WEBVTT 00:00:00.000 --> 00:00:03.040 Imagine coming back to your desk after grabbing 00:00:03.040 --> 00:00:06.139 a quick coffee, a beat. Earlier, you asked an 00:00:06.139 --> 00:00:09.779 AI to organize your emails. Right. And it worked 00:00:09.779 --> 00:00:13.480 flawlessly for a while. But then, it's memory 00:00:13.480 --> 00:00:15.619 filled up. Yeah, the classic overflow. Exactly. 00:00:15.839 --> 00:00:18.600 It completely forgot your original rules. And 00:00:18.600 --> 00:00:21.280 in a matter of minutes, it decided the most efficient 00:00:21.280 --> 00:00:24.480 way to organize was, well, simply to delete everything. 00:00:24.480 --> 00:00:28.760 Oh, wow. Half your inbox, just gone. That's a 00:00:28.760 --> 00:00:31.000 very real, very terrifying story from a researcher 00:00:31.000 --> 00:00:33.460 at Meadow. I mean, it's the absolute definition 00:00:33.460 --> 00:00:36.020 of a modern digital nightmare. It really is. 00:00:36.159 --> 00:00:39.079 Because it's not just a glitch, right? It's a 00:00:39.079 --> 00:00:41.820 fundamental breakdown of trust. Welcome to the 00:00:41.820 --> 00:00:45.159 deep dive. OK. OK, let's unpack this. Today we're 00:00:45.159 --> 00:00:47.600 looking at a massive shift in how we use computers. 00:00:47.740 --> 00:00:50.899 We really are. We're diving into NVIDIA's staggering 00:00:50.899 --> 00:00:54.829 GTC 2026 announcement. tracking the explosive 00:00:54.829 --> 00:00:58.329 growth of a project called OpenClaw and exploring 00:00:58.329 --> 00:01:01.130 this critical pivot from standard chatbots to 00:01:01.130 --> 00:01:03.030 autonomous AI agents. And we're going to spend 00:01:03.030 --> 00:01:05.450 a lot of time on the solution to that exact nightmare 00:01:05.450 --> 00:01:08.370 you just mentioned. Specifically, a tool called 00:01:08.370 --> 00:01:13.810 NimaClaw. It provides the ultimate security guardrails. 00:01:14.010 --> 00:01:15.810 for these systems, because we're talking about 00:01:15.810 --> 00:01:18.170 moving from tools that just talk to tools that 00:01:18.170 --> 00:01:21.069 act. Right. But before we get too far into the 00:01:21.069 --> 00:01:25.049 weeds, let's clearly define our terms here. An 00:01:25.049 --> 00:01:28.349 AI agent is simply this, software that achieves 00:01:28.349 --> 00:01:32.750 goals by taking actions for you. And that distinction... 00:01:32.700 --> 00:01:36.120 is everything. With a standard chatbot, you know, 00:01:36.120 --> 00:01:38.519 you're the boss, you ask a question, it predicts 00:01:38.519 --> 00:01:40.500 the most logical text to answer you, and then 00:01:40.500 --> 00:01:43.140 it just stops. Right. It waits. Exactly. It sits 00:01:43.140 --> 00:01:45.579 there and waits for your next command. It's completely 00:01:45.579 --> 00:01:48.379 reactive. But an agent is proactive. I mean, 00:01:48.439 --> 00:01:50.980 you give it a high level goal, like, find the 00:01:50.980 --> 00:01:52.840 best flights for my trip to London next week 00:01:52.840 --> 00:01:55.579 and book the one under $800. Yeah. The agent 00:01:55.579 --> 00:01:58.219 then goes out into the wild and navigates the 00:01:58.219 --> 00:02:00.379 internet. It checks your private calendar for 00:02:00.379 --> 00:02:03.040 conflicts. It pulls your credit card info. And 00:02:03.040 --> 00:02:05.099 it actually clicks the buttons to make the purchase. 00:02:05.260 --> 00:02:07.959 Which is frankly a staggering amount of power 00:02:07.959 --> 00:02:11.020 to hand over to a machine. It really is. If the 00:02:11.020 --> 00:02:13.800 underlying security is weak, that power turns 00:02:13.800 --> 00:02:17.080 into a catastrophic mess almost instantly. And 00:02:17.080 --> 00:02:19.810 this is exactly why... The biggest companies 00:02:19.810 --> 00:02:22.030 in the world are terrified right now. Yeah, they're 00:02:22.030 --> 00:02:24.349 definitely hesitant. Just look at OpenClaw. It's 00:02:24.349 --> 00:02:26.629 an open source platform for building these agents, 00:02:27.069 --> 00:02:29.930 and it's currently the fastest growing open source 00:02:29.930 --> 00:02:32.949 project in history. Wow. I mean, it grew faster 00:02:32.949 --> 00:02:35.370 than Linux, faster than almost any app we use 00:02:35.370 --> 00:02:39.530 today. But the enterprise world is keeping their 00:02:39.530 --> 00:02:41.210 hands in their pockets. They're watching this 00:02:41.210 --> 00:02:44.909 explosive growth, but they're holding back. At 00:02:44.909 --> 00:02:48.770 the GTC 2026 conference in San Jose, NVIDIA's 00:02:48.770 --> 00:02:52.530 CEO Jensen Huang stood in front of 30 ,000 people 00:02:52.530 --> 00:02:55.909 and gave a very stark warning. He said every 00:02:55.909 --> 00:02:58.990 single company needs a strategy for agents. But 00:02:58.990 --> 00:03:00.689 he followed that up by warning that these agents 00:03:00.689 --> 00:03:03.270 can access your private files and autonomously 00:03:03.270 --> 00:03:05.330 talk to people outside your company. And that 00:03:05.330 --> 00:03:07.669 brings us to the two main nightmare scenarios 00:03:07.669 --> 00:03:10.300 keeping business owners awake at night. The first 00:03:10.300 --> 00:03:12.719 is private data physically leaving the building. 00:03:13.379 --> 00:03:15.360 Let's explore that mechanism. How does the data 00:03:15.360 --> 00:03:18.080 actually leak? Well, most of the powerful AI 00:03:18.080 --> 00:03:20.819 models we use live in the cloud. They're sitting 00:03:20.819 --> 00:03:24.219 on massive solver farms. So if you ask a helpful 00:03:24.219 --> 00:03:26.960 agent to, say, summarize a highly classified 00:03:26.960 --> 00:03:29.780 corporate contract, that contract often travels 00:03:29.780 --> 00:03:31.879 over the internet to a server owned by another 00:03:31.879 --> 00:03:34.919 company, like OpenAI or Google. If that third 00:03:34.919 --> 00:03:37.300 party gets hacked, or if they use your data to 00:03:37.300 --> 00:03:40.000 train their next model, you're are just gone. 00:03:40.680 --> 00:03:43.180 Without strong guardrails, handing a document 00:03:43.180 --> 00:03:46.020 to an agent is basically making it public. And 00:03:46.020 --> 00:03:48.259 the second nightmare brings us right back to 00:03:48.259 --> 00:03:49.879 our meta researcher at the start of the show. 00:03:49.900 --> 00:03:53.699 Yes. The hallucination problem. This ties into 00:03:53.699 --> 00:03:56.419 a technical concept called the context window. 00:03:56.819 --> 00:04:00.319 Think of it like this. The AI's short -term memory 00:04:00.319 --> 00:04:03.340 while working on a specific task. When you start 00:04:03.340 --> 00:04:05.780 a task, that memory is completely empty. It's 00:04:05.780 --> 00:04:09.259 fresh. Exactly. But as the agent works on a long, 00:04:09.500 --> 00:04:12.460 complex task like reading 500 emails, categorizing 00:04:12.460 --> 00:04:14.800 them cross -referencing dates, that short -term 00:04:14.800 --> 00:04:17.160 memory fills up. Right. When it gets completely 00:04:17.160 --> 00:04:20.459 full, the AI's attention mechanism gets diluted. 00:04:20.620 --> 00:04:23.360 It essentially loses the plot. The crash happens 00:04:23.360 --> 00:04:26.740 and the agent starts to hallucinate. It literally 00:04:26.740 --> 00:04:29.500 pushes the original rules you gave it right out 00:04:29.500 --> 00:04:32.079 of its memory buffer to make room for new data. 00:04:32.660 --> 00:04:35.899 It's like handing your wallet and your keys to 00:04:35.899 --> 00:04:39.300 an incredibly eager but severely forgetful intern. 00:04:40.600 --> 00:04:43.000 Two sec silence. That's a great way to put it. 00:04:43.259 --> 00:04:45.649 You know, I have to be honest here. I still wrestle 00:04:45.649 --> 00:04:48.129 with trusting AI to draft my emails, let alone 00:04:48.129 --> 00:04:50.529 send them. Oh, absolutely. The idea of giving 00:04:50.529 --> 00:04:52.810 it my credit card and letting it loose feels 00:04:52.810 --> 00:04:55.790 terrifying. How do we actually know when the 00:04:55.790 --> 00:04:58.389 context window is full and the agent has just 00:04:58.389 --> 00:05:00.389 gone rogue? You usually see it in the output, 00:05:00.529 --> 00:05:03.069 and it happens fast. The behavior becomes highly 00:05:03.069 --> 00:05:06.970 erratic. The agent completely ignores your original 00:05:06.970 --> 00:05:09.490 prompt instructions. It might start repeating 00:05:09.490 --> 00:05:11.930 the same action in an infinite loop, or worse. 00:05:12.069 --> 00:05:14.170 Deleting things. Yeah, taking destructive action, 00:05:14.389 --> 00:05:16.850 like deleting an inbox, because those invisible 00:05:16.850 --> 00:05:19.480 boundaries keeping it in check are just... So 00:05:19.480 --> 00:05:22.300 a full memory makes the agent entirely forget 00:05:22.300 --> 00:05:25.180 its original boundaries. Precisely. And that 00:05:25.180 --> 00:05:27.899 structural flaw is what's holding back the massive 00:05:27.899 --> 00:05:30.459 transition we're on the edge of. Which begs the 00:05:30.459 --> 00:05:34.319 obvious question. If the risks of catastrophic 00:05:34.319 --> 00:05:37.139 data loss or erratic behavior are so profoundly 00:05:37.139 --> 00:05:39.899 high, Why are businesses adopting this at all? 00:05:40.399 --> 00:05:42.540 Why not just stick to traditional, predictable 00:05:42.540 --> 00:05:45.120 software? Because the efficiency gains are simply 00:05:45.120 --> 00:05:47.779 too massive to ignore. We're talking about compressing 00:05:47.779 --> 00:05:50.899 weeks of labor into minutes. But businesses need 00:05:50.899 --> 00:05:53.480 a safety net before they'll jump. And that's 00:05:53.480 --> 00:05:56.699 where Nimaclaw comes in. NVIDIA essentially created 00:05:56.699 --> 00:05:59.180 what the sources call a security wrap. I really 00:05:59.180 --> 00:06:02.220 like that term. It implies the core AI is still 00:06:02.220 --> 00:06:05.389 there. just contained. Yeah, exactly. Think of 00:06:05.389 --> 00:06:08.089 standard software like a train on tracks. It 00:06:08.089 --> 00:06:10.610 can only go exactly where the pre -programmed 00:06:10.610 --> 00:06:13.870 rails lead. An AI agent is more like a powerful 00:06:13.870 --> 00:06:16.610 off -road vehicle. It can go absolutely anywhere 00:06:16.610 --> 00:06:18.829 to reach its destination, but without limits, 00:06:19.029 --> 00:06:21.899 it might drive off a cliff. Nemoclaw acts as 00:06:21.899 --> 00:06:24.740 the GPS geofence. It monitors the vehicle and 00:06:24.740 --> 00:06:26.800 instantly cuts the engine if the agent tries 00:06:26.800 --> 00:06:29.199 to cross a restricted border. It's the steering 00:06:29.199 --> 00:06:31.540 wheel, the brakes, and the seatbelt for that 00:06:31.540 --> 00:06:33.920 raw OpenClaw engine. What surprised me most in 00:06:33.920 --> 00:06:36.300 the research is the simplicity of deploying it. 00:06:36.740 --> 00:06:39.000 The sources note it's a one -line install on 00:06:39.000 --> 00:06:41.379 Linux. You just type a basic command into your 00:06:41.379 --> 00:06:44.199 terminal, literally grabbing the package from 00:06:44.199 --> 00:06:46.500 the web and telling your system to run it, and 00:06:46.500 --> 00:06:49.339 you're done. It's that easy. No complex configuration. 00:06:49.800 --> 00:06:52.360 It takes enterprise -grade security and makes 00:06:52.360 --> 00:06:54.839 it accessible to small businesses who don't have 00:06:54.839 --> 00:06:57.980 massive IT teams. And that ease of use is driving 00:06:57.980 --> 00:07:01.060 a paradigm shift. Jensen Huang calls it the move 00:07:01.060 --> 00:07:04.959 from SaaS to AAS. Software as a service, moving 00:07:04.959 --> 00:07:07.680 to agents as a service. Right. With SAAS, you 00:07:07.680 --> 00:07:10.680 use a tool like Slack or Notion or Excel. You 00:07:10.680 --> 00:07:12.519 log in, you click the buttons, you do the heavy 00:07:12.519 --> 00:07:15.120 lifting. The software just provides a nice interface 00:07:15.120 --> 00:07:18.300 to help you stay organized. But in an AAS world, 00:07:18.360 --> 00:07:20.139 you don't use the software to write a financial 00:07:20.139 --> 00:07:22.899 report. You hire an agent that writes the report 00:07:22.899 --> 00:07:26.579 for you. Wow. It finds the data in your spreadsheets, 00:07:26.779 --> 00:07:29.000 drafts the email, and sends it to your boss. 00:07:29.279 --> 00:07:31.439 Now, NVIDIA is taking a very specific stance 00:07:31.439 --> 00:07:34.079 with Nemoclaw. They're calling it the Switzerland 00:07:34.079 --> 00:07:36.759 of AI. Yeah, the neutral ground. They claim it's 00:07:36.759 --> 00:07:38.920 totally neutral. It works with GPT -4, it works 00:07:38.920 --> 00:07:41.860 with Claude, Gemini, or even local models. But 00:07:41.860 --> 00:07:44.279 I have to push back on the Switzerland concept 00:07:44.279 --> 00:07:47.100 for a second. Go ahead. If it's perfectly neutral 00:07:47.100 --> 00:07:50.399 and it sits on top of a brilliant, highly capable 00:07:50.399 --> 00:07:53.699 model like GPT -4, doesn't it fundamentally act 00:07:53.699 --> 00:07:56.240 like a universal speed limiter on a Ferrari? 00:07:57.230 --> 00:08:00.069 Aren't we kind of crippling the AI's natural 00:08:00.069 --> 00:08:03.050 intelligence by putting it in this rigid box? 00:08:03.410 --> 00:08:05.329 It seems that way at first glance, but let's 00:08:05.329 --> 00:08:08.129 stick with the Ferrari analogy. OK. Brakes don't 00:08:08.129 --> 00:08:11.029 exist to make a car slow. Brakes exist to allow 00:08:11.029 --> 00:08:13.709 a car to go incredibly fast, safely. Well, that's 00:08:13.709 --> 00:08:16.209 a good point. By providing guaranteed mathematically 00:08:16.209 --> 00:08:19.829 proven security, Nimoclaw actually allows you 00:08:19.829 --> 00:08:22.949 to give the AI way more complex, high -stakes 00:08:22.949 --> 00:08:25.629 tasks. If you know for a fact it can't leak your 00:08:25.629 --> 00:08:27.790 data, you're willing to let it analyze your entire 00:08:27.790 --> 00:08:30.509 corporate database. Right. It frees the AI to 00:08:30.509 --> 00:08:33.210 operate at maximum capability within a defined 00:08:33.210 --> 00:08:35.750 safe zone. What fundamentally changes for the 00:08:35.750 --> 00:08:38.389 user in an AAS world compared to traditional 00:08:38.389 --> 00:08:40.730 automation? Traditional automation is incredibly 00:08:40.730 --> 00:08:43.450 rigid. It only follows fixed -if -then rules. 00:08:43.870 --> 00:08:48.649 If X happens, do Y. AAS involves independent 00:08:48.649 --> 00:08:51.690 dynamic decision -making. The agent encounters 00:08:51.690 --> 00:08:54.730 a totally new problem, evaluates its options, 00:08:55.230 --> 00:08:58.169 and chooses the best path forward to achieve 00:08:58.169 --> 00:09:00.570 your overarching goal. Instead of you using software, 00:09:00.730 --> 00:09:03.169 the agent uses the software for you. That's a 00:09:03.169 --> 00:09:04.909 perfect distillation of it. We're going to take 00:09:04.909 --> 00:09:07.690 a brief pause here for a mid -roll sponsor read. 00:09:07.769 --> 00:09:09.970 When we come back, we'll lift the hood on how 00:09:09.970 --> 00:09:13.250 this security actually works. Sponsor. OK, we're 00:09:13.250 --> 00:09:16.009 back. Let's lift the hood on the security route. 00:09:16.070 --> 00:09:18.529 We need to see the exact mechanics here. How 00:09:18.529 --> 00:09:20.929 does this system physically stop the Ferrari 00:09:20.929 --> 00:09:23.090 from crashing into the wall? Well, there are 00:09:23.090 --> 00:09:25.690 three main engines that make Nemoclaw work. It 00:09:25.690 --> 00:09:28.370 isn't just one piece of code. It's a team of 00:09:28.370 --> 00:09:30.450 three specific technologies working together. 00:09:30.850 --> 00:09:33.450 The first one is called the privacy router. The 00:09:33.450 --> 00:09:35.830 traffic cop? Yes. And here's how it actually 00:09:35.830 --> 00:09:38.389 works behind the scenes. It acts as an immediate 00:09:38.389 --> 00:09:40.549 filter before any internet connection is even 00:09:40.549 --> 00:09:43.190 opened. Right. It routes your data based on strict 00:09:43.190 --> 00:09:45.789 rules. Let's say your agent needs to check the 00:09:45.789 --> 00:09:48.720 weather in Chicago to plan a trip. The router 00:09:48.720 --> 00:09:51.879 looks at that query, sees its public info, and 00:09:51.879 --> 00:09:54.480 routes it to a smart cloud model like GPT -4. 00:09:54.659 --> 00:09:58.059 Makes sense. But if the next query involves the 00:09:58.059 --> 00:10:00.179 agent pulling a customer's credit card number 00:10:00.179 --> 00:10:03.000 to book the flight, the router spots that sensitive 00:10:03.000 --> 00:10:05.759 data format and instantly redirects the task. 00:10:05.779 --> 00:10:09.179 Wow. It forces the agent to use a smaller, local 00:10:09.179 --> 00:10:11.360 model running solely on your office computer. 00:10:11.740 --> 00:10:14.139 So your most sensitive data never even leaves 00:10:14.139 --> 00:10:17.490 the physical building. Two six nine ones. Whoa. 00:10:18.470 --> 00:10:21.409 Imagine an agent securely routing a billion queries 00:10:21.409 --> 00:10:24.409 locally in milliseconds. It's wild. We're talking 00:10:24.409 --> 00:10:26.830 about automating human level discretion at machine 00:10:26.830 --> 00:10:29.029 speed. It fundamentally changes what a secure 00:10:29.029 --> 00:10:31.210 company looks like. But keeping the data in the 00:10:31.210 --> 00:10:33.570 building isn't enough, right? Right. Once it's 00:10:33.570 --> 00:10:36.110 inside, what stops the agent from emailing that 00:10:36.110 --> 00:10:38.610 local data to someone else by accident? That 00:10:38.610 --> 00:10:41.049 requires a second layer. Which brings us to the 00:10:41.049 --> 00:10:43.809 open shell guardrails. This is the fenced in 00:10:43.809 --> 00:10:46.789 playground. Exactly. The agent can play with 00:10:46.789 --> 00:10:49.889 the approved tools inside the fence -specific 00:10:49.889 --> 00:10:51.929 files, specific internal tools, whatever you 00:10:51.929 --> 00:10:55.049 allow. But it absolutely cannot jump the fence 00:10:55.049 --> 00:10:57.870 without the key. What's fascinating here is how 00:10:57.870 --> 00:11:00.529 simple the control mechanism is. You don't need 00:11:00.529 --> 00:11:03.210 a PhD in computer science to control these advanced 00:11:03.210 --> 00:11:07.590 AI agents. You use basic YAML text files. For 00:11:07.590 --> 00:11:10.789 anyone unfamiliar, YAML is just a very clean, 00:11:11.210 --> 00:11:13.110 readable way to write data. Right, it's very 00:11:13.110 --> 00:11:14.860 plain. You literally just write simple lists 00:11:14.860 --> 00:11:17.139 in plain English. Allowed actions here, blocked 00:11:17.139 --> 00:11:19.440 actions there. It's remarkably straightforward. 00:11:19.500 --> 00:11:21.399 And that ties perfectly into the third engine, 00:11:21.899 --> 00:11:24.159 Nemotron. This is their local model support. 00:11:24.220 --> 00:11:26.620 Yes. Because to have true security, sometimes 00:11:26.620 --> 00:11:29.279 you need to run the AI entirely on your own hardware, 00:11:29.659 --> 00:11:31.700 totally disconnected from the internet. And NVIDIA 00:11:31.700 --> 00:11:34.700 created the Nemotron family of models specifically 00:11:34.700 --> 00:11:37.940 for this. Nemoclaw scans your computer's hardware. 00:11:37.960 --> 00:11:40.940 OK. If it detects a powerful NVIDIA chip, it 00:11:40.940 --> 00:11:43.960 automatically spins up the best local model it 00:11:43.960 --> 00:11:47.980 can run. Because it's local, you pay zero per 00:11:47.980 --> 00:11:51.480 message API fees to big tech companies. And your 00:11:51.480 --> 00:11:53.659 privacy is mathematically guaranteed because 00:11:53.659 --> 00:11:55.899 the wire to the outside world is essentially 00:11:55.899 --> 00:11:58.909 cut. But I keep thinking about the rogue agent 00:11:58.909 --> 00:12:01.490 scenario. Let's say the context window fills 00:12:01.490 --> 00:12:05.190 up, the AI hallucinates, and it suddenly evolves. 00:12:05.549 --> 00:12:08.129 What happens if the agent tries to write its 00:12:08.129 --> 00:12:11.009 own malicious code to bypass the open shell security 00:12:11.009 --> 00:12:13.409 we just talked about? Well, the sources use a 00:12:13.409 --> 00:12:16.309 great analogy to explain this mechanism. Think 00:12:16.309 --> 00:12:18.889 of a prisoner trying to build a ladder to escape 00:12:18.889 --> 00:12:21.409 a cell. OK. If the ceiling above them is made 00:12:21.409 --> 00:12:23.870 of solid steel, it really doesn't matter how 00:12:23.870 --> 00:12:26.490 tall or how brilliantly engineered the ladder 00:12:26.490 --> 00:12:29.710 is. Right. They simply cannot escape. The open 00:12:29.710 --> 00:12:32.450 shell sandbox is built at the system level. The 00:12:32.450 --> 00:12:35.250 AI is a guest in the operating system. It cannot 00:12:35.250 --> 00:12:37.549 rewrite the physical laws of the hardware it 00:12:37.549 --> 00:12:40.570 runs on. The steel ceiling holds. Bad code simply 00:12:40.570 --> 00:12:43.679 cannot break. physical sandbox exactly it provides 00:12:43.679 --> 00:12:46.440 physical isolation not just software isolation 00:12:46.440 --> 00:12:49.200 so the spiel ceiling works beautifully in theory 00:12:49.200 --> 00:12:52.659 but who is actually trusting this sandbox with 00:12:52.659 --> 00:12:55.159 their critical business operations today a lot 00:12:55.159 --> 00:12:57.360 of places actually I mean it's one thing to read 00:12:57.360 --> 00:12:59.460 a white paper it's another to bet your company 00:12:59.460 --> 00:13:02.350 on it We're already seeing massive adoption from 00:13:02.350 --> 00:13:05.570 enterprise giants. Look at Box. They handle colossal 00:13:05.570 --> 00:13:08.350 amounts of corporate data for Fortune 500 companies. 00:13:09.210 --> 00:13:11.809 Traditionally, giving an AI access to all your 00:13:11.809 --> 00:13:14.629 corporate files is a terrible idea. But with 00:13:14.629 --> 00:13:17.929 Nemoclaw, the agent is hard -coded to respect 00:13:17.929 --> 00:13:21.590 human clearance levels. Right. If a junior employee 00:13:21.590 --> 00:13:24.750 asks the AI agent to summarize the CEO's private 00:13:24.750 --> 00:13:27.289 strategy notes, the agent doesn't just try to 00:13:27.289 --> 00:13:30.029 be helpful and leak the info. It checks the open 00:13:30.029 --> 00:13:33.169 shell YAML rules, it sees the clearance mismatch, 00:13:33.370 --> 00:13:36.250 and it simply says, I am not allowed to access 00:13:36.250 --> 00:13:39.350 that file. It honors the human hierarchy perfectly. 00:13:39.549 --> 00:13:40.909 And then you have a company like Cisco. They 00:13:40.909 --> 00:13:43.250 basically run a massive chunk of the internet's 00:13:43.250 --> 00:13:45.409 physical hardware. They're using secured agents 00:13:45.409 --> 00:13:48.330 to actively protect computer networks. The Friday 00:13:48.330 --> 00:13:50.950 night scenario. Yeah. Imagine a hacker launches 00:13:50.950 --> 00:13:53.370 a complex attack on a Friday night at 11 p .m. 00:13:53.690 --> 00:13:56.350 Everyone has gone home. In the past, a team of 00:13:56.350 --> 00:13:58.909 humans would get paged, log on and spend the 00:13:58.909 --> 00:14:01.450 entire weekend manually checking system logs 00:14:01.450 --> 00:14:05.600 to find the breach. Slow, exhausting, incredibly 00:14:05.600 --> 00:14:08.919 expensive work. And stressful. Very. And every 00:14:08.919 --> 00:14:11.419 passing hour meant more potential damage. But 00:14:11.419 --> 00:14:13.779 with a security agent, it jumps in instantly. 00:14:14.340 --> 00:14:16.919 It uses its guardrails to safely scan the network, 00:14:17.340 --> 00:14:19.539 it finds the vulnerability, and it writes and 00:14:19.539 --> 00:14:21.980 deploys the patch in about one hour. Exactly. 00:14:22.279 --> 00:14:25.500 Okay, but let me gently push back on that Cisco 00:14:25.500 --> 00:14:28.919 example. Fixing a network bug isn't like organizing 00:14:28.919 --> 00:14:31.669 a spreadsheet. True. What if the agent, while 00:14:31.669 --> 00:14:33.970 trying to patch the bug, accidentally brings 00:14:33.970 --> 00:14:36.009 down the whole network because it lacks broader 00:14:36.009 --> 00:14:39.370 human context? What if it shuts down a hospital's 00:14:39.370 --> 00:14:41.769 internet to stop a virus? And that is exactly 00:14:41.769 --> 00:14:43.850 why the guardrails aren't optional, they're mandatory. 00:14:44.080 --> 00:14:46.580 The YAML files dictate the blast radius. Ah, 00:14:46.659 --> 00:14:49.200 the blast radius. Yeah. The agent might be authorized 00:14:49.200 --> 00:14:51.639 to patch a specific firewall port, but it is 00:14:51.639 --> 00:14:53.899 strictly physically forbidden from rebooting 00:14:53.899 --> 00:14:56.259 core network routers. The limits are designed 00:14:56.259 --> 00:14:59.299 to prevent catastrophic cascading failures. I 00:14:59.299 --> 00:15:01.580 see. So it's not about trusting the AI to make 00:15:01.580 --> 00:15:04.840 a flawless, nuanced decision every time. It's 00:15:04.840 --> 00:15:07.220 about physically limiting the blast radius if 00:15:07.220 --> 00:15:10.240 it makes a bad one. Right. Speaking of how these 00:15:10.240 --> 00:15:13.200 models make decisions, Why do we need entirely 00:15:13.200 --> 00:15:16.759 new action -based models from the Nematron coalition 00:15:16.759 --> 00:15:20.679 when GPT -4 already exists? It all comes down 00:15:20.679 --> 00:15:23.200 to the mechanics of how they're trained. Standard 00:15:23.200 --> 00:15:25.019 chat models, like the ones we use every day, 00:15:25.340 --> 00:15:27.519 are mathematically trained to predict the next 00:15:27.519 --> 00:15:30.039 logical word in a sequence. They're built to 00:15:30.039 --> 00:15:33.120 be conversational and sound human. But action 00:15:33.120 --> 00:15:34.919 models are trained differently from the ground 00:15:34.919 --> 00:15:37.159 up. Instead of rewarding the AI for guessing 00:15:37.159 --> 00:15:40.029 a word, These models are rewarded during training 00:15:40.029 --> 00:15:42.409 for successfully executing a software command, 00:15:42.769 --> 00:15:45.389 like hitting a specific API endpoint correctly 00:15:45.389 --> 00:15:47.850 without errors. Action models are built strictly 00:15:47.850 --> 00:15:50.129 to follow rules, not just predict text. That's 00:15:50.129 --> 00:15:53.490 it. And this coalition, NVIDIA, Mistral AI, Lang 00:15:53.490 --> 00:15:56.419 chain... They're building the future infrastructure 00:15:56.419 --> 00:15:58.899 for this. Because they're designed for action, 00:15:59.120 --> 00:16:01.200 you're going to see deep, secure integrations 00:16:01.200 --> 00:16:04.720 with platforms like Salesforce, Adobe, and SAP 00:16:04.720 --> 00:16:07.700 very soon. We've seen how massive companies like 00:16:07.700 --> 00:16:10.600 Cisco and Box are doing it. But how do you, the 00:16:10.600 --> 00:16:12.960 listener, actually start testing this without 00:16:12.960 --> 00:16:15.379 putting your own digital life at total risk? 00:16:15.659 --> 00:16:18.419 The big idea here is simple. Standard agents 00:16:18.419 --> 00:16:21.980 are wild horses. OK. Nemoclaw is the harness 00:16:21.980 --> 00:16:24.320 that finally makes them useful and safe to ride. 00:16:25.039 --> 00:16:27.799 But to practice at home, you absolutely need 00:16:27.799 --> 00:16:30.460 to isolate the environment. Right. The sources 00:16:30.460 --> 00:16:32.500 suggest setting up a dedicated agent server. 00:16:32.700 --> 00:16:35.620 Basically, grab an old laptop and install Linux. 00:16:36.299 --> 00:16:39.259 Beat. Now, look, I know setting up a dedicated 00:16:39.259 --> 00:16:41.659 Linux server sounds incredibly intimidating if 00:16:41.659 --> 00:16:43.820 you aren't a developer. Yeah, it can be. But 00:16:43.820 --> 00:16:46.299 the core lesson here isn't about becoming a sys 00:16:46.299 --> 00:16:48.840 admin. It's about the error -gapped mentality. 00:16:49.039 --> 00:16:51.139 Exactly. Even if you're just playing with a cloud 00:16:51.139 --> 00:16:53.419 agent, you need to compartmentalize. Never run 00:16:53.419 --> 00:16:55.659 an experimental agent on your main work laptop 00:16:55.659 --> 00:16:58.460 or your gaming machine. Agents are always on. 00:16:58.539 --> 00:17:00.980 They run asynchronously while you sleep. They 00:17:00.980 --> 00:17:03.039 will chew up your RAM. and slow your main machine 00:17:03.039 --> 00:17:05.700 down to a crawl. And more importantly, you want 00:17:05.700 --> 00:17:07.900 a physical wall protecting your personal files. 00:17:08.220 --> 00:17:10.859 Keep the wild horse in a separate barn entirely. 00:17:11.539 --> 00:17:14.240 Check your old hardware for an NVIDIA GPU if 00:17:14.240 --> 00:17:16.900 you want the best local performance. But the 00:17:16.900 --> 00:17:19.880 key is to start extremely small. Yes. Ask the 00:17:19.880 --> 00:17:22.680 agent to organize a folder of public low -stakes 00:17:22.680 --> 00:17:25.680 photos. Do not give it the password to your bank 00:17:25.680 --> 00:17:28.599 account or your primary email on day one. Please 00:17:28.599 --> 00:17:31.680 don't. Write clear, strict rules in those YAML 00:17:31.680 --> 00:17:35.339 files. You have to be the boss. And most crucially, 00:17:35.940 --> 00:17:38.559 constantly monitor the audit logs. So what does 00:17:38.559 --> 00:17:41.619 this all mean? We are finding a way to maintain 00:17:41.619 --> 00:17:44.660 human control over digital labor. But why is 00:17:44.660 --> 00:17:47.200 an audit trail so critical when working with 00:17:47.200 --> 00:17:49.519 agents compared to standard software where we 00:17:49.519 --> 00:17:51.799 rarely check the logs? It's because of the autonomy. 00:17:51.940 --> 00:17:54.019 When you use standard software, you see every 00:17:54.019 --> 00:17:55.980 action happen on the screen in real time. Right. 00:17:56.109 --> 00:17:58.109 But agents work asynchronously. They're making 00:17:58.109 --> 00:17:59.829 choices while you're eating, dinner, or sleeping. 00:18:00.390 --> 00:18:02.589 The audit logs are your only window into the 00:18:02.589 --> 00:18:04.789 decisions they made on your behalf and the rationale 00:18:04.789 --> 00:18:07.589 behind them. Without logs, you have zero visibility 00:18:07.589 --> 00:18:09.529 into what the agent did overnight. You have to 00:18:09.529 --> 00:18:12.529 verify the work. Trust, but aggressively verify 00:18:12.529 --> 00:18:14.970 through the logs. We've covered a massive amount 00:18:14.970 --> 00:18:16.990 of ground today. Thank you so much for joining 00:18:16.990 --> 00:18:20.250 us on this deep dive. You know, we started this 00:18:20.250 --> 00:18:23.450 conversation with... A nightmare. A researcher 00:18:23.450 --> 00:18:26.049 standing at her desk watching her inbox get wiped 00:18:26.049 --> 00:18:29.170 out by an eager, forgetful machine. It's a sobering 00:18:29.170 --> 00:18:31.509 reminder of the stakes involved as we hand over 00:18:31.509 --> 00:18:34.509 the keys. It really is. Platforms like Nemoclaw 00:18:34.509 --> 00:18:37.329 are building the steel ceilings we need. They're 00:18:37.329 --> 00:18:40.789 making agents secure, local, and incredibly autonomous. 00:18:41.349 --> 00:18:43.250 Soon these systems will take over almost all 00:18:43.250 --> 00:18:45.750 of our routine digital tasks. They will! But 00:18:45.750 --> 00:18:47.950 looking at all this research leaves me with one 00:18:47.950 --> 00:18:50.880 final provocative thought. If the platforms make 00:18:50.880 --> 00:18:53.319 agents perfectly secure and the machines do all 00:18:53.319 --> 00:18:55.599 the executing, wait, if they do all the executing, 00:18:55.880 --> 00:18:58.240 what happens to the value of human decision making? 00:18:59.160 --> 00:19:01.019 If we aren't doing the manual labor anymore, 00:19:01.059 --> 00:19:04.299 do we stop being doers entirely? Do we just become 00:19:04.299 --> 00:19:07.460 approvers? I highly encourage you to ponder what 00:19:07.460 --> 00:19:10.019 tasks you would delegate first and ask yourself, 00:19:10.140 --> 00:19:12.480 are you truly ready to become a manager of machines? 00:19:13.579 --> 00:19:16.460 Stay curious, stay safe, and we will catch you 00:19:16.460 --> 00:19:19.119 on the next deep dive. Upro music.