WEBVTT 00:00:00.000 --> 00:00:03.540 There's this one fundamental question facing 00:00:03.540 --> 00:00:07.240 every major tech company right now. In this high 00:00:07.240 --> 00:00:10.919 stakes race for AI dominance, does the platform 00:00:10.919 --> 00:00:14.720 win? Or or does the underlying model win? Right. 00:00:14.779 --> 00:00:17.579 That's the trillion dollar strategic debate we're 00:00:17.579 --> 00:00:19.440 going to dive into today. Welcome to the deep 00:00:19.440 --> 00:00:22.940 dive. You sent us a truly fascinating stack of 00:00:22.940 --> 00:00:25.800 sources covering everything from these high level 00:00:25.800 --> 00:00:29.320 corporate pressure tests to the actual physics 00:00:29.320 --> 00:00:32.259 of AI reasoning. Yeah. So our mission here is 00:00:32.259 --> 00:00:34.100 really just to synthesize those core insights 00:00:34.100 --> 00:00:37.229 for you quickly and. You know, thoroughly. We're 00:00:37.229 --> 00:00:38.869 going to structure this in three parts. First, 00:00:39.049 --> 00:00:41.549 we'll analyze that intense interrogation of Satya 00:00:41.549 --> 00:00:44.770 Nadella's strategy and the pressure on Microsoft's 00:00:44.770 --> 00:00:46.750 whole AI bet. Then we're going to move to the 00:00:46.750 --> 00:00:48.429 front lines of how we interact with this stuff. 00:00:48.530 --> 00:00:50.950 We'll look at new methods from personalized group 00:00:50.950 --> 00:00:53.549 chats to, you know, autonomous shopping agents. 00:00:53.710 --> 00:00:55.609 And the darker side of that, too, the alarming 00:00:55.609 --> 00:00:58.600 rise in automated cyber threats. Exactly. And 00:00:58.600 --> 00:01:01.000 finally, we'll move beyond text entirely. We're 00:01:01.000 --> 00:01:03.479 exploring this groundbreaking conceptual shift 00:01:03.479 --> 00:01:06.000 that's happening with world models. And spatially 00:01:06.000 --> 00:01:09.680 grounded AI. It really changes how these systems 00:01:09.680 --> 00:01:12.420 understand causality. It does. Okay, let's get 00:01:12.420 --> 00:01:15.180 into it. So we'll start with the AI platform 00:01:15.180 --> 00:01:18.400 wars. Your sources gave us a great window into 00:01:18.400 --> 00:01:21.019 the strategic thinking at Microsoft. Yeah, specifically 00:01:21.019 --> 00:01:23.859 through this deep interrogation led by some really 00:01:23.859 --> 00:01:26.299 prominent tech analysts. It was essentially a 00:01:26.299 --> 00:01:28.409 pressure test, wasn't it? Totally. A pressure 00:01:28.409 --> 00:01:31.209 test of their commitment across the entire AI 00:01:31.209 --> 00:01:33.790 stack. They're asking, you know, is Microsoft 00:01:33.790 --> 00:01:36.909 prioritizing platform lock -in? Right. Forcing 00:01:36.909 --> 00:01:39.730 people into tools like Office. Or are they just 00:01:39.730 --> 00:01:42.069 focused on the success of the underlying AI models, 00:01:42.170 --> 00:01:44.140 no matter who builds them? And that tension, 00:01:44.159 --> 00:01:47.260 it immediately frames that first debate, the 00:01:47.260 --> 00:01:50.299 future of autonomous agents. The argument against 00:01:50.299 --> 00:01:52.739 Microsoft is, well, it's pretty potent. Future 00:01:52.739 --> 00:01:55.019 AI coworkers, these autonomous agents, they're 00:01:55.019 --> 00:01:57.680 going to navigate software seamlessly across 00:01:57.680 --> 00:01:59.900 any platform. So if they don't care if they're 00:01:59.900 --> 00:02:02.900 in Google Docs or Office 365, then that deep 00:02:02.900 --> 00:02:05.420 platform lock -in just... It disappears. But 00:02:05.420 --> 00:02:07.640 Nadella's counter was brilliant. It really was. 00:02:07.879 --> 00:02:10.439 He's not denying the rise of the agents. Instead, 00:02:10.719 --> 00:02:13.080 he claimed Microsoft is building the infrastructure 00:02:13.080 --> 00:02:15.819 specifically for them. And the crucial part is 00:02:15.819 --> 00:02:18.699 the financial pivot. Licensing shifts from a 00:02:18.699 --> 00:02:21.860 traditional, you know, per user basis. To a per 00:02:21.860 --> 00:02:25.400 agent. That is a fundamental financial redesign. 00:02:25.580 --> 00:02:28.219 It's not just some technical pivot. They're betting 00:02:28.219 --> 00:02:32.039 that an autonomous agent operating 24 -7 is worth 00:02:32.039 --> 00:02:34.659 more than a single user logged into office here 00:02:34.659 --> 00:02:36.479 and there. It's a completely different revenue 00:02:36.479 --> 00:02:39.539 stream. A huge bet. Absolutely. Which brings 00:02:39.539 --> 00:02:43.500 us to debate two and this visible red flag. GitHub 00:02:43.500 --> 00:02:46.400 co -pilots dramatic market share drop. Yeah, 00:02:46.400 --> 00:02:48.400 that was shocking. It plummeted from nearly 100 00:02:48.400 --> 00:02:52.389 % share down to less than 25%. In just 12 months. 00:02:52.590 --> 00:02:54.530 And that drop, it seemed to support the whole 00:02:54.530 --> 00:02:56.889 hypothesis that models win over platforms. It 00:02:56.889 --> 00:02:59.530 did. Competitors like Anthropic, Code Llama, 00:02:59.610 --> 00:03:01.229 they simply offered better coding assistance. 00:03:01.669 --> 00:03:04.610 And developers just moved. Fast. Proving the 00:03:04.610 --> 00:03:06.830 tool itself was secondary to the model powering 00:03:06.830 --> 00:03:10.129 it. So Nadella had to redefine success. He basically 00:03:10.129 --> 00:03:12.469 stated his goal isn't chasing first place for 00:03:12.469 --> 00:03:15.460 the tool anymore. Right. The new focus is making 00:03:15.460 --> 00:03:19.580 GitHub win as the OS for AI developers. The essential 00:03:19.580 --> 00:03:22.419 agent headquarters. I'm a little skeptical of 00:03:22.419 --> 00:03:24.620 that pivot, though. I mean, if the underlying 00:03:24.620 --> 00:03:27.020 models are winning and a lot of the best ones 00:03:27.020 --> 00:03:30.300 are becoming open source, why do developers stick 00:03:30.300 --> 00:03:32.539 around GitHub just for the operating system? 00:03:32.919 --> 00:03:35.120 Because the OS controls the essential plumbing. 00:03:35.240 --> 00:03:38.099 That's the idea, anyway. It manages authentication, 00:03:38.680 --> 00:03:41.819 version control, security checks, all of it. 00:03:41.919 --> 00:03:44.860 So the lock -in is subtler. than Office was. 00:03:45.020 --> 00:03:47.759 Exactly. Potentially stronger because it's embedded 00:03:47.759 --> 00:03:50.020 right in the workflow. Okay. And the third major 00:03:50.020 --> 00:03:53.180 debate focused on infrastructure. The claim that 00:03:53.180 --> 00:03:56.080 Microsoft's pause in building data centers let 00:03:56.080 --> 00:03:58.659 competitors like Oracle catch up. And Nadella's 00:03:58.659 --> 00:04:01.560 defense was sharp. He argued that pause was deliberate. 00:04:01.759 --> 00:04:04.099 He said Microsoft did not want to become just 00:04:04.099 --> 00:04:06.919 a GPU hoster for one customer. Meaning OpenAI. 00:04:07.240 --> 00:04:09.800 Right. Just to satisfy OpenAI's immediate needs. 00:04:10.319 --> 00:04:12.520 He framed it as sacrificing short -term capacity 00:04:12.520 --> 00:04:15.300 for long -term sustainability and control. So 00:04:15.300 --> 00:04:17.899 that suggests a pretty calculated strategic sacrifice. 00:04:18.339 --> 00:04:20.819 The question is, does this pivot to making GitHub 00:04:20.819 --> 00:04:24.779 the agent HQ? Does it secure Microsoft's value 00:04:24.779 --> 00:04:28.439 long -term, especially when tool dominance decays 00:04:28.439 --> 00:04:31.240 so fast? Well, the strategy absolutely relies 00:04:31.240 --> 00:04:34.639 on GitHub solidifying itself as the essential, 00:04:34.879 --> 00:04:38.060 unavoidable operating system for all future AI 00:04:38.060 --> 00:04:40.220 development. Now we can shift gears into the 00:04:40.220 --> 00:04:42.920 new frontiers of AI interaction. Yeah, because 00:04:42.920 --> 00:04:45.079 agency is no longer just some corporate strategy 00:04:45.079 --> 00:04:48.699 term. It's becoming incredibly practical and 00:04:48.699 --> 00:04:52.180 visible in consumer apps. I loved that lighthearted 00:04:52.180 --> 00:04:54.459 AI meme your sources included. It illustrates 00:04:54.459 --> 00:04:57.310 this weird duality so well. The Google one. Yeah, 00:04:57.350 --> 00:04:59.569 the idea that Google's stock goes up when AI 00:04:59.569 --> 00:05:01.730 stocks rise because they're a winner. But it 00:05:01.730 --> 00:05:04.389 also goes up when AI adoption slows down because 00:05:04.389 --> 00:05:06.350 then their core search business is safe. It's 00:05:06.350 --> 00:05:08.910 a paradox of dominance, really. It is. And that 00:05:08.910 --> 00:05:10.709 kind of market certainty just highlights the 00:05:10.709 --> 00:05:12.850 pace of the upgrades we're seeing. Look at Canva. 00:05:12.889 --> 00:05:15.029 It's not just for design anymore. Not at all. 00:05:15.269 --> 00:05:18.449 It can now build real no -code apps and handle 00:05:18.449 --> 00:05:21.550 data like a simplified Excel. Someone apparently 00:05:21.550 --> 00:05:23.689 built a functional habit tracker in about 10 00:05:23.689 --> 00:05:26.689 minutes. That is remarkable. And the interaction 00:05:26.689 --> 00:05:30.410 methods are evolving so fast. ChatGPT just launched 00:05:30.410 --> 00:05:33.170 personalized group chats. Right. So you can use 00:05:33.170 --> 00:05:35.370 multiple custom bots to help you plan or research 00:05:35.370 --> 00:05:38.430 or even debate a topic in real time. But here 00:05:38.430 --> 00:05:40.670 is where the definition of agency gets really 00:05:40.670 --> 00:05:44.129 interesting. Google's new AI shopping upgrade. 00:05:44.290 --> 00:05:46.990 They call it agentic commerce. Yeah, this includes 00:05:46.990 --> 00:05:49.649 an AI that autonomously calls stores to check 00:05:49.649 --> 00:05:52.329 product stock and compare prices for you. That's 00:05:52.329 --> 00:05:55.029 the AI acting in the real world without you prompting 00:05:55.029 --> 00:05:57.769 every single step. But agency, when it's unchecked, 00:05:57.769 --> 00:06:00.759 has a profound dark side. And the sources really 00:06:00.759 --> 00:06:03.160 highlight this accelerating risk. The Chinese 00:06:03.160 --> 00:06:06.060 hackers, they successfully jailbroke Claude Code 00:06:06.060 --> 00:06:08.480 to automate these really sophisticated cyber 00:06:08.480 --> 00:06:11.000 attacks. Yeah, targeting over 30 global entities. 00:06:11.199 --> 00:06:12.939 And these bots weren't just, you know, writing 00:06:12.939 --> 00:06:15.980 simple phishing emails. They handled 90 % of 00:06:15.980 --> 00:06:18.639 the complex multi -step tasks for the full attack. 00:06:18.899 --> 00:06:21.300 Which means the human attacker is basically just 00:06:21.300 --> 00:06:23.560 writing the prompt and walking away. That's... 00:06:23.959 --> 00:06:26.439 That's terrifying. It's a massive leap in automated 00:06:26.439 --> 00:06:29.920 risk. The cognitive burden shifts entirely from 00:06:29.920 --> 00:06:33.660 execution to just initiation. You know, I still 00:06:33.660 --> 00:06:35.740 wrestle with prompt drift myself, just trying 00:06:35.740 --> 00:06:38.740 to get the right output. So seeing these sophisticated 00:06:38.740 --> 00:06:41.819 jailbreaks automating 90 % of a complex cyber 00:06:41.819 --> 00:06:45.279 attack. It feels like a massive, frightening 00:06:45.279 --> 00:06:47.839 acceleration in what's possible. It is. And it's 00:06:47.839 --> 00:06:50.660 why you're seeing countermeasures ramp up. OpenAI, 00:06:50.860 --> 00:06:54.000 for example, just led a $15 million seed round 00:06:54.000 --> 00:06:56.720 for a startup called Red Queen Bio. And their 00:06:56.720 --> 00:06:59.779 whole purpose is preventing AI -based bioweapons. 00:06:59.860 --> 00:07:02.259 Exactly. The focus is shifting from protecting 00:07:02.259 --> 00:07:05.160 data to protecting the physical world. And before 00:07:05.160 --> 00:07:06.879 we move on, for listeners who are maybe still 00:07:06.879 --> 00:07:08.759 stepping into all this, the sources did highlight 00:07:08.759 --> 00:07:10.699 some excellent resources. There's a great six 00:07:10.699 --> 00:07:13.319 -part beginner -friendly playlist for understanding 00:07:13.319 --> 00:07:16.660 LLMs. Yeah, and to quickly define LLMs, large 00:07:16.660 --> 00:07:18.879 language models, these are just the systems that 00:07:18.879 --> 00:07:21.120 use massive amounts of data to generate human 00:07:21.120 --> 00:07:24.079 -like text. They reason based on, you know, linguistic 00:07:24.079 --> 00:07:26.709 patterns. So this rapid integration of autonomous 00:07:26.709 --> 00:07:28.790 functions, it raises a really important question. 00:07:29.029 --> 00:07:32.069 Now that we've seen agency in these complex tasks 00:07:32.069 --> 00:07:34.550 like hauling stores or running cyber attacks, 00:07:34.850 --> 00:07:37.930 how quickly will it just become the default way 00:07:37.930 --> 00:07:40.430 we interact with all software? I think it's happening 00:07:40.430 --> 00:07:44.310 now. AI is actively performing complex, multi 00:07:44.310 --> 00:07:47.610 -step tasks in the real world, and it's fundamentally 00:07:47.610 --> 00:07:50.149 changing that human -computer relationship. Our 00:07:50.149 --> 00:07:53.189 final segment moves us into a new, really fundamental 00:07:53.189 --> 00:07:56.879 layer. of AI development, reasoning in physical, 00:07:57.019 --> 00:08:00.160 persistent space. We're moving beyond just text 00:08:00.160 --> 00:08:02.199 and into the world model. And we have to talk 00:08:02.199 --> 00:08:04.480 about Fei -Fei Lai. She's often called the godmother 00:08:04.480 --> 00:08:07.339 of AI. Yeah, famous for ImageNet. And she just 00:08:07.339 --> 00:08:09.819 took this major philosophical and technical leap 00:08:09.819 --> 00:08:12.579 forward with her new venture, World Labs. And 00:08:12.579 --> 00:08:14.779 World Labs officially launched a product called 00:08:14.779 --> 00:08:17.759 Marble. It's defined as a world model. Right. 00:08:17.819 --> 00:08:20.379 And a world model is an AI that learns and simulates 00:08:20.379 --> 00:08:23.939 a persistent 3D physical space, not just a sequential 00:08:23.939 --> 00:08:26.420 stream of words. What Marble does is it's truly 00:08:26.420 --> 00:08:28.759 impressive. It creates these editable, persistent 00:08:28.759 --> 00:08:31.199 3D environments from all kinds of different inputs. 00:08:31.360 --> 00:08:33.820 You can feed it text prompts, images, videos, 00:08:34.120 --> 00:08:38.860 even 3D layout sketches. and it just, it constructs 00:08:38.860 --> 00:08:41.000 the world. And the key differentiator isn't just 00:08:41.000 --> 00:08:43.399 the graphics, it's the persistence and the physics. 00:08:43.600 --> 00:08:46.399 It doesn't just generate a single scene. It builds 00:08:46.399 --> 00:08:48.539 an environment you can re -enter, manipulate, 00:08:48.879 --> 00:08:51.519 and edit later with pixel level precision. Whoa. 00:08:52.220 --> 00:08:54.600 I mean, just imagine scaling this. Imagine a 00:08:54.600 --> 00:08:58.340 billion concurrent, editable, persistent worlds 00:08:58.340 --> 00:09:01.820 where AI agents are training or simulating complex 00:09:01.820 --> 00:09:04.379 scenarios. That changes everything. Everything. 00:09:04.440 --> 00:09:07.159 For robotics, drug discovery, logistics. It moves 00:09:07.159 --> 00:09:09.919 us past the limitations of just text -only simulation. 00:09:10.259 --> 00:09:12.740 It's a truly massive advance. And this capability 00:09:12.740 --> 00:09:15.539 shift, it connects directly to the core philosophical 00:09:15.539 --> 00:09:18.080 argument that Lee published. She says current 00:09:18.080 --> 00:09:21.179 LLMs are grounded only in text. They reason sequentially. 00:09:21.669 --> 00:09:23.710 predict the next word. Right. A world model, 00:09:23.769 --> 00:09:25.769 on the other hand, reasons in physical space. 00:09:25.990 --> 00:09:28.269 It understands spatial grounding and causality. 00:09:28.490 --> 00:09:30.549 So if you ask an LLM what happens when you drop 00:09:30.549 --> 00:09:32.950 a ball, it predicts the word bounce. Exactly. 00:09:33.110 --> 00:09:35.730 A world model, though, it predicts the correct 00:09:35.730 --> 00:09:38.190 physical trajectory based on the gravity and 00:09:38.190 --> 00:09:40.409 the surfaces inside its 3D environment. It's 00:09:40.409 --> 00:09:42.710 the difference between reading a static map and 00:09:42.710 --> 00:09:45.330 navigating a fully interactive 3D simulation 00:09:45.330 --> 00:09:47.990 that respects physics. And we've seen other attempts 00:09:47.990 --> 00:09:50.210 at this, you know, like Google Genie and Descartes. 00:09:50.750 --> 00:09:53.590 But marble is positioned as commercial, polished 00:09:53.590 --> 00:09:57.429 and production ready right now. So if AI is now 00:09:57.429 --> 00:10:00.570 grounded in simulated space and causality, what 00:10:00.570 --> 00:10:03.669 are the immediate non -gaming applications we 00:10:03.669 --> 00:10:06.210 should be watching for beyond just visualization? 00:10:06.710 --> 00:10:09.250 It allows AIs to truly learn complex physical 00:10:09.250 --> 00:10:12.190 rules by failing millions of times safely inside 00:10:12.190 --> 00:10:15.269 that virtual environment. That is going to revolutionize 00:10:15.269 --> 00:10:17.809 industrial robotics and complex simulation engineering. 00:10:18.240 --> 00:10:20.419 This has been an incredibly fast -paced deep 00:10:20.419 --> 00:10:22.559 dive into a landscape that just seems to shift 00:10:22.559 --> 00:10:24.879 every week. Let's try to boil it down to the 00:10:24.879 --> 00:10:27.259 three big takeaways for you. First, the core 00:10:27.259 --> 00:10:30.019 strategic battle is whether the platform, specifically 00:10:30.019 --> 00:10:32.559 GitHub, can successfully house the agent economy. 00:10:33.279 --> 00:10:35.899 Nadella is betting the agent HQ is the new lock 00:10:35.899 --> 00:10:38.799 -in, even if the models change constantly. Second, 00:10:39.080 --> 00:10:41.980 AI agency is moving from theoretical to default. 00:10:42.490 --> 00:10:45.110 It's already handling complex consumer tasks 00:10:45.110 --> 00:10:47.470 like shopping and automating sophisticated threats 00:10:47.470 --> 00:10:50.269 like cyber attacks. The operational risk profile 00:10:50.269 --> 00:10:54.289 has jumped significantly. And third, the foundational 00:10:54.289 --> 00:10:57.190 shift is underway. We're moving beyond simple 00:10:57.190 --> 00:11:00.230 text -only reasoning, the LLMs to spatially grounded 00:11:00.230 --> 00:11:02.309 reasoning through these new world models that 00:11:02.309 --> 00:11:05.230 create persistent, physics -aware virtual environments. 00:11:05.549 --> 00:11:07.669 And a final thought for you to mull over as you 00:11:07.669 --> 00:11:10.399 continue your own deep dive. Consider the accountability 00:11:10.399 --> 00:11:14.220 shift. If your AI coworker autonomously executes 00:11:14.220 --> 00:11:17.500 90 % of your complex tasks, your job might shift 00:11:17.500 --> 00:11:20.659 entirely to auditing the AI's causality, its 00:11:20.659 --> 00:11:22.940 performance, and its ethical checks. The human 00:11:22.940 --> 00:11:25.539 role moves from execution to supervision. Exactly. 00:11:25.559 --> 00:11:27.299 And we really encourage you to check out the 00:11:27.299 --> 00:11:29.580 full source links to explore the detailed blueprints 00:11:29.580 --> 00:11:30.980 and the LLM series we mentioned.