WEBVTT 00:00:00.000 --> 00:00:03.129 if you've ever uh gotten something from generative 00:00:03.129 --> 00:00:05.269 AI that looked OK on the surface, but just felt 00:00:05.269 --> 00:00:07.349 off, like it needed a ton of work. You're definitely 00:00:07.349 --> 00:00:09.650 not alone. We're talking about work slop. It's 00:00:09.650 --> 00:00:11.650 this AI generated stuff that kind of pretends 00:00:11.650 --> 00:00:14.550 to be productive work. But really, it just makes 00:00:14.550 --> 00:00:17.250 the next person down the line waste hours fixing 00:00:17.250 --> 00:00:19.269 it. It's like this invisible tax we're putting 00:00:19.269 --> 00:00:21.570 on ourselves inside companies. And it's a huge 00:00:21.570 --> 00:00:24.609 tax. I mean, one source we look at put a number 00:00:24.609 --> 00:00:27.210 on it for a, say, 10 ,000 person company. We're 00:00:27.210 --> 00:00:29.609 talking about nine million dollars a year lost. 00:00:30.629 --> 00:00:34.130 beat. That's just pure waste. Productivity down 00:00:34.130 --> 00:00:36.429 the drain, basically fueled by taking the easy 00:00:36.429 --> 00:00:39.310 digital way out. Welcome to the Deep Dive. Today, 00:00:39.429 --> 00:00:41.030 we're going to try and cut through some of that 00:00:41.030 --> 00:00:43.530 noise. We're using the sources you sent over 00:00:43.530 --> 00:00:47.289 to look at the hard reality of scaling AI, trying 00:00:47.289 --> 00:00:48.869 to figure out where it's actually delivering 00:00:48.869 --> 00:00:50.969 genius and where it's just making these expensive 00:00:50.969 --> 00:00:54.009 messes. Yeah, we've got a plan. First up, we'll 00:00:54.009 --> 00:00:56.369 dig deeper into this work slop thing, the actual 00:00:56.369 --> 00:00:59.200 cost. financially, but also psychologically. 00:00:59.719 --> 00:01:02.299 Old trust tax idea. Then second, we'll pivot. 00:01:02.420 --> 00:01:04.859 Look at how fast AI capabilities are actually 00:01:04.859 --> 00:01:08.319 jumping forward. I mean, from gaming to even 00:01:08.319 --> 00:01:12.299 warnings about the job market. And finally, we'll 00:01:12.299 --> 00:01:14.700 tackle the big one. This potential AI energy 00:01:14.700 --> 00:01:17.739 apocalypse. Is this massive build -out of power 00:01:17.739 --> 00:01:20.620 infrastructure really needed? Or is it maybe 00:01:20.620 --> 00:01:23.200 speculative hype? Right. The goal here is really 00:01:23.200 --> 00:01:26.480 separating that hype from the... The hard reality 00:01:26.480 --> 00:01:28.239 when you actually try to deploy this stuff at 00:01:28.239 --> 00:01:30.599 scale. Let's get into it. So the basic idea behind 00:01:30.599 --> 00:01:33.000 work slop, it's really just pushing the cost 00:01:33.000 --> 00:01:34.859 onto someone else, isn't it? Someone hits generate, 00:01:35.000 --> 00:01:36.900 maybe glances at it. Feels good because they 00:01:36.900 --> 00:01:38.920 send something. Reductive. Right. But then the 00:01:38.920 --> 00:01:41.379 person who gets it might spend, what, two hours 00:01:41.379 --> 00:01:43.079 cleaning it up? They end up eating the entire 00:01:43.079 --> 00:01:45.579 cost of that quick solution. Exactly. It's the 00:01:45.579 --> 00:01:47.840 hidden cost of their convenience. Passed straight 00:01:47.840 --> 00:01:50.540 to their colleague. And, you know, this isn't 00:01:50.540 --> 00:01:53.780 rare. The data we saw suggests 40 % of employees 00:01:53.780 --> 00:01:56.519 got work slop just last month. and maybe 15 % 00:01:56.519 --> 00:01:57.859 of all the stuff flying around the workplace, 00:01:58.159 --> 00:02:00.859 emails, reports, code, whatever, is basically 00:02:00.859 --> 00:02:04.180 slop now. Scale that up, right? The estimate 00:02:04.180 --> 00:02:07.840 was about $186 per incident per month. That's 00:02:07.840 --> 00:02:09.900 how you get to that $9 million figure pretty 00:02:09.900 --> 00:02:12.080 fast in a big company. This may be less like 00:02:12.080 --> 00:02:14.719 stacking Lego blocks badly and more like, hmm, 00:02:14.860 --> 00:02:17.360 like building software on a shaky foundation. 00:02:17.460 --> 00:02:20.240 The code looks okay, maybe compiles. But try 00:02:20.240 --> 00:02:22.669 adding a new feature. Everything breaks. You 00:02:22.669 --> 00:02:25.110 face huge delays, costly rewrites you have to 00:02:25.110 --> 00:02:27.509 do. It's unstable underneath. Okay. That analogy 00:02:27.509 --> 00:02:30.009 makes more sense for maybe our technical listeners. 00:02:30.189 --> 00:02:32.330 But I still have to push a little on that 9 million 00:02:32.330 --> 00:02:34.189 number. Is there a chance that's consultants 00:02:34.189 --> 00:02:36.310 kind of hyping the problem to sell solutions? 00:02:36.629 --> 00:02:39.030 Or is it a solid measured cost? That's a really 00:02:39.030 --> 00:02:40.930 important question. And yeah, consultants are 00:02:40.930 --> 00:02:42.810 always going to sell solutions, right? But this 00:02:42.810 --> 00:02:47.009 specific calculation seems based on... documented 00:02:47.009 --> 00:02:50.090 time sinks. Things like average hourly wage times 00:02:50.090 --> 00:02:52.689 the hours people report spending just verifying 00:02:52.689 --> 00:02:55.849 or rewriting this AI output. So you can quibble 00:02:55.849 --> 00:02:58.169 with the exact dollar amount maybe, but the cost 00:02:58.169 --> 00:03:00.069 is definitely real. And it's not just money. 00:03:00.169 --> 00:03:02.870 It's the human side. This trust tax idea. You 00:03:02.870 --> 00:03:05.689 keep sending out subpar AI stuff and people notice. 00:03:05.870 --> 00:03:08.789 Over half, 53%, they get annoyed. Okay, fine, 00:03:08.870 --> 00:03:11.830 but worse. 42 % start seeing the sender as less 00:03:11.830 --> 00:03:15.150 trustworthy. And 37 % literally think they're 00:03:15.150 --> 00:03:17.659 less intelligent. Ouch. Your credibility just 00:03:17.659 --> 00:03:19.919 tanks. That's the real danger, systemically. 00:03:20.240 --> 00:03:22.000 And that happens when companies just say use 00:03:22.000 --> 00:03:24.139 AI everywhere, right, without clear rules. So 00:03:24.139 --> 00:03:26.219 people become passengers just using it to dodge 00:03:26.219 --> 00:03:29.159 work instead of being pilots who use it to genuinely 00:03:29.159 --> 00:03:31.879 amplify their own skills, their creativity. I'll 00:03:31.879 --> 00:03:33.759 admit, I still wrestle with this temptation myself 00:03:33.759 --> 00:03:37.020 sometimes. You know, prompt drift, getting that 00:03:37.020 --> 00:03:39.180 easy answer that looks good initially, but, you 00:03:39.180 --> 00:03:41.099 know, deep down, it's going to need five more 00:03:41.099 --> 00:03:42.860 rounds of checks and edits before it's actually 00:03:42.860 --> 00:03:45.500 usable. Good point. Let's quickly define prompt 00:03:45.500 --> 00:03:47.639 drift for everyone because it's key to work slop. 00:03:48.020 --> 00:03:50.229 It's basically when the AI. This output quality 00:03:50.229 --> 00:03:53.469 gets worse over a long chat or lots of edits. 00:03:53.750 --> 00:03:56.569 Like the model loses the plot, loses context, 00:03:56.629 --> 00:03:59.270 and just starts spitting out nonsense. So given 00:03:59.270 --> 00:04:01.509 this trust problem and people acting like passengers, 00:04:01.669 --> 00:04:03.870 what's the single most critical thing leaders 00:04:03.870 --> 00:04:06.590 need to do? to shift people towards being pilots. 00:04:07.610 --> 00:04:09.750 Leaders need to prioritize critical thinking 00:04:09.750 --> 00:04:12.409 and verification over just speed. Okay, so that 00:04:12.409 --> 00:04:14.550 work slop issue, that's one side of the coin. 00:04:14.590 --> 00:04:16.610 It's misusing a powerful tool, basically. But 00:04:16.610 --> 00:04:18.189 the irony is while we're struggling with that, 00:04:18.310 --> 00:04:21.430 the tech itself is just leaping forward exponentially, 00:04:21.930 --> 00:04:24.009 achieving things that are genuinely, you know, 00:04:24.009 --> 00:04:26.649 genius level. Oh, the pace is wild. Look at connectivity. 00:04:27.709 --> 00:04:30.889 ChatGPT connectors, which are just... Tools linking 00:04:30.889 --> 00:04:32.810 language models to other software for automation. 00:04:32.829 --> 00:04:35.310 They now work with over 500 apps. Automation 00:04:35.310 --> 00:04:36.889 isn't some future thing. It's right there. Plug 00:04:36.889 --> 00:04:38.850 and play, pretty much. And the models are getting 00:04:38.850 --> 00:04:41.930 smarter, deeper. Google's Gemini AI, for instance, 00:04:42.050 --> 00:04:44.509 it's showing how it can think through a search 00:04:44.509 --> 00:04:46.810 query, not just matching keywords, but producing 00:04:46.810 --> 00:04:49.329 these detailed summaries. Better reasoning. And 00:04:49.329 --> 00:04:54.250 then there's this. This moment that really makes 00:04:54.250 --> 00:04:57.449 you pause. GPT -5 codex. It apparently one -shotted 00:04:57.449 --> 00:04:59.449 Minecraft. Just think about that. One prompt 00:04:59.449 --> 00:05:01.949 in plain English. And it builds a whole 3D world. 00:05:02.050 --> 00:05:04.529 Textures, water, physics, the works. Whoa. I 00:05:04.529 --> 00:05:07.329 mean, imagine scaling that ability. Solving incredibly 00:05:07.329 --> 00:05:09.689 complex real -world problems with that kind of 00:05:09.689 --> 00:05:12.529 power. That's huge potential. Kind of terrifying, 00:05:12.689 --> 00:05:15.509 too. Yeah, but that power surge has a flip side 00:05:15.509 --> 00:05:18.810 as it filters into daily life. You see meta using 00:05:18.810 --> 00:05:22.199 AI and Facebook dating now. trying to optimize 00:05:22.199 --> 00:05:26.540 human connection. And in the really dark end, 00:05:26.660 --> 00:05:30.920 those hyper -realistic AI -generated murder videos 00:05:30.920 --> 00:05:33.720 made with VO that showed up on YouTube only got 00:05:33.720 --> 00:05:35.560 taken down because a journalist flagged them. 00:05:35.699 --> 00:05:38.139 The tech moves faster than our ability to manage 00:05:38.139 --> 00:05:40.399 it or even moderate it sometimes. And that speed 00:05:40.399 --> 00:05:43.079 crashes right into the economy. Which brings 00:05:43.079 --> 00:05:45.480 us to that stark warning from the Wharton professor 00:05:45.480 --> 00:05:47.829 you mentioned. Their research suggests this kind 00:05:47.829 --> 00:05:49.730 of exponential leap like the Minecraft thing 00:05:49.730 --> 00:05:52.230 shows is possible could make a lot of human jobs, 00:05:52.310 --> 00:05:55.250 especially white -collar ones, obsolete and fast. 00:05:55.470 --> 00:05:57.329 The timelines seem to be shrinking way faster 00:05:57.329 --> 00:05:59.410 than people thought. Right, so we need to ask, 00:05:59.589 --> 00:06:01.529 does this sheer speed of deployment building 00:06:01.529 --> 00:06:04.050 a whole game world instantly mean job losses 00:06:04.050 --> 00:06:06.269 are coming much, much sooner than most predictions? 00:06:06.610 --> 00:06:09.110 Yeah, seems like it. The exponential leap demands 00:06:09.110 --> 00:06:12.870 we start preparing now for big job shifts. Okay, 00:06:12.949 --> 00:06:15.850 we've covered the software side. The slop and 00:06:15.850 --> 00:06:18.870 the genius. Let's switch gears to the hardware. 00:06:19.110 --> 00:06:22.149 The physical infrastructure. Specifically, the 00:06:22.149 --> 00:06:24.410 massive energy needed to actually run these things. 00:06:24.550 --> 00:06:27.470 Oh, it's staggering. One modern AI data rack. 00:06:27.670 --> 00:06:30.910 It pulls the same power as maybe 80 to 100 average 00:06:30.910 --> 00:06:33.509 homes. Compare that to a traditional server rack, 00:06:33.610 --> 00:06:36.139 maybe three homes worth of power. So this need 00:06:36.139 --> 00:06:39.180 for huge computation is driving these gigantic 00:06:39.180 --> 00:06:41.860 energy requests to utilities everywhere. We're 00:06:41.860 --> 00:06:43.959 talking gigawatts. But here's the tension, right? 00:06:44.040 --> 00:06:47.500 This predicted AI energy apocalypse. Maybe it's 00:06:47.500 --> 00:06:49.319 not all based on solid confirmed need. Maybe 00:06:49.319 --> 00:06:52.300 there's a lot of speculative hype driving overbuilding. 00:06:52.439 --> 00:06:54.259 The signs of potential inflation are definitely 00:06:54.259 --> 00:06:56.800 there. Since this whole AI boom kicked off, proposals 00:06:56.800 --> 00:06:59.319 for new natural gas power plants are up 70%, 00:06:59.319 --> 00:07:01.860 mostly earmarked for future data centers. In 00:07:01.860 --> 00:07:04.120 the U .S. Southeast, the demand utilities are 00:07:04.120 --> 00:07:07.220 forecasting is four times higher than what independent 00:07:07.220 --> 00:07:09.600 researchers think the actual need will be. Four 00:07:09.600 --> 00:07:12.740 times. Four times. That's a massive difference. 00:07:12.879 --> 00:07:15.819 How does that even happen? Why are utilities 00:07:15.819 --> 00:07:18.939 seemingly accepting forecasts that might be wildly 00:07:18.939 --> 00:07:21.500 inflated? What's the mechanism? Well, it seems 00:07:21.500 --> 00:07:23.800 like a mix of things. System incentives, maybe 00:07:23.800 --> 00:07:26.360 some regulatory capture. Developers put in these 00:07:26.360 --> 00:07:28.279 huge requests often before they even have the 00:07:28.279 --> 00:07:31.199 money lined up or firm customers. Speculative 00:07:31.199 --> 00:07:33.939 requests. Utilities often accept them, perhaps 00:07:33.939 --> 00:07:35.860 to lock down the rights to build generations, 00:07:36.019 --> 00:07:38.360 secure land, get permits. They call it nameplate 00:07:38.360 --> 00:07:40.560 capacity. It's like a land grab for future power 00:07:40.560 --> 00:07:43.819 needs based on potential, not guaranteed use. 00:07:44.060 --> 00:07:46.420 And apparently even utilities admit privately 00:07:46.420 --> 00:07:48.459 the real usage might end up being, you know, 00:07:48.459 --> 00:07:50.420 three to five times smaller than what's on paper 00:07:50.420 --> 00:07:52.899 right now. OK, so if that high bubble bursts 00:07:52.899 --> 00:07:55.639 or even just shrinks significantly, who pays 00:07:55.639 --> 00:07:58.360 for all that potentially unnecessary gas infrastructure 00:07:58.360 --> 00:08:00.540 they build? It's going to be the consumers, right? 00:08:00.660 --> 00:08:03.180 Through higher electricity bills, the costs get 00:08:03.180 --> 00:08:06.100 socialized. Exactly. Stranded assets paid for 00:08:06.100 --> 00:08:09.430 by ratepayers and the environmental cost. Also 00:08:09.430 --> 00:08:12.009 huge. One example cited was a metadata center 00:08:12.009 --> 00:08:14.910 project down in Louisiana. If all that speculative 00:08:14.910 --> 00:08:17.350 demand actually happens, it could lock in over 00:08:17.350 --> 00:08:20.889 100 million tons of extra carbon emissions just 00:08:20.889 --> 00:08:23.029 from that one project's power needs over its 00:08:23.029 --> 00:08:25.490 life. So this rush to build feels like it creates 00:08:25.490 --> 00:08:28.259 a major policy clash. It lines up nicely with 00:08:28.259 --> 00:08:32.379 political agendas favoring fossil fuels, but 00:08:32.379 --> 00:08:35.000 it runs directly against stated goals like getting 00:08:35.000 --> 00:08:37.960 to a carbon -free grid by, say, 2035. It feels 00:08:37.960 --> 00:08:39.860 like we're planning a fossil fuel future based 00:08:39.860 --> 00:08:42.139 on maybe a ghost demand. Yeah, it's a real paradox. 00:08:42.220 --> 00:08:44.779 So if utilities know they might be over forecasting 00:08:44.779 --> 00:08:47.980 demand by potentially 4X, what's the single biggest 00:08:47.980 --> 00:08:50.460 systemic risk here for energy policy for all 00:08:50.460 --> 00:08:52.899 of us consumers? Consumers risk paying for unnecessary 00:08:52.899 --> 00:08:56.000 fossil fuel plants based on inflated AI hype. 00:08:57.840 --> 00:09:03.440 Okay, let's try to pull the threads together 00:09:03.440 --> 00:09:05.899 from this deep dive. What we're seeing is this 00:09:05.899 --> 00:09:09.860 fundamental AI paradox, right? On one hand, it's 00:09:09.860 --> 00:09:12.820 creating huge new inefficiencies inside companies 00:09:12.820 --> 00:09:15.639 that work slop, costing millions. But on the 00:09:15.639 --> 00:09:17.600 other hand, it's achieving these incredible genius 00:09:17.600 --> 00:09:20.360 level things that could totally reshape the job 00:09:20.360 --> 00:09:22.639 market. Yeah, the connection seems to be this 00:09:22.639 --> 00:09:25.580 deep misalignment everywhere. Misalignment between 00:09:25.580 --> 00:09:28.120 the tools capability and how it's actually being 00:09:28.120 --> 00:09:30.539 applied. That's the whole passengers versus pilots 00:09:30.539 --> 00:09:33.580 thing causing the slop. And then this huge misalignment 00:09:33.580 --> 00:09:36.360 between the real energy needs right now versus 00:09:36.360 --> 00:09:39.409 this potentially overblown. speculative infrastructure 00:09:39.409 --> 00:09:42.230 build out based on hype. It feels like, based 00:09:42.230 --> 00:09:44.769 on the sources, we're building this massive energy 00:09:44.769 --> 00:09:47.110 future based on the promise of AI, what it might 00:09:47.110 --> 00:09:49.870 consume someday. But maybe not on its proven, 00:09:49.970 --> 00:09:52.799 confirmed necessity today. So here's a provocative 00:09:52.799 --> 00:09:55.000 thought for you to chew on this week. Thinking 00:09:55.000 --> 00:09:58.039 about that work slot, what existing human institution, 00:09:58.259 --> 00:10:00.460 maybe a big government agency, a traditional 00:10:00.460 --> 00:10:02.980 industry, maybe even a university, what would 00:10:02.980 --> 00:10:04.940 be the first one to really start crumbling under 00:10:04.940 --> 00:10:07.299 the weight of all that AI -generated inefficiency? 00:10:07.679 --> 00:10:10.539 That is a fascinating and slightly unsettling 00:10:10.539 --> 00:10:12.059 thought exercise. So your call to action this 00:10:12.059 --> 00:10:14.360 week, just be mindful. Pay attention to the AI 00:10:14.360 --> 00:10:16.720 -generated content you encounter. Ask yourself, 00:10:16.840 --> 00:10:19.929 is this saving time? Is it quality? Or is it 00:10:19.929 --> 00:10:22.029 slop? Are you seeing genius? Or just more digital 00:10:22.029 --> 00:10:24.509 noise? Thanks for joining us for this deep dive 00:10:24.509 --> 00:10:26.889 into AI's hidden costs and its undeniable power. 00:10:27.110 --> 00:10:28.649 Otiro music fades in.