WEBVTT 00:00:00.000 --> 00:00:02.899 it is uh it's pretty strange to think about sometimes 00:00:02.899 --> 00:00:05.540 yeah it really is your devices are constantly 00:00:05.540 --> 00:00:08.679 mapping your daily reality but they are also 00:00:08.679 --> 00:00:11.580 actively trying to reshape it right they record 00:00:11.580 --> 00:00:14.539 our habits quietly in the background then they 00:00:14.539 --> 00:00:17.280 subtly try to alter our future behavior and it 00:00:17.280 --> 00:00:19.920 happens without us even really noticing it welcome 00:00:19.920 --> 00:00:23.570 to this personalized deep dive we have a uh a 00:00:23.570 --> 00:00:26.550 truly fascinating journey ahead for you. We really 00:00:26.550 --> 00:00:28.789 do. We're covering some wild territory today. 00:00:28.910 --> 00:00:31.309 When you see an app crunching your life's data 00:00:31.309 --> 00:00:34.130 overnight, it seems, you know, kind of quirky. 00:00:34.350 --> 00:00:36.549 Yeah, it's a fun little consumer toy. Exactly. 00:00:36.590 --> 00:00:38.850 But you zoom out and you see the bigger picture. 00:00:39.070 --> 00:00:42.049 The sheer computing power required is driving 00:00:42.049 --> 00:00:45.630 a massive infrastructure race. A $700 billion 00:00:45.630 --> 00:00:48.950 race, to be exact. Right. And that massive race 00:00:48.950 --> 00:00:51.609 is creating a powerful counter -movement. We're 00:00:51.609 --> 00:00:54.140 seeing a shift toward... entirely private, open 00:00:54.140 --> 00:00:57.020 source, local models. Bringing all that power 00:00:57.020 --> 00:00:59.119 right back to your own laptop. Two sec silence. 00:00:59.979 --> 00:01:02.200 So we're going to begin at the extreme consumer 00:01:02.200 --> 00:01:04.920 edge today. This is where massive tech meets 00:01:04.920 --> 00:01:06.840 your morning routine. Yeah, right at the breakfast 00:01:06.840 --> 00:01:09.560 table. Google Labs recently dropped a highly 00:01:09.560 --> 00:01:12.219 experimental new app. It's currently available 00:01:12.219 --> 00:01:15.840 for both iOS and Android platforms. The project 00:01:15.840 --> 00:01:18.459 is officially called Dream Beans. Which, I gotta 00:01:18.459 --> 00:01:21.120 say, sounds like a complete fever dream. It really 00:01:21.120 --> 00:01:23.299 does sound a bit bizarre. But the technology 00:01:23.299 --> 00:01:25.180 running behind it is absolutely fascinating. 00:01:25.540 --> 00:01:27.780 Dream Beans relies heavily on your permission 00:01:27.780 --> 00:01:31.200 -backed Google data. It curates highly personalized 00:01:31.200 --> 00:01:33.879 stories every single morning. And we aren't just 00:01:33.879 --> 00:01:36.189 talking about a couple of alerts. It generates 00:01:36.189 --> 00:01:39.590 roughly 10 to 14 of these stories. And they aren't 00:01:39.590 --> 00:01:41.870 just simple text updates either. They're fully 00:01:41.870 --> 00:01:44.549 animated, narrative -driven stories about your 00:01:44.549 --> 00:01:48.269 actual life. It is wild. The app crunches your 00:01:48.269 --> 00:01:50.849 massive data footprint while you sleep. Right. 00:01:50.890 --> 00:01:52.870 It looks across your entire connected Google 00:01:52.870 --> 00:01:54.989 ecosystem. It checks your calendar, your maps, 00:01:55.069 --> 00:01:57.629 your search history. Then it distills everything 00:01:57.629 --> 00:02:00.530 into a concentrated morning drop. Let's look 00:02:00.530 --> 00:02:02.370 at a specific example from the source material. 00:02:02.609 --> 00:02:04.620 Okay, yeah. Say you have... Get a puppy on your 00:02:04.620 --> 00:02:07.099 calendar. The app actively sees that specific 00:02:07.099 --> 00:02:10.639 calendar event. It then generates a custom animated 00:02:10.639 --> 00:02:13.719 guide just for you. Right. It shows exactly what 00:02:13.719 --> 00:02:15.819 to expect that first week. It basically creates 00:02:15.819 --> 00:02:17.979 a customized narrative out of your own schedule. 00:02:18.219 --> 00:02:20.780 Yeah. And it pulls in relevant tips and weather 00:02:20.780 --> 00:02:23.659 data. It even integrates your location info seamlessly. 00:02:24.319 --> 00:02:26.419 It's like having a personal Pixar studio on your 00:02:26.419 --> 00:02:28.740 phone, just animating your daily to -do list 00:02:28.740 --> 00:02:31.300 every single morning. Yet is a perfect way to 00:02:31.300 --> 00:02:33.819 visualize the experience. It makes your mundane 00:02:33.819 --> 00:02:36.860 daily tasks feel much more engaging. Where did 00:02:36.860 --> 00:02:38.759 that bizarre name actually come from anyway? 00:02:39.080 --> 00:02:41.860 Product lead Gazda Osner explained the origin 00:02:41.860 --> 00:02:44.139 quite recently. Okay, what was the reasoning? 00:02:44.599 --> 00:02:47.240 Well, the dream part of the title is quite literal. 00:02:47.340 --> 00:02:49.319 The system does all its heavy lifting while you 00:02:49.319 --> 00:02:52.379 sleep. Ah, so it processes your digital life 00:02:52.379 --> 00:02:55.379 while you're actually dreaming. Exactly. And 00:02:55.379 --> 00:02:57.520 what about the beans part of the name? Let me 00:02:57.520 --> 00:03:01.199 guess. Coffee. You got it. It represents a freshly 00:03:01.199 --> 00:03:04.560 brewed cup of morning coffee. It's meant to kickstart 00:03:04.560 --> 00:03:06.939 your early morning with energy. Now, privacy 00:03:06.939 --> 00:03:09.580 is obviously a massive question right here. You're 00:03:09.580 --> 00:03:12.240 giving an app access to your entire digital life. 00:03:12.500 --> 00:03:14.979 That requires an enormous amount of personal 00:03:14.979 --> 00:03:17.879 trust. It definitely requires a massive leap 00:03:17.879 --> 00:03:21.469 of faith. But according to Google, the data stays 00:03:21.469 --> 00:03:24.810 localized to you alone. You maintain total control 00:03:24.810 --> 00:03:27.250 over your own personal information. Right. You 00:03:27.250 --> 00:03:29.389 can delete your entire history at any given time. 00:03:29.530 --> 00:03:31.949 Right now, consumer access is still pretty limited, 00:03:32.030 --> 00:03:34.210 though. Yeah, it's only available to U .S.-based 00:03:34.210 --> 00:03:37.210 users at the moment. And specifically, you need 00:03:37.210 --> 00:03:40.710 a Google AI Ultra subscription to use it. But 00:03:40.710 --> 00:03:42.849 there is a public waitlist open right now. Yep. 00:03:42.889 --> 00:03:45.150 Anyone with a personal Google account can actually 00:03:45.150 --> 00:03:47.389 join it. You can wait in line to see your life 00:03:47.389 --> 00:03:49.590 cartoonified. I'm curious about the underlying 00:03:49.590 --> 00:03:51.830 psychology of this tool, though. What's the actual 00:03:51.830 --> 00:03:54.930 goal of this quirky little app? Google essentially 00:03:54.930 --> 00:03:57.050 wants to give you a quick burst of inspiration. 00:03:57.270 --> 00:03:59.930 The idea is to deeply motivate you with these 00:03:59.930 --> 00:04:02.469 personalized stories. Then you confidently put 00:04:02.469 --> 00:04:04.849 the phone down to live your life. So a digital 00:04:04.849 --> 00:04:08.289 push to go live your actual life. Let's move 00:04:08.289 --> 00:04:09.949 away from personal data curation for a moment. 00:04:10.069 --> 00:04:12.509 We really need to look at the massive industry 00:04:12.509 --> 00:04:15.370 consequences here. Yeah, because we rely so incredibly 00:04:15.370 --> 00:04:17.899 heavily on these intelligence systems now. The 00:04:17.899 --> 00:04:20.439 shift from personal tools to global infrastructure 00:04:20.439 --> 00:04:23.980 is staggering. It completely changes how we interact 00:04:23.980 --> 00:04:27.199 with information daily. Anthropic recently shared 00:04:27.199 --> 00:04:29.860 a very interesting prompt system regarding this. 00:04:29.920 --> 00:04:33.300 It's specifically designed to keep humans actively 00:04:33.300 --> 00:04:36.620 thinking alongside AI. They're growing increasingly 00:04:36.620 --> 00:04:39.000 worried about our collective cognitive habits. 00:04:39.240 --> 00:04:42.459 Extremely worried. Letting an AI entirely think 00:04:42.459 --> 00:04:45.420 for you is quite dangerous. It can quietly weaken 00:04:45.420 --> 00:04:47.899 your own critical judgment over time. I still 00:04:47.899 --> 00:04:50.399 wrestle with prompt drift myself, honestly. I 00:04:50.399 --> 00:04:52.480 catch myself letting AI do my deep thinking. 00:04:53.720 --> 00:04:56.339 It honestly happens to everyone who uses these 00:04:56.339 --> 00:04:59.139 tools regularly. You just slowly start outsourcing 00:04:59.139 --> 00:05:01.279 your own critical thought processes. Right. And 00:05:01.279 --> 00:05:03.560 Anthropic desperately wants you to remain actively 00:05:03.560 --> 00:05:05.319 part of it. Yeah, they want you in the cognitive 00:05:05.319 --> 00:05:08.360 loop. We're also seeing intense new rules emerging 00:05:08.360 --> 00:05:11.660 globally. Over in the UK, new regulations are 00:05:11.660 --> 00:05:14.040 pushing back on Google. Content publishers are 00:05:14.040 --> 00:05:16.000 finally demanding more control over their work. 00:05:16.319 --> 00:05:18.980 They're tired of their data being endlessly scraped 00:05:18.980 --> 00:05:21.079 for free. For the first time, they can actually 00:05:21.079 --> 00:05:23.540 block their content. They can stop it from appearing 00:05:23.540 --> 00:05:26.160 in AI overviews completely. And they can also 00:05:26.160 --> 00:05:29.259 block it from AI mode entirely. Right. And block 00:05:29.259 --> 00:05:31.639 AI -generated search answers as well. It's a 00:05:31.639 --> 00:05:33.639 massive shift in intellectual property control 00:05:33.639 --> 00:05:36.379 and digital rights. Speaking of massive shifts, 00:05:36.519 --> 00:05:38.259 let's talk about Amazon for a second. They are 00:05:38.259 --> 00:05:41.120 rolling out a very strange new feature right 00:05:41.120 --> 00:05:43.899 now. Oh, yeah. When you search for items, things 00:05:43.899 --> 00:05:46.560 might look quite different. You might start seeing 00:05:46.560 --> 00:05:49.579 AI -generated product images first. These appear 00:05:49.579 --> 00:05:52.639 before you even see the real physical products. 00:05:52.779 --> 00:05:55.199 They're showing fake product concepts to help 00:05:55.199 --> 00:05:58.110 you shop. Yes. Fake product images. to help you 00:05:58.110 --> 00:06:01.029 find real ones. I have to push back on this Amazon 00:06:01.029 --> 00:06:03.329 feature, honestly. Why look at fake products 00:06:03.329 --> 00:06:06.029 to find real ones? It sounds weird, I know. It 00:06:06.029 --> 00:06:08.189 seems completely counterintuitive to the whole 00:06:08.189 --> 00:06:10.269 shopping experience. If I want a toaster, I want 00:06:10.269 --> 00:06:12.870 a real toaster. I don't want an AI hallucination 00:06:12.870 --> 00:06:15.490 of a toaster. It really does sound deeply confusing 00:06:15.490 --> 00:06:18.069 on the surface. But they want to help you visualize 00:06:18.069 --> 00:06:21.430 abstract lifestyle concepts. They think it helps 00:06:21.430 --> 00:06:23.550 you narrow down a specific aesthetic faster. 00:06:24.220 --> 00:06:26.379 Well, Google is also aggressively dealing with 00:06:26.379 --> 00:06:29.259 fake content right now. They're rolling out AI 00:06:29.259 --> 00:06:32.040 -powered fake call detection software. Which 00:06:32.040 --> 00:06:34.399 is becoming a highly crucial security feature 00:06:34.399 --> 00:06:37.300 right now. It can intelligently check if a trusted 00:06:37.300 --> 00:06:40.000 contact is really calling. It verifies if the 00:06:40.000 --> 00:06:42.459 familiar voice on the line is authentic. Voice 00:06:42.459 --> 00:06:44.480 cloning has become a massive global security 00:06:44.480 --> 00:06:47.860 problem lately. Scammers can perfectly mimic 00:06:47.860 --> 00:06:50.579 the voices of your loved ones. This defensive 00:06:50.579 --> 00:06:53.339 feature is now available on Android 14 devices. 00:06:53.579 --> 00:06:56.860 It runs entirely natively on Android 14 plus 00:06:56.860 --> 00:06:59.420 devices globally. Doing all of this requires 00:06:59.420 --> 00:07:01.899 an unbelievable amount of money, though. The 00:07:01.899 --> 00:07:04.939 financial scale of this specific industry is 00:07:04.939 --> 00:07:07.560 just staggering. Alphabet recently announced 00:07:07.560 --> 00:07:10.660 plans to raise $80 billion. They're expanding 00:07:10.660 --> 00:07:13.839 their AI infrastructure as rapidly as humanly 00:07:13.839 --> 00:07:16.040 possible. They actually expect to spend up to 00:07:16.040 --> 00:07:20.079 $190 billion. That's just on AI -related capital 00:07:20.079 --> 00:07:22.519 expenditures for this year alone. The broader 00:07:22.519 --> 00:07:25.000 tech sector numbers are even crazier to comprehend. 00:07:25.339 --> 00:07:28.240 Tech giants could collectively invest an estimated 00:07:28.240 --> 00:07:32.000 $700 billion. That's nearly a trillion dollars 00:07:32.000 --> 00:07:35.000 in raw physical infrastructure. It requires massive 00:07:35.000 --> 00:07:37.639 data centers, cooling systems, and enormous power 00:07:37.639 --> 00:07:40.680 grids. With all that power, societal impact becomes 00:07:40.680 --> 00:07:44.699 a huge concern. Why is Anthropic hiring an AI 00:07:44.699 --> 00:07:48.540 rule and law team for 345K? They urgently need 00:07:48.540 --> 00:07:51.790 to study how AI affects our society. Specifically, 00:07:51.990 --> 00:07:54.310 they're researching the impacts on courts and 00:07:54.310 --> 00:07:57.069 global elections. Building guardrails before 00:07:57.069 --> 00:08:00.170 the system breaks our civic institutions. Sponsor. 00:08:00.670 --> 00:08:03.189 We just talked about Alphabet spending $190 billion. 00:08:03.550 --> 00:08:06.209 What does that kind of astronomical money actually 00:08:06.209 --> 00:08:09.089 buy? Well, it buys immense computational power 00:08:09.089 --> 00:08:11.209 and vital research breakthroughs. It completely 00:08:11.209 --> 00:08:13.050 changes the physical limits of what computers 00:08:13.050 --> 00:08:15.470 can do. It buys the incredible ability to shrink 00:08:15.470 --> 00:08:18.569 massive cloud power. You can now fit that power 00:08:18.569 --> 00:08:21.449 locally on your own laptop. This brings us directly 00:08:21.449 --> 00:08:24.550 to a major open source release. Google DeepMind 00:08:24.550 --> 00:08:27.230 just officially dropped the Gemma 412B model. 00:08:27.529 --> 00:08:30.230 This model handles complex agentic workflows 00:08:30.230 --> 00:08:33.149 entirely locally. You can run it directly on 00:08:33.149 --> 00:08:35.730 your own personal machine. It's the exact model 00:08:35.730 --> 00:08:38.750 many developers have been waiting for. It bridges 00:08:38.750 --> 00:08:41.809 the huge gap between massive clouds and local 00:08:41.809 --> 00:08:44.649 hardware. DeepMind completely re -engineered 00:08:44.649 --> 00:08:46.909 how the model actually processes information. 00:08:47.389 --> 00:08:50.250 They fundamentally changed how it sees and hues 00:08:50.250 --> 00:08:53.149 the world. Vision and audio now flow directly 00:08:53.149 --> 00:08:55.710 into the main backbone. They no longer have to 00:08:55.710 --> 00:08:58.149 be awkwardly translated into text first. This 00:08:58.149 --> 00:09:00.690 drastically cuts latency across the entire computing 00:09:00.690 --> 00:09:03.570 system. It also reduces the heavy memory usage 00:09:03.570 --> 00:09:05.850 quite significantly. This is their very first 00:09:05.850 --> 00:09:09.059 mid -sized model to handle audio natively. This 00:09:09.059 --> 00:09:11.879 multimodal data processing is absolutely fascinating 00:09:11.879 --> 00:09:14.379 to me. It's like stacking Lego blocks of data. 00:09:14.639 --> 00:09:16.200 Yeah, that's a great way to describe it. You 00:09:16.200 --> 00:09:18.799 connect text, vision, and audio directly together 00:09:18.799 --> 00:09:21.000 without translating them. They simplified the 00:09:21.000 --> 00:09:23.539 underlying architecture so much to achieve this. 00:09:24.240 --> 00:09:26.820 raw audio signals project directly into the exact 00:09:26.820 --> 00:09:29.840 same space they seamlessly occupy the exact same 00:09:29.840 --> 00:09:32.820 space as text tokens do this allows it to translate 00:09:32.820 --> 00:09:35.759 and transcribe entirely offline it doesn't need 00:09:35.759 --> 00:09:40.419 to ping a massive server farm anymore beat. Imagine 00:09:40.419 --> 00:09:43.899 native raw audio processing running entirely 00:09:43.899 --> 00:09:47.080 offline on a regular consumer laptop. It's a 00:09:47.080 --> 00:09:49.860 truly massive leap forward for local computing 00:09:49.860 --> 00:09:53.080 power. The benchmark performance is also shockingly 00:09:53.080 --> 00:09:55.820 impressive for its size. It performs very close 00:09:55.820 --> 00:09:59.279 to Google's much larger, heavier models. Specifically, 00:09:59.480 --> 00:10:02.600 it rivals the massive 26B mixture of experts 00:10:02.600 --> 00:10:05.009 model. But it still comfortably runs locally 00:10:05.009 --> 00:10:08.009 on standard consumer laptops. That kind of efficiency 00:10:08.009 --> 00:10:10.570 was practically unheard of last year. It also 00:10:10.570 --> 00:10:12.870 features something called MTP drafters built 00:10:12.870 --> 00:10:15.490 right in. This keeps the generation of complex 00:10:15.490 --> 00:10:18.450 text incredibly fast. Let's quickly clarify that 00:10:18.450 --> 00:10:21.149 concept for a moment. What are MTP drafters exactly? 00:10:21.509 --> 00:10:23.809 They fundamentally change how the AI writes its 00:10:23.809 --> 00:10:26.269 responses. Helpers that guess upcoming words 00:10:26.269 --> 00:10:28.570 faster to speed up the whole system. Exactly. 00:10:28.710 --> 00:10:30.909 It predicts multiple tokens at once to save crucial 00:10:30.909 --> 00:10:33.169 time. It doesn't waste time agonizing over a 00:10:33.169 --> 00:10:35.230 single word choice anymore. The absolute best 00:10:35.230 --> 00:10:38.210 part is how accessible this model is. It's completely 00:10:38.210 --> 00:10:41.809 open under a permissive Apache 2 .0 license. 00:10:43.279 --> 00:10:45.340 the way it's on Hugging Face right now, you can 00:10:45.340 --> 00:10:48.019 also find them hosted on Kaggle immediately today. 00:10:48.379 --> 00:10:50.820 You can spin it up instantly in familiar developer 00:10:50.820 --> 00:10:54.980 tools, tools like Llama, LM Studio, and Llama 00:10:54.980 --> 00:10:57.899 .cpp. The developer community is already building 00:10:57.899 --> 00:11:00.220 truly amazing things with it. The open source 00:11:00.220 --> 00:11:02.500 world moves at an absolutely blistering pace. 00:11:02.779 --> 00:11:04.919 There is another tool mentioned alongside this 00:11:04.919 --> 00:11:08.299 release. What exactly is the Gemma Skills Repository 00:11:08.299 --> 00:11:10.659 mentioned in the source? It's an official developer 00:11:10.659 --> 00:11:13.399 toolkit launching alongside the main model. It 00:11:13.399 --> 00:11:15.879 helps you build autonomous, multi -step agents 00:11:15.879 --> 00:11:18.399 right out of the box. Basically a starter kit 00:11:18.399 --> 00:11:22.120 for building your own offline AI workers. Beat. 00:11:22.340 --> 00:11:24.700 We're in a highly weird transitional phase right 00:11:24.700 --> 00:11:27.000 now. The tech industry is spending nearly a trillion 00:11:27.000 --> 00:11:29.440 dollars collectively. They're rapidly building 00:11:29.440 --> 00:11:32.299 out massive, highly centralized cloud infrastructure 00:11:32.299 --> 00:11:35.500 globally. They desperately want to hoard our 00:11:35.500 --> 00:11:38.940 cloud data for morning cartoons. Apps like DreamBeans 00:11:38.940 --> 00:11:41.879 rely entirely on this massive centralized ecosystem. 00:11:42.460 --> 00:11:45.240 They need your data living on their servers to 00:11:45.240 --> 00:11:47.159 function properly. But they're simultaneously 00:11:47.159 --> 00:11:49.940 moving rapidly in the exact opposite direction. 00:11:50.259 --> 00:11:52.320 They're giving away the keys to the kingdom at 00:11:52.320 --> 00:11:54.700 the same time. They're democratizing offline 00:11:54.700 --> 00:11:57.740 open source power globally with Gemma 4. You 00:11:57.740 --> 00:12:00.000 can now run incredibly smart agents completely 00:12:00.000 --> 00:12:02.759 on your laptop. The deep tension between centralized 00:12:02.759 --> 00:12:06.299 cloud and local privacy is fascinating. We're 00:12:06.299 --> 00:12:08.720 watching. two entirely different philosophies 00:12:08.720 --> 00:12:10.779 battle for dominance right now. It really makes 00:12:10.779 --> 00:12:12.840 you wonder about the long -term future of these 00:12:12.840 --> 00:12:16.159 tools. If models like Gemma 4 get so incredibly 00:12:16.159 --> 00:12:19.259 good natively, if they can understand our daily 00:12:19.259 --> 00:12:21.879 reality completely offline, will we eventually 00:12:21.879 --> 00:12:24.360 abandon cloud -dependent apps like DreamBeans 00:12:24.360 --> 00:12:27.039 completely? Might we trade them for completely 00:12:27.039 --> 00:12:30.220 private, entirely local AI companions? It's a 00:12:30.220 --> 00:12:32.480 profoundly important question about who truly 00:12:32.480 --> 00:12:34.700 controls our data. And it's a question we'll 00:12:34.700 --> 00:12:36.899 have to answer very soon. Thank you so much for 00:12:36.899 --> 00:12:39.500 joining us on this deep dive. We really appreciate 00:12:39.500 --> 00:12:42.740 you exploring these complex ideas alongside us 00:12:42.740 --> 00:12:42.940 today.