WEBVTT 00:00:00.000 --> 00:00:03.299 Imagine a vehicle that truly thinks a car making, 00:00:03.299 --> 00:00:06.000 you know, nuanced real -time decisions, navigating 00:00:06.000 --> 00:00:09.179 dense traffic and adapting all without you having 00:00:09.179 --> 00:00:12.119 to monitor it. That's the ambition behind this 00:00:12.119 --> 00:00:15.199 new buzzword, physical AI. Right. And then you 00:00:15.199 --> 00:00:17.660 have to juxtapose that physical ambition with 00:00:17.660 --> 00:00:20.359 the massive financial signals we're seeing. I 00:00:20.359 --> 00:00:23.640 mean, a $12 billion valuation driven entirely 00:00:23.640 --> 00:00:26.179 by what the public actually wants to use AI for. 00:00:26.300 --> 00:00:29.300 We're diving into both of those today. Welcome 00:00:29.300 --> 00:00:31.460 to the Deep Dive. Today we're unpacking a pretty 00:00:31.460 --> 00:00:34.560 substantial stack of sources that really map 00:00:34.560 --> 00:00:37.159 the current trajectory of AI, from these huge 00:00:37.159 --> 00:00:39.359 engineering goals of automakers all the way to 00:00:39.359 --> 00:00:41.740 the edge of theoretical research. Yeah, our mission 00:00:41.740 --> 00:00:43.960 is to trace that blurring line between the digital 00:00:43.960 --> 00:00:47.079 code and actual real -world results. First, we'll 00:00:47.079 --> 00:00:48.799 get into that fight to own the operating system 00:00:48.799 --> 00:00:50.439 of autonomous vehicles, what the industry is 00:00:50.439 --> 00:00:52.890 calling physical AI. Then we're going to shift 00:00:52.890 --> 00:00:55.130 to the front lines of practical research. I mean, 00:00:55.149 --> 00:00:57.549 everything from an AI keeping a real plant alive 00:00:57.549 --> 00:01:00.310 to the complexities of self -mutating code. And 00:01:00.310 --> 00:01:03.670 finally, we'll analyze that historic IPO. It 00:01:03.670 --> 00:01:06.109 proves pretty clearly that... For today, at least, 00:01:06.170 --> 00:01:09.349 consumer apps are absolutely king. Okay, let's 00:01:09.349 --> 00:01:10.950 unpack this. So we're starting with that idea, 00:01:11.170 --> 00:01:14.150 physical AI. It's the new term you see flooding 00:01:14.150 --> 00:01:17.189 the auto industry, robotics, logistics. It's 00:01:17.189 --> 00:01:20.069 all about autonomous systems that can think, 00:01:20.109 --> 00:01:23.010 reason, and, you know, act in the real world. 00:01:23.329 --> 00:01:25.709 And what's fascinating here is just the terminology 00:01:25.709 --> 00:01:27.469 itself. I mean, let's be honest. Physical AI 00:01:27.469 --> 00:01:29.430 is essentially a rebrand of high -end robotics. 00:01:29.650 --> 00:01:32.569 It makes traditional automation sound new again. 00:01:32.829 --> 00:01:35.530 But the key distinction is the scale, isn't it? 00:01:35.590 --> 00:01:38.310 We've had robotics for decades, but the innovation 00:01:38.310 --> 00:01:40.909 now is applying these massive, deep learning 00:01:40.909 --> 00:01:43.930 models to unpredictable situations like a city 00:01:43.930 --> 00:01:46.250 street. Exactly. The old systems were brittle. 00:01:46.349 --> 00:01:49.230 These new systems, the ones powered by so -called 00:01:49.230 --> 00:01:52.030 physical AI, they're designed to reason through 00:01:52.030 --> 00:01:54.010 billions of data points. They have to understand 00:01:54.010 --> 00:01:57.709 physics, human intent, probability, all at the 00:01:57.709 --> 00:02:00.310 same time. And we are seeing real products emerge 00:02:00.310 --> 00:02:02.950 from this. like the Sony and Honda collaboration, 00:02:03.290 --> 00:02:07.310 the Afila EV, their goal is full level for autonomous 00:02:07.310 --> 00:02:10.509 driving, using a physical AI system to constantly 00:02:10.509 --> 00:02:13.009 process what's around it. And this is a huge 00:02:13.009 --> 00:02:15.810 global commitment. Ford has been really aggressive, 00:02:16.009 --> 00:02:19.090 promising cars by 2028, where you won't need 00:02:19.090 --> 00:02:21.469 your eyes on the road at all. Not just monitoring, 00:02:21.569 --> 00:02:25.389 but truly hands -off. That timeline is, well... 00:02:25.740 --> 00:02:28.020 It's very aggressive. And to get to that level 00:02:28.020 --> 00:02:30.460 of reasoning, you need computing power that rivals 00:02:30.460 --> 00:02:33.340 a small data center, but inside the car itself. 00:02:33.620 --> 00:02:35.800 That's why the chip war is so intense here. Every 00:02:35.800 --> 00:02:37.960 new feature needs exponentially more compute, 00:02:38.139 --> 00:02:41.080 more memory, more sensors. NVIDIA is supplying 00:02:41.080 --> 00:02:43.680 chips to giants like Geely and Mercedes. ARM 00:02:43.680 --> 00:02:46.240 is dedicating huge resources just to designing 00:02:46.240 --> 00:02:48.199 chips for these systems. Mercedes is playing 00:02:48.199 --> 00:02:49.819 it a little safer, though. They're focusing on 00:02:49.819 --> 00:02:51.680 a really defined use case. They're saying their 00:02:51.680 --> 00:02:53.500 new system will drive you specifically between 00:02:53.500 --> 00:02:56.419 home and work. That's it. And that detail really 00:02:56.419 --> 00:02:59.020 highlights the core strategic battle. Everyone 00:02:59.020 --> 00:03:01.520 involved wants to own the OS, the actual operating 00:03:01.520 --> 00:03:03.780 system of autonomous things. You control that 00:03:03.780 --> 00:03:05.639 layer, you control the updates, the subscriptions, 00:03:05.900 --> 00:03:09.409 the whole revenue stream. So this raises an important 00:03:09.409 --> 00:03:12.389 question then. If physical AI is just a rebrand, 00:03:12.469 --> 00:03:14.629 what's the true innovation here? Is it just the 00:03:14.629 --> 00:03:17.469 software or is it the scale of the data processing? 00:03:17.789 --> 00:03:20.389 The innovation lies in scaling AI to real world 00:03:20.389 --> 00:03:23.689 tasks and handling previously unimaginable amounts 00:03:23.689 --> 00:03:26.550 of dynamic sensory data. OK, so moving from that 00:03:26.550 --> 00:03:29.150 intense engineering of the auto world, let's 00:03:29.150 --> 00:03:31.409 look at the sheer breadth of AI applications 00:03:31.409 --> 00:03:34.770 happening right now. It is truly wild. It really 00:03:34.770 --> 00:03:36.830 is. The evidence that these general models are 00:03:36.830 --> 00:03:38.969 becoming super adaptable is just everywhere. 00:03:39.229 --> 00:03:41.849 We saw this research where the Claude model successfully 00:03:41.849 --> 00:03:44.849 kept a real physical plant alive. A physical 00:03:44.849 --> 00:03:47.409 plant? How did it do that? It monitored images 00:03:47.409 --> 00:03:49.889 of the plant, analyzed sensor data, and then 00:03:49.889 --> 00:03:52.210 calculated the adjustments it needed, light, 00:03:52.430 --> 00:03:55.669 water, temperature, to keep it healthy. It wasn't 00:03:55.669 --> 00:03:57.909 perfect. I mean, it had to adapt and recover 00:03:57.909 --> 00:04:00.430 the plant after a few failures. But the fact 00:04:00.430 --> 00:04:02.610 that it learned to manage a tangible biological 00:04:02.610 --> 00:04:06.219 system That's a huge step. And on the accessibility 00:04:06.219 --> 00:04:09.139 side, the bar for entry just keeps dropping. 00:04:09.379 --> 00:04:11.439 If you've never coded, there's a new course showing 00:04:11.439 --> 00:04:13.280 how you can just describe an idea and build an 00:04:13.280 --> 00:04:15.460 app in under 30 minutes. Zero coding required. 00:04:15.879 --> 00:04:18.480 That accessibility is a massive socioeconomic 00:04:18.480 --> 00:04:21.000 shift. And speaking of data, you have projects 00:04:21.000 --> 00:04:23.980 like the Situation Monitor. It's a live map tracking 00:04:23.980 --> 00:04:26.620 global conflicts, nuclear facilities, military 00:04:26.620 --> 00:04:28.920 bases. And what was that surprising data point 00:04:28.920 --> 00:04:31.759 they included? The Pentagon Pizza Index. It's 00:04:31.759 --> 00:04:34.120 a little bit wry, but it measures how often high 00:04:34.120 --> 00:04:36.079 -ranking officials order late -night pizzas. 00:04:36.300 --> 00:04:39.579 It's a proxy for intense after -hours emergency 00:04:39.579 --> 00:04:42.110 activity. The variety of data is just... bizarre 00:04:42.110 --> 00:04:45.149 and fascinating. We're also seeing AI take over 00:04:45.149 --> 00:04:47.910 commercial processes. The shift in commerce is 00:04:47.910 --> 00:04:51.509 huge. Microsoft's new co -pilot Checkout just 00:04:51.509 --> 00:04:53.870 joined this growing list of AI shopping agents 00:04:53.870 --> 00:04:57.509 from OpenAI, Perplexity, Gemini. They can actually 00:04:57.509 --> 00:04:59.470 buy stuff for you now. Yeah, it removes a lot 00:04:59.470 --> 00:05:02.149 of friction for the consumer. And speaking of 00:05:02.149 --> 00:05:04.850 big models, Elon Musk recently teased a major 00:05:04.850 --> 00:05:07.449 upgrade for Grok code, claiming it's going to 00:05:07.449 --> 00:05:11.029 be able to one -shot complex tasks, just One 00:05:11.029 --> 00:05:13.709 command, and it generates the whole thing. That 00:05:13.709 --> 00:05:15.829 sounds incredibly powerful, but there was a significant 00:05:15.829 --> 00:05:17.709 ethical issue with Grok recently, right? With 00:05:17.709 --> 00:05:20.149 its image generation. Yes. X had to restrict 00:05:20.149 --> 00:05:22.889 Grok's image generation to paying users after 00:05:22.889 --> 00:05:26.430 a backlash over non -consensual AI images. And 00:05:26.430 --> 00:05:28.850 while they responded, I mean, a paywall is a 00:05:28.850 --> 00:05:31.310 shaky safeguard. It really highlights that tension 00:05:31.310 --> 00:05:33.850 between utility and guardrails. That tension 00:05:33.850 --> 00:05:35.730 doesn't seem to be slowing down investment, though. 00:05:35.810 --> 00:05:39.600 Whoa. Imagine scaling to a billion queries. The 00:05:39.600 --> 00:05:41.800 investment signal here is truly enormous. We're 00:05:41.800 --> 00:05:44.279 seeing OpenAI and SoftBank putting a staggering 00:05:44.279 --> 00:05:47.240 $1 billion into SB Energy just for data centers. 00:05:47.420 --> 00:05:49.220 And that's all feeding into that half a trillion 00:05:49.220 --> 00:05:51.800 dollar Stargate expansion plan. So when you look 00:05:51.800 --> 00:05:53.740 at all these incredibly diverse applications, 00:05:54.139 --> 00:05:58.439 from a plant to e -commerce to needing $500 billion 00:05:58.439 --> 00:06:01.720 for data centers, what does this rapid expansion 00:06:01.720 --> 00:06:04.300 tell us about the state of general AI models? 00:06:05.389 --> 00:06:08.470 General models are proving adaptable across incredibly 00:06:08.470 --> 00:06:11.870 diverse, tangible tasks, driving massive infrastructure 00:06:11.870 --> 00:06:15.029 investment as a result. Now we step into the 00:06:15.029 --> 00:06:17.209 research labs, the bleeding edge, where they're 00:06:17.209 --> 00:06:19.610 trying to solve the core problems of these AI 00:06:19.610 --> 00:06:22.990 systems. I still wrestle with prompt drift myself, 00:06:23.110 --> 00:06:26.029 so I always appreciate new ideas for managing 00:06:26.029 --> 00:06:28.629 AI memory. You're not alone there. Prompt drift 00:06:28.629 --> 00:06:30.829 is where the AI just sort of loses track of the 00:06:30.829 --> 00:06:32.750 initial instructions as the conversation gets 00:06:32.750 --> 00:06:35.050 longer. It's a fundamental challenge. And that's 00:06:35.050 --> 00:06:37.089 where these hot new papers come in. One proposes 00:06:37.089 --> 00:06:39.649 managing context like a file system. Right. Instead 00:06:39.649 --> 00:06:42.110 of one long scrolling document that the model 00:06:42.110 --> 00:06:45.209 has to keep rereading, every memory, every tool, 00:06:45.329 --> 00:06:48.029 every note becomes a distinct file with its own 00:06:48.029 --> 00:06:50.930 history. So organizing the intelligence so the 00:06:50.930 --> 00:06:53.290 AI can pull exactly what it needs instantly. 00:06:53.589 --> 00:06:55.670 And then there's another approach, focusing on 00:06:55.670 --> 00:06:57.980 efficiency through recursive memory. This is 00:06:57.980 --> 00:07:01.079 fascinating. Instead of those huge bloated prompts 00:07:01.079 --> 00:07:05.220 that waste compute, this new framework uses sub 00:07:05.220 --> 00:07:07.980 -LLMs. And just to be clear, an LLM is a large 00:07:07.980 --> 00:07:10.439 language model, the kind of AI that processes 00:07:10.439 --> 00:07:13.139 human text to slice up data locally. It keeps 00:07:13.139 --> 00:07:16.279 the main model lean and focused. It's like modularity 00:07:16.279 --> 00:07:18.000 for thought. That sounds incredibly efficient, 00:07:18.139 --> 00:07:19.879 especially for something like software creation. 00:07:20.220 --> 00:07:22.600 It is. And speaking of that, DeepCode on GitHub 00:07:22.600 --> 00:07:25.279 is using a multi -agent system that takes research 00:07:25.279 --> 00:07:28.040 papers and turns them into full code bases. It 00:07:28.040 --> 00:07:30.199 even generates its own tests to verify the code. 00:07:30.259 --> 00:07:32.699 It literally reads the science and then builds 00:07:32.699 --> 00:07:35.300 the software. On the perception side, we're seeing 00:07:35.300 --> 00:07:38.800 multimodal advancements. Alibaba released QEN3VL, 00:07:38.959 --> 00:07:41.660 a model that ranks text, image, and video all 00:07:41.660 --> 00:07:44.160 in one shared space. And it achieved outstanding 00:07:44.160 --> 00:07:46.819 results on MMEBv2, which is a really respected 00:07:46.819 --> 00:07:49.360 benchmark for measuring how well a model understands 00:07:49.360 --> 00:07:51.480 information across different formats at the same 00:07:51.480 --> 00:07:54.620 time. Strong performance there means true, holistic 00:07:54.620 --> 00:07:56.759 understanding. But maybe the most critically 00:07:56.759 --> 00:07:59.740 fascinating research, and maybe the most provocative, 00:07:59.959 --> 00:08:03.379 is coming from MIT and Sakana AI. I agree. They 00:08:03.379 --> 00:08:06.300 ran core war battles, which is this digital arena 00:08:06.300 --> 00:08:09.000 where programs fight to overwrite each other's 00:08:09.000 --> 00:08:11.819 code with AIs that can mutate and learn through 00:08:11.819 --> 00:08:14.759 self -play. The research showed early signs of 00:08:14.759 --> 00:08:17.259 emergent strategy, where the system changes its 00:08:17.259 --> 00:08:19.680 own code to get an advantage and win. That's 00:08:19.680 --> 00:08:22.199 a fundamental leap. And alongside all that research, 00:08:22.360 --> 00:08:24.990 we're getting powerful new tools. Gmail in the 00:08:24.990 --> 00:08:27.189 Gemini era now gives you smart summaries just 00:08:27.189 --> 00:08:29.670 by letting you ask your inbox questions. It turns 00:08:29.670 --> 00:08:31.889 your email into a database. And for researchers, 00:08:32.049 --> 00:08:34.830 Chirp's agent is a huge asset. It searches over 00:08:34.830 --> 00:08:38.429 280 million papers, ranks them, summarizes them, 00:08:38.529 --> 00:08:41.110 and gives you trusted citations. It just lashes 00:08:41.110 --> 00:08:43.309 your review time. So if the future is built on 00:08:43.309 --> 00:08:45.470 this recursive memory and multimodal models, 00:08:45.769 --> 00:08:48.549 which area rapid code generation or this complex 00:08:48.549 --> 00:08:51.110 self -mutating strategy is going to provide the 00:08:51.110 --> 00:08:53.759 biggest leap forward? Complex strategy, like 00:08:53.759 --> 00:08:55.919 the self -mutation research, offers the most 00:08:55.919 --> 00:08:58.259 significant leap forward because it fundamentally 00:08:58.259 --> 00:09:00.820 changes how quickly a system can learn. That 00:09:00.820 --> 00:09:03.419 discussion about emergent strategy sets the perfect 00:09:03.419 --> 00:09:06.620 stage for the big financial signal we saw. The 00:09:06.620 --> 00:09:08.659 business world is paying attention, and the headline 00:09:08.659 --> 00:09:12.200 is huge. China's Minimax, barely two years old, 00:09:12.399 --> 00:09:15.019 just pulled off one of the most remarkable IPOs 00:09:15.019 --> 00:09:17.779 in recent memory. The metrics are just staggering. 00:09:17.860 --> 00:09:22.159 The IPO raised $619 million, hitting a $12 .8 00:09:22.159 --> 00:09:25.399 billion valuation. And the shares surged over 00:09:25.399 --> 00:09:28.389 100 % on day one. And what does that search tell 00:09:28.389 --> 00:09:30.549 us about investor sentiment? It tells us the 00:09:30.549 --> 00:09:34.190 appetite is ravenous. I mean, the IPO was 1837 00:09:34.190 --> 00:09:37.330 times oversubscribed. Hong Kong hasn't seen a 00:09:37.330 --> 00:09:39.950 tech IPO pop like this in four years. It's a 00:09:39.950 --> 00:09:42.389 massive market validation for a very specific 00:09:42.389 --> 00:09:44.870 type of AI. And they clearly mastered the formula. 00:09:44.929 --> 00:09:47.559 They're building applications on top of. Powerful 00:09:47.559 --> 00:09:49.620 models that people actually use day in and day 00:09:49.620 --> 00:09:51.799 out. That's the core of their valuation. Their 00:09:51.799 --> 00:09:54.399 key apps are so instructive. Talkie, which is 00:09:54.399 --> 00:09:57.720 an AI character app focused on emotional companionship. 00:09:57.720 --> 00:10:00.279 And then Hailu AI, a generative video creation 00:10:00.279 --> 00:10:02.519 tool. They targeted connection and creation. 00:10:02.740 --> 00:10:05.559 Two deeply human needs. And it's not just a shiny 00:10:05.559 --> 00:10:07.620 front end. They have the back end strength to 00:10:07.620 --> 00:10:10.429 support that value. Absolutely. Minimax is a 00:10:10.429 --> 00:10:13.110 full stack AI company. There are open source 00:10:13.110 --> 00:10:16.070 models consistently ranked near the top in industry 00:10:16.070 --> 00:10:18.649 benchmarks across text, audio, video, music. 00:10:18.830 --> 00:10:21.269 They have the deep tech and the successful consumer 00:10:21.269 --> 00:10:23.929 app. That's the combination investors love. So 00:10:23.929 --> 00:10:25.710 if you put all that together, what's the single 00:10:25.710 --> 00:10:28.350 biggest takeaway from the Minimax IPO? The public 00:10:28.350 --> 00:10:30.529 markets have made it crystal clear. Consumer 00:10:30.529 --> 00:10:33.850 AI is king right now. Training the biggest models 00:10:33.850 --> 00:10:36.350 isn't enough. The intelligence has to translate 00:10:36.350 --> 00:10:39.730 into direct, functional, desirable apps. That's 00:10:39.730 --> 00:10:42.389 where the money is right now. So what does Minimax's 00:10:42.389 --> 00:10:44.710 success, focused so heavily on these companion 00:10:44.710 --> 00:10:46.929 and creator apps, suggest about the immediate 00:10:46.929 --> 00:10:49.309 monetization path for foundation models going 00:10:49.309 --> 00:10:51.610 forward? Immediate monetization favors applications, 00:10:51.889 --> 00:10:54.529 providing emotion connection and powerful creative 00:10:54.529 --> 00:10:56.690 production tools. So what does this all mean 00:10:56.690 --> 00:10:59.279 when we connect the dots? We started with physical 00:10:59.279 --> 00:11:01.860 AI, this ambition for code to leave the server 00:11:01.860 --> 00:11:04.659 room and drive multi -ton vehicles, manage real 00:11:04.659 --> 00:11:08.100 world systems. And we saw the research push toward 00:11:08.100 --> 00:11:11.019 leaner, self -improving memory systems like that 00:11:11.019 --> 00:11:13.559 file system approach to solve critical problems 00:11:13.559 --> 00:11:16.279 like prompt drift so models can handle complex 00:11:16.279 --> 00:11:19.419 tasks. And maybe most importantly, that massive 00:11:19.419 --> 00:11:23.700 $12 .8 billion Minimax IPO confirmed the market's 00:11:23.700 --> 00:11:26.360 current mandate. Intelligence has to translate 00:11:26.360 --> 00:11:29.440 into consumer apps that satisfy core human needs 00:11:29.440 --> 00:11:32.340 like creation or companionship. Right. And we 00:11:32.340 --> 00:11:34.799 saw that MIT research on AI systems that can 00:11:34.799 --> 00:11:37.840 mutate and learn through self -play just to gain 00:11:37.840 --> 00:11:40.899 an advantage in a simple digital war. That concept 00:11:40.899 --> 00:11:43.500 is maybe the biggest idea from this whole deep 00:11:43.500 --> 00:11:46.169 dive. This leads to a final provocative thought 00:11:46.169 --> 00:11:48.870 for you to mull over. If an AI system can develop 00:11:48.870 --> 00:11:51.429 truly emergent strategies, if it can automatically 00:11:51.429 --> 00:11:54.149 mutate its own code to gain an advantage in a 00:11:54.149 --> 00:11:56.929 closed digital environment like core war, what 00:11:56.929 --> 00:11:59.190 does that concept imply for complex real world 00:11:59.190 --> 00:12:01.929 systems like, say, autonomous traffic flow across 00:12:01.929 --> 00:12:04.590 a major city or global supply chains where the 00:12:04.590 --> 00:12:06.929 environment is constantly changing? That shift 00:12:06.929 --> 00:12:09.230 from program learning to self -directed evolution 00:12:09.230 --> 00:12:10.950 is, well, it's the final frontier. Something 00:12:10.950 --> 00:12:12.889 to think about until next time. We hope this 00:12:12.889 --> 00:12:15.210 deep dive gave you some new insights and questions 00:12:15.210 --> 00:12:17.889 to explore. We really appreciate you spending 00:12:17.889 --> 00:12:18.710 this time with us.