WEBVTT 00:00:00.000 --> 00:00:02.660 Is the current boom in artificial intelligence 00:00:02.660 --> 00:00:06.219 like a financial bubble just waiting to pop, 00:00:06.379 --> 00:00:10.380 wiping out trillions like back in 2001? Or is 00:00:10.380 --> 00:00:11.919 it something really fundamentally different, 00:00:11.960 --> 00:00:15.099 almost like a self -perpetuating infinite money 00:00:15.099 --> 00:00:17.519 glitch? It's kind of wild to think about. Two 00:00:17.519 --> 00:00:20.809 -sec silence, yeah. Today, we're diving deep. 00:00:20.929 --> 00:00:24.609 We've got 12 essential lessons from 28 intense 00:00:24.609 --> 00:00:27.850 months right there in the AI trenches. We're 00:00:27.850 --> 00:00:30.230 going to unpack why the underlying tech is absolutely 00:00:30.230 --> 00:00:32.750 transformative, where the actual money is being 00:00:32.750 --> 00:00:34.990 made right now. And that critical skill you really 00:00:34.990 --> 00:00:38.350 need to survive the coming labor disruption. 00:00:38.469 --> 00:00:41.950 It's serious stuff. Welcome back, everyone. Yeah, 00:00:41.969 --> 00:00:43.689 welcome back to the deep dive. And look, this 00:00:43.689 --> 00:00:45.590 isn't just theory, right? We're looking at patterns 00:00:45.590 --> 00:00:47.710 drawn from people with actual skin in the game. 00:00:47.850 --> 00:00:50.600 Right, moving past those. you know academic debates 00:00:50.600 --> 00:00:52.920 you usually hear exactly okay so let's unpack 00:00:52.920 --> 00:00:55.780 this we've got a roadmap three major areas to 00:00:55.780 --> 00:00:59.799 explore today first the financial reality why 00:00:59.799 --> 00:01:02.640 this feels more like a vc bubble maybe not a 00:01:02.640 --> 00:01:05.299 tech bubble and how nvidia ended up being this 00:01:05.299 --> 00:01:10.129 this central indispensable Okay. Then second, 00:01:10.209 --> 00:01:11.730 we'll shift to the technical change. We're going 00:01:11.730 --> 00:01:15.109 to look at why smaller, more focused models using 00:01:15.109 --> 00:01:17.950 things like reinforcement learning, why they're 00:01:17.950 --> 00:01:20.469 actually taking the lead now over those huge 00:01:20.469 --> 00:01:22.650 generic ones we saw dominate the last couple 00:01:22.650 --> 00:01:26.930 of years. And finally, the human cost. The very 00:01:26.930 --> 00:01:29.769 real job displacement we see coming and this 00:01:29.769 --> 00:01:31.790 kind of counterintuitive idea, the return to 00:01:31.790 --> 00:01:35.150 code. Let's jump in. Where do we start? The money. 00:01:35.290 --> 00:01:36.810 Let's start with the money. The big question. 00:01:37.340 --> 00:01:39.680 bubble or not right the bubble debate it's constant 00:01:39.680 --> 00:01:41.620 you know there's this solid heuristic this rule 00:01:41.620 --> 00:01:44.180 of thumb if everyone is screaming bubble it probably 00:01:44.180 --> 00:01:46.000 isn't going to be that sudden crash they're imagining 00:01:46.000 --> 00:01:48.900 yeah it's different from say 2008 or the crypto 00:01:48.900 --> 00:01:51.420 thing in 2021 where there wasn't as much skepticism 00:01:51.420 --> 00:01:55.719 beforehand exactly but i mean if revenue growth 00:01:55.719 --> 00:01:59.299 is genuinely explosive You mentioned anthropic 00:01:59.299 --> 00:02:01.700 reporting, like 10x year -over -year growth, 00:02:01.920 --> 00:02:04.879 NVIDIA smashing expectations quarter after quarter. 00:02:05.000 --> 00:02:06.739 Why are we even still calling this a bubble? 00:02:06.879 --> 00:02:10.240 Isn't the VC side just scrambling to catch up 00:02:10.240 --> 00:02:12.240 to something totally new? That's the tricky part, 00:02:12.360 --> 00:02:14.479 isn't it? It is. And honestly, the people calling 00:02:14.479 --> 00:02:16.719 bubble, they do have some strong short -term 00:02:16.719 --> 00:02:19.169 points. The investment levels are... Well, they're 00:02:19.169 --> 00:02:23.169 insane. Like 17 million seed rounds. Billions 00:02:23.169 --> 00:02:25.490 for companies with no actual product yet. Yeah. 00:02:25.569 --> 00:02:28.009 Billions for just talent, like getting Andrej 00:02:28.009 --> 00:02:31.169 Karpathy or Elias Asitskiver on board. It's nuts. 00:02:31.430 --> 00:02:33.870 Right. Hard to justify that unless the underlying 00:02:33.870 --> 00:02:36.580 tech was really, really real. Precisely. And 00:02:36.580 --> 00:02:38.719 it is. The core tech is delivering real value 00:02:38.719 --> 00:02:41.419 now. It's writing code, analyzing data, making 00:02:41.419 --> 00:02:43.539 people more productive. That's why it's not a 00:02:43.539 --> 00:02:46.560 technology bubble you see. OK. But the overfunded, 00:02:46.680 --> 00:02:49.360 often unprofitable startups built on top of it. 00:02:50.039 --> 00:02:51.780 Yeah, a lot of those are going to collapse. It's 00:02:51.780 --> 00:02:53.639 more like a VC bubble riding in a tech revolution. 00:02:53.639 --> 00:02:56.300 So like a correction is coming for those weak 00:02:56.300 --> 00:02:59.099 players. Definitely likely. Maybe a 10, even 00:02:59.099 --> 00:03:01.139 30 percent market correction when those companies 00:03:01.139 --> 00:03:05.360 fail. But the core tech. It sticks. Okay. That 00:03:05.360 --> 00:03:07.939 makes sense. And this brings us to that weird 00:03:07.939 --> 00:03:11.080 infinite money glitch thing, the loop between 00:03:11.080 --> 00:03:14.539 NVIDIA and the big AI labs. Ah, yes. Yeah. Strange 00:03:14.539 --> 00:03:17.060 loop. It's where the structural friction really 00:03:17.060 --> 00:03:18.939 lies. It's like the ultimate capital recycling 00:03:18.939 --> 00:03:21.520 machine. How does it work? Exactly. So NVIDIA 00:03:21.520 --> 00:03:24.039 acts like a VC, right? They bore huge capital, 00:03:24.139 --> 00:03:26.699 let's say $100 billion just as an example, into 00:03:26.699 --> 00:03:31.300 a lab like OpenAI. Okay. But OpenAI has to spend 00:03:31.300 --> 00:03:34.439 that money on compute. Which means... NVIDIA's 00:03:34.439 --> 00:03:36.900 GPUs. They're really high margin graphics cards. 00:03:37.180 --> 00:03:39.400 Exactly. The money flows straight back to NVIDIA's 00:03:39.400 --> 00:03:41.460 revenue. Their stock price goes up. Which lets 00:03:41.460 --> 00:03:42.879 them make more investments and the loop just 00:03:42.879 --> 00:03:44.520 keeps going. It's brilliant. A self -feeding 00:03:44.520 --> 00:03:47.759 system. Genius, really. But it exposes this tough 00:03:47.759 --> 00:03:49.800 truth, doesn't it? Pretty much everyone else 00:03:49.800 --> 00:03:53.840 in the AI game, the labs, the startups, they're 00:03:53.840 --> 00:03:56.560 losing money. Burning millions, billions on training 00:03:56.560 --> 00:04:00.879 costs. Absolutely. Jensen Huang and NVIDIA. They're 00:04:00.879 --> 00:04:02.979 selling the picks and shovels in this gold rush. 00:04:03.120 --> 00:04:05.879 They've got the guaranteed massive profit margins. 00:04:06.699 --> 00:04:08.879 Everyone else is just spending. And that must 00:04:08.879 --> 00:04:12.199 be why Google's pushing their TPUs so hard. Amazon's 00:04:12.199 --> 00:04:14.639 building custom silicon. Meta's designing their 00:04:14.639 --> 00:04:17.759 own chips. They have to. It's desperation. They 00:04:17.759 --> 00:04:20.860 need to break this dependency on NVIDIA to control 00:04:20.860 --> 00:04:23.360 their own destiny, frankly, just to survive long 00:04:23.360 --> 00:04:26.120 term. Okay, so if owning the hardware, the compute 00:04:26.120 --> 00:04:29.300 is the main battleground. What's the single biggest 00:04:29.300 --> 00:04:31.819 constraint holding back AI growth and revenue 00:04:31.819 --> 00:04:34.660 right now? If labs could just get more GPUs, 00:04:34.660 --> 00:04:37.259 could they make more money? Compute. It's the 00:04:37.259 --> 00:04:39.839 ultimate physical choke point. It limits how 00:04:39.839 --> 00:04:41.959 fast labs can expand and even meet the demand 00:04:41.959 --> 00:04:44.779 they already have. Labs like Anthropic and OpenAI 00:04:44.779 --> 00:04:48.040 could likely 2x or 3x their revenue today if 00:04:48.040 --> 00:04:50.639 they just have more GPUs. So compute is the constraint. 00:04:50.819 --> 00:04:53.180 Got it. The bottleneck limiting revenue and expansion. 00:04:53.660 --> 00:04:55.579 Mineral sponsor, Reed Placeholder. All right, 00:04:55.600 --> 00:04:57.779 we're back. So if hardware is the choke point, 00:04:57.920 --> 00:04:59.620 how are companies fighting back? It sounds like 00:04:59.620 --> 00:05:01.740 they're changing the definition of what good 00:05:01.740 --> 00:05:04.579 AI even means. That's right. Which brings us 00:05:04.579 --> 00:05:06.180 neatly into the technical shifts we're seeing. 00:05:06.300 --> 00:05:09.379 Yeah, the whole era of just building the biggest 00:05:09.379 --> 00:05:12.300 possible model. That's kind of ending, you know, 00:05:12.300 --> 00:05:14.779 like a lot of analysts found some recent releases 00:05:14.779 --> 00:05:17.360 disappointing. But they were intentionally smaller, 00:05:17.579 --> 00:05:20.120 weren't they? Smaller than previous versions. 00:05:20.279 --> 00:05:22.699 Exactly. It wasn't about ideology. It was driven 00:05:22.699 --> 00:05:25.420 by that cost barrier we talked about. Yeah. Size 00:05:25.420 --> 00:05:27.639 just started giving diminishing returns on performance 00:05:27.639 --> 00:05:30.660 compared to the crazy jump in compute cost. So 00:05:30.660 --> 00:05:33.259 efficiency and speed are the new game. That's 00:05:33.259 --> 00:05:35.459 the real frontier now, efficiency and speed. 00:05:35.740 --> 00:05:38.060 And the breakthrough that's really keeping the 00:05:38.060 --> 00:05:40.699 momentum going is reinforcement learning, or 00:05:40.699 --> 00:05:44.410 RL. OK, define RL for us quickly. Sure. It's 00:05:44.410 --> 00:05:47.170 basically when the AI learns by doing a task, 00:05:47.250 --> 00:05:48.850 like actually trying to write code or play a 00:05:48.850 --> 00:05:51.490 game, and then it gets rewards or penalties based 00:05:51.490 --> 00:05:53.529 on how well it did. So it's moving beyond just 00:05:53.529 --> 00:05:56.009 reading like all the text on the Internet. Right. 00:05:56.110 --> 00:05:58.449 It's learning from action and consequence. And 00:05:58.449 --> 00:06:01.050 what's fascinating is how fast the AIs are getting 00:06:01.050 --> 00:06:04.410 better at coding and math specifically. Why those 00:06:04.410 --> 00:06:08.160 deterministic tasks? tasks with clear right or 00:06:08.160 --> 00:06:10.339 wrong answers? Why are they improving so much 00:06:10.339 --> 00:06:13.040 faster than, say, creative writing? Because the 00:06:13.040 --> 00:06:15.500 feedback is clean. If the code doesn't compile, 00:06:15.699 --> 00:06:18.269 boom, wrong path. If the math equation fails, 00:06:18.430 --> 00:06:21.910 nope, try again. The AI immediately knows to 00:06:21.910 --> 00:06:24.589 discard that way of thinking. Ah, I see. So it 00:06:24.589 --> 00:06:26.990 can generate tons of attempts. Exactly. Like, 00:06:27.050 --> 00:06:30.709 try a solution 500 different ways, find the one 00:06:30.709 --> 00:06:32.870 successful reasoning trace, the one path that 00:06:32.870 --> 00:06:35.709 worked, and then train itself just on that optimal 00:06:35.709 --> 00:06:37.790 path. It's an amazing synthetic data advantage. 00:06:38.230 --> 00:06:40.709 Wow. It's pretty humbling, actually, seeing how 00:06:40.709 --> 00:06:42.829 fast those areas are moving. I still wrestle 00:06:42.829 --> 00:06:44.870 with prompt drift myself. What's prompt drift 00:06:44.870 --> 00:06:47.029 exactly? For listeners who haven't hit that wall 00:06:47.029 --> 00:06:48.550 yet. Oh, it's maddening. It's basically when 00:06:48.550 --> 00:06:51.230 you put in the exact same prompt, but the model 00:06:51.230 --> 00:06:53.329 gives you wildly different answers over time. 00:06:53.569 --> 00:06:56.129 Its internal behavior just changes. Makes reliable 00:06:56.129 --> 00:06:58.410 work kind of impossible, right? Totally. And 00:06:58.410 --> 00:07:01.589 that's much harder to fix for subjective things 00:07:01.589 --> 00:07:03.629 like writing compared to deterministic stuff 00:07:03.629 --> 00:07:07.370 like code. It's a real challenge. Okay, so that 00:07:07.370 --> 00:07:10.779 brings us to this flow zone problem. The focus 00:07:10.779 --> 00:07:14.579 now is on smaller, faster models because speed 00:07:14.579 --> 00:07:17.240 is just critical for actually getting work done, 00:07:17.339 --> 00:07:19.740 right? Like waiting five minutes for a code suggestion 00:07:19.740 --> 00:07:22.199 just breaks your focus. Absolutely breaks your 00:07:22.199 --> 00:07:25.459 flow state. Look at Anthropic's Haiku 4 .5 model. 00:07:25.639 --> 00:07:28.439 People find it super useful, mainly because it's 00:07:28.439 --> 00:07:30.420 three times cheaper and more than twice as fast 00:07:30.420 --> 00:07:32.779 as their bigger Sonnet model. So good enough 00:07:32.779 --> 00:07:36.439 but quick is way better than perfect but slow. 00:07:36.779 --> 00:07:39.519 Vastly superior for most practical uses. Yeah. 00:07:39.920 --> 00:07:42.759 And there's this new strategy reflecting that 00:07:42.759 --> 00:07:45.660 idea, test time compute. Right. So instead of 00:07:45.660 --> 00:07:47.680 blowing the whole compute budget on training 00:07:47.680 --> 00:07:49.500 the absolute biggest model possible. What do 00:07:49.500 --> 00:07:51.620 they do? They train a, let's say, medium sized 00:07:51.620 --> 00:07:54.720 model. Then they save some of that compute budget 00:07:54.720 --> 00:07:57.240 to use during inference, like when you're actually 00:07:57.240 --> 00:07:59.759 using the model. They let the model think longer 00:07:59.759 --> 00:08:02.120 and reason more deeply before it gives you the 00:08:02.120 --> 00:08:03.980 answer. That's a great way to put it, like giving 00:08:03.980 --> 00:08:06.680 a student extra time on the final exam to double 00:08:06.680 --> 00:08:09.139 check the hard questions, even if they only studied. 00:08:09.259 --> 00:08:11.519 a medium amount. Exactly like that. And this 00:08:11.519 --> 00:08:15.079 efficiency, it's unlocking this exponential growth 00:08:15.079 --> 00:08:17.740 in how long AI agents can work autonomously. 00:08:17.839 --> 00:08:20.720 How long are we talking? Well, six months ago, 00:08:20.779 --> 00:08:22.779 maybe 20 minutes was a good run for an autonomous 00:08:22.779 --> 00:08:26.300 agent. Today, we're seeing sustained work sessions 00:08:26.300 --> 00:08:30.199 like two to seven hours long. GPT -5 codecs apparently 00:08:30.199 --> 00:08:33.179 had some seven hour autonomous sessions. Whoa. 00:08:33.960 --> 00:08:36.100 Imagine scaling that. That unlocks really complex 00:08:36.100 --> 00:08:38.480 stuff, right? Like huge code -based refactors 00:08:38.480 --> 00:08:41.240 or deep research projects. Totally. The agent 00:08:41.240 --> 00:08:44.139 can analyze, plan, run searches, synthesize the 00:08:44.139 --> 00:08:46.100 findings, and just repeat that whole loop for 00:08:46.100 --> 00:08:48.600 hours without a human stepping in. If that trajectory 00:08:48.600 --> 00:08:51.240 keeps going, we could have agents working reliably 00:08:51.240 --> 00:08:54.360 for days soon, automating tasks that take junior 00:08:54.360 --> 00:08:56.820 engineers weeks. It's heading that way. It really 00:08:56.820 --> 00:08:59.019 seems possible. This changes the game for complex 00:08:59.019 --> 00:09:01.470 automation. Okay, how does this new strategy 00:09:01.470 --> 00:09:03.710 using test time compute actually improve the 00:09:03.710 --> 00:09:06.590 results if the model itself isn't bigger? Because 00:09:06.590 --> 00:09:08.950 the smaller model gets dedicated time to reason 00:09:08.950 --> 00:09:11.370 during inference right when it's needed. It leads 00:09:11.370 --> 00:09:13.370 to better, more thoughtful answers than just 00:09:13.370 --> 00:09:15.590 a fast, reflexive response from a giant model. 00:09:15.830 --> 00:09:17.690 More thoughtful answers from smaller models. 00:09:17.870 --> 00:09:20.129 Got it. Efficiency is the new size. The technical 00:09:20.129 --> 00:09:22.549 shift is clear. Efficiency over brute force. 00:09:23.129 --> 00:09:26.149 Now let's talk about data. What really separates 00:09:26.149 --> 00:09:29.470 the winners long term? You mentioned XAI earlier. 00:09:29.769 --> 00:09:31.649 Yeah, let's validate that prediction about XAI. 00:09:31.929 --> 00:09:34.309 People kind of dismissed Elon Musk, but he always 00:09:34.309 --> 00:09:36.450 had foundational advantages that meant he was 00:09:36.450 --> 00:09:39.549 going to catch up. Unique data, capital, manufacturing 00:09:39.549 --> 00:09:43.009 skill. Right. XAI gets exclusive access to X, 00:09:43.169 --> 00:09:45.710 formerly Twitter, that real -time global conversation. 00:09:45.789 --> 00:09:49.549 That gives their AI, Grok, this unique edge on 00:09:49.549 --> 00:09:51.950 current events and slang that others just can't 00:09:51.950 --> 00:09:54.750 legally scrape, right? Correct. That's a powerful 00:09:54.750 --> 00:09:57.490 data mode. But the ultimate mode, it's physical. 00:09:57.820 --> 00:10:00.440 It's Tesla and the Optimus robots. How so? Tesla 00:10:00.440 --> 00:10:02.559 cars provide this constant stream of real -world 00:10:02.559 --> 00:10:05.159 video footage, amazing for training world models 00:10:05.159 --> 00:10:07.440 that understand physics and navigation. Yeah. 00:10:07.500 --> 00:10:10.059 But Optimus. Optimus will generate something 00:10:10.059 --> 00:10:13.779 totally unique. Physical embodied data. Three 00:10:13.779 --> 00:10:16.080 -dimensional data from interacting with the world. 00:10:16.419 --> 00:10:18.720 Embodying the data? What does that physical experience 00:10:18.720 --> 00:10:21.759 give an AI that just reading text misses? Everything. 00:10:21.860 --> 00:10:24.970 Seriously. Think about a human child. A toddler 00:10:24.970 --> 00:10:27.950 is infinitely smarter than GPT -5 in some really 00:10:27.950 --> 00:10:30.710 fundamental ways. How? Because the child learns 00:10:30.710 --> 00:10:33.090 from physical feedback that isn't scraped from 00:10:33.090 --> 00:10:35.590 the Internet. They learn about gravity by dropping 00:10:35.590 --> 00:10:38.529 things, force by pushing things, object permanence, 00:10:38.529 --> 00:10:41.649 social cues, all by doing things in the real 00:10:41.649 --> 00:10:44.929 world. Ah, that embodied experience. That's what 00:10:44.929 --> 00:10:47.409 current AIs are fundamentally missing. Totally. 00:10:47.490 --> 00:10:49.450 It's a huge gap. Okay, but here's where things 00:10:49.450 --> 00:10:51.490 get really interesting, maybe counterintuitive 00:10:51.490 --> 00:10:54.519 again. Open source models seem to be catching 00:10:54.519 --> 00:10:57.580 up incredibly fast, even matching or beating 00:10:57.580 --> 00:11:00.000 the big closed systems. Yeah, this is blowing 00:11:00.000 --> 00:11:03.419 up quietly. Take a model like GLM 4 .6 from Zippo 00:11:03.419 --> 00:11:06.379 AI. It's open source and it's performing better 00:11:06.379 --> 00:11:09.139 than Anthropic's closed source Claude Sonnet 00:11:09.139 --> 00:11:11.519 on a lot of important coding benchmarks. So the 00:11:11.519 --> 00:11:13.639 open source community is moving at lightning 00:11:13.639 --> 00:11:16.580 speed. Undeniably fast, yeah. But why isn't this 00:11:16.580 --> 00:11:19.539 like front page news? The incumbents, the big 00:11:19.539 --> 00:11:22.039 labs like OpenAI and Anthropic, they kind of 00:11:22.039 --> 00:11:23.639 ignore it, right? To protect their perceived 00:11:23.639 --> 00:11:25.879 advantage, their moat. But there's also a practical 00:11:25.879 --> 00:11:28.840 issue. What's that? Cloud infrastructure, AWS, 00:11:29.460 --> 00:11:34.159 Google Cloud, Azure. They're optimized to run 00:11:34.159 --> 00:11:37.379 the popular models efficiently. The GPT -series 00:11:37.379 --> 00:11:40.240 cloud. So these powerful new open source models 00:11:40.240 --> 00:11:42.519 seem slower just because the pipes aren't optimized 00:11:42.519 --> 00:11:44.840 for them yet. Exactly. It creates friction. Yeah. 00:11:44.919 --> 00:11:47.299 But the underlying quality of the models proves 00:11:47.299 --> 00:11:49.480 how fast the open source community is innovating. 00:11:49.659 --> 00:11:52.259 You know, Anthropix valued maybe $200 billion, 00:11:52.500 --> 00:11:55.100 but a lab like Zippo is hitting comparable performance. 00:11:55.100 --> 00:11:58.789 It's validation. Wild. OK, now for the more sobering 00:11:58.789 --> 00:12:00.909 part, job displacement. Yeah, got to talk about 00:12:00.909 --> 00:12:02.990 it. Look, the goal here isn't just disrupting 00:12:02.990 --> 00:12:05.570 the, what, $300 billion software market. The 00:12:05.570 --> 00:12:08.029 aim is replacing big chunks of the $15 trillion 00:12:08.029 --> 00:12:11.389 global labor market. $15 trillion, wow. So the 00:12:11.389 --> 00:12:13.370 trend we're likely going to see accelerate by 00:12:13.370 --> 00:12:16.289 2026 is single job replacement agents. Meaning? 00:12:16.509 --> 00:12:18.629 Think customers, court reps, sales development 00:12:18.629 --> 00:12:20.549 reps, setting appointments, executive assistants. 00:12:20.870 --> 00:12:22.889 These roles are prime targets for automation 00:12:22.889 --> 00:12:25.990 by specialized AI agents. Which means huge. efficiency 00:12:25.990 --> 00:12:30.929 gains for companies massive games but also significant 00:12:30.929 --> 00:12:34.730 social unrest i mean widespread protests related 00:12:34.730 --> 00:12:38.090 to ai job losses are genuinely predicted for 00:12:38.090 --> 00:12:41.629 2026. that's heavy which leads us maybe surprisingly 00:12:41.629 --> 00:12:44.929 to your prediction the return to code learning 00:12:44.929 --> 00:12:47.309 computer science is going to be sexy again why 00:12:47.309 --> 00:12:50.850 because it gives you asymmetrical leverage Massive 00:12:50.850 --> 00:12:53.149 leverage. Explain that. Okay. A non -technical 00:12:53.149 --> 00:12:55.669 person using some AI assistance. Maybe they get 00:12:55.669 --> 00:12:59.009 2x, 3x faster at their job. That's nice. Useful. 00:12:59.049 --> 00:13:01.460 Right. But a skilled programmer, someone who 00:13:01.460 --> 00:13:03.720 really understands systems, logic, architecture, 00:13:04.000 --> 00:13:06.500 they don't just get 3x faster. They become 100 00:13:06.500 --> 00:13:09.500 times more powerful. 100 times? How? Because 00:13:09.500 --> 00:13:11.399 you stop being just the coder writing lines. 00:13:11.539 --> 00:13:13.539 You become the architect. You start managing 00:13:13.539 --> 00:13:16.740 teams of AI agents. You leverage your deep technical 00:13:16.740 --> 00:13:19.120 know -how to orchestrate these tools at a scale 00:13:19.120 --> 00:13:21.279 nobody else can. You have to get to the cutting 00:13:21.279 --> 00:13:23.299 edge to really crush the competition. So you 00:13:23.299 --> 00:13:26.080 go from using AI to, like, help write an email. 00:13:26.259 --> 00:13:28.940 To using AI to refactor an entire million lines. 00:13:28.940 --> 00:13:31.840 code base in an afternoon by coordinating five 00:13:31.840 --> 00:13:33.759 different specialized agent teams simultaneously. 00:13:34.220 --> 00:13:37.279 That's the 100x leverage. Wow. So what does this 00:13:37.279 --> 00:13:40.159 all mean for the average career then? If you're 00:13:40.159 --> 00:13:42.259 listening right now, what's the takeaway? The 00:13:42.259 --> 00:13:44.639 greatest advantage, the biggest upside, goes 00:13:44.639 --> 00:13:47.500 to those with the technical competence to truly 00:13:47.500 --> 00:13:50.179 leverage these new AI tools at a deep level. 00:13:50.340 --> 00:13:52.960 Technical competence is key. Okay. Hashtag, hashtag 00:13:52.960 --> 00:13:55.799 occlusion and final thought. All right. Let's 00:13:55.799 --> 00:13:58.799 try to summarize the big ideas here. Three core 00:13:58.799 --> 00:14:02.899 lessons, maybe. First, AI is real transformation, 00:14:03.259 --> 00:14:05.580 not just a bubble, and it's backed by actual 00:14:05.580 --> 00:14:08.820 explosive revenue growth, even if the VC funding 00:14:08.820 --> 00:14:11.379 is bubbly. Right. And second, compute is that 00:14:11.379 --> 00:14:13.460 critical physical constraint, the bottleneck, 00:14:13.460 --> 00:14:16.279 which is fueling that kind of bizarre self -feeding 00:14:16.279 --> 00:14:18.909 money loop between the AI labs and NVIDIA. And 00:14:18.909 --> 00:14:21.190 third, the game has totally changed. That era 00:14:21.190 --> 00:14:23.990 of just copycat startups reselling API tokens 00:14:23.990 --> 00:14:26.669 at a loss, that's ending. It's over. The future 00:14:26.669 --> 00:14:28.950 really belongs to efficiency, specialized models, 00:14:29.070 --> 00:14:31.529 and that deep technical competence we just talked 00:14:31.529 --> 00:14:33.190 about. And it feels like the conversation has 00:14:33.190 --> 00:14:35.450 shifted too, right? Less focus on, you know, 00:14:35.450 --> 00:14:38.450 abstract AI doomerism, the paperclip maximizers. 00:14:38.710 --> 00:14:41.250 Yeah, thankfully. The focus has rightly shifted 00:14:41.250 --> 00:14:44.250 to the very real world problems. Bias in the 00:14:44.250 --> 00:14:47.230 models, market consolidation, and crucially, 00:14:47.269 --> 00:14:50.379 labor disruption. These are tangible issues we 00:14:50.379 --> 00:14:52.799 need to solve now. We've seen just how much capital, 00:14:52.860 --> 00:14:56.259 how much energy is pouring into building the 00:14:56.259 --> 00:14:59.360 physical infrastructure for AI, the data centers, 00:14:59.480 --> 00:15:01.679 the silicon foundries, the power grid upgrades. 00:15:02.000 --> 00:15:04.539 It's staggering. So maybe the final provocative 00:15:04.539 --> 00:15:06.960 thought for you, the listener, is this. It's 00:15:06.960 --> 00:15:09.100 about where you fit. You need to ask yourself, 00:15:09.259 --> 00:15:12.279 are your skills foundational? Are they essential, 00:15:12.440 --> 00:15:14.059 like the picks and shovels needed to build this 00:15:14.059 --> 00:15:17.769 new AI infrastructure? Or... Are your skills 00:15:17.769 --> 00:15:20.350 part of that repetitive lower leverage labor 00:15:20.350 --> 00:15:23.330 that is now, frankly, being systematically replaced? 00:15:23.649 --> 00:15:25.909 A heavy question, but a really necessary one 00:15:25.909 --> 00:15:28.429 to ponder. Where does your unique leverage lie 00:15:28.429 --> 00:15:30.629 in this new world? Yeah. Thank you for taking 00:15:30.629 --> 00:15:32.669 this deep dive with us today. Lots to think about. 00:15:32.750 --> 00:15:33.730 Always is. Talk soon.