WEBVTT 00:00:00.000 --> 00:00:03.259 Apple didn't win the AI innovation race. They 00:00:03.259 --> 00:00:07.040 didn't build the smartest model. But last year, 00:00:07.099 --> 00:00:09.199 they quietly collected nearly a billion dollars 00:00:09.199 --> 00:00:13.660 from it anyway. Beat. Today, we look at the invisible 00:00:13.660 --> 00:00:16.140 toll booths of the future. Yeah, it's a completely 00:00:16.140 --> 00:00:18.920 wild strategy when you actually look at the numbers. 00:00:19.160 --> 00:00:20.980 Welcome to this deep dive. I'm really glad you 00:00:20.980 --> 00:00:23.660 are joining us. Yeah. We are exploring how AI 00:00:23.660 --> 00:00:26.370 is kind of, well... escaping the chat box today 00:00:26.370 --> 00:00:28.649 escaping the chat box and moving into the real 00:00:28.649 --> 00:00:32.369 world from pixels to actual physics exactly and 00:00:32.369 --> 00:00:35.189 our roadmap for today is pretty fascinating we'll 00:00:35.189 --> 00:00:37.310 start with the sheer economics of ai distribution 00:00:37.310 --> 00:00:40.030 right the silent victory apple pulled off yeah 00:00:40.030 --> 00:00:43.289 then we will move into the wild ways people are 00:00:43.289 --> 00:00:45.289 deploying these tools right now from selling 00:00:45.289 --> 00:00:48.149 houses to building these crazy ceo agents putting 00:00:48.149 --> 00:00:50.869 ai in charge of things right and finally we will 00:00:50.869 --> 00:00:52.869 hit The absolute frontier. We're talking about 00:00:52.869 --> 00:00:55.609 world models. This is where AI basically learns 00:00:55.609 --> 00:00:58.270 the physical laws of our universe. It's kind 00:00:58.270 --> 00:01:00.369 of the holy grail of where this tech is heading. 00:01:00.549 --> 00:01:03.170 It really is. So let's start with segment one, 00:01:03.310 --> 00:01:06.109 the invisible tollbooth, Apple's strategy. Which 00:01:06.109 --> 00:01:08.469 is basically no strategy, right? Or at least 00:01:08.469 --> 00:01:10.870 it looks that way on the surface. Right. They 00:01:10.870 --> 00:01:13.209 seem totally behind in the generative AI arms 00:01:13.209 --> 00:01:16.170 race. Yeah. But we have this new data from AppMagic. 00:01:16.170 --> 00:01:17.950 The Wall Street Journal just reported on it. 00:01:18.140 --> 00:01:20.780 The numbers are just staggering. They are. Generative 00:01:20.780 --> 00:01:24.519 AI apps paid Apple around $900 million in app 00:01:24.519 --> 00:01:27.200 store fees for 2025. Almost a billion dollars. 00:01:27.299 --> 00:01:29.500 And they didn't even build the core technology. 00:01:29.780 --> 00:01:33.219 Exactly. Yeah. And about 75 % of that revenue 00:01:33.219 --> 00:01:37.140 came directly from ChatGPT. Wow. Yeah. And just 00:01:37.140 --> 00:01:40.000 for context, XAI's grok was only about 5%. But 00:01:40.000 --> 00:01:41.799 OpenAI is doing all the heavy lifting, paying 00:01:41.799 --> 00:01:44.840 for all the compute. Yep. And Apple's just collecting 00:01:44.840 --> 00:01:47.159 the toll. The fee structure is what makes it 00:01:47.159 --> 00:01:49.480 so lucrative. They take a standard 30 % cut. 00:01:49.659 --> 00:01:52.439 Right. Or 15 % if you're a small business making 00:01:52.439 --> 00:01:55.140 under a million dollars. But for the big players, 00:01:55.299 --> 00:01:58.500 it's 30 % of every single subscription. And the 00:01:58.500 --> 00:02:01.480 growth trajectory here is just crazy. In January, 00:02:01.599 --> 00:02:04.879 it was around $35 million. And then it peaked 00:02:04.879 --> 00:02:08.159 in August at $101 million. In a single month. 00:02:08.340 --> 00:02:11.060 In a single month. It stabilized a bit after 00:02:11.060 --> 00:02:13.770 that, but still. The takeaway here is really 00:02:13.770 --> 00:02:17.909 profound. Apple invests far less in AI infrastructure 00:02:17.909 --> 00:02:21.229 than its rivals. Oh, massively less. But they 00:02:21.229 --> 00:02:23.069 prove that, you know, it's like stacking Lego 00:02:23.069 --> 00:02:25.750 blocks of data. Apple didn't mold the plastic. 00:02:25.810 --> 00:02:28.189 They just own the table everyone is building 00:02:28.189 --> 00:02:30.870 on. That is a perfect analogy. They own the table 00:02:30.870 --> 00:02:33.969 because distribution is just so much more profitable 00:02:33.969 --> 00:02:36.009 than invention right now. It really makes you 00:02:36.009 --> 00:02:38.289 wonder about the long game. Is controlling distribution 00:02:38.289 --> 00:02:41.310 always going to beat raw model intelligence in 00:02:41.310 --> 00:02:44.229 the long run? I mean, I think so. Yes. Because 00:02:44.229 --> 00:02:46.110 intelligence itself is becoming a commodity. 00:02:46.310 --> 00:02:48.349 The models are all kind of converging in quality. 00:02:48.610 --> 00:02:50.669 Right. Everyone has a smart model now. Exactly. 00:02:50.969 --> 00:02:54.689 But access to a user's daily habits, the fact 00:02:54.689 --> 00:02:56.710 that your credit card is already tied to your 00:02:56.710 --> 00:02:59.870 face ID on your iPhone. Yeah. That friction or 00:02:59.870 --> 00:03:02.289 lack of friction is the ultimate bottleneck. 00:03:02.389 --> 00:03:04.729 So owning the user relationship matters more 00:03:04.729 --> 00:03:07.310 than having the smartest math. Precisely. The 00:03:07.310 --> 00:03:09.189 math is just a background feature at this point. 00:03:09.289 --> 00:03:11.610 It's beat. Which brings us to our second segment. 00:03:12.129 --> 00:03:14.469 Because people are using these distributed tools 00:03:14.469 --> 00:03:17.189 in some really disruptive ways. The real -world 00:03:17.189 --> 00:03:19.270 takeovers. Yeah, the real -world takeovers and 00:03:19.270 --> 00:03:21.550 this whole remix culture. Because the tools are 00:03:21.550 --> 00:03:24.250 getting so accessible, the lines are blurring 00:03:24.250 --> 00:03:27.509 between who is building and who is just remixing. 00:03:27.509 --> 00:03:29.530 The cursor is a prime example of this. Right, 00:03:29.530 --> 00:03:32.340 the coding tool. Cursor recently admitted that 00:03:32.340 --> 00:03:34.900 their new Frontier model is basically partly 00:03:34.900 --> 00:03:38.800 built on Moonshot's Kimi 2 .5. Which is a powerful 00:03:38.800 --> 00:03:41.599 Chinese model. Yeah. Plus, you know, some heavy 00:03:41.599 --> 00:03:43.819 training on top of it. But it raises this big 00:03:43.819 --> 00:03:46.620 question. Who is actually building from scratch? 00:03:46.939 --> 00:03:49.219 Almost nobody, right? Everyone is remixing everyone 00:03:49.219 --> 00:03:52.210 else. It's true. I have to admit, I still wrestle 00:03:52.210 --> 00:03:54.810 with prompt drift myself, let alone knowing who 00:03:54.810 --> 00:03:56.870 is actually building versus remixing. Prompt 00:03:56.870 --> 00:03:58.849 drift is the worst. When the AI just kind of 00:03:58.849 --> 00:04:00.830 forgets how you trained it. It draws me crazy. 00:04:01.150 --> 00:04:04.449 But even with those flaws, everyday people are 00:04:04.449 --> 00:04:06.870 doing incredible things. They are completely 00:04:06.870 --> 00:04:10.270 eliminating middlemen. Like that tech CEO. Yes. 00:04:10.810 --> 00:04:14.930 A tech CEO recently used ChatGPT to completely 00:04:14.930 --> 00:04:18.620 bypass realtors. He sold his house for $100 ,000 00:04:18.620 --> 00:04:21.279 over the estimates. Completely on his own. Yeah, 00:04:21.300 --> 00:04:24.310 he just used the AI. For pricing strategies, 00:04:24.529 --> 00:04:26.769 for staging the house, even getting paint ideas. 00:04:27.089 --> 00:04:29.610 He saved tens of thousands in commission. Exactly. 00:04:29.790 --> 00:04:32.629 Yeah. And the AI companies know this value is 00:04:32.629 --> 00:04:35.490 huge, which is why we're seeing a shift in monetization. 00:04:35.810 --> 00:04:38.709 The ads are finally here. They are. OpenAI is 00:04:38.709 --> 00:04:41.649 rolling out clear sponsored ads for their free 00:04:41.649 --> 00:04:43.990 and go tier users in the U .S. It was inevitable, 00:04:44.149 --> 00:04:46.370 really. The compute costs are just too astronomically 00:04:46.370 --> 00:04:49.329 high to stay totally free. Right. But taking 00:04:49.329 --> 00:04:52.009 a step back, the craziest real world takeover 00:04:52.009 --> 00:04:55.139 right now. is over at Meta. Oh, Mark Zuckerberg's 00:04:55.139 --> 00:04:57.639 personal agent. Yeah. Zuck is actively building 00:04:57.639 --> 00:05:00.360 a personal AI agent to help him be CEO. Just 00:05:00.360 --> 00:05:02.839 a digital clone managing his day. But it gets 00:05:02.839 --> 00:05:05.220 deeper. Meta employees are already using personal 00:05:05.220 --> 00:05:07.699 agents internally. And these agents are communicating 00:05:07.699 --> 00:05:10.139 directly with each other. Agent to agent. No 00:05:10.139 --> 00:05:13.420 humans in the loop. None. They are just negotiating 00:05:13.420 --> 00:05:15.579 schedules, sorting out logistics, all in the 00:05:15.579 --> 00:05:18.870 background. But it makes me wonder, when Meta's 00:05:18.870 --> 00:05:20.990 internal employee agents start talking to each 00:05:20.990 --> 00:05:24.850 other, how do we prevent a closed loop of AI 00:05:24.850 --> 00:05:27.110 hallucinations? Well, that's the thing. The human's 00:05:27.110 --> 00:05:29.589 job fundamentally changes. You aren't doing the 00:05:29.589 --> 00:05:31.990 task anymore. You are managing the system of 00:05:31.990 --> 00:05:35.069 agents. We shift from doing the work to babysitting 00:05:35.069 --> 00:05:37.689 the algorithms. Exactly. You are the high -level 00:05:37.689 --> 00:05:39.769 editor, not the writer. Which is a great transition 00:05:39.769 --> 00:05:42.209 to our third segment, stepping outside the screen. 00:05:42.569 --> 00:05:46.189 Because AI is moving beyond just text and software. 00:05:46.370 --> 00:05:48.670 It's bleeding into physical and immersive spaces 00:05:48.670 --> 00:05:50.970 now. Yeah. Look at OpenArt. They just launched 00:05:50.970 --> 00:05:54.050 a tool called Worlds. Oh, I saw this. It's wild. 00:05:54.310 --> 00:05:57.810 It turns a single text prompt into a fully explorable 00:05:57.810 --> 00:06:00.550 3D environment. You can literally walk inside 00:06:00.550 --> 00:06:03.329 the world it generates. From text to a full 3D 00:06:03.329 --> 00:06:05.610 spatial reality in seconds. Right. But it's not 00:06:05.610 --> 00:06:07.829 just virtual reality. It's happening in the literal 00:06:07.829 --> 00:06:10.910 dirt. Yeah. In agriculture. Holter. Yes, Holter. 00:06:11.110 --> 00:06:13.569 They are an AI startup for livestock management. 00:06:13.730 --> 00:06:16.670 And they just hit a $2 billion valuation. Led 00:06:16.670 --> 00:06:19.149 by Founders Fund, which shows you how serious 00:06:19.149 --> 00:06:21.790 the smart money is about this. Absolutely. They 00:06:21.790 --> 00:06:25.569 use AI. to manage physical herds of cows out 00:06:25.569 --> 00:06:27.709 in the real world. Tracking their health, managing 00:06:27.709 --> 00:06:30.509 grazing patterns. It's a massive physical deployment 00:06:30.509 --> 00:06:32.649 of AI. And even in the corporate world, you have 00:06:32.649 --> 00:06:35.990 tools like Masterclass On Call. It actively analyzes 00:06:35.990 --> 00:06:38.389 live meetings and gives you instant leadership 00:06:38.389 --> 00:06:41.069 guidance. So it's observing human behavior in 00:06:41.069 --> 00:06:44.709 real time and coaching you. Exactly. AI is becoming 00:06:44.709 --> 00:06:48.170 an active participant in human reality. But here 00:06:48.170 --> 00:06:51.500 is my concern. Moving from generating text to 00:06:51.500 --> 00:06:54.720 managing a $2 billion farm is a fundamentally 00:06:54.720 --> 00:06:56.899 different type of intelligence. It requires a 00:06:56.899 --> 00:06:58.660 totally different architecture. Are we moving 00:06:58.660 --> 00:07:01.100 too fast by putting AI in charge of physical 00:07:01.100 --> 00:07:03.579 assets like livestock before it truly understands 00:07:03.579 --> 00:07:06.339 the physical world? I mean, yes, that is the 00:07:06.339 --> 00:07:08.199 exact limitation the industry is hitting right 00:07:08.199 --> 00:07:10.300 now. Text models don't know what a fence actually 00:07:10.300 --> 00:07:12.579 is. They just know the word. Right. Text models 00:07:12.579 --> 00:07:15.519 can't truly understand a messy physical farm. 00:07:15.680 --> 00:07:18.560 Exactly. Which is why the biggest players are... 00:07:18.569 --> 00:07:20.949 We're all racing toward the next paradigm. Beat 00:07:20.949 --> 00:07:24.850 the final frontier. Segment four, world models. 00:07:25.069 --> 00:07:27.230 This is the big one. Let's define the jargon 00:07:27.230 --> 00:07:29.990 here really clearly. An AI system trained to 00:07:29.990 --> 00:07:33.529 simulate reality and physical laws. Yes. Not 00:07:33.529 --> 00:07:36.149 just predicting words, predicting physics. Right. 00:07:36.389 --> 00:07:38.649 Captures things like gravity affecting objects, 00:07:38.889 --> 00:07:42.420 human motion. cause and effect over time. Because 00:07:42.420 --> 00:07:45.639 an LLM, a large language model, it doesn't know 00:07:45.639 --> 00:07:48.339 that if you drop a glass on concrete, it shatters. 00:07:48.560 --> 00:07:51.259 It just knows the text associations. Right. But 00:07:51.259 --> 00:07:54.300 embodied intelligence, like a robot in your house, 00:07:54.560 --> 00:07:58.300 needs to deeply understand time, space, and motion. 00:07:58.620 --> 00:08:00.819 Paki McCormick wrote a brilliant piece on this. 00:08:00.939 --> 00:08:03.259 He argues that world models will dramatically 00:08:03.259 --> 00:08:06.680 accelerate AI in everything physical. Robotics, 00:08:06.680 --> 00:08:08.779 logistics, manufacturing, healthcare. Because 00:08:08.779 --> 00:08:10.959 of the simulation loop. Explain that loop for 00:08:10.959 --> 00:08:13.160 a second. So instead of a physical robot trying 00:08:13.160 --> 00:08:15.540 a million times to grasp a cup and dropping it, 00:08:15.639 --> 00:08:18.300 the world model simulates the environment perfectly. 00:08:18.560 --> 00:08:21.060 It tests the action, learns from the mistake, 00:08:21.240 --> 00:08:23.699 improves the parameters, and repeats millions 00:08:23.699 --> 00:08:28.500 of times per second. Whoa. Two sec silence. Imagine 00:08:28.500 --> 00:08:31.519 an AI scaling to a billion queries, simulating 00:08:31.519 --> 00:08:33.940 centuries of robotic movement in an afternoon. 00:08:34.429 --> 00:08:36.950 before ever moving a physical arm. It's mind 00:08:36.950 --> 00:08:39.389 -bending. The robot is essentially fully trained 00:08:39.389 --> 00:08:41.610 before the hardware is even switched on. But 00:08:41.610 --> 00:08:44.409 the real world is messy. What happens when a 00:08:44.409 --> 00:08:47.230 simulated world model encounters a totally unpredictable 00:08:47.230 --> 00:08:50.009 human action in the real world? Well, the simulation 00:08:50.009 --> 00:08:53.629 loop doesn't stop. It takes that anomaly, feeds 00:08:53.629 --> 00:08:56.330 it back into the engine in real time, and adjusts 00:08:56.330 --> 00:08:58.950 its understanding of reality instantly. It treats 00:08:58.950 --> 00:09:00.929 our unpredictable chaos as just another data 00:09:00.929 --> 00:09:03.750 point. Exactly. We are just new variables in 00:09:03.750 --> 00:09:06.149 its physics engine. Beat. Okay, we need to take 00:09:06.149 --> 00:09:08.110 a quick break here, sponsor break. Welcome back. 00:09:08.289 --> 00:09:10.830 Beat. So let's synthesize today's big idea. We 00:09:10.830 --> 00:09:12.509 really went on a journey here. We covered a lot 00:09:12.509 --> 00:09:15.039 of ground. We started with Apple. They're quietly 00:09:15.039 --> 00:09:18.720 raking in $900 million just by owning the distribution 00:09:18.720 --> 00:09:21.600 of AI, the invisible toll booth. They own the 00:09:21.600 --> 00:09:23.840 table everyone is building on. Right. Then we 00:09:23.840 --> 00:09:26.539 looked at how everyday people and tech giants 00:09:26.539 --> 00:09:28.899 are using these distributed tools to completely 00:09:28.899 --> 00:09:31.440 replace middlemen. Real estate agents, assistants, 00:09:31.840 --> 00:09:34.059 we're building networks of personal agents. Yeah. 00:09:34.179 --> 00:09:37.399 But the ultimate takeaway today is this massive 00:09:37.399 --> 00:09:41.470 leap from the screen to the soil. AI is moving 00:09:41.470 --> 00:09:43.889 into the physical world. Building world models 00:09:43.889 --> 00:09:47.029 to understand gravity, mass, and motion. Preparing 00:09:47.029 --> 00:09:49.909 to operate in our actual physical reality. It's 00:09:49.909 --> 00:09:52.370 the most exciting and kind of terrifying shift 00:09:52.370 --> 00:09:53.929 in technology right now. Yeah, it really is. 00:09:53.970 --> 00:09:56.830 So as we wrapped up, I want you to think about 00:09:56.830 --> 00:10:00.539 this. As AI shifts from generating text to testing 00:10:00.539 --> 00:10:03.200 actions in simulated realities, think about your 00:10:03.200 --> 00:10:05.620 own daily workflow. What would you actually hand 00:10:05.620 --> 00:10:08.460 over to it? Exactly. When your personal AI agent 00:10:08.460 --> 00:10:11.659 finally connects to a physical world model, what 00:10:11.659 --> 00:10:13.539 real -world action would you trust it to take 00:10:13.539 --> 00:10:16.279 first? And more importantly, who is responsible 00:10:16.279 --> 00:10:19.019 if it drops the ball? That is the million -dollar 00:10:19.019 --> 00:10:21.740 liability question. It really is. Thank you for 00:10:21.740 --> 00:10:23.600 joining us on this deep dive. We'll see you next 00:10:23.600 --> 00:10:24.899 time. Out T -Row music.