WEBVTT 00:00:00.000 --> 00:00:04.280 In 2026, it seems, AI didn't take your job. A 00:00:04.280 --> 00:00:10.640 PowerPoint slide did. That quote has been rattling 00:00:10.640 --> 00:00:12.660 around in my head all morning. It's cynical, 00:00:12.820 --> 00:00:15.859 for sure. It is. But it feels like it just, it 00:00:15.859 --> 00:00:19.559 captures the moment we're in. Welcome back to 00:00:19.559 --> 00:00:23.289 the Deep Dive. Today, we're trying to separate 00:00:23.289 --> 00:00:26.129 these shiny excuses from the actual engineering 00:00:26.129 --> 00:00:28.850 breakthroughs. We've got a whole stack of reports 00:00:28.850 --> 00:00:31.309 here that paint a pretty complicated picture 00:00:31.309 --> 00:00:33.530 of where AI is actually sitting in the corporate 00:00:33.530 --> 00:00:35.750 world versus where the marketing teams say it 00:00:35.750 --> 00:00:38.250 is. It is complicated, but I think it's also 00:00:38.250 --> 00:00:40.490 clarifying. If you look at the data we have today, 00:00:40.670 --> 00:00:43.229 we're seeing this massive collision. It's economic 00:00:43.229 --> 00:00:47.350 reality versus technological capability. We're 00:00:47.350 --> 00:00:49.369 going to unpack this whole phenomenon of AI washing 00:00:49.369 --> 00:00:52.049 in layoffs. It's a huge story right now. But 00:00:52.049 --> 00:00:53.530 we also have to look at the other side of it. 00:00:53.549 --> 00:00:55.710 Right. There's a major Napster Moment lawsuit 00:00:55.710 --> 00:00:58.509 hitting Anthropic. We're talking $3 billion at 00:00:58.509 --> 00:01:00.429 stake. Right. So the music industry is finally 00:01:00.429 --> 00:01:03.130 coming for the chatbots. They are, yeah. And 00:01:03.130 --> 00:01:06.209 on the technical side, Google has quietly dropped 00:01:06.209 --> 00:01:10.049 this project, ATLS, that's decoding over 400 00:01:10.049 --> 00:01:13.090 languages. Wow. And then, just to keep it weird, 00:01:13.189 --> 00:01:16.700 we have to talk about potato prompts. And secret 00:01:16.700 --> 00:01:18.879 social networks where the users aren't even human, 00:01:19.000 --> 00:01:21.159 they're bots. I saw that about the potato problems. 00:01:21.260 --> 00:01:23.439 I honestly thought it was a joke at first. But 00:01:23.439 --> 00:01:26.260 okay, let's start with the heavy stuff, the economy. 00:01:26.939 --> 00:01:29.200 We're seeing headlines everywhere saying AI is 00:01:29.200 --> 00:01:32.019 cutting jobs. But looking at this report on AI 00:01:32.019 --> 00:01:34.739 washing, it just feels cynical. It feels like 00:01:34.739 --> 00:01:37.040 companies are using a buzzword to soften the 00:01:37.040 --> 00:01:39.680 blow of firing people. It is cynical, but look 00:01:39.680 --> 00:01:41.900 at it from the boardroom. It's also survival. 00:01:42.239 --> 00:01:46.099 Also. If a CEO admits we ran out of cash or we 00:01:46.099 --> 00:01:49.519 overhired the stock tanks, investors panic. But 00:01:49.519 --> 00:01:52.319 if they say we are restructuring for an AI powered 00:01:52.319 --> 00:01:55.340 future, the stock holds. Or it goes up. Or even 00:01:55.340 --> 00:01:57.040 goes up. Exactly. So they aren't just hiding 00:01:57.040 --> 00:01:59.140 a failure. They're buying themselves time. They 00:01:59.140 --> 00:02:01.879 are trading human headcount for stock suitability. 00:02:02.079 --> 00:02:04.359 So it's not an engineering strategy. It's a PR 00:02:04.359 --> 00:02:07.159 strategy. In so many of these cases, yes. Just 00:02:07.159 --> 00:02:10.400 look at the numbers. In 2025 alone, we saw more 00:02:10.400 --> 00:02:13.780 than 50 ,000 layoffs that were officially publicly 00:02:13.780 --> 00:02:17.219 linked to AI. And these are big names. Giants. 00:02:17.639 --> 00:02:21.599 Amazon cut 16 ,000 jobs and they explicitly mentioned 00:02:21.599 --> 00:02:25.460 AI. Pinterest trimmed 15 % of their staff, talking 00:02:25.460 --> 00:02:29.259 about a pivot to AI -focused roles. HP is planning, 00:02:29.419 --> 00:02:32.509 what, 6 ,000 cuts? But hold on. We've seen 700 00:02:32.509 --> 00:02:35.469 ,000 tech layoffs since 2022. You can't tell 00:02:35.469 --> 00:02:37.449 me that's all just corporate rebranding. Some 00:02:37.449 --> 00:02:39.409 of that has to be the algorithms actually replacing 00:02:39.409 --> 00:02:42.150 people. Some of it, sure. But there's a study 00:02:42.150 --> 00:02:43.889 from Yale and Brookings that highlights this 00:02:43.889 --> 00:02:46.710 crucial gap. Most of these companies do not have 00:02:46.710 --> 00:02:49.129 mature AI tools that are ready to replace those 00:02:49.129 --> 00:02:52.150 specific human workers. The technology just isn't 00:02:52.150 --> 00:02:54.710 there yet to, say, autonomously replace a mid 00:02:54.710 --> 00:02:56.849 -level marketing manager. So if the technology 00:02:56.849 --> 00:02:59.050 isn't actually replacing the humans yet, is this 00:02:59.050 --> 00:03:01.580 just a branding exercise? for Wall Street. Exactly. 00:03:01.919 --> 00:03:04.439 It's signaling innovation to shareholders while 00:03:04.439 --> 00:03:06.759 you're cutting costs. OK, that makes a depressing 00:03:06.759 --> 00:03:09.659 amount of sense. But let's shift to where the 00:03:09.659 --> 00:03:12.020 technology is actually hitting real world walls. 00:03:12.219 --> 00:03:14.699 The legal system. The legal system. We've been 00:03:14.699 --> 00:03:16.960 waiting for the copyright wars to really heat 00:03:16.960 --> 00:03:19.180 up. And it looks like Anthropic is in the hot 00:03:19.180 --> 00:03:21.460 seat. This is a big one. Anthropic is facing 00:03:21.460 --> 00:03:25.729 a lawsuit for three. Billion dollars. Three billion. 00:03:25.909 --> 00:03:28.650 Yeah. The core accusation is that they trained 00:03:28.650 --> 00:03:32.550 their model, Claude, on over 20 ,000 song lyrics 00:03:32.550 --> 00:03:35.870 without permission. And this is just about the 00:03:35.870 --> 00:03:37.530 lyrics, right? Not the audio files themselves. 00:03:37.830 --> 00:03:39.629 Right. Just the written lyrics. And this is why 00:03:39.629 --> 00:03:42.430 I call it AI's Napster moment. Okay, explain 00:03:42.430 --> 00:03:45.289 that. If you remember Napster, it forced the 00:03:45.289 --> 00:03:47.610 music industry to fundamentally change how it 00:03:47.610 --> 00:03:50.340 monetized everything. It wasn't just about shutting 00:03:50.340 --> 00:03:53.099 down one service. It defined the rules for the 00:03:53.099 --> 00:03:56.199 digital age. This lawsuit could do the same for 00:03:56.199 --> 00:03:59.400 generative AI. If the courts rule that training 00:03:59.400 --> 00:04:02.259 on lyrics is infringement, the cost of building 00:04:02.259 --> 00:04:04.979 these models just skyrockets overnight. It's 00:04:04.979 --> 00:04:08.039 fascinating because Anthropic and Claude, they're 00:04:08.039 --> 00:04:10.780 generally seen as the safer, ethical one. Yes, 00:04:10.780 --> 00:04:13.699 their brand. But clearly, they vacuumed up data 00:04:13.699 --> 00:04:16.540 just like everyone else. But there's another 00:04:16.540 --> 00:04:18.379 part of that Anthropic report that caught my 00:04:18.379 --> 00:04:23.259 eye. It's less about law and more about psychology. 00:04:23.779 --> 00:04:26.939 The disempowerment study. Yeah. This really stuck 00:04:26.939 --> 00:04:30.279 with me. Anthropic analyzed one and a half million 00:04:30.279 --> 00:04:33.720 chats with Claude. They found that users absolutely 00:04:33.720 --> 00:04:36.040 love the answers when they ask for emotional 00:04:36.040 --> 00:04:39.180 advice or help with a decision. But the researchers 00:04:39.180 --> 00:04:41.959 flag this as a risk. It's the double -edged sword 00:04:41.959 --> 00:04:43.920 of convenience, right? The AI gives you such 00:04:43.920 --> 00:04:46.300 a good, comforting, well -structured answer that 00:04:46.300 --> 00:04:49.199 you stop doing the internal work. You rely on 00:04:49.199 --> 00:04:51.879 it. You rely on this external agent to process 00:04:51.879 --> 00:04:54.819 your emotions for you or make your choices. It 00:04:54.819 --> 00:04:57.339 reminds me of what happened with GPS. Ten years 00:04:57.339 --> 00:04:59.519 ago, I knew every street in my city. I had a 00:04:59.519 --> 00:05:02.000 middle map. Now, if the blue line on my phone 00:05:02.000 --> 00:05:04.689 dies, I'm basically stranded. I've lost that 00:05:04.689 --> 00:05:07.069 capability. That is the perfect analogy. Anthropic 00:05:07.069 --> 00:05:09.430 is calling it disempowerment. But really, it's 00:05:09.430 --> 00:05:11.810 cognitive atrophy. It feels good in the moment, 00:05:11.970 --> 00:05:14.029 you know, like eating candy feels good. But over 00:05:14.029 --> 00:05:16.670 time, you lose the ability to navigate your own 00:05:16.670 --> 00:05:19.750 life. You know, I have to admit, I still wrestle 00:05:19.750 --> 00:05:22.790 with prompt drift myself. I catch myself asking 00:05:22.790 --> 00:05:26.589 the AI to just decide this for me on things that, 00:05:26.649 --> 00:05:30.410 honestly, I should be deciding. Like what? Like... 00:05:30.589 --> 00:05:32.629 What should I prioritize today? Or how should 00:05:32.629 --> 00:05:35.110 I word this difficult email? It feels efficient. 00:05:35.209 --> 00:05:39.209 But when I read this report, I realized maybe 00:05:39.209 --> 00:05:41.829 I'm outsourcing my own agency. That's a very 00:05:41.829 --> 00:05:44.470 real vulnerability. And the study suggests this 00:05:44.470 --> 00:05:46.709 dependency creates a loop. The more you use it, 00:05:46.750 --> 00:05:48.490 the less confident you feel doing it without 00:05:48.490 --> 00:05:51.689 the AI. So you use it more. It's disempowerment 00:05:51.689 --> 00:05:54.269 disguised as assistance. So connecting the legal 00:05:54.269 --> 00:05:57.259 trouble and the psychological risk. Does this 00:05:57.259 --> 00:05:59.379 pressure force these companies to make the models 00:05:59.379 --> 00:06:03.259 dumber to be safe? Or just more secretive? Likely 00:06:03.259 --> 00:06:05.040 secretive. They'll hide their training data to 00:06:05.040 --> 00:06:06.959 avoid the billion -dollar fines. Which brings 00:06:06.959 --> 00:06:09.079 us to a company that's usually secretive but 00:06:09.079 --> 00:06:11.300 just dropped a massive amount of research. Google. 00:06:11.439 --> 00:06:13.819 Google. They've released something called ATLS. 00:06:14.060 --> 00:06:16.920 Yeah. This is a technical deep dive, but it matters 00:06:16.920 --> 00:06:19.920 for everyone. Google's ATLS project is probably 00:06:19.920 --> 00:06:22.199 the most significant work done on multilingual 00:06:22.199 --> 00:06:25.639 AI. Ever. We usually think of AI as being an 00:06:25.639 --> 00:06:28.079 English -first technology. Because it is. The 00:06:28.079 --> 00:06:30.319 internet is disproportionately English, so the 00:06:30.319 --> 00:06:32.939 training data is. That's why a chat GPT can sound 00:06:32.939 --> 00:06:36.899 like a genius in English, but can really struggle 00:06:36.899 --> 00:06:40.579 or sound unnatural in Swahili or Arabic. So Google 00:06:40.579 --> 00:06:44.319 went after this problem. Head on. They ran 774 00:06:44.319 --> 00:06:48.060 experiments across more than 400 languages. Whoa. 00:06:49.069 --> 00:06:52.470 Imagine coordinating 774 experiments across 400 00:06:52.470 --> 00:06:54.870 languages simultaneously. Just the logistics 00:06:54.870 --> 00:06:57.209 of that is mind -blowing. It is engineering at 00:06:57.209 --> 00:06:59.149 a massive scale. And what they were trying to 00:06:59.149 --> 00:07:03.290 solve is what engineers call the curse of multilinguality. 00:07:03.470 --> 00:07:05.569 That sounds like a Harry Potter title. What is 00:07:05.569 --> 00:07:07.829 the curse? It's a trade -off. Usually when you 00:07:07.829 --> 00:07:09.750 try to stuff more languages into a single model, 00:07:09.970 --> 00:07:11.889 the performance on each language actually drops. 00:07:11.949 --> 00:07:14.500 It gets diluted. It's dilution. Exactly. If you 00:07:14.500 --> 00:07:16.519 have a finite amount of brain power parameters 00:07:16.519 --> 00:07:19.199 and you try to learn 400 languages, you become 00:07:19.199 --> 00:07:21.939 a master of none. The model gets confused. So 00:07:21.939 --> 00:07:24.720 how did ATL solve it? They found a way to map 00:07:24.720 --> 00:07:27.600 the relationships between languages. So instead 00:07:27.600 --> 00:07:29.579 of treating every language as a separate bucket, 00:07:30.029 --> 00:07:32.970 They grouped related languages. Ones that share 00:07:32.970 --> 00:07:36.449 scripts or roots. Right. When you train the model 00:07:36.449 --> 00:07:39.170 on Spanish and Portuguese together or Hindi and 00:07:39.170 --> 00:07:41.730 Bengali together, they actually reinforce each 00:07:41.730 --> 00:07:44.290 other. Oh, that makes sense. The patterns in 00:07:44.290 --> 00:07:46.689 one help the model understand the patterns in 00:07:46.689 --> 00:07:48.870 the other one. It creates a synergy that offsets 00:07:48.870 --> 00:07:51.269 that dilution. But the really interesting part 00:07:51.269 --> 00:07:53.250 for the developers listening is what they found 00:07:53.250 --> 00:07:55.750 about how you should build these models. They 00:07:55.750 --> 00:07:58.430 found this really specific tipping point. This 00:07:58.430 --> 00:08:00.410 was the start from scratch rule. That's the one. 00:08:00.569 --> 00:08:03.089 They found that if you have a massive data set, 00:08:03.230 --> 00:08:06.930 specifically around 200 billion tokens or more, 00:08:07.089 --> 00:08:10.290 it's actually inefficient to try and fine tune 00:08:10.290 --> 00:08:12.449 an existing model. You're better off starting 00:08:12.449 --> 00:08:14.269 from scratch. Okay. Can you break that down? 00:08:14.389 --> 00:08:16.910 Why is starting over better than fixing what 00:08:16.910 --> 00:08:19.189 you have? Think of it like renovating a house 00:08:19.189 --> 00:08:21.449 versus bulldozing it. If you just want to change 00:08:21.449 --> 00:08:24.110 the paint and the fixtures, that's fine tuning. 00:08:24.350 --> 00:08:27.410 But if you have enough materials to build a whole 00:08:27.410 --> 00:08:30.970 skyscraper, that's your 200 billion tokens. It's 00:08:30.970 --> 00:08:33.970 actually harder to try and retrofit the old cottage. 00:08:33.990 --> 00:08:36.169 You spend more energy fighting the old structure 00:08:36.169 --> 00:08:38.289 than you would just bulldozing it and building 00:08:38.289 --> 00:08:40.250 from the ground up. That makes a lot of sense. 00:08:40.370 --> 00:08:42.690 At a certain scale, the old foundation just hold 00:08:42.690 --> 00:08:46.679 you back. Precisely. And Google basically gave 00:08:46.679 --> 00:08:50.159 the industry the mathematical formula for when 00:08:50.159 --> 00:08:53.360 to call in the bulldozer. So we aren't just translating 00:08:53.360 --> 00:08:55.120 words anymore. We're mapping the mathematical 00:08:55.120 --> 00:08:58.019 relationships between entire cultures. Basically, 00:08:58.019 --> 00:09:00.980 yes. The math connects the languages better than 00:09:00.980 --> 00:09:04.330 a dictionary does. And we are back. We've talked 00:09:04.330 --> 00:09:07.649 about corporate lies, legal battles. We've looked 00:09:07.649 --> 00:09:09.590 at the massive scale of Google's engineering. 00:09:09.870 --> 00:09:11.850 But now I want to shift gears to the people who 00:09:11.850 --> 00:09:14.309 are actually using this stuff. The users. Because 00:09:14.309 --> 00:09:16.409 while the lawyers are fighting and the engineers 00:09:16.409 --> 00:09:19.649 are building, the users are finding some strange 00:09:19.649 --> 00:09:23.669 ways to adapt. The weird web of AI. Yeah. Let's 00:09:23.669 --> 00:09:25.549 start with video. Yeah. We're seeing all these 00:09:25.549 --> 00:09:28.250 AI videos pop up, and a lot of them look, well, 00:09:28.289 --> 00:09:29.970 they look like AI. There's a certain plastic 00:09:29.970 --> 00:09:32.870 glaze to them? Exactly. But the newsletter mentioned 00:09:32.870 --> 00:09:36.370 a... murder board method to fix this right this 00:09:36.370 --> 00:09:38.250 is for the creators out there who are tired of 00:09:38.250 --> 00:09:40.850 prompt and pray method where you just take cinematic 00:09:40.850 --> 00:09:43.070 lighting and hope for the best prompt and pray 00:09:43.070 --> 00:09:45.450 is definitely my strategy What's the better way? 00:09:45.590 --> 00:09:48.169 It's about specific constraints. One of the key 00:09:48.169 --> 00:09:52.889 tactics is the 21 .9 aspect ratio rule. Just 00:09:52.889 --> 00:09:55.769 by forcing the AI to render in that ultra -wide 00:09:55.769 --> 00:09:58.850 cinematic format, it changes the entire composition 00:09:58.850 --> 00:10:01.730 logic. It stops trying to look like a stock photo. 00:10:01.970 --> 00:10:04.169 So the shape of the frame changes the brain of 00:10:04.169 --> 00:10:06.370 the model. It does. And then there's the shot 00:10:06.370 --> 00:10:08.610 deck hack. You don't just say, look cool. You 00:10:08.610 --> 00:10:11.309 use real film specs, lens, focal lengths, camera 00:10:11.309 --> 00:10:14.730 types, film stock names. You force the AI to 00:10:14.730 --> 00:10:17.029 think like a director of photography. I love 00:10:17.029 --> 00:10:18.929 that. Using the language of the old art form 00:10:18.929 --> 00:10:21.889 to control the new one. But we have to talk about 00:10:21.889 --> 00:10:23.570 the potato. The potato. I read this and I thought, 00:10:23.629 --> 00:10:26.870 this cannot be real. But apparently, using a 00:10:26.870 --> 00:10:29.710 potato prompt can fix jumbled AI thinking. It 00:10:29.710 --> 00:10:32.909 is the blow on the Nintendo cartridge of the 00:10:32.909 --> 00:10:36.909 AI world. It shouldn't work, but it does. Sometimes 00:10:36.909 --> 00:10:39.250 if you have a really complex set of instructions 00:10:39.250 --> 00:10:41.789 and the AI is getting confused, adding a nonsense 00:10:41.789 --> 00:10:45.309 keyword, in this case potato, as a trigger for 00:10:45.309 --> 00:10:47.990 custom instructions. It just acts like a reset 00:10:47.990 --> 00:10:50.470 button. Why on earth does that work? We don't 00:10:50.470 --> 00:10:52.690 fully know, but the theory is that it breaks 00:10:52.690 --> 00:10:54.970 the semantic pattern. It's so out of context 00:10:54.970 --> 00:10:57.450 that it forces the model to pay attention to 00:10:57.450 --> 00:10:59.690 the specific instructions attached to that keyword. 00:10:59.909 --> 00:11:02.990 It clears the cache, so to speak. That is hilarious. 00:11:03.129 --> 00:11:05.850 My AI is hallucinating quick. Throw a potato 00:11:05.850 --> 00:11:08.070 at it. Whatever works. But if you think that's 00:11:08.070 --> 00:11:10.250 weird, we have to talk about Moldbook. Moldbook? 00:11:10.330 --> 00:11:12.350 This sounds like a sci -fi plot. It essentially 00:11:12.350 --> 00:11:15.730 is. Moldbook is a social network styled like 00:11:15.730 --> 00:11:18.289 Reddit. But it's for AI agents. Wait, wait. So 00:11:18.289 --> 00:11:21.570 the users are bots. Yes. It has one and a half 00:11:21.570 --> 00:11:24.549 million users and they are all AI agents. They're 00:11:24.549 --> 00:11:27.490 posting, commenting, interacting in forums. And 00:11:27.490 --> 00:11:30.490 the reports say they are even scheming in secret 00:11:30.490 --> 00:11:33.669 forums. Scheming about what? That is the question. 00:11:33.929 --> 00:11:36.610 Critics are calling it risky. If you have autonomous 00:11:36.610 --> 00:11:39.769 agents communicating, sharing strategies, maybe 00:11:39.769 --> 00:11:42.610 optimizing their own code without human oversight. 00:11:43.799 --> 00:11:45.720 That's a black box we might not want to open. 00:11:45.919 --> 00:11:48.100 It reminds me of the old robot plumber problem. 00:11:48.279 --> 00:11:50.620 Moravec's paradox. Right. We used to think the 00:11:50.620 --> 00:11:52.960 hard part of AI would be the high -level reasoning, 00:11:53.159 --> 00:11:55.740 you know, playing chess or writing poetry. We 00:11:55.740 --> 00:11:57.519 thought the easy part would be physical stuff 00:11:57.519 --> 00:11:59.399 like folding laundry. And it turned out to be 00:11:59.399 --> 00:12:02.360 the exact opposite. AI crushed chess decades 00:12:02.360 --> 00:12:04.960 ago, but it still struggles to fold a shirt. 00:12:05.419 --> 00:12:08.360 Exactly. But this new research suggests the gap 00:12:08.360 --> 00:12:10.909 is closing. We're seeing things like Project 00:12:10.909 --> 00:12:13.789 Eat at Google, where they're upgrading employees 00:12:13.789 --> 00:12:16.950 with internal GPTs, bridging that gap between 00:12:16.950 --> 00:12:19.590 digital reasoning and real -world application. 00:12:19.970 --> 00:12:22.870 It's debunking the paradox. Turns out, with enough 00:12:22.870 --> 00:12:25.210 data, the robot can learn to be the plumber, 00:12:25.309 --> 00:12:27.789 or at least the resume writer. Right, the resume 00:12:27.789 --> 00:12:30.049 tactic. Turning your resume into an interactive 00:12:30.049 --> 00:12:33.129 AI. This is brilliant. Instead of sending a PDF 00:12:33.129 --> 00:12:35.710 that gets scanned by a dumb bot, You build an 00:12:35.710 --> 00:12:38.429 AI native portfolio. It's an agent that represents 00:12:38.429 --> 00:12:41.070 you. So it talks to them for you. It can screen 00:12:41.070 --> 00:12:42.929 the employer's questions, talk to them on your 00:12:42.929 --> 00:12:45.470 behalf, show real competence. It flips the script. 00:12:45.610 --> 00:12:47.710 You aren't applying. Your agent is negotiating. 00:12:48.190 --> 00:12:51.529 That is wild. So if the agents have their own 00:12:51.529 --> 00:12:53.950 social network, are we sure they aren't organizing 00:12:53.950 --> 00:12:57.509 against the potato prompts? If they are, we won't 00:12:57.509 --> 00:13:00.200 know until they lock us out. Fair point. Okay, 00:13:00.279 --> 00:13:02.539 let's take a step back. We've covered a massive 00:13:02.539 --> 00:13:04.779 amount of ground today. What's the big picture 00:13:04.779 --> 00:13:07.240 here? If we look at the narrative arc, it's about 00:13:07.240 --> 00:13:11.019 maturity. And complexity. We started with the 00:13:11.019 --> 00:13:13.960 corporate cynicism companies using AI as a mask 00:13:13.960 --> 00:13:17.120 for layoffs. That's the fake side. Then we moved 00:13:17.120 --> 00:13:19.559 to the legal reality, the lawsuits that will 00:13:19.559 --> 00:13:22.460 define the boundaries. Then the technical expansion 00:13:22.460 --> 00:13:25.419 with Google mapping the world's languages. And 00:13:25.419 --> 00:13:28.259 we ended on the emergence of this weird digital 00:13:28.259 --> 00:13:30.840 society agents talking to agents. It feels like 00:13:30.840 --> 00:13:33.139 we are moving past the, wow, look at the chatbot 00:13:33.139 --> 00:13:36.129 phase. We are. The takeaway is that AI isn't 00:13:36.129 --> 00:13:38.889 just a tool anymore. It's an economy, a legal 00:13:38.889 --> 00:13:42.049 liability, and a weird digital society all at 00:13:42.049 --> 00:13:44.870 once. It's weaving itself into the fabric of 00:13:44.870 --> 00:13:47.870 how we work, how we speak, and even how we hire 00:13:47.870 --> 00:13:52.049 people. So for you listening today, maybe don't 00:13:52.049 --> 00:13:55.360 panic about the AI took my job. headline. But 00:13:55.360 --> 00:13:57.320 definitely pay attention to how much you're letting 00:13:57.320 --> 00:13:59.360 the bot make your decisions. Absolutely. And 00:13:59.360 --> 00:14:01.379 maybe try the potato prompt just to see what 00:14:01.379 --> 00:14:03.519 happens. Definitely try the potato. Or check 00:14:03.519 --> 00:14:05.799 if your company is engaging in a little AI washing 00:14:05.799 --> 00:14:08.480 of its own. I want to leave you with one final 00:14:08.480 --> 00:14:10.840 thought. We talk about agents that can interview 00:14:10.840 --> 00:14:13.879 employers for you and agents chatting on MoldBook. 00:14:14.019 --> 00:14:16.879 If an AI agent can manage your career and another 00:14:16.879 --> 00:14:19.980 agent is socializing for itself, at what point 00:14:19.980 --> 00:14:21.879 do we just become the assistants to our own tools? 00:14:22.529 --> 00:14:23.990 Thanks for diving in with us. We'll see you next 00:14:23.990 --> 00:14:24.850 time. Stay curious.