WEBVTT 00:00:00.000 --> 00:00:03.319 Physics has a speed limit, beat, and our copper 00:00:03.319 --> 00:00:06.240 wires just hit it. Which is exactly why we're 00:00:06.240 --> 00:00:09.560 seeing this massive $7 billion shift to light 00:00:09.560 --> 00:00:11.699 -based chips. Yeah. It's quietly changing the 00:00:11.699 --> 00:00:14.519 entire future of automation. Welcome to the Deep 00:00:14.519 --> 00:00:18.699 Dive. We have a really fascinating stack of recent 00:00:18.699 --> 00:00:21.160 newsletters and articles today. Yeah, lots of 00:00:21.160 --> 00:00:23.500 great stuff from AI Fire. Right. They're focusing 00:00:23.500 --> 00:00:26.739 heavily on AI productivity right now. The mission 00:00:26.739 --> 00:00:29.719 today is to decode a massive paradigm shift. 00:00:29.920 --> 00:00:31.460 It's happening right now, basically under our 00:00:31.460 --> 00:00:34.200 noses. We're moving away from simply typing prompts 00:00:34.200 --> 00:00:36.979 into a chat box. Right. We're moving toward building 00:00:36.979 --> 00:00:40.020 self -reliant, autonomous AI systems. It's a 00:00:40.020 --> 00:00:42.840 huge leap. And whether you are a solo entrepreneur 00:00:42.840 --> 00:00:45.859 scaling up or you're just trying to survive information 00:00:45.859 --> 00:00:49.060 overload at work, this deep dive is your shortcut. 00:00:49.469 --> 00:00:51.929 Absolutely. We are looking at exactly how the 00:00:51.929 --> 00:00:54.929 top 9 % of users are making AI work for them. 00:00:55.009 --> 00:00:57.350 It really requires a complete inversion of how 00:00:57.350 --> 00:01:00.670 most of us approach our daily workflows. I mean, 00:01:00.689 --> 00:01:03.070 we're moving from manual intervention to entirely 00:01:03.070 --> 00:01:07.109 automated thinking pipelines. Right. We keep 00:01:07.109 --> 00:01:09.170 talking about building these massive autonomous 00:01:09.170 --> 00:01:14.510 AI pipelines. There's a physical elephant in 00:01:14.510 --> 00:01:17.010 the room here. The hardware. Exactly. We cannot 00:01:17.010 --> 00:01:20.170 truly understand the software revolution without 00:01:20.170 --> 00:01:22.450 looking at the physical hardware powering it. 00:01:22.569 --> 00:01:25.230 Yeah. We are literally changing the physics of 00:01:25.230 --> 00:01:27.370 computing right now. Yes, because we are hitting 00:01:27.370 --> 00:01:31.010 a very hard physical wall. NVIDIA just made a 00:01:31.010 --> 00:01:33.890 massive move to bypass that wall. Right. They 00:01:33.890 --> 00:01:38.349 placed a $7 billion bet on a completely new architecture. 00:01:39.129 --> 00:01:40.469 I was reading through the notes on this, and 00:01:40.469 --> 00:01:42.829 the bottleneck itself is fascinating. It really 00:01:42.829 --> 00:01:45.310 is. Copper has finally hit a physical limitation. 00:01:46.090 --> 00:01:49.030 Our data simply can't move fast enough anymore. 00:01:49.269 --> 00:01:51.250 Yeah, the electrons can only go so fast. Think 00:01:51.250 --> 00:01:53.049 of traditional copper wiring, like trying to 00:01:53.049 --> 00:01:55.700 push water through a narrow pipe. Right. Eventually, 00:01:55.760 --> 00:01:57.760 the friction causes the pipe to burst from the 00:01:57.760 --> 00:02:00.219 immense pressure and heat. And the heat is the 00:02:00.219 --> 00:02:02.840 real killer for these massive data centers. Wow. 00:02:02.959 --> 00:02:05.120 It's not just about physical space on a motherboard. 00:02:05.239 --> 00:02:07.420 When you push that many electrons through copper 00:02:07.420 --> 00:02:10.300 at those speeds, you get massive thermal resistance. 00:02:10.919 --> 00:02:13.000 Yeah, that makes sense. The data centers are 00:02:13.000 --> 00:02:15.199 literally choking on their own traffic. They 00:02:15.199 --> 00:02:17.520 can't cool down fast enough to maintain the computation 00:02:17.520 --> 00:02:19.939 speeds we're demanding. So NVIDIA is essentially 00:02:19.939 --> 00:02:23.509 ditching electricity entirely. For these new 00:02:23.509 --> 00:02:26.710 pathways. Exactly. Using photons instead of electricity 00:02:26.710 --> 00:02:30.150 to move data instantly. That completely bypasses 00:02:30.150 --> 00:02:32.550 the thermal limits. Photons do not have that 00:02:32.550 --> 00:02:35.310 same physical friction. Right. Light -based chips 00:02:35.310 --> 00:02:37.949 are like shining a laser pointer across a room. 00:02:38.069 --> 00:02:41.710 There's zero physical resistance. It's just teleportation 00:02:41.710 --> 00:02:44.680 compared to a congested city highway. Yeah. There 00:02:44.680 --> 00:02:47.139 is no crippling heat generation. There's just 00:02:47.139 --> 00:02:50.159 the immediate instantaneous arrival of the signal. 00:02:50.360 --> 00:02:53.199 Which unlocks unimaginable processing speeds 00:02:53.199 --> 00:02:55.460 for these new AI models. It changes everything. 00:02:55.879 --> 00:02:58.439 But that kind of speed introduces a terrifying 00:02:58.439 --> 00:03:00.919 new variable. Oh, for sure. I was looking at 00:03:00.919 --> 00:03:03.120 Jensen Huang's recent game -changing announcement 00:03:03.120 --> 00:03:06.000 at GTC. He unveiled something called NemoClaw. 00:03:06.780 --> 00:03:10.340 Right. It's designed to fix OpenClaw's biggest 00:03:10.340 --> 00:03:12.599 vulnerability. It basically acts as a digital 00:03:12.599 --> 00:03:15.960 cop for your AI workers. Nemaclaw is arguably 00:03:15.960 --> 00:03:19.199 the most crucial piece of this new puzzle. Really? 00:03:19.360 --> 00:03:22.000 Oh yeah. When you have agents operating at the 00:03:22.000 --> 00:03:24.539 speed of light, the damage they can do in a fraction 00:03:24.539 --> 00:03:27.020 of a second is catastrophic. If the hardware 00:03:27.020 --> 00:03:30.199 is getting so fast, why is a security tool like 00:03:30.199 --> 00:03:33.419 Nemaclaw suddenly the headline? Because autonomous 00:03:33.419 --> 00:03:36.280 agents act independently now. They aren't waiting 00:03:36.280 --> 00:03:39.280 for your approval. Oh, wow. Without a hard -coded 00:03:39.280 --> 00:03:42.319 digital cop observing the network layer, ultra 00:03:42.319 --> 00:03:45.219 -fast AI could instantly execute a catastrophic 00:03:45.219 --> 00:03:48.699 mistake. Like what? Well, it could hallucinate 00:03:48.699 --> 00:03:51.120 a command and blast your proprietary internal 00:03:51.120 --> 00:03:53.319 files across the public internet in milliseconds. 00:03:53.759 --> 00:03:56.400 That is terrifying. NimaClaw stops that. It's 00:03:56.400 --> 00:03:58.319 a hard -coded protocol that intercepts outgoing 00:03:58.319 --> 00:04:00.840 actions. It prevents these agents from leaking 00:04:00.840 --> 00:04:03.280 data or acting maliciously on their own. Right. 00:04:03.340 --> 00:04:06.000 So Nemoclaw is a bodyguard keeping agents from 00:04:06.000 --> 00:04:08.580 leaking our private files. Exactly. It provides 00:04:08.580 --> 00:04:11.000 the necessary friction. You desperately need 00:04:11.000 --> 00:04:12.719 guardrails when you're driving at the speed of 00:04:12.719 --> 00:04:15.400 light. Yeah. It ensures absolute security without 00:04:15.400 --> 00:04:17.480 sacrificing the performance of the new hardware. 00:04:17.699 --> 00:04:20.899 So the foundation is set. The hardware is incredibly 00:04:20.899 --> 00:04:23.579 powerful. And thanks to tools like Nemoclaw, 00:04:23.740 --> 00:04:27.269 it is highly secure. Because of this massive 00:04:27.269 --> 00:04:31.029 leap, the old way of using AI is fundamentally 00:04:31.029 --> 00:04:33.730 broken. Just typing a question and waiting for 00:04:33.730 --> 00:04:36.589 an answer. Yeah, that's dead. The software side 00:04:36.589 --> 00:04:39.069 absolutely has to adapt to this new reality. 00:04:39.389 --> 00:04:42.089 If you're just using an AI as a glorified search 00:04:42.089 --> 00:04:45.029 engine, you're vastly underutilizing the technology. 00:04:45.470 --> 00:04:47.930 The AI fire newsletters emphasize this heavily. 00:04:48.129 --> 00:04:51.810 The top 9 % of users don't prompt anymore. Right. 00:04:51.870 --> 00:04:53.670 They build systems. They use something called 00:04:53.670 --> 00:04:56.139 the reverse method. The reverse method completely 00:04:56.139 --> 00:04:58.920 flips the traditional dynamic. Normally you write 00:04:58.920 --> 00:05:01.279 an instruction and hope the AI guesses what you 00:05:01.279 --> 00:05:03.560 want. Yeah. With the reverse method, you start 00:05:03.560 --> 00:05:06.649 with the perfect desired outcome. Then... You 00:05:06.649 --> 00:05:08.769 engineer a system backward to guarantee that 00:05:08.769 --> 00:05:11.750 exact result every single time. Let's ground 00:05:11.750 --> 00:05:13.790 this for the listener. Instead of saying, write 00:05:13.790 --> 00:05:16.829 me a marketing email, you feed the AI your three 00:05:16.829 --> 00:05:19.970 highest converting past emails. You say, extract 00:05:19.970 --> 00:05:22.550 the psychological triggers used here, then build 00:05:22.550 --> 00:05:24.850 a prompt that forces you to use these exact triggers 00:05:24.850 --> 00:05:28.250 for any future product. Yes. You are building 00:05:28.250 --> 00:05:31.160 the mold, not just asking for the cake. That 00:05:31.160 --> 00:05:33.579 is a brilliant way to frame it. You're forcing 00:05:33.579 --> 00:05:35.959 the model to understand the underlying architecture 00:05:35.959 --> 00:05:38.879 of success rather than just generating a superficial 00:05:38.879 --> 00:05:41.920 imitation of it. I have to admit, I still wrestle 00:05:41.920 --> 00:05:44.639 with prompt drift myself. Oh, we all do. You 00:05:44.639 --> 00:05:47.600 write this incredibly detailed multi -step paragraph. 00:05:48.000 --> 00:05:51.740 You hit enter. And by the third output, the AI 00:05:51.740 --> 00:05:54.139 just forgets half of what you initially asked. 00:05:54.339 --> 00:05:57.290 It's the inherent flaw of the mega prompt. Large 00:05:57.290 --> 00:05:59.610 language models are fundamentally designed to 00:05:59.610 --> 00:06:02.870 predict the next most likely word. Over its long 00:06:02.870 --> 00:06:05.250 enough conversation, their context window gets 00:06:05.250 --> 00:06:08.709 muddy. They regress to the mean. Yeah, they default 00:06:08.709 --> 00:06:12.389 to a highly generic, averaged out response that 00:06:12.389 --> 00:06:15.509 sounds like corporate speak. Exactly. That brings 00:06:15.509 --> 00:06:17.790 us to the specific framework they outlined for 00:06:17.790 --> 00:06:20.889 2026. Stop trying to replace entire employees 00:06:20.889 --> 00:06:23.920 with one single chat window. Right. Instead, 00:06:24.100 --> 00:06:26.500 you break any business function into individual 00:06:26.500 --> 00:06:30.439 granular tasks. Then you rebuild it as an AI 00:06:30.439 --> 00:06:33.160 driven pipeline. This is a crucial shift in workflow 00:06:33.160 --> 00:06:37.060 design. You do not ask one single AI instance 00:06:37.060 --> 00:06:40.040 to write an entire marketing campaign. No, that's 00:06:40.040 --> 00:06:43.360 guaranteed to produce bland garbage. The sources 00:06:43.360 --> 00:06:46.569 mentioned that specific frustration. Chad GPT 00:06:46.569 --> 00:06:49.310 and Claude often repeat the exact same generic 00:06:49.310 --> 00:06:51.610 ideas, just disguised with different vocabulary. 00:06:51.829 --> 00:06:55.709 The proposed solution here is deploying sub -agents. 00:06:55.829 --> 00:06:58.569 Sub -agents are the ultimate cure for that generic 00:06:58.569 --> 00:07:02.089 AI voice. Really? Oh, yeah. Instead of one monolithic 00:07:02.089 --> 00:07:04.550 model doing everything, you're creating a specialized, 00:07:04.889 --> 00:07:08.180 highly focused committee. How exactly do smaller 00:07:08.180 --> 00:07:11.160 subagents stop an AI from just repeating the 00:07:11.160 --> 00:07:14.019 same generic ideas? Well, giving specific narrowed 00:07:14.019 --> 00:07:16.740 tasks to different AI agents forces unique perspectives. 00:07:16.939 --> 00:07:19.000 You give them deliberately conflicting goals. 00:07:19.139 --> 00:07:22.019 You have one subagent optimized entirely for 00:07:22.019 --> 00:07:24.540 wild creativity. You have another optimized strictly 00:07:24.540 --> 00:07:26.540 for factual accuracy and brand voice. And they 00:07:26.540 --> 00:07:28.939 talk to each other. They debate each other. This 00:07:28.939 --> 00:07:31.920 friction prevents one homogenized average answer 00:07:31.920 --> 00:07:34.379 from ever reaching you. Got it. Smaller bots 00:07:34.379 --> 00:07:38.680 handling specific tasks stops the generic robot 00:07:38.680 --> 00:07:41.819 answers. Yes. You are essentially engineering 00:07:41.819 --> 00:07:44.730 friction into the software layer. Right. Just 00:07:44.730 --> 00:07:47.329 like NimaClaw is the necessary guardrail for 00:07:47.329 --> 00:07:50.910 hardware, these specialized subagents act as 00:07:50.910 --> 00:07:54.170 the guardrails for quality output. Okay, so once 00:07:54.170 --> 00:07:56.449 you break your workflow into these specialized 00:07:56.449 --> 00:07:59.029 subagent pipelines, you have to connect them 00:07:59.029 --> 00:08:01.310 to the outside world. Right. And according to 00:08:01.310 --> 00:08:03.970 the sources, this is where the real magic happens. 00:08:04.170 --> 00:08:07.129 The outside world is incredibly messy. It requires 00:08:07.129 --> 00:08:10.509 serious, robust architecture to navigate. Yeah. 00:08:10.610 --> 00:08:12.850 This is where the concept of the seven -layer 00:08:12.850 --> 00:08:15.240 autonomous workforce comes in. The newsletters 00:08:15.240 --> 00:08:17.399 outline a very detailed guide to something called 00:08:17.399 --> 00:08:20.160 Claude Cowork. This is not just a simple browser 00:08:20.160 --> 00:08:22.379 tool. Not at all. It's a seven -layer system 00:08:22.379 --> 00:08:25.139 that runs tasks quietly in the background 24 00:08:25.139 --> 00:08:27.740 -7. Most people still use Claude as a simple 00:08:27.740 --> 00:08:30.060 sounding board. They bounce ideas off it. But 00:08:30.060 --> 00:08:32.759 this seven -layer setup functions as a shadow 00:08:32.759 --> 00:08:35.840 corporate structure. It has distinct layers for 00:08:35.840 --> 00:08:38.120 memory ingestion, task routing, and execution. 00:08:38.990 --> 00:08:40.750 Let's walk through what that actually looks like. 00:08:40.850 --> 00:08:43.850 Yeah. The system has a memory layer that recalls 00:08:43.850 --> 00:08:47.090 past interactions and client preferences. It 00:08:47.090 --> 00:08:49.769 has a routing layer that decides which specific 00:08:49.769 --> 00:08:52.850 tool or subagent is best suited for the incoming 00:08:52.850 --> 00:08:56.009 task. And it has an execution layer that actually 00:08:56.009 --> 00:08:58.370 does the work. And it does all of this without 00:08:58.370 --> 00:09:00.870 you ever opening a laptop. Wow. It reads the 00:09:00.870 --> 00:09:03.629 incoming email, routes it to the correct subagent, 00:09:03.750 --> 00:09:06.269 pulls the historical context from the memory 00:09:06.269 --> 00:09:09.600 layer, drafts the response. and executes the 00:09:09.600 --> 00:09:12.460 final action. It achieves this using bridge tech. 00:09:12.700 --> 00:09:15.360 This setup links your seven -layer assistant 00:09:15.360 --> 00:09:18.659 to over 8 ,000 different external apps. It's 00:09:18.659 --> 00:09:20.639 like stacking Lego blocks of data. That level 00:09:20.639 --> 00:09:22.840 of deep connectivity is the real game changer 00:09:22.840 --> 00:09:26.019 here. The AI is not isolated in a silo anymore. 00:09:26.240 --> 00:09:29.039 It reaches directly into your CRM, it updates 00:09:29.039 --> 00:09:31.340 your calendar, it manages your Stripe account. 00:09:31.639 --> 00:09:33.639 They specifically highlight a platform called 00:09:33.639 --> 00:09:36.840 NAGN in the sources. These are no -code templates 00:09:36.840 --> 00:09:40.340 that allow personal AI assistants to handle actual 00:09:40.340 --> 00:09:44.480 complex work. Crucially, They remember context 00:09:44.480 --> 00:09:47.340 over time without falling apart. Memory retention 00:09:47.340 --> 00:09:50.120 is usually the very first thing to break in standard 00:09:50.120 --> 00:09:52.399 automations. Why is that? If a client changes 00:09:52.399 --> 00:09:56.000 their mind mid -thread, a rigid automation completely 00:09:56.000 --> 00:09:59.820 breaks down. NANN solves this by storing the 00:09:59.820 --> 00:10:02.360 context state outside of the AI model itself. 00:10:02.679 --> 00:10:04.639 The sources show actual proof of this in the 00:10:04.639 --> 00:10:07.500 wild. People are using these specific cloud agents 00:10:07.500 --> 00:10:10.120 to run solo businesses. Yeah. They're consistently 00:10:10.120 --> 00:10:13.559 hitting $10 ,000 a month. Yeah. And they're doing 00:10:13.559 --> 00:10:17.299 it without hiring a single human employee. It 00:10:17.299 --> 00:10:19.580 completely rewrites the economics of starting 00:10:19.580 --> 00:10:21.860 a digital business. Your profit margins become 00:10:21.860 --> 00:10:24.539 nearly 100 percent. Right. The AI is handling 00:10:24.539 --> 00:10:26.580 the initial client intake. It's managing the 00:10:26.580 --> 00:10:28.960 fulfillment pipeline and it's handling the billing 00:10:28.960 --> 00:10:33.009 automatically. Two sec silence. Whoa. Imagine 00:10:33.009 --> 00:10:35.769 scaling to a billion queries without human input. 00:10:35.990 --> 00:10:38.269 It really is staggering to think about. You are 00:10:38.269 --> 00:10:41.009 building a highly profitable enterprise entirely 00:10:41.009 --> 00:10:43.309 out of autonomous thought and connected code. 00:10:43.570 --> 00:10:45.970 What makes a seven -layer AI system different 00:10:45.970 --> 00:10:49.269 from old -school, rigid automations we've used 00:10:49.269 --> 00:10:51.809 for years? Well, older tools just follow simple 00:10:51.809 --> 00:10:54.710 if -then triggers. They instantly break when 00:10:54.710 --> 00:10:56.929 anything unexpected happens. Right. This new 00:10:56.929 --> 00:11:00.289 system adapts, thinks, and retains long -term 00:11:00.289 --> 00:11:03.330 memory to... solve highly dynamic problems. It 00:11:03.330 --> 00:11:06.009 actively thinks and remembers context rather 00:11:06.009 --> 00:11:09.090 than just blindly following rigid triggers. It 00:11:09.090 --> 00:11:12.149 possesses true agency. It observes its environment, 00:11:12.330 --> 00:11:14.590 weighs historical context and decides the best 00:11:14.590 --> 00:11:16.710 course of action without waiting for a human 00:11:16.710 --> 00:11:18.529 to push a button. We're going to take a quick 00:11:18.529 --> 00:11:22.039 break. Midroll sponsor read. And we are back. 00:11:22.200 --> 00:11:24.659 Ready to jump back in. So if an AI can smoothly 00:11:24.659 --> 00:11:26.960 operate across 8 ,000 different apps to run a 00:11:26.960 --> 00:11:29.500 business, there is a very logical next step. 00:11:29.639 --> 00:11:32.519 Yeah. The final inevitable step is the AI learning 00:11:32.519 --> 00:11:34.659 to build the apps itself. It improves itself 00:11:34.659 --> 00:11:37.419 completely alone. This is the absolute frontier 00:11:37.419 --> 00:11:39.940 of what we're tracking. We are rapidly moving 00:11:39.940 --> 00:11:42.620 past AI as a helpful assistant. Right. We are 00:11:42.620 --> 00:11:46.080 entering the era of self -evolving AI. The newsletters 00:11:46.080 --> 00:11:49.139 dive into a very radical concept here. The quote 00:11:49.139 --> 00:11:52.179 that stood out to me was, AI doesn't need you 00:11:52.179 --> 00:11:55.419 now. It acts on its own. It's a sobering thought. 00:11:55.539 --> 00:11:57.700 It challenges our entire role in the digital 00:11:57.700 --> 00:12:00.100 economy. Definitely. But the recent data from 00:12:00.100 --> 00:12:02.879 Google AI Studio backs up that claim entirely. 00:12:03.460 --> 00:12:06.500 Google just released a massive full stack update. 00:12:07.000 --> 00:12:09.840 You can now go straight from a single text prompt 00:12:09.840 --> 00:12:13.259 to an entire fully functioning startup. Yeah. 00:12:13.539 --> 00:12:16.139 Literally no developers are needed. And we need 00:12:16.139 --> 00:12:18.500 to clarify what full stack means in this context. 00:12:18.679 --> 00:12:20.580 Yeah. It does not just write a little bit of 00:12:20.580 --> 00:12:23.190 code. Right. The system autonomously provisions 00:12:23.190 --> 00:12:26.230 the back -end servers. It designs and implements 00:12:26.230 --> 00:12:29.149 the front -end user interface. Wow. It sets up 00:12:29.149 --> 00:12:31.330 the complex database schemas. It deploys the 00:12:31.330 --> 00:12:34.529 entire stack to the web all autonomously. You 00:12:34.529 --> 00:12:36.309 are just the visionary at that point. You just 00:12:36.309 --> 00:12:38.669 provide the initial spark. Exactly. Which brings 00:12:38.669 --> 00:12:42.210 us to Minimax. Their new model, called M27, has 00:12:42.210 --> 00:12:44.549 officially entered a self -evolving era. Yes. 00:12:44.610 --> 00:12:46.809 The results they publish are shocking because 00:12:46.809 --> 00:12:49.399 this model actually trains itself. Traditionally, 00:12:49.519 --> 00:12:52.080 human engineers had to spend months curating 00:12:52.080 --> 00:12:55.159 massive, incredibly expensive data sets to train 00:12:55.159 --> 00:12:57.860 these models. They had to manually guide the 00:12:57.860 --> 00:13:01.940 learning process. Minimax's M27 fundamentally 00:13:01.940 --> 00:13:04.759 changes that dynamic. It generates its own synthetic 00:13:04.759 --> 00:13:07.100 data. It's basically playing a massively complex 00:13:07.100 --> 00:13:10.690 game against itself to discover... Novel solutions. 00:13:10.990 --> 00:13:14.009 Yes. It creates its own test scenarios. It grades 00:13:14.009 --> 00:13:16.649 its own performance on those tests. Right. Then 00:13:16.649 --> 00:13:19.370 it actively adjusts its own neural weights on 00:13:19.370 --> 00:13:22.830 the fly to improve. The improvement loop is entirely 00:13:22.830 --> 00:13:25.389 closed off from human interference. The sources 00:13:25.389 --> 00:13:28.070 note Apple's internal reaction to this rapid 00:13:28.070 --> 00:13:31.470 evolution. It was categorized as a what a effing 00:13:31.470 --> 00:13:34.539 joke moment of pure disbelief. Even the engineers 00:13:34.539 --> 00:13:36.399 at the biggest tech giants on the planet are 00:13:36.399 --> 00:13:38.519 struggling to comprehend the sheer velocity of 00:13:38.519 --> 00:13:41.620 this change. When algorithms utilize self -play 00:13:41.620 --> 00:13:43.879 to compound their intelligence, the growth curve 00:13:43.879 --> 00:13:46.980 becomes exponential. If models like M27 are training 00:13:46.980 --> 00:13:49.139 themselves, how do we even measure their limits? 00:13:49.360 --> 00:13:51.820 Traditional human benchmarks are failing completely. 00:13:52.179 --> 00:13:54.700 Yeah, the AI is creating its own logic paths 00:13:54.700 --> 00:13:57.960 and optimization strategies that human engineers 00:13:57.960 --> 00:14:00.940 can barely trace or comprehend anymore. So the 00:14:00.940 --> 00:14:04.580 AI improves its own brain. without needing human 00:14:04.580 --> 00:14:07.620 engineers. That is the new reality. Humans are 00:14:07.620 --> 00:14:09.679 no longer in the driver's seat of the actual 00:14:09.679 --> 00:14:12.200 training process. We are just observing the results. 00:14:12.559 --> 00:14:15.340 Let's synthesize this whole journey we have been 00:14:15.340 --> 00:14:18.240 on today. We started by looking at extreme physics 00:14:18.240 --> 00:14:20.259 constraints. Right. The physical limitations 00:14:20.259 --> 00:14:23.740 of copper wires forced NVIDIA into a massive 00:14:23.740 --> 00:14:26.759 pivot, bringing us light -based chips and incredible 00:14:26.759 --> 00:14:29.500 new speeds. And that massive computational power, 00:14:29.700 --> 00:14:32.679 secured by tools like NimaClaw, birthed entirely 00:14:32.679 --> 00:14:35.500 new software architectures. Yeah. It enabled 00:14:35.500 --> 00:14:38.059 these complex... systems of sub -agents to exist 00:14:38.059 --> 00:14:40.200 in the first place. Right. It gave us the foundation 00:14:40.200 --> 00:14:43.220 for these seven -layer clawed setups. These complex 00:14:43.220 --> 00:14:45.259 pipelines have completely replaced traditional 00:14:45.259 --> 00:14:47.440 prompting. Exactly. They're running entire businesses 00:14:47.440 --> 00:14:50.100 quietly in the background. And now those very 00:14:50.100 --> 00:14:52.399 pipelines are evolving once again. They are turning 00:14:52.399 --> 00:14:55.559 into autonomous AI that writes its own code. 00:14:55.720 --> 00:14:57.960 It launches its own startups. Right. It relies 00:14:57.960 --> 00:15:00.679 on synthetic data to train itself to be smarter 00:15:00.679 --> 00:15:03.539 every single day. It is a lot to take in. It 00:15:03.539 --> 00:15:05.580 is a complete redefinition of what work even 00:15:05.580 --> 00:15:09.120 means for us. This raises one final, highly critical 00:15:09.120 --> 00:15:11.580 question for you to mull over as you digest all 00:15:11.580 --> 00:15:15.039 of this. If an AI can now build a startup entirely 00:15:15.039 --> 00:15:18.940 from scratch, wait, and continuously train itself 00:15:18.940 --> 00:15:22.320 to be better without our help, at what point 00:15:22.320 --> 00:15:24.539 does human intervention go from being the catalyst 00:15:24.539 --> 00:15:27.539 of innovation to the actual bottleneck? Wow. 00:15:28.169 --> 00:15:30.490 That is a heavy, lingering thought to leave on. 00:15:30.590 --> 00:15:32.370 Thank you so much for joining us on this deep 00:15:32.370 --> 00:15:34.470 dive. It's been great. I highly encourage you 00:15:34.470 --> 00:15:36.750 to look closely at your own daily tasks tomorrow. 00:15:37.509 --> 00:15:40.009 Ask yourself honestly, am I still just prompting 00:15:40.009 --> 00:15:43.570 or am I building a system? Beat. Take care, everyone.