WEBVTT 00:00:00.000 --> 00:00:04.599 Imagine a single machine, just one. But this 00:00:04.599 --> 00:00:07.000 machine draws more electrical power than the 00:00:07.000 --> 00:00:09.660 entire city of San Francisco. Wow. And that's 00:00:09.660 --> 00:00:11.919 not, you know, science fiction. As of right now, 00:00:12.060 --> 00:00:15.359 January 19th, 2026, it is a physical reality 00:00:15.359 --> 00:00:18.039 humming away on the grid. That is the kind of 00:00:18.039 --> 00:00:19.679 stat that just stops you in your tracks. Yeah. 00:00:19.719 --> 00:00:21.359 I mean, we are not talking about a server farm 00:00:21.359 --> 00:00:23.600 anymore. We're talking about energy use on the 00:00:23.600 --> 00:00:26.140 scale of a, well, a small country. Welcome to 00:00:26.140 --> 00:00:28.859 the Deep Dive. Today we're navigating a world 00:00:28.859 --> 00:00:31.920 of some pretty wild extremes. We've reached a 00:00:31.920 --> 00:00:34.539 point where the infrastructure for intelligence 00:00:34.539 --> 00:00:37.420 is becoming truly titanic. Literal city -sized 00:00:37.420 --> 00:00:40.079 brain. Exactly. And yet at the exact same moment 00:00:40.079 --> 00:00:42.179 people are what fleeing to the countryside to 00:00:42.179 --> 00:00:44.460 knit sweaters. It's just we have the biggest 00:00:44.460 --> 00:00:46.979 computers in history and also a massive spike 00:00:46.979 --> 00:00:50.159 in yarn sales. It's a fascinating dichotomy. 00:00:50.280 --> 00:00:52.820 It is. And we're going to unpack that tension. 00:00:52.960 --> 00:00:55.140 We have a stack of reports today covering the 00:00:55.140 --> 00:00:58.920 activation of XAI's Colossus 2, this strange 00:00:58.920 --> 00:01:02.079 cultural thing called the analog backlash, and 00:01:02.079 --> 00:01:04.760 some really interesting research from DeepMind 00:01:04.760 --> 00:01:07.599 on how to keep these massive systems safe without 00:01:07.599 --> 00:01:10.099 going broke. Plus, on a much more practical level, 00:01:10.219 --> 00:01:12.739 we really need to talk about why copying and 00:01:12.739 --> 00:01:15.819 pasting from an AI still completely ruins your 00:01:15.819 --> 00:01:17.900 formatting. Yes. Let's start with the heavy metal, 00:01:17.980 --> 00:01:19.819 the hardware. We're looking at these reports 00:01:19.819 --> 00:01:23.260 on XAI, Elon Musk's AI company, and this new 00:01:23.260 --> 00:01:26.420 cluster, Colossus 2. Cluster just feels like 00:01:26.420 --> 00:01:28.739 such a small word for what this actually is. 00:01:28.859 --> 00:01:31.859 I mean, to set the scene, XAI has just activated 00:01:31.859 --> 00:01:35.000 a one gigawatt supercomputer. One gigawatt. Yeah. 00:01:35.099 --> 00:01:37.920 And it's purpose built for one thing, training 00:01:37.920 --> 00:01:40.400 the Grok series of models. And just for context, 00:01:40.540 --> 00:01:42.879 that's roughly the output of a large nuclear 00:01:42.879 --> 00:01:45.480 reactor, right? Exactly. They have built a machine 00:01:45.480 --> 00:01:47.920 that effectively requires a nuclear reactor's 00:01:47.920 --> 00:01:50.840 worth of juice just to turn on. And the timeline 00:01:50.840 --> 00:01:53.900 is just while Colossus 1 took, what, about 122 00:01:53.900 --> 00:01:56.400 days to bring online, Colossus 2 followed immediately 00:01:56.400 --> 00:01:58.859 after. Just bang, bang. And looking at the specs 00:01:58.859 --> 00:02:01.000 here, it's not even stopping at one gigawatt. 00:02:01.079 --> 00:02:03.120 Oh, no, that's just the starting line. The reports 00:02:03.120 --> 00:02:05.159 are saying it's already set to expand to one 00:02:05.159 --> 00:02:08.080 and a half gigawatts by April and then hit two 00:02:08.080 --> 00:02:10.650 gigawatts soon after that. The logistics of that 00:02:10.650 --> 00:02:12.650 are, well, they're mind -bending. It has to sit 00:02:12.650 --> 00:02:15.050 on dedicated grid lines. It has its own on -site 00:02:15.050 --> 00:02:18.909 substations. And the heat, just imagine the thermodynamics 00:02:18.909 --> 00:02:22.990 of, what, 555 ,000 GPUs running at full tilt. 00:02:23.169 --> 00:02:25.250 Yeah, the cooling systems alone must be an engineering 00:02:25.250 --> 00:02:28.849 marvel. It's liquid cooling on just a ridiculous 00:02:28.849 --> 00:02:32.710 scale. And you mentioned the GPUs, 555 ,000 of 00:02:32.710 --> 00:02:34.909 them. The procurement number is just eye -watering. 00:02:35.180 --> 00:02:37.699 The cost for that hardware alone is estimated 00:02:37.699 --> 00:02:41.580 at roughly $18 billion. $18 billion. That suddenly 00:02:41.580 --> 00:02:44.240 explains why XAI blew right past their original 00:02:44.240 --> 00:02:46.800 $15 billion funding goal. Yeah, if you want to 00:02:46.800 --> 00:02:48.599 play at this table, the buy -in is astronomical. 00:02:48.979 --> 00:02:51.520 They ended up raising a $20 billion Series E, 00:02:51.719 --> 00:02:54.280 backed by, you know, all the heavy hitters, Fidelity, 00:02:54.400 --> 00:02:57.879 Nvidia, Cisco. And the stated mission here, the 00:02:57.879 --> 00:03:00.900 philosophy behind spending all this money, is 00:03:00.900 --> 00:03:03.580 to understand the universe. They claim that to 00:03:03.580 --> 00:03:05.460 do that, you need the world's biggest brain. 00:03:05.719 --> 00:03:07.960 Right. And practically speaking, this machine 00:03:07.960 --> 00:03:11.120 is what's going to power Grok 4, Grok Voice, 00:03:11.300 --> 00:03:14.219 and the upcoming Grok 5. Technologically, though, 00:03:14.300 --> 00:03:16.319 what stands out to me is that they've kind of 00:03:16.319 --> 00:03:19.300 hit a ceiling, but not on hardware. It's a software 00:03:19.300 --> 00:03:21.960 ceiling. Right. The reports say that the global 00:03:21.960 --> 00:03:25.199 batch sizes for training are now limited by optimization, 00:03:25.500 --> 00:03:28.139 not by compute. They effectively have enough 00:03:28.139 --> 00:03:30.599 hardware to do continuous trillion parameter 00:03:30.599 --> 00:03:33.699 training. It's just there. It's brute force on 00:03:33.699 --> 00:03:36.949 a celestial scale, which I have to ask. Does 00:03:36.949 --> 00:03:38.849 building a brain the size of a city actually 00:03:38.849 --> 00:03:41.370 guarantee understanding? Or are we just building 00:03:41.370 --> 00:03:43.490 the most expensive pattern matching machine in 00:03:43.490 --> 00:03:45.870 human history? That is the multibillion dollar 00:03:45.870 --> 00:03:47.389 question, isn't it? But I guess when you have 00:03:47.389 --> 00:03:49.810 half a million GPUs, the strategy is pretty clear 00:03:49.810 --> 00:03:51.810 that quantity might just have a quality all its 00:03:51.810 --> 00:03:54.430 own. The hope is that understanding just kind 00:03:54.430 --> 00:03:56.349 of emerges from the math if you make the math 00:03:56.349 --> 00:03:59.889 big enough. OK, so while XAI is building the 00:03:59.889 --> 00:04:02.620 digital equivalent of a Dyson sphere. the rest 00:04:02.620 --> 00:04:04.939 of us, the human world, we seem to be reacting 00:04:04.939 --> 00:04:07.199 in a very, very different way. Yeah. Welcome 00:04:07.199 --> 00:04:10.740 to what some are calling the 2026 paradox. On 00:04:10.740 --> 00:04:13.819 one side, the digital economy is absolutely on 00:04:13.819 --> 00:04:17.459 fire. OpenAI has revealed they hit $20 billion 00:04:17.459 --> 00:04:21.060 in annualized revenue for 2025. That's triple 00:04:21.060 --> 00:04:23.439 the year before. The numbers are vertical. Totally. 00:04:23.519 --> 00:04:26.740 And Gemini API traffic doubled in just five months. 00:04:26.839 --> 00:04:28.680 But then you look at the cultural data. And you 00:04:28.680 --> 00:04:31.540 get yarn. You get yarn. I'm not kidding. Sales 00:04:31.540 --> 00:04:34.939 of knitting kits at Michaels have jumped 1 ,000. 00:04:35.599 --> 00:04:38.199 200%. Wow. Searches for analog lifestyle have 00:04:38.199 --> 00:04:40.939 just exploded. It's this weird vibe where for 00:04:40.939 --> 00:04:43.540 a lot of people, 2026 feels a lot like 1996. 00:04:44.079 --> 00:04:46.540 It feels like a retreat. Like while the algorithms 00:04:46.540 --> 00:04:48.399 are eating the Internet, people are just craving 00:04:48.399 --> 00:04:50.660 something tactile, something they can, you know, 00:04:50.680 --> 00:04:52.839 control from start to finish without a prompt 00:04:52.839 --> 00:04:55.079 window. Well, there's some real economic anxiety 00:04:55.079 --> 00:04:57.540 under that, too. Economists are warning that 00:04:57.540 --> 00:04:59.839 2026 is the year this all hits the political 00:04:59.839 --> 00:05:03.139 radar. AI isn't necessarily taking everyone's 00:05:03.139 --> 00:05:06.540 job overnight. What it is doing is deleting the 00:05:06.540 --> 00:05:08.819 on -ramp. The entry -level positions. Exactly. 00:05:08.819 --> 00:05:11.060 All the junior work is being done by the models 00:05:11.060 --> 00:05:13.319 now. So it's getting harder and harder for beginners 00:05:13.319 --> 00:05:15.379 to break into these industries. You have this 00:05:15.379 --> 00:05:18.319 squeeze where the digital elite are generating 00:05:18.319 --> 00:05:20.899 massive revenue, and a lot of other people are 00:05:20.899 --> 00:05:23.000 feeling locked out. And they knit. They knit. 00:05:23.199 --> 00:05:25.939 It's a way to feel competent, I think. When you 00:05:25.939 --> 00:05:28.040 knit a scarf, you know exactly how it was made. 00:05:28.139 --> 00:05:31.680 There are no hallucinations in wool. No hallucinations 00:05:31.680 --> 00:05:34.560 in wool. I like that. But for those of us who 00:05:34.560 --> 00:05:37.399 do have to work in this digital storm, we can't 00:05:37.399 --> 00:05:39.540 just go knit all day. We have to use these tools. 00:05:39.980 --> 00:05:43.740 And honestly, it's still kind of a mess. It really 00:05:43.740 --> 00:05:45.160 is. I was reading through the source material 00:05:45.160 --> 00:05:48.060 on perfectly formatted reports, and it just... 00:05:48.670 --> 00:05:50.910 it hit home we talk about super intelligence 00:05:50.910 --> 00:05:53.149 but most of us are just struggling to get a document 00:05:53.149 --> 00:05:55.810 to look right you spend hours on a prompt you 00:05:55.810 --> 00:05:58.269 get great text you paste it into google docs 00:05:58.269 --> 00:06:02.129 and It just explodes. The formatting nightmare. 00:06:02.449 --> 00:06:05.529 Giant bold text where a header should be. Bullet 00:06:05.529 --> 00:06:08.009 points that turn into random dashes. It just 00:06:08.009 --> 00:06:10.550 looks so amateur. I still wrestle with prompt 00:06:10.550 --> 00:06:12.829 drift myself. You spend like three hours trying 00:06:12.829 --> 00:06:14.490 to get the headers just right, and the model 00:06:14.490 --> 00:06:17.310 just gives you bold text instead of proper H1 00:06:17.310 --> 00:06:20.370 tags. It's infuriating. The friction is so real. 00:06:20.589 --> 00:06:23.269 The whole promise of AI is speed. But if you 00:06:23.269 --> 00:06:25.810 spend 20 minutes reformatting font sizes and 00:06:25.810 --> 00:06:27.850 bullet points, you haven't actually saved any 00:06:27.850 --> 00:06:30.399 time. So what's the fix? The sources mentioned 00:06:30.399 --> 00:06:33.500 some new empowered tools. Well, Google has rolled 00:06:33.500 --> 00:06:35.899 out this personal intelligence thing in Gemini. 00:06:36.279 --> 00:06:39.439 It's supposed to act like a, quote, weirdly well 00:06:39.439 --> 00:06:41.959 -informed best friend that remembers your context. 00:06:42.240 --> 00:06:44.019 But for the formatting problem specifically, 00:06:44.259 --> 00:06:46.480 the solution seems to be learning how to force 00:06:46.480 --> 00:06:49.000 the output. What do you mean by force it? You 00:06:49.000 --> 00:06:51.759 have to stop asking nicely. You need to be incredibly 00:06:51.759 --> 00:06:54.319 specific about the underlying code. You have 00:06:54.319 --> 00:06:56.639 to tell it, I'll put this in Markdown that maps 00:06:56.639 --> 00:07:00.899 to the specific H1 and H2 tags. Or you use these 00:07:00.899 --> 00:07:03.519 new bridge tools like Noodle Seed or Flow Genie 00:07:03.519 --> 00:07:06.699 that structure the data before it even hits the 00:07:06.699 --> 00:07:09.279 document. So we have to learn to speak the machine's 00:07:09.279 --> 00:07:11.399 language just to get it to speak ours properly. 00:07:11.579 --> 00:07:14.040 Pretty much. Until the AI can intuitively understand 00:07:14.040 --> 00:07:16.839 your company's brand style guide, you have to 00:07:16.839 --> 00:07:20.110 manually override its default chaos. It really 00:07:20.110 --> 00:07:22.029 brings up a question of utility then. If we have 00:07:22.029 --> 00:07:24.589 to fight the AI to format a simple page, are 00:07:24.589 --> 00:07:27.350 we really more productive yet? Only if you master 00:07:27.350 --> 00:07:29.670 those force formatting prompts. Otherwise, yeah, 00:07:29.730 --> 00:07:31.790 it's often just a very messy copy paste job. 00:07:32.029 --> 00:07:33.829 Now, speaking of looking under the hood, let's 00:07:33.829 --> 00:07:35.850 pivot to something arguably more critical than 00:07:35.850 --> 00:07:39.709 font sizes, safety. With models as big as Colossus 00:07:39.709 --> 00:07:43.290 2, the question of control is, it's massive. 00:07:43.569 --> 00:07:45.529 Yeah, and DeepMind just dropped a paper on this 00:07:45.529 --> 00:07:47.610 that is... Honestly, a game changer. This is 00:07:47.610 --> 00:07:50.649 the activation probes paper? Yes. And if you 00:07:50.649 --> 00:07:52.930 care about AI safety, but you also care about, 00:07:53.009 --> 00:07:56.810 you know, not lighting money on fire, this is 00:07:56.810 --> 00:07:58.990 huge. The problem they're really solving here 00:07:58.990 --> 00:08:02.569 is cost, right? Monitoring these huge models 00:08:02.569 --> 00:08:05.550 for safety usually costs a fortune. An insane 00:08:05.550 --> 00:08:08.350 amount. Traditionally, to check if an AI is being 00:08:08.350 --> 00:08:10.889 toxic or dangerous, you have to run its output 00:08:10.889 --> 00:08:13.959 through another AI to check it. It's like hiring 00:08:13.959 --> 00:08:17.079 a supervisor to watch every single employee 24 00:08:17.079 --> 00:08:19.699 -7. And that supervisor demands a big salary. 00:08:19.860 --> 00:08:22.720 A huge one in compute power. So DeepMind's solution 00:08:22.720 --> 00:08:25.439 is to stop watching the output and start watching 00:08:25.439 --> 00:08:27.839 the brain itself. So as it's thinking. Precisely. 00:08:28.089 --> 00:08:30.449 As the AI is processing a request, it's doing 00:08:30.449 --> 00:08:32.870 all this internal matrix math. So DeepMind developed 00:08:32.870 --> 00:08:35.529 what they call activation probes. And you can 00:08:35.529 --> 00:08:38.809 think of them like tiny sensors or like a tiny 00:08:38.809 --> 00:08:41.090 brain scan. They scan the model's internal activity 00:08:41.090 --> 00:08:42.750 while it's thinking. So it's like reading the 00:08:42.750 --> 00:08:44.669 mind before the mouth opens. That's the perfect 00:08:44.669 --> 00:08:47.090 analogy. It's eavesdropping on the inner monologue. 00:08:47.350 --> 00:08:49.250 And because it's just reading data that is already 00:08:49.250 --> 00:08:50.990 being computed anyway, it's incredibly cheap. 00:08:51.070 --> 00:08:54.269 How cheap are we talking? We're talking 10 ,000 00:08:54.269 --> 00:08:56.450 times cheaper than running a full safety monitor 00:08:56.450 --> 00:09:00.309 on top. 10 ,000 times. That effectively makes 00:09:00.309 --> 00:09:03.549 safety free or close to it. Basically. And it 00:09:03.549 --> 00:09:06.149 works using a few clever layers. They use these 00:09:06.149 --> 00:09:08.450 multi -max probes to pick out signals and huge 00:09:08.450 --> 00:09:11.149 prompts, up to a million tokens. And then something 00:09:11.149 --> 00:09:13.629 called rolling mean attention to filter out all 00:09:13.629 --> 00:09:15.690 the noise and spot, you know, the toxic thoughts. 00:09:16.110 --> 00:09:18.309 And what if a probe isn't sure about something? 00:09:18.769 --> 00:09:21.799 Then... And only then does it call in the supervisor. 00:09:22.080 --> 00:09:24.720 It's a cascade system. If the probe sees a red 00:09:24.720 --> 00:09:27.440 flag, it sends just that little snippet to a 00:09:27.440 --> 00:09:29.899 lightweight model like Gemini Flash to get a 00:09:29.899 --> 00:09:31.820 second opinion. So you cut your overall safety 00:09:31.820 --> 00:09:34.919 costs by 50x or more without really losing reliability. 00:09:35.629 --> 00:09:37.409 That's a brilliant piece of engineering. It's 00:09:37.409 --> 00:09:39.409 moving safety from an external policing layer 00:09:39.409 --> 00:09:42.029 to an internal awareness layer. We're essentially 00:09:42.029 --> 00:09:44.409 installing a conscience into the AI that costs 00:09:44.409 --> 00:09:47.269 almost nothing to run. A lightweight moral compass 00:09:47.269 --> 00:09:49.710 embedded right there in the matrix math. I want 00:09:49.710 --> 00:09:52.250 to take just a brief pause here. We've covered 00:09:52.250 --> 00:09:54.789 city -sized computers, the return of knitting, 00:09:54.990 --> 00:09:57.129 the struggle of formatting a simple document, 00:09:57.309 --> 00:10:00.289 and now the internal conscience of the machine. 00:10:00.690 --> 00:10:06.309 We'll be right back. And we are back. We've definitely 00:10:06.309 --> 00:10:08.710 covered a lot of ground today. Let's try to synthesize 00:10:08.710 --> 00:10:10.210 this for a moment, connect some of these dots. 00:10:10.509 --> 00:10:13.009 Okay, the big picture. We've got this massive 00:10:13.009 --> 00:10:15.230 convergence happening. On one side, you have 00:10:15.230 --> 00:10:18.250 a company like XAI. They are building physical 00:10:18.250 --> 00:10:20.470 infrastructure at a scale that challenges the 00:10:20.470 --> 00:10:24.909 power grids of major American cities. 1 .5 gigawatts. 00:10:25.979 --> 00:10:29.679 $18 billion. Pure brute force. And then on the 00:10:29.679 --> 00:10:31.299 complete other end of the spectrum, you have 00:10:31.299 --> 00:10:33.360 DeepMind. Right. They're looking at the microscopic 00:10:33.360 --> 00:10:35.799 level. They're finding these tiny efficiencies, 00:10:36.080 --> 00:10:39.220 inserting tiny, cheap probes into the math to 00:10:39.220 --> 00:10:41.320 make these things safe without bankrupting the 00:10:41.320 --> 00:10:42.899 companies that run them. So you have this macro 00:10:42.899 --> 00:10:44.940 expansion and this micro optimization happening 00:10:44.940 --> 00:10:48.019 at the exact same time. Exactly. And who is sandwiched 00:10:48.019 --> 00:10:50.679 in the middle of all this silicon and code? Us. 00:10:51.289 --> 00:10:54.149 The humans. Yeah, the humans who are generating 00:10:54.149 --> 00:10:57.549 billions in revenue for open AI, but also the 00:10:57.549 --> 00:10:59.929 humans who are feeling totally overwhelmed by 00:10:59.929 --> 00:11:03.149 it all. The humans buying all that yarn. It's 00:11:03.149 --> 00:11:05.289 a real tug of war. We're building these, you 00:11:05.289 --> 00:11:08.110 know, gods in the cloud. But down here on Earth, 00:11:08.210 --> 00:11:10.870 we just want a nice sweater and a document that 00:11:10.870 --> 00:11:13.370 formats correctly. And honestly, the fact that 00:11:13.370 --> 00:11:15.169 we can't get the document to format correctly 00:11:15.169 --> 00:11:17.330 is probably why we're all knitting the sweater. 00:11:17.490 --> 00:11:20.529 It's a need to feel competence. To feel like 00:11:20.529 --> 00:11:22.970 we can actually finish a task. I think that's 00:11:22.970 --> 00:11:25.470 going to be the theme in 2026. Competence in 00:11:25.470 --> 00:11:28.350 an age of automation. Before we wrap up, for 00:11:28.350 --> 00:11:30.570 anyone listening who is still stuck in that formatting 00:11:30.570 --> 00:11:33.509 loop. Yes, please do yourself a favor. Look up 00:11:33.509 --> 00:11:35.370 the formatting guides for these models. Learn 00:11:35.370 --> 00:11:37.590 the force formatting prompts. I'm telling you, 00:11:37.629 --> 00:11:40.389 it's a small skill, but it will change your daily 00:11:40.389 --> 00:11:42.850 quality of life. And as you head out into your 00:11:42.850 --> 00:11:44.809 week, I want to leave you with one final thought 00:11:44.809 --> 00:11:48.559 on that. analog backlash. Is this surge in yarn 00:11:48.559 --> 00:11:52.220 sales just a temporary fad, a little blip? Or 00:11:52.220 --> 00:11:54.740 are we seeing the beginning of a permanent bifurcation, 00:11:55.059 --> 00:11:57.600 a real split between a digital elite who live 00:11:57.600 --> 00:12:00.460 entirely in the stream and a kind of analog resistance 00:12:00.460 --> 00:12:03.019 that chooses to disconnect? That is something 00:12:03.019 --> 00:12:05.460 to think about. Are we heading for a world where 00:12:05.460 --> 00:12:08.700 true luxury just means being offline? I think 00:12:08.700 --> 00:12:10.840 we might be. Thanks for diving in with us. See 00:12:10.840 --> 00:12:11.320 you in the deep end.