WEBVTT 00:00:00.000 --> 00:00:02.480 I want you to imagine a specific moment. You're 00:00:02.480 --> 00:00:04.580 sitting in a room, maybe a high -stakes meeting, 00:00:04.799 --> 00:00:08.279 and someone turns to you and asks a question 00:00:08.279 --> 00:00:10.740 you weren't expecting, a hard question. Before 00:00:10.740 --> 00:00:13.339 you answer, there's this silence. It's a split 00:00:13.339 --> 00:00:16.420 -second pause. Inside your head, you're auditing 00:00:16.420 --> 00:00:19.160 yourself. You're asking, do I actually know this? 00:00:19.320 --> 00:00:22.120 Am I about to bluff? How confident am I, really? 00:00:22.579 --> 00:00:25.960 That pause, that is the mechanism of intelligence. 00:00:26.480 --> 00:00:28.519 It's the difference between a reflex and a reason. 00:00:29.309 --> 00:00:32.049 And for the history of AI so far, well, we haven't 00:00:32.049 --> 00:00:33.789 had that pause. We've just had hallucinations. 00:00:33.869 --> 00:00:37.350 Today, that starts to change. Welcome back to 00:00:37.350 --> 00:00:39.429 the Deep Dive. I'm glad you're here. We have 00:00:39.429 --> 00:00:41.829 a stack of sources today that I think point to 00:00:41.829 --> 00:00:44.170 a fundamental shift in how these systems operate. 00:00:44.630 --> 00:00:47.429 We're moving from the era of the black box, where 00:00:47.429 --> 00:00:49.789 the AI just spits out an answer and we hope it's 00:00:49.789 --> 00:00:52.509 right, to a system that essentially checks its 00:00:52.509 --> 00:00:55.060 own math. We're looking at a new... Framework 00:00:55.060 --> 00:00:58.340 for AI metacognition. Metacognition. It sounds 00:00:58.340 --> 00:01:00.159 so academic, but it's really just thinking about 00:01:00.159 --> 00:01:02.240 thinking. It's the AI looking in the mirror. 00:01:02.460 --> 00:01:05.519 Exactly. And then we're going to take that concept 00:01:05.519 --> 00:01:08.120 of modeling the self and apply it to the physical 00:01:08.120 --> 00:01:10.719 world. We're looking at how AI is modeling the 00:01:10.719 --> 00:01:13.079 Earth's atmosphere. Oh, yeah. NVIDIA has entered 00:01:13.079 --> 00:01:15.120 the weather wars, and the implications for who 00:01:15.120 --> 00:01:18.019 owns the forecast are, well, they're complicated. 00:01:18.099 --> 00:01:20.159 It is wild stuff. And to keep the energy up, 00:01:20.200 --> 00:01:22.810 we also have to talk about vibe coding. Apparently 00:01:22.810 --> 00:01:25.650 resumes are dead, syntax is dead, and we're just 00:01:25.650 --> 00:01:27.730 hiring people based on how well they can vibe 00:01:27.730 --> 00:01:31.250 with an LLM. Plus some big hardware shifts from 00:01:31.250 --> 00:01:34.069 Microsoft. So here's the roadmap for our conversation. 00:01:34.629 --> 00:01:38.129 First, metacognition, the new sensors that let 00:01:38.129 --> 00:01:40.870 an AI reflect on its own uncertainty. Second, 00:01:41.170 --> 00:01:44.049 the toolkit, updates from Claude, Microsoft, 00:01:44.109 --> 00:01:46.849 and this rise of the vibe coder. And finally, 00:01:46.989 --> 00:01:50.049 Earth 2, the battle to simulate the climate. 00:01:50.409 --> 00:01:52.090 Let's unpack this. Let's get into it. I want 00:01:52.090 --> 00:01:54.189 to start with this paper on metacognition and 00:01:54.189 --> 00:01:57.049 the next wave of AI evolution. The researchers 00:01:57.049 --> 00:01:58.750 here have dropped a framework that I think is 00:01:58.750 --> 00:02:00.510 really important to understand, but I want to 00:02:00.510 --> 00:02:02.730 be precise with our language. When we say the 00:02:02.730 --> 00:02:05.810 AI is thinking about its thinking, we aren't 00:02:05.810 --> 00:02:07.530 talking about consciousness. We aren't talking 00:02:07.530 --> 00:02:10.729 about a soul or a ghost in the machine. No, absolutely 00:02:10.729 --> 00:02:14.150 not. We have to be so careful not to anthropomorphize. 00:02:14.719 --> 00:02:17.300 We are not near Skynet. This is engineering, 00:02:17.500 --> 00:02:20.780 not philosophy. Right. It's a metacognitive state 00:02:20.780 --> 00:02:24.319 vector. Which, again, sounds like Star Trek technobabble, 00:02:24.340 --> 00:02:26.580 but it's actually pretty practical. Think about 00:02:26.580 --> 00:02:29.360 how large language models, LLMs, usually work. 00:02:29.539 --> 00:02:31.580 You give a prompt, it predicts the next word. 00:02:31.719 --> 00:02:34.219 It's a probability engine. It's incredibly fast. 00:02:34.479 --> 00:02:37.539 Right. In psychology, specifically in the work 00:02:37.539 --> 00:02:39.919 of Daniel Kahneman, we'd call that system one 00:02:39.919 --> 00:02:43.400 thinking. Fast, intuitive, reactive. If I ask 00:02:43.400 --> 00:02:46.500 you, what is 2 plus 2? Hmm, you don't calculate 00:02:46.500 --> 00:02:48.360 it. The number 4 just appears in your head. That's 00:02:48.360 --> 00:02:50.719 system 1. Exactly. But if I ask you, what is 00:02:50.719 --> 00:02:55.659 17 multiplied by 24? You stop. The answer doesn't 00:02:55.659 --> 00:02:58.000 just pop up. You have to engage a different gear. 00:02:58.120 --> 00:03:00.060 You have to grind through the logic. That's system 00:03:00.060 --> 00:03:04.780 2. That's system 2. It's slow. deliberate reasoning. 00:03:05.020 --> 00:03:07.939 And up until now, AI has been kind of stuck in 00:03:07.939 --> 00:03:10.580 system one. It just blazes through confident 00:03:10.580 --> 00:03:14.180 even when it's completely wrong. This new framework 00:03:14.180 --> 00:03:17.360 gives the AI a dashboard to see when it needs 00:03:17.360 --> 00:03:19.520 to pump the brakes and switch gears. I love that 00:03:19.520 --> 00:03:21.500 distinction, the ability to switch gears. So 00:03:21.500 --> 00:03:23.919 let's look at the dashboard. This framework introduces 00:03:23.919 --> 00:03:27.180 five specific dimension sensors, essentially, 00:03:27.400 --> 00:03:29.939 that the AI tracks internally. The first one 00:03:29.939 --> 00:03:33.080 is the most obvious, confidence. How sure am 00:03:33.080 --> 00:03:35.759 I? But it's not just a binary yes or no. It's 00:03:35.759 --> 00:03:38.020 a gradient. The system tracks the probability 00:03:38.020 --> 00:03:41.039 variance of its own tokens. If the confidence 00:03:41.039 --> 00:03:43.879 score drops below a certain threshold, say, it's 00:03:43.879 --> 00:03:46.580 only 60 % sure of the next logical step. The 00:03:46.580 --> 00:03:48.520 system flags it. So it basically says, wait, 00:03:48.580 --> 00:03:50.340 I'm entering low probability territory here. 00:03:50.379 --> 00:03:52.379 I need to verify this before I speak. Precisely. 00:03:52.400 --> 00:03:54.539 The second sensor is conflict detection. Am I 00:03:54.539 --> 00:03:57.300 contradicting myself? This is huge because we've 00:03:57.300 --> 00:03:59.439 all seen an AI write an essay where the first 00:03:59.439 --> 00:04:01.659 paragraph claims one thing. And the conclusion 00:04:01.659 --> 00:04:03.699 claims the complete opposite. Yeah. It's like 00:04:03.699 --> 00:04:05.879 the AI is listening to its own output stream 00:04:05.879 --> 00:04:08.919 and comparing it against what it just said. It's 00:04:08.919 --> 00:04:11.840 checking for logical dissonance. If paragraph 00:04:11.840 --> 00:04:14.400 one says the project is under budget and paragraph 00:04:14.400 --> 00:04:17.100 three says costs have overrun, the conflict sensor 00:04:17.100 --> 00:04:19.740 spikes. In a standard model, it would just keep 00:04:19.740 --> 00:04:21.959 going, just keep hallucinating. Right. In this 00:04:21.959 --> 00:04:24.990 model, that spike. forces a correction then there's 00:04:24.990 --> 00:04:27.589 experience matching have i seen this before this 00:04:27.589 --> 00:04:30.410 grounds the reasoning it checks the current query 00:04:30.410 --> 00:04:32.990 against clusters of its training data if the 00:04:32.990 --> 00:04:35.910 problem is highly novel something totally outside 00:04:35.910 --> 00:04:38.930 its distribution it recognizes that it's guessing 00:04:38.930 --> 00:04:41.449 and it lowers its own confidence score it's the 00:04:41.449 --> 00:04:44.310 ability to admit ignorance which is surprisingly 00:04:44.310 --> 00:04:46.750 difficult to engineer it really is the fourth 00:04:46.750 --> 00:04:49.050 one is interesting to me emotional awareness 00:04:49.050 --> 00:04:52.250 is this content loaded this acts as a safety 00:04:52.250 --> 00:04:55.339 valve It's not feeling emotion, obviously, but 00:04:55.339 --> 00:04:58.519 it is detecting the stakes of the language. If 00:04:58.519 --> 00:05:01.139 a user is aggressive or the topic is politically 00:05:01.139 --> 00:05:03.240 charged or sensitive, like a medical crisis, 00:05:03.500 --> 00:05:07.139 the AI detects that weight. It signals that this 00:05:07.139 --> 00:05:09.720 isn't just a casual chat. It requires a higher 00:05:09.720 --> 00:05:13.199 degree of precision and care. And that ties into 00:05:13.199 --> 00:05:16.639 the fifth one, problem importance. Is this worth 00:05:16.639 --> 00:05:19.279 slowing down for? That's the efficiency key. 00:05:19.459 --> 00:05:22.779 Exactly. You don't need... Deep, slow, expensive 00:05:22.779 --> 00:05:26.000 reasoning to write a haiku about a cat. That's 00:05:26.000 --> 00:05:29.279 a waste of compute. But you do need it if you're 00:05:29.279 --> 00:05:32.120 analyzing a legal contract. This sensor tells 00:05:32.120 --> 00:05:35.160 the AI when to spend the resources. So when you 00:05:35.160 --> 00:05:37.480 combine these confidence, conflict, experience, 00:05:37.860 --> 00:05:40.500 emotion, importance, you get the state vector. 00:05:40.899 --> 00:05:43.199 And the result is that instead of a black box 00:05:43.199 --> 00:05:45.660 giving you an answer because I said so. you get 00:05:45.660 --> 00:05:47.839 explainable steps. Right. You get an output that 00:05:47.839 --> 00:05:50.459 looks something like, I chose this strategy because 00:05:50.459 --> 00:05:53.240 I was 92 % confident, I detected no internal 00:05:53.240 --> 00:05:55.879 conflict, and the problem importance triggered 00:05:55.879 --> 00:05:57.740 a deeper review. You know, I have to admit something 00:05:57.740 --> 00:06:00.800 here. I still struggle with prompt drift. I'll 00:06:00.800 --> 00:06:03.019 be working with an AI on a long thread, maybe 00:06:03.019 --> 00:06:05.100 for coding or writing, and after 10 minutes, 00:06:05.139 --> 00:06:07.360 it starts getting confused. Or honestly, I get 00:06:07.360 --> 00:06:09.180 confused about what I originally asked. Yeah, 00:06:09.220 --> 00:06:11.699 that happens. The idea that the AI could stop 00:06:11.699 --> 00:06:14.129 and say, hey, I'm feeling low confidence here, 00:06:14.209 --> 00:06:16.850 or I think we're contradicting the original goal, 00:06:17.410 --> 00:06:19.949 that is such a relief. It changes the dynamic 00:06:19.949 --> 00:06:23.310 completely. It removes the burden of you being 00:06:23.310 --> 00:06:25.870 the only adult in the room. Yeah. You're just 00:06:25.870 --> 00:06:27.730 prompting. You're collaborating with a system 00:06:27.730 --> 00:06:30.189 that has a sense of its own limitations. But 00:06:30.189 --> 00:06:33.290 let me ask you this. If AI tells us why it's 00:06:33.290 --> 00:06:35.810 confident, does that actually fix the hallucination 00:06:35.810 --> 00:06:38.389 problem? Or does it just give us a better excuse? 00:06:38.910 --> 00:06:40.930 That's the big question. And the short answer 00:06:40.930 --> 00:06:44.110 is... It doesn't fix it perfectly. The AI can 00:06:44.110 --> 00:06:47.089 still be confident and wrong, but it makes the 00:06:47.089 --> 00:06:49.949 error transparent. It moves us from silent failure 00:06:49.949 --> 00:06:53.750 to auditable failure. Transparency over perfection. 00:06:54.089 --> 00:06:56.810 I can work with that. Okay, let's shift gears. 00:06:57.339 --> 00:06:59.480 We look at the internal wiring, the mind of the 00:06:59.480 --> 00:07:02.000 machine. Now let's look at the tools on our desk. 00:07:02.319 --> 00:07:04.639 The landscape is moving so fast. We've got updates 00:07:04.639 --> 00:07:06.540 from the big players and some really interesting 00:07:06.540 --> 00:07:08.839 new concepts in how we actually work. Let's start 00:07:08.839 --> 00:07:11.259 with Claude. Anthropic has been busy. They've 00:07:11.259 --> 00:07:13.720 opened up Claude for Excel to pro users now. 00:07:13.959 --> 00:07:15.819 Finally, and this isn't just about reading a 00:07:15.819 --> 00:07:18.439 spreadsheet. The big deal here is memory and 00:07:18.439 --> 00:07:21.759 integrity. Context rot. Exactly. Historically, 00:07:21.819 --> 00:07:25.720 you'd feed a CSV to an AI, ask it to fix a column. 00:07:26.089 --> 00:07:28.550 and it hallucinates the rest of the sheet or 00:07:28.550 --> 00:07:32.569 just deletes rows. This update allows it to handle 00:07:32.569 --> 00:07:36.089 multi -sheet workbooks and, crucially, it doesn't 00:07:36.089 --> 00:07:38.490 overwrite your data. It edits in place. Yes, 00:07:38.569 --> 00:07:40.589 it keeps the context of the whole workbook. Just 00:07:40.589 --> 00:07:42.189 the difference between a toy and a real tool. 00:07:42.649 --> 00:07:45.170 And they've also launched interactive apps. You 00:07:45.170 --> 00:07:48.490 can run Slack. Figma, and Canva directly inside 00:07:48.490 --> 00:07:50.529 the cloud chat. It's becoming an operating system. 00:07:50.629 --> 00:07:53.050 You aren't just chatting, you're executing. Speaking 00:07:53.050 --> 00:07:55.649 of execution, Microsoft is hardening the infrastructure. 00:07:56.310 --> 00:07:59.329 They just unveiled the Maya 200 chip. It's rolling 00:07:59.329 --> 00:08:01.930 out in US data centers to power Copilot. 30 % 00:08:01.930 --> 00:08:04.230 faster, same cost. Yeah. So why does that matter 00:08:04.230 --> 00:08:06.389 to the listener? It matters because of the token 00:08:06.389 --> 00:08:08.730 tax. Every time you ask a deeper question, it 00:08:08.730 --> 00:08:12.110 costs time. Faster chips mean the AI feels less 00:08:12.110 --> 00:08:14.370 like a loading bar and more like a real conversation. 00:08:14.889 --> 00:08:16.870 It's the invisible plumbing that makes everything 00:08:16.870 --> 00:08:19.550 else possible. And on the creative side? CREA 00:08:19.550 --> 00:08:22.649 AI. They've launched real -time photo editing. 00:08:22.810 --> 00:08:25.290 You aren't waiting for a render. You tweak the 00:08:25.290 --> 00:08:28.389 prompt. The image changes instantly. It's fluid. 00:08:28.509 --> 00:08:31.850 And Synthesia just raised $200 million. Wow. 00:08:32.210 --> 00:08:34.409 They're building agents that interact with training 00:08:34.409 --> 00:08:36.769 videos. Imagine a corporate training video that 00:08:36.769 --> 00:08:38.669 answers your questions instead of just lecturing 00:08:38.669 --> 00:08:40.929 you. That's the vision. But the story that really 00:08:40.929 --> 00:08:44.029 caught my eye in this deck, vibe coding. I knew 00:08:44.029 --> 00:08:45.409 you were going to bring this up. I mean, come 00:08:45.409 --> 00:08:48.269 on. There's a startup called Anything, and they 00:08:48.269 --> 00:08:51.669 are hiring. But the job posting is unique. No 00:08:51.669 --> 00:08:54.730 resume, no GitHub repo. You just reply to the 00:08:54.730 --> 00:08:57.669 post. They're looking for vibe coders. The whole 00:08:57.669 --> 00:08:59.929 premise is fascinating. They argue that with 00:08:59.929 --> 00:09:02.389 the current state of LLMs, you don't really need 00:09:02.389 --> 00:09:05.970 to know Python or C++ syntax anymore. You need 00:09:05.970 --> 00:09:07.769 to know how to talk to the AI to get it to write 00:09:07.769 --> 00:09:10.029 the code. It suggests that coding is shifting 00:09:10.029 --> 00:09:12.490 substrates. It's moving from syntax, you know, 00:09:12.490 --> 00:09:14.470 knowing where to put the semicolon to pure intent. 00:09:14.809 --> 00:09:17.009 It's about being able to articulate a vision 00:09:17.009 --> 00:09:20.269 so clearly that the machine can build it. It's 00:09:20.269 --> 00:09:23.340 the democratization of software creation. But 00:09:23.340 --> 00:09:25.200 it's also probably a little terrifying for the 00:09:25.200 --> 00:09:27.919 purists. So here's a question. With vibe coding 00:09:27.919 --> 00:09:30.340 and tools like Verdant, which is another new 00:09:30.340 --> 00:09:32.840 one that coordinates multiple AI agents to code 00:09:32.840 --> 00:09:35.500 while you step away, are we seeing the end of 00:09:35.500 --> 00:09:38.940 the developer or is this just an evolution? I 00:09:38.940 --> 00:09:40.779 think we're seeing the evolution of the architect. 00:09:41.639 --> 00:09:44.940 The person who lays the bricks, the syntax, is 00:09:44.940 --> 00:09:47.620 being replaced by the machine. But the person 00:09:47.620 --> 00:09:50.269 who designs the cathedral. who understands the 00:09:50.269 --> 00:09:53.649 flow and the purpose. That role is more important 00:09:53.649 --> 00:09:56.730 than ever. From bricklayer to conductor. Exactly. 00:09:56.909 --> 00:09:59.450 You aren't typing code. You are orchestrating 00:09:59.450 --> 00:10:02.190 intelligence. Okay, we've looked at the inner 00:10:02.190 --> 00:10:04.830 mind of the AI metacognition. We've looked at 00:10:04.830 --> 00:10:07.049 the tools, chips, and vibe coding. Now, I want 00:10:07.049 --> 00:10:10.470 to zoom out. Way out. To the atmosphere of the 00:10:10.470 --> 00:10:12.549 planet itself. This is one of those stories that 00:10:12.549 --> 00:10:14.730 feels like it belongs in the future. But it happened 00:10:14.730 --> 00:10:16.769 last week at the American Meteorological Society 00:10:16.769 --> 00:10:19.750 meeting. NVIDIA, a chip company, announced something 00:10:19.750 --> 00:10:22.730 called Earth 2. Earth 2. It sounds like a backup 00:10:22.730 --> 00:10:24.889 planet, doesn't it? It's a fully open source 00:10:24.889 --> 00:10:28.340 suite of AI weather models. So how does a graphics 00:10:28.340 --> 00:10:31.139 card company predict the weather? Usually we 00:10:31.139 --> 00:10:33.720 do this with physics. Right. Traditional forecasting 00:10:33.720 --> 00:10:37.039 is physics -based. We use these massive supercomputers 00:10:37.039 --> 00:10:40.559 to solve fluid dynamics equations. It's incredibly 00:10:40.559 --> 00:10:42.679 accurate, but it's computationally heavy and 00:10:42.679 --> 00:10:45.139 it is slow. NVIDIA is taking a different approach. 00:10:45.220 --> 00:10:47.879 They're using AI. They're ingesting public Earth 00:10:47.879 --> 00:10:50.500 observation data, satellites, weather balloons, 00:10:50.700 --> 00:10:53.500 all of it. Billions of chaotic data points. And 00:10:53.500 --> 00:10:55.820 this is the moment of wonder for me. Think about 00:10:55.820 --> 00:10:58.519 the atmosphere. It's messy. It's noisy. There 00:10:58.519 --> 00:11:01.539 are gaps in the data. The AI takes this chaotic 00:11:01.539 --> 00:11:04.519 input clouds moving, pressure dropping, and it 00:11:04.519 --> 00:11:07.120 smooths it. It creates a continuous estimate 00:11:07.120 --> 00:11:09.240 of the atmospheric state. It's like taking a 00:11:09.240 --> 00:11:12.860 blurry pixelated image and using AI to upscale 00:11:12.860 --> 00:11:15.460 it to 4K resolution instantly. That's a perfect 00:11:15.460 --> 00:11:19.169 analogy. And because it's AI, it is fast. NVIDIA 00:11:19.169 --> 00:11:22.009 claims Earth 2 now outperforms those traditional 00:11:22.009 --> 00:11:24.610 physics -based models for short -term precipitation 00:11:24.610 --> 00:11:27.269 forecasting. That's a massive claim. It is. And 00:11:27.269 --> 00:11:29.529 the context here is vital. The newsletter mentions 00:11:29.529 --> 00:11:31.490 that federal funding for traditional forecasting 00:11:31.490 --> 00:11:34.889 is, well, it's drying up. It is. The public systems, 00:11:35.070 --> 00:11:37.409 the ones that farmers, wildfire response teams, 00:11:37.490 --> 00:11:39.830 and flight planners all rely on, they're under 00:11:39.830 --> 00:11:42.850 pressure. The satellites are aging. And into 00:11:42.850 --> 00:11:46.149 that void steps NVIDIA, along with Google, Microsoft 00:11:46.149 --> 00:11:49.529 and Huawei. They are all entering these weather 00:11:49.529 --> 00:11:52.490 wars. They see the value cap. Climate risk is 00:11:52.490 --> 00:11:54.909 the single biggest variable for the global economy. 00:11:55.230 --> 00:11:57.509 If you can predict the storm better than the 00:11:57.509 --> 00:12:00.529 government, you hold the keys to a lot of value. 00:12:00.710 --> 00:12:02.750 It raises a pretty serious question about access, 00:12:02.850 --> 00:12:05.710 though. If federal funding is drying up and the 00:12:05.710 --> 00:12:07.750 tech giants are taking over the weather models, 00:12:07.929 --> 00:12:11.059 what happens to public access to that data? That 00:12:11.059 --> 00:12:13.559 is the big question. Does climate safety become 00:12:13.559 --> 00:12:16.639 a paid premium service? Does a hedge fund get 00:12:16.639 --> 00:12:19.080 the NVIDIA forecast that says it will rain at 00:12:19.080 --> 00:12:22.019 2 .04 p .m. while the public gets a generic chance 00:12:22.019 --> 00:12:24.659 of showers? It turns weather into an information 00:12:24.659 --> 00:12:27.879 asymmetry. It's a sobering thought. It is. But 00:12:27.879 --> 00:12:30.360 on the flip side, the technology itself is incredible. 00:12:30.500 --> 00:12:32.320 The ability to model the Earth with this level 00:12:32.320 --> 00:12:34.919 of fidelity could save countless lives in disaster 00:12:34.919 --> 00:12:37.200 zones. It's just a matter of who holds the controls. 00:12:37.519 --> 00:12:40.179 So let's try to pull this all together. What's 00:12:40.179 --> 00:12:43.299 the thread connecting these segments? We started 00:12:43.299 --> 00:12:46.100 with the micro giving AI an internal sensor system 00:12:46.100 --> 00:12:48.440 to check its own confidence. We moved to the 00:12:48.440 --> 00:12:51.440 macro Earth, too, using AI to sensor the entire 00:12:51.440 --> 00:12:54.259 planet. I think the theme is precision and reliability. 00:12:54.799 --> 00:12:57.460 We're moving past the guesswork phase of generative 00:12:57.460 --> 00:13:00.539 AI. We're moving past the novelty phase. Now 00:13:00.539 --> 00:13:03.899 we're asking, can it check its own math? Can 00:13:03.899 --> 00:13:07.720 it reliably code an app? Can it predict a hurricane 00:13:07.720 --> 00:13:09.879 better than a physics engine? We're demanding 00:13:09.879 --> 00:13:12.059 that the systems double check themselves. Exactly. 00:13:13.059 --> 00:13:15.279 Metacognition builds trust internally. Earth 00:13:15.279 --> 00:13:17.820 2 builds trust externally. We're trying to build 00:13:17.820 --> 00:13:20.200 a digital substrate that actually maps to reality. 00:13:20.580 --> 00:13:22.299 I want to leave you with a thought to mull over. 00:13:22.460 --> 00:13:24.659 We talked about the AI checking its own confusion. 00:13:24.840 --> 00:13:27.320 We talked about it modeling the chaotic atmosphere 00:13:27.320 --> 00:13:30.519 better than physics equations. If AI can detect 00:13:30.519 --> 00:13:32.440 its own blind spots and if it can understand 00:13:32.440 --> 00:13:35.220 the physical world better than we can, at what 00:13:35.220 --> 00:13:37.320 point do we stop double checking the AI? And 00:13:37.320 --> 00:13:40.019 start having the AI double check us. Exactly. 00:13:40.019 --> 00:13:42.360 Would you trust a weather forecast from NVIDIA 00:13:42.360 --> 00:13:44.940 over the National Weather Service? It's not a 00:13:44.940 --> 00:13:47.860 hypothetical anymore. I'd check both. For now. 00:13:48.100 --> 00:13:50.620 But I know which one is learning faster. Let 00:13:50.620 --> 00:13:52.860 us know what you think. Until next time. See 00:13:52.860 --> 00:13:53.000 you then.