WEBVTT 00:00:00.000 --> 00:00:01.940 Okay, so just imagine this for a second. You're 00:00:01.940 --> 00:00:06.019 running a pharma startup, right? And you're using 00:00:06.019 --> 00:00:08.699 one of these next -gen AI models, let's say, 00:00:08.699 --> 00:00:10.839 I don't know, something way past GPT -5. Right. 00:00:11.019 --> 00:00:14.199 And this thing, it's not just, you know, drafting 00:00:14.199 --> 00:00:16.399 your emails. It's running millions of simulations. 00:00:16.519 --> 00:00:19.460 It hypothesizes a new chemical compound, and 00:00:19.460 --> 00:00:22.280 it basically discovers a cure for a rare disease. 00:00:22.640 --> 00:00:26.199 Wow. You stand to make billions, but then... 00:00:26.800 --> 00:00:29.940 You get this tap on the shoulder. The AI company 00:00:29.940 --> 00:00:32.640 doesn't want your $20 a month subscription fee. 00:00:32.759 --> 00:00:35.320 They want a percentage. A cut of the cure. Cut 00:00:35.320 --> 00:00:38.140 of the cure. They want royalties. Because in 00:00:38.140 --> 00:00:40.579 their view. Without them, you're just a scientist 00:00:40.579 --> 00:00:42.920 with a hunch. And that's the paradigm shift we're 00:00:42.920 --> 00:00:44.880 really waking up to this week. It creates such 00:00:44.880 --> 00:00:47.100 a wild precedent, doesn't it? I mean, we are 00:00:47.100 --> 00:00:50.520 moving from AI as a tool, like a really, really 00:00:50.520 --> 00:00:53.679 smart calculator, to AI as a partner. A stakeholder. 00:00:53.960 --> 00:00:55.960 A stakeholder. And if your stakeholders are digital, 00:00:56.179 --> 00:00:58.880 the economics of everything are about to get 00:00:58.880 --> 00:01:01.899 very weird. Welcome to the Deep Dive. It is Sunday, 00:01:02.119 --> 00:01:05.900 January 25th, 2026. Today, we're unpacking a 00:01:05.900 --> 00:01:07.739 stack of sources that suggest the rules of the 00:01:07.739 --> 00:01:10.859 game are changing and fast. We're not just talking 00:01:10.859 --> 00:01:12.819 about better chatbots anymore. Not at all. We 00:01:12.819 --> 00:01:16.299 are looking at a fundamental shift in business 00:01:16.299 --> 00:01:19.760 models, a huge leap in how AI perceives physical 00:01:19.760 --> 00:01:23.359 reality, and honestly, a bit of a crisis when 00:01:23.359 --> 00:01:25.939 it comes to trusting what these things even tell 00:01:25.939 --> 00:01:27.879 us. We have a lot to get through. We're going 00:01:27.879 --> 00:01:30.340 to start with that open AI profit sharing news, 00:01:30.519 --> 00:01:32.280 which I think is probably the biggest economic 00:01:32.280 --> 00:01:34.780 story in tech so far this year. Yeah, for sure. 00:01:34.920 --> 00:01:36.640 Then we'll look at what people are calling the 00:01:36.640 --> 00:01:39.659 GPT -2 moment for world models, basically AI 00:01:39.659 --> 00:01:41.900 that can build physics aware realities from scratch. 00:01:42.159 --> 00:01:45.379 Which is just mind blowing tech. It is. But we 00:01:45.379 --> 00:01:47.959 also have to look at the messy side of it. We've 00:01:47.959 --> 00:01:50.739 got reports of Google's AI prioritizing YouTube 00:01:50.739 --> 00:01:54.019 over, say, the Mayo Clinic for medical advice. 00:01:54.200 --> 00:01:56.569 Yikes. And some internal drugs. between deep 00:01:56.569 --> 00:01:58.609 mind and open AI that's starting to spill out 00:01:58.609 --> 00:02:00.969 into public. And then finally, we'll ground all 00:02:00.969 --> 00:02:03.590 of it with some practical tools. Google have 00:02:03.590 --> 00:02:05.329 a new breakdown of prompt engineering that's 00:02:05.329 --> 00:02:07.689 actually useful, and Apple is, believe it or 00:02:07.689 --> 00:02:10.960 not, finally giving serious. So let's start there. 00:02:11.000 --> 00:02:13.280 Let's unpack this open AI news. We're seeing 00:02:13.280 --> 00:02:16.840 reports surface about a major strategy shift, 00:02:16.860 --> 00:02:19.120 specifically coming from some comments by Sarah 00:02:19.120 --> 00:02:22.560 Fryer, the COO. What exactly is being proposed 00:02:22.560 --> 00:02:25.139 here? So the gist is that the flat fee model, 00:02:25.280 --> 00:02:27.979 the whole pay 20 bucks a month and use the bot, 00:02:28.180 --> 00:02:31.960 is becoming kind of outdated for high level enterprise 00:02:31.960 --> 00:02:35.360 use. Open AI is exploring a licensing model. 00:02:36.019 --> 00:02:38.520 And the argument, their argument, is that modern 00:02:38.520 --> 00:02:39.939 models, especially the ones they have in the 00:02:39.939 --> 00:02:42.400 lab right now, they aren't just summarizing PDFs 00:02:42.400 --> 00:02:45.560 anymore. They're acting like an AI researcher 00:02:45.560 --> 00:02:47.840 that never sleeps. That phrase, that really stood 00:02:47.840 --> 00:02:49.460 out to me in the notes, an AI researcher that 00:02:49.460 --> 00:02:53.360 never sleeps. It implies a level of agency that 00:02:53.360 --> 00:02:55.979 we haven't really given to software before. Exactly. 00:02:56.120 --> 00:02:57.419 I mean, think about what these models are doing 00:02:57.419 --> 00:03:00.780 in 2026. They're running millions of hypothesis 00:03:00.780 --> 00:03:03.900 tests across decades of research papers. They're 00:03:03.900 --> 00:03:06.219 stitching together data from simulations, from 00:03:06.219 --> 00:03:08.919 real world experiments, and they're proposing 00:03:08.919 --> 00:03:11.080 experiments that actually work. We're seeing 00:03:11.080 --> 00:03:13.699 this in biotech, in energy, materials, science. 00:03:14.000 --> 00:03:16.240 So the logic is, if the AI is doing the real 00:03:16.240 --> 00:03:18.659 heavy lifting of the invention, the creator of 00:03:18.659 --> 00:03:21.400 that AI deserves a slice of the pie. Right. If 00:03:21.400 --> 00:03:23.639 you hire a world -class scientist and she invents 00:03:23.639 --> 00:03:25.840 a new type of battery, she usually gets a bonus. 00:03:26.379 --> 00:03:28.460 Maybe some stock options, right? Sure. OpenAI 00:03:28.460 --> 00:03:30.439 is basically saying, we provided the digital 00:03:30.439 --> 00:03:32.580 scientist. We want in on the downstream success. 00:03:33.159 --> 00:03:36.060 It's treating the model as a collaborator, not 00:03:36.060 --> 00:03:38.460 just a utility you rent. But it changes the definition 00:03:38.460 --> 00:03:40.759 of collaboration, doesn't it? I mean, usually 00:03:40.759 --> 00:03:43.300 collaboration implies two humans sharing risk, 00:03:43.400 --> 00:03:45.620 sharing effort. Here, one party is a software 00:03:45.620 --> 00:03:48.680 vendor. But I suppose if the software is providing 00:03:48.680 --> 00:03:52.000 genuine cognitive labor. Cognitive labor. That's 00:03:52.000 --> 00:03:54.699 the key word. If a model suggests a financial 00:03:54.699 --> 00:03:57.379 strategy that makes a hedge fund an extra $100 00:03:57.379 --> 00:04:01.080 million, OpenAI looks at that and says, that 00:04:01.080 --> 00:04:04.360 wasn't just a tool like Excel. That was insight. 00:04:04.780 --> 00:04:07.639 But the complexities here seem massive. Well, 00:04:07.659 --> 00:04:10.060 they are. Who owns the IP? How do you even track 00:04:10.060 --> 00:04:13.719 value? That's what I'm wondering. If I use, say, 00:04:13.780 --> 00:04:16.399 five different AI tools, one for coding, one 00:04:16.399 --> 00:04:18.860 for biology, one for data analysis, and I have 00:04:18.860 --> 00:04:21.529 a breakthrough, who gets the check? Do they all 00:04:21.529 --> 00:04:23.529 split it? It sounds like a legal nightmare. Oh, 00:04:23.550 --> 00:04:26.730 absolutely. It is incredibly messy. But the signal 00:04:26.730 --> 00:04:28.930 is what's important here, and it's crystal clear. 00:04:29.529 --> 00:04:32.670 OpenAI sees itself as a stakeholder in your breakthroughs. 00:04:33.029 --> 00:04:35.629 They are betting their models will be so integral 00:04:35.629 --> 00:04:38.350 to the next wave of scientific discovery, and 00:04:38.350 --> 00:04:39.889 they don't want to leave that value on the table 00:04:39.889 --> 00:04:42.350 for $20 a month. It's a bold move. It speaks 00:04:42.350 --> 00:04:44.930 to a level of confidence in their tech that's... 00:04:45.439 --> 00:04:47.560 Well, it's almost arrogant, but maybe it's justified 00:04:47.560 --> 00:04:49.439 if the results are there. It really is. It's 00:04:49.439 --> 00:04:51.139 them saying, we know we're going to help you 00:04:51.139 --> 00:04:53.519 win, so we want our cut. So here's the question 00:04:53.519 --> 00:04:57.720 then. Does taking a cut of the win incentivize 00:04:57.720 --> 00:05:00.100 OpenAI to build drastically better models for 00:05:00.100 --> 00:05:02.800 us? Or does it just create this massive barrier 00:05:02.800 --> 00:05:05.300 where companies stop using them to protect their 00:05:05.300 --> 00:05:08.360 own innovation? It aligns incentives, but might 00:05:08.360 --> 00:05:10.779 scare off companies protecting their IP. Okay, 00:05:10.819 --> 00:05:12.959 let's pivot from the business side to the tech 00:05:12.959 --> 00:05:15.389 itself. While the business guys are figuring 00:05:15.389 --> 00:05:17.730 out how to charge for this intelligence, the 00:05:17.730 --> 00:05:19.949 engineers are figuring out how to make that intelligence 00:05:19.949 --> 00:05:22.389 understand the physical world. Yes. We're hearing 00:05:22.389 --> 00:05:25.029 a lot about world models. This is the fun stuff. 00:05:25.089 --> 00:05:27.689 We're seeing what some experts are calling the 00:05:27.689 --> 00:05:30.810 GPT -2 moment for world models. Okay. So for 00:05:30.810 --> 00:05:32.829 listeners who might not remember 2019 all that 00:05:32.829 --> 00:05:35.670 clearly, just remind us why GPT -2 is the benchmark 00:05:35.670 --> 00:05:38.329 here. Right. So GPT -2 wasn't the model that 00:05:38.329 --> 00:05:40.389 changed the world that was really GPT -3 and 00:05:40.389 --> 00:05:43.420 4. But GPT -2 was the first time that text generation 00:05:43.420 --> 00:05:46.319 felt stable enough to build on. It wasn't perfect, 00:05:46.459 --> 00:05:49.220 but it worked. We've just hit that same milestone. 00:05:49.579 --> 00:05:52.839 But for generating video in 3D worlds? And we 00:05:52.839 --> 00:05:54.980 have two specific players mentioned here driving 00:05:54.980 --> 00:05:58.019 this, Odyssey and World Labs. Right. Let's start 00:05:58.019 --> 00:06:01.519 with Odyssey 2 Pro. This is a new API that generates 00:06:01.519 --> 00:06:05.019 interactive, physics -aware video. And I really 00:06:05.019 --> 00:06:07.579 want to stress physics -aware. Okay. You aren't 00:06:07.579 --> 00:06:09.600 just generating a video file that plays from 00:06:09.600 --> 00:06:12.079 start to finish. You're generating a simulation. 00:06:12.579 --> 00:06:15.600 So what happens if I type in, say, a laughing 00:06:15.600 --> 00:06:18.160 baby? Okay, so it generates the scene almost 00:06:18.160 --> 00:06:20.639 instantly. We're talking 720p at about 22 frames 00:06:20.639 --> 00:06:23.779 per second. But here's the kicker. The model 00:06:23.779 --> 00:06:27.139 is actively predicting the next frame like a 00:06:27.139 --> 00:06:29.060 physics engine. Wow. It understands gravity, 00:06:29.259 --> 00:06:32.089 light. how things move. And you can interact 00:06:32.089 --> 00:06:33.930 with it in real time. It's not a movie. It's 00:06:33.930 --> 00:06:36.970 a little sandbox. That distinction feels crucial. 00:06:37.110 --> 00:06:39.629 It's moving from just capturing reality to actively 00:06:39.629 --> 00:06:42.029 simulating it. And World Labs is doing something 00:06:42.029 --> 00:06:45.029 similar, but in 3D space. Exactly. World Labs 00:06:45.029 --> 00:06:47.350 just dropped their marble model. You can upload 00:06:47.350 --> 00:06:49.509 a single photo or even just describe a scene. 00:06:49.810 --> 00:06:51.970 And in about five minutes, it gives you a fully 00:06:51.970 --> 00:06:54.610 navigable 3D world. Five minutes? Five minutes. 00:06:54.629 --> 00:06:56.189 It figures out the layout, the lighting, the 00:06:56.189 --> 00:06:58.889 depth. It uses this tech called Gaussian Splats 00:06:58.889 --> 00:07:02.009 to... render it all gaussian splats it sounds 00:07:02.009 --> 00:07:04.209 like a messy painting technique but i know it's 00:07:04.209 --> 00:07:06.550 a big deal technically it's surprisingly elegant 00:07:06.550 --> 00:07:09.509 the simple version is it's a way of representing 00:07:09.509 --> 00:07:14.509 3d scenes as these sort of fuzzy blobs the splats 00:07:14.509 --> 00:07:17.370 that all blend together to look solid right and 00:07:17.370 --> 00:07:19.730 it's way faster to render than traditional polygons 00:07:19.730 --> 00:07:23.370 so you go from a flat 2d picture of a living 00:07:23.370 --> 00:07:26.180 room to literally Walking around inside that 00:07:26.180 --> 00:07:27.959 living room in five minutes. That just feels 00:07:27.959 --> 00:07:30.860 like a massive leap for industries like gaming 00:07:30.860 --> 00:07:33.160 or architecture. You don't need to code the wall. 00:07:33.259 --> 00:07:36.839 You just dream the wall. Whoa. And imagine scaling 00:07:36.839 --> 00:07:39.040 that. Imagine just saying, I want a city that 00:07:39.040 --> 00:07:41.639 looks like 1920s Paris, but with neon lights. 00:07:42.110 --> 00:07:44.509 and then having a walkable simulation in minutes. 00:07:44.769 --> 00:07:47.829 We're moving from generating a picture to stepping 00:07:47.829 --> 00:07:49.310 into the simulation. It's a holodeck. It's a 00:07:49.310 --> 00:07:50.930 holodeck just on your screen. It's incredible. 00:07:51.050 --> 00:07:53.649 But it also blurs the line between creation and 00:07:53.649 --> 00:07:56.129 hallucination. If we can generate physics -aware 00:07:56.129 --> 00:07:58.949 reality on the fly, what happens to the whole 00:07:58.949 --> 00:08:01.889 concept of filming a movie or coding a game? 00:08:01.990 --> 00:08:04.569 Traditional production dies. We move from capturing 00:08:04.569 --> 00:08:07.269 reality to hallucinating consistent realities. 00:08:08.120 --> 00:08:11.459 That idea, hallucinating realities, it brings 00:08:11.459 --> 00:08:13.600 us to a more grounding and maybe more concerning 00:08:13.600 --> 00:08:16.699 topic. We have to talk about trust because while 00:08:16.699 --> 00:08:18.879 these models are busy building worlds, they're 00:08:18.879 --> 00:08:20.740 also answering our questions about our health, 00:08:20.819 --> 00:08:24.339 about history. And the latest reports are mixed, 00:08:24.480 --> 00:08:27.120 to say the least. Mixed is a very polite way 00:08:27.120 --> 00:08:28.939 to put it. Yeah. We've got a bit of a crisis 00:08:28.939 --> 00:08:31.240 of trust brewing here. There's a new study out 00:08:31.240 --> 00:08:34.259 about Google's AI overviews, you know, the summaries 00:08:34.259 --> 00:08:36.639 at the top of a search. Right. It turns out. 00:08:36.889 --> 00:08:39.730 When it comes to health information, the AI is 00:08:39.730 --> 00:08:42.730 citing YouTube videos more often than actual 00:08:42.730 --> 00:08:45.389 medical sites. Yeah. More than the Mayo Clinic. 00:08:45.549 --> 00:08:48.070 That is deeply unsettling. I mean, YouTube is 00:08:48.070 --> 00:08:50.870 fantastic if you need to fix a leaky sink. But 00:08:50.870 --> 00:08:53.289 for medical advice, it's a complete minefield 00:08:53.289 --> 00:08:57.190 of anecdotes and unverified claims. Right. And 00:08:57.190 --> 00:08:59.230 the studies show that even reputable medical 00:08:59.230 --> 00:09:01.690 sources were ranking lower than video content. 00:09:02.049 --> 00:09:04.669 It looks like the algorithm is prioritizing engagement. 00:09:05.370 --> 00:09:07.610 what people click on over actual medical accuracy. 00:09:07.990 --> 00:09:10.350 So it's the Internet's popularity contest problem, 00:09:10.509 --> 00:09:12.149 but now it's being presented as an authoritative 00:09:12.149 --> 00:09:15.350 answer by an AI. Precisely. And this isn't just 00:09:15.350 --> 00:09:17.110 a Google problem. We're seeing tension between 00:09:17.110 --> 00:09:21.190 the big labs, too. DeepMind's CEO publicly called 00:09:21.190 --> 00:09:23.929 out OpenAI recently. Yeah, that was spicy. What 00:09:23.929 --> 00:09:25.909 happened there? He basically questioned them, 00:09:25.950 --> 00:09:29.789 rushing ads into ChatGPT. His point was, you 00:09:29.789 --> 00:09:32.090 know, how can you claim to be a trusted assistant 00:09:32.090 --> 00:09:34.730 if you're also trying to sell me something? It's 00:09:34.730 --> 00:09:37.110 a conflict of interest. If I ask my assistant 00:09:37.110 --> 00:09:39.750 for the best running shoe, I want the best shoe, 00:09:39.850 --> 00:09:41.730 not the one that paid for a placement. That makes 00:09:41.730 --> 00:09:44.529 sense. And then there's this issue of models 00:09:44.529 --> 00:09:46.289 learning from other models. We saw that with 00:09:46.289 --> 00:09:48.970 GPT -5 .2, right? Oh, the Grokopedia incident. 00:09:49.090 --> 00:09:52.730 Yes. So reportedly, GPT -5 .2 started referencing 00:09:52.730 --> 00:09:55.769 Elon Musk's Grokopedia, which is the data source 00:09:55.769 --> 00:09:57.950 for his Grok model. Okay. And the problem with 00:09:57.950 --> 00:10:00.389 that is? The problem is Grok has a very specific, 00:10:00.490 --> 00:10:03.049 let's call it an edgy worldview. So all of a 00:10:03.049 --> 00:10:06.190 sudden, GPT -5 .2 starts giving these weird controversial 00:10:06.190 --> 00:10:09.210 takes on really sensitive topics like AIDS or 00:10:09.210 --> 00:10:16.309 slavery. It's like a contagion effect. One model 00:10:16.309 --> 00:10:18.450 hallucinates or carries a bias, and the next 00:10:18.450 --> 00:10:21.429 model comes along, scrapes that output, and treats 00:10:21.429 --> 00:10:24.509 it as fact. Exactly. It's a huge data contamination 00:10:24.509 --> 00:10:27.710 problem. If the internet is just flooding with 00:10:27.710 --> 00:10:30.870 AI -generated content, then new models are just 00:10:30.870 --> 00:10:33.889 training on old models' outputs. It's like making 00:10:33.889 --> 00:10:36.590 a photocopy of a photocopy. The image just... 00:10:36.779 --> 00:10:39.539 degrades over time. So if major models are citing 00:10:39.539 --> 00:10:41.899 YouTube and each other's hallucinations, are 00:10:41.899 --> 00:10:44.240 we just entering a feedback loop of misinformation? 00:10:44.799 --> 00:10:47.960 Yes. It's a digital echo chamber degrading the 00:10:47.960 --> 00:10:49.899 quality of truth. We're going to take a brief 00:10:49.899 --> 00:10:51.720 pause here. All right. We've talked about the 00:10:51.720 --> 00:10:54.759 high level economics, the futuristic tech. Let's 00:10:54.759 --> 00:10:56.720 try to bring this back down to earth a bit. What 00:10:56.720 --> 00:10:58.820 can people listening actually use this week? 00:10:58.860 --> 00:11:01.379 We've got some updates from Google, Apple and 00:11:01.379 --> 00:11:03.820 a few interesting startups. OK, let's start with 00:11:03.820 --> 00:11:07.700 learning. Google has this six -hour prompt engineering 00:11:07.700 --> 00:11:10.200 course, which I know sounds exhausting. Yeah, 00:11:10.220 --> 00:11:12.039 that's a commitment. But a newsletter author 00:11:12.039 --> 00:11:15.100 we follow distilled it down beautifully. And 00:11:15.100 --> 00:11:17.440 the core value isn't just a list of magic words. 00:11:17.779 --> 00:11:20.580 It's about understanding these five key principles 00:11:20.580 --> 00:11:23.379 that help you start to think like the model thinks. 00:11:23.840 --> 00:11:26.460 Thinking like the model. That really does seem 00:11:26.460 --> 00:11:29.139 to be the critical skill for 2026. It's not about 00:11:29.139 --> 00:11:31.419 memorizing commands. It's about understanding 00:11:31.419 --> 00:11:35.000 the logic flow. Exactly. It's about context, 00:11:35.059 --> 00:11:38.320 constraints, iteration. And speaking of iteration, 00:11:38.620 --> 00:11:41.340 I have to make a vulnerable admission here. Go 00:11:41.340 --> 00:11:43.740 for it. I still wrestle with prompt drift myself. 00:11:44.159 --> 00:11:46.200 You know, you get a great result and then you 00:11:46.200 --> 00:11:48.440 change one tiny word and suddenly the AI just 00:11:48.440 --> 00:11:50.440 goes completely off the rails. Oh, it happens 00:11:50.440 --> 00:11:52.480 to everyone. You think you finally mastered it 00:11:52.480 --> 00:11:54.639 and then the model decides that concise means 00:11:54.639 --> 00:11:58.120 be rude. Precisely. So these courses are actually 00:11:58.120 --> 00:12:00.799 pretty vital. But for people who just want the 00:12:00.799 --> 00:12:03.259 tech to work better, Apple is finally catching 00:12:03.259 --> 00:12:05.330 up. We're hearing that the Apple intelligence 00:12:05.330 --> 00:12:07.789 updates coming in February are pretty significant. 00:12:08.110 --> 00:12:10.350 Siri will finally be able to read on -screen 00:12:10.350 --> 00:12:12.309 content. That seems like such a basic feature. 00:12:12.389 --> 00:12:15.230 It's amazing that's been missing. It's huge for 00:12:15.230 --> 00:12:17.070 context. If you're looking at an email and you 00:12:17.070 --> 00:12:19.289 say, remind me about this, Siri will finally 00:12:19.289 --> 00:12:22.570 know what this is. And there's a bigger, more 00:12:22.570 --> 00:12:25.370 chatbot -style update planned for June. We also 00:12:25.370 --> 00:12:27.570 saw a big acquisition in the business world. 00:12:27.809 --> 00:12:30.470 Yelp bought a startup called Hatch for, what, 00:12:30.470 --> 00:12:33.730 $270 million? Yeah, this is a smart move. Hatch 00:12:33.730 --> 00:12:37.230 is an AI for service businesses. So think plumbers, 00:12:37.350 --> 00:12:39.470 movers, that sort of thing. It helps them auto 00:12:39.470 --> 00:12:41.450 -reply to customers. Yelp is basically saying, 00:12:41.549 --> 00:12:44.110 we don't just want to list your business. We 00:12:44.110 --> 00:12:46.090 want to handle your frontline customer service. 00:12:46.309 --> 00:12:49.110 It's automation for the less glamorous but essential 00:12:49.110 --> 00:12:52.210 parts of the economy. Totally. And any smaller 00:12:52.210 --> 00:12:54.779 tools that caught your eye this week? Two, really 00:12:54.779 --> 00:12:57.379 quickly. One is called Pliable. It's an AI -native 00:12:57.379 --> 00:12:59.940 analytics platform. It helps you see the forest 00:12:59.940 --> 00:13:02.600 for the trees with your data, but without needing 00:13:02.600 --> 00:13:05.059 a data science degree. The other one is for creators. 00:13:05.340 --> 00:13:08.460 Thumbfy .est. Thumbfast. Yeah. You drop in a 00:13:08.460 --> 00:13:12.080 face, describe a vision, and it generates a YouTube 00:13:12.080 --> 00:13:14.480 thumbnail instantly. You can grab inspiration 00:13:14.480 --> 00:13:16.299 from thumbnails that are already out there and 00:13:16.299 --> 00:13:18.539 just iterate until it's perfect. It's pure, simple 00:13:18.539 --> 00:13:21.960 utility. It's fascinating. We have A .I. negotiating 00:13:21.960 --> 00:13:25.240 million dollar drug discoveries on one end and 00:13:25.240 --> 00:13:27.600 A .I. making YouTube thumbnails on the other. 00:13:27.759 --> 00:13:30.139 The range is just staggering. That's the ecosystem 00:13:30.139 --> 00:13:32.980 right now. Yeah. From the sublime. all the way 00:13:32.980 --> 00:13:35.159 to the clickbait. With tools automating everything 00:13:35.159 --> 00:13:38.179 from thumbnails to customer service, what is 00:13:38.179 --> 00:13:40.899 the remaining role for human intuition? Curating 00:13:40.899 --> 00:13:43.399 the output and defining the creative vision. 00:13:43.600 --> 00:13:45.279 Let's try to pull all of this together. We've 00:13:45.279 --> 00:13:47.580 covered a lot of ground today. We have. And I 00:13:47.580 --> 00:13:49.200 think there really is a through line here. We're 00:13:49.200 --> 00:13:52.440 seeing the industry mature, maybe. Maybe it's 00:13:52.440 --> 00:13:54.879 losing some of its innocence. That's a good way 00:13:54.879 --> 00:13:57.870 to put it. Economically, we're moving from simple 00:13:57.870 --> 00:14:00.750 subscriptions to this complex idea of value sharing. 00:14:01.090 --> 00:14:04.269 Open AI wanting royalties is a sign they believe 00:14:04.269 --> 00:14:06.570 their product isn't just software anymore, it's 00:14:06.570 --> 00:14:08.769 a co -founder. And technologically, we're graduating 00:14:08.769 --> 00:14:11.509 from static images to these physics -aware worlds. 00:14:11.769 --> 00:14:15.470 That GPT -2 moment for world models means we 00:14:15.470 --> 00:14:18.350 are right on the cusp of generated reality being 00:14:18.350 --> 00:14:22.429 standard. But culturally, we're... are really 00:14:22.429 --> 00:14:25.049 struggling with this messy middle. The trust 00:14:25.049 --> 00:14:27.990 issues at Google, the data contamination with 00:14:27.990 --> 00:14:30.850 Grokopedia. It shows that as these models get 00:14:30.850 --> 00:14:33.190 more powerful, they also get more dangerous if 00:14:33.190 --> 00:14:35.730 we aren't careful about what we feed them. Right. 00:14:35.850 --> 00:14:38.490 The models are becoming less like tools and more 00:14:38.490 --> 00:14:42.629 like partners. Like any partner, sometimes they're 00:14:42.629 --> 00:14:44.789 brilliant -like when they discover a new material, 00:14:44.929 --> 00:14:47.690 and sometimes they're just completely unreliable 00:14:47.690 --> 00:14:50.190 citing a random YouTuber for medical advice. 00:14:50.409 --> 00:14:52.490 It forces us to be sharper, I think. We can't 00:14:52.490 --> 00:14:54.610 just passively consume the output anymore, we 00:14:54.610 --> 00:14:57.360 have to audit it. Absolutely. The human in the 00:14:57.360 --> 00:14:59.360 loop is more important than ever, even if that 00:14:59.360 --> 00:15:01.679 loop is getting faster and more automated. If 00:15:01.679 --> 00:15:03.399 you want to dive a little deeper on any of this, 00:15:03.519 --> 00:15:06.059 I'd really recommend checking out that distillation 00:15:06.059 --> 00:15:07.879 of the Google prompt course. It's a low stakes 00:15:07.879 --> 00:15:09.899 way to just sharpen your skills. Yeah. And if 00:15:09.899 --> 00:15:12.120 you have access, play around with Odyssey or 00:15:12.120 --> 00:15:15.139 World Labs. Seeing a static image turn into a 00:15:15.139 --> 00:15:17.360 3D world is it's something you really have to 00:15:17.360 --> 00:15:19.649 experience to understand. It's a trip. Highly 00:15:19.649 --> 00:15:21.649 recommended. I want to leave you with one final 00:15:21.649 --> 00:15:24.730 thought today. We talked about OpenAI wanting 00:15:24.730 --> 00:15:28.669 a cut of your profits because their AI helped 00:15:28.669 --> 00:15:31.590 you think. It's an interesting argument, but 00:15:31.590 --> 00:15:34.330 it raises the flip side of that coin. If OpenAI 00:15:34.330 --> 00:15:36.789 wants a share of the win because their model 00:15:36.789 --> 00:15:39.570 was a partner, does that mean they're also liable 00:15:39.570 --> 00:15:42.710 if their AI helps you fail? If the researcher 00:15:42.710 --> 00:15:45.370 that never sleeps hallucinates a strategy that 00:15:45.370 --> 00:15:47.679 bankrupts your company. Do they share in that 00:15:47.679 --> 00:15:50.340 loss? That is the billion dollar question, isn't 00:15:50.340 --> 00:15:51.960 it? Thanks for listening. We'll see you in the 00:15:51.960 --> 00:15:53.379 deep end next time. Bye for now.