WEBVTT 00:00:00.000 --> 00:00:02.540 AI is no longer just text on a screen. Beat. 00:00:02.980 --> 00:00:06.019 It is demanding massive power grids. It is causing 00:00:06.019 --> 00:00:09.900 global geopolitical friction. And now, beat. 00:00:10.339 --> 00:00:13.939 It is literally growing hands and legs to walk 00:00:13.939 --> 00:00:16.539 around the physical world. Yeah, it is wild to 00:00:16.539 --> 00:00:18.219 see. I mean, we're watching code crash straight 00:00:18.219 --> 00:00:21.960 into reality here. We are. Two sec silence. Welcome 00:00:21.960 --> 00:00:24.019 to today's Deep Dive. I am glad you are here 00:00:24.019 --> 00:00:26.199 with us. We're taking a stack of fresh newsletter 00:00:26.199 --> 00:00:28.820 sources to explore how AI is breaking out of 00:00:28.820 --> 00:00:30.940 the software layer and crashing into physical 00:00:30.940 --> 00:00:32.799 reality. Right. And we've got a lot to cover 00:00:32.799 --> 00:00:35.420 today. We're looking at Anthropix's massive new 00:00:35.420 --> 00:00:38.219 climate commitments, for one. And then there's 00:00:38.219 --> 00:00:41.820 this intense global political fight over an American 00:00:41.820 --> 00:00:45.119 AI off switch. Plus, Alibaba's new breakthrough 00:00:45.119 --> 00:00:48.890 in giving AI physical bodies. Before AI can walk, 00:00:49.070 --> 00:00:51.990 it has to eat. Beat. And right now it is eating 00:00:51.990 --> 00:00:53.950 the power grid. Oh, absolutely. I mean, training 00:00:53.950 --> 00:00:56.850 these frontier models requires just immense uninterrupted 00:00:56.850 --> 00:00:59.289 electricity. Which brings us to Anthropic. They 00:00:59.289 --> 00:01:01.710 actually just became the first AI startup to 00:01:01.710 --> 00:01:03.929 join the Frontier Carbon Coalition. I want to 00:01:03.929 --> 00:01:05.890 break down what this coalition actually does. 00:01:06.069 --> 00:01:08.250 Beat. Because it was originally founded by major 00:01:08.250 --> 00:01:10.730 tech players. Companies like Stripe, Google, 00:01:10.849 --> 00:01:13.569 and Shopify. Yeah, exactly. And it basically 00:01:13.569 --> 00:01:16.670 acts as a massive vetting machine for... Carbon 00:01:16.670 --> 00:01:19.269 removal credits. And we're not talking about 00:01:19.269 --> 00:01:22.370 like planting a few trees here. Anthropic is 00:01:22.370 --> 00:01:26.689 contributing to a new $915 million funding tranche. 00:01:26.730 --> 00:01:30.120 Wow. That is a significant amount of capital. 00:01:30.280 --> 00:01:32.980 It really is. That fresh money nearly doubles 00:01:32.980 --> 00:01:37.120 Frontier's total pledges to $1 .8 billion. And 00:01:37.120 --> 00:01:39.900 to date, the coalition has contracted about $700 00:01:39.900 --> 00:01:43.400 million to remove 1 .8 million tons of carbon 00:01:43.400 --> 00:01:45.840 across 50 different projects. That feels like 00:01:45.840 --> 00:01:48.799 a major philosophical shift for Anthropic. They 00:01:48.799 --> 00:01:50.780 previously leaned on an all -of -the -above energy 00:01:50.780 --> 00:01:53.620 approach, which naturally implies utilizing some 00:01:53.620 --> 00:01:56.120 polluting power sources to get that constant 00:01:56.120 --> 00:01:58.200 baseload power. Right. Because data centers, 00:01:58.319 --> 00:01:59.819 they just kind of can't wait for the wind to 00:01:59.819 --> 00:02:03.340 blow. They need 247 power. So this is Anthropic's 00:02:03.340 --> 00:02:05.739 first major climate deal. And Frontier itself 00:02:05.739 --> 00:02:08.259 is fundamentally shifting its strategy right 00:02:08.259 --> 00:02:10.639 now. They are moving away from funding smaller 00:02:10.639 --> 00:02:13.080 experiments. Yeah, exactly. They are moving away 00:02:13.080 --> 00:02:15.219 from dozens of small projects. They want fewer 00:02:15.219 --> 00:02:18.219 massive projects now. The explicit goal is removing 00:02:18.219 --> 00:02:20.919 a gigaton of CO2 annually on these 8 - to 10 00:02:20.919 --> 00:02:24.710 -year contracts. A gigaton. Beat. That is one 00:02:24.710 --> 00:02:28.210 billion metric tons. It is like shifting from 00:02:28.210 --> 00:02:30.469 planting a few trees in the backyard to building 00:02:30.469 --> 00:02:33.590 a global industrial scale carbon recycling plant. 00:02:33.810 --> 00:02:36.789 That is a great analogy. You need direct air 00:02:36.789 --> 00:02:39.650 capture facilities like giant industrial fans 00:02:39.650 --> 00:02:41.569 sucking in the atmosphere, chemically binding 00:02:41.569 --> 00:02:44.610 the CO2 and burying it underground. But to reach 00:02:44.610 --> 00:02:46.669 that gigaton scale, they're trying to guarantee 00:02:46.669 --> 00:02:48.650 a market exists so these massive factories can 00:02:48.650 --> 00:02:51.150 actually get built. Hold on. Beat. There is a 00:02:51.150 --> 00:02:53.340 strict new rule here that gives me pause. Frontier 00:02:53.340 --> 00:02:55.180 requires these projects to show a clear path 00:02:55.180 --> 00:02:58.180 to government subsidy by 2040. Why the strict 00:02:58.180 --> 00:03:00.219 requirement to hand the ultimate bill to the 00:03:00.219 --> 00:03:02.060 public sector? Well, that's the critical tension, 00:03:02.139 --> 00:03:04.259 right? Private tech money can really only kickstart 00:03:04.259 --> 00:03:06.900 the market. Removing a gigaton of carbon annually 00:03:06.900 --> 00:03:09.620 requires infrastructure on the scale of, like, 00:03:09.680 --> 00:03:12.439 the global highway system. Tech companies can 00:03:12.439 --> 00:03:15.659 de -risk the early technology, sure. But maintaining 00:03:15.659 --> 00:03:18.919 utility -scale public works over decades. That 00:03:18.919 --> 00:03:20.960 structurally requires government -level funding. 00:03:21.360 --> 00:03:24.520 So tech funds the start, but taxpayers must eventually 00:03:24.520 --> 00:03:27.120 take the wheel. That is the exact gamble they 00:03:27.120 --> 00:03:29.599 are making. It's a multi -decade baton pass. 00:03:29.860 --> 00:03:31.939 If governments don't step in, the infrastructure 00:03:31.939 --> 00:03:34.039 just eventually collapses under its own operating 00:03:34.039 --> 00:03:36.479 costs. But building these gigaton -scale power 00:03:36.479 --> 00:03:40.099 grids introduces a new vulnerability. That massive 00:03:40.099 --> 00:03:43.039 infrastructure is not just about energy. It is 00:03:43.039 --> 00:03:46.020 driving billions of dollars and creating incredibly 00:03:46.020 --> 00:03:49.020 high stakes global politics. The geopolitical 00:03:49.020 --> 00:03:51.159 panic is so palpable right now. I mean, look 00:03:51.159 --> 00:03:54.419 at the sudden Fable 5 ban overseas. A new letter 00:03:54.419 --> 00:03:56.780 from the U .S. Commerce Secretary to Dario Amadei 00:03:56.780 --> 00:03:59.759 just surfaced, giving hidden context on why access 00:03:59.759 --> 00:04:02.139 to those specific frontier models was frozen 00:04:02.139 --> 00:04:05.000 so suddenly for international users. And we are 00:04:05.000 --> 00:04:07.639 seeing immediate pushback on the global stage. 00:04:09.150 --> 00:04:12.770 At the G7 summit, Emmanuel Macron and Narendra 00:04:12.770 --> 00:04:15.189 Modi explicitly pushed back against these U .S. 00:04:15.210 --> 00:04:17.649 restrictions, the same restrictions that froze 00:04:17.649 --> 00:04:20.110 Claude Fabel and mythos. Right. World leaders 00:04:20.110 --> 00:04:22.990 are waking up to this harsh reality. They desperately 00:04:22.990 --> 00:04:25.889 want American A .I. capabilities to drive their 00:04:25.889 --> 00:04:28.750 economies, but they explicitly do not want an 00:04:28.750 --> 00:04:31.029 American off switch. They're looking at their 00:04:31.029 --> 00:04:33.649 critical infrastructure like banking, health 00:04:33.649 --> 00:04:36.269 care, logistics, running on intelligence hosted 00:04:36.269 --> 00:04:38.629 in data centers controlled by Washington. We 00:04:38.629 --> 00:04:40.769 should look at this impartially. The QBES clearly 00:04:40.769 --> 00:04:43.410 has profound national security concerns regarding 00:04:43.410 --> 00:04:46.610 these frontier models. Conversely, the G7 partners 00:04:46.610 --> 00:04:49.230 have entirely valid national sovereignty concerns. 00:04:49.389 --> 00:04:52.449 They fear sudden economic blackouts. Both sides 00:04:52.449 --> 00:04:54.370 are operating on fundamentally different yet 00:04:54.370 --> 00:04:57.259 entirely rational priorities. Two sec silence. 00:04:57.579 --> 00:04:59.779 Absolutely. And that geopolitical bottleneck 00:04:59.779 --> 00:05:02.319 is, well, it's forcing a massive hardware pivot. 00:05:02.439 --> 00:05:05.319 If nations can't rely on cloud APIs, they need 00:05:05.319 --> 00:05:07.100 local compute. I mean, look at China's response. 00:05:07.300 --> 00:05:10.379 DeepSeek just raised $7 .4 billion at a $50 billion 00:05:10.379 --> 00:05:13.180 valuation. Making them China's most valuable 00:05:13.180 --> 00:05:16.839 AI startup. Beat. And the founder backed it with 00:05:16.839 --> 00:05:19.980 $3 billion of his own money. Yeah. Liying Wenfeng. 00:05:20.300 --> 00:05:22.639 It's immense conviction to secure independent 00:05:22.639 --> 00:05:25.980 compute power. This bottleneck isn't just global. 00:05:26.120 --> 00:05:28.220 It's hitting everyday consumer devices right 00:05:28.220 --> 00:05:31.500 now. Apple is facing what industry insiders are 00:05:31.500 --> 00:05:34.540 calling Ramageddon. Ramageddon, Pete. That is 00:05:34.540 --> 00:05:36.959 a wonderfully vivid term. It is less wonderful 00:05:36.959 --> 00:05:39.980 for consumers, though. Apple indicates a massive 00:05:39.980 --> 00:05:42.680 memory price increase is essentially unavoidable. 00:05:42.779 --> 00:05:46.180 It could force a $270 price hike across future 00:05:46.180 --> 00:05:49.319 iPhones, Macs, and iPads. That is a staggering 00:05:49.319 --> 00:05:52.529 jump for consumer hardware. Why is Apple suddenly 00:05:52.529 --> 00:05:54.589 facing Remageddon right now? It comes down to 00:05:54.589 --> 00:05:57.250 how a neural network physically operates. Traditional 00:05:57.250 --> 00:05:59.449 apps, they just sit memory. You load a photo 00:05:59.449 --> 00:06:01.990 and you close it. But running a frontier AI model 00:06:01.990 --> 00:06:04.069 locally means storing billions of parameters. 00:06:04.470 --> 00:06:06.649 The entire neural network basically has to sit 00:06:06.649 --> 00:06:09.329 in active, high bandwidth memory simultaneously 00:06:09.329 --> 00:06:12.709 to generate a single word. Bigger AI models simply 00:06:12.709 --> 00:06:15.069 demand more memory to run on our personal devices. 00:06:15.470 --> 00:06:18.310 Exactly. You can't just page AI weights to a 00:06:18.310 --> 00:06:21.139 slow hard drive. But while hardware gets expensive, 00:06:21.439 --> 00:06:24.079 the everyday integration of these models is accelerating 00:06:24.079 --> 00:06:27.060 fast. Have you seen the viral index from artificial 00:06:27.060 --> 00:06:30.620 analysis? I have. BEAT. It maps top models by 00:06:30.620 --> 00:06:33.180 capability versus their actual running cost. 00:06:33.379 --> 00:06:35.819 Yeah, and it proves a vital point. Not all frontier 00:06:35.819 --> 00:06:38.939 models are expensive. And not all expensive models 00:06:38.939 --> 00:06:41.180 are necessarily the best. The deployment is getting 00:06:41.180 --> 00:06:44.959 incredibly smooth. Beat. Anthropic's Applied 00:06:44.959 --> 00:06:47.839 AI team showed how cloud -managed agents are 00:06:47.839 --> 00:06:49.519 now running in production. Right. They are no 00:06:49.519 --> 00:06:51.579 longer theoretical. Yeah. And the competition 00:06:51.579 --> 00:06:54.939 is driving massive value. Copilot Cowork is now 00:06:54.939 --> 00:06:56.759 generally available. Half of the Fortune 500 00:06:56.759 --> 00:06:59.420 already use it. And it claims to be 30 % to 40 00:06:59.420 --> 00:07:02.759 % cheaper than Cloud Cowork. Beat. It is embedding 00:07:02.759 --> 00:07:05.699 into everything. Grok is now a free add -in for 00:07:05.699 --> 00:07:08.079 PowerPoint. Yeah, it takes a rough outline and 00:07:08.079 --> 00:07:10.519 just generates a research deck. And Google just 00:07:10.519 --> 00:07:13.279 launched a new $99 home speaker that handles 00:07:13.279 --> 00:07:15.500 completely messy commands. Like, you can stutter, 00:07:15.579 --> 00:07:17.439 change your mind mid -sentence, and it just follows 00:07:17.439 --> 00:07:20.180 the chaotic human logic perfectly. I have a vulnerable 00:07:20.180 --> 00:07:23.339 admission to make here. Beat. I still wrestle 00:07:23.339 --> 00:07:25.959 with prompt drift myself. Seeing how smoothly 00:07:25.959 --> 00:07:29.439 these new tools handle messy, real -world human 00:07:29.439 --> 00:07:32.519 workflows is inspiring. Oh, me too. It really 00:07:32.519 --> 00:07:34.959 shows we're moving past the era of rigid prompt 00:07:34.959 --> 00:07:36.899 engineering. We aren't learning to speak to the 00:07:36.899 --> 00:07:39.199 machine anymore. The machine is finally learning 00:07:39.199 --> 00:07:41.819 to speak to us. But handling these massive local 00:07:41.819 --> 00:07:45.019 models requires totally new infrastructure. Beat. 00:07:45.620 --> 00:07:47.420 Developers are building a totally new software 00:07:47.420 --> 00:07:50.199 layer purely for AI agents. Yes, the tooling 00:07:50.199 --> 00:07:52.959 ecosystem is just exploding right now. Look at 00:07:52.959 --> 00:07:55.949 the new Minimax M3. It's an open -weight model 00:07:55.949 --> 00:07:58.569 that combines frontier -level coding, a 1 million 00:07:58.569 --> 00:08:01.470 token context window, and multimodal support. 00:08:01.790 --> 00:08:03.670 For the listener who isn't a developer, beat, 00:08:03.889 --> 00:08:05.810 what does an open -weight model actually mean? 00:08:06.370 --> 00:08:09.209 Think of a closed model like a baked cake. You 00:08:09.209 --> 00:08:11.370 can eat it, but you can't unbake it to change 00:08:11.370 --> 00:08:13.850 the recipe. An open -weight model gives you the 00:08:13.850 --> 00:08:17.110 exact ratio of flour, sugar, and eggs. It gives 00:08:17.110 --> 00:08:18.810 you the neural network's parameters so you can 00:08:18.810 --> 00:08:21.089 tweak the recipe for your own needs. Anyone can 00:08:21.089 --> 00:08:23.230 download and modify the core model architecture 00:08:23.230 --> 00:08:26.610 for their own specific needs. Perfectly. And 00:08:26.610 --> 00:08:28.730 that customizability is empowering developers 00:08:28.730 --> 00:08:32.730 to build wild new platforms. Take Cursor's Origin. 00:08:32.889 --> 00:08:35.549 It is a Git hosting platform built specifically 00:08:35.549 --> 00:08:39.029 for AI agents. Code management designed for bots. 00:08:39.289 --> 00:08:42.299 Beat. The documentation says it allows automated 00:08:42.299 --> 00:08:44.740 conflict resolution. How does that actually work 00:08:44.740 --> 00:08:47.860 without a human? Well, if Agent A changes the 00:08:47.860 --> 00:08:50.220 database schema and Agent B builds a new website 00:08:50.220 --> 00:08:53.200 button, Origin understands the purpose of both. 00:08:53.379 --> 00:08:56.220 It automatically rewrites the button's code to 00:08:56.220 --> 00:08:58.460 match the new database schema without a human 00:08:58.460 --> 00:09:01.360 ever mediating. It's true agent -to -agent collaboration. 00:09:02.159 --> 00:09:04.799 That is fascinating. Beat. Then there's Hermes 00:09:04.799 --> 00:09:06.740 Agent. It adds official Stripe skills directly 00:09:06.740 --> 00:09:08.620 into the model. Yeah, to let agents make real 00:09:08.620 --> 00:09:10.019 financial payments, they can launch their own 00:09:10.019 --> 00:09:12.740 SaaS product safely. And look at Crilly. It's 00:09:12.740 --> 00:09:15.039 finding creators based on real -time habits and 00:09:15.039 --> 00:09:17.220 moving conversations into structured deal workflows 00:09:17.220 --> 00:09:20.179 automatically. Giving agents the ability to negotiate 00:09:20.179 --> 00:09:23.379 legal deals or swipe a credit card. Beat. It 00:09:23.379 --> 00:09:25.580 feels like a massive threshold crossed. It definitely 00:09:25.580 --> 00:09:27.779 raises the stakes. I mean, we are building an 00:09:27.779 --> 00:09:30.600 autonomous economic layer. So agents finally 00:09:30.600 --> 00:09:33.139 understand the purpose of our code and our money. 00:09:33.279 --> 00:09:35.399 They do. And the thresholds just keep falling. 00:09:35.519 --> 00:09:38.080 We're crossing the biggest one right now. Midroll 00:09:38.080 --> 00:09:41.360 sponsor read. We have seen AI consume global 00:09:41.360 --> 00:09:44.620 power grids. Beat. It has disrupted international 00:09:44.620 --> 00:09:47.299 geopolitics. It is managing our code and our 00:09:47.299 --> 00:09:50.659 money. Now, beat. Alibaba is finally giving it 00:09:50.659 --> 00:09:52.779 a physical body. This is where the science fiction 00:09:52.779 --> 00:09:56.330 becomes our daily reality. Alibaba just unveiled 00:09:56.330 --> 00:09:58.850 the Quinn Robot Suite to achieve physical world 00:09:58.850 --> 00:10:00.769 intelligence. They broke this massive problem 00:10:00.769 --> 00:10:03.389 down into three specialized foundation models. 00:10:03.590 --> 00:10:06.850 Beat. First, there is Quinn Robot Manip. Right, 00:10:06.909 --> 00:10:09.409 basically the hands. It's trained on over 38 00:10:09.409 --> 00:10:12.169 ,000 hours of pure robotics data to translate 00:10:12.169 --> 00:10:14.889 visual inputs and natural language directly into 00:10:14.889 --> 00:10:17.549 precise physical manipulation. It learns how 00:10:17.549 --> 00:10:20.330 to grasp, turn, and pull. Next is Quinn Robotnov. 00:10:20.509 --> 00:10:23.210 Beat. These are the legs. Exactly. It handles 00:10:23.210 --> 00:10:25.330 complex spatial reasoning to navigate unseen 00:10:25.330 --> 00:10:27.870 environments. It avoids obstacles it has never 00:10:27.870 --> 00:10:30.129 encountered before. Finally, there's Quinn Robot 00:10:30.129 --> 00:10:34.029 World. Beat. This is the simulator, a video world 00:10:34.029 --> 00:10:37.250 model. And this is the real breakthrough. Traditional 00:10:37.250 --> 00:10:40.049 robots learn by physical trial and error, which 00:10:40.049 --> 00:10:42.649 means they fall over constantly. The simulator 00:10:42.649 --> 00:10:45.769 model uses real -time visual data to predict 00:10:45.769 --> 00:10:48.610 future trajectories in a digital space. How does 00:10:48.610 --> 00:10:50.809 the simulator model change the way robots learn? 00:10:51.049 --> 00:10:54.370 It allows the AI to test physics digitally. The 00:10:54.370 --> 00:10:57.490 AI watches a video of a cup falling, learns the 00:10:57.490 --> 00:11:00.370 implicit rule of gravity, and simulates the physics 00:11:00.370 --> 00:11:03.350 in its mind. It runs countless virtual iterations 00:11:03.350 --> 00:11:05.870 before moving a single physical servo. It lets 00:11:05.870 --> 00:11:08.549 robots safely practice physical tasks in a digital 00:11:08.549 --> 00:11:11.889 dream world first. Yes, a digital dream. And 00:11:11.889 --> 00:11:14.850 according to Alibaba, This framework excels at 00:11:14.850 --> 00:11:17.169 out -of -distribution generalization. Adapting 00:11:17.169 --> 00:11:20.009 to totally new, unexpected situations without 00:11:20.009 --> 00:11:22.730 explicit training. Right. If you drop it in a 00:11:22.730 --> 00:11:25.070 kitchen, it is never seen. It can still open 00:11:25.070 --> 00:11:27.429 a drawer it has never encountered. It's not just 00:11:27.429 --> 00:11:29.509 executing a hard -coded script, it's exhibiting 00:11:29.509 --> 00:11:33.190 genuine spatial exploration. Whoa. P -Land. Imagine 00:11:33.190 --> 00:11:35.789 scaling to a billion queries. Those queries are 00:11:35.789 --> 00:11:38.470 robots making physical decisions. Two secs, silence. 00:11:38.750 --> 00:11:41.169 It shifts our entire relationship with technology. 00:11:41.899 --> 00:11:43.960 AI is stepping out of the browser and into the 00:11:43.960 --> 00:11:45.899 living room. Let's pull all these threads together. 00:11:46.139 --> 00:11:48.379 The big takeaway today is the physicalization 00:11:48.379 --> 00:11:52.580 of AI. Beat. This technology started as simple 00:11:52.580 --> 00:11:55.830 chatbots. Today, we saw the entire physical pipeline 00:11:55.830 --> 00:11:58.409 emerge. It requires immense physical energy, 00:11:58.629 --> 00:12:02.009 driving 1 .8 million tons of carbon removal through 00:12:02.009 --> 00:12:04.389 the Frontier Coalition just to power the data 00:12:04.389 --> 00:12:06.970 centers. Then it hits our physical hardware. 00:12:07.190 --> 00:12:09.850 Beat. It triggers a ramageddon on our personal 00:12:09.850 --> 00:12:13.250 phones, potentially forcing a $270 hardware tax. 00:12:13.509 --> 00:12:16.409 Which escalates into physical geopolitics. G7 00:12:16.409 --> 00:12:18.570 leaders pushing back against American off switches 00:12:18.570 --> 00:12:20.830 because their local compute is constrained. And 00:12:20.830 --> 00:12:24.090 ultimately, it steps directly into reality. Beat, 00:12:24.250 --> 00:12:26.769 Alibaba's Quinn Foundation models are driving 00:12:26.769 --> 00:12:30.190 robotic hands and legs. AI is no longer a software 00:12:30.190 --> 00:12:32.450 abstraction. It is sharing our physical space 00:12:32.450 --> 00:12:34.289 now. I want to leave you with a new question 00:12:34.289 --> 00:12:38.090 to mull over. Beat. If an autonomous robot with 00:12:38.090 --> 00:12:40.629 physical world intelligence causes damage in 00:12:40.629 --> 00:12:43.029 a foreign country, but its actions were suddenly 00:12:43.029 --> 00:12:46.029 frozen by an American off switch built into its 00:12:46.029 --> 00:12:48.850 underlying frontier model, who ultimately holds 00:12:48.850 --> 00:12:51.389 the legal liability to sex violence? That is 00:12:51.389 --> 00:12:53.190 the multi -trillion dollar question we have to 00:12:53.190 --> 00:12:55.149 answer next. Thank you for joining us on this 00:12:55.149 --> 00:12:57.789 deep dive. Beat, keep questioning everything 00:12:57.789 --> 00:13:01.129 you see. We will see you next time. Outro music.