WEBVTT 00:00:00.000 --> 00:00:02.080 It creates this fascinating tension, doesn't 00:00:02.080 --> 00:00:05.700 it? On one side, you have this massive grassroots 00:00:05.700 --> 00:00:10.220 movement triggered by a $25 million donation. 00:00:10.880 --> 00:00:14.400 People are actually canceling subscriptions in 00:00:14.400 --> 00:00:17.670 protest. trying to stop the machine. And at the 00:00:17.670 --> 00:00:20.089 exact same time. At literally the same time, 00:00:20.149 --> 00:00:22.210 on the other side of the country, you have a 00:00:22.210 --> 00:00:25.329 $10 billion construction project breaking ground. 00:00:25.489 --> 00:00:27.530 Right. It's the ultimate contrast. You have the 00:00:27.530 --> 00:00:29.949 human layer trying to pull the emergency brake, 00:00:30.070 --> 00:00:32.630 and then you have the industrial layer hitting 00:00:32.630 --> 00:00:35.630 the gas harder than we have ever seen. Hello 00:00:35.630 --> 00:00:37.649 and welcome back to the Deep Dive. I've been 00:00:37.649 --> 00:00:39.850 sitting with this stack of sources you sent over 00:00:39.850 --> 00:00:42.590 for a few hours now, and I have to be honest, 00:00:42.750 --> 00:00:45.630 the signal -to -noise ratio in the AI space this 00:00:45.630 --> 00:00:48.770 week has been overwhelming. It feels that way. 00:00:48.950 --> 00:00:50.590 It feels like the industry is splitting in two 00:00:50.590 --> 00:00:52.909 directions at once. We have a cultural reckoning 00:00:52.909 --> 00:00:55.329 happening in the software and a physical explosion 00:00:55.329 --> 00:00:57.950 happening in the hardware. It really is an inflection 00:00:57.950 --> 00:01:00.329 point. I mean, usually when we sit down, we're 00:01:00.329 --> 00:01:02.409 obsessing over the models. Is it smarter? Is 00:01:02.409 --> 00:01:06.049 it faster? But today's sources paint a much, 00:01:06.090 --> 00:01:09.069 much messier picture. To really get what's happening 00:01:09.069 --> 00:01:11.349 this week, we have to look at three specific 00:01:11.349 --> 00:01:13.799 pillars. Okay, walk us through the roadmap. First, 00:01:13.879 --> 00:01:15.239 we have to talk about the cultural backlash. 00:01:15.560 --> 00:01:18.299 We're seeing the rise of this quit GPT movement. 00:01:18.620 --> 00:01:21.859 This is activism finally colliding with algorithms. 00:01:22.040 --> 00:01:25.159 Second, the infrastructure explosion. We aren't 00:01:25.159 --> 00:01:27.459 talking about code here. We're talking concrete, 00:01:27.680 --> 00:01:30.959 chips, and power grids, the physical reality 00:01:30.959 --> 00:01:34.200 of the cloud. And third, despite all that noise, 00:01:34.299 --> 00:01:37.319 the capability leap is still happening. We have 00:01:37.319 --> 00:01:40.599 new tools from Alibaba and OpenAI that fundamentally 00:01:40.599 --> 00:01:43.140 change how we actually work. Let's start with 00:01:43.140 --> 00:01:45.859 the friction, the cultural backlash. For a long 00:01:45.859 --> 00:01:48.140 time, the public sentiment around AI was just 00:01:48.140 --> 00:01:51.519 pure awe. Look what this thing can do. But reading 00:01:51.519 --> 00:01:54.000 these notes on QuitGPT, this feels different. 00:01:54.099 --> 00:01:55.840 This isn't just people being afraid of the Terminator. 00:01:55.859 --> 00:01:58.739 This is organized economic pressure. What sparked 00:01:58.739 --> 00:02:01.480 this? It is different because it's political. 00:02:02.250 --> 00:02:04.790 The catalyst here wasn't a software bug or a 00:02:04.790 --> 00:02:08.490 bad answer. It was a $25 million donation to 00:02:08.490 --> 00:02:11.449 Magia Inc. by Greg Brockman, the president of 00:02:11.449 --> 00:02:13.969 OpenAI, and his wife. Okay, let's pause on that. 00:02:14.009 --> 00:02:16.330 So we have a direct financial link from the leadership 00:02:16.330 --> 00:02:21.930 of the leading AI company to a very polarizing 00:02:21.930 --> 00:02:24.750 political campaign. Exactly. And that donation 00:02:24.750 --> 00:02:27.750 basically lit the match. But if you look at the 00:02:27.750 --> 00:02:29.610 source material, it's not just that one check. 00:02:29.689 --> 00:02:32.349 It's a compounding effect. You have the donation, 00:02:32.430 --> 00:02:35.330 but then you also have this new partnership between. 00:02:36.090 --> 00:02:39.310 OpenAI and IC immigration and customs enforcement. 00:02:39.889 --> 00:02:43.629 They're using ChatGPT4 for hiring and for administrative 00:02:43.629 --> 00:02:47.009 processing. Which for a significant portion of 00:02:47.009 --> 00:02:49.349 the tech workforce and that early adopter user 00:02:49.349 --> 00:02:52.750 base represents a massive red line. A huge red 00:02:52.750 --> 00:02:54.610 line. Remember, a lot of the people who pay for 00:02:54.610 --> 00:02:57.810 ChatGPT Plus are in that creative tech adjacent 00:02:57.810 --> 00:03:00.810 demographic. They feel a sense of ownership over 00:03:00.810 --> 00:03:03.669 the tool. So when they see the leadership funding 00:03:03.669 --> 00:03:07.159 one political side. and then the technology being 00:03:07.159 --> 00:03:09.719 deployed by ICE, it just feels like a betrayal 00:03:09.719 --> 00:03:12.680 of that original open and beneficial mission. 00:03:13.419 --> 00:03:15.699 And Scott Galloway is involved now. I saw his 00:03:15.699 --> 00:03:18.199 name in the notes. Yeah, the NYU professor. He's 00:03:18.199 --> 00:03:20.699 very good at capturing the zeitgeist of these 00:03:20.699 --> 00:03:24.039 things. He went viral urging people to resist 00:03:24.039 --> 00:03:27.199 and unsubscribe. Right. He's framing it not just 00:03:27.199 --> 00:03:30.080 as a boycott, but as a resistance against big 00:03:30.080 --> 00:03:33.039 tech tools fueling specific political agendas. 00:03:33.400 --> 00:03:35.900 And it's resonating. The sources say there are 00:03:35.900 --> 00:03:39.479 about 17 ,000 pledges on the QuitGPT site so 00:03:39.479 --> 00:03:41.620 far. Okay, 17 ,000. Let's look at the numbers. 00:03:41.699 --> 00:03:43.400 On a human level, that's... That's a stadium 00:03:43.400 --> 00:03:45.560 full of angry people. That's significant. It 00:03:45.560 --> 00:03:49.139 is. But ChatGPT has something like 900 million 00:03:49.139 --> 00:03:53.199 weekly users. Does 17 ,000 cancellations actually 00:03:53.199 --> 00:03:56.360 matter to a company like OpenAI? Or is this just 00:03:56.360 --> 00:03:59.159 screaming into the void? Financially, no. The 00:03:59.159 --> 00:04:01.379 math is brutal. Only about 5 or 6 % of those 00:04:01.379 --> 00:04:03.340 900 million users are even paid subscribers. 00:04:03.699 --> 00:04:06.900 So even if they lost 100 ,000 subscribers, the 00:04:06.900 --> 00:04:09.879 revenue hit is just a rounding error. It won't 00:04:09.879 --> 00:04:12.340 stop the lights from coming on. OpenAI is burning 00:04:12.340 --> 00:04:15.479 cash on compute, not relying on $20 subscriptions 00:04:15.479 --> 00:04:18.180 to survive. So if it doesn't hurt the wallet, 00:04:18.279 --> 00:04:21.779 is it just performative? I wouldn't call it performative. 00:04:21.779 --> 00:04:24.389 I'd call it a signal. A cultural signal. OK. 00:04:24.490 --> 00:04:26.350 You have to remember, OpenAI isn't just fighting 00:04:26.350 --> 00:04:28.910 for customers. They are fighting for talent. 00:04:29.670 --> 00:04:32.509 The engineers who build these models are often 00:04:32.509 --> 00:04:34.910 very politically active, very ethics focused. 00:04:35.310 --> 00:04:38.660 If the brand becomes toxic. If working for OpenAI 00:04:38.660 --> 00:04:41.259 becomes synonymous with funding the bad guys 00:04:41.259 --> 00:04:44.180 in the eyes of their peers, that hurts them way 00:04:44.180 --> 00:04:47.079 more than losing some subscription revenue. That's 00:04:47.079 --> 00:04:49.019 a great point. The talent war is the real war. 00:04:49.180 --> 00:04:51.879 Exactly. If the smartest 25 year old researcher 00:04:51.879 --> 00:04:54.180 refuses to sign the offer letter because of the 00:04:54.180 --> 00:04:57.339 brand reputation, that's the actual cost. Precisely. 00:04:57.339 --> 00:04:59.060 And it's not just consumers versus companies 00:04:59.060 --> 00:05:01.180 anymore. The sources highlight something fascinating. 00:05:02.329 --> 00:05:04.050 The companies are starting to fight each other 00:05:04.050 --> 00:05:07.290 politically, too. It's a full -on war of the 00:05:07.290 --> 00:05:10.610 pigs. I saw that bit about Anthropic. They frame 00:05:10.610 --> 00:05:13.370 themselves as the safety -first, constitutional 00:05:13.370 --> 00:05:16.529 AI alternative. What are they doing? They just 00:05:16.529 --> 00:05:19.769 funded a $20 million super piece. $80 million? 00:05:20.189 --> 00:05:23.560 Specifically to push for AI regulation. That 00:05:23.560 --> 00:05:25.459 is not a small amount of change for a policy 00:05:25.459 --> 00:05:28.399 push. No. And it's directly opposing the political 00:05:28.399 --> 00:05:31.100 interests backed by open eyes. Fifty million 00:05:31.100 --> 00:05:35.399 dollar push. So you have these two. massive entities 00:05:35.399 --> 00:05:38.560 who are competitors in code now becoming competitors 00:05:38.560 --> 00:05:41.120 in the political lobby. They're literally spending 00:05:41.120 --> 00:05:43.939 tens of millions of dollars to shape the laws 00:05:43.939 --> 00:05:46.360 that will govern the other. It really highlights 00:05:46.360 --> 00:05:48.879 that AI isn't just a tool anymore. It's a political 00:05:48.879 --> 00:05:51.259 battlefield. You can't just build it and they 00:05:51.259 --> 00:05:52.920 will come. You have to build it, you have to 00:05:52.920 --> 00:05:54.699 lobby for it, and you have to defend it from 00:05:54.699 --> 00:05:57.360 your own user base. That's the reality of 2026. 00:05:57.699 --> 00:06:00.180 So just to cap this segment off, does this consumer 00:06:00.180 --> 00:06:02.959 pressure actually work in an age of enterprise 00:06:02.959 --> 00:06:05.600 dominance? It creates a morale crisis, which 00:06:05.600 --> 00:06:07.680 creates a tenant crisis, which eventually creates 00:06:07.680 --> 00:06:10.720 a product crisis. So, yes, it matters. Interesting. 00:06:10.759 --> 00:06:13.279 So while the culture is fighting over donations 00:06:13.279 --> 00:06:16.220 and hiring policies and the lobbyists are fighting 00:06:16.220 --> 00:06:19.720 over regulation, the actual physical machine 00:06:19.720 --> 00:06:22.819 is just getting bigger. Oh, yeah. I want to pivot 00:06:22.819 --> 00:06:25.379 to the infrastructure piece because the numbers 00:06:25.379 --> 00:06:29.360 here are staggering. The hardware reality. This 00:06:29.360 --> 00:06:31.439 is the stuff you can kick. I was looking at this 00:06:31.439 --> 00:06:35.790 note about meta in Indiana. $10 billion for a 00:06:35.790 --> 00:06:37.829 data center. It's monumental. Meta just broke 00:06:37.829 --> 00:06:40.649 ground on this. And data center feels like too 00:06:40.649 --> 00:06:43.910 small a word. Right. It's a campus. 13 buildings, 00:06:44.290 --> 00:06:47.629 4 ,000 construction jobs. This is a physical 00:06:47.629 --> 00:06:50.589 manifestation of their next -gen ambitions. When 00:06:50.589 --> 00:06:52.370 Zuckerberg says they are building the future, 00:06:52.550 --> 00:06:54.889 they are literally pouring the concrete for it 00:06:54.889 --> 00:06:57.470 in Indiana. It changes how you visualize the 00:06:57.470 --> 00:06:59.290 cloud, doesn't it? We tend to think of the cloud 00:06:59.290 --> 00:07:01.410 as this ethereal thing floating above us. Right. 00:07:01.759 --> 00:07:04.920 But it's actually 13 massive buildings in Indiana 00:07:04.920 --> 00:07:07.240 sucking up electricity. And that electricity 00:07:07.240 --> 00:07:09.879 part is becoming the friction point. The power 00:07:09.879 --> 00:07:11.980 demands are so high that these companies are 00:07:11.980 --> 00:07:14.399 having to play good neighbor in a very, very 00:07:14.399 --> 00:07:18.180 expensive way. I saw that in the notes. Anthropic, 00:07:18.379 --> 00:07:22.879 Microsoft, and OpenAI. are all promising to pay 00:07:22.879 --> 00:07:25.379 the extra power bills. They have to. Imagine 00:07:25.379 --> 00:07:27.819 you live in a town, a data center moves in, and 00:07:27.819 --> 00:07:30.120 suddenly your residential electricity rate goes 00:07:30.120 --> 00:07:33.459 up 20 % because demand spiked. That's a political 00:07:33.459 --> 00:07:36.060 nightmare. It is. So they are preemptively saying, 00:07:36.279 --> 00:07:38.620 we will cover the difference. It shows you how 00:07:38.620 --> 00:07:40.399 desperate they are to keep these things running 00:07:40.399 --> 00:07:42.980 without local backlash. They're essentially paying 00:07:42.980 --> 00:07:45.399 rent on the grid itself. It's almost like a pollution 00:07:45.399 --> 00:07:48.439 tax, but for electron consumption. Exactly. And 00:07:48.439 --> 00:07:50.879 speaking of desperate need for speed, we have 00:07:50.879 --> 00:07:53.920 to talk about this OpenAI and Cerebras deal. 00:07:54.120 --> 00:07:56.699 This connects the hardware to the actual user 00:07:56.699 --> 00:07:59.800 experience. This is Codex Spark. Right. OpenAI 00:07:59.800 --> 00:08:02.399 launched Codex Spark. It's a lighter version 00:08:02.399 --> 00:08:05.800 of GPT -5 .3 designed specifically for real -time 00:08:05.800 --> 00:08:08.560 coding. But the moment of wonder here isn't the 00:08:08.560 --> 00:08:11.459 software. It's the chip. It's the chip powering 00:08:11.459 --> 00:08:14.899 it. The Cerebras chip. I saw the Spex 4T transistors. 00:08:14.980 --> 00:08:18.040 Four trillion. Can we contextualize that? Because 00:08:18.040 --> 00:08:21.019 trillion is one of those numbers that human brains 00:08:21.019 --> 00:08:24.220 aren't good at visualizing. What does a four 00:08:24.220 --> 00:08:26.699 trillion transistor chip actually allow you to 00:08:26.699 --> 00:08:30.029 do that a normal chip doesn't? Speed. Specifically, 00:08:30.189 --> 00:08:32.490 inference speed. So training is teaching the 00:08:32.490 --> 00:08:35.009 AI. That takes months. Inference is when you 00:08:35.009 --> 00:08:37.289 ask it a question and it answers. Right. That's 00:08:37.289 --> 00:08:40.330 the daily usage. Usually there's a lag. You type, 00:08:40.350 --> 00:08:43.029 the bubbles spin, the text streams out. With 00:08:43.029 --> 00:08:45.769 a chip this size, the memory bandwidth is so 00:08:45.769 --> 00:08:48.649 high that the latency basically vanishes. So 00:08:48.649 --> 00:08:51.529 it moves from email correspondence to instant 00:08:51.529 --> 00:08:54.710 messaging. It moves to thought speed. If you're 00:08:54.710 --> 00:08:58.029 a coder using Codex Spark. The AI isn't pausing 00:08:58.029 --> 00:09:01.250 to think. It's typing with you. It predicts your 00:09:01.250 --> 00:09:03.690 next logic block before you even finish the syntax. 00:09:04.070 --> 00:09:06.429 Wow. It changes the psychological relationship 00:09:06.429 --> 00:09:08.730 with the tool. It feels less like a tool and 00:09:08.730 --> 00:09:10.830 more like an extension of your own fingers. That's 00:09:10.830 --> 00:09:12.549 incredible. And this ties into that valuation 00:09:12.549 --> 00:09:15.009 for Modal Labs, right? Exactly. Modal Labs is 00:09:15.009 --> 00:09:18.389 raising at a $2 .5 billion valuation. They focus 00:09:18.389 --> 00:09:21.330 entirely on inference infrastructure, basically 00:09:21.330 --> 00:09:23.450 the plumbing that allows these massive models 00:09:23.450 --> 00:09:26.600 to run quickly and cheaply. The market is betting 00:09:26.600 --> 00:09:28.919 heavily that this is the year we stop just training 00:09:28.919 --> 00:09:30.940 models and start running them at massive scale. 00:09:31.120 --> 00:09:34.679 So are we underestimating the physical footprint, 00:09:34.980 --> 00:09:37.220 the concrete and electricity required for digital 00:09:37.220 --> 00:09:39.919 intelligence? Absolutely. Software is now driving 00:09:39.919 --> 00:09:42.659 the largest industrial construction boom in recent 00:09:42.659 --> 00:09:44.620 history. Which brings us to the tools themselves. 00:09:44.879 --> 00:09:47.139 We built the data centers. We fought the political 00:09:47.139 --> 00:09:50.090 battles. Yeah. What can we actually do with this 00:09:50.090 --> 00:09:52.490 stuff today that we couldn't do yesterday? This 00:09:52.490 --> 00:09:55.490 is the fun part. The capabilities are shifting 00:09:55.490 --> 00:09:59.789 from just chatting to doing. And the big story 00:09:59.789 --> 00:10:03.169 here is Alibaba. Quinn Image 2 .0. Catchy name, 00:10:03.190 --> 00:10:05.649 right? Rolls right off the tongue. The note here 00:10:05.649 --> 00:10:07.870 says it's a PowerPoint native model. What does 00:10:07.870 --> 00:10:09.909 that mean exactly? Okay, think about how most 00:10:09.909 --> 00:10:12.389 image generators work right now. Mid -journey, 00:10:12.389 --> 00:10:15.210 daily. You ask for a picture, you get a picture. 00:10:15.330 --> 00:10:18.799 It's art. But if you try to put text on it, it 00:10:18.799 --> 00:10:20.419 looks like alien hieroglyphics. If you try to 00:10:20.419 --> 00:10:22.720 make it a specific resolution for a slide deck, 00:10:22.879 --> 00:10:25.659 you have to crop it. Right. It's creative. But 00:10:25.659 --> 00:10:28.320 it's not productive. Exactly. Quen changes that. 00:10:28.419 --> 00:10:31.980 It merges generation and editing in one shot. 00:10:32.220 --> 00:10:36.000 It handles 2048 by 2048 resolution. So no more 00:10:36.000 --> 00:10:39.320 stitching tiles together. And crucially, it has 00:10:39.320 --> 00:10:41.559 high text fidelity. So you can actually read 00:10:41.559 --> 00:10:43.960 it. You can tell it. Make me a slide about Q3 00:10:43.960 --> 00:10:46.419 earnings with a bar graph and the title growth 00:10:46.419 --> 00:10:49.860 vector in Arial font. And it just does it. So 00:10:49.860 --> 00:10:52.580 it's an AI competitor to Canva. not just an AI 00:10:52.580 --> 00:10:55.039 computer to an artist. That's the promise. It's 00:10:55.039 --> 00:10:57.500 moving the workflow. You aren't just generating 00:10:57.500 --> 00:10:59.399 assets that you then have to take into Photoshop 00:10:59.399 --> 00:11:02.720 to fix. You are producing a final asset right 00:11:02.720 --> 00:11:04.659 in the chat. That is a significant shift for 00:11:04.659 --> 00:11:06.720 enterprise users. I know I struggle with that, 00:11:06.799 --> 00:11:08.779 constantly getting the image right, but having 00:11:08.779 --> 00:11:12.019 the text look ridiculous. Right. And while Alibaba 00:11:12.019 --> 00:11:14.779 is fixing images, Anthropic is coming for your 00:11:14.779 --> 00:11:17.860 Office suite. They just unlocked file tools for 00:11:17.860 --> 00:11:21.179 free users. Excel, PowerPoint, Word, no sh**. 00:11:21.129 --> 00:11:23.409 Previously, you had to pay to have Claude analyze 00:11:23.409 --> 00:11:25.909 these. Now it's open to everyone. That is a big 00:11:25.909 --> 00:11:28.629 democratization move. Suddenly, having an AI 00:11:28.629 --> 00:11:30.570 analyst go through your spreadsheets isn't a 00:11:30.570 --> 00:11:32.610 premium feature. It's a baseline expectation. 00:11:32.830 --> 00:11:35.750 Right. Exactly. But, and there is always a but 00:11:35.750 --> 00:11:38.230 in this industry, with these new capabilities 00:11:38.230 --> 00:11:41.370 comes a darker side. We have to talk about what 00:11:41.370 --> 00:11:43.750 happened to Google's Gemini. The stealth attack. 00:11:44.460 --> 00:11:46.779 I read this twice because I wasn't sure I understood 00:11:46.779 --> 00:11:49.120 it. This wasn't a hack in the traditional sense, 00:11:49.299 --> 00:11:51.159 right? They didn't break a password. No, they 00:11:51.159 --> 00:11:52.879 didn't break in. They invited themselves in. 00:11:53.019 --> 00:11:55.779 It's called a model copying campaign or model 00:11:55.779 --> 00:11:57.799 distillation. Walk us through how that works. 00:11:58.240 --> 00:12:01.080 So Google's Gemini is a trillion dollar brain. 00:12:01.549 --> 00:12:05.250 It's huge, expensive, and smart. A competitor 00:12:05.250 --> 00:12:08.450 or a bad actor wants that intelligence but doesn't 00:12:08.450 --> 00:12:11.129 have the money to train it. Okay. So they hit 00:12:11.129 --> 00:12:14.669 Gemini with over 100 ,000 very specific complex 00:12:14.669 --> 00:12:18.509 prompts. Explain quantum physics. Write a legal 00:12:18.509 --> 00:12:21.480 brief. Debug this code. And they just record 00:12:21.480 --> 00:12:24.139 all the answers? Record the answers. Then they 00:12:24.139 --> 00:12:26.299 take those high -quality answers and feed them 00:12:26.299 --> 00:12:29.139 into their own much smaller, cheaper model. They 00:12:29.139 --> 00:12:31.399 are teaching their cheap students using Google's 00:12:31.399 --> 00:12:33.980 expensive textbook. Wow. Effectively, they are 00:12:33.980 --> 00:12:36.059 siphoning off the intelligence of the master 00:12:36.059 --> 00:12:39.799 model to create a clone. That sounds incredibly 00:12:39.799 --> 00:12:43.159 difficult to police. It is. Google is calling 00:12:43.159 --> 00:12:45.860 it theft. They are warning startups that if it 00:12:45.860 --> 00:12:47.639 can happen to Gemini, it can happen to anyone. 00:12:47.879 --> 00:12:50.440 But the legal framework here is nonexistent. 00:12:50.639 --> 00:12:54.000 Is asking a question theft? Is learning from 00:12:54.000 --> 00:12:57.279 an answer copyright infringement? It raises this 00:12:57.279 --> 00:13:00.059 massive question about what intellectual property 00:13:00.059 --> 00:13:02.299 even looks like when you can steal it just by 00:13:02.299 --> 00:13:05.419 asking questions. If I can distill your secret 00:13:05.419 --> 00:13:08.799 sauce just by talking to your bot, do you even 00:13:08.799 --> 00:13:11.000 have a moat? That is the billion -dollar question. 00:13:11.240 --> 00:13:13.240 And it connects back to that feeling of vulnerability. 00:13:13.559 --> 00:13:16.000 We are building these massive systems, but the 00:13:16.000 --> 00:13:18.460 edges are porous. Speaking of vulnerability, 00:13:18.840 --> 00:13:20.440 reading through these sources made me think about 00:13:20.440 --> 00:13:23.500 my own usage. I have to admit, I still wrestle 00:13:23.500 --> 00:13:26.840 with what they call confident guessing. Oh, hallucinations. 00:13:26.840 --> 00:13:29.080 Yeah, but specifically in high -stakes stuff, 00:13:29.139 --> 00:13:31.860 like I'll use AI for a contract review or a compliance 00:13:31.860 --> 00:13:34.879 check, and it sounds so sure of itself. But there's 00:13:34.879 --> 00:13:37.340 that nagging fear. Is it making this up? The 00:13:37.340 --> 00:13:39.000 source material mentioned a safety prompt rule 00:13:39.000 --> 00:13:41.740 to help with this. Yes, this is a crucial takeaway 00:13:41.740 --> 00:13:44.600 for you listening. The source mentioned a specific 00:13:44.600 --> 00:13:47.000 framework for stopping that confident guessing. 00:13:47.940 --> 00:13:50.279 Essentially, when you are doing something high 00:13:50.279 --> 00:13:53.000 stakes, checking a legal doc, looking for citations, 00:13:53.200 --> 00:13:56.259 you have to explicitly instruct the model on 00:13:56.259 --> 00:13:58.379 the negative constraint. Meaning what? You tell 00:13:58.379 --> 00:14:01.259 it. If you do not find the specific clause in 00:14:01.259 --> 00:14:03.679 the text provided, state that you cannot find 00:14:03.679 --> 00:14:06.919 it, do not infer, do not guess. It sounds so 00:14:06.919 --> 00:14:09.659 simple. But we forget to do it. Yeah. We trust 00:14:09.659 --> 00:14:12.019 the chat interface too much. We treat it like 00:14:12.019 --> 00:14:14.159 a conversation where it's rude to be that blunt. 00:14:14.320 --> 00:14:16.299 But you have to. We need to force it to prove 00:14:16.299 --> 00:14:18.320 its claim. Absolutely. We have to move from being 00:14:18.320 --> 00:14:21.340 passive users, hey, tell me about this, to active 00:14:21.340 --> 00:14:23.820 supervisors. Analyze this. Improve your work. 00:14:24.200 --> 00:14:26.460 So with things like Quinn merging editing and 00:14:26.460 --> 00:14:28.779 generation, what does this actually mean for 00:14:28.779 --> 00:14:31.519 creative workflows? We are moving from generating 00:14:31.519 --> 00:14:35.000 ideas to producing final assets without leaving 00:14:35.000 --> 00:14:37.539 the chat interface. It's a brave new world. It 00:14:37.539 --> 00:14:39.519 certainly is. We're going to take a very quick 00:14:39.519 --> 00:14:41.340 break to thank our partners who help keep this 00:14:41.340 --> 00:14:43.879 deep dive running. And when we come back, we're 00:14:43.879 --> 00:14:46.340 going to try to synthesize all this, the politics, 00:14:46.580 --> 00:14:50.000 the concrete, and the code into one big picture. 00:14:50.639 --> 00:14:54.299 Stay with us. Okay, we're back. We've covered 00:14:54.299 --> 00:14:57.179 a lot of ground today. From the cancellation 00:14:57.179 --> 00:14:59.659 parties of QuitGPT to the massive construction 00:14:59.659 --> 00:15:02.220 sites in Indiana to the model thieves stealing 00:15:02.220 --> 00:15:05.019 intelligence prompt by prompt. Yeah. When you 00:15:05.019 --> 00:15:08.120 look at this entire stack of stories, what is 00:15:08.120 --> 00:15:10.419 the through line for you? I think we are seeing 00:15:10.419 --> 00:15:13.039 a massive divergence. That's the word that keeps 00:15:13.039 --> 00:15:15.039 coming to mind. Avergence between what and what? 00:15:15.159 --> 00:15:17.179 Between the human layer and the machine layer. 00:15:17.320 --> 00:15:18.960 On one hand, you have the human layer pushing 00:15:18.960 --> 00:15:21.460 back. You have the quit GPT movement. You have 00:15:21.460 --> 00:15:23.740 the political maneuvering, the ethical concerns 00:15:23.740 --> 00:15:27.120 about ICE and funding. It's messy. It's emotional. 00:15:27.840 --> 00:15:31.039 It's loud. Right. It's people saying, wait, stop. 00:15:31.200 --> 00:15:32.980 Does this align with our values? And on the other 00:15:32.980 --> 00:15:35.259 hand. On the other hand, the machine layer is 00:15:35.259 --> 00:15:37.460 just accelerating with total indifference to 00:15:37.460 --> 00:15:40.559 that noise. Ten billion dollar chips, massive 00:15:40.559 --> 00:15:43.299 data centers, models that can clone each other. 00:15:43.399 --> 00:15:45.580 Yeah. The technology is no longer just software 00:15:45.580 --> 00:15:48.240 on a screen. It is becoming a physical industrial 00:15:48.240 --> 00:15:51.200 force and a political lightning rod simultaneously. 00:15:51.720 --> 00:15:54.019 It's almost like the technology has gained enough 00:15:54.019 --> 00:15:57.909 momentum that it's. It's decoupling from the 00:15:57.909 --> 00:16:00.370 public sentiment. The boycott is happening, but 00:16:00.370 --> 00:16:02.649 the cement trucks are still pouring. Exactly. 00:16:02.649 --> 00:16:05.350 The train has left the station, and now we are 00:16:05.350 --> 00:16:07.850 just arguing about who gets to sit in the conductor's 00:16:07.850 --> 00:16:11.169 chair. That is a sobering thought. The scale 00:16:11.169 --> 00:16:13.350 of the investment, the sheer physical weight 00:16:13.350 --> 00:16:16.210 of it, makes it very hard to stop or even steer. 00:16:16.490 --> 00:16:19.029 It does. And if I can leave you with one final 00:16:19.029 --> 00:16:21.269 thing that's been sticking in my brain, it's 00:16:21.269 --> 00:16:24.250 that story about the model copying. The Gemini 00:16:24.250 --> 00:16:27.340 theft. Yeah. Think about it. If a trillion -dollar 00:16:27.340 --> 00:16:30.059 company like Google, with all its defenses, can 00:16:30.059 --> 00:16:33.179 have its intelligence stolen via 100 ,000 text 00:16:33.179 --> 00:16:36.200 prompts, what does intellectual property even 00:16:36.200 --> 00:16:39.399 mean in the age of fluid intelligence? If knowledge 00:16:39.399 --> 00:16:41.779 can be decanted from one machine to another just 00:16:41.779 --> 00:16:44.019 by asking questions, the economics of this whole 00:16:44.019 --> 00:16:45.779 industry might be more fragile than we think. 00:16:46.320 --> 00:16:48.580 That is a question I think we're going to be 00:16:48.580 --> 00:16:50.500 wrestling with for the next decade. I think so, 00:16:50.580 --> 00:16:52.620 too. If you want to protect your own workflows, 00:16:53.650 --> 00:16:55.950 especially if you're using these tools for contracts 00:16:55.950 --> 00:16:59.009 or anything legal, I highly recommend you check 00:16:59.009 --> 00:17:02.029 out that safety prompt rule we mentioned. We'll 00:17:02.029 --> 00:17:04.549 have the details and the specific phrasing in 00:17:04.549 --> 00:17:08.250 the show notes. It's a small step, but it might 00:17:08.250 --> 00:17:11.190 save you from a very confident, very wrong answer. 00:17:11.349 --> 00:17:13.450 Worth a click. Thanks for diving Dean with us 00:17:13.450 --> 00:17:16.490 today. It's a complex world out there, but hey, 00:17:16.609 --> 00:17:18.750 at least we're figuring it out together. See 00:17:18.750 --> 00:17:19.829 you next time. Take care, everyone.