WEBVTT 00:00:00.000 --> 00:00:02.359 Imagine having the smartest supercomputer in 00:00:02.359 --> 00:00:05.700 the world right in your pocket, beat. It's ready 00:00:05.700 --> 00:00:08.500 to solve almost any problem. But, you know, you 00:00:08.500 --> 00:00:10.099 realize you're actually falling behind. Why? 00:00:11.019 --> 00:00:12.800 Well, because you simply don't know the right 00:00:12.800 --> 00:00:14.960 way to talk to it. Yeah, and the landscape is 00:00:14.960 --> 00:00:17.300 shifting so aggressively that old habits are 00:00:17.300 --> 00:00:20.679 already obsolete. So, welcome to today's deep 00:00:20.679 --> 00:00:23.219 dive. We've got a fascinating stack of updates 00:00:23.219 --> 00:00:25.870 to get through. We really do. OK, so let's unpack 00:00:25.870 --> 00:00:28.710 this. We're looking at a massive structural shift 00:00:28.710 --> 00:00:31.390 in artificial intelligence today. Our mission 00:00:31.390 --> 00:00:34.189 is to explore how models are literally shrinking 00:00:34.189 --> 00:00:37.450 to fit our personal devices. We'll also examine 00:00:37.450 --> 00:00:39.909 how a new skills gap is quietly dividing the 00:00:39.909 --> 00:00:42.289 modern workforce. And finally, we'll look at 00:00:42.289 --> 00:00:44.469 how major tech players are making ruthless bets 00:00:44.469 --> 00:00:46.869 on our physical future. The rules are being rewritten 00:00:46.869 --> 00:00:49.750 daily. It's incredibly fast paced. It truly is. 00:00:49.869 --> 00:00:51.590 So let's start with the physical constraints 00:00:51.590 --> 00:00:53.590 of AI. We all know these models are incredibly 00:00:53.590 --> 00:00:55.799 powerful. But running a large language model 00:00:55.799 --> 00:00:57.899 locally is exhausting for standard computers. 00:00:58.219 --> 00:01:01.420 Right. You typically need massive, highly expensive 00:01:01.420 --> 00:01:06.500 GPUs and you need enormous amounts of RAM. Exactly. 00:01:06.560 --> 00:01:10.200 That hardware barrier is the fundamental bottleneck. 00:01:10.359 --> 00:01:12.939 It's why AI has remained so incredibly centralized. 00:01:12.939 --> 00:01:15.379 You can't just run a frontier model on a standard 00:01:15.379 --> 00:01:18.700 laptop. No, you need vast, expensive data centers. 00:01:18.959 --> 00:01:22.030 Well, until now, apparently. Yeah. Google Research 00:01:22.030 --> 00:01:24.030 just introduced something called TurboQuant, 00:01:24.109 --> 00:01:27.030 and it completely changes the math on this. It's 00:01:27.030 --> 00:01:29.650 a massive breakthrough. TurboQuant is this radical 00:01:29.650 --> 00:01:32.530 new compression method. It shrinks the memory 00:01:32.530 --> 00:01:36.609 usage of a model by up to six times. Wow. And 00:01:36.609 --> 00:01:39.150 at the exact same time, it speeds up attention 00:01:39.150 --> 00:01:42.430 computation by eight times. Let's slow down and 00:01:42.430 --> 00:01:44.810 clarify that. Attention computation is essentially 00:01:44.810 --> 00:01:47.700 how the AI... figures out which words matter 00:01:47.700 --> 00:01:50.260 most. Yeah. It's the core engine of how these 00:01:50.260 --> 00:01:52.939 models understand context. So making that process 00:01:52.939 --> 00:01:55.459 eight times faster is staggering. It's like packing 00:01:55.459 --> 00:01:57.980 a giant suitcase of data into a tiny carry -on 00:01:57.980 --> 00:02:00.159 without wrinkling a single shirt. Right. And 00:02:00.159 --> 00:02:02.019 to take that carry -on analogy further, most 00:02:02.019 --> 00:02:04.140 compression methods usually destroy the quality 00:02:04.140 --> 00:02:06.180 of the model's output. It's like folding the 00:02:06.180 --> 00:02:08.439 shirt so tightly you just ruin the fabric completely. 00:02:08.960 --> 00:02:11.650 But TurboQuant doesn't do that. it maintains 00:02:11.650 --> 00:02:14.330 the output quality perfectly. How exactly are 00:02:14.330 --> 00:02:17.430 they pulling that off? Because in software, compression 00:02:17.430 --> 00:02:20.270 almost always demands a heavy trade -off. What's 00:02:20.270 --> 00:02:22.449 fascinating here is the underlying architecture 00:02:22.449 --> 00:02:25.870 shift. Instead of storing large vector data in 00:02:25.870 --> 00:02:28.810 traditional grid formats, they pivoted. They 00:02:28.810 --> 00:02:31.310 used something completely different called polar 00:02:31.310 --> 00:02:34.169 quant. Vector data being mathematical lists representing 00:02:34.169 --> 00:02:37.310 how concepts relate to each other. Spot on. So 00:02:37.310 --> 00:02:40.530 PolarQuant works by changing how that math is 00:02:40.530 --> 00:02:43.270 actually mapped. Instead of plotting data on 00:02:43.270 --> 00:02:45.669 a standard grid, which takes up massive space, 00:02:46.050 --> 00:02:48.750 it maps the data using polar coordinates. Using 00:02:48.750 --> 00:02:51.969 angles and distances instead. Exactly. It's vastly 00:02:51.969 --> 00:02:54.210 more efficient for the computer to process. And 00:02:54.210 --> 00:02:56.389 it works smoothly on existing open models. You 00:02:56.389 --> 00:02:58.789 can apply it to models like Gemma or Mistral 00:02:58.789 --> 00:03:01.050 without any retraining. That's incredible. Yeah, 00:03:01.069 --> 00:03:03.189 it compresses cache precision down to just three 00:03:03.189 --> 00:03:05.680 or four bits. Meaning it drastically... reduces 00:03:05.680 --> 00:03:08.400 the mathematical footprint of the AI's short 00:03:08.400 --> 00:03:11.259 -term memory? Yes, and it does this without making 00:03:11.259 --> 00:03:14.319 it forget how to speak intelligently. There's 00:03:14.319 --> 00:03:20.460 zero drop in output quality. Whoa! Imagine running 00:03:20.460 --> 00:03:23.520 a massive brain locally without a giant data 00:03:23.520 --> 00:03:27.479 center. Two -sec silence. Does this mean cloud 00:03:27.479 --> 00:03:30.759 computing for AI is suddenly obsolete? Not obsolete, 00:03:30.840 --> 00:03:33.740 no. But this pushes inference directly onto cheaper 00:03:33.740 --> 00:03:35.960 hardware and mobile devices. So we train models 00:03:35.960 --> 00:03:38.539 centrally, but run them locally right in our 00:03:38.539 --> 00:03:41.400 pockets. That is exactly the shift we're seeing 00:03:41.400 --> 00:03:43.460 right now. Because these models can now run on 00:03:43.460 --> 00:03:46.080 much smaller hardware, developers are shifting 00:03:46.080 --> 00:03:48.659 focus. They're aggressively building tools that 00:03:48.659 --> 00:03:51.219 let you control AI locally. The power is moving 00:03:51.219 --> 00:03:53.259 out to the edges. We saw this clearly in the 00:03:53.259 --> 00:03:55.699 recent live OpenClaw Mastery workshop. I happened 00:03:55.699 --> 00:03:58.139 there. It was an intensive live build session. 00:03:58.280 --> 00:04:00.879 It provided a complete blueprint for users to 00:04:00.879 --> 00:04:03.259 own their private AI agents. Right. It showed 00:04:03.259 --> 00:04:05.939 developers how to connect to Cloud or GPT or 00:04:05.939 --> 00:04:08.400 even how to run open models entirely locally 00:04:08.400 --> 00:04:10.439 on their own machines. What really stood out 00:04:10.439 --> 00:04:12.439 to me was the deployment strategy. They were 00:04:12.439 --> 00:04:14.360 deploying these private agents directly into 00:04:14.360 --> 00:04:17.180 apps that people already use daily. Yeah. Platforms 00:04:17.180 --> 00:04:19.759 like WhatsApp, Telegram, and Zollo. That's incredibly 00:04:19.759 --> 00:04:22.360 smart. You aren't asking a user to completely 00:04:22.360 --> 00:04:25.100 change their daily habits. You're putting the 00:04:25.100 --> 00:04:27.759 intelligence right into their existing communication 00:04:27.759 --> 00:04:30.300 flow. If we connect this to the bigger picture, 00:04:30.500 --> 00:04:33.959 friction is the enemy. The wave of new empowered 00:04:33.959 --> 00:04:36.860 tools hitting the market right now is just relentless. 00:04:37.300 --> 00:04:38.779 Here's where it gets really interesting. Let's 00:04:38.779 --> 00:04:40.220 look at a few of these tools to see how they 00:04:40.220 --> 00:04:43.040 function. First, we have Cloud Codes Auto Mode. 00:04:43.279 --> 00:04:46.000 This lets Cloud approve file rights and bash 00:04:46.000 --> 00:04:49.079 commands inside isolated environments. Bash commands 00:04:49.079 --> 00:04:51.759 being instructions typed directly into your computer's 00:04:51.759 --> 00:04:54.120 operating system. Exactly. It's a massive time 00:04:54.120 --> 00:04:56.620 saver for software engineers. It automates the 00:04:56.620 --> 00:04:58.959 tedious back -end work. Then there's Agent Place. 00:04:59.240 --> 00:05:01.699 This tool lets you build highly specialized... 00:05:01.709 --> 00:05:04.490 agents easily. You can deploy them for lead routing, 00:05:04.670 --> 00:05:07.610 document analysis, or complex scheduling. We're 00:05:07.610 --> 00:05:09.730 also seeing unique tools like LayerProof Mat. 00:05:09.970 --> 00:05:12.589 It automatically repurposes your social media 00:05:12.589 --> 00:05:15.290 posts. It generates unique platform -specific 00:05:15.290 --> 00:05:18.370 formats for LinkedIn, X, Instagram, and TikTok 00:05:18.370 --> 00:05:20.709 instantly. And for the visual builders, there's 00:05:20.709 --> 00:05:23.389 Alma. It seamlessly combines language models 00:05:23.389 --> 00:05:26.379 with 3D generation. You can use parallel agents 00:05:26.379 --> 00:05:28.860 to create interactive apps and 3D assets quickly. 00:05:29.100 --> 00:05:32.019 It's wild. Sounds magical. But I have to admit 00:05:32.019 --> 00:05:35.470 something here. Beat. I still wrestle with prompt 00:05:35.470 --> 00:05:37.990 drift myself when trying to chain these tools 00:05:37.990 --> 00:05:40.649 together. It's genuinely difficult. Prompt drift 00:05:40.649 --> 00:05:43.389 is when an agent slowly forgets its original 00:05:43.389 --> 00:05:46.910 goal over time. Yeah. The orchestration is undeniably 00:05:46.910 --> 00:05:48.850 the hardest part. You're basically managing a 00:05:48.850 --> 00:05:51.649 complex team of digital workers. So let's circle 00:05:51.649 --> 00:05:54.790 back to that OpenClaw workshop. Sure. Why is 00:05:54.790 --> 00:05:57.610 the OpenClaw community focusing so heavily on 00:05:57.610 --> 00:06:00.949 deploying to chat apps like WhatsApp? Well, because 00:06:00.949 --> 00:06:04.120 users absolutely hate downloading new apps. They 00:06:04.120 --> 00:06:06.100 want AI assistance natively where they already 00:06:06.100 --> 00:06:08.860 talk. Meeting users in their existing chats rather 00:06:08.860 --> 00:06:11.379 than forcing them into new apps. Exactly. Keep 00:06:11.379 --> 00:06:13.319 the intelligence right where they already live. 00:06:13.480 --> 00:06:15.399 Yeah. So we have these incredible local models 00:06:15.399 --> 00:06:17.199 now. We have powerful agent tools everywhere. 00:06:17.600 --> 00:06:19.879 But what happens to the humans actually trying 00:06:19.879 --> 00:06:22.339 to use them? This raises an incredibly important 00:06:22.339 --> 00:06:24.500 question. The workforce is shifting underneath 00:06:24.500 --> 00:06:27.480 us. Encropix head of economics, Peter McCrory, 00:06:27.620 --> 00:06:30.220 just released some sobering new research. The 00:06:30.220 --> 00:06:32.360 findings are actually quite counterintuitive. 00:06:32.600 --> 00:06:35.319 Everyone initially expected immediate mass layoffs 00:06:35.319 --> 00:06:37.540 across the board. Right. But there are no widespread 00:06:37.540 --> 00:06:41.040 job losses yet. Unemployment and AI exposed roles 00:06:41.040 --> 00:06:44.019 is completely stable right now. Companies aren't 00:06:44.019 --> 00:06:46.459 firing people, but the baseline expectations 00:06:46.459 --> 00:06:49.639 are changing rapidly. The real shift is that 00:06:49.639 --> 00:06:51.740 power users are pulling significantly ahead. 00:06:51.920 --> 00:06:54.899 It's a massive divergence. Early adopters are 00:06:54.899 --> 00:06:57.480 using AI for rapid iteration. They use it for 00:06:57.480 --> 00:07:00.300 instant feedback and complex problem solving. 00:07:00.519 --> 00:07:03.319 This makes them vastly faster and more capable 00:07:03.319 --> 00:07:05.699 than their peers. They're doing weeks of complex 00:07:05.699 --> 00:07:09.240 work in mere days. Anthropic CEO Dario Amadei 00:07:09.240 --> 00:07:12.000 issued a very specific warning about this trend. 00:07:12.180 --> 00:07:15.019 He believes entry level white collar roles will 00:07:15.019 --> 00:07:17.800 deeply feel this impact within five years. The 00:07:17.800 --> 00:07:20.519 productivity floor is rising. The baseline of 00:07:20.519 --> 00:07:23.139 what is expected is just much higher now. Yes. 00:07:23.379 --> 00:07:26.459 And the usage data shows another highly concerning 00:07:26.459 --> 00:07:30.269 trend. AI adoption is currently skewed much higher 00:07:30.269 --> 00:07:32.529 in wealthier regions. Which means the capability 00:07:32.529 --> 00:07:35.470 gap between different populations of users is 00:07:35.470 --> 00:07:38.250 actively growing. It's compounding daily. The 00:07:38.250 --> 00:07:40.329 people learning these tools are accelerating 00:07:40.329 --> 00:07:42.790 rapidly. Everyone else is basically standing 00:07:42.790 --> 00:07:45.889 still. Is this gap just about learning a new 00:07:45.889 --> 00:07:48.269 software interface like we did with Excel? No. 00:07:48.810 --> 00:07:51.089 It is fundamentally different. It is about learning 00:07:51.089 --> 00:07:54.149 how to successfully delegate cognition and strategy. 00:07:54.430 --> 00:07:56.889 It's less about clicking buttons and more about 00:07:56.889 --> 00:07:59.310 managing a digital brain. That's a perfect way 00:07:59.310 --> 00:08:01.230 to put it. You're stepping into a managerial 00:08:01.230 --> 00:08:04.889 role over logic itself. Sponsor. Welcome back. 00:08:04.930 --> 00:08:07.009 We were just talking about the human skills gap. 00:08:07.480 --> 00:08:09.959 But while everyday workers try to navigate this, 00:08:10.139 --> 00:08:12.519 the tech giants are making entirely different 00:08:12.519 --> 00:08:14.399 calculations. Yeah, the companies controlling 00:08:14.399 --> 00:08:16.540 the actual infrastructure are making massive, 00:08:16.779 --> 00:08:19.920 ruthless pivots. Meta is the prime example of 00:08:19.920 --> 00:08:22.600 this dramatic whiplash right now. Mark Zuckerberg 00:08:22.600 --> 00:08:25.379 is pushing incredibly aggressively for, quote, 00:08:25.519 --> 00:08:28.300 super intelligence growth. The internal math 00:08:28.300 --> 00:08:30.980 there is brutal. Look at two staggering facts 00:08:30.980 --> 00:08:34.830 from our sources today. Meta recently laid off 00:08:34.830 --> 00:08:38.750 700 employees. At the exact same time, they offered 00:08:38.750 --> 00:08:42.129 their executives massive stock packages. Packages 00:08:42.129 --> 00:08:46.149 worth up to $921 million. It's a massive realignment 00:08:46.149 --> 00:08:49.750 of capital. Smaller core teams, much bigger financial 00:08:49.750 --> 00:08:52.070 bets on infrastructure. They're also looking 00:08:52.070 --> 00:08:54.389 outward, though. They just launched Meta Small 00:08:54.389 --> 00:08:57.259 Business. It's a new initiative to help entrepreneurs 00:08:57.259 --> 00:09:00.559 build and grow using their specific AI tools. 00:09:00.840 --> 00:09:02.600 They want everyone locked into their ecosystem 00:09:02.600 --> 00:09:05.960 early. But, you know, not every major pivot works 00:09:05.960 --> 00:09:08.519 out perfectly. Look at OpenAI's recent public 00:09:08.519 --> 00:09:11.240 stumble. This was deeply fascinating. ChatGPT 00:09:11.240 --> 00:09:13.240 essentially tried to become a shopping destination. 00:09:13.950 --> 00:09:15.690 They want to be the next Amazon. You completely 00:09:15.690 --> 00:09:17.870 flopped. Users just didn't bite. They didn't 00:09:17.870 --> 00:09:20.309 want to buy things that way. Does ChatGPT's shopping 00:09:20.309 --> 00:09:23.250 failure expose a fundamental limit to what people 00:09:23.250 --> 00:09:26.330 want from LLMs? Well, it shows people trust AI 00:09:26.330 --> 00:09:29.230 to help them brainstorm and discover, but not 00:09:29.230 --> 00:09:31.350 to actually swipe their credit cards. People 00:09:31.350 --> 00:09:34.769 want an AI shopping advisor, not an automated 00:09:34.769 --> 00:09:37.669 AI checkout register. Right. Trust is highly 00:09:37.669 --> 00:09:40.970 contextual. We trust the machine to think. We 00:09:40.970 --> 00:09:42.690 just don't trust it with our wallets quite yet. 00:09:43.070 --> 00:09:45.350 But the physical domain of AI is expanding anyway. 00:09:45.629 --> 00:09:47.750 We've seen it shrink into our phones. We've seen 00:09:47.750 --> 00:09:50.789 it heavily disrupt our offices. Its final frontier 00:09:50.789 --> 00:09:53.629 is breaking out of the screen entirely. It's 00:09:53.629 --> 00:09:55.950 aggressively entering the physical world. The 00:09:55.950 --> 00:09:59.210 financial investments here are massive. QCraft 00:09:59.210 --> 00:10:02.789 just announced a staggering $100 million fundraise. 00:10:02.889 --> 00:10:06.190 They're scaling physical AI and autonomous driving. 00:10:06.450 --> 00:10:08.330 Their QPilot system is already running in over 00:10:08.330 --> 00:10:11.210 a million vehicles right now. It seamlessly covers 00:10:11.210 --> 00:10:13.529 about 30 different vehicle models. And they plan 00:10:13.529 --> 00:10:16.769 to add 50 more models by 2026. They're also moving 00:10:16.769 --> 00:10:19.730 into level four logistics. Meaning vehicles operating 00:10:19.730 --> 00:10:22.590 entirely without human input in specific areas. 00:10:22.889 --> 00:10:26.289 Exactly. They are also actively launching robotaxi 00:10:26.289 --> 00:10:29.320 pilots. So AI is driving. our cars, but it's 00:10:29.320 --> 00:10:31.379 also manifesting in our culture and our creativity. 00:10:31.679 --> 00:10:34.639 Google just released Lyria 3 Pro. They integrated 00:10:34.639 --> 00:10:37.899 it directly into Gemini, Vertex AI and Studio. 00:10:38.179 --> 00:10:40.700 What does it actually do? It generates highly 00:10:40.700 --> 00:10:43.279 structured, completely original three -minute 00:10:43.279 --> 00:10:46.659 music tracks. It gives creators immense granular 00:10:46.659 --> 00:10:50.379 control over the composition. The AI is literally 00:10:50.379 --> 00:10:53.100 writing our pop songs now. The cultural integration 00:10:53.100 --> 00:10:55.720 is happening everywhere, even in politics and 00:10:55.720 --> 00:10:58.299 education. For example, at a recent White House 00:10:58.299 --> 00:11:01.480 event, we saw this clearly. Former First Lady 00:11:01.480 --> 00:11:04.899 Melania Trump was seen walking alongside a humanoid 00:11:04.899 --> 00:11:08.259 robot. She stated her belief that children should 00:11:08.259 --> 00:11:11.159 be educated by humanoid educators. It really 00:11:11.159 --> 00:11:13.320 shows how quickly this hardware is entering mainstream 00:11:13.320 --> 00:11:15.679 cultural conversations. We're moving rapidly 00:11:15.679 --> 00:11:17.899 past just software on a screen. It's all deeply 00:11:17.899 --> 00:11:20.179 connected. What connects an autonomous robotaxi, 00:11:20.179 --> 00:11:23.879 a three -minute AI pop song, and a humanoid teacher? 00:11:24.159 --> 00:11:26.740 They all represent the outsourcing of complex, 00:11:27.019 --> 00:11:30.519 previously human -only tasks, driving. art, and 00:11:30.519 --> 00:11:33.340 care into physical or structured algorithms. 00:11:33.559 --> 00:11:35.659 We are handing over creativity, transportation, 00:11:35.960 --> 00:11:38.720 and education to autonomous algorithms. It's 00:11:38.720 --> 00:11:40.860 a profound shift in what we consider uniquely 00:11:40.860 --> 00:11:43.700 human territory. Let's take a step back and look 00:11:43.700 --> 00:11:45.860 at the big picture we've painted today. The physical 00:11:45.860 --> 00:11:48.299 constraints of hardware are dissolving. Innovations 00:11:48.299 --> 00:11:50.500 like TurboQuant are making models much smaller, 00:11:50.639 --> 00:11:53.860 faster, and cheaper to run locally. Because of 00:11:53.860 --> 00:11:56.379 this breakthrough, builders are suddenly able 00:11:56.379 --> 00:12:00.250 to create highly personalized agents. These agents 00:12:00.250 --> 00:12:02.289 are smoothly moving into the apps we already 00:12:02.289 --> 00:12:05.129 use daily. But this rapid technological acceleration 00:12:05.129 --> 00:12:08.289 is quietly dividing us. The modern workforce 00:12:08.289 --> 00:12:11.389 is splitting into power users and everyone else. 00:12:11.610 --> 00:12:13.690 All while big tech places their billion -dollar 00:12:13.690 --> 00:12:16.830 bets. They're laying off staff to fund a future 00:12:16.830 --> 00:12:19.710 of physical and super -intelligent AI. The world 00:12:19.710 --> 00:12:22.250 is being rewired from the inside out. Thank you 00:12:22.250 --> 00:12:23.750 so much for taking the time to learn with us 00:12:23.750 --> 00:12:26.299 today. Understanding these dramatic shifts is 00:12:26.299 --> 00:12:28.580 the first step to navigating them. Always keep 00:12:28.580 --> 00:12:32.279 questioning the data you're given. Beat. If memory 00:12:32.279 --> 00:12:34.259 keeps shrinking to the point where advanced models 00:12:34.259 --> 00:12:36.940 live entirely on your phone, what happens when 00:12:36.940 --> 00:12:39.200 your local AI agent eventually knows your habits, 00:12:39.379 --> 00:12:41.960 your work, and your mind better than you do? 00:12:42.360 --> 00:12:43.460 Out to your own music.