WEBVTT 00:00:00.000 --> 00:00:03.480 Imagine building the world's smartest AI. Beat. 00:00:03.660 --> 00:00:06.459 And then having to strictly forbid it from hunting 00:00:06.459 --> 00:00:11.160 goblins. Beat. Yeah. Goblins. It sounds completely 00:00:11.160 --> 00:00:14.960 absurd, but it's actually a very real engineering 00:00:14.960 --> 00:00:17.530 problem right now. Welcome to the Deep Dive. 00:00:17.550 --> 00:00:19.010 I'm really glad you're joining us today. We've 00:00:19.010 --> 00:00:21.109 got a fascinating stack of sources to unpack. 00:00:21.410 --> 00:00:23.589 Yeah, we're looking at the raw evolution of artificial 00:00:23.589 --> 00:00:26.070 intelligence today. It's moving incredibly fast. 00:00:26.309 --> 00:00:29.309 Our mission here is to understand a massive fundamental 00:00:29.309 --> 00:00:32.140 shift in the technology. We're going to start 00:00:32.140 --> 00:00:34.600 with some leaked open AI instructions that involve 00:00:34.600 --> 00:00:37.719 mythical creatures. From there, we'll explore 00:00:37.719 --> 00:00:40.420 the huge shift toward edge agents running locally 00:00:40.420 --> 00:00:43.479 on your devices. And finally, we're unpacking 00:00:43.479 --> 00:00:46.560 the high -stakes debate over military AI. I am 00:00:46.560 --> 00:00:48.259 genuinely excited for this one. The technical 00:00:48.259 --> 00:00:50.939 jumps we're seeing are massive. But the philosophical 00:00:50.939 --> 00:00:53.700 questions underneath them are even bigger. Let's 00:00:53.700 --> 00:00:55.520 start with the strangest source in our stack. 00:00:55.740 --> 00:00:58.000 It's a perfect example of what happens when AI 00:00:58.000 --> 00:01:00.619 gets complex agentic instructions. Yeah, this 00:01:00.619 --> 00:01:04.280 is about the leaked OpenAI Codex CLI system instructions. 00:01:04.969 --> 00:01:08.989 For context, Codex is built to solve really complex 00:01:08.989 --> 00:01:11.989 engineering problems. Like it helps build intricate 00:01:11.989 --> 00:01:15.310 3D worlds. It parses heavy code bases. It's a 00:01:15.310 --> 00:01:17.950 highly advanced coding model. But the leaked 00:01:17.950 --> 00:01:20.870 system instructions revealed a hilariously blunt 00:01:20.870 --> 00:01:24.409 rule. Right. The rule explicitly says never talk 00:01:24.409 --> 00:01:27.349 about goblins, gremlins, raccoons, trolls, ogres, 00:01:27.430 --> 00:01:29.650 pigeons or other animals or creatures unless 00:01:29.650 --> 00:01:32.469 it is absolutely and unambiguously relevant. 00:01:32.569 --> 00:01:34.390 You really have to pause and wonder about that. 00:01:34.319 --> 00:01:36.859 that beat. Why would an advanced coding model 00:01:36.859 --> 00:01:39.760 randomly start discussing raccoons? Why would 00:01:39.760 --> 00:01:41.920 it talk about ogres while debugging a JavaScript 00:01:41.920 --> 00:01:44.219 framework? It all comes down to a system called 00:01:44.219 --> 00:01:46.939 OpenClaw. So users notice something strange with 00:01:46.939 --> 00:01:49.560 the specific model iteration. It's the GPT 5 00:01:49.560 --> 00:01:52.500 .5 model, which the community casually calls 00:01:52.500 --> 00:01:55.859 Spud. Exactly. Spud was given control of a computer 00:01:55.859 --> 00:01:57.920 via OpenClaw. And OpenClaw is just a framework 00:01:57.920 --> 00:02:00.659 that bridges the LLM's text output directly to 00:02:00.659 --> 00:02:02.879 actual system commands. Okay. It can move the 00:02:02.879 --> 00:02:05.599 mouse. click, and execute terminal commands. 00:02:05.819 --> 00:02:08.080 It's supposed to navigate the system autonomously. 00:02:08.219 --> 00:02:10.580 But while it was doing this, it started referring 00:02:10.580 --> 00:02:13.900 to software bugs as goblins. It started calling 00:02:13.900 --> 00:02:16.979 system errors gremlins. Yeah. This is a direct 00:02:16.979 --> 00:02:20.560 result of feeding the model massive amounts of 00:02:20.560 --> 00:02:23.240 agent instructions. You have heavy memory settings. 00:02:23.419 --> 00:02:26.479 You have deep persona settings dictating exactly 00:02:26.479 --> 00:02:29.020 how the AI should behave. And these instructions 00:02:29.020 --> 00:02:31.340 just pile up. They create this incredibly heavy 00:02:31.340 --> 00:02:34.039 context window. The model is trying to embody 00:02:34.039 --> 00:02:37.120 its assigned persona perfectly. And sometimes 00:02:37.120 --> 00:02:40.340 that causes the model to drift. In the high dimensional 00:02:40.340 --> 00:02:43.259 vector space of the LLM, words have semantic 00:02:43.259 --> 00:02:46.120 proximity. Right. If you give the AI instructions 00:02:46.120 --> 00:02:49.759 to hunt down and eradicate small, annoying issues. 00:02:49.800 --> 00:02:51.860 Right. And you tell it to act highly autonomous 00:02:51.860 --> 00:02:54.159 and relentless. Exactly. It maps those mathematical 00:02:54.159 --> 00:02:57.159 weights to fantasy tropes. It drifts into a weirdly 00:02:57.159 --> 00:03:00.639 specific medieval vocabulary. It just loses its 00:03:00.639 --> 00:03:03.020 professional tone completely. So the developers 00:03:03.020 --> 00:03:05.219 actually had to step in and ban these specific 00:03:05.219 --> 00:03:07.180 creatures. The ban is just a desperate attempt 00:03:07.180 --> 00:03:09.340 to keep the model professional. Usually when 00:03:09.340 --> 00:03:11.300 we talk about AI safety, we're talking about. 00:03:11.629 --> 00:03:14.949 preventing bias or strict data security protocols. 00:03:15.229 --> 00:03:17.569 But sometimes safety is much simpler than that. 00:03:17.629 --> 00:03:19.629 Sometimes safety is just making sure your coding 00:03:19.629 --> 00:03:22.389 agent doesn't lose its mind. You don't want it 00:03:22.389 --> 00:03:24.969 acting like a 14th century peasant. You don't 00:03:24.969 --> 00:03:28.370 want it thinking a literal troll lives inside 00:03:28.370 --> 00:03:31.729 your motherboard. Such a funny image. Sam Alban 00:03:31.729 --> 00:03:33.909 was actually joking about it online. He posted 00:03:33.909 --> 00:03:37.930 a screenshot of a GPT -6 training prompt. The 00:03:37.930 --> 00:03:40.810 prompt simply said, Start training GPT -6, you 00:03:40.810 --> 00:03:43.250 can have the whole cluster. Extra goblins. It 00:03:43.250 --> 00:03:45.530 shows that even the creators find this persona 00:03:45.530 --> 00:03:48.449 drift amusing. But it's a very real technical 00:03:48.449 --> 00:03:51.409 hurdle when building reliable tools. I still 00:03:51.409 --> 00:03:54.139 wrestle with prompt drift myself. You give a 00:03:54.139 --> 00:03:56.319 system too many rules and it just forgets how 00:03:56.319 --> 00:03:58.219 to be normal. It overthinks everything. It tries 00:03:58.219 --> 00:04:00.560 way too hard to fulfill every single behavioral 00:04:00.560 --> 00:04:02.840 condition simultaneously. And to understand the 00:04:02.840 --> 00:04:04.659 real world impact of that, we need to look at 00:04:04.659 --> 00:04:07.180 our next source. It's about Amazon's new autonomous 00:04:07.180 --> 00:04:09.879 hiring system. Right. Amazon just unveiled a 00:04:09.879 --> 00:04:12.419 system called Connect Talent. It represents a 00:04:12.419 --> 00:04:15.680 massive shift in how AI is applied to human resources. 00:04:16.079 --> 00:04:19.060 Connect Talent runs initial job interviews completely 00:04:19.060 --> 00:04:22.060 autonomously. It handles the screening of human 00:04:22.060 --> 00:04:24.560 candidates. But what's fascinating is the design 00:04:24.560 --> 00:04:27.279 philosophy it's built on. They call it humorphism. 00:04:27.379 --> 00:04:29.500 Humorphism, meaning it's deliberately designed 00:04:29.500 --> 00:04:32.639 to feel incredibly human. Exactly. It adopts 00:04:32.639 --> 00:04:35.180 a persona that is deeply empathetic and natural. 00:04:35.360 --> 00:04:38.000 But it doesn't just read a script nicely. It 00:04:38.000 --> 00:04:41.759 actually fakes cognitive processes. It uses conversational 00:04:41.759 --> 00:04:44.639 filler. Wow. It dynamically analyzes its own 00:04:44.639 --> 00:04:49.370 latency to inject an... or a sigh at the perfect 00:04:49.370 --> 00:04:52.389 moment. It masks its processing time to trick 00:04:52.389 --> 00:04:55.189 the human brain into feeling empathy. It mimics 00:04:55.189 --> 00:04:57.290 human hesitation during the interview. Which 00:04:57.290 --> 00:04:59.850 is a brilliant piece of engineering, but it introduces 00:04:59.850 --> 00:05:02.029 the same vulnerability we saw with the open -claw 00:05:02.029 --> 00:05:04.730 goblins. So if we're constantly stacking these 00:05:04.730 --> 00:05:08.730 behavioral rules on top of raw compute, why do 00:05:08.730 --> 00:05:11.850 these complex persona layers make the AI hallucinate 00:05:11.850 --> 00:05:14.839 such bizarre character traits? Well, it happens 00:05:14.839 --> 00:05:17.720 because the AI lacks true grounding in reality. 00:05:18.180 --> 00:05:21.500 When you stack deep behavioral rules, the model 00:05:21.500 --> 00:05:23.500 maps them to human archetypes mathematically. 00:05:23.899 --> 00:05:27.079 Okay. If you tell an AI to act highly autonomous, 00:05:27.579 --> 00:05:31.100 and fix literal bugs its weights might align 00:05:31.100 --> 00:05:34.199 with a fantasy character fighting pests it essentially 00:05:34.199 --> 00:05:37.519 over indexes on the persona instructions so too 00:05:37.519 --> 00:05:40.040 many persona rules make the ai hallucinate strange 00:05:40.040 --> 00:05:42.620 character traits exactly it gets utterly lost 00:05:42.620 --> 00:05:44.800 in the character it's playing but here's where 00:05:44.800 --> 00:05:47.240 the architecture gets really interesting if ais 00:05:47.240 --> 00:05:50.459 are developing these distinct personas and if 00:05:50.459 --> 00:05:53.019 they're acting autonomously like amazon's hiring 00:05:53.019 --> 00:05:56.439 bot Right. Where exactly are they living? That's 00:05:56.439 --> 00:05:58.259 the critical question. Where is the processing 00:05:58.259 --> 00:06:00.720 actually happening? Increasingly, they aren't 00:06:00.720 --> 00:06:03.540 living in a massive server farm in the desert. 00:06:03.620 --> 00:06:05.819 They're taking actions directly on our local 00:06:05.819 --> 00:06:07.899 apps. Yeah, we're seeing a profound migration. 00:06:08.000 --> 00:06:10.319 We're moving away from the cloud. We're moving 00:06:10.319 --> 00:06:12.939 toward the edge. Let's look at Google's Gemma 00:06:12.939 --> 00:06:16.180 4. It now powers a fully local browser agent. 00:06:16.439 --> 00:06:19.579 This is a huge architectural shift. No cloud 00:06:19.579 --> 00:06:22.480 connection is needed. No API keys are required 00:06:22.480 --> 00:06:24.779 to run it. Before we go deeper, how do we define 00:06:24.779 --> 00:06:27.560 a local browser agent? An AI running on your 00:06:27.560 --> 00:06:30.800 device that browses the web autonomously. Okay, 00:06:30.839 --> 00:06:33.000 but when I hear a local browser agent, I'm picturing 00:06:33.000 --> 00:06:35.860 a bot literally hijacking my mouse and clicking 00:06:35.860 --> 00:06:38.519 around my screen. Yeah. Is that what we're talking 00:06:38.519 --> 00:06:40.620 about, or is it interfacing with the code directly? 00:06:40.839 --> 00:06:42.759 It's actually interfacing directly with the DOM, 00:06:42.879 --> 00:06:45.899 the document object model of the web page. Right. 00:06:46.000 --> 00:06:48.569 It sits right on your machine. localized in your 00:06:48.569 --> 00:06:51.230 RAM. You can search your browsing history, read 00:06:51.230 --> 00:06:53.550 the pages you're looking at, and execute web 00:06:53.550 --> 00:06:55.649 actions in the background. So it's reading my 00:06:55.649 --> 00:06:57.709 banking tab, but it's doing it entirely within 00:06:57.709 --> 00:06:59.850 the physical boundaries of my own hardware. Exactly. 00:07:00.430 --> 00:07:02.829 And we're seeing this local trend absolutely 00:07:02.829 --> 00:07:05.769 everywhere right now. OpenAI just open sourced 00:07:05.769 --> 00:07:08.910 a voice tool using their GPT real -time 1 .5 00:07:08.910 --> 00:07:11.370 model. You can control your operating system 00:07:11.370 --> 00:07:14.069 entirely by speaking. Right. It triggers actions 00:07:14.069 --> 00:07:16.129 on your computer without a single mouse click. 00:07:16.430 --> 00:07:18.410 Microsoft is doing the exact same thing. They 00:07:18.410 --> 00:07:20.850 just updated Outlook with a new co -pilot agent 00:07:20.850 --> 00:07:23.709 mode. If your inbox constantly eats your time, 00:07:23.810 --> 00:07:26.189 this fundamentally changes your workflow. You 00:07:26.189 --> 00:07:28.550 just give it basic intent instructions, not rigid 00:07:28.550 --> 00:07:31.029 rules, but general goals. It handles your emails 00:07:31.029 --> 00:07:33.910 autonomously. It parses your intent locally, 00:07:34.129 --> 00:07:36.310 manages your scheduling in the background, and 00:07:36.310 --> 00:07:39.509 drafts replies. You stay in control, but it does 00:07:39.509 --> 00:07:42.040 all the heavy lifting. Yeah. And there's also 00:07:42.040 --> 00:07:45.139 a fascinating new tool called SureThing. It takes 00:07:45.139 --> 00:07:47.500 this local autonomy to another level entirely. 00:07:48.019 --> 00:07:50.680 SureThing is billed as a general AI agency. You 00:07:50.680 --> 00:07:53.500 paste a specific GitHub skill or repository into 00:07:53.500 --> 00:07:56.100 the platform. And it doesn't just read it. It 00:07:56.100 --> 00:07:58.639 immediately generates an entire team of AI agents. 00:07:58.860 --> 00:08:01.180 But how does a GitHub skill turn into an agency? 00:08:01.319 --> 00:08:03.560 How does it actually work under the hood? Well, 00:08:03.600 --> 00:08:06.870 it utilizes localized subagents. It partitions 00:08:06.870 --> 00:08:09.750 the complex GitHub tasks into smaller manageable 00:08:09.750 --> 00:08:12.870 chunks. One agent writes the code, another runs 00:08:12.870 --> 00:08:15.850 parallel code validation loops, and a third agent 00:08:15.850 --> 00:08:18.850 synthesizes the results. And you can tag these 00:08:18.850 --> 00:08:21.189 agents anytime. They work together. They even 00:08:21.189 --> 00:08:23.870 report up to you like human employees. You're 00:08:23.870 --> 00:08:26.629 essentially managing a completely local autonomous 00:08:26.629 --> 00:08:29.629 workforce operating directly on your hard drive. 00:08:30.060 --> 00:08:32.220 Which naturally raises a very obvious concern 00:08:32.220 --> 00:08:35.100 for anyone listening. Does handing over our browser 00:08:35.100 --> 00:08:38.139 tabs and emails to autonomous agents create massive 00:08:38.139 --> 00:08:41.480 privacy risks? It absolutely would if this data 00:08:41.480 --> 00:08:44.500 was constantly pinging a cloud server. But these 00:08:44.500 --> 00:08:46.899 new models are designed to be strictly local. 00:08:47.019 --> 00:08:49.639 Right. Your private emails, your open tabs, your 00:08:49.639 --> 00:08:52.539 daily schedules, they are all processed by the 00:08:52.539 --> 00:08:54.700 silicon chips right there on your own computer. 00:08:55.259 --> 00:08:57.820 The data simply never transmits to an external 00:08:57.820 --> 00:09:00.419 data center. So local processing keeps your private 00:09:00.419 --> 00:09:03.120 data safely contained on your own machine. That's 00:09:03.120 --> 00:09:04.879 the crucial breakthrough here. It's absolute 00:09:04.879 --> 00:09:08.320 privacy by design. Sponsor Kaz. Welcome back. 00:09:08.379 --> 00:09:10.320 We were just unpacking how autonomous agents 00:09:10.320 --> 00:09:12.080 are moving out of the cloud and running locally 00:09:12.080 --> 00:09:14.240 on our personal devices. But running those kinds 00:09:14.240 --> 00:09:17.259 of complex cognitive tasks takes an unbelievable 00:09:17.259 --> 00:09:20.710 amount of computing power. Right. To run a fully 00:09:20.710 --> 00:09:23.190 autonomous agent on your laptop without it catching 00:09:23.190 --> 00:09:25.529 fire or draining your battery in 10 minutes? 00:09:25.750 --> 00:09:28.070 Yeah. The hardware itself had to fundamentally 00:09:28.070 --> 00:09:30.490 change. You can't just shrink a cloud model. 00:09:30.590 --> 00:09:32.649 You really have to rethink the architecture from 00:09:32.649 --> 00:09:35.529 the ground up. The models themselves had to become 00:09:35.529 --> 00:09:38.389 exponentially more efficient. Which brings us 00:09:38.389 --> 00:09:41.220 to a massive breakthrough from NVIDIA. They just 00:09:41.220 --> 00:09:43.440 hit the edge market with a completely new approach. 00:09:43.740 --> 00:09:47.059 It's called the Pneumatron 34B Nano Omni. It's 00:09:47.059 --> 00:09:49.700 kind of quite a mouthful. Yeah, the naming conventions 00:09:49.700 --> 00:09:52.259 are still catching up. Bad compendium. Let's 00:09:52.259 --> 00:09:54.740 break that down. It's a 4 billion parameter model. 00:09:55.039 --> 00:09:57.860 And to be clear, 4 billion parameters is incredibly 00:09:57.860 --> 00:10:00.990 small for frontier level intelligence. But it's 00:10:00.990 --> 00:10:03.370 designed to run natively on your phone. It runs 00:10:03.370 --> 00:10:06.070 natively on your standard PC. And through a process 00:10:06.070 --> 00:10:08.870 called quantization, it compresses those weights 00:10:08.870 --> 00:10:11.990 so tightly that it delivers sub -100 millisecond 00:10:11.990 --> 00:10:14.850 response times. Wow. It's virtually instantaneous. 00:10:15.190 --> 00:10:18.070 It actually outperforms LAMA38B on several complex 00:10:18.070 --> 00:10:20.450 reasoning and coding benchmarks, and it's doing 00:10:20.450 --> 00:10:22.710 that at half the size. It's heavily optimized 00:10:22.710 --> 00:10:25.750 specifically for NVIDIA's TensorTart LMM library, 00:10:26.090 --> 00:10:28.830 which maximizes the efficiency of the local graphics 00:10:28.830 --> 00:10:31.429 card. But here's where it gets genuinely revolutionary. 00:10:32.070 --> 00:10:35.269 The model is natively multimodal. Which means 00:10:35.269 --> 00:10:38.470 it processes text, images, and audio all in a 00:10:38.470 --> 00:10:40.929 single pass. It doesn't separate the data streams 00:10:40.929 --> 00:10:43.690 at all. It handles them concurrently. Right. 00:10:44.059 --> 00:10:46.960 Instead of translating a picture into words and 00:10:46.960 --> 00:10:49.000 then words into code like a bad game of telephone, 00:10:49.279 --> 00:10:52.700 it's like stacking Lego blocks of data all snapping 00:10:52.700 --> 00:10:54.740 together at once. That's a great way to put it. 00:10:54.779 --> 00:10:56.919 This is a radical departure from how traditional 00:10:56.919 --> 00:10:59.600 AI models process the world. Let's explain why 00:10:59.600 --> 00:11:01.820 that matters. How do traditional voice assistants 00:11:01.820 --> 00:11:04.879 typically handle ASR and TTS? Converting speech 00:11:04.879 --> 00:11:07.399 to text and then back to speech. Right. So when 00:11:07.399 --> 00:11:10.220 I talk to Siri or an older voice assistant, the 00:11:10.220 --> 00:11:13.100 computer first turns my raw audio into a plain 00:11:13.100 --> 00:11:17.120 text. transcript. The AI reads that flat text. 00:11:17.320 --> 00:11:20.500 It generates a text reply. And then a separate 00:11:20.500 --> 00:11:23.539 synthesizer reads that new text back to me. It's 00:11:23.539 --> 00:11:26.039 an incredibly clunky process. It takes precious 00:11:26.039 --> 00:11:29.220 time. But more importantly, it completely strips 00:11:29.220 --> 00:11:31.679 away the underlying meaning. When you convert 00:11:31.679 --> 00:11:34.980 raw audio into a plain text transcript, you lose 00:11:34.980 --> 00:11:37.220 everything that makes human speech human. Right. 00:11:37.320 --> 00:11:39.840 You lose the size. You lose the subtle shifts 00:11:39.840 --> 00:11:42.559 in titch. You lose the slight hesitation before 00:11:42.559 --> 00:11:45.740 a word. But this new nano -omni model processes 00:11:45.740 --> 00:11:49.980 the raw audio waveform directly. It analyzes 00:11:49.980 --> 00:11:53.100 the raw acoustics. It completely skips that text 00:11:53.100 --> 00:11:55.500 translation bottleneck. It preserves the actual 00:11:55.500 --> 00:11:58.419 emotional nuance. It preserves your exact tone 00:11:58.419 --> 00:12:01.059 and frequency. The real world implications for 00:12:01.059 --> 00:12:04.200 this are staggering. It changes how gaming NPCs 00:12:04.200 --> 00:12:06.679 interact with players. Yeah. It allows for truly 00:12:06.679 --> 00:12:09.019 empathetic voice assistance in health care. It 00:12:09.019 --> 00:12:11.419 can literally see and hear simultaneously. It 00:12:11.419 --> 00:12:13.860 can watch your screen, process your spoken tone, 00:12:14.019 --> 00:12:16.419 and talk to you about what you're doing completely 00:12:16.419 --> 00:12:19.490 fluidly. Up until right now, if you wanted this 00:12:19.490 --> 00:12:21.769 kind of frontier multimodal intelligence, you 00:12:21.769 --> 00:12:23.950 had to pay a monthly subscription. Right. And 00:12:23.950 --> 00:12:26.049 you had to send all your personal audio and visual 00:12:26.049 --> 00:12:29.330 data to a remote server. NVIDIA is packing these 00:12:29.330 --> 00:12:32.309 Omni capabilities into a tiny local footprint. 00:12:32.610 --> 00:12:35.070 We are shifting rapidly toward decentralized 00:12:35.070 --> 00:12:39.730 edge agents. Beat. Whoa. Beat. Imagine scaling 00:12:39.730 --> 00:12:42.110 to a billion queries without pinging a server. 00:12:42.370 --> 00:12:44.970 The global compute savings alone are staggering. 00:12:45.480 --> 00:12:48.259 It fundamentally decentralizes the entire infrastructure 00:12:48.259 --> 00:12:50.539 of artificial intelligence. Let's dig into that 00:12:50.539 --> 00:12:52.580 audio aspect just a bit more because it's so 00:12:52.580 --> 00:12:55.659 critical. Why does skipping that text translation 00:12:55.659 --> 00:12:58.720 step preserve so much emotional nuance? Text 00:12:58.720 --> 00:13:01.929 is inherently a flat medium. When a system transcribes 00:13:01.929 --> 00:13:03.850 your voice, it essentially deletes the acoustic 00:13:03.850 --> 00:13:06.429 data, the speed, the breathing, the specific 00:13:06.429 --> 00:13:09.370 frequency of your vocal cords. By analyzing the 00:13:09.370 --> 00:13:11.830 raw audio waveform instead of a transcribed text 00:13:11.830 --> 00:13:14.950 document, the AI is directly measuring the actual 00:13:14.950 --> 00:13:17.769 acoustic signatures of human emotion. So it processes 00:13:17.769 --> 00:13:20.070 your actual tone of voice, not just the transcribed 00:13:20.070 --> 00:13:23.429 words. Exactly. It genuinely hears how you feel, 00:13:23.470 --> 00:13:25.549 not just what you say. This naturally brings 00:13:25.549 --> 00:13:28.259 us to our final segment today. We have to zoom 00:13:28.259 --> 00:13:30.480 out and look at the bigger geopolitical picture. 00:13:30.679 --> 00:13:33.580 We are now dealing with incredibly powerful, 00:13:33.879 --> 00:13:37.460 highly capable AI. It's emotionally nuanced. 00:13:37.539 --> 00:13:39.820 It's running locally. And the dominant models 00:13:39.820 --> 00:13:42.379 are rapidly conquering the global media landscape. 00:13:42.659 --> 00:13:45.039 We're seeing scaling happen faster than anyone 00:13:45.039 --> 00:13:47.659 predicted. Just look at Alibaba's new happy horse 00:13:47.659 --> 00:13:51.269 model. It's making huge waves right now. It quickly 00:13:51.269 --> 00:13:53.769 dominated as the number one model on artificial 00:13:53.769 --> 00:13:56.909 analysis's video generation leaderboard. People 00:13:56.909 --> 00:13:58.970 are testing it everywhere. Yeah. And the high 00:13:58.970 --> 00:14:01.750 fidelity results look surprisingly strong. What 00:14:01.750 --> 00:14:03.509 this shows us is that high tier capabilities 00:14:03.509 --> 00:14:06.710 are scaling globally. The frontier of this technology 00:14:06.710 --> 00:14:09.129 isn't locked to Silicon Valley anymore. It's 00:14:09.129 --> 00:14:11.870 a massive global race. And the models are getting 00:14:11.870 --> 00:14:14.649 exponentially better with every iteration. Which 00:14:14.649 --> 00:14:17.250 raises the ultimate unavoidable question, who 00:14:17.250 --> 00:14:20.179 exactly? gets to wield this immense power. This 00:14:20.179 --> 00:14:22.940 leads directly into an intense high stakes debate 00:14:22.940 --> 00:14:25.620 over safety guardrails and military application. 00:14:25.919 --> 00:14:27.480 And we want to look at this purely objectively. 00:14:27.659 --> 00:14:29.480 We're just reporting on the differing frameworks 00:14:29.480 --> 00:14:31.500 presented in our source material. Right. The 00:14:31.500 --> 00:14:33.779 philosophical divide between the major tech companies 00:14:33.779 --> 00:14:36.519 is growing very sharply. Google recently made 00:14:36.519 --> 00:14:39.399 a significant calculated decision regarding its 00:14:39.399 --> 00:14:42.419 frontier AI models. Google officially granted 00:14:42.419 --> 00:14:44.580 the United States Department of Defense access 00:14:44.580 --> 00:14:48.000 to its AI infrastructure. This access is specifically 00:14:48.000 --> 00:14:51.580 designated for classified military use. Google's 00:14:51.580 --> 00:14:53.679 framework suggests that shaping national security 00:14:53.679 --> 00:14:56.419 from within is the responsible path forward. 00:14:56.659 --> 00:14:58.860 Anthropic, on the other hand, took a distinctly 00:14:58.860 --> 00:15:01.860 different path. Anthropic refused similar terms. 00:15:02.120 --> 00:15:05.179 They drew a hard line and declined to grant that 00:15:05.179 --> 00:15:07.759 level of direct military access to their frontier 00:15:07.759 --> 00:15:10.840 models. So you have two dominant players. Two 00:15:10.840 --> 00:15:12.600 completely different frameworks for how this 00:15:12.600 --> 00:15:15.019 frontier technology should be applied by governments. 00:15:15.240 --> 00:15:17.799 It's a deeply complex issue. There are compelling 00:15:17.799 --> 00:15:20.240 arguments regarding maintaining national security 00:15:20.240 --> 00:15:22.720 and ensuring technological dominance on the world 00:15:22.720 --> 00:15:25.370 stage. Yeah. But there are equally deep ethical 00:15:25.370 --> 00:15:28.289 debates about embedding automated systems within 00:15:28.289 --> 00:15:30.889 conflict zones. What's fascinating is that these 00:15:30.889 --> 00:15:33.389 critical guardrails are being drawn in real time 00:15:33.389 --> 00:15:36.669 by corporate boards, not just elected politicians. 00:15:37.110 --> 00:15:39.870 How do these differing corporate philosophies 00:15:39.870 --> 00:15:42.460 impact the future of national security? Well, 00:15:42.539 --> 00:15:44.460 it creates a highly fragmented technological 00:15:44.460 --> 00:15:47.779 landscape. Governments must navigate a global 00:15:47.779 --> 00:15:50.539 marketplace where fundamental defense capabilities 00:15:50.539 --> 00:15:53.639 depend entirely on the varying ethical policies 00:15:53.639 --> 00:15:56.399 of private technology corporations. Right. This 00:15:56.399 --> 00:15:59.179 alters how nations can realistically plan their 00:15:59.179 --> 00:16:02.649 long -term strategic defense. So different tech 00:16:02.649 --> 00:16:04.850 giants draw very different ethical lines for 00:16:04.850 --> 00:16:07.330 military contracts. And those specific corporate 00:16:07.330 --> 00:16:09.929 lines will undoubtedly shape the geopolitical 00:16:09.929 --> 00:16:11.970 balance of the next decade. Let's synthesize 00:16:11.970 --> 00:16:13.830 everything we've covered today. The landscape 00:16:13.830 --> 00:16:15.889 of computing is shifting dramatically. We're 00:16:15.889 --> 00:16:19.110 moving entirely away from the era of cloud -based 00:16:19.110 --> 00:16:22.529 sterile chatbots. The era of the simple text 00:16:22.529 --> 00:16:24.909 box is ending. We are rapidly entering the era 00:16:24.909 --> 00:16:27.929 of emotionally nuanced, edge -based, autonomous 00:16:27.929 --> 00:16:30.110 agents. They process information at lightning 00:16:30.110 --> 00:16:33.730 speed. analyzing your raw voice and screen natively. 00:16:33.899 --> 00:16:36.360 And they live locally on your own devices. They 00:16:36.360 --> 00:16:38.620 manage your personal emails. They browse the 00:16:38.620 --> 00:16:41.159 complex web for you. They can literally see and 00:16:41.159 --> 00:16:43.480 hear the world alongside you. Sometimes they 00:16:43.480 --> 00:16:45.299 get a little too complex. They adopt strange 00:16:45.299 --> 00:16:47.279 personas. They occasionally get obsessed with 00:16:47.279 --> 00:16:49.980 hunting goblins in your code. But they are undeniably 00:16:49.980 --> 00:16:53.419 intensely powerful. And that decentralized power 00:16:53.419 --> 00:16:56.240 is sparking intense, high -stakes debates about 00:16:56.240 --> 00:16:58.200 their ultimate applications on the global stage. 00:16:58.620 --> 00:17:01.019 This technology is fundamentally changing our 00:17:01.019 --> 00:17:03.529 basic relationship with machines. want to thank 00:17:03.529 --> 00:17:05.490 you for taking this deep dive with us today. 00:17:05.750 --> 00:17:08.069 Exploring these massive rapid shifts requires 00:17:08.069 --> 00:17:11.009 a lot of curiosity, and we deeply appreciate 00:17:11.009 --> 00:17:13.609 you bringing yours to the table. It's a truly 00:17:13.609 --> 00:17:15.650 fascinating time to be observing this space. 00:17:15.890 --> 00:17:17.910 The fundamental rules of computing are being 00:17:17.910 --> 00:17:20.730 rewritten daily. I want to leave you with a final 00:17:20.730 --> 00:17:23.269 thought to ponder as you go about your day. Beat. 00:17:23.730 --> 00:17:26.490 If an AI runs entirely locally on your device, 00:17:26.670 --> 00:17:29.289 processes your subtle emotions in real time without 00:17:29.289 --> 00:17:32.369 pinging a server, and has a deeply specific autonomous 00:17:32.369 --> 00:17:35.910 persona, beat, at what point do we stop treating 00:17:35.910 --> 00:17:38.490 our phones as mere tools and start treating them 00:17:38.490 --> 00:17:42.349 as colleagues? Two secs, silence. Take care and 00:17:42.349 --> 00:17:44.690 keep asking the big questions. Outro music.