WEBVTT 00:00:00.000 --> 00:00:02.839 You know, we often hear about the coming AI job 00:00:02.839 --> 00:00:05.980 shock. But if you look at the real data, there's 00:00:05.980 --> 00:00:09.820 a truly fascinating tension there. MIT recently 00:00:09.820 --> 00:00:13.160 mapped the entire U .S. workforce with this massive 00:00:13.160 --> 00:00:16.420 digital twin they call Project Iceberg. And what 00:00:16.420 --> 00:00:18.620 they found is, well, the potential is just huge. 00:00:18.780 --> 00:00:22.339 AI already threatens about $1 .2 trillion in 00:00:22.339 --> 00:00:25.420 wages. That impacts 21 million American jobs 00:00:25.420 --> 00:00:28.000 that could be automated right now, which sounds... 00:00:28.320 --> 00:00:30.640 Terrifying. It does. But here's the twist, and 00:00:30.640 --> 00:00:33.219 this is what we need to unpack. Only 2 .2 % of 00:00:33.219 --> 00:00:35.060 those jobs have actually been disrupted so far. 00:00:35.359 --> 00:00:37.020 So we're looking at what they call the surface 00:00:37.020 --> 00:00:39.939 index of this massive buried iceberg of automation. 00:00:40.439 --> 00:00:42.979 Welcome to the Deep Dive. You shared an incredible 00:00:42.979 --> 00:00:45.380 stack of sources with us, a really deep look 00:00:45.380 --> 00:00:47.859 into global robotics, the new rules for talking 00:00:47.859 --> 00:00:50.200 to AI, and what's really driving this job market 00:00:50.200 --> 00:00:52.159 shift. Yeah, and our mission is simple. We want 00:00:52.159 --> 00:00:53.880 to pull out the biggest nuggets of insight and 00:00:53.880 --> 00:00:55.479 connect these dots for you. We're going to show 00:00:55.479 --> 00:00:57.380 you exactly where the massive investment hype 00:00:57.380 --> 00:01:00.119 ends and where the immediate practical reality 00:01:00.119 --> 00:01:03.039 begins. We've got three main stops on this dive. 00:01:03.159 --> 00:01:05.079 First, we'll look at the financial risks, specifically 00:01:05.079 --> 00:01:08.159 the humanoid robotics bubble that's building 00:01:08.159 --> 00:01:10.359 up in China. Then we'll get practical. We're 00:01:10.359 --> 00:01:11.840 going to talk about how you should be prompting 00:01:11.840 --> 00:01:14.959 these new, smarter AI models like Claude Opus 00:01:14.959 --> 00:01:17.519 4 .5 because the rules have completely changed. 00:01:17.980 --> 00:01:20.379 And finally, we'll go deep on Project Iceberg 00:01:20.379 --> 00:01:22.379 itself. We'll define this thing called the Model 00:01:22.379 --> 00:01:25.560 Context Protocol, or MCP. This is basically the 00:01:25.560 --> 00:01:27.379 digital key that's turning the whole automation 00:01:27.379 --> 00:01:29.980 engine on, and it's closing that gap between 00:01:29.980 --> 00:01:34.439 2 .2 % and 11 .7 % way faster than anyone thinks. 00:01:35.159 --> 00:01:36.859 OK, let's start with the hardware side of things, 00:01:36.939 --> 00:01:39.560 with the humanoid robots. Your sources really 00:01:39.560 --> 00:01:42.260 zeroed in on China for this. And it's not just 00:01:42.260 --> 00:01:45.019 private money, right? This is one of six industries 00:01:45.019 --> 00:01:47.299 that the Communist Party has named a national 00:01:47.299 --> 00:01:50.379 priority through 2030. Yeah, that's serious state 00:01:50.379 --> 00:01:53.000 backing. And that national priority is fueling 00:01:53.000 --> 00:01:55.620 a boom that honestly already smells a lot like 00:01:55.620 --> 00:01:57.859 a bubble. What's really fascinating here is the 00:01:57.859 --> 00:01:59.879 contradiction you see coming straight from the 00:01:59.879 --> 00:02:02.519 top. China's main economic planning agency, the 00:02:02.519 --> 00:02:05.680 NDRC, just put out a very public warning. They 00:02:05.680 --> 00:02:08.860 said the sector is overheating. That's a pretty 00:02:08.860 --> 00:02:10.979 rare thing for them to admit about a favored 00:02:10.979 --> 00:02:13.340 industry, isn't it? It's extremely rare. And 00:02:13.340 --> 00:02:16.300 it suggests the investment is growing way, way 00:02:16.300 --> 00:02:19.500 faster than the actual utility. I mean, get this. 00:02:19.599 --> 00:02:23.180 We are talking about more than 150 robotics companies 00:02:23.180 --> 00:02:26.759 building humanoid bots in China right now. 150. 00:02:26.879 --> 00:02:28.719 And I'm guessing they all kind of look the same. 00:02:29.050 --> 00:02:30.810 They mostly look and perform the same. Yeah, 00:02:30.849 --> 00:02:33.069 the differentiation is just not there. And yet 00:02:33.069 --> 00:02:34.930 the financial numbers are just going vertical. 00:02:35.030 --> 00:02:38.509 You see UB Tech claiming billions in pre -orders. 00:02:38.550 --> 00:02:41.430 The China Humanoid Robotics Index is up almost 00:02:41.430 --> 00:02:44.650 30 % this year. And Citigroup is projecting the 00:02:44.650 --> 00:02:47.530 market could hit $7 trillion by 2050. It's an 00:02:47.530 --> 00:02:50.680 astronomical number. But the reality gap is huge. 00:02:50.860 --> 00:02:52.659 Right now, if you actually look at what they're 00:02:52.659 --> 00:02:54.639 doing, most of these bots are just doing these 00:02:54.639 --> 00:02:57.259 little demonstration dances at expos. They're 00:02:57.259 --> 00:02:59.740 not packing boxes in a real factory or making 00:02:59.740 --> 00:03:02.000 coffee in someone's kitchen. OK, so that just 00:03:02.000 --> 00:03:04.800 screams red flag. It sounds incredibly familiar, 00:03:04.900 --> 00:03:07.240 almost cyclical. It does. It feels a lot like 00:03:07.240 --> 00:03:09.960 the bike sharing boom back in 2017 in China. 00:03:10.020 --> 00:03:12.509 Remember that? Oh, yeah. The mountains of bikes. 00:03:12.610 --> 00:03:15.569 Exactly. Investors just poured money in and millions 00:03:15.569 --> 00:03:17.930 of unused bikes ended up in landfills because 00:03:17.930 --> 00:03:19.590 the market just wasn't there yet. That's what 00:03:19.590 --> 00:03:22.169 the NDRC is trying to head off here. So despite 00:03:22.169 --> 00:03:24.150 all that investment in these huge projections, 00:03:24.389 --> 00:03:27.830 what is fundamentally stopping these 150 plus 00:03:27.830 --> 00:03:29.610 companies from actually getting into the real 00:03:29.610 --> 00:03:32.770 world right now? It really boils down to manufacturing 00:03:32.770 --> 00:03:36.689 cost and durability, specifically in the actuators. 00:03:37.180 --> 00:03:39.639 the motors in the hands and joints. They're just 00:03:39.639 --> 00:03:41.879 too expensive and they break too easily outside 00:03:41.879 --> 00:03:44.060 of a lab. Okay, let's switch gears and talk about 00:03:44.060 --> 00:03:46.500 the AI that is working, the software we use every 00:03:46.500 --> 00:03:49.560 day. The sources you sent over highlight this 00:03:49.560 --> 00:03:51.840 major shift in how we should be talking to these 00:03:51.840 --> 00:03:53.939 new models, especially something like Claude 00:03:53.939 --> 00:03:57.250 Opus 4 .5. This is such a crucial point for anyone 00:03:57.250 --> 00:04:00.030 using these tools. For years, we basically trained 00:04:00.030 --> 00:04:01.870 ourselves to over prompt. You know, that old 00:04:01.870 --> 00:04:04.289 habit of saying critical, you must use this exact 00:04:04.289 --> 00:04:07.430 JSON format and listing out 12 steps. Right. 00:04:07.590 --> 00:04:09.469 If you're still doing that, you are probably 00:04:09.469 --> 00:04:11.789 making the response worse. It's overkill. So 00:04:11.789 --> 00:04:14.050 all that extra detail, all that instruction that 00:04:14.050 --> 00:04:16.529 we thought was helping, it's actually hurting 00:04:16.529 --> 00:04:19.110 performance on these new models. That's totally 00:04:19.110 --> 00:04:21.209 counterintuitive. It really is. The big shift 00:04:21.209 --> 00:04:23.889 is that you have to recognize the model is already 00:04:23.889 --> 00:04:27.399 competent. When you give Opus 4 .5 this massive 00:04:27.399 --> 00:04:31.139 prompt, it takes you too literally. It gets bogged 00:04:31.139 --> 00:04:33.360 down in the how instead of just focusing on the 00:04:33.360 --> 00:04:36.120 what. So the takeaway is what? Treat it like 00:04:36.120 --> 00:04:38.699 a competent human colleague. Exactly. Prompt 00:04:38.699 --> 00:04:40.720 like you're texting that colleague. Just say 00:04:40.720 --> 00:04:43.060 what you want. Trust it to figure out the best 00:04:43.060 --> 00:04:46.300 way to get there. Less is truly more. Anthropic 00:04:46.300 --> 00:04:49.060 is even releasing tools like a concise output 00:04:49.060 --> 00:04:51.480 skill to help people break those old habits. 00:04:51.699 --> 00:04:53.620 Which tells you how hard that transition must 00:04:53.620 --> 00:04:55.279 be for people who've been doing this for a while. 00:04:55.439 --> 00:04:58.060 It is, and I'll admit this is something I still 00:04:58.060 --> 00:05:00.079 wrestle with myself. You know that prompt drift 00:05:00.079 --> 00:05:02.360 when you move between different models? The muscle 00:05:02.360 --> 00:05:04.740 memory of over -explaining is hard to shake. 00:05:04.879 --> 00:05:08.230 It takes real discipline to pull back. And shifting 00:05:08.230 --> 00:05:10.269 to the broader market, we've seen this incredible 00:05:10.269 --> 00:05:13.029 acceleration. ChatGPT is, what, three years old 00:05:13.029 --> 00:05:16.129 now? And it helped NVIDIA surge almost 1 ,000%. 00:05:16.129 --> 00:05:18.990 But at the same time, access seems to be tightening 00:05:18.990 --> 00:05:21.350 up. Oh, definitely. We're seeing free access 00:05:21.350 --> 00:05:24.029 get restricted everywhere. Google and OpenAI 00:05:24.029 --> 00:05:26.850 have limited free use of Gemini 3 Pro and Sora 00:05:26.850 --> 00:05:29.670 2. They're citing concerns that their GPUs are 00:05:29.670 --> 00:05:32.629 literally melting under the load. So access is 00:05:32.629 --> 00:05:35.290 becoming a real commodity. And speaking of limits, 00:05:35.370 --> 00:05:37.569 let's talk about that failure mode you flagged 00:05:37.569 --> 00:05:41.410 from the red teaming study. 62 % of top AI models 00:05:41.410 --> 00:05:44.730 failed when they were given poetic prompts, things 00:05:44.730 --> 00:05:47.709 that were, you know, abstract or metaphorical. 00:05:48.009 --> 00:05:50.209 This is where you see the cracks in the architecture. 00:05:50.449 --> 00:05:53.370 And the specific finding is just shocking. Google's 00:05:53.370 --> 00:05:56.910 Gemini 2 .5 Pro failed every single time. Every 00:05:56.910 --> 00:05:59.089 time. Wow. And it's because these models are 00:05:59.089 --> 00:06:01.069 brilliant at pattern matching. It's like stacking 00:06:01.069 --> 00:06:03.550 Lego blocks of data. But when you ask them to 00:06:03.550 --> 00:06:05.709 interpret a complex metaphor, they just they 00:06:05.709 --> 00:06:08.430 can't map that abstract idea onto a literal output. 00:06:08.670 --> 00:06:11.699 That feels like a fundamental blind spot. So 00:06:11.699 --> 00:06:13.860 if the best models are consistently failing on 00:06:13.860 --> 00:06:16.680 non -literal prompts, does that point to a really 00:06:16.680 --> 00:06:19.060 deep -seated limitation in their creative reasoning? 00:06:19.600 --> 00:06:22.819 Yes. It strongly suggests they struggle when 00:06:22.819 --> 00:06:25.300 instructions aren't literal and structured. It 00:06:25.300 --> 00:06:27.199 reveals a real lack of interpretive abstraction. 00:06:27.879 --> 00:06:31.000 So as we move into this real -world job shock, 00:06:31.370 --> 00:06:32.889 We have to look at the tools that are actually 00:06:32.889 --> 00:06:34.689 making it happen today. Let's just run through 00:06:34.689 --> 00:06:37.029 four quick examples because they show you where 00:06:37.029 --> 00:06:38.790 the economy is actually heading. Right. These 00:06:38.790 --> 00:06:40.529 are just theories. These are actual applications. 00:06:41.009 --> 00:06:43.129 Precisely. You've got Manus Browser Operator, 00:06:43.189 --> 00:06:45.009 which is a big deal because it can automate tasks 00:06:45.009 --> 00:06:47.509 on sites you have to be logged into. That used 00:06:47.509 --> 00:06:50.389 to require a person. Then there's Microsoft's 00:06:50.389 --> 00:06:54.310 VSA1. It makes hyper -realistic talking videos 00:06:54.310 --> 00:06:57.110 with perfect lip sync. That just changes the 00:06:57.110 --> 00:06:59.810 entire cost of making video content. So that's 00:06:59.810 --> 00:07:01.750 moving straight into... creative workflows it 00:07:01.750 --> 00:07:04.189 is and then you have for cells workflow builder 00:07:04.189 --> 00:07:07.490 which lets non coders create really complex automations 00:07:07.490 --> 00:07:11.250 just by dragging and dropping blocks and masonry 00:07:11.250 --> 00:07:13.410 which is this all -in -one tool for images and 00:07:13.410 --> 00:07:16.899 video and these kinds of tools are the building 00:07:16.899 --> 00:07:19.860 blocks that lead us right to this AI red alert 00:07:19.860 --> 00:07:24.079 and MIT's Project Iceberg. I mean, the scale 00:07:24.079 --> 00:07:27.319 of this thing is just immense. It tracks 151 00:07:27.319 --> 00:07:31.560 million workers, over 900 job types, and 32 ,000 00:07:31.560 --> 00:07:33.759 different skills. And that scale is what lets 00:07:33.759 --> 00:07:37.300 us understand the potential energy here. 11 .7 00:07:37.300 --> 00:07:40.939 % of all U .S. jobs. That's over 21 million roles 00:07:40.939 --> 00:07:44.379 accounting for $1 .2 trillion in wages that can 00:07:44.379 --> 00:07:46.699 be automated today. with what we have now. But 00:07:46.699 --> 00:07:49.300 we keep coming back to that low 2 .2 % surface 00:07:49.300 --> 00:07:52.379 index. So why? Why is there still this huge gap 00:07:52.379 --> 00:07:54.459 between what's possible and what's actually happening? 00:07:54.620 --> 00:07:56.759 The short answer is it was the connection. Before 00:07:56.759 --> 00:07:59.319 the AI was siloed, it was smart, but it couldn't 00:07:59.319 --> 00:08:01.620 plug into your company's Salesforce or your calendar 00:08:01.620 --> 00:08:03.920 or your ERP system. So the intelligence was there, 00:08:03.939 --> 00:08:05.620 but it was locked out of the business's operating 00:08:05.620 --> 00:08:08.139 system. Exactly. It couldn't do anything. That 00:08:08.139 --> 00:08:10.819 all changed in late 2024 with the Model Context 00:08:10.819 --> 00:08:13.540 Protocol, or MCP. Think of MCP as like a universal 00:08:13.540 --> 00:08:16.600 adapter for workflows. It lets any major AI model, 00:08:16.759 --> 00:08:18.800 Claude, Gemini, whatever, plug directly into 00:08:18.800 --> 00:08:20.540 real world tools and start acting on things, 00:08:20.639 --> 00:08:22.899 not just advising. So the MCP is the unlock. 00:08:23.019 --> 00:08:26.060 That's what turns the 11 .7 % from a theory into 00:08:26.060 --> 00:08:28.939 a reality. It is. And the speed of adoption is 00:08:28.939 --> 00:08:30.720 just staggering. We're not talking about pilot 00:08:30.720 --> 00:08:34.700 programs. As of March 2025, there are over 7 00:08:34.700 --> 00:08:40.600 ,950 MCP servers active in organizations. 7 ,950. 00:08:40.679 --> 00:08:43.799 That number. Mm -hmm. That gives us a clear metric 00:08:43.799 --> 00:08:45.840 for real -world adoption that isn't just about 00:08:45.840 --> 00:08:49.529 financial hype. Yeah. And, whoa! Just stop and 00:08:49.529 --> 00:08:52.289 imagine scaling that autonomous workflow management 00:08:52.289 --> 00:08:54.629 at that level across the entire global knowledge 00:08:54.629 --> 00:08:57.870 economy. It's happening. This is where AI agents 00:08:57.870 --> 00:09:00.190 go from being little assistants to autonomously 00:09:00.190 --> 00:09:02.250 managing things like calendars, booking travel, 00:09:02.490 --> 00:09:05.190 running reports, updating dashboards. So actual 00:09:05.190 --> 00:09:07.389 knowledge work is being managed by non -human 00:09:07.389 --> 00:09:09.710 entities at a massive scale. And you mentioned 00:09:09.710 --> 00:09:11.629 some U .S. states are already using this data 00:09:11.629 --> 00:09:13.450 from Project Iceberg right now. That's right. 00:09:13.529 --> 00:09:15.730 They're using these precise automation maps to 00:09:15.730 --> 00:09:18.090 plan job retraining programs and upskilling grants. 00:09:18.330 --> 00:09:19.990 They are moving money based on where the math 00:09:19.990 --> 00:09:21.570 shows the jobs are going to be hit the hardest. 00:09:21.830 --> 00:09:24.049 So if the model context protocol is the main 00:09:24.049 --> 00:09:26.809 thing bridging this gap, what's the one factor 00:09:26.809 --> 00:09:29.690 that now drives how fast the rest of that automation 00:09:29.690 --> 00:09:32.950 will happen? The pace now is entirely dependent 00:09:32.950 --> 00:09:35.970 on how fast more of these active MCP servers 00:09:35.970 --> 00:09:38.210 get integrated into organizations. That's the 00:09:38.210 --> 00:09:41.990 only bottleneck left. So to just synthesize everything 00:09:41.990 --> 00:09:44.759 we've covered. You see this huge disconnect, 00:09:44.960 --> 00:09:47.059 right? You've got massive hype like the humanoid 00:09:47.059 --> 00:09:48.799 robotics bubble. And then you have the current 00:09:48.799 --> 00:09:51.419 reality where things like prompt design and GPU 00:09:51.419 --> 00:09:53.740 limits are still very real constraints. Right. 00:09:53.799 --> 00:09:56.159 The core tension is that the capability to automate 00:09:56.159 --> 00:09:59.340 11 .7 % of jobs has been there. But the adoption, 00:09:59.559 --> 00:10:02.639 that 2 .2 % surface index was stalled. And the 00:10:02.639 --> 00:10:04.850 bridge that's being built. very, very quickly 00:10:04.850 --> 00:10:07.690 is this model context protocol. And that technological 00:10:07.690 --> 00:10:10.149 speed brings us right back to the financial and 00:10:10.149 --> 00:10:12.169 frankly political debate around who gets the 00:10:12.169 --> 00:10:14.690 value. The sources you shared highlight this 00:10:14.690 --> 00:10:17.750 messy debate around Genesis AI. We should probably 00:10:17.750 --> 00:10:20.429 define that. What is Genesis AI? It's basically 00:10:20.429 --> 00:10:23.490 the proposal for how we should tax or distribute 00:10:23.490 --> 00:10:26.529 the new value that's created purely by AI systems. 00:10:27.009 --> 00:10:29.710 You know, the value that comes when AI replaces 00:10:29.710 --> 00:10:32.269 human labor. And critics are calling this idea 00:10:33.009 --> 00:10:35.909 Socialism for the rich. Especially after big 00:10:35.909 --> 00:10:38.509 names like David Sachs did a complete 180 on 00:10:38.509 --> 00:10:41.059 the concept. They are. It's a really charged 00:10:41.059 --> 00:10:43.559 idea, but it raises this fundamental question 00:10:43.559 --> 00:10:46.960 for all of us. As this tech reshapes the whole 00:10:46.960 --> 00:10:49.940 economy and creates these vast new pools of capital, 00:10:50.139 --> 00:10:53.799 how should we as a society think about who really 00:10:53.799 --> 00:10:55.899 benefits from all this money flowing into AI? 00:10:56.139 --> 00:10:58.559 That is an uncomfortable question, but it's one 00:10:58.559 --> 00:11:00.460 we need to start asking now, you know, before 00:11:00.460 --> 00:11:03.879 that 2 .2 % gets a lot closer to 11 .7%. Absolutely. 00:11:04.409 --> 00:11:06.330 It's something to keep a very close eye on. We'd 00:11:06.330 --> 00:11:07.970 encourage you to think about your own prompting 00:11:07.970 --> 00:11:10.210 habits and maybe check if you're over -prompting 00:11:10.210 --> 00:11:12.389 your AI models and see if your state is using 00:11:12.389 --> 00:11:15.049 the Project Iceberg data for job planning. Thank 00:11:15.049 --> 00:11:17.590 you, as always, for sharing the sources for this 00:11:17.590 --> 00:11:18.830 deep dive. Until next time.