WEBVTT 00:00:00.000 --> 00:00:01.480 Okay, wait, hold on a second. I just need to 00:00:01.480 --> 00:00:03.459 make sure I heard you read the pre -show. You 00:00:03.459 --> 00:00:07.160 said Sam Altman's kids were supposed to run the 00:00:07.160 --> 00:00:10.500 Mars colony. Yes, that is exactly what the leaked 00:00:10.500 --> 00:00:13.880 email said. It's part of the whole lawsuit discovery. 00:00:14.259 --> 00:00:17.500 Altman proposed this $80 billion project for 00:00:17.500 --> 00:00:20.539 a Mars colony. And yeah, the text implies that... 00:00:20.839 --> 00:00:24.480 the governance structure that involves his children. 00:00:24.579 --> 00:00:26.760 Your biological children or his AGI children. 00:00:27.000 --> 00:00:29.800 You know, knowing that whole circle of tech philosophy, 00:00:29.920 --> 00:00:32.060 genuinely hard to tell. Maybe they see them as 00:00:32.060 --> 00:00:33.979 the same thing eventually. But here's the real 00:00:33.979 --> 00:00:36.939 kicker. Elon Musk's reply to this whole pitch 00:00:36.939 --> 00:00:40.210 was just, it was brutal. I can only imagine. 00:00:40.429 --> 00:00:42.869 He just replied with a single statistic. He called 00:00:42.869 --> 00:00:46.170 the probability of success zero percent. Zero. 00:00:46.270 --> 00:00:48.770 Zero. Okay. Welcome back to the Deem Time. I'm 00:00:48.770 --> 00:00:51.009 really glad you're here with us. We usually spend 00:00:51.009 --> 00:00:52.929 our time trying to filter out all the noise, 00:00:53.090 --> 00:00:54.729 you know, find the signal in these huge data 00:00:54.729 --> 00:00:58.609 dumps. But today. I mean, the noise is incredibly 00:00:58.609 --> 00:01:01.130 loud. It is loud. But if you listen really closely, 00:01:01.270 --> 00:01:04.409 there's a very distinct rhythm underneath it 00:01:04.409 --> 00:01:06.549 all. Right. And we aren't just talking about 00:01:06.549 --> 00:01:08.730 the usual chatbot headlines today. We're looking 00:01:08.730 --> 00:01:11.549 at a fundamental shift in competence. We're talking 00:01:11.549 --> 00:01:14.209 about AI putting on a lab coat at Stanford. Yeah. 00:01:14.290 --> 00:01:17.030 To cure diseases. We're talking about AI literally 00:01:17.030 --> 00:01:20.420 breaking out of the screen. To what? Fix your 00:01:20.420 --> 00:01:22.840 dishwasher. And maybe most significantly, we're 00:01:22.840 --> 00:01:25.040 talking about a massive geopolitical shift in 00:01:25.040 --> 00:01:27.420 hardware. Something coming out of China that 00:01:27.420 --> 00:01:30.579 just changes the entire board. It feels like 00:01:30.579 --> 00:01:35.280 the theme for this dive is doing work. We're 00:01:35.280 --> 00:01:37.840 moving past the era of, you know, generating 00:01:37.840 --> 00:01:40.200 interesting text and into the era of executing 00:01:40.200 --> 00:01:43.700 complex tasks from the biology lab to the factory 00:01:43.700 --> 00:01:45.319 floor and then all the way to the global supply 00:01:45.319 --> 00:01:47.700 chain. There's so much to unpack here. Let's 00:01:47.700 --> 00:01:49.680 start with the scientists. Let's do it. So we 00:01:49.680 --> 00:01:51.599 have this really fascinating study from Anthropic. 00:01:51.599 --> 00:01:54.319 It's titled The AI Frontier, Scientific Research 00:01:54.319 --> 00:01:56.760 and Physical Evolution. And I think for a lot 00:01:56.760 --> 00:01:58.099 of people listening, this is going to require 00:01:58.099 --> 00:02:01.420 a real mental shift about what an LLM actually 00:02:01.420 --> 00:02:04.769 is. Ugh, a huge shift. Because usually when we 00:02:04.769 --> 00:02:07.349 talk about large language models, you know, Claude, 00:02:07.430 --> 00:02:09.969 GPT, whatever, we think of them as writing assistants. 00:02:10.169 --> 00:02:13.969 Like, Claude, summarize this PDF. Or, hey, write 00:02:13.969 --> 00:02:16.650 me a polite email to my boss. That kind of thing. 00:02:16.830 --> 00:02:18.930 Exactly. It's a retrieval tool, a formatting 00:02:18.930 --> 00:02:21.729 tool. It's basically a sophisticated autocomplete. 00:02:21.930 --> 00:02:24.909 Right. But what Anthropic did was profile three 00:02:24.909 --> 00:02:27.509 different labs, and these are serious, high -level 00:02:27.509 --> 00:02:29.930 research labs, where they're using Claude to 00:02:29.930 --> 00:02:32.909 do work that... That normally takes human scientists 00:02:32.909 --> 00:02:37.729 months, and the AI is doing it in minutes. And 00:02:37.729 --> 00:02:39.669 it's not just doing the grunt work. It's actually 00:02:39.669 --> 00:02:42.210 doing the thinking. I mean, take Stanford's Biomne 00:02:42.210 --> 00:02:44.610 project as the prime example. Biomne is fascinating. 00:02:44.710 --> 00:02:46.710 They call this a research agent. Okay, before 00:02:46.710 --> 00:02:48.490 we go further, we should probably clarify that 00:02:48.490 --> 00:02:50.569 term. When we say agent here, we're not just 00:02:50.569 --> 00:02:52.270 talking about a chatbot with a personality, are 00:02:52.270 --> 00:02:54.729 we? No, definitely not. A research agent is a 00:02:54.729 --> 00:02:57.449 system that doesn't just chat. It uses tools 00:02:57.449 --> 00:03:01.139 to execute tasks. Think of it this way. A chatbot 00:03:01.139 --> 00:03:05.090 is a brain in a jar. It can talk to you. An agent 00:03:05.090 --> 00:03:08.270 is a brain that's been given digital hands that 00:03:08.270 --> 00:03:12.710 can open files, run code, search databases. It 00:03:12.710 --> 00:03:15.270 can execute a whole workflow on its own. So it 00:03:15.270 --> 00:03:17.270 has agency. It can actually interact with its 00:03:17.270 --> 00:03:20.289 environment. Exactly. And Biomni is working across 00:03:20.289 --> 00:03:23.569 25 different scientific fields. It's building 00:03:23.569 --> 00:03:26.610 this memory bank of skills. It learns expert 00:03:26.610 --> 00:03:28.689 workflows so it doesn't have to start from scratch 00:03:28.689 --> 00:03:31.430 every single time. See, that's the key differentiator. 00:03:31.509 --> 00:03:33.569 It's not just retrieving data. It's figuring 00:03:33.569 --> 00:03:35.909 out. how to do the science. Yeah. But the example 00:03:35.909 --> 00:03:37.610 that really stood out to me, the one that really 00:03:37.610 --> 00:03:40.389 made me pause, was from MIT. Oh, the Matsuram 00:03:40.389 --> 00:03:42.729 Project. Such a great name, by the way. Scientists 00:03:42.729 --> 00:03:44.789 love a good pun. They really do. So they were 00:03:44.789 --> 00:03:47.270 using it for CRISPR screens. For anyone listening 00:03:47.270 --> 00:03:49.550 who might not be a biologist, CRISPR is basically 00:03:49.550 --> 00:03:53.490 gene editing. And these screens, they're usually 00:03:53.490 --> 00:03:56.000 a brute force process, right? Oh, incredibly 00:03:56.000 --> 00:03:58.979 brute force. You basically have to disable thousands 00:03:58.979 --> 00:04:01.360 of genes one by one just to see what breaks. 00:04:01.539 --> 00:04:04.699 It's tedious. It's expensive. And it's so prone 00:04:04.699 --> 00:04:06.979 to human error because you're just staring at 00:04:06.979 --> 00:04:09.960 these massive data sets for weeks. And this is 00:04:09.960 --> 00:04:12.900 where the story gets a little wild. Because Claude 00:04:12.900 --> 00:04:15.039 didn't just speed up the spreadsheet work. No. 00:04:15.319 --> 00:04:17.339 This is the part that just stopped me in my tracks. 00:04:17.439 --> 00:04:20.709 I mean, whoa. Yeah. Get this. Claude actually 00:04:20.709 --> 00:04:23.850 analyzed the data and it spotted an RNA modification 00:04:23.850 --> 00:04:26.870 pathway that the human scientists had completely 00:04:26.870 --> 00:04:29.769 missed. Wait, hold on. Let's pause right there. 00:04:29.870 --> 00:04:32.470 It found a biological reality that the human 00:04:32.470 --> 00:04:35.389 experts overlooked. Yes. When they benchmarked 00:04:35.389 --> 00:04:37.410 it, you know, against other models and against 00:04:37.410 --> 00:04:40.050 the human analysis, Claude identified a pathway 00:04:40.050 --> 00:04:42.269 that wasn't obvious. It saw a pattern in the 00:04:42.269 --> 00:04:44.490 noise that the humans just didn't see. That feels 00:04:44.490 --> 00:04:46.810 like a pivotal moment. It's one thing to say 00:04:46.810 --> 00:04:49.829 AI can calculate faster than me. We're used to 00:04:49.829 --> 00:04:52.069 that. It's a totally different thing to say the 00:04:52.069 --> 00:04:54.430 AI noticed something about biology that I didn't. 00:04:54.449 --> 00:04:56.490 It completely changes the value proposition. 00:04:56.610 --> 00:04:59.170 It moves from assistant to collaborator. And 00:04:59.170 --> 00:05:00.790 there's a third example, right? It touched on 00:05:00.790 --> 00:05:03.149 the economics of all this. The Lindbergh lab 00:05:03.149 --> 00:05:06.779 at Stanford. Right. This is all about cost. Gene 00:05:06.779 --> 00:05:09.259 screens are incredibly expensive. We're talking 00:05:09.259 --> 00:05:12.259 something like $20 ,000 per screen. 20 grand. 00:05:12.459 --> 00:05:14.480 Yeah. So you can't just test everything. You 00:05:14.480 --> 00:05:16.420 have to be picky. You have to place your bets 00:05:16.420 --> 00:05:19.759 very, very carefully. Usually, scientists use 00:05:19.759 --> 00:05:22.899 a mix of spreadsheets, some intuition, and existing 00:05:22.899 --> 00:05:25.819 research to guess which genes to target. It's 00:05:25.819 --> 00:05:29.819 an educated guess, but at $20 ,000 a pop, it's 00:05:29.819 --> 00:05:32.540 still a big gamble. So how did the agent change 00:05:32.540 --> 00:05:35.980 that gamble? Well, they had Claude navigate a 00:05:35.980 --> 00:05:38.879 molecule relationship map. So instead of just 00:05:38.879 --> 00:05:41.759 looking at lists, the AI analyzed this whole 00:05:41.759 --> 00:05:44.540 network of connections between molecules, and 00:05:44.540 --> 00:05:46.939 it suggested the targets based on those relationships. 00:05:47.259 --> 00:05:49.540 So it's replacing intuition with network analysis. 00:05:50.000 --> 00:05:52.660 Precisely. And the study emphasizes this point 00:05:52.660 --> 00:05:55.139 over and over again. None of these breakthroughs 00:05:55.139 --> 00:05:56.939 came from just chatting with the bot. Right. 00:05:57.000 --> 00:05:59.420 That goes back to your agent definition. Yeah, 00:05:59.480 --> 00:06:01.500 the breakthrough is connecting Claude to tools, 00:06:01.680 --> 00:06:04.779 giving it a custom workflow, adding guardrails 00:06:04.779 --> 00:06:06.620 so it doesn't just hallucinate and make things 00:06:06.620 --> 00:06:08.660 up. It's like the difference between asking a 00:06:08.660 --> 00:06:11.860 carpenter to describe a table and giving a carpenter 00:06:11.860 --> 00:06:14.519 a hammer and wood and saying, build it. That's 00:06:14.519 --> 00:06:16.519 a perfect analogy. So this brings up the big 00:06:16.519 --> 00:06:19.540 question for me here. If AI can cut the time 00:06:19.540 --> 00:06:21.339 from months to minutes and find things that we 00:06:21.339 --> 00:06:25.459 miss, does this make science cheaper? Or does 00:06:25.459 --> 00:06:27.519 it just mean we start asking much, much harder 00:06:27.519 --> 00:06:30.480 questions? It removes the grunt work so brains 00:06:30.480 --> 00:06:32.420 can focus on the breakthrough. So we just level 00:06:32.420 --> 00:06:34.300 up the difficulty because we have a smarter teammate. 00:06:34.519 --> 00:06:36.480 We have to. All the easy problems are already 00:06:36.480 --> 00:06:39.759 solved. All right. Let's move from the sterile 00:06:39.759 --> 00:06:41.720 lab environment to something a little messier. 00:06:41.959 --> 00:06:45.319 The kitchen, the garage, the factory floor. The 00:06:45.319 --> 00:06:47.579 real world. Exactly. We're talking about physical 00:06:47.579 --> 00:06:50.529 AI. Yeah. Now, I have to admit, when I first 00:06:50.529 --> 00:06:53.069 saw that term in the source material, my marketing 00:06:53.069 --> 00:06:55.889 alarm bells went off. Physical AI just sounds 00:06:55.889 --> 00:06:57.649 like a buzzword somebody made up to sell more 00:06:57.649 --> 00:07:00.350 chips. It definitely has that ring to it. I get 00:07:00.350 --> 00:07:03.350 it. But if you look at the actual capital expenditure, 00:07:03.649 --> 00:07:06.910 the money is telling a different story. The source 00:07:06.910 --> 00:07:10.550 material is adamant about this. 2026 won't be 00:07:10.550 --> 00:07:13.649 on screens. That is a very bold claim considering 00:07:13.649 --> 00:07:15.910 how much of our lives are on screens. It is. 00:07:16.069 --> 00:07:18.189 But look at the investment stack. We're looking 00:07:18.189 --> 00:07:23.170 at a $123 billion ecosystem. Yeah. NVIDIA, Tesla, 00:07:23.329 --> 00:07:26.649 the entire robotics supply chain, they are all 00:07:26.649 --> 00:07:29.029 racing towards this. So when they say physical 00:07:29.029 --> 00:07:32.050 AI, what are they actually building? It's really 00:07:32.050 --> 00:07:34.689 about embodiment. For the last few years, AI 00:07:34.689 --> 00:07:37.269 has been a brain in a jar, right? It's a server 00:07:37.269 --> 00:07:39.670 farm in Virginia. It outputs text or images. 00:07:40.189 --> 00:07:42.370 Physical AI is about giving that brain a body. 00:07:42.470 --> 00:07:44.149 We're putting that brain inside the bodies we 00:07:44.149 --> 00:07:46.790 already have. You know, our cars, our appliances. 00:07:47.110 --> 00:07:49.649 Right. The source material had the slightly hilarious 00:07:49.649 --> 00:07:51.970 example of a future where your fridge is negotiating 00:07:51.970 --> 00:07:54.629 with your dishwasher. I saw that. Yeah. And honestly, 00:07:54.750 --> 00:07:56.670 it sounds like a bad cartoon. Why do I need my 00:07:56.670 --> 00:07:58.779 appliances to gossip with each other? I know, 00:07:58.899 --> 00:08:01.680 it sounds ridiculous. But think about the internet 00:08:01.680 --> 00:08:04.459 of things. We've been hearing about IoT for a 00:08:04.459 --> 00:08:07.079 decade. Your toaster will talk to the internet. 00:08:07.339 --> 00:08:10.399 And for 10 years, all that really meant was that 00:08:10.399 --> 00:08:12.459 your toaster needed a firmware update and had 00:08:12.459 --> 00:08:15.220 a security vulnerability. It was just dumb connectivity. 00:08:15.600 --> 00:08:20.240 Exactly. Physical AI is about competence. It's 00:08:20.240 --> 00:08:23.259 not just connected. It has agency. It's about 00:08:23.259 --> 00:08:25.800 the fridge, analyzing what's inside, realizing 00:08:25.800 --> 00:08:28.399 you're out of milk, checking the dishwasher cycle 00:08:28.399 --> 00:08:30.839 to coordinate energy use to the cheapest time 00:08:30.839 --> 00:08:33.539 of day, and maybe ordering groceries based on 00:08:33.539 --> 00:08:35.399 your diet plan. So it's the fulfillment of the 00:08:35.399 --> 00:08:38.179 smart home promise, but like actually smart this 00:08:38.179 --> 00:08:40.580 time. And it's about the supply chain. If you're 00:08:40.580 --> 00:08:43.419 a consumer or a creator, you are now part of 00:08:43.419 --> 00:08:45.970 this loop. This isn't just about cool robots 00:08:45.970 --> 00:08:48.309 doing backflips. It's about the entire physical 00:08:48.309 --> 00:08:51.149 world getting a new operating system. That is 00:08:51.149 --> 00:08:53.509 both exciting and, frankly, a little terrifying. 00:08:53.789 --> 00:08:55.710 I'm just picturing my fridge judging my late 00:08:55.710 --> 00:08:57.789 -night snack choices. Oh, it absolutely will. 00:08:57.889 --> 00:08:59.509 I'm sorry, Dave. I can't open the door. You've 00:08:59.509 --> 00:09:01.710 exceeded your calorie limit for the day. See, 00:09:01.750 --> 00:09:03.649 that's where I pull the plug. But a serious question, 00:09:03.769 --> 00:09:07.370 then. Yeah. Are we ready for our appliances to 00:09:07.370 --> 00:09:10.409 have agency? Or do we just want better automation? 00:09:10.830 --> 00:09:12.809 We want them to do the dishes, not argue about 00:09:12.809 --> 00:09:15.500 the soap. Right. We want the labor, not the debate. 00:09:15.820 --> 00:09:18.779 Exactly. Okay. Let's pivot from the domestic 00:09:18.779 --> 00:09:22.240 to the geopolitical. Because while the U .S. 00:09:22.240 --> 00:09:25.000 is focusing on robots, there's this massive story 00:09:25.000 --> 00:09:28.019 coming out of China about the chips that run 00:09:28.019 --> 00:09:30.620 them. Hardware wars. We have a story here about 00:09:30.620 --> 00:09:33.899 Zippo AI launching a new model. It's called GLM 00:09:33.899 --> 00:09:36.559 Image. And on the surface, okay, it's an image 00:09:36.559 --> 00:09:39.500 generator, like Midjourney or DALI. Right. It's 00:09:39.500 --> 00:09:43.200 a 16 billion parameter model. Decent size, pretty 00:09:43.200 --> 00:09:45.539 capable. But the specs of the model aren't the 00:09:45.539 --> 00:09:48.059 real headline here. The headline is the silicon 00:09:48.059 --> 00:09:50.100 it was trained on. Exactly. Usually when you 00:09:50.100 --> 00:09:51.679 read a paper like this, you scroll down to the 00:09:51.679 --> 00:09:53.139 hardware section and it always says, you know, 00:09:53.159 --> 00:09:55.980 trained on NVIDIA H100 clusters. Industry standard. 00:09:56.179 --> 00:09:59.960 But this one, no NVIDIA, no AMD, no US tech at 00:09:59.960 --> 00:10:02.620 all. This was trained entirely on Huawei hardware. 00:10:03.120 --> 00:10:05.679 That is a huge, huge signal. We've been talking 00:10:05.679 --> 00:10:08.179 about the chip bans for years. The U .S. government 00:10:08.179 --> 00:10:11.120 placed these really strict restrictions to try 00:10:11.120 --> 00:10:14.759 and limit China's access to high -end AI training 00:10:14.759 --> 00:10:17.759 chips. And the prevailing theory in Washington 00:10:17.759 --> 00:10:20.080 was that this would, you know, cripple their 00:10:20.080 --> 00:10:22.740 ability to train foundation models. It was supposed 00:10:22.740 --> 00:10:26.480 to be a chokehold. But Zipu AI just proved that 00:10:26.480 --> 00:10:28.860 that chokehold might be slipping. They build 00:10:28.860 --> 00:10:31.259 a full domestic stack. This is what China has 00:10:31.259 --> 00:10:34.259 been chasing for years. Independence. Let's look 00:10:34.259 --> 00:10:36.039 under the hood for a second, because the way 00:10:36.039 --> 00:10:38.179 they did this is technically pretty interesting. 00:10:38.320 --> 00:10:41.000 It's not just a copy paste of a U .S. model. 00:10:41.139 --> 00:10:42.960 No, the architecture is quite clever, actually. 00:10:43.000 --> 00:10:46.080 They use a two stage system. OK. Bring that down 00:10:46.080 --> 00:10:50.080 for us. So stage one uses what's called an autoregressive 00:10:50.080 --> 00:10:52.039 transformer. All right, jargon alert. Let's define 00:10:52.039 --> 00:10:54.299 autoregressive transformer for everyone listening. 00:10:54.500 --> 00:10:56.940 Yeah, the simplest definition is it's a model 00:10:56.940 --> 00:10:59.559 that predicts the next piece of data in a sequence. 00:10:59.860 --> 00:11:02.179 Kind of like how chat GPT predicts the next word 00:11:02.179 --> 00:11:05.309 in a sentence. Exactly like that. But here it's 00:11:05.309 --> 00:11:07.470 predicting the layout of an image. It's kind 00:11:07.470 --> 00:11:09.570 of acting like an architect. So it creates the 00:11:09.570 --> 00:11:11.870 blueprint first. Right. It predicts what they 00:11:11.870 --> 00:11:15.029 call semantic VQ tokens. Basically, it figures 00:11:15.029 --> 00:11:16.929 out the meaning and the structure of the image 00:11:16.929 --> 00:11:19.730 first. That's a nine billion parameter model 00:11:19.730 --> 00:11:22.529 doing the planning. And then stage two. Stage 00:11:22.529 --> 00:11:25.450 two is a diffusion transformer. This is the painter. 00:11:25.629 --> 00:11:28.169 It takes that blueprint and it renders the final 00:11:28.169 --> 00:11:31.549 pixels. So logic first, then aesthetics. Yes. 00:11:32.429 --> 00:11:34.769 Because of this split approach, it's really good 00:11:34.769 --> 00:11:37.230 at text -heavy images, which has been a weak 00:11:37.230 --> 00:11:39.190 point for some other models. Now, the source 00:11:39.190 --> 00:11:42.929 material does mention that in terms of pure artistic 00:11:42.929 --> 00:11:47.690 vibes, it might not beat the absolute top -tier 00:11:47.690 --> 00:11:50.149 Western models yet. You know, models like Nano 00:11:50.149 --> 00:11:52.990 Banana or Seadream. Maybe not yet, no. But it 00:11:52.990 --> 00:11:55.710 proves the concept. It proves you can train a 00:11:55.710 --> 00:11:58.370 massive functional commercial -grade model without 00:11:58.370 --> 00:12:01.340 ever touching American silicon. So if the hardware 00:12:01.340 --> 00:12:03.879 bands were meant to stop them, did they just 00:12:03.879 --> 00:12:05.740 force them to accelerate their own development? 00:12:06.299 --> 00:12:08.419 Necessity is the mother of invention. We force 00:12:08.419 --> 00:12:10.740 them to build their own. And now that stack exists. 00:12:10.980 --> 00:12:12.779 And once it exists, you can't un -invent it. 00:12:12.899 --> 00:12:15.399 The divergence is real now. We basically have 00:12:15.399 --> 00:12:18.460 two separate hardware ecosystems developing in 00:12:18.460 --> 00:12:21.799 parallel. Which complicates, well, everything. 00:12:22.139 --> 00:12:24.100 All right, let's take a quick breath. That was 00:12:24.100 --> 00:12:27.039 a lot of heavy lifting. Science agents, robotic 00:12:27.039 --> 00:12:30.480 supply chains, and semiconductor geopolitics. 00:12:30.559 --> 00:12:33.279 The trifecta of modern anxiety. Let's move to 00:12:33.279 --> 00:12:36.679 our rapid fire segment. Culture, chaos, and what 00:12:36.679 --> 00:12:38.759 we're calling cursed phrases. Oh, I love this 00:12:38.759 --> 00:12:42.860 list. First up, Leo Tolstoy. Poor Tolstoy. Can't 00:12:42.860 --> 00:12:45.759 even rest in peace. People on X, formerly Twitter, 00:12:45.960 --> 00:12:48.899 are accusing Leo Tolstoy of using AI. Because 00:12:48.899 --> 00:12:51.440 of the em dashes. Apparently using em dashes, 00:12:51.440 --> 00:12:53.700 that long dash you use for a break in a sentence, 00:12:53.860 --> 00:12:56.679 is now considered a sign of AI writing. Never 00:12:56.679 --> 00:12:58.399 mind that War and Peace was published in the 00:12:58.399 --> 00:13:01.320 1860s. Right. Or that, you know, he died in 1910. 00:13:01.500 --> 00:13:03.639 The paranoia is real. People are seeing ghosts 00:13:03.639 --> 00:13:05.539 in the machine everywhere. It's becoming a witch 00:13:05.539 --> 00:13:08.559 hunt. If you use good grammar or complex punctuation, 00:13:08.779 --> 00:13:11.000 you must be a bot. Which leads us right to the 00:13:11.000 --> 00:13:13.980 next item, the cursed phrases list. Ah, yes. 00:13:14.679 --> 00:13:17.379 There's a new list of words circulating that 00:13:17.379 --> 00:13:21.860 supposedly out you as an AI. Words that the models 00:13:21.860 --> 00:13:25.039 are statistically more likely to use. I actually 00:13:25.039 --> 00:13:26.860 checked this list before we started recording. 00:13:27.100 --> 00:13:29.480 And did you pass? I have to be honest. Yeah. 00:13:29.580 --> 00:13:32.320 I failed miserably. I used Delve and Tapestry 00:13:32.320 --> 00:13:34.860 all the time. You're a bot, confirmed. I know. 00:13:34.919 --> 00:13:37.809 I felt so called out. But it's a real problem 00:13:37.809 --> 00:13:39.649 for writers now. I actually still wrestle with 00:13:39.649 --> 00:13:41.509 this. I find myself pruning my own vocabulary, 00:13:41.950 --> 00:13:44.789 deleting words I love just because I'm afraid 00:13:44.789 --> 00:13:47.090 someone will think I use chat GPT to write it. 00:13:47.169 --> 00:13:49.730 It's this weird feedback loop of anxiety. It's 00:13:49.730 --> 00:13:52.029 a crisis of authenticity. We're stripping down 00:13:52.029 --> 00:13:53.929 our language just to prove we're human, which 00:13:53.929 --> 00:13:56.690 is so ironic. Item three. We touch on this in 00:13:56.690 --> 00:13:59.090 the intro. Elon versus Sam. The lawsuit that 00:13:59.090 --> 00:14:01.210 keeps on giving. We mentioned the Mars colony 00:14:01.210 --> 00:14:03.009 email, but there was another detail that really 00:14:03.009 --> 00:14:06.129 stuck out to me. Elon Musk. in that correspondence 00:14:06.129 --> 00:14:10.470 called OpenAI 0 % likely to succeed before he 00:14:10.470 --> 00:14:12.750 left. That is such a definitive statement, not 00:14:12.750 --> 00:14:15.509 unlikely, zero. It really just highlights the 00:14:15.509 --> 00:14:17.669 split in vision, doesn't it? You have Altman 00:14:17.669 --> 00:14:21.889 with this expansive, almost hallucinogenic optimism, 00:14:22.210 --> 00:14:27.409 and then Musk with this brutal probability -based 00:14:27.409 --> 00:14:30.210 skepticism. And yet, strangely, they're both 00:14:30.210 --> 00:14:32.629 pushing the frontier harder than anyone else. 00:14:32.889 --> 00:14:36.480 True. Item four. Energy. The elephant in the 00:14:36.480 --> 00:14:38.379 room. The Trump administration is reportedly 00:14:38.379 --> 00:14:42.679 asking tech giants for $15 billion. For power 00:14:42.679 --> 00:14:44.799 plants. Power plants they might never use. They 00:14:44.799 --> 00:14:46.879 want an upfront commitment. And the tech companies 00:14:46.879 --> 00:14:48.460 aren't thrilled. They say they weren't even consulted 00:14:48.460 --> 00:14:50.679 on the price tag. But it just shows the scale 00:14:50.679 --> 00:14:52.259 of the infrastructure problem. We're not talking 00:14:52.259 --> 00:14:53.980 about coding anymore. We're talking about AI 00:14:53.980 --> 00:14:56.539 consuming electricity on a nation state level. 00:14:56.720 --> 00:14:58.539 The constraints aren't software anymore. It's 00:14:58.539 --> 00:15:02.899 physics. It's electrons. And finally, the bear 00:15:02.899 --> 00:15:06.330 case. Michael Burry. The big short guy himself. 00:15:06.710 --> 00:15:09.129 He's quoting Warren Buffett, and he's predicting 00:15:09.129 --> 00:15:12.490 the AI stock boom, specifically NVIDIA, is going 00:15:12.490 --> 00:15:15.350 to end badly. He thinks it's a bubble. He thinks 00:15:15.350 --> 00:15:18.090 the hype has just completely outpaced the reality. 00:15:18.409 --> 00:15:20.750 So is Burry right, or is he just early again? 00:15:20.950 --> 00:15:23.149 He's betting against the momentum of the entire 00:15:23.149 --> 00:15:27.649 world. That's a bold move. Very bold. OK, we're 00:15:27.649 --> 00:15:29.570 back. Let's try to pull all of these threads 00:15:29.570 --> 00:15:31.830 together before we go. We covered a lot of ground 00:15:31.830 --> 00:15:34.649 today. We did. But I think there's a pretty clear 00:15:34.649 --> 00:15:37.289 through line here. I agree. It starts with competence. 00:15:37.570 --> 00:15:39.730 That's Claude. It's not just chatting. It's doing 00:15:39.730 --> 00:15:43.009 real science at MIT and Stanford. It's discovering 00:15:43.009 --> 00:15:45.889 biology that we missed. Then you have embodiment, 00:15:45.970 --> 00:15:49.769 the whole physical layer. That $123 billion stack, 00:15:50.029 --> 00:15:53.830 the fridge, the car, the factory. AI is growing 00:15:53.830 --> 00:15:56.129 legs and entering the supply chain. And then 00:15:56.129 --> 00:15:58.730 divergence. China building a non -U .S. stack 00:15:58.730 --> 00:16:01.690 with Zipu and Huawei. The hardware world is splitting 00:16:01.690 --> 00:16:03.889 in two, and it's proving that bands might just 00:16:03.889 --> 00:16:06.809 accelerate independence. And finally, hype versus 00:16:06.809 --> 00:16:10.450 reality. Mars colonies run by kids versus the 00:16:10.450 --> 00:16:13.149 reality of crumbling power grids and these $15 00:16:13.149 --> 00:16:15.970 billion demands for electricity. It really feels 00:16:15.970 --> 00:16:18.289 like the play phase is over. Yeah, it's not a 00:16:18.289 --> 00:16:20.610 toy anymore. Now the hard work and the hardware 00:16:20.610 --> 00:16:23.929 begins. So here's my question for you, listening 00:16:23.929 --> 00:16:26.529 to this. We talked about research agents and 00:16:26.529 --> 00:16:31.139 physical AI. Are you using these tools to chat 00:16:31.139 --> 00:16:34.960 or are you using them to build workflows? That's 00:16:34.960 --> 00:16:37.500 the shift you really need to make. Because if 00:16:37.500 --> 00:16:40.000 you're just chatting, you might be missing the 00:16:40.000 --> 00:16:41.879 whole point. Oh, and definitely check the show 00:16:41.879 --> 00:16:44.379 notes for that cursed phrases list. Save yourself. 00:16:44.480 --> 00:16:47.039 Don't be a tapestry person. Please don't. I want 00:16:47.039 --> 00:16:48.899 to leave you with one final thought. We spent 00:16:48.899 --> 00:16:50.840 a lot of time on the code and the chips today. 00:16:51.519 --> 00:16:55.039 But think about that $15 billion demand for power 00:16:55.039 --> 00:16:58.149 plants. It's a massive number. The ultimate constraint 00:16:58.149 --> 00:17:00.710 on AI isn't code anymore. It isn't even really 00:17:00.710 --> 00:17:03.230 chips, as China just proved. It's electricity. 00:17:03.750 --> 00:17:05.890 The smartest model in the world is completely 00:17:05.890 --> 00:17:08.150 useless if you can't plug it in. Keep diving. 00:17:08.309 --> 00:17:08.930 See you next time.