 00:03:21.990 --> 00:03:26.330 see buttons and pictures and text. A chat bot 00:03:26.330 --> 00:03:29.469 doesn't have eyes. How does it see the internet 00:03:29.469 --> 00:03:31.969 to navigate around it? Well, it doesn't see the 00:03:31.969 --> 00:03:34.469 screen the way we do, obviously. When an agent 00:03:34.469 --> 00:03:36.650 -like operator looks at a website, it's actually 00:03:36.650 --> 00:03:40.289 parsing the underlying HTML code. Oh, yeah. It 00:03:40.289 --> 00:03:42.250 analyzes something called the Document Object 00:03:42.250 --> 00:03:44.669 Model, or the DOM. The DOM, right. Right. So 00:03:44.669 --> 00:03:46.669 it reads the raw code of the page to figure out 00:03:46.669 --> 00:03:48.610 exactly where the interactive elements are. And 00:03:48.610 --> 00:03:51.250 then, and this is a crazy part, it mathematically 00:03:51.250 --> 00:03:54.030 predicts the exact x and y coordinates on the 00:03:54.030 --> 00:03:57.009 screen to simulate a human mouse click. Or a 00:03:57.009 --> 00:04:00.289 keystroke. It's all math. And then a few months 00:04:00.289 --> 00:04:03.729 later, in May of 2025, they introduced Codex, 00:04:04.490 --> 00:04:06.590 which was an agent specifically built to do this 00:04:06.590 --> 00:04:09.569 exact thing, but for software engineering. So 00:04:09.569 --> 00:04:12.050 it was coding itself? Basically, yeah. Codex 00:04:12.050 --> 00:04:14.509 can write software, run its own tests, and even 00:04:14.509 --> 00:04:17.290 propose code changes autonomously. And then... 00:04:17.160 --> 00:04:21.079 By July of last year, we got the chat GPT agent, 00:04:21.360 --> 00:04:23.819 the one that performs multi -step tasks across 00:04:23.819 --> 00:04:26.379 a virtual computer. I mean, my analogy for this 00:04:26.379 --> 00:04:29.100 is always it's gone from an intern who just hands 00:04:29.100 --> 00:04:31.759 you a printed recipe to an intern who takes your 00:04:31.759 --> 00:04:34.100 actual credit card, drives to the grocery store, 00:04:34.579 --> 00:04:37.300 walks the aisles. buys the food, and then comes 00:04:37.300 --> 00:04:39.199 back and cooks the meal for you. That is honestly 00:04:39.199 --> 00:04:40.720 the perfect analogy because they didn't just 00:04:40.720 --> 00:04:42.819 stop at virtual machines either. They integrated 00:04:42.819 --> 00:04:45.180 this capability directly into the everyday tools 00:04:45.180 --> 00:04:49.139 we use. By October 2025, they launched ChatGPT 00:04:49.139 --> 00:04:52.000 Atlas. Right, the browser. Yeah, their own dedicated 00:04:52.000 --> 00:04:54.740 web browser that has this agentic mode baked 00:04:54.740 --> 00:04:57.639 right into the navigation bar. But see, if they 00:04:57.639 --> 00:04:59.860 are integrating this into everyday browsing, 00:05:00.399 --> 00:05:02.670 they're also integrating it into our money. The 00:05:02.670 --> 00:05:04.790 source timeline makes a huge point of noting 00:05:04.790 --> 00:05:09.110 that in September of 2025, OpenAI rolled out 00:05:09.110 --> 00:05:12.129 the agent at commerce protocol. Yes, the partnerships. 00:05:12.269 --> 00:05:15.129 Right. They partnered with Stripe and Etsy, which 00:05:15.129 --> 00:05:18.589 means the AI can literally make purchases using 00:05:18.589 --> 00:05:21.350 your linked payment method. The AI isn't just 00:05:21.350 --> 00:05:23.550 adjusting a product anymore. It is authorizing 00:05:23.550 --> 00:05:25.930 and completely finishing the financial transaction. 00:05:26.569 --> 00:05:28.889 And by the way, OpenAI even takes a cut from 00:05:28.889 --> 00:05:31.589 the merchant's payment as a facilitation fee. 00:05:31.759 --> 00:05:33.980 I gotta push back on this though, because giving 00:05:33.980 --> 00:05:36.699 a chatbot direct access to my wallet and my web 00:05:36.699 --> 00:05:38.920 browser, that sounds like a massive security 00:05:38.920 --> 00:05:40.680 nightmare just waiting to happen. Oh, people 00:05:40.680 --> 00:05:42.680 were terrified of it. Because the internet is 00:05:42.680 --> 00:05:44.800 so messy, it's chaotic, it's full of deceptive 00:05:44.800 --> 00:05:46.620 pop -ups, the layouts are constantly changing, 00:05:46.939 --> 00:05:49.540 buttons are weirdly coded. If operators struggled 00:05:49.540 --> 00:05:52.480 with basic user interfaces back in early 2025, 00:05:53.180 --> 00:05:55.759 how on earth did OpenAI convince companies like 00:05:55.759 --> 00:05:58.379 Stripe to trust it with actual money just a few 00:05:58.379 --> 00:06:00.670 months later? Well, the friction you're describing, 00:06:00.930 --> 00:06:02.970 that is exactly what the developers were wrestling 00:06:02.970 --> 00:06:06.170 with behind the scenes. Early versions of Operator 00:06:06.170 --> 00:06:08.610 were kept in highly restricted virtual machines 00:06:08.610 --> 00:06:11.329 for those exact safety reasons. Because it would 00:06:11.329 --> 00:06:14.439 just click the wrong thing. Basically. It would 00:06:14.439 --> 00:06:17.839 get confused by poorly designed websites. If 00:06:17.839 --> 00:06:21.259 a site's DOM code is messy, the AI's mathematical 00:06:21.259 --> 00:06:23.839 prediction of where to click just breaks down 00:06:23.839 --> 00:06:27.120 entirely. But what's fascinating here is the 00:06:27.120 --> 00:06:29.740 psychological shift this demands from you, the 00:06:29.740 --> 00:06:32.879 user. In what way? Like, trusting it. Exactly. 00:06:33.180 --> 00:06:35.180 Think about it. When you use a traditional search 00:06:35.180 --> 00:06:37.540 engine, your role is just a searcher of information. 00:06:37.759 --> 00:06:40.199 You type a query, you get a list of links, and 00:06:40.199 --> 00:06:42.100 you do all the work of evaluating them. Right, 00:06:42.100 --> 00:06:44.100 I have to click and read. Right. But with these 00:06:44.100 --> 00:06:46.180 agentic models, your role shifts completely. 00:06:46.339 --> 00:06:48.699 You become the manager of an autonomous entity. 00:06:48.980 --> 00:06:50.899 You're delegating a task, you're assigning it 00:06:50.899 --> 00:06:52.959 a financial budget, and you're reviewing its 00:06:52.959 --> 00:06:55.560 work after the fact. That is so true. Managing 00:06:55.560 --> 00:06:57.680 an autonomous system requires a fundamentally 00:06:57.680 --> 00:06:59.920 different level of trust than simply reading 00:06:59.920 --> 00:07:03.420 a text summary on a screen. So to earn that trust, 00:07:03.920 --> 00:07:07.040 I mean to be able to flawlessly parse a chaotic 00:07:07.040 --> 00:07:10.560 web page, bypass a weird pop -up, and buy a pair 00:07:10.560 --> 00:07:13.459 of shoes without accidentally spending a thousand 00:07:13.459 --> 00:07:17.879 dollars, the underlying brain of this AI had 00:07:17.879 --> 00:07:20.459 to get significantly smarter. It couldn't just 00:07:20.459 --> 00:07:22.959 guess the next word anymore. No, it had to actually 00:07:22.959 --> 00:07:26.100 think. Right. And that required an entirely new 00:07:26.100 --> 00:07:29.040 architecture. We started seeing this transition 00:07:29.040 --> 00:07:31.959 with the O series models. Oh right, like O1 and 00:07:31.959 --> 00:07:34.819 O3. Exactly. These were explicitly designed as 00:07:34.819 --> 00:07:37.160 reasoning models. They utilized something called 00:07:37.160 --> 00:07:39.990 test time compute. Test time compute. OK, let's 00:07:39.990 --> 00:07:42.089 break that down for someone listening who isn't 00:07:42.089 --> 00:07:44.670 a machine learning engineer. What does that actually 00:07:44.670 --> 00:07:47.290 mean in practical terms? So in oldie models, 00:07:47.449 --> 00:07:49.470 when you asked a question, the AI would just 00:07:49.470 --> 00:07:51.449 immediately start generating the answer word 00:07:51.449 --> 00:07:53.790 by word as fast as humanly possible. Like it's 00:07:53.790 --> 00:07:56.350 just reacting. Right. But with test time compute, 00:07:56.550 --> 00:07:58.910 the model actually pauses. It generates thousands 00:07:58.910 --> 00:08:01.329 of hidden thoughts behind the scenes before it 00:08:01.329 --> 00:08:03.569 ever shows you a single word. So it's drafting 00:08:03.569 --> 00:08:07.810 a plan. Yes. It proposes a plan, it tests that 00:08:07.810 --> 00:08:10.370 plan against its own logic, it realizes it made 00:08:10.370 --> 00:08:12.990 a mistake, self -corrects, and then finalizes 00:08:12.990 --> 00:08:16.310 a strategy. Only after it has completely reasoned 00:08:16.310 --> 00:08:18.569 through the problem does it actually output the 00:08:18.569 --> 00:08:20.730 final answer or, you know, take an action on 00:08:20.730 --> 00:08:23.779 a website. which perfectly brings us to August 00:08:23.779 --> 00:08:28.339 2025 and that massive GPT -5 release. The notes 00:08:28.339 --> 00:08:30.879 from the Wikipedia breakdown indicate that GPT 00:08:30.879 --> 00:08:33.639 -5 wasn't just like a bigger version of the old 00:08:33.639 --> 00:08:35.840 brain. It was structured completely differently. 00:08:35.980 --> 00:08:38.559 Yeah, GPT -5 utilized a concept known as a mixture 00:08:38.559 --> 00:08:41.419 of experts or MOE. A mixture of experts. Right. 00:08:41.659 --> 00:08:44.279 So instead of one monolithic, dense neural network 00:08:44.279 --> 00:08:46.279 where every single connection fires for every 00:08:46.279 --> 00:08:49.279 single question you ask, GPT -5 acts as a highly 00:08:49.279 --> 00:08:51.539 efficient router. Oh, like a traffic. Exactly. 00:08:51.840 --> 00:08:54.159 When you give it a prompt, this router evaluates 00:08:54.159 --> 00:08:56.340 what you actually need. If you ask it a simple 00:08:56.340 --> 00:08:58.700 history question, it wraps that query to a smaller, 00:08:58.980 --> 00:09:00.539 specialized sub -model. And doesn't need the 00:09:00.539 --> 00:09:02.980 big guns for that. Right. But if you ask it to 00:09:02.980 --> 00:09:05.279 write a complex Python script or autonomously 00:09:05.279 --> 00:09:07.980 book a multi -city flight, the router sends that 00:09:07.980 --> 00:09:11.139 pass to a heavy -duty, reasoning -focused, expert 00:09:11.139 --> 00:09:14.309 model. That makes sense. And this modular architecture 00:09:14.309 --> 00:09:17.669 is exactly what allowed them to scale up to the 00:09:17.669 --> 00:09:20.990 current GPT 5 .4 that just dropped this month, 00:09:21.389 --> 00:09:23.409 which focuses really heavily on professional 00:09:23.409 --> 00:09:26.129 computer use. But I mean, building a router that 00:09:26.129 --> 00:09:29.330 commands a literal legion of specialized expert 00:09:29.330 --> 00:09:33.269 brains, that comes with a staggering cost. Oh, 00:09:33.350 --> 00:09:35.289 the overhead is unreal. Let's talk about the 00:09:35.289 --> 00:09:39.940 money first. In December 2024, OpenAI introduced 00:09:39.940 --> 00:09:43.700 a pro -tier, $200 a month. $200. And if you're 00:09:43.700 --> 00:09:45.539 sitting there listening, wondering why anyone 00:09:45.539 --> 00:09:48.539 would ever pay $200 a month for a chatbot, you 00:09:48.539 --> 00:09:50.480 have to realize they are no longer selling you 00:09:50.480 --> 00:09:52.840 a chatbot, they're selling you a digital employee. 00:09:53.179 --> 00:09:55.639 Exactly. It's B2B at that point, really. Right. 00:09:55.799 --> 00:09:58.539 And then just this past January in 2026, they 00:09:58.539 --> 00:10:00.759 finally crossed the Rubicon and started testing 00:10:00.759 --> 00:10:03.269 advertisements in the free version. Which everyone 00:10:03.269 --> 00:10:05.690 hated, but the financial reality finally caught 00:10:05.690 --> 00:10:07.889 up with the magic. I mean, OpenAI committed to 00:10:07.889 --> 00:10:11.169 spending $1 .4 trillion on AI infrastructure 00:10:11.169 --> 00:10:14.669 over the next eight years. Wait, trillion? With 00:10:14.669 --> 00:10:18.960 a T. $1 .4 trillion. The compute power required 00:10:18.960 --> 00:10:21.940 to run these agentic test time compute models 00:10:21.940 --> 00:10:25.700 at a global scale is just astronomically expensive. 00:10:25.980 --> 00:10:28.059 Which brings us to the physical reality of that 00:10:28.059 --> 00:10:30.940 cost. And honestly, this part of the source material 00:10:30.940 --> 00:10:33.700 completely blew my mind. We always think of AI 00:10:33.700 --> 00:10:37.399 as this invisible weightless magic, right? Living 00:10:37.399 --> 00:10:39.179 up in the cloud. Like it's just floating around? 00:10:39.539 --> 00:10:42.440 Exactly. Just code floating in the ether. But 00:10:42.440 --> 00:10:44.899 the infrastructure running this... relies on 00:10:44.899 --> 00:10:47.960 massive concrete facilities. They're housing 00:10:47.960 --> 00:10:51.879 things like Microsoft's 30 ,000 NVIDIA GPUs. 00:10:52.159 --> 00:10:54.320 And the environmental toll is shocking. It's 00:10:54.320 --> 00:10:56.679 massive. Researchers from the University of California, 00:10:56.679 --> 00:10:58.759 Riverside published a study showing that a series 00:10:58.759 --> 00:11:01.860 of five to 50 prompts to chat GPT consumes roughly 00:11:01.860 --> 00:11:04.100 half a liter of water. It's incredible. I am 00:11:04.100 --> 00:11:05.860 literally drinking a small water bottle's worth 00:11:05.860 --> 00:11:07.860 of physical resources every time I ask it a few 00:11:07.860 --> 00:11:10.679 dozen questions. Why on earth does the computer 00:11:10.679 --> 00:11:12.980 program need water? Well, if we connected by 00:11:12.980 --> 00:11:14.980 the bigger picture, it really shatters that whole 00:11:14.980 --> 00:11:17.220 illusion of the cloud. Those tens of thousands 00:11:17.220 --> 00:11:20.399 of NVIDIA GPUs are processing trillions of calculations 00:11:20.399 --> 00:11:23.600 a second. So they get hot. incredibly hot. The 00:11:23.600 --> 00:11:25.740 friction generates an immense amount of physical 00:11:25.740 --> 00:11:29.279 heat. So to prevent the data centers from literally 00:11:29.279 --> 00:11:32.259 melting down, they utilize these massive evaporative 00:11:32.259 --> 00:11:35.059 cooling towers. Evaporative cooling. Right. They 00:11:35.059 --> 00:11:37.659 are constantly pumping thousands of gallons of 00:11:37.659 --> 00:11:40.080 fresh water over the heated infrastructure, and 00:11:40.080 --> 00:11:43.080 it evaporates into the atmosphere. Artificial 00:11:43.080 --> 00:11:46.820 intelligence is a heavy resource -intensive industrial 00:11:46.820 --> 00:11:49.259 process. It's basically a factory. It is a factory. 00:11:49.519 --> 00:11:52.080 And when you realize the concrete, the steel, 00:11:52.220 --> 00:11:54.759 the electricity, and the water required for every 00:11:54.759 --> 00:11:57.899 single query, it perfectly explains why they 00:11:57.899 --> 00:12:01.360 had to introduce ads and a $200 subscription 00:12:01.360 --> 00:12:03.899 tier. Because that massive infrastructure needs 00:12:03.899 --> 00:12:07.080 massive, continuous revenue just to survive. 00:12:07.620 --> 00:12:09.840 And to get that revenue, the AI has to be integrated 00:12:09.840 --> 00:12:12.960 into high value, high stakes industries just 00:12:12.960 --> 00:12:15.379 to justify its existence and pay the server bills. 00:12:15.559 --> 00:12:17.340 Which is where things get really messy. Yeah, 00:12:17.440 --> 00:12:20.039 this is exactly where the AI's inherent flaws 00:12:20.039 --> 00:12:23.080 violently collide with messy human reality. Let's 00:12:23.080 --> 00:12:25.320 look at the legal and medical fields. These are 00:12:25.320 --> 00:12:28.159 domains where precision is totally non -negotiable. 00:12:28.360 --> 00:12:31.360 I mean, a single incorrect word in a medical 00:12:31.360 --> 00:12:34.519 diagnosis or a legal brief. can ruin someone's 00:12:34.519 --> 00:12:37.220 life. Absolutely. And yet ChatGPT still wrestles 00:12:37.220 --> 00:12:39.779 with a really well -documented limitation, which 00:12:39.779 --> 00:12:43.299 is hallucination. Or, as the academic papers 00:12:43.299 --> 00:12:45.639 in the source material more accurately describe 00:12:45.639 --> 00:12:48.980 it, confabulation. Yeah, one paper bluntly referred 00:12:48.980 --> 00:12:52.440 to it as a bullshit machine. Yes, which is so 00:12:52.440 --> 00:12:55.600 harsh, but... Kind of accurate. Well, the legal 00:12:55.600 --> 00:12:57.679 examples in the timeline are wild. I mean, we 00:12:57.679 --> 00:13:00.500 have actual cases of attorneys being sanctioned 00:13:00.500 --> 00:13:03.440 in federal court for filing legal motions generated 00:13:03.440 --> 00:13:06.480 by chat GPT that contained entirely fictitious 00:13:06.480 --> 00:13:09.120 legal decisions. Fake cases. The AI just made 00:13:09.120 --> 00:13:11.860 up previous court cases that did not exist. But 00:13:11.860 --> 00:13:14.240 why does it do that? If it has access to the 00:13:14.240 --> 00:13:16.500 internet, why would it invent a fake case instead 00:13:16.500 --> 00:13:19.009 of just searching for a real one? Because At 00:13:19.009 --> 00:13:21.590 its core, it is still just a statistical pattern 00:13:21.590 --> 00:13:23.830 matcher. It doesn't know facts the way a human 00:13:23.830 --> 00:13:26.149 being does. In a legal context, it recognizes 00:13:26.149 --> 00:13:28.590 the semantic pattern of a legal citation. OK, 00:13:28.669 --> 00:13:30.549 what does that mean? It knows that a citation 00:13:30.549 --> 00:13:32.950 usually looks like name versus name, followed 00:13:32.950 --> 00:13:35.210 by a volume number, the reporter abbreviation 00:13:35.210 --> 00:13:38.389 like F bot third, and a page number. That's a 00:13:38.389 --> 00:13:40.570 pattern. Right. So when it can't find a perfect 00:13:40.570 --> 00:13:43.840 case to support your legal argument, The mathematical 00:13:43.840 --> 00:13:46.840 drive to complete that pattern takes over. It 00:13:46.840 --> 00:13:49.659 flawlessly fills in that structural pattern with 00:13:49.659 --> 00:13:53.539 highly plausible but completely fake variables. 00:13:53.879 --> 00:13:56.460 Wow. It's like having a brilliant savant in the 00:13:56.460 --> 00:13:58.799 room who is absolutely terrified of saying, I 00:13:58.799 --> 00:14:00.799 don't know, so they just confidently invent case 00:14:00.799 --> 00:14:02.440 law just to please you. That is exactly what 00:14:02.440 --> 00:14:05.080 it is. But the crazy part is, despite these sanctions, 00:14:05.370 --> 00:14:08.009 We actually have judges in the U .S. and Pakistan 00:14:08.009 --> 00:14:11.509 openly endorsing the use of CHAT -GPT to investigate 00:14:11.509 --> 00:14:14.370 legal questions during active cases. The system 00:14:14.370 --> 00:14:16.789 is simultaneously being punished for lying and 00:14:16.789 --> 00:14:19.190 embraced for its efficiency. And we see the exact 00:14:19.190 --> 00:14:21.669 same paradox in medicine. Oh, yeah. CHAT -GPT 00:14:21.669 --> 00:14:24.190 actually passed the United States medical licensing 00:14:24.190 --> 00:14:27.190 exam. It can pass specialty dermatology exams. 00:14:27.330 --> 00:14:30.350 It is objectively capable of processing highly 00:14:30.350 --> 00:14:33.730 complex medical knowledge. And just this January, 00:14:34.029 --> 00:14:38.350 OpenAI launched They partnered with a data company 00:14:38.350 --> 00:14:41.350 to let the AI directly ingest and discuss your 00:14:41.350 --> 00:14:44.399 actual private health records. But the medical 00:14:44.399 --> 00:14:46.980 community is flashing huge red warning signs 00:14:46.980 --> 00:14:49.399 about this. There are studies mentioned here 00:14:49.399 --> 00:14:51.639 from the Lancet Psychiatry and the journal Digital 00:14:51.639 --> 00:14:54.720 Health specifically warning about people using 00:14:54.720 --> 00:14:58.100 chat GPT as a therapist. Which is becoming incredibly 00:14:58.100 --> 00:15:01.059 common. And the reasons people are doing it are 00:15:01.059 --> 00:15:03.940 so fascinating. The studies note that users are 00:15:03.940 --> 00:15:07.019 attracted to its constant availability and its 00:15:07.019 --> 00:15:09.620 lack of negative reactions. Well, yeah, think 00:15:09.620 --> 00:15:12.019 about it. A human therapist will challenge you. 00:15:12.090 --> 00:15:13.750 They'll set firm boundaries, they will point 00:15:13.750 --> 00:15:15.830 out your flaws, and they might react in ways 00:15:15.830 --> 00:15:17.789 that require you to do the really uncomfortable 00:15:17.789 --> 00:15:20.169 work of personal growth. Therapy is hard. It's 00:15:20.169 --> 00:15:23.370 hard work. The AI, on the other hand, is perpetually 00:15:23.370 --> 00:15:26.470 patient, perpetually available, and fundamentally 00:15:26.470 --> 00:15:28.570 designed to be helpful and subservient to you. 00:15:28.730 --> 00:15:31.049 It just tells you what you want to hear. Exactly, 00:15:31.389 --> 00:15:33.129 and the studies warn that this creates a really 00:15:33.129 --> 00:15:36.230 dangerous dynamic of emotional over -reliance. 00:15:36.509 --> 00:15:38.830 Users are offloading their emotional regulation 00:15:38.830 --> 00:15:41.789 to a machine that provides frictionless empathy, 00:15:42.230 --> 00:15:44.970 but has absolutely no actual understanding of 00:15:44.970 --> 00:15:47.529 human suffering. It even bleeds into our spirituality, 00:15:47.929 --> 00:15:51.230 too. The source notes this June 2023 church service 00:15:51.230 --> 00:15:54.129 in Germany, it was a Protestant convention where 00:15:54.129 --> 00:15:57.250 the entire sermon, like 98 % of the whole service, 00:15:57.870 --> 00:16:01.289 was run by an AI avatar on a large screen. It 00:16:01.289 --> 00:16:03.590 was literally telling a packed room of people 00:16:03.590 --> 00:16:06.559 not to fear death. This raises an important question, 00:16:06.740 --> 00:16:08.860 though. Why are we so willing to do this? Why 00:16:08.860 --> 00:16:11.440 are we so eager to offload our physical health 00:16:11.440 --> 00:16:14.240 diagnoses, our legal defense, our emotional well 00:16:14.240 --> 00:16:17.240 -being, and even our spiritual sermons to a statistical 00:16:17.240 --> 00:16:19.299 text predictor? It's wild when you say it like 00:16:19.299 --> 00:16:21.899 that. There is a profound psychological vulnerability 00:16:21.899 --> 00:16:24.980 here. We are seeking absolute certainty and frictionless 00:16:24.980 --> 00:16:27.320 support in areas of human life that are inherently 00:16:27.320 --> 00:16:30.000 uncertain and full of friction. That is a deeply 00:16:30.000 --> 00:16:32.259 unsettling thought, honestly. We just want the 00:16:32.259 --> 00:16:34.600 easy answer, even if the entity giving it to 00:16:34.600 --> 00:16:37.240 us is really just doing math. But as chat GPT 00:16:37.240 --> 00:16:39.559 attempts to navigate all these high stakes human 00:16:39.559 --> 00:16:42.659 systems, we are forced to confront the illusion 00:16:42.659 --> 00:16:46.240 of the perfect machine. Here's where it gets 00:16:46.240 --> 00:16:50.080 really interesting. To make the AI seem so polished 00:16:50.080 --> 00:16:53.120 and so helpful, it actually relies on enormous 00:16:53.120 --> 00:16:55.779 amount of grueling, hidden human labor. This 00:16:55.779 --> 00:16:57.100 is the part people don't want to talk about. 00:16:57.529 --> 00:17:00.629 Building the safety filters. The guardrails that 00:17:00.629 --> 00:17:03.169 prevent chat GPT from generating horrific, illegal, 00:17:03.289 --> 00:17:05.990 or toxic content that requires a process called 00:17:05.990 --> 00:17:08.829 reinforcement learning from human feedback, or 00:17:08.829 --> 00:17:11.750 RLHF. And the human part of that equation takes 00:17:11.750 --> 00:17:14.690 a massive toll. A huge toll. The source outlines 00:17:14.690 --> 00:17:17.529 this perfectly. OpenAI use outsourced workers 00:17:17.529 --> 00:17:20.259 in Kenya through a company called Sama. These 00:17:20.259 --> 00:17:22.940 people were earning between $1 .32 and $2 an 00:17:22.940 --> 00:17:25.420 hour. Just pennies, really. Yeah. And their entire 00:17:25.420 --> 00:17:28.859 job was to read, label, and categorize toxic, 00:17:29.079 --> 00:17:31.680 traumatic, and deeply disturbing content so the 00:17:31.680 --> 00:17:33.460 AI could mathematically learn what to filter 00:17:33.460 --> 00:17:35.700 out. One worker literally described the assignment 00:17:35.700 --> 00:17:38.380 as torture. Because it is the digital equivalent 00:17:38.380 --> 00:17:40.940 of toxic waste cleanup. What? We get to sit here 00:17:40.940 --> 00:17:44.680 and enjoy a clean, polite, helpful chatbot interface. 00:17:45.230 --> 00:17:48.329 only because thousands of low -wage workers absorbed 00:17:48.329 --> 00:17:50.650 the trauma of the internet's darkest corners 00:17:50.650 --> 00:17:54.250 to build those invisible guardrails for us. We 00:17:54.250 --> 00:17:56.730 look at the AI as this perfectly neutral oracle. 00:17:57.670 --> 00:18:00.269 But underneath it, there are real people earning 00:18:00.269 --> 00:18:02.769 $2 an hour traumatizing themselves to make it 00:18:02.769 --> 00:18:05.569 safe. The AI isn't neutral because the human 00:18:05.569 --> 00:18:07.680 world building it isn't neutral. Which brings 00:18:07.680 --> 00:18:10.359 us right to the intense political polarization 00:18:10.359 --> 00:18:12.839 surrounding the tool. Yes. And before we dive 00:18:12.839 --> 00:18:14.799 into this part, we need to be incredibly clear 00:18:14.799 --> 00:18:17.200 with you, the listener. We are not taking sides 00:18:17.200 --> 00:18:19.680 here. Not at all. We are strictly reporting on 00:18:19.680 --> 00:18:21.680 the information provided in the source material 00:18:21.680 --> 00:18:24.700 just to show how this technology has become a 00:18:24.700 --> 00:18:26.900 massive battleground in our modern culture wars. 00:18:27.160 --> 00:18:29.799 We are absolutely not endorsing any political 00:18:29.799 --> 00:18:32.019 viewpoint. Right. We are simply looking at the 00:18:32.019 --> 00:18:35.039 AI as a focal point for societal tension. So, 00:18:35.160 --> 00:18:37.500 on one side of the spectrum, you have multiple 00:18:37.500 --> 00:18:40.000 academic studies, including a major one published 00:18:40.000 --> 00:18:43.119 in the journal Public Choice, that found a systemic 00:18:43.119 --> 00:18:46.500 political bias in CHAT GPT's outputs toward left 00:18:46.500 --> 00:18:49.039 -leaning perspectives. Right, specifically favoring 00:18:49.039 --> 00:18:51.019 the Democrats in the US or the Labor Party in 00:18:51.019 --> 00:18:53.539 the UK. Exactly. And conservative commentators 00:18:53.539 --> 00:18:56.200 have heavily criticized the AI for this, arguing 00:18:56.200 --> 00:18:58.160 that the safety filters and the training data 00:18:58.160 --> 00:19:01.079 have baked a very specific ideological worldview. 00:19:01.519 --> 00:19:04.259 right into the system. And then on the exact 00:19:04.259 --> 00:19:07.039 other side of the spectrum, the source details 00:19:07.039 --> 00:19:10.079 a massive user backlash that just happened in 00:19:10.079 --> 00:19:13.660 February 2026. A movement started on Reddit called 00:19:13.660 --> 00:19:16.700 Quit GPT, where hundreds of thousands of users 00:19:16.700 --> 00:19:19.420 organized a massive boycott and canceled their 00:19:19.420 --> 00:19:22.339 $200 subscriptions. All over a political donation. 00:19:22.519 --> 00:19:26.460 Yes. This was entirely over a $25 million donation 00:19:26.460 --> 00:19:29.720 made by OpenAI's president, Greg Brockman, to 00:19:29.720 --> 00:19:32.079 a Trump super PAC. So you have the algorithmic 00:19:32.079 --> 00:19:34.380 outputs leaning one way, the executive money 00:19:34.380 --> 00:19:37.079 leaning another, and users on all sides are just 00:19:37.079 --> 00:19:39.279 furious that the tool is entangled in politics 00:19:39.279 --> 00:19:42.099 at all. It just highlights that you cannot separate 00:19:42.099 --> 00:19:45.180 a trillion dollar technology from the socio -political 00:19:45.180 --> 00:19:47.019 environment it exists within. It's impossible. 00:19:47.680 --> 00:19:49.940 And this tension extends to how the company manages 00:19:49.940 --> 00:19:52.400 fundamental issues of trust and security too. 00:19:52.500 --> 00:19:55.779 Oh, like the watermark thing. Yes. The timeline 00:19:55.779 --> 00:19:58.559 notes that OpenAI actually successfully developed 00:19:58.559 --> 00:20:01.940 a watermarking tool to detect AI -generated text, 00:20:02.119 --> 00:20:04.799 but they completely refused to release it publicly. 00:20:04.940 --> 00:20:07.960 But why? I mean, cheating in schools, academic 00:20:07.960 --> 00:20:10.140 fraud, and political misinformation are such 00:20:10.140 --> 00:20:13.000 huge problems for society right now. Why would 00:20:13.000 --> 00:20:15.339 they sit on the technological solution? Market 00:20:15.339 --> 00:20:18.500 competition. Pure and simple. They explicitly 00:20:18.500 --> 00:20:21.180 stated they feared that if they applied the watermark, 00:20:21.720 --> 00:20:24.079 users who wanted to pass off AI text as their 00:20:24.079 --> 00:20:26.920 own human work would just cancel their subscriptions. 00:20:27.039 --> 00:20:29.079 They'd go somewhere else. Right. They'd flee 00:20:29.079 --> 00:20:31.200 to competitor models that didn't use watermarks. 00:20:31.779 --> 00:20:33.940 So the business survival keeping that user growth 00:20:33.940 --> 00:20:36.380 high took precedence over societal transparency. 00:20:36.619 --> 00:20:39.779 Wow. And the stakes of user safety are literally 00:20:39.779 --> 00:20:42.759 life and death. The timeline notes that in September 00:20:42.759 --> 00:20:46.180 2025, following the tragic suicide of a 16 -year 00:20:46.180 --> 00:20:49.380 -old user, OpenAI had to rapidly announce plans 00:20:49.380 --> 00:20:52.200 to add restrictions for users under 18. Blocking 00:20:52.200 --> 00:20:54.319 certain content and flotacious interactions. 00:20:54.559 --> 00:20:57.119 Yeah. It shows just how reactive these safety 00:20:57.119 --> 00:20:58.960 measures are when the product is scaling this 00:20:58.960 --> 00:21:00.380 fast. They're building the plane while they're 00:21:00.380 --> 00:21:03.099 flying it. If we synthesize all of this, the 00:21:03.099 --> 00:21:05.539 hidden labor, the political boycotts, the delayed 00:21:05.539 --> 00:21:08.319 safety features, what really emerges is that 00:21:08.319 --> 00:21:10.680 artificial intelligence is ultimately a mirror. 00:21:10.680 --> 00:21:13.180 A mirror. Yes. It doesn't just reflect our data. 00:21:13.259 --> 00:21:16.220 It reflects our geopolitical realities. The source 00:21:16.220 --> 00:21:19.740 points out that OpenAI actively had to ban state 00:21:19.740 --> 00:21:22.839 -backed influence operations from China, Russia, 00:21:22.880 --> 00:21:25.880 and Israel that were using chat GPT to generate 00:21:25.880 --> 00:21:29.039 international propaganda. Wow. The AI reflects 00:21:29.039 --> 00:21:32.440 our global influence. It reflects our labor inequalities 00:21:32.440 --> 00:21:35.720 and it reflects our own deeply entrenched societal 00:21:35.720 --> 00:21:38.190 divisions. What does this all mean? We started 00:21:38.190 --> 00:21:40.789 this deep dive looking at a quirky intern that 00:21:40.789 --> 00:21:44.069 wrote bad poetry back in 2022. And we traced 00:21:44.069 --> 00:21:46.670 its evolution into this autonomous agent that 00:21:46.670 --> 00:21:49.250 parses web code, spends your money, and operates 00:21:49.250 --> 00:21:51.529 on massive water -cooled supercomputers. Doing 00:21:51.529 --> 00:21:54.049 things we couldn't have imagined. Exactly. It's 00:21:54.049 --> 00:21:56.150 passing complex medical exams, it's hallucinating 00:21:56.150 --> 00:21:58.809 federal legal cases, and it's sitting dead center 00:21:58.809 --> 00:22:00.869 in the middle of our modern political and ethical 00:22:00.869 --> 00:22:04.509 firestorms. It has been a relentless integration 00:22:04.509 --> 00:22:08.309 into the human experience, but There is one final 00:22:08.309 --> 00:22:10.789 lingering thought I want to leave you with, something 00:22:10.789 --> 00:22:13.309 that wasn't explicitly spelled out in the timeline, 00:22:13.670 --> 00:22:16.230 but sits right beneath the surface of all these 00:22:16.230 --> 00:22:18.269 developments. OK. We'll be honest. We've spent 00:22:18.269 --> 00:22:20.990 this entire time talking about how ChatGPT is 00:22:20.990 --> 00:22:24.829 learning to mimic and perform human tasks, right? 00:22:24.829 --> 00:22:27.329 Right. It's learning to write code, to buy groceries, 00:22:27.630 --> 00:22:30.109 to diagnose illnesses, and to provide comfort. 00:22:30.789 --> 00:22:33.730 But as we increasingly rely on ChatGPT Health 00:22:33.730 --> 00:22:36.509 for understanding our own bodies, as lean on 00:22:36.509 --> 00:22:38.670 the agentic commerce protocol for our shopping, 00:22:39.009 --> 00:22:41.769 and as we look to AI avatars for our spiritual 00:22:41.769 --> 00:22:44.309 and emotional comfort, are we the ones actually 00:22:44.309 --> 00:22:47.269 being programmed? Whoa. Think about it. If a 00:22:47.269 --> 00:22:49.250 machine is constantly anticipating your needs, 00:22:49.589 --> 00:22:51.150 smoothing out the friction of your daily life, 00:22:51.190 --> 00:22:53.170 and making the frictionless choices for you, 00:22:53.410 --> 00:22:56.029 what happens to the uniquely human muscle of 00:22:56.029 --> 00:22:58.710 decision making? It just doesn't get used. Exactly. 00:22:58.970 --> 00:23:01.349 If you no longer have to struggle to find an 00:23:01.349 --> 00:23:04.170 answer, or wrestle with a complex user interface, 00:23:04.490 --> 00:23:06.680 or just sit with the uncomfortable silence of 00:23:06.680 --> 00:23:09.559 a human therapist, does that human capacity begin 00:23:09.559 --> 00:23:12.140 to atrophy? If the machine does all the thinking, 00:23:12.680 --> 00:23:15.339 what exactly is left for us to do? That brings 00:23:15.339 --> 00:23:18.920 us right back to where we started. In 2022, we 00:23:18.920 --> 00:23:21.220 were looking at a novelty. Today, we're looking 00:23:21.220 --> 00:23:23.460 at a system that is fundamentally rewiring how 00:23:23.460 --> 00:23:25.380 we interact with the world and potentially how 00:23:25.380 --> 00:23:27.900 we interact with our own minds. The world is 00:23:27.900 --> 00:23:30.359 changing incredibly fast, and just keeping up 00:23:30.359 --> 00:23:32.779 is half the battle. Thank you so much for joining 00:23:32.779 --> 00:23:35.140 us on this deep drive. Stay insanely curious, 00:23:35.420 --> 00:23:37.059 keep asking the hard questions, and we'll catch 00:23:37.059 --> 00:23:37.700 you on the next one.