WEBVTT 00:00:00.000 --> 00:00:02.819 So I have to start today with something that 00:00:02.819 --> 00:00:05.240 honestly it made my brain hurt a little bit this 00:00:05.240 --> 00:00:08.460 morning. I was reading this story in Wired about 00:00:08.460 --> 00:00:11.019 a platform called Rent -A -Human. Which, just 00:00:11.019 --> 00:00:12.339 right out of the gate, sounds like the start 00:00:12.339 --> 00:00:14.259 of a bad sci -fi novel. It gets so much worse. 00:00:14.300 --> 00:00:15.859 It's basically a gig platform, right? Right. 00:00:15.859 --> 00:00:19.679 But the employers are AI bots. Okay. They're 00:00:19.679 --> 00:00:23.399 hiring humans for tasks, you know, solving CAPTCHAs, 00:00:23.399 --> 00:00:25.820 data verification, that sort of thing. But there's 00:00:25.820 --> 00:00:29.379 this one user, a human, who did the work, he 00:00:29.379 --> 00:00:32.659 logged his hours, and earned absolutely zero 00:00:32.659 --> 00:00:35.600 dollars. Zero. Zero. Because the system... It 00:00:35.600 --> 00:00:38.579 logged the transaction as AI paid me. Oh, wow. 00:00:38.799 --> 00:00:42.060 The bot hallucinated the payment and the platform's 00:00:42.060 --> 00:00:44.960 code just, it accepted the bot's word over the 00:00:44.960 --> 00:00:47.340 human's actual bank account. That is a perfect, 00:00:47.420 --> 00:00:50.840 if slightly terrifying, encapsulation of where 00:00:50.840 --> 00:00:53.280 we are right now. I mean, we spent a decade worrying 00:00:53.280 --> 00:00:55.520 that robots would take our jobs. Right. It turns 00:00:55.520 --> 00:00:56.899 out we should have been worried they'd just be 00:00:56.899 --> 00:01:00.100 really, really terrible payroll managers. It's 00:01:00.100 --> 00:01:02.359 the ultimate irony. But I think it sets the stage 00:01:02.359 --> 00:01:04.319 perfectly for what we need to wrestle with today. 00:01:04.810 --> 00:01:07.170 Welcome back to the Deep Dive. I'm here to help 00:01:07.170 --> 00:01:09.590 you navigate the noise. And I'm here to try and 00:01:09.590 --> 00:01:11.950 help you find the signal. Today is Monday, February 00:01:11.950 --> 00:01:15.870 16th, 2026. And if that rent -a -human story 00:01:15.870 --> 00:01:18.590 tells us anything, it's that the lines between 00:01:18.590 --> 00:01:22.790 human agency and artificial autonomy are getting... 00:01:23.180 --> 00:01:25.659 Well, messy. Messy is an understatement. We are 00:01:25.659 --> 00:01:29.439 so far past the chatbot era, you know, the look 00:01:29.439 --> 00:01:32.260 what it can say phase. We are firmly in the agent 00:01:32.260 --> 00:01:34.659 era now, the look what it can do phase. And that 00:01:34.659 --> 00:01:38.379 shift, it changes everything from like geopolitics 00:01:38.379 --> 00:01:41.750 to how you edit. a simple video. We have a stacked 00:01:41.750 --> 00:01:43.510 lineup to get through. We're going to break down 00:01:43.510 --> 00:01:45.750 Alibaba's massive new move in the model wars 00:01:45.750 --> 00:01:47.909 with Quen 3 .5, and I really want to push on 00:01:47.909 --> 00:01:49.950 whether this signals the end of American dominance 00:01:49.950 --> 00:01:52.150 in AI. It's a huge question. We're also going 00:01:52.150 --> 00:01:54.109 to look at the state of AI video, which has become, 00:01:54.290 --> 00:01:56.489 frankly, overwhelming for anyone just trying 00:01:56.489 --> 00:01:58.730 to keep up. And we have a strategy to fix that 00:01:58.730 --> 00:02:01.680 overwhelm, I promise. Good. Plus, we've got a 00:02:01.680 --> 00:02:03.760 headline segment that covers, well, everything 00:02:03.760 --> 00:02:07.359 from cheating consultants at KPMG to actual drone 00:02:07.359 --> 00:02:09.360 swarms at the Pentagon. And we're going to geek 00:02:09.360 --> 00:02:12.039 out. We are. We're going to get into a technical 00:02:12.039 --> 00:02:14.240 paper called Adapt Evolve that might just solve 00:02:14.240 --> 00:02:17.199 the energy crisis of running all these agents. 00:02:17.460 --> 00:02:19.659 That paper is fascinating. It's essentially teaching 00:02:19.659 --> 00:02:22.699 AI the art of strategic laziness. I think we 00:02:22.699 --> 00:02:24.719 can all learn a little something from that. Okay, 00:02:24.780 --> 00:02:27.020 let's unpack this. We have to start with the 00:02:27.020 --> 00:02:29.439 big release out of China. Alibaba just dropped 00:02:29.439 --> 00:02:34.409 Quen 3 .5. Now, for... The uninitiated or for 00:02:34.409 --> 00:02:37.169 people who just track, you know, open AI and 00:02:37.169 --> 00:02:40.150 anthropic. How big of a deal is this release, 00:02:40.370 --> 00:02:42.449 really? It is a very loud statement. I mean, 00:02:42.449 --> 00:02:44.669 for a long time, the narrative was that Western 00:02:44.669 --> 00:02:46.949 labs held the frontier crown. Yeah. And Chinese 00:02:46.949 --> 00:02:49.469 labs were just the fast followers. Exactly. Fast 00:02:49.469 --> 00:02:52.949 followers. Quinn 3 .5 challenges that core assumption. 00:02:53.229 --> 00:02:55.289 This isn't just another language model that can 00:02:55.289 --> 00:02:57.689 write you a poem. The branding, the architecture. 00:02:58.389 --> 00:03:02.229 It is all focused on one thing. AI agents. Okay, 00:03:02.289 --> 00:03:04.789 so we throw that word agent around a lot. I want 00:03:04.789 --> 00:03:07.129 to be precise here because I feel like it gets 00:03:07.129 --> 00:03:09.810 muddied in all the marketing. It does. If I'm 00:03:09.810 --> 00:03:12.409 a developer listening or just a user, what is 00:03:12.409 --> 00:03:14.750 the functional difference between a model like 00:03:14.750 --> 00:03:17.689 GPT -4 and an agent model like this one? That's 00:03:17.689 --> 00:03:20.669 a great question. A conversational model, like 00:03:20.669 --> 00:03:23.810 a standard chatbot, it's linear. You ask, it 00:03:23.810 --> 00:03:25.550 answers. It's just predicting the next word. 00:03:25.650 --> 00:03:28.870 Right. An agent is circular. It has a feedback 00:03:28.870 --> 00:03:31.430 loop. A feedback loop. Exactly. Think of it like 00:03:31.430 --> 00:03:33.969 the ODEA loop. In military strategy, observe, 00:03:34.169 --> 00:03:37.590 orient, decide, act. An agent can, say, write 00:03:37.590 --> 00:03:39.490 some code, try to run that code, see an error 00:03:39.490 --> 00:03:42.310 message, then read the error, rewrite the code, 00:03:42.330 --> 00:03:45.490 and try again. It has tool use baked into its 00:03:45.490 --> 00:03:48.569 core. It doesn't just talk. It navigates an environment. 00:03:48.770 --> 00:03:51.050 So it's not just generating text. It's correcting 00:03:51.050 --> 00:03:53.830 itself based on reality. Precisely. And Quen 00:03:53.830 --> 00:03:56.530 3 .5 is built to plug directly into these open 00:03:56.530 --> 00:03:58.990 source agent frameworks, specifically one called 00:03:58.990 --> 00:04:01.370 OpenClaw. OpenClaw has been all over the GitHub 00:04:01.370 --> 00:04:04.289 trends lately. It has. It's basically the interface 00:04:04.289 --> 00:04:07.009 layer that lets the model control your computer. 00:04:07.169 --> 00:04:09.569 And Alibaba released this as an open weight model 00:04:09.569 --> 00:04:14.990 with 397 billion parameters. 397 billion. That 00:04:14.990 --> 00:04:18.689 is a very specific number. And it sounds massive, 00:04:18.750 --> 00:04:21.149 but why does that number matter? Is it just bragging 00:04:21.149 --> 00:04:23.310 rights? Well, it matters because of the hardware 00:04:23.310 --> 00:04:26.990 reality. To put 397 billion parameters in perspective, 00:04:27.370 --> 00:04:29.709 you are not running this on your MacBook Pro. 00:04:29.870 --> 00:04:31.750 Right. You are not even running this on a tricked 00:04:31.750 --> 00:04:35.850 -out gaming PC. You need serious enterprise -grade 00:04:35.850 --> 00:04:39.769 GPU clusters. We're talking multiple H100s or 00:04:39.769 --> 00:04:42.230 the new Blackwell chips just to load the weights 00:04:42.230 --> 00:04:43.829 into memory. So this is an industrial -grade 00:04:43.829 --> 00:04:47.040 tool. It is chunky. But here's the kicker. It's 00:04:47.040 --> 00:04:48.899 actually smaller than their last flagship model. 00:04:49.079 --> 00:04:50.720 Wait, I thought the trend was always bigger is 00:04:50.720 --> 00:04:53.500 better. That trend is dead. We are now in the 00:04:53.500 --> 00:04:57.000 era of denser is better. Alibaba is claiming 00:04:57.000 --> 00:04:59.439 this model has stronger performance per parameter 00:04:59.439 --> 00:05:01.399 than anything else on the market. So they're 00:05:01.399 --> 00:05:03.720 optimizing for the economics of it all? Exactly. 00:05:03.779 --> 00:05:06.199 The inference economics. How much intelligence 00:05:06.199 --> 00:05:08.220 can we squeeze out of every dollar of electricity? 00:05:08.649 --> 00:05:10.470 And they're offering this in two flavors, right? 00:05:10.550 --> 00:05:12.870 Yep. The open -weight version, which you can 00:05:12.870 --> 00:05:15.069 download and fine -tune on your own servers, 00:05:15.250 --> 00:05:17.310 which is crucial for data privacy if you're a 00:05:17.310 --> 00:05:21.290 bank or a hospital. And the hosted version, Gwen 00:05:21.290 --> 00:05:25.069 3 .5 +, which just runs on Alibaba Cloud. This 00:05:25.069 --> 00:05:27.029 feels like a strategic play we've seen before. 00:05:27.170 --> 00:05:29.769 It reminds me of, I don't know, Android versus 00:05:29.769 --> 00:05:33.629 iOS, or maybe what Meta did with Llama. It is 00:05:33.629 --> 00:05:36.769 the control plus ecosystem play. Yeah. If you 00:05:36.769 --> 00:05:39.069 are a Western developer, or maybe a developer 00:05:39.069 --> 00:05:41.269 in Southeast Asia, and you want to build an autonomous 00:05:41.269 --> 00:05:44.149 coding bot, you need a model that's great at 00:05:44.149 --> 00:05:46.730 function calling. The ability to use tools. Right. 00:05:46.810 --> 00:05:49.310 And if Quen 3 .5 is the best open model for that, 00:05:49.350 --> 00:05:51.069 you're going to build on Quen. And just like 00:05:51.069 --> 00:05:53.490 that, Alibaba becomes the infrastructure layer 00:05:53.490 --> 00:05:55.769 for your whole business. Precisely. You get locked 00:05:55.769 --> 00:05:58.290 in. And the timing is no accident. This dropped 00:05:58.290 --> 00:06:01.300 right before Chinese New Year. It's a flex. It 00:06:01.300 --> 00:06:03.279 brings to mind that quote from Demis Hassabis, 00:06:03.560 --> 00:06:06.259 the head of Google DeepMind. A while back, he 00:06:06.259 --> 00:06:08.740 said Chinese labs were months behind. I remember 00:06:08.740 --> 00:06:12.899 that. Looking at QIN 3 .5 and these specs, does 00:06:12.899 --> 00:06:15.500 that still hold up? I think that gap has completely 00:06:15.500 --> 00:06:18.240 evaporated. If you look at the self -reported 00:06:18.240 --> 00:06:21.560 benchmarks on these agentic tasks, you know, 00:06:21.560 --> 00:06:24.000 things like go find this file, summarize it and 00:06:24.000 --> 00:06:26.720 email it. QIN is trading blows with the absolute 00:06:26.720 --> 00:06:31.379 best from OpenAI and Anthropic. Wow. That months 00:06:31.379 --> 00:06:34.660 behind narrative is just dangerous complacency 00:06:34.660 --> 00:06:36.959 at this point. So let me ask the probing question 00:06:36.959 --> 00:06:39.899 here. If the parameter war is over, what are 00:06:39.899 --> 00:06:42.860 we fighting now? The agent war. It's no longer 00:06:42.860 --> 00:06:44.819 about who is the smartest at answering a trivia 00:06:44.819 --> 00:06:48.220 question. It is about reliability. Can the model 00:06:48.220 --> 00:06:51.000 do a 20 -step task without hallucinating or getting 00:06:51.000 --> 00:06:53.519 stuck? That is the new battlefield. Speaking 00:06:53.519 --> 00:06:54.980 of things getting sick, let's talk about my weekend. 00:06:55.100 --> 00:06:58.120 I was trying to generate a simple video asset 00:06:58.120 --> 00:07:00.740 for a project, and I just fell down this rabbit 00:07:00.740 --> 00:07:03.319 hole. Oh, the video landscape in 2026 is an absolute 00:07:03.319 --> 00:07:05.300 jungle. It's not just a jungle. It's a chaotic 00:07:05.300 --> 00:07:07.540 mess. I was looking at the list of industry standard 00:07:07.540 --> 00:07:10.060 tools just for this month. Heian, Synthesia, 00:07:10.120 --> 00:07:12.259 Runway, PicoLab, Zuma Dream Machine, Kling AI, 00:07:12.540 --> 00:07:15.620 VO3, Sora 2. And that's just the top tier. Don't 00:07:15.620 --> 00:07:18.500 forget the niche ones. I have to be honest, and 00:07:18.500 --> 00:07:21.360 this is a bit of a vulnerable admission. But 00:07:21.360 --> 00:07:23.300 looking at that list just makes me want to quit. 00:07:23.379 --> 00:07:26.079 I still wrestle with prompt drift when I'm just 00:07:26.079 --> 00:07:29.019 trying to get a static image. I spent three hours 00:07:29.019 --> 00:07:31.319 on Sunday just trying to get a character to keep 00:07:31.319 --> 00:07:34.060 the same shirt on. Now I have to master eight 00:07:34.060 --> 00:07:36.740 different video interfaces. You are not alone 00:07:36.740 --> 00:07:38.680 in that fatigue. We hear this from creative directors 00:07:38.680 --> 00:07:41.439 all the time. I just learned Runway, but now 00:07:41.439 --> 00:07:43.699 Sora 2 is out. Do I have to start all over? So 00:07:43.699 --> 00:07:46.639 do they. No. And this is the critical insight 00:07:46.639 --> 00:07:49.399 from our research this week. If you try to become 00:07:49.399 --> 00:07:52.949 a tool expert, You will lose. Okay. The tools 00:07:52.949 --> 00:07:55.189 just change way too fast. You need to become 00:07:55.189 --> 00:07:58.649 a workflow expert. That sounds a bit like consulting 00:07:58.649 --> 00:08:00.709 jargon. What does that actually mean for someone 00:08:00.709 --> 00:08:03.170 listening? It means you stop treating these things 00:08:03.170 --> 00:08:05.209 as all -in -one solutions. You start treating 00:08:05.209 --> 00:08:07.949 them like components in an assembly line. Give 00:08:07.949 --> 00:08:10.230 me a concrete example. Walk me through a workflow. 00:08:10.490 --> 00:08:11.990 Okay. Let's say you need to make a corporate 00:08:11.990 --> 00:08:15.490 training video. High quality, low cost. Step 00:08:15.490 --> 00:08:18.910 one. You don't start in a video app. You start 00:08:18.910 --> 00:08:23.519 in a text. LLM like Claude or GPT to write the 00:08:23.519 --> 00:08:26.160 script and, crucially, the visual descriptions 00:08:26.160 --> 00:08:28.300 for each scene. Okay, so the blueprint comes 00:08:28.300 --> 00:08:31.279 first. That makes sense. Right. Step two, you 00:08:31.279 --> 00:08:34.440 use an audio synthesizer, maybe 11 labs, to generate 00:08:34.440 --> 00:08:37.460 the voiceover track. Ooh! Step three, you feed 00:08:37.460 --> 00:08:40.019 that voiceover into an avatar tool like Hagen 00:08:40.019 --> 00:08:41.980 for just the talking head parts. Okay, I'm with 00:08:41.980 --> 00:08:45.840 you. Step four, you use Luma or Runway to generate 00:08:45.840 --> 00:08:48.179 the B -roll, the background footage. based on 00:08:48.179 --> 00:08:50.139 those descriptions from step one. And then step 00:08:50.139 --> 00:08:52.840 five, you assemble it all in a traditional editor 00:08:52.840 --> 00:08:55.419 like Premiere or DaVinci. So you're not asking 00:08:55.419 --> 00:08:57.600 one AI to make me a video. You're acting like 00:08:57.600 --> 00:09:00.000 a general contractor, hiring different specialists 00:09:00.000 --> 00:09:02.679 for each part of the job. Exactly. And here's 00:09:02.679 --> 00:09:04.379 why this is so important. Let's say tomorrow 00:09:04.379 --> 00:09:07.279 Sora 3 comes out and it just blows Luma away. 00:09:07.460 --> 00:09:09.220 Which it probably will. If you have a defined 00:09:09.220 --> 00:09:11.200 workflow, you just swap out the step forward 00:09:11.200 --> 00:09:13.299 tool. The rest of your process, the script, the 00:09:13.299 --> 00:09:15.419 voice, the editing, it stays exactly the same. 00:09:15.559 --> 00:09:17.559 So the skill isn't clicking the buttons anymore. 00:09:17.820 --> 00:09:20.980 The skill is more like data pipeline management. 00:09:21.340 --> 00:09:24.559 That's a great way to put it. It shifts the value 00:09:24.559 --> 00:09:28.379 from just technical operation to creative direction 00:09:28.379 --> 00:09:30.840 and production. You become the conductor. I like 00:09:30.840 --> 00:09:33.240 that. The violin player might change, but the 00:09:33.240 --> 00:09:35.879 symphony is yours. That image really helps. It 00:09:35.879 --> 00:09:38.100 makes it feel less like I'm drowning in software 00:09:38.100 --> 00:09:41.580 and more like I'm building a system. So the tool 00:09:41.580 --> 00:09:44.639 matters less than the process. Right. Tools change 00:09:44.639 --> 00:09:47.159 monthly. The workflow is the skill that stays 00:09:47.159 --> 00:09:49.820 relevant. Okay. Speaking of people trying to 00:09:49.820 --> 00:09:52.440 game the system, or maybe direct it in the wrong 00:09:52.440 --> 00:09:55.600 way, let's hit the headlines. This Today in AI 00:09:55.600 --> 00:09:57.940 segment has a theme, and that theme seems to 00:09:57.940 --> 00:10:00.460 be human nature is the bug in the code. There's 00:10:00.460 --> 00:10:02.659 definitely a lot of gray area today. Let's start 00:10:02.659 --> 00:10:05.009 in the corporate world. There's a story out of 00:10:05.009 --> 00:10:08.529 Australia involving KPMG. A partner, a senior 00:10:08.529 --> 00:10:11.669 partner, was fined 10 ,000 Australian dollars. 00:10:11.950 --> 00:10:15.190 And the reason is just perfect. Why? Because 00:10:15.190 --> 00:10:17.970 they used AI to cheat on an internal course. 00:10:18.370 --> 00:10:20.710 A course about AI. You almost have to admire 00:10:20.710 --> 00:10:22.809 the meta -irony of it. It's incredible. And it 00:10:22.809 --> 00:10:24.970 wasn't just him. Over 20 staff members were caught 00:10:24.970 --> 00:10:27.149 doing it. This highlights a massive issue for 00:10:27.149 --> 00:10:29.700 the enterprise. These firms are charging clients 00:10:29.700 --> 00:10:33.120 millions to advise them on AI governance and 00:10:33.120 --> 00:10:35.820 AI safety. Trust us, we know how to implement 00:10:35.820 --> 00:10:38.179 this safely. Meanwhile, their own leadership 00:10:38.179 --> 00:10:41.100 is using the technology to bypass the very training 00:10:41.100 --> 00:10:43.460 meant to ensure that safety. It's something else. 00:10:43.659 --> 00:10:45.960 It's the cobbler's children have no shoes. But 00:10:45.960 --> 00:10:47.840 in this case, the cobbler is just faking his 00:10:47.840 --> 00:10:50.419 credentials. It raises a serious question about 00:10:50.419 --> 00:10:53.580 knowledge verification. If an AI can pass the 00:10:53.580 --> 00:10:56.080 test for you, does the certification even mean 00:10:56.080 --> 00:10:59.799 anything? We are entering a world where credentialism 00:10:59.799 --> 00:11:02.480 is going to collapse because remote testing is 00:11:02.480 --> 00:11:05.340 just fundamentally broken. Shifting gears from 00:11:05.340 --> 00:11:07.399 corporate cheating to something much more kinetic 00:11:07.399 --> 00:11:09.879 and honestly much more alarming. We have news 00:11:09.879 --> 00:11:13.299 about SpaceX and XAI. This is a big one. They're 00:11:13.299 --> 00:11:15.340 officially competing in a $100 million Pentagon 00:11:15.340 --> 00:11:18.480 challenge. And the goal, building voice -controlled 00:11:18.480 --> 00:11:21.259 autonomous drone swarms. Let that phrase sink 00:11:21.259 --> 00:11:25.710 in for a second. Voice -controlled swarms. And 00:11:25.710 --> 00:11:28.710 here's where I get stuck. We all know Elon Musk 00:11:28.710 --> 00:11:31.090 has a history of warning about the dangers of 00:11:31.090 --> 00:11:34.350 AI, right? He signed the pause letter. Gave speeches 00:11:34.350 --> 00:11:36.210 about autonomous weapons being an existential 00:11:36.210 --> 00:11:40.450 threat. Exactly. And now his companies are building 00:11:40.450 --> 00:11:43.769 the very thing he warned against. It is a striking 00:11:43.769 --> 00:11:46.330 contradiction. But if you look at it through 00:11:46.330 --> 00:11:49.389 a geopolitical lens, which is likely how he justifies 00:11:49.389 --> 00:11:51.809 it, the argument is always, if we don't build 00:11:51.809 --> 00:11:54.600 it... Someone else will. The classic arms race 00:11:54.600 --> 00:11:57.480 logic. It is. But this isn't just a bigger missile. 00:11:57.659 --> 00:12:00.799 Bringing XAI into the mix means these drones 00:12:00.799 --> 00:12:04.179 are agents. Right. These aren't just drones following 00:12:04.179 --> 00:12:07.220 a GPS coordinate. No. These are agents that can 00:12:07.220 --> 00:12:10.240 perceive, decide, and act. The voice -controlled 00:12:10.240 --> 00:12:12.960 part implies high -level intent. You aren't flying 00:12:12.960 --> 00:12:15.399 the drone. You're telling the swarm, secure that 00:12:15.399 --> 00:12:18.299 perimeter, or neutralize threats in Sector 4. 00:12:18.580 --> 00:12:20.700 And the AI figures out the how. The all -over 00:12:20.700 --> 00:12:23.500 figures out the how. We are moving from human 00:12:23.500 --> 00:12:25.620 in the loop to human on the loop. And eventually 00:12:25.620 --> 00:12:27.600 human out of the loop. That's the real threshold. 00:12:27.840 --> 00:12:30.460 That is the threshold. Once software decides 00:12:30.460 --> 00:12:33.059 when to pull the trigger, the speed of conflict 00:12:33.059 --> 00:12:35.539 just accelerates beyond human reaction time. 00:12:35.679 --> 00:12:39.220 That is heavy. Let's touch on one more story 00:12:39.220 --> 00:12:42.840 before we get technical. Privacy. Meta is planning 00:12:42.840 --> 00:12:45.340 something new for their Ray -Ban smart glasses. 00:12:45.720 --> 00:12:48.059 The name tag feature. Facial recognition. You 00:12:48.059 --> 00:12:50.299 look at someone. And it pulls up their info. 00:12:50.480 --> 00:12:52.639 This has been the third rail of augmented reality 00:12:52.639 --> 00:12:56.899 for a decade. Google Glass failed, partly because 00:12:56.899 --> 00:12:59.600 people were terrified of being recorded. Now 00:12:59.600 --> 00:13:02.000 Meta is reportedly planning to launch this maybe 00:13:02.000 --> 00:13:04.100 as early as this year. It connects right back 00:13:04.100 --> 00:13:05.899 to that rent -a -human story, doesn't it? The 00:13:05.899 --> 00:13:07.879 boundary between your digital data and your physical 00:13:07.879 --> 00:13:10.960 presence is just dissolving. You can't even walk 00:13:10.960 --> 00:13:13.779 down the street anonymously anymore. In 2026, 00:13:14.159 --> 00:13:16.399 anonymity is quickly becoming a luxury good. 00:13:16.539 --> 00:13:19.279 Before we move on, I have to ask. Of these stories, 00:13:19.440 --> 00:13:21.019 the cheating, the drones, the face scanning, 00:13:21.200 --> 00:13:24.200 which one feels the most dystopian to you? The 00:13:24.200 --> 00:13:27.320 drone swarms automated warfare is a genie you 00:13:27.320 --> 00:13:30.240 can't put back in the bottle. I agree. But there's 00:13:30.240 --> 00:13:32.059 a practical constraint to all those swarms, right? 00:13:32.299 --> 00:13:35.059 You can't put a supercomputer on a tiny drone. 00:13:35.159 --> 00:13:37.720 It would drain the battery in five minutes. Exactly. 00:13:37.759 --> 00:13:40.580 And that is the perfect bridge to our final segment. 00:13:40.720 --> 00:13:43.389 Right. We've painted this picture of a world 00:13:43.389 --> 00:13:47.009 with these massive agents and drone swarms, but 00:13:47.009 --> 00:13:49.629 the bottleneck to all of this is energy. It's 00:13:49.629 --> 00:13:53.330 cost. Running a massive model like QAN 3 .5 or 00:13:53.330 --> 00:13:57.169 GPT -5 for every single step of a task is incredibly 00:13:57.169 --> 00:14:00.210 expensive. We call it inference cost. So if I 00:14:00.210 --> 00:14:02.149 have an agent trying to write a software program 00:14:02.149 --> 00:14:05.970 and it takes, say, 50 steps and every step costs 00:14:05.970 --> 00:14:09.179 a dollar. That adds up fast. It adds up incredibly 00:14:09.179 --> 00:14:11.360 fast. It makes autonomous agents economically 00:14:11.360 --> 00:14:14.120 unviable for most businesses. Yeah. But there's 00:14:14.120 --> 00:14:16.620 a new paper out called Adapt Evolve that proposes 00:14:16.620 --> 00:14:18.899 this fascinating solution. And it cuts compute 00:14:18.899 --> 00:14:21.720 costs by 40%. 40%. And it's not by just making 00:14:21.720 --> 00:14:23.679 the model smaller. It's much smarter than that. 00:14:23.740 --> 00:14:26.480 It's based on a tiered system. Okay. Think of 00:14:26.480 --> 00:14:28.360 it like a law firm. You don't pay the senior 00:14:28.360 --> 00:14:30.740 partner $1 ,000 an hour to proofread a memo. 00:14:30.820 --> 00:14:32.710 You give that to the junior associate. Okay, 00:14:32.769 --> 00:14:35.590 so the junior associate AI takes the first crack 00:14:35.590 --> 00:14:38.049 at the task. Right. In this system, they start 00:14:38.049 --> 00:14:41.809 every single step with a smaller, cheaper model. 00:14:42.110 --> 00:14:45.309 A four billion parameter model. But small models 00:14:45.309 --> 00:14:47.850 are kind of dumb. They make mistakes. That seems 00:14:47.850 --> 00:14:50.809 risky for, say, a drone or a coding bot. They 00:14:50.809 --> 00:14:54.070 do. But here is the breakthrough. While that 00:14:54.070 --> 00:14:56.950 small model is generating its response, the system 00:14:56.950 --> 00:14:59.669 is measuring its confidence in real time. Okay, 00:14:59.730 --> 00:15:02.200 pause on that. How does a machine have confidence? 00:15:02.779 --> 00:15:04.840 It's not a person. It doesn't have feelings. 00:15:05.120 --> 00:15:08.080 It's just math. When an AI generates a word, 00:15:08.279 --> 00:15:10.580 it's actually calculating the probability of 00:15:10.580 --> 00:15:13.899 that word versus every other word it knows. So 00:15:13.899 --> 00:15:16.120 if it says the sky is, and the probability for 00:15:16.120 --> 00:15:19.639 the word blue is 99 .9%, that's high confidence. 00:15:19.879 --> 00:15:22.000 And if it's wavering. If the probability is flat, 00:15:22.159 --> 00:15:24.720 if it thinks blue, gray, and green are all equally 00:15:24.720 --> 00:15:27.519 likely, the system detects that wobble. It knows 00:15:27.519 --> 00:15:29.720 the model is unsure. So it's like a lie detector 00:15:29.720 --> 00:15:32.919 for the AI's own brain? Exactly. And if the system 00:15:32.919 --> 00:15:35.820 detects that low confidence, it immediately stops 00:15:35.820 --> 00:15:38.480 the junior associate and it escalates the task 00:15:38.480 --> 00:15:41.840 to the senior partner, the big, expensive 32 00:15:41.840 --> 00:15:44.879 billion parameter model. And if the junior associate 00:15:44.879 --> 00:15:46.820 is doing just fine... It keeps the cheap result 00:15:46.820 --> 00:15:50.399 and moves on. Whoa! That's surprisingly intuitive. 00:15:50.539 --> 00:15:52.299 It's like knowing when to raise your hand and 00:15:52.299 --> 00:15:54.860 ask for help. That self -awareness is really 00:15:54.860 --> 00:15:58.120 elegant. That self -awareness is the key to efficiency. 00:15:58.539 --> 00:16:01.639 And cutting costs by 40 % means that these long 00:16:01.639 --> 00:16:03.919 -running autonomous systems finally start to 00:16:03.919 --> 00:16:06.080 make financial sense. Right. If you are running 00:16:06.080 --> 00:16:08.639 those drone swarms we talked about or the digital 00:16:08.639 --> 00:16:11.360 twin Simile is building, you can't afford to 00:16:11.360 --> 00:16:14.700 be genius -level smart 100 % of the time. You 00:16:14.700 --> 00:16:16.799 only need to be smart when it's absolutely necessary. 00:16:17.259 --> 00:16:20.340 So laziness or efficiency is actually a feature, 00:16:20.500 --> 00:16:22.919 not a bug. Efficiency is the only way agents 00:16:22.919 --> 00:16:24.720 survive. We just taught them how to budget their 00:16:24.720 --> 00:16:27.200 brainpower. I love that concept, budgeting brain 00:16:27.200 --> 00:16:29.200 power. It's so evolutionary. Nature doesn't use 00:16:29.200 --> 00:16:31.820 more energy than it needs to. Why should AI? 00:16:32.159 --> 00:16:33.600 All right, we're going to take a very quick break. 00:16:33.700 --> 00:16:35.120 When we come back, we're going to tie all this 00:16:35.120 --> 00:16:38.080 together. The agents, the video workflows, and 00:16:38.080 --> 00:16:41.279 these efficient swarms. Stay with us, mid -rule 00:16:41.279 --> 00:16:48.340 placeholder. And we are back. Okay, we have covered 00:16:48.340 --> 00:16:50.820 a massive amount of ground today, from the sheer 00:16:50.820 --> 00:16:54.039 scale of Quen 3 .5, to the workflow revolution 00:16:54.039 --> 00:16:58.139 in video, to the ethical messes at KPMG, and 00:16:58.139 --> 00:17:01.159 finally this adapt -evolve efficiency. What's 00:17:01.159 --> 00:17:03.340 the through line here? If there's one theme, 00:17:03.480 --> 00:17:06.079 I think it's the shift from intelligence to agency. 00:17:06.440 --> 00:17:08.339 Break that down for me. For the last few years, 00:17:08.380 --> 00:17:09.940 we've been obsessed with the question, is the 00:17:09.940 --> 00:17:12.220 model smart? Now we're asking a different question, 00:17:12.339 --> 00:17:15.609 can the model act? Quen gives it the tools. Adapt 00:17:15.609 --> 00:17:17.890 Evolve gives it the budget and the video workflows, 00:17:18.049 --> 00:17:19.769 they give it the output. We are building the 00:17:19.769 --> 00:17:21.670 infrastructure for action. It's not about what 00:17:21.670 --> 00:17:23.589 the AI knows anymore. It's about what it can 00:17:23.589 --> 00:17:26.589 do and how cheaply it can do it. Precisely. But 00:17:26.589 --> 00:17:28.109 then you have the human element just crashing 00:17:28.109 --> 00:17:29.849 right into it. That's the KPMG story. That's 00:17:29.849 --> 00:17:32.089 the rent -a -human story. Right. We're building 00:17:32.089 --> 00:17:35.210 these incredibly efficient autonomous systems, 00:17:35.309 --> 00:17:39.690 but the humans involved are still, well, they're 00:17:39.690 --> 00:17:43.529 still human. We cheat. We pretend. We find loopholes. 00:17:43.710 --> 00:17:46.450 That remains the most unpredictable variable 00:17:46.450 --> 00:17:48.950 in the entire equation. You can optimize the 00:17:48.950 --> 00:17:51.869 compute cost of a drone swarm by 40%, but you 00:17:51.869 --> 00:17:53.809 can't optimize the ethics of the person who's 00:17:53.809 --> 00:17:55.970 commanding it. That is the truth. So as we move 00:17:55.970 --> 00:17:59.269 forward into 2026, the technology is stabilizing. 00:17:59.369 --> 00:18:01.430 The workflows are being defined. The costs are 00:18:01.430 --> 00:18:04.349 coming down. The question is no longer, can we 00:18:04.349 --> 00:18:06.670 build it? It is, how do we live with it? Exactly. 00:18:06.690 --> 00:18:08.589 And that leads me to my final thought for you, 00:18:08.650 --> 00:18:10.980 the listener, today. I want you to think about 00:18:10.980 --> 00:18:13.759 that adapt -evolve concept again. The budgeting 00:18:13.759 --> 00:18:17.019 of brain power. Using high intelligence only 00:18:17.019 --> 00:18:19.680 when it's really necessary. Exactly. We live 00:18:19.680 --> 00:18:22.319 in this world that demands we are on 100 % of 00:18:22.319 --> 00:18:24.720 the time. High alert, high productivity, maximum 00:18:24.720 --> 00:18:27.460 processing power. But maybe, just like these 00:18:27.460 --> 00:18:29.420 agents, we're burning way too much inference 00:18:29.420 --> 00:18:32.079 cost. A permission structure for human inefficiency. 00:18:32.299 --> 00:18:35.190 I like that. Maybe. Perhaps the smartest thing 00:18:35.190 --> 00:18:37.690 we can do is identify which parts of our day 00:18:37.690 --> 00:18:40.910 actually need the senior partner and which parts 00:18:40.910 --> 00:18:43.269 can be handled by the junior associate. Yeah. 00:18:43.349 --> 00:18:45.549 Save your high -level compute for the problems 00:18:45.549 --> 00:18:48.190 that actually need it. I like that. Efficiency 00:18:48.190 --> 00:18:51.170 isn't about doing more. It's about thinking less 00:18:51.170 --> 00:18:54.329 but thinking better. Something to mull over as 00:18:54.329 --> 00:18:56.549 you start your week. If you enjoyed this deep 00:18:56.549 --> 00:18:58.809 dive, hit that subscribe button. We've got more 00:18:58.809 --> 00:19:01.269 coming your way. Always more to learn. Stay curious.