WEBVTT 00:00:00.000 --> 00:00:04.059 There's this UK analyst claiming the AI financial 00:00:04.059 --> 00:00:07.540 boom right now. It's 17 times worse than the 00:00:07.540 --> 00:00:11.089 dot -com bubble. 17 times that number. Well, 00:00:11.189 --> 00:00:13.150 it really makes you stop and think, doesn't it? 00:00:13.390 --> 00:00:15.570 Deserves a proper look. It really does. Yeah. 00:00:15.630 --> 00:00:17.269 And thanks for sharing those sources with us. 00:00:17.329 --> 00:00:19.670 It's clear we're dealing with this massive contradiction, 00:00:19.949 --> 00:00:22.829 right? You've got these sky high AI valuations 00:00:22.829 --> 00:00:26.530 and then you look for the actual profitable uses 00:00:26.530 --> 00:00:28.530 right now. And it's well, it's a different picture. 00:00:28.710 --> 00:00:30.710 Exactly. So today, that's what we're going to 00:00:30.710 --> 00:00:32.909 do. We'll unpack that bubble claim, see where 00:00:32.909 --> 00:00:35.289 AI is actually delivering ROI and also touch 00:00:35.289 --> 00:00:38.049 on this really intense work culture popping up 00:00:38.049 --> 00:00:39.969 again in Silicon Valley to try and meet. these 00:00:39.969 --> 00:00:42.590 huge expectations. OK, let's dig into that big 00:00:42.590 --> 00:00:45.609 number first. The 17 times bubble. This analyst, 00:00:45.770 --> 00:00:48.350 Julian Guerin, he's not mincing words. 17 times 00:00:48.350 --> 00:00:51.890 the dotcom era. Four times the 2008 crash. Beep. 00:00:51.890 --> 00:00:55.000 That's a heavy statement. Mm hmm. And. It's crucial 00:00:55.000 --> 00:00:57.140 to get why he's saying that the dotcom thing 00:00:57.140 --> 00:01:00.020 that was built on maybe shaky revenue projections. 00:01:00.219 --> 00:01:04.060 Right. But this A .I. boom, he argues it's built 00:01:04.060 --> 00:01:06.879 more on just sheer market cap momentum. It's 00:01:06.879 --> 00:01:10.480 exploded way, way past any kind of near term 00:01:10.480 --> 00:01:13.739 profit reality. Yeah. You've got what, 10. major 00:01:13.739 --> 00:01:15.799 AI startups, together they've added something 00:01:15.799 --> 00:01:18.560 like a trillion dollars in market value. A trillion. 00:01:18.819 --> 00:01:21.920 But you look closer and almost all of them are 00:01:21.920 --> 00:01:24.599 still, you know, deeply unprofitable. It's this 00:01:24.599 --> 00:01:26.920 profitability paradox. It really is. So you got 00:01:26.920 --> 00:01:29.879 to ask, who is making money here? And right now 00:01:29.879 --> 00:01:32.650 the answer seems pretty clear. It's mostly NVIDIA 00:01:32.650 --> 00:01:34.989 selling the hardware, the shovels for the gold 00:01:34.989 --> 00:01:36.829 rush. Basically, they're getting consistent returns. 00:01:37.109 --> 00:01:39.150 Pretty much everyone else developing the core 00:01:39.150 --> 00:01:41.530 AI models. They're basically just, well, bleeding 00:01:41.530 --> 00:01:43.950 cash, funding operations based on what they hope 00:01:43.950 --> 00:01:45.689 will happen down the line. And all that cash 00:01:45.689 --> 00:01:47.450 burn. Yeah. It's making the traditional money 00:01:47.450 --> 00:01:50.090 people, the VCs, a bit nervous. You're seeing 00:01:50.090 --> 00:01:52.090 them pull back now. They look at the valuations 00:01:52.090 --> 00:01:54.269 and just think they're, well, kind of absurd. 00:01:54.430 --> 00:01:56.530 Right. So the whole ecosystem is increasingly 00:01:56.530 --> 00:02:00.200 propped up by just a few huge players. The mega 00:02:00.200 --> 00:02:02.480 backers, as Garen calls them. Yeah, not your 00:02:02.480 --> 00:02:05.140 usual VCs. We're talking SoftBank, sovereign 00:02:05.140 --> 00:02:08.080 wealth funds. Even NVIDIA itself is investing 00:02:08.080 --> 00:02:10.219 heavily back into the ecosystem. They're the 00:02:10.219 --> 00:02:13.080 capital life support right now. And the bet seems 00:02:13.080 --> 00:02:15.840 to be if they just keep pouring money in, eventually 00:02:15.840 --> 00:02:17.699 one of these companies will crack artificial 00:02:17.699 --> 00:02:21.460 general intelligence or AGI. So Garen lays out 00:02:21.460 --> 00:02:24.280 two possible futures, basically. Option one. 00:02:24.419 --> 00:02:27.020 Yeah. The money keeps flowing, but it never really 00:02:27.020 --> 00:02:29.199 delivers that massive value everyone expects. 00:02:29.719 --> 00:02:33.060 It's just this capital treadmill or option two. 00:02:33.539 --> 00:02:36.960 Someone actually does achieve true AGI, a genuine 00:02:36.960 --> 00:02:38.900 breakthrough. And he's betting on option one, 00:02:38.939 --> 00:02:40.819 isn't he? He seems to think the market hype has 00:02:40.819 --> 00:02:42.740 just gotten way out ahead of the actual tech 00:02:42.740 --> 00:02:44.800 capabilities right now. I'm going to say just 00:02:44.800 --> 00:02:46.500 trying to understand the tech itself can be tough. 00:02:46.599 --> 00:02:48.719 Honestly, sometimes I wrestle with prompt drift 00:02:48.719 --> 00:02:51.060 myself, you know, getting consistent results. 00:02:51.360 --> 00:02:53.219 So when you talk about these billion dollar market 00:02:53.219 --> 00:02:55.219 shifts. Yeah. Yeah. It's a lot to wrap your head 00:02:55.219 --> 00:02:57.520 around. So given all that cash burn right across 00:02:57.520 --> 00:03:00.509 the board. What's the single biggest factor keeping 00:03:00.509 --> 00:03:03.990 these valuations so incredibly high? It really 00:03:03.990 --> 00:03:06.550 seems to boil down to those mega backers. Them 00:03:06.550 --> 00:03:11.310 plus this persistent belief, this hope that true 00:03:11.310 --> 00:03:15.569 AGI is just around the corner. OK, let's shift 00:03:15.569 --> 00:03:17.669 gears a bit away from the financials and towards 00:03:17.669 --> 00:03:19.650 some more practical stuff. If you're trying to 00:03:19.650 --> 00:03:22.129 get a better handle on the tech itself, Stanford 00:03:22.129 --> 00:03:24.349 actually has this really good four part lecture 00:03:24.349 --> 00:03:26.069 series out. It's about five and a half hours 00:03:26.069 --> 00:03:28.169 total. Gives you a pretty solid kind of medium 00:03:28.169 --> 00:03:30.789 depth understanding of LLMs, you know, large 00:03:30.789 --> 00:03:32.650 language models, the AIs that generate human 00:03:32.650 --> 00:03:34.569 like text. And we are seeing some pretty cool 00:03:34.569 --> 00:03:36.710 jumps in what these things can do. There's a 00:03:36.710 --> 00:03:39.310 new tool, DeepSeq OCR, for reading documents. 00:03:39.569 --> 00:03:41.490 Apparently it's really impressing people. It 00:03:41.490 --> 00:03:43.650 managed to accurately read some properly messy 00:03:43.650 --> 00:03:47.349 handwritten. written letters, which is not easy. 00:03:47.530 --> 00:03:50.150 Yeah, that's huge. Imagine for archives or just 00:03:50.150 --> 00:03:53.569 extracting data. But there's always this ethical 00:03:53.569 --> 00:03:55.449 tension simmering underneath, isn't there? Like 00:03:55.449 --> 00:03:58.449 Microsoft's AI chief coming out and publicly 00:03:58.449 --> 00:04:01.449 criticizing ChatGPT's ability to generate erotica. 00:04:01.569 --> 00:04:05.210 That feels kind of awkward given Microsoft put, 00:04:05.310 --> 00:04:09.069 what, $13 billion into open AI? It does. And 00:04:09.069 --> 00:04:10.610 you also mentioned the Australian government. 00:04:11.050 --> 00:04:14.389 suing Microsoft over confusion about AI pricing 00:04:14.389 --> 00:04:16.670 changes. It just shows how regulators are scrambling 00:04:16.670 --> 00:04:18.610 to keep up, trying to figure out how to handle 00:04:18.610 --> 00:04:23.189 the ambiguity of these new AI systems. And here's 00:04:23.189 --> 00:04:26.149 something that feels genuinely concerning in 00:04:26.149 --> 00:04:28.350 the sources you shared. This trend of extreme 00:04:28.350 --> 00:04:30.430 overwork making a comeback in Silicon Valley. 00:04:30.839 --> 00:04:33.180 AI startups are apparently bringing back China's 00:04:33.180 --> 00:04:37.500 infamous 996 grind. That's 9 a .m. to 9 p .m. 00:04:37.500 --> 00:04:39.800 six days a week. Wow. Yeah, and there's even 00:04:39.800 --> 00:04:41.480 that dark joke going around among researchers 00:04:41.480 --> 00:04:44.500 about the 002 schedule, zero sleep, zero days 00:04:44.500 --> 00:04:47.199 off, just two hours maybe. The pressure to deliver, 00:04:47.360 --> 00:04:50.079 to hit those short -term milestones, maybe justify 00:04:50.079 --> 00:04:52.300 those crazy valuations we talked about, it's 00:04:52.300 --> 00:04:54.699 pushing people really, really hard. It seems 00:04:54.699 --> 00:04:56.980 like a dangerous game. That kind of pressure 00:04:56.980 --> 00:04:58.839 cooker environment might get quick results, but... 00:04:59.420 --> 00:05:02.120 Burning out your best people. Yeah, and it introduces 00:05:02.120 --> 00:05:04.439 a huge risk of mistakes, right? Errors creeping 00:05:04.439 --> 00:05:06.620 into the very foundations of the models they're 00:05:06.620 --> 00:05:09.420 building. Not good long term. Meanwhile, the 00:05:09.420 --> 00:05:11.139 big players are still pushing into new areas. 00:05:11.680 --> 00:05:14.439 OpenAI has apparently got a secret project, building 00:05:14.439 --> 00:05:17.680 a generative music tool. Ah, okay. Like SunoAI 00:05:17.680 --> 00:05:21.199 and those existing tools. Turning text prompts 00:05:21.199 --> 00:05:23.980 or even audio snippets into full songs. Always 00:05:23.980 --> 00:05:26.459 looking for that next big creative market to 00:05:26.459 --> 00:05:30.290 disrupt. So if this 996 work culture, this intense 00:05:30.290 --> 00:05:33.490 grind, is becoming widespread again, how does 00:05:33.490 --> 00:05:35.709 that really impact AI innovation in the long 00:05:35.709 --> 00:05:37.689 run? Well, you might get some rapid gains, sure, 00:05:37.790 --> 00:05:40.589 short bursts of progress. But the risk of burnout, 00:05:40.790 --> 00:05:43.370 costly errors, and ultimately maybe less thoughtful, 00:05:43.509 --> 00:05:47.529 lower quality innovation seems pretty high. Okay, 00:05:47.569 --> 00:05:49.990 let's focus now on some specific tools and maybe 00:05:49.990 --> 00:05:51.910 more importantly, some really impactful applications. 00:05:52.329 --> 00:05:55.269 We saw Google talking about vibe coding. in their 00:05:55.269 --> 00:05:57.670 AI studio. Sounds like they're trying to streamline 00:05:57.670 --> 00:05:59.470 how engineers work, make development faster, 00:05:59.569 --> 00:06:01.970 more intuitive. But the applications that really 00:06:01.970 --> 00:06:04.189 jumped out, I think, were in biodefense and health. 00:06:04.389 --> 00:06:07.230 There's this company, Valfos, backed by OpenAI 00:06:07.230 --> 00:06:10.029 folks. They launched with $30 million specifically 00:06:10.029 --> 00:06:13.509 to build AI biodefense systems. Yes, systems 00:06:13.509 --> 00:06:15.509 designed to detect and figure out how to counter 00:06:15.509 --> 00:06:19.889 new pathogens. In just hours. Yeah. Hours. I 00:06:19.889 --> 00:06:21.850 mean, think about that. That's a massive leap 00:06:21.850 --> 00:06:25.370 in how quickly we could respond to, say, a new 00:06:25.370 --> 00:06:27.509 pandemic threat. It really is. And then there 00:06:27.509 --> 00:06:30.329 was that other study from Utah showing AI could 00:06:30.329 --> 00:06:32.930 spot parasites in human stool samples really 00:06:32.930 --> 00:06:37.699 quickly and accurately. Pause. Whoa. I mean, 00:06:37.699 --> 00:06:39.939 just imagine scaling that kind of detection speed, 00:06:40.139 --> 00:06:42.300 applying it globally to public health surveillance. 00:06:42.639 --> 00:06:45.000 That's tangible value. That justifies serious 00:06:45.000 --> 00:06:47.579 investment, you know? Absolutely. And the support 00:06:47.579 --> 00:06:49.319 system for all this is growing, too. Amazon, 00:06:50.000 --> 00:06:52.819 OpenAI, NVIDIA, they're teaming up to turn Cal 00:06:52.819 --> 00:06:55.279 State into a big AI training center, recognizing 00:06:55.279 --> 00:06:57.160 we need more people who actually know how to 00:06:57.160 --> 00:06:58.980 build and use this stuff. There's even a paper 00:06:58.980 --> 00:07:01.620 mapping out like 90 different AI coding tools 00:07:01.620 --> 00:07:03.779 available right now. So thinking about all those 00:07:03.779 --> 00:07:06.509 tools and applications. Beyond the big names, 00:07:06.649 --> 00:07:09.350 what was the most surprising practical use case 00:07:09.350 --> 00:07:11.930 we came across this time? For me, it was that 00:07:11.930 --> 00:07:15.410 health stuff. The AI detecting pathogens and 00:07:15.410 --> 00:07:18.410 parasites so quickly and accurately that speed 00:07:18.410 --> 00:07:20.829 just feels like a potential game changer for 00:07:20.829 --> 00:07:23.870 public health. Mineral sponsor read, insert sponsor 00:07:23.870 --> 00:07:25.689 read here. All right. So we've talked about the 00:07:25.689 --> 00:07:28.689 bubble fears, the potential, the pressures. Let's 00:07:28.689 --> 00:07:30.850 look at some hard data now, specifically from 00:07:30.850 --> 00:07:33.430 industries already using this tech. We're pulling 00:07:33.430 --> 00:07:36.329 this from the Artificial Analysis 2025 State 00:07:36.329 --> 00:07:39.149 of Generative Media report, based on input from 00:07:39.149 --> 00:07:41.430 about 300 developers and creators using these 00:07:41.430 --> 00:07:44.170 tools right now. The data really highlights Google's 00:07:44.170 --> 00:07:46.790 position, especially in creative AI, for making 00:07:46.790 --> 00:07:50.089 images. Google Gemini is number one, 74 % usage 00:07:50.089 --> 00:07:53.149 among those surveyed. OpenAI's model, GPT image, 00:07:53.370 --> 00:07:57.129 is next at 64%. So, strong competition there. 00:07:57.389 --> 00:07:59.610 Yeah, and that lead carries over into video generation 00:07:59.610 --> 00:08:03.089 too. Google VO. Leading the pack with 69 % usage 00:08:03.089 --> 00:08:05.069 in this group, competition's clearly driving 00:08:05.069 --> 00:08:07.589 things forward. And what's driving the choice 00:08:07.589 --> 00:08:09.949 for these creators? What makes them pick one 00:08:09.949 --> 00:08:12.449 model over another? Overwhelmingly, it's quality. 00:08:12.649 --> 00:08:15.269 That's the number one factor, cited by like 76 00:08:15.269 --> 00:08:18.209 % to 82%, depending on the task. Makes sense. 00:08:18.410 --> 00:08:21.569 Cost comes second, yeah. But the message is clear. 00:08:22.439 --> 00:08:24.959 If the output isn't good enough, isn't high quality, 00:08:25.180 --> 00:08:27.680 then people just aren't going to use it regardless 00:08:27.680 --> 00:08:30.120 of price. OK, so let's connect this back to that 00:08:30.120 --> 00:08:32.600 bubble debate we started with. This feels really 00:08:32.600 --> 00:08:35.659 important. Sixty five percent, almost two thirds 00:08:35.659 --> 00:08:38.519 of the organizations surveyed said they're already 00:08:38.519 --> 00:08:40.779 seeing a return on their investment in generative 00:08:40.779 --> 00:08:43.740 AI or they expect to see it within the next year. 00:08:44.120 --> 00:08:47.039 Right now. OK, that's likely self -reported data. 00:08:47.120 --> 00:08:49.279 You got to take it with maybe a small grain of 00:08:49.279 --> 00:08:52.610 salt. But still. 65%. That's pretty strong validation, 00:08:52.809 --> 00:08:55.190 isn't it? It suggests that investment in these 00:08:55.190 --> 00:08:58.330 specific high quality focused applications like 00:08:58.330 --> 00:09:01.169 generative media is paying off for many right 00:09:01.169 --> 00:09:03.210 now. Yeah, it kind of counterbalances the hype 00:09:03.210 --> 00:09:05.710 around the purely research focused arms that 00:09:05.710 --> 00:09:08.830 are still burning cash. There's real measurable 00:09:08.830 --> 00:09:12.379 value being created in certain sectors. If quality 00:09:12.379 --> 00:09:14.919 is the absolute top priority, does that automatically 00:09:14.919 --> 00:09:18.200 mean the big expensive proprietary models from, 00:09:18.279 --> 00:09:21.019 say, Google and OpenAI will just always dominate 00:09:21.019 --> 00:09:23.840 the open source alternatives? Well, quality is 00:09:23.840 --> 00:09:26.000 definitely driving adoption now. But the report 00:09:26.000 --> 00:09:28.220 suggests that the ability to fine tune models 00:09:28.220 --> 00:09:31.259 for specific needs and dealing with IP and licensing 00:09:31.259 --> 00:09:33.960 stuff, those are becoming really important secondary 00:09:33.960 --> 00:09:37.039 factors pretty quickly. So bringing it all together. 00:09:37.960 --> 00:09:39.960 What does this mean for you listening? We've 00:09:39.960 --> 00:09:41.620 really grappled with the central tension today, 00:09:41.679 --> 00:09:43.799 haven't we? On one hand, you've got this immense 00:09:43.799 --> 00:09:46.860 financial speculation that potential 17X bubble 00:09:46.860 --> 00:09:49.620 Garen warned about. Huge risk. But on the other 00:09:49.620 --> 00:09:52.399 hand, you have clear, demonstrable ROI already 00:09:52.399 --> 00:09:54.820 happening in fields like generative media and 00:09:54.820 --> 00:09:56.899 these incredible potential breakthroughs in areas 00:09:56.899 --> 00:09:59.240 like biodefense. Yeah, it's a paradox, like we 00:09:59.240 --> 00:10:02.220 said. Massive financial risk running alongside 00:10:02.220 --> 00:10:05.299 absolutely essential functional progress. The 00:10:05.299 --> 00:10:07.759 debate isn't really if AI is valuable anymore, 00:10:07.899 --> 00:10:09.940 is it? It seems more about where the sustainable 00:10:09.940 --> 00:10:13.340 value is actually landing. Is it with the companies 00:10:13.340 --> 00:10:15.600 selling the infrastructure like NVIDIA and Google 00:10:15.600 --> 00:10:17.899 through its cloud? Or is it eventually going 00:10:17.899 --> 00:10:19.840 to come from those research arms chasing AGI 00:10:19.840 --> 00:10:22.519 even though they're burning cash now? And maybe 00:10:22.519 --> 00:10:25.600 here's a final thought to chew on. If 55 % of 00:10:25.600 --> 00:10:27.600 organizations say they're already getting ROI 00:10:27.600 --> 00:10:31.610 from AI, What do the other 35 % need to figure 00:10:31.610 --> 00:10:34.110 out? What needs to change in their strategy or 00:10:34.110 --> 00:10:36.350 maybe the specific way they're applying AI so 00:10:36.350 --> 00:10:38.789 they don't get left behind? That's the challenge 00:10:38.789 --> 00:10:40.570 of applying this knowledge in real time, isn't 00:10:40.570 --> 00:10:42.669 it? Absolutely. Well, thanks again for sharing 00:10:42.669 --> 00:10:44.570 the sources and letting us dive deep into all 00:10:44.570 --> 00:10:46.269 this with you today. Definitely check out those 00:10:46.269 --> 00:10:48.009 Stanford lectures if you're interested and maybe 00:10:48.009 --> 00:10:50.490 take a look at that generative media report data. 00:10:50.669 --> 00:10:53.169 Lots to think about. Out to your own music.