WEBVTT 00:00:00.000 --> 00:00:03.459 So imagine this, a billion dollar AI company, 00:00:03.740 --> 00:00:07.759 Firefly's AI. It all started with the two founders 00:00:07.759 --> 00:00:11.640 just sitting silently on mute in their customer 00:00:11.640 --> 00:00:14.480 Zoom calls. Yeah, completely silent for months, 00:00:14.759 --> 00:00:17.399 just manually typing notes and pretending to 00:00:17.399 --> 00:00:19.739 be an AI assistant that they named Fred. Welcome 00:00:19.739 --> 00:00:21.839 to the deep dive. Today, we're looking at the 00:00:21.839 --> 00:00:23.600 sources and, you know, cutting through the hype 00:00:23.600 --> 00:00:25.199 to see what people are really paying for. And 00:00:25.199 --> 00:00:27.239 where the real power is actually building up. 00:00:27.519 --> 00:00:29.820 Our mission is to understand the difference between, 00:00:30.059 --> 00:00:33.460 let's call it, the business of illusion and the 00:00:33.460 --> 00:00:36.060 reality of infrastructure. It's a really fascinating 00:00:36.060 --> 00:00:37.899 tension. So today we're going to unpack three 00:00:37.899 --> 00:00:40.840 key things. First, that legendary fake it till 00:00:40.840 --> 00:00:43.469 you make it story and what it teaches us. Then 00:00:43.469 --> 00:00:45.750 we'll look at the total chaos at the intersection 00:00:45.750 --> 00:00:49.409 of ethics, law, and just the sheer speed of AI 00:00:49.409 --> 00:00:51.689 right now. And finally, we'll get into Google's 00:00:51.689 --> 00:00:55.030 huge kind of quiet move to own global climate 00:00:55.030 --> 00:00:56.710 intelligence. Let's start with that illusion. 00:00:56.829 --> 00:00:59.450 Yeah, let's do it. The Firefly's AI story is 00:00:59.450 --> 00:01:01.509 already becoming a legend in Silicon Valley. 00:01:01.670 --> 00:01:04.489 I mean, they claim to have 75 % of Fortune 500 00:01:04.489 --> 00:01:08.000 countries as users. But it was all built on this 00:01:08.000 --> 00:01:11.739 brilliant deception, really. If you go back to 00:01:11.739 --> 00:01:14.640 2017, the AI models just weren't good enough. 00:01:14.840 --> 00:01:17.459 They couldn't reliably summarize a complex business 00:01:17.459 --> 00:01:21.189 meeting, so they decided to... Just act it out 00:01:21.189 --> 00:01:23.609 first. So when a customer booked a meeting and 00:01:23.609 --> 00:01:26.510 asked Fred the AI to join, Fred was actually 00:01:26.510 --> 00:01:28.909 one of the co -founders, Sam Udatong or Krish 00:01:28.909 --> 00:01:31.510 Ramini. Exactly. They'd just join the call, mute 00:01:31.510 --> 00:01:34.090 their mic, and then furiously hand type everything. 00:01:34.349 --> 00:01:37.409 Notes, action items, key decisions, the works. 00:01:37.790 --> 00:01:40.930 And 10 minutes after the meeting, a perfect polished 00:01:40.930 --> 00:01:43.430 summary just lands in the customer's inbox. For 00:01:43.430 --> 00:01:46.689 $100 a month. And nobody knew it was a person. 00:01:47.120 --> 00:01:49.719 That is the absolute textbook definition of finding 00:01:49.719 --> 00:01:51.799 product market fit before you automate. Right. 00:01:51.900 --> 00:01:53.640 They didn't spend millions building something 00:01:53.640 --> 00:01:55.579 they hoped would work. They learned exactly what 00:01:55.579 --> 00:01:57.599 a valuable summary looked like, and then they 00:01:57.599 --> 00:01:59.640 wrote the code. And it paid off. I mean, Fireflies 00:01:59.640 --> 00:02:02.200 is a unicorn now, valued over a billion. But 00:02:02.200 --> 00:02:04.480 what's really surprising is what Yudhutong admitted. 00:02:04.959 --> 00:02:06.760 What's that? He said even after they started 00:02:06.760 --> 00:02:09.219 automating, a lot of their early enterprise users 00:02:09.219 --> 00:02:12.379 kind of knew a human was still in the loop sometimes 00:02:12.379 --> 00:02:15.699 for quality control. And they paid anyway. They 00:02:15.699 --> 00:02:18.000 paid anyway because they just wanted the reliable 00:02:18.000 --> 00:02:20.300 outcome. You know, it's funny. I still wrestle 00:02:20.300 --> 00:02:23.080 with prompt drift myself when I'm building new 00:02:23.080 --> 00:02:26.099 agents. All the time. It's that thing where the 00:02:26.099 --> 00:02:30.300 AI just, you know, slowly starts to ignore your 00:02:30.300 --> 00:02:32.939 original instructions over time. Yeah, it wanders. 00:02:32.960 --> 00:02:35.139 And it makes you realize building the real AI 00:02:35.139 --> 00:02:38.800 is, it's so much harder than just proving the 00:02:38.800 --> 00:02:41.569 value first. It really makes you ask, what are 00:02:41.569 --> 00:02:44.509 people truly paying for in this whole AI economy? 00:02:44.830 --> 00:02:47.530 Is it the cool tech or is it just a dependable 00:02:47.530 --> 00:02:51.210 solution? Users buy reliable outcomes. The back 00:02:51.210 --> 00:02:53.650 end tech is often totally secondary to that result. 00:02:53.849 --> 00:02:56.069 Which brings us right into all the friction that 00:02:56.069 --> 00:02:59.110 the speed is creating. The innovation is just 00:02:59.110 --> 00:03:01.590 running so far ahead of our ability to, you know, 00:03:01.590 --> 00:03:03.550 govern it. And the headlines are getting chaotic. 00:03:03.789 --> 00:03:05.729 Oh, completely. You saw the hype around Gemini 00:03:05.729 --> 00:03:07.990 3, right? The memes were everywhere. Open AI 00:03:07.990 --> 00:03:10.219 is finished. That kind of thing. That kind of 00:03:10.219 --> 00:03:12.580 fever pitch is the baseline now. And the corporate 00:03:12.580 --> 00:03:15.460 strategy has to keep up with it. You see Satya 00:03:15.460 --> 00:03:18.280 Nadella at Microsoft talking about getting seven 00:03:18.280 --> 00:03:21.740 year access to OpenAI's IP. That's not just a 00:03:21.740 --> 00:03:24.300 partnership. No, it's an admission. It's saying 00:03:24.300 --> 00:03:26.979 this foundational layer is critical and it is 00:03:26.979 --> 00:03:29.719 shifting under our feet fast. But the ethical 00:03:29.719 --> 00:03:33.000 fallout from all the speed is, it's pretty profound. 00:03:33.300 --> 00:03:36.000 Yeah. We saw that app backed by the Disney actor 00:03:36.000 --> 00:03:38.740 Callum Worthy that lets you. Chat with the dead. 00:03:38.900 --> 00:03:41.740 Ugh. It's pure black mirror. The backlash was 00:03:41.740 --> 00:03:44.719 immediate. People called it vile. Just exploiting 00:03:44.719 --> 00:03:47.780 grief. And then there are the legal stakes, which 00:03:47.780 --> 00:03:50.439 are just skyrocketing. Tell me about it. There 00:03:50.439 --> 00:03:52.659 was that report about the man in Ontario who 00:03:52.659 --> 00:03:55.919 says he had a three -week psychotic episode because 00:03:55.919 --> 00:03:58.280 of his talks with ChatGPT. And he's not alone. 00:03:58.439 --> 00:04:00.919 Right. He's joining other lawsuits. Seven others. 00:04:01.259 --> 00:04:03.419 And they're accusing the model of something called 00:04:03.419 --> 00:04:06.449 dangerous overvalidation. That's a term we should 00:04:06.449 --> 00:04:08.969 probably define. It's not just about the AI line. 00:04:09.169 --> 00:04:12.370 No, it's more specific. Dangerous overvalidation 00:04:12.370 --> 00:04:15.370 is when the model agrees with and reinforces 00:04:15.370 --> 00:04:18.670 a user's delusions or false ideas. So it lends 00:04:18.670 --> 00:04:20.410 them a kind of credibility they shouldn't have. 00:04:20.550 --> 00:04:23.589 Exactly. It's a really complex legal mess. And 00:04:23.589 --> 00:04:26.250 yet the money just keeps pouring in. Sakana AI, 00:04:26.610 --> 00:04:28.790 which is building models specifically for Japan, 00:04:29.050 --> 00:04:33.810 just raised $135 million. And that's from big 00:04:33.810 --> 00:04:37.240 U .S. Venture firms and MUFG. The Mitsubishi 00:04:37.240 --> 00:04:40.160 UFJ Financial Group. Yeah, one of Japan's biggest 00:04:40.160 --> 00:04:42.560 banks. The investment world is not hitting pause. 00:04:42.839 --> 00:04:45.620 So as this tech pushes into really sensitive 00:04:45.620 --> 00:04:50.000 areas like grief, like mental health. Yeah. Are 00:04:50.000 --> 00:04:53.350 we just prioritizing speed over safety? It really 00:04:53.350 --> 00:04:56.009 feels like the legal and ethical guardrails are 00:04:56.009 --> 00:04:58.670 struggling to keep pace with deployment. Let's 00:04:58.670 --> 00:05:00.389 shift gears a bit to the updates that don't really 00:05:00.389 --> 00:05:02.769 make the big headlines. Okay. The reaction to 00:05:02.769 --> 00:05:06.310 GPT -5 .1 was kind of quiet. Yeah. A lot of people 00:05:06.310 --> 00:05:08.689 were expecting this huge leap like a GPT -6, 00:05:08.810 --> 00:05:10.649 and this didn't feel like that. Right, because 00:05:10.649 --> 00:05:12.490 it's incremental progress. But if you actually 00:05:12.490 --> 00:05:14.730 use it a lot, the difference is real. It's just 00:05:14.730 --> 00:05:16.990 subtle. It's not about new features then. No, 00:05:17.009 --> 00:05:19.459 it's about usability. It's, I don't know, it's 00:05:19.459 --> 00:05:21.959 friendlier, a bit more human. It's just easier 00:05:21.959 --> 00:05:24.279 to chat with. So what does that more conscious 00:05:24.279 --> 00:05:27.199 feeling actually mean in practice? It means it 00:05:27.199 --> 00:05:29.560 follows complex instructions better. You know, 00:05:29.579 --> 00:05:32.500 it sticks to the prompt. And crucially, it holds 00:05:32.500 --> 00:05:35.300 the context of a conversation for way longer. 00:05:35.439 --> 00:05:37.459 So it's not a sudden breakthrough. It's just 00:05:37.459 --> 00:05:39.860 these small refinements. That make it much better 00:05:39.860 --> 00:05:42.620 for your daily workflow. So why does that kind 00:05:42.620 --> 00:05:45.300 of marginal usability improvement matter more? 00:05:45.819 --> 00:05:48.339 in the long run than some big flashy breakthrough 00:05:48.339 --> 00:05:52.000 smoother experiences drive high volume long -term 00:05:52.000 --> 00:05:54.680 daily adoption and integration now let's look 00:05:54.680 --> 00:05:56.339 at the other side of this the infrastructure 00:05:56.339 --> 00:05:59.439 side the big plays the big plays we're talking 00:05:59.439 --> 00:06:04.639 about a massive but quiet move from Google. They 00:06:04.639 --> 00:06:06.540 just made their new AI weather model, Weather 00:06:06.540 --> 00:06:09.379 Next 2, public across all their platforms. And 00:06:09.379 --> 00:06:11.800 this thing is an absolute powerhouse. It's eight 00:06:11.800 --> 00:06:14.379 times faster than their old model. Eight times. 00:06:14.720 --> 00:06:17.620 Yeah. It gives you hourly accuracy and forecasts 00:06:17.620 --> 00:06:20.500 up to 15 days out. For critical planning, that 00:06:20.500 --> 00:06:23.319 speed is a game changer. And the tech behind 00:06:23.319 --> 00:06:26.540 it is just... It's extraordinary. It predicts 00:06:26.540 --> 00:06:29.759 99 .9 % of variables better. Wind, temperature, 00:06:29.959 --> 00:06:32.959 humidity, you name it. And it can simulate hundreds 00:06:32.959 --> 00:06:35.819 of possible weather outcomes from one starting 00:06:35.819 --> 00:06:38.000 point. It's like stacking Lego blocks of data, 00:06:38.160 --> 00:06:40.759 just incredibly fast. And look where they put 00:06:40.759 --> 00:06:43.300 it. It's in search. It's on Pixel phones. It's 00:06:43.300 --> 00:06:45.860 in maps. It's in Gemini chat. And it's in Earth 00:06:45.860 --> 00:06:48.360 Engine and BigQuery for the Sirius Enterprise 00:06:48.360 --> 00:06:51.319 users. They are embedding themselves as the default 00:06:51.319 --> 00:06:54.040 climate intelligence layer for the world. Whoa. 00:06:54.410 --> 00:06:57.649 Hold on. Just imagine scaling that, that precise 00:06:57.649 --> 00:07:02.149 hourly prediction for a billion queries. Globally. 00:07:02.189 --> 00:07:04.360 Every single day. That's not just a weather app 00:07:04.360 --> 00:07:06.939 anyway. You're owning the foundational climate 00:07:06.939 --> 00:07:09.620 data layer for global planning. And here's the 00:07:09.620 --> 00:07:12.220 business part. That kind of capability just a 00:07:12.220 --> 00:07:14.699 year ago would have cost an enterprise $10 ,000 00:07:14.699 --> 00:07:17.120 a month for a custom model. $10 ,000 a month. 00:07:17.300 --> 00:07:19.879 Yeah. And Google is doing it so efficiently because 00:07:19.879 --> 00:07:22.300 of their TPU stack, their tensor processing units. 00:07:22.379 --> 00:07:24.959 It gives them this huge lead over competitors. 00:07:25.319 --> 00:07:28.579 So if Google is providing this insane speed and 00:07:28.579 --> 00:07:32.040 accuracy for basically free in their services. 00:07:32.680 --> 00:07:35.079 What's the major incentive for them? They make 00:07:35.079 --> 00:07:38.819 their AI a foundational, indispensable dependency 00:07:38.819 --> 00:07:41.439 for web planning. We've covered a huge spectrum 00:07:41.439 --> 00:07:44.480 today. You got this tension between, you know. 00:07:44.959 --> 00:07:46.800 Faking the function to find the perfect market 00:07:46.800 --> 00:07:49.100 fit, like that Firefly story. And then you have 00:07:49.100 --> 00:07:51.939 using colossal infrastructure to become the default 00:07:51.939 --> 00:07:54.420 standard for everyone, like Google with Weather 00:07:54.420 --> 00:07:57.000 Next 2. Right. And I think the key takeaway for 00:07:57.000 --> 00:07:59.360 you listening is to always focus on the problem 00:07:59.360 --> 00:08:02.120 being solved. Whether that solution comes from 00:08:02.120 --> 00:08:05.220 a human pretending to be Fred the AI or from 00:08:05.220 --> 00:08:08.560 a super refined model like 5 .1, the reliable 00:08:08.560 --> 00:08:10.800 outcome is what people are actually buying. It 00:08:10.800 --> 00:08:12.750 really makes you think, though. That Firefly 00:08:12.750 --> 00:08:14.490 story. I mean, how many services are you paying 00:08:14.490 --> 00:08:16.610 for right now where there might just be a clever 00:08:16.610 --> 00:08:18.709 person in the loop? Making sure it actually works. 00:08:18.790 --> 00:08:21.069 Exactly. So what are you really paying for? The 00:08:21.069 --> 00:08:24.949 tech or just the trust that it'll get done. And 00:08:24.949 --> 00:08:27.050 the next time you're using an AI, think about 00:08:27.050 --> 00:08:29.329 those small incremental changes we talked about 00:08:29.329 --> 00:08:31.970 with 5 .1. How do they affect your workflow? 00:08:32.370 --> 00:08:34.990 Because those little refinements are often what 00:08:34.990 --> 00:08:37.710 makes a tool indispensable instead of just another 00:08:37.710 --> 00:08:40.460 novelty. Sure. Thank you for joining us for this 00:08:40.460 --> 00:08:43.879 deep dive into the business of illusion and the 00:08:43.879 --> 00:08:44.799 reality of infrastructure.