WEBVTT 00:00:00.000 --> 00:00:01.780 What if we've been looking at the search for 00:00:01.780 --> 00:00:06.860 AGI all wrong? How do you mean? Well, we always 00:00:06.860 --> 00:00:09.679 picture this massive fundamental leap forward, 00:00:09.839 --> 00:00:12.439 you know? Right. A whole new architecture, some 00:00:12.439 --> 00:00:15.160 huge breakthrough in processing power. Exactly. 00:00:15.519 --> 00:00:20.739 But what if AGI isn't some expensive grand overhaul? 00:00:20.859 --> 00:00:23.339 What if it's just... A better wrapper. A wrapper. 00:00:23.600 --> 00:00:26.039 A coordination layer built right around the AI 00:00:26.039 --> 00:00:28.739 skills we already have. Ah, I see where you're 00:00:28.739 --> 00:00:31.239 going. That tension, you know, between the incredible 00:00:31.239 --> 00:00:34.579 skills of our current models and them just not 00:00:34.579 --> 00:00:37.039 having a proper manager. That's what we're diving 00:00:37.039 --> 00:00:39.119 into today. It's the perfect way to frame it. 00:00:39.320 --> 00:00:42.259 Welcome to the deep dive. You sent over a fascinating 00:00:42.259 --> 00:00:44.280 stack of sources this week, and they all seem 00:00:44.280 --> 00:00:46.280 to be pointing in that same direction. It really 00:00:46.280 --> 00:00:48.060 feels like the tech is finally catching up with 00:00:48.060 --> 00:00:50.340 the theory. It really does. So our mission today 00:00:50.340 --> 00:00:52.619 is to unpack this shift. We're going to start 00:00:52.619 --> 00:00:55.020 with the rise of what's being called the agentic 00:00:55.020 --> 00:00:58.079 desktop. And a new system from Anthropic, Cloud 00:00:58.079 --> 00:01:00.740 Cowork. Exactly. And what that means for your 00:01:00.740 --> 00:01:03.780 local files. Then we'll hit some of the industry 00:01:03.780 --> 00:01:06.599 dynamics, career moves, acquisitions, some pretty 00:01:06.599 --> 00:01:09.140 significant regulatory blocks. All things that 00:01:09.140 --> 00:01:11.599 reinforce this need for. Well, for oversight. 00:01:11.939 --> 00:01:14.379 For that wrapper. And finally, we'll get to the 00:01:14.379 --> 00:01:17.000 big one. The theoretical breakthrough from Stanford 00:01:17.000 --> 00:01:20.340 that suggests our current LLMs are already pattern 00:01:20.340 --> 00:01:24.019 engines with the raw skills for AGI. Okay, let's 00:01:24.019 --> 00:01:27.180 start there with the agentic desktop. This feels 00:01:27.180 --> 00:01:29.959 like a huge step. It is. For a while, we've had 00:01:29.959 --> 00:01:32.379 powerful tools. I mean, think of Anthropic's 00:01:32.379 --> 00:01:35.980 last agent, Claude Code. It was amazing. but 00:01:35.980 --> 00:01:38.359 only if you were a developer. Right. It was transformative 00:01:38.359 --> 00:01:41.719 for coders, but for most people it was inaccessible. 00:01:42.000 --> 00:01:43.939 Powerful, but frustrating. Kind of stuck in its 00:01:43.939 --> 00:01:46.959 own world. Yeah. But now they've repackaged that 00:01:46.959 --> 00:01:49.239 power into something called Claude Cowork. And 00:01:49.239 --> 00:01:51.540 this is the significant part. We are officially, 00:01:51.680 --> 00:01:54.459 I think, entering the agentic desktop era. Meaning 00:01:54.459 --> 00:01:57.980 the AI is leaving the browser sandbox. it's leaving 00:01:57.980 --> 00:02:01.659 that secure controlled environment co -work wraps 00:02:01.659 --> 00:02:04.560 all that existing power in a friendly approachable 00:02:04.560 --> 00:02:08.139 ui so now non -coders get access to automation 00:02:08.139 --> 00:02:10.419 that used to require you know complex scripting 00:02:10.419 --> 00:02:12.620 and what really sets us apart from a standard 00:02:12.620 --> 00:02:16.080 gpt agent is how it interacts with your local 00:02:16.080 --> 00:02:19.099 data your files deeply and directly that's the 00:02:19.099 --> 00:02:22.590 essential technical jump here it's using internal 00:02:22.590 --> 00:02:25.210 APIs to talk to your operating system. So it's 00:02:25.210 --> 00:02:27.530 not just generating text in a little box anymore. 00:02:27.750 --> 00:02:30.250 No, it's manipulating real -world digital stuff. 00:02:30.629 --> 00:02:32.870 We're talking about an agent that can organize 00:02:32.870 --> 00:02:36.830 your files. As in, rename, sort, delete. And 00:02:36.830 --> 00:02:39.889 move documents across folders, all based on a 00:02:39.889 --> 00:02:42.400 single high -level command you give it. So I 00:02:42.400 --> 00:02:44.560 could just say something like, draft the Q3 report, 00:02:44.860 --> 00:02:47.060 pull the sales figures from those receipts I 00:02:47.060 --> 00:02:49.439 scanned, and cross -reference them with the notes 00:02:49.439 --> 00:02:51.620 in my Google Drive. And it would do it. It can 00:02:51.620 --> 00:02:53.939 draft that report from scattered notes. It can 00:02:53.939 --> 00:02:56.699 pull structured data out of a messy PDF. And 00:02:56.699 --> 00:02:58.780 it can use connectors to get into your Google 00:02:58.780 --> 00:03:01.340 Drive or Slack. It's like it's stacking these 00:03:01.340 --> 00:03:03.439 little Lego blocks of data from all over the 00:03:03.439 --> 00:03:07.319 place into a finished product. And crucially... 00:03:07.800 --> 00:03:10.379 It keeps you in the loop. It gives you progress 00:03:10.379 --> 00:03:13.060 updates step by step. Like a real teammate would. 00:03:13.199 --> 00:03:16.460 Exactly. It's the first real do -anything assistant 00:03:16.460 --> 00:03:18.659 that actually lives on your system. It's the 00:03:18.659 --> 00:03:21.180 difference between having a single wrench, a 00:03:21.180 --> 00:03:24.419 very specific AI tool, and having a whole workshop 00:03:24.419 --> 00:03:27.180 installed right on your computer. The utility 00:03:27.180 --> 00:03:30.060 just jumps exponentially. That's a perfect analogy. 00:03:30.479 --> 00:03:32.580 And the sources mention that some of the alpha 00:03:32.580 --> 00:03:34.599 versions of these agents can run projects on 00:03:34.599 --> 00:03:37.900 their own for... Months. Months. Months. And 00:03:37.900 --> 00:03:39.639 even build their own kind of functional identity 00:03:39.639 --> 00:03:43.039 over time. Wow. Okay. So if these agents are 00:03:43.039 --> 00:03:46.520 running that long on their own, touching our 00:03:46.520 --> 00:03:49.580 personal files, what's the biggest risk here? 00:03:49.740 --> 00:03:52.259 When an agent is touching a user's local files 00:03:52.259 --> 00:03:55.379 like that, the primary danger is unintended data 00:03:55.379 --> 00:03:57.819 manipulation. That's why it demands really strict 00:03:57.819 --> 00:03:59.780 testing and very clear boundaries. Right. And 00:03:59.780 --> 00:04:02.300 understanding those risks forces us to look at 00:04:02.300 --> 00:04:04.819 the bigger picture. Regulation, career growth, 00:04:05.000 --> 00:04:08.419 all of it. Precisely. Let's pivot there. Because 00:04:08.419 --> 00:04:11.500 a tool like Cowork changes the required skills 00:04:11.500 --> 00:04:14.300 for a job overnight. That's a great point. Yeah, 00:04:14.360 --> 00:04:16.720 there's this great piece of career advice from 00:04:16.720 --> 00:04:20.459 a Google AI product manager. They said, be like 00:04:20.459 --> 00:04:22.730 a crab. Like a crab? I know it sounds a little 00:04:22.730 --> 00:04:25.149 silly, but the idea is that lateral moves are 00:04:25.149 --> 00:04:27.230 often the fastest way to grow in this field. 00:04:27.509 --> 00:04:29.509 Ah, so instead of trying to climb straight up 00:04:29.509 --> 00:04:32.610 the pure AI research ladder. Exactly. You find 00:04:32.610 --> 00:04:35.329 a role that bridges what you already know, healthcare, 00:04:35.610 --> 00:04:38.970 finance, whatever, with these new AI tools. You 00:04:38.970 --> 00:04:41.629 become the human part of that coordination layer. 00:04:41.850 --> 00:04:44.050 That's really smart. You're leveraging what you 00:04:44.050 --> 00:04:46.750 already know to manage the new patterns the AI 00:04:46.750 --> 00:04:48.750 is seeing. And even the older tools are still 00:04:48.750 --> 00:04:51.519 incredibly useful. Claude Code is great for data 00:04:51.519 --> 00:04:54.139 visualization. You can build these powerful dashboards 00:04:54.139 --> 00:04:57.339 from, say, Google Analytics data almost instantly. 00:04:57.540 --> 00:04:59.509 So there's an efficiency gain there. For sure. 00:04:59.629 --> 00:05:01.629 But there's also the opposite, right? The research 00:05:01.629 --> 00:05:04.149 pointed out that AI agents are often overused. 00:05:04.370 --> 00:05:07.569 People are spending a fortune on complex AI when 00:05:07.569 --> 00:05:09.970 a simpler workflow would have been fine. Which 00:05:09.970 --> 00:05:12.310 brings us right back to the quality control problem, 00:05:12.470 --> 00:05:16.329 the AI slop. The slop, yeah. That robotic generic 00:05:16.329 --> 00:05:18.990 fluff that just fills up everything now. I'll 00:05:18.990 --> 00:05:21.009 admit, I still wrestle with prompt drift myself, 00:05:21.410 --> 00:05:23.529 you know, especially when I'm trying to get a 00:05:23.529 --> 00:05:25.990 unique voice out of it across a long output. 00:05:26.189 --> 00:05:28.519 You're not alone in that. That drift is a real 00:05:28.519 --> 00:05:31.220 -world sign of low reliability, but tools are 00:05:31.220 --> 00:05:33.839 starting to pop up to fix it. The notes mentioned 00:05:33.839 --> 00:05:35.920 an extension designed to help you write like 00:05:35.920 --> 00:05:38.480 a human again by stripping out that robotic language. 00:05:38.779 --> 00:05:41.100 Okay, so that helps with the content side. But 00:05:41.100 --> 00:05:43.560 this coordination layer idea isn't just inside 00:05:43.560 --> 00:05:46.439 the machine, it's external too. We saw some big 00:05:46.439 --> 00:05:48.699 regulatory news that supports that. Absolutely. 00:05:48.759 --> 00:05:51.459 The decision by Malaysia and Indonesia to block 00:05:51.459 --> 00:05:54.689 Grok. That's a critical precedent. And the concern 00:05:54.689 --> 00:05:56.829 wasn't just hypothetical. Right. Not at all. 00:05:56.870 --> 00:06:00.009 It was explicitly about AI -powered deepfake 00:06:00.009 --> 00:06:02.970 porn. The governments basically said Musk's team 00:06:02.970 --> 00:06:04.990 wasn't doing enough to stop the tool from being 00:06:04.990 --> 00:06:08.290 used to, quote, unmask people. So the regulatory 00:06:08.290 --> 00:06:11.550 block is a form of external coordination. forcing 00:06:11.550 --> 00:06:14.470 better ethical boundaries. It is. And then linking 00:06:14.470 --> 00:06:17.250 back to this need for rock solid reliability, 00:06:17.769 --> 00:06:20.269 we saw that big acquisition in health care. Right. 00:06:20.310 --> 00:06:23.209 Open AI buying Torch. For about $60 million. 00:06:23.410 --> 00:06:26.350 And what Torch does is pull all this incredibly 00:06:26.350 --> 00:06:29.490 complex medical data patient histories, test 00:06:29.490 --> 00:06:34.670 results into one secure central place. And why 00:06:34.670 --> 00:06:37.750 health care? Because if any field needs what 00:06:37.750 --> 00:06:39.949 we'll later call zone three reliability with 00:06:39.949 --> 00:06:42.990 no fluff, no hallucinations, it's medicine. This 00:06:42.990 --> 00:06:44.870 shows they're getting serious about high stakes 00:06:44.870 --> 00:06:46.810 applications. Yeah, it all just keeps coming 00:06:46.810 --> 00:06:48.610 back to that idea of autonomy. It's hard to shake 00:06:48.610 --> 00:06:50.290 what you said about those alpha agents. I know. 00:06:50.370 --> 00:06:53.149 The sources mentioned do anything's alpha agents, 00:06:53.370 --> 00:06:56.110 systems that run autonomously on projects for 00:06:56.110 --> 00:06:58.290 months, and they build and maintain their own 00:06:58.290 --> 00:07:00.970 operational identities over that time. Imagine 00:07:00.970 --> 00:07:03.769 scaling a project to run on its own for months 00:07:03.769 --> 00:07:06.420 where the AI manages. its own identity and trajectory 00:07:06.420 --> 00:07:08.899 without a human stepping in. That's a little 00:07:08.899 --> 00:07:11.120 chilling. A digital identity just running on 00:07:11.120 --> 00:07:13.540 your desktop. Did the sources touch on the legal 00:07:13.540 --> 00:07:16.600 side of that? Not explicitly, no. But the question 00:07:16.600 --> 00:07:19.180 is hanging in the air. If an autonomous agent 00:07:19.180 --> 00:07:22.300 causes harm, who's responsible? That's the whole 00:07:22.300 --> 00:07:24.480 ballgame. But OK, back to the slap problem for 00:07:24.480 --> 00:07:27.360 a second. We need real ways to fight it beyond 00:07:27.360 --> 00:07:30.699 a simple extension. How can users actively fight 00:07:30.699 --> 00:07:33.199 against that robotic content right now? I think 00:07:33.199 --> 00:07:35.319 the best approach is to stop giving it generic 00:07:35.319 --> 00:07:39.180 prompts. Use very specific tools and, more importantly, 00:07:39.379 --> 00:07:42.500 constraints that force the model out of its easy, 00:07:42.639 --> 00:07:46.000 default pattern matching mode. Constraints. Which 00:07:46.000 --> 00:07:47.920 is the perfect segue to the theory behind all 00:07:47.920 --> 00:07:50.620 of this. Stanford's research on the AGI pattern 00:07:50.620 --> 00:07:52.670 engine. This is where all clicks into place. 00:07:52.850 --> 00:07:56.050 The Stanford paper argues that our existing LLMs 00:07:56.050 --> 00:07:59.649 GPT -4 clawed are already incredibly powerful 00:07:59.649 --> 00:08:01.949 pattern engines. They basically digested all 00:08:01.949 --> 00:08:04.009 of human knowledge. Think of it like a massive 00:08:04.009 --> 00:08:06.550 digital encyclopedia. It can see billions of 00:08:06.550 --> 00:08:08.810 connections between concepts. It knows what to 00:08:08.810 --> 00:08:11.569 do. But, and this is the key part, it doesn't 00:08:11.569 --> 00:08:14.310 reliably know when to do it. The skills are there, 00:08:14.389 --> 00:08:17.279 but the manager is missing. Correct. So the missing 00:08:17.279 --> 00:08:20.360 piece isn't more data or a faster processor necessarily. 00:08:20.939 --> 00:08:23.819 It's what they call a coordination layer. The 00:08:23.819 --> 00:08:26.420 expert librarian for the encyclopedia. That's 00:08:26.420 --> 00:08:28.420 a great way to put it. It's a slower, smarter 00:08:28.420 --> 00:08:31.699 system sitting on top, picking the right patterns, 00:08:31.800 --> 00:08:34.259 enforcing the goals you set, and keeping track 00:08:34.259 --> 00:08:36.740 of a task over time. But wait, are you saying 00:08:36.740 --> 00:08:39.559 AGI is basically just a glorified operating system? 00:08:39.700 --> 00:08:42.460 That feels, I don't know, a bit simplistic. It's 00:08:42.460 --> 00:08:45.200 more subtle. The idea is that the potential for 00:08:45.200 --> 00:08:47.820 goal -directed reasoning is already baked into 00:08:47.820 --> 00:08:51.120 the LLM's structure. The coordination layer isn't 00:08:51.120 --> 00:08:53.799 just an OS. It's more like the executive brain. 00:08:53.980 --> 00:08:55.960 The part that provides grounded decision -making. 00:08:56.159 --> 00:08:58.419 And makes sure the outputs stay consistent and 00:08:58.419 --> 00:09:00.720 relevant over a long period, which is where current 00:09:00.720 --> 00:09:03.200 models really struggle. Okay. And the Stanford 00:09:03.200 --> 00:09:06.299 team found a way to measure this. The anchoring 00:09:06.299 --> 00:09:09.440 strength score. Yes. Exactly. It's a score that 00:09:09.440 --> 00:09:12.580 measures how locked in the model is to a reliable 00:09:12.580 --> 00:09:14.899 answer. It's like its internal confidence meter. 00:09:15.220 --> 00:09:17.460 We need to know when we can actually trust it. 00:09:17.559 --> 00:09:19.600 So how do you increase that score? How do you 00:09:19.600 --> 00:09:22.450 get a stronger anchor? Three things. First, you 00:09:22.450 --> 00:09:24.629 give it crystal clear goals and constraints. 00:09:24.889 --> 00:09:27.309 The clearer the instructions, the higher the 00:09:27.309 --> 00:09:29.889 anchor. Okay. Second, the evidence available 00:09:29.889 --> 00:09:32.470 has to clearly point to one path over the others. 00:09:32.610 --> 00:09:34.950 And third, this is the crucial one, the anchor 00:09:34.950 --> 00:09:37.429 gets stronger when the answers stay stable, even 00:09:37.429 --> 00:09:39.090 if you tweak the prompt a little bit. Because 00:09:39.090 --> 00:09:41.649 that shows real reasoning. Not just mimicry. 00:09:41.929 --> 00:09:44.470 Precisely. So low anchoring strength is where 00:09:44.470 --> 00:09:46.509 we get the frustration, the prompt drift, the 00:09:46.509 --> 00:09:49.409 fluff, the hallucinations. The Stanford team 00:09:49.409 --> 00:09:52.090 laid out three zones based on this. They did. 00:09:52.330 --> 00:09:54.570 Zone one is weak anchoring. That's just useless 00:09:54.570 --> 00:09:57.769 noise. Pure slop. Zone two is the unstable middle 00:09:57.769 --> 00:10:01.210 ground. Small prompt changes lead to big, unpredictable 00:10:01.210 --> 00:10:04.169 changes in behavior. But zone three is the goal. 00:10:04.549 --> 00:10:06.850 That's the sweet spot. Strong anchoring. That's 00:10:06.850 --> 00:10:08.570 where you get reliable, goal -directed reasoning. 00:10:08.690 --> 00:10:10.809 The kind of output you'd need for medicine or 00:10:10.809 --> 00:10:14.149 finance. The implication here is just, it's massive. 00:10:14.389 --> 00:10:17.429 Yeah. It suggests AGI isn't this long, slow climb 00:10:17.429 --> 00:10:19.250 up a mountain. It could be more like a switch. 00:10:19.509 --> 00:10:22.149 You either cross that reliability threshold or 00:10:22.149 --> 00:10:25.789 you don't. The raw skills, that giant encyclopedia, 00:10:26.049 --> 00:10:29.190 it's already here. It just needs that smarter 00:10:29.190 --> 00:10:32.580 wrapper to get to zone three consistently. And 00:10:32.580 --> 00:10:34.620 what's so cool is this connects directly to your 00:10:34.620 --> 00:10:37.360 own experience. It explains why some AI outputs 00:10:37.360 --> 00:10:40.139 you get are brilliant. You accidentally provided 00:10:40.139 --> 00:10:42.659 constraints that pushed it to zone three and 00:10:42.659 --> 00:10:44.700 why others are useless. They're stuck down in 00:10:44.700 --> 00:10:46.919 zone one. Because the quality of the output is 00:10:46.919 --> 00:10:48.799 directly tied to the quality of the management. 00:10:48.960 --> 00:10:52.700 And right now that manager is you. It's the prompts 00:10:52.700 --> 00:10:55.559 you provide. So if the skills are already here. 00:10:56.000 --> 00:10:58.100 What's the one practical behavior that will tell 00:10:58.100 --> 00:11:00.899 us we finally crossed the AGI switch? I think 00:11:00.899 --> 00:11:02.860 it's one that goal -directed reasoning becomes 00:11:02.860 --> 00:11:05.600 reliable across many, many different and diverse 00:11:05.600 --> 00:11:07.899 constraints. Okay, let's wrap this up. What's 00:11:07.899 --> 00:11:10.139 the big idea to take away? The big idea is that 00:11:10.139 --> 00:11:11.720 we're seeing these two threads come together. 00:11:11.860 --> 00:11:14.139 We have the rise of practical desktop agents 00:11:14.139 --> 00:11:16.480 like CoWork that are finally starting to use 00:11:16.480 --> 00:11:18.980 the raw pattern skills that Stanford's research 00:11:18.980 --> 00:11:21.379 has identified. So the theory and the tools are 00:11:21.379 --> 00:11:24.590 meeting. They're meeting. And the path to AGI 00:11:24.590 --> 00:11:28.210 might not be some totally new architecture. It 00:11:28.210 --> 00:11:30.009 might just be adding that crucial coordination 00:11:30.009 --> 00:11:32.850 layer, that executive brain, to the amazing pattern 00:11:32.850 --> 00:11:35.629 engines we already have. It's about management, 00:11:35.730 --> 00:11:39.110 not just memory. And that idea changes everything 00:11:39.110 --> 00:11:42.909 about how we use these tools. It means we are 00:11:42.909 --> 00:11:45.210 a part of that first coordination layer when 00:11:45.210 --> 00:11:47.580 we write our prompts. We are. If you want to 00:11:47.580 --> 00:11:49.419 put this into practice and start building those 00:11:49.419 --> 00:11:52.240 strong goal -directed constraints yourself, you 00:11:52.240 --> 00:11:54.000 should check out the Higgs Field Cinema Challenge. 00:11:54.379 --> 00:11:58.259 The deadline is January 24th, 2026. That's a 00:11:58.259 --> 00:12:00.379 great way to get hands -on experience. It's a 00:12:00.379 --> 00:12:02.580 fantastic way to practice prompt design with 00:12:02.580 --> 00:12:05.779 real deadlines and clear goals. By doing that... 00:12:05.950 --> 00:12:08.210 you are actively learning how to build Zone 3 00:12:08.210 --> 00:12:10.090 anchors. That's a really good point. So here's 00:12:10.090 --> 00:12:11.450 the final thought we want to leave you with. 00:12:11.549 --> 00:12:14.230 If HEI is really just a smarter wrapper around 00:12:14.230 --> 00:12:17.090 the pattern engines we already have, what responsibilities 00:12:17.090 --> 00:12:20.210 do we, the users, have to provide the clear, 00:12:20.230 --> 00:12:22.230 goal -directed constraints that system needs 00:12:22.230 --> 00:12:23.889 to reach anchoring strength Zone 3?