WEBVTT 00:00:00.000 --> 00:00:02.720 Here, let's try to unpack this. Just, what, four 00:00:02.720 --> 00:00:04.500 years ago, one of the world's biggest companies, 00:00:04.660 --> 00:00:07.299 Meta, bet the farm, everything on the Metaverse. 00:00:07.440 --> 00:00:09.759 And now we're seeing one of the most dramatic 00:00:09.759 --> 00:00:12.359 corporate reversals in, I don't know, a decade. 00:00:12.599 --> 00:00:15.560 Yeah. They are pulling back tens of billions 00:00:15.560 --> 00:00:19.519 of dollars from that massive VR future and redirecting 00:00:19.519 --> 00:00:22.579 all of it entirely to another massive future. 00:00:23.300 --> 00:00:27.660 That pivot, that $72 billion pivot, really defines 00:00:27.660 --> 00:00:29.920 this moment we're in. It's a fundamental change 00:00:29.920 --> 00:00:32.600 in Silicon Valley's priorities, away from isolated 00:00:32.600 --> 00:00:36.140 virtual reality, toward pervasive intelligence. 00:00:36.560 --> 00:00:39.079 Welcome to the Deep Dive. We have a pretty dense 00:00:39.079 --> 00:00:41.340 stack of sources for you today. We've got articles, 00:00:41.520 --> 00:00:44.200 technical guides, research papers, all detailing 00:00:44.200 --> 00:00:46.320 this huge corporate pivot. We're not just looking 00:00:46.320 --> 00:00:47.740 at the money, we're looking at where the energy 00:00:47.740 --> 00:00:49.920 is going. And our mission here is to give you 00:00:49.920 --> 00:00:52.920 a shortcut to understanding this rapid reorientation. 00:00:53.020 --> 00:00:55.179 We're going to unpack what this huge shift in 00:00:55.179 --> 00:00:57.859 capital means for developers, for content creators 00:00:57.859 --> 00:01:00.020 in the trenches, and especially for the future 00:01:00.020 --> 00:01:03.280 of specialized AI tools, particularly in high 00:01:03.280 --> 00:01:05.340 stakes fields like healthcare. So we'll start 00:01:05.340 --> 00:01:07.799 with the specifics of Meta's retreat. The financial 00:01:07.799 --> 00:01:09.959 stakes are just huge. Then we'll get into the 00:01:09.959 --> 00:01:13.079 practical and sometimes messy world of AI tools 00:01:13.079 --> 00:01:15.799 for creators. Right. After that, we'll tackle 00:01:15.799 --> 00:01:18.099 regulation and the rise of agentic workflow in 00:01:18.099 --> 00:01:20.000 the enterprise. And finally, we'll look at some 00:01:20.000 --> 00:01:22.959 groundbreaking medical AI that could really reshape 00:01:22.959 --> 00:01:25.799 diagnostics. The details of the Metaverse retreat 00:01:25.799 --> 00:01:27.920 are striking. And I want to stress that word. 00:01:28.650 --> 00:01:31.870 retreat. This is not a slow budget adjustment. 00:01:32.290 --> 00:01:35.069 It's a full -on downsizing. That's exactly right. 00:01:35.150 --> 00:01:38.769 They laid off 1 ,500 staff. That's 10 % of their 00:01:38.769 --> 00:01:42.390 Reality Labs unit. But the real signal, you know, 00:01:42.409 --> 00:01:44.829 is what they chose to cut. It's not just the 00:01:44.829 --> 00:01:47.310 headcount. It's the specific projects. They shut 00:01:47.310 --> 00:01:50.590 down three VR game studios, Armature, Twisted 00:01:50.590 --> 00:01:52.840 Pixel, and Sanzaru. These are the studios that 00:01:52.840 --> 00:01:54.879 were supposed to deliver the high concept, high 00:01:54.879 --> 00:01:58.239 cost, AAA gaming ambition for VR. And they moved 00:01:58.239 --> 00:02:00.739 their fitness app, Supernatural, into basically 00:02:00.739 --> 00:02:03.140 maintenance mode. They even killed a bunch of 00:02:03.140 --> 00:02:05.920 social VR features inside Horizon Worlds. It 00:02:05.920 --> 00:02:09.719 all signals a move away from that vision of high 00:02:09.719 --> 00:02:12.580 fidelity, highly immersive virtual spaces that, 00:02:12.639 --> 00:02:14.960 well, frankly, very few people were actually 00:02:14.960 --> 00:02:17.740 visiting. Or staying in. So it's not like they're 00:02:17.740 --> 00:02:19.400 killing the metaverse entirely, but they are 00:02:19.400 --> 00:02:23.259 trading that huge, empty, virtual world idea 00:02:23.259 --> 00:02:26.719 for what our sources call a Roblox -style downgrade. 00:02:26.919 --> 00:02:29.159 Yeah, shifting focus from the grand fantasy to 00:02:29.159 --> 00:02:32.439 a more pragmatic, utilized version of VR and 00:02:32.439 --> 00:02:35.379 AR. It's all about utility now. And the financial 00:02:35.379 --> 00:02:37.659 discipline needed to make those cuts is just. 00:02:38.189 --> 00:02:40.210 It's eclipsed by the spending on the other side. 00:02:40.289 --> 00:02:42.810 Oh, for sure. Contrast those cuts with the AI 00:02:42.810 --> 00:02:45.150 double down. The spending boost is genuinely 00:02:45.150 --> 00:02:48.009 massive. We're talking $72 billion in capital 00:02:48.009 --> 00:02:51.090 expenditure CapEx plan just for 2025 through 00:02:51.090 --> 00:02:53.629 2026. And that money is overwhelmingly focused 00:02:53.629 --> 00:02:55.689 on AI infrastructure. So we're talking chips, 00:02:55.930 --> 00:02:58.590 servers, data centers. And that infrastructure 00:02:58.590 --> 00:03:01.509 supports the people and the models. They acquired 00:03:01.509 --> 00:03:05.030 the AI startup Manas for over $2 billion. They 00:03:05.030 --> 00:03:07.449 made massive strategic hires, including Alexander. 00:03:07.500 --> 00:03:10.199 Xander Wang from Scale AI to Lead Strategy. Right. 00:03:10.319 --> 00:03:12.960 And they have their next -gen large language 00:03:12.960 --> 00:03:15.960 model, or LLM, which they're calling Avocado, 00:03:16.180 --> 00:03:19.840 planned for a Q1 release. Wait, you used LM there. 00:03:19.979 --> 00:03:21.639 For our listener, can you just briefly define 00:03:21.639 --> 00:03:25.240 that? Oh, sure. An LLM is an AI trained on massive 00:03:25.240 --> 00:03:28.439 data to understand and generate text. Think of 00:03:28.439 --> 00:03:30.919 it like a super smart chatbot. Got it. Thank 00:03:30.919 --> 00:03:33.639 you. Now, here's where the strategy seems to 00:03:33.639 --> 00:03:36.520 be working already. Unlike those high -cost, 00:03:36.560 --> 00:03:39.960 isolated VR efforts, their physical, wearable 00:03:39.960 --> 00:03:42.879 AI products are actually succeeding. The Ray 00:03:42.879 --> 00:03:45.379 -Ban MetaSmart glasses are selling extremely 00:03:45.379 --> 00:03:47.639 well. Extremely well. They're backordered in 00:03:47.639 --> 00:03:49.719 the U .S. and they're tracking toward 10 million 00:03:49.719 --> 00:03:52.719 units way sooner than anyone anticipated. And 00:03:52.719 --> 00:03:55.080 this is the critical difference. These glasses 00:03:55.080 --> 00:03:57.939 integrate AI into daily life. Right. You can 00:03:57.939 --> 00:03:59.740 interact with the world without being isolated 00:03:59.740 --> 00:04:02.259 inside a headset. It bridges the physical and 00:04:02.259 --> 00:04:05.180 the digital. And that shows success. In building 00:04:05.180 --> 00:04:06.680 things, people actually want to wear and use 00:04:06.680 --> 00:04:09.000 in the real world. Yeah. It solves that isolation 00:04:09.000 --> 00:04:11.259 problem that plagued the Quest headsets. It's 00:04:11.259 --> 00:04:13.439 a completely different and I think much more 00:04:13.439 --> 00:04:17.100 viable market dynamic. So does this massive capital 00:04:17.100 --> 00:04:20.480 expenditure guarantee Meta's eventual dominance 00:04:20.480 --> 00:04:23.000 in AI? Well, the spending shows commitment for 00:04:23.000 --> 00:04:26.699 sure, but their LLM, Avocado, has to deliver 00:04:26.699 --> 00:04:29.000 some breakthrough results. That financial commitment 00:04:29.000 --> 00:04:31.800 from Meta is huge, but let's shift focus. What 00:04:31.800 --> 00:04:33.639 does this all mean for the person trying to use 00:04:33.639 --> 00:04:35.480 these n****s? new tools every day. Let's talk 00:04:35.480 --> 00:04:37.740 about the creator trenches. Right. Moving from 00:04:37.740 --> 00:04:40.379 the corporate boardroom to the developer's workbench. 00:04:40.639 --> 00:04:43.300 The AI landscape for content creators right now 00:04:43.300 --> 00:04:46.360 is just in a phase of hyper saturation. I still 00:04:46.360 --> 00:04:48.259 wrestle with prompt drift myself. Yeah. I mean, 00:04:48.279 --> 00:04:50.100 if you've ever burned 20 bucks in credits trying 00:04:50.100 --> 00:04:53.319 to generate a usable five to 10 second video 00:04:53.319 --> 00:04:55.620 clip only to have the character suddenly grow 00:04:55.620 --> 00:04:57.959 a third arm or the background just melts. Yeah. 00:04:58.000 --> 00:05:00.779 Yeah. Same. That frustration is so real. That 00:05:00.779 --> 00:05:03.660 experience, that difficulty and expense in getting 00:05:03.660 --> 00:05:06.100 usable output, that's the pain point driving 00:05:06.100 --> 00:05:09.079 the market. Last year, we had maybe two or three 00:05:09.079 --> 00:05:12.420 viable video generation tools. Yeah, maybe. Now, 00:05:12.480 --> 00:05:15.319 there are literally 10 plus competing platforms 00:05:15.319 --> 00:05:18.480 all claiming to be the best. This means testing 00:05:18.480 --> 00:05:21.480 is absolutely crucial. And our sources detailed 00:05:21.480 --> 00:05:23.779 a great methodology for cutting through the noise. 00:05:23.899 --> 00:05:26.879 A researcher spent seven hours testing. They 00:05:26.879 --> 00:05:29.279 used the same exact prompt across all the major 00:05:29.279 --> 00:05:32.180 video models. And they created a clear, measurable 00:05:32.180 --> 00:05:35.660 A to D tier ranking system. And this is so important 00:05:35.660 --> 00:05:38.899 because it's based on objective criteria. The 00:05:38.899 --> 00:05:41.459 top tier tools were ranked on clean motion, consistent 00:05:41.459 --> 00:05:44.019 characters across shots, and minimizing that 00:05:44.019 --> 00:05:47.620 frustrating artifact we all call AI slop. This 00:05:47.620 --> 00:05:49.800 moves the industry past just marketing hype. 00:05:50.569 --> 00:05:53.069 Creators can now invest based on measurable results, 00:05:53.250 --> 00:05:56.329 on documented consistency. And that is a huge 00:05:56.329 --> 00:05:58.649 step forward for workflow. And this whole search 00:05:58.649 --> 00:06:01.389 for consistency and quality feeds directly into 00:06:01.389 --> 00:06:03.509 content strategy. And here's where it gets really 00:06:03.509 --> 00:06:05.509 interesting for the content game. The sources 00:06:05.509 --> 00:06:07.610 highlight some really high -impact, actionable 00:06:07.610 --> 00:06:10.629 findings. One is the ability to build an automated 00:06:10.629 --> 00:06:13.709 digital content empire using Google AIs from 00:06:13.709 --> 00:06:16.629 scratch without immediate spending. And the so 00:06:16.629 --> 00:06:18.730 -called YouTube cloning hack is a really compelling 00:06:18.730 --> 00:06:20.819 example of that strategy. It lets you reverse 00:06:20.819 --> 00:06:23.139 engineer any viral channel in about 10 minutes. 00:06:23.300 --> 00:06:26.910 10 minutes? How does it do that so quickly? By 00:06:26.910 --> 00:06:29.209 turning the channel's entire library of video 00:06:29.209 --> 00:06:32.389 transcripts into a searchable AI database, the 00:06:32.389 --> 00:06:35.250 AI can then index everything, identify what specific 00:06:35.250 --> 00:06:38.410 hooks drive view duration, and pinpoint retention 00:06:38.410 --> 00:06:40.889 points within the videos themselves. Wow. So 00:06:40.889 --> 00:06:42.649 it's like having an instant competitive analysis 00:06:42.649 --> 00:06:45.910 tool that breaks down the why behind viral success. 00:06:46.290 --> 00:06:48.790 Exactly. And the ecosystem is providing immediate 00:06:48.790 --> 00:06:52.329 utility too. For instance, PDF .beauty. This 00:06:52.329 --> 00:06:55.269 tool turns image -based PDFs or slides things 00:06:55.269 --> 00:06:58.209 that are usually locked down into fully editable 00:06:58.209 --> 00:07:01.449 PowerPoint files. Huge time saver. And on the 00:07:01.449 --> 00:07:04.730 security side, PinDrop was mentioned. It detects 00:07:04.730 --> 00:07:06.949 fake voices or videos in real time during live 00:07:06.949 --> 00:07:10.129 calls with 99 % accuracy. These are practical 00:07:10.129 --> 00:07:12.870 applications solving real problems today. For 00:07:12.870 --> 00:07:15.089 the average creator, is consistency or quality 00:07:15.089 --> 00:07:17.269 more important right now? Consistency is key 00:07:17.269 --> 00:07:19.589 for workflow, but quality is what defines the 00:07:19.589 --> 00:07:22.029 output. You really need both. We've talked corporate 00:07:22.029 --> 00:07:24.430 bets and creator tools. Now we're crossing into 00:07:24.430 --> 00:07:27.889 this increasingly complex territory of regulatory 00:07:27.889 --> 00:07:31.250 headaches and high -profile conflicts. Just look 00:07:31.250 --> 00:07:33.189 at the controversy around Grok. It's a perfect 00:07:33.189 --> 00:07:35.769 example of the duality of AI deployment right 00:07:35.769 --> 00:07:39.509 now. This AI is being integrated into Pentagon 00:07:39.509 --> 00:07:42.730 military systems in the U .S. While at the same 00:07:42.730 --> 00:07:45.009 time being banned in places like Indonesia and 00:07:45.009 --> 00:07:48.310 Malaysia. Yeah. And that's due to really serious 00:07:48.310 --> 00:07:50.889 concerns over the generation of sexual imagery. 00:07:51.009 --> 00:07:54.230 The lack of reliable guardrails has immediate 00:07:54.230 --> 00:07:57.250 global political consequences. Right. And it's 00:07:57.250 --> 00:07:59.910 not just Grok. Google's Gemini is also involved 00:07:59.910 --> 00:08:02.810 in high stakes military systems. So this intersection 00:08:02.810 --> 00:08:05.649 of emerging AI and national security is now just 00:08:05.649 --> 00:08:08.529 fully baked into the global discussion. And we 00:08:08.529 --> 00:08:10.550 should note the California attorney general is 00:08:10.550 --> 00:08:12.589 actively investigating Grok regarding claims 00:08:12.589 --> 00:08:14.930 about the generation of naked underage images 00:08:14.930 --> 00:08:17.649 citing Musk's claims of unawareness. Yeah. So 00:08:17.649 --> 00:08:19.990 this raises a big question. How do we manage 00:08:19.990 --> 00:08:21.949 these powerful tools that are being adopted at 00:08:21.949 --> 00:08:23.790 the highest levels but still struggle with basic 00:08:23.790 --> 00:08:26.810 safety filters? That gap is really hard to bridge 00:08:26.810 --> 00:08:29.379 quickly. And while that ethical and regulatory 00:08:29.379 --> 00:08:32.240 battle plays out, the practical side agentic 00:08:32.240 --> 00:08:34.379 workflow is moving very quickly into the enterprise 00:08:34.379 --> 00:08:37.379 space. Slackbot, for example, has been completely 00:08:37.379 --> 00:08:39.519 rebuilt. And this is where we probably need a 00:08:39.519 --> 00:08:42.679 quick definition. Agentic AI just means an AI 00:08:42.679 --> 00:08:45.759 that plans and executes complex tasks using other 00:08:45.759 --> 00:08:48.580 software tools all on its own. Okay, that makes 00:08:48.580 --> 00:08:51.179 sense. So the new Slackbot isn't just a smarter 00:08:51.179 --> 00:08:54.169 search engine. Exactly. It now acts as a true 00:08:54.169 --> 00:08:57.250 agent. It can find info across linked apps. It 00:08:57.250 --> 00:08:59.830 can draft emails. And critically, it can schedule 00:08:59.830 --> 00:09:02.269 meetings without you ever having to leave Slack. 00:09:02.549 --> 00:09:05.230 It initiates actions instead of just answering 00:09:05.230 --> 00:09:08.090 questions. And tools like Vellum are mentioned 00:09:08.090 --> 00:09:10.049 as ways you can build these working AI agents 00:09:10.049 --> 00:09:13.029 using just plain English commands to automate 00:09:13.029 --> 00:09:15.850 your own boring, repetitive tasks. It's about 00:09:15.850 --> 00:09:17.929 empowering people. And this integration goes 00:09:17.929 --> 00:09:21.279 deep into existing systems like HubSpot. Connectors 00:09:21.279 --> 00:09:24.419 now let LLMs like Claude and ChatGPT create and 00:09:24.419 --> 00:09:27.820 update CRM records, log activities, and do deep 00:09:27.820 --> 00:09:29.919 research with citations that link right back 00:09:29.919 --> 00:09:32.240 to the HubSpot records. The citations are huge 00:09:32.240 --> 00:09:34.559 for trust and verification in a business setting. 00:09:34.919 --> 00:09:36.879 They absolutely are. And think about specialized 00:09:36.879 --> 00:09:40.000 integration, too. Descript integration can move 00:09:40.000 --> 00:09:42.820 AI -generated videos between HubSpot libraries 00:09:42.820 --> 00:09:46.279 for easier editing, and this is the key for ROI 00:09:46.279 --> 00:09:49.200 tracking. It ties creative output directly to 00:09:49.200 --> 00:09:51.679 business metrics. So will ethical concerns slow 00:09:51.679 --> 00:09:53.519 down the deployment of these powerful enterprise 00:09:53.519 --> 00:09:56.580 agents? I think regulation is playing catch -up, 00:09:56.580 --> 00:09:59.059 so deployment speed is going to vary widely by 00:09:59.059 --> 00:10:00.980 region. It'll create a patchwork of adoption 00:10:00.980 --> 00:10:03.159 rates. Let's shift focus pretty dramatically 00:10:03.159 --> 00:10:06.309 now. from enterprise productivity to specialized 00:10:06.309 --> 00:10:09.210 AI in healthcare. Google just launched two new 00:10:09.210 --> 00:10:13.830 open models, MedGemma 1 .5 and MedSR. And this 00:10:13.830 --> 00:10:15.690 feels less like an iteration and more like a 00:10:15.690 --> 00:10:18.289 breakthrough. It absolutely is. These open models 00:10:18.289 --> 00:10:20.269 are pushing medical AI from that hypothetical 00:10:20.269 --> 00:10:23.210 research -only phase into actually usable clinical 00:10:23.210 --> 00:10:25.909 scenarios. And here's the key detail. Both of 00:10:25.909 --> 00:10:28.309 them are free for commercial use. That completely 00:10:28.309 --> 00:10:30.210 changes the market. Can you break down the difference 00:10:30.210 --> 00:10:33.559 between the two? Sure. Majema 1 .5 is a smaller, 00:10:33.659 --> 00:10:37.159 faster, multimodal model. It handles images and 00:10:37.159 --> 00:10:39.759 text together. MedSR is much more specialized. 00:10:40.120 --> 00:10:42.519 It's a speech -to -text model tuned specifically 00:10:42.519 --> 00:10:45.519 for real -world medical dictation. Which involves 00:10:45.519 --> 00:10:48.120 a lot of jargon, fast -speaking background noise. 00:10:48.379 --> 00:10:51.379 Exactly. Now, earlier versions of medical AI 00:10:51.379 --> 00:10:54.940 often only handled... simple static inputs like 00:10:54.940 --> 00:10:58.279 an x -ray or a photo of a skin condition. MedGemma 00:10:58.279 --> 00:11:02.259 1 .5 capabilities are just, they're huge, a major 00:11:02.259 --> 00:11:03.960 technical leap. It supports high -dimensional 00:11:03.960 --> 00:11:07.600 scans, CTs, MRIs, histopathology slides. These 00:11:07.600 --> 00:11:10.179 are really complex, data -rich images. Right. 00:11:10.220 --> 00:11:12.320 They require significant processing power and 00:11:12.320 --> 00:11:14.620 real domain expertise to read correctly. But 00:11:14.620 --> 00:11:16.659 the real technical leap based on the source material 00:11:16.659 --> 00:11:19.139 seems to be its ability to handle a time series 00:11:19.139 --> 00:11:21.360 of chest x -rays. What does that mean in practice? 00:11:21.620 --> 00:11:24.279 It means the AI maintains memory. It's not just 00:11:24.279 --> 00:11:26.519 analyzing one's static image. It's comparing 00:11:26.519 --> 00:11:28.740 a patient's current scan against a history of 00:11:28.740 --> 00:11:31.659 previous scans. So it can identify trends, compare 00:11:31.659 --> 00:11:33.899 changes in lung capacity, or see the subtle growth 00:11:33.899 --> 00:11:36.639 of a nodule over months or years. That context 00:11:36.639 --> 00:11:38.419 is absolutely vital for an accurate diagnosis. 00:11:38.779 --> 00:11:41.419 That is genuinely advanced. And beyond imaging, 00:11:41.659 --> 00:11:44.080 what else does it handle? It performs anatomical 00:11:44.080 --> 00:11:47.090 localization. So accurately finding specific 00:11:47.090 --> 00:11:50.490 organs in images. And crucially, it facilitates 00:11:50.490 --> 00:11:53.750 lab report parsing and EHR Q &A. So it helps 00:11:53.750 --> 00:11:56.730 the AI chew through these vast amounts of unstructured 00:11:56.730 --> 00:11:59.970 data, like messy PDF lab reports, and pull out 00:11:59.970 --> 00:12:02.649 structured, actionable data instantly. That's 00:12:02.649 --> 00:12:05.750 its most immediate real -world use case. Extracting 00:12:05.750 --> 00:12:07.909 structured data from mountains of unstructured 00:12:07.909 --> 00:12:10.190 PDFs that would otherwise take hours of human 00:12:10.190 --> 00:12:12.809 triage. It streamlines the whole administrative 00:12:12.809 --> 00:12:16.460 and diagnostic pathway. Whoa. Imagine scaling 00:12:16.460 --> 00:12:18.860 this model, which is free for commercial use, 00:12:19.000 --> 00:12:22.320 to analyze a billion medical queries. Providing 00:12:22.320 --> 00:12:24.519 specialized insight instantly everywhere from 00:12:24.519 --> 00:12:26.740 a major hospital to a rural clinic. That is a 00:12:26.740 --> 00:12:29.120 massive leap forward for accessibility and speed. 00:12:29.399 --> 00:12:31.320 And Google is actively encouraging this development. 00:12:31.500 --> 00:12:33.159 They've started the MedGemma Impact Challenge. 00:12:33.320 --> 00:12:36.100 It's a Kaggle competition with $100 ,000 in prizes. 00:12:36.440 --> 00:12:38.600 They even released full tutorials for developers 00:12:38.600 --> 00:12:41.580 to encourage remixing these new tools. So what 00:12:41.580 --> 00:12:44.259 is the most immediate real -world use case for 00:12:44.259 --> 00:12:47.970 MedGemma 1? I'd say extracting structured data 00:12:47.970 --> 00:12:51.090 from vast amounts of unstructured PDF lab reports 00:12:51.090 --> 00:12:54.110 quickly and accurately. That was a tremendous 00:12:54.110 --> 00:12:56.809 sweep of sources. We covered massive corporate 00:12:56.809 --> 00:12:59.590 spending, the messy world of creation, and the 00:12:59.590 --> 00:13:02.250 specialized frontier of medicine. To recap, I 00:13:02.250 --> 00:13:04.809 think we've extracted three major insights on 00:13:04.809 --> 00:13:07.049 where AI, energy, and capital are focusing right 00:13:07.049 --> 00:13:10.190 now. First, that massive corporate pivot is definitive. 00:13:10.840 --> 00:13:13.259 The metaverse is now secondary to AI, attracting 00:13:13.259 --> 00:13:15.759 billions in dedicated capital from giants like 00:13:15.759 --> 00:13:18.279 Meta. That money is focused on infrastructure 00:13:18.279 --> 00:13:21.220 and training the next generation of LLMs. Second, 00:13:21.399 --> 00:13:24.059 creator tools are maturing and fast. They're 00:13:24.059 --> 00:13:26.200 moving from general models that generate AI slop 00:13:26.200 --> 00:13:29.059 to highly specialized measurable ranking systems. 00:13:29.220 --> 00:13:31.440 That ensures quality control and consistency 00:13:31.440 --> 00:13:34.539 for the average user. And third, agentic AI is 00:13:34.539 --> 00:13:36.899 moving rapidly into enterprise and critical fields 00:13:36.899 --> 00:13:39.649 like healthcare. The idea of an AI that takes 00:13:39.649 --> 00:13:41.929 initiative exemplified by the specialized open 00:13:41.929 --> 00:13:45.090 source power of Med Gemma 1 .5 shows that specialized 00:13:45.090 --> 00:13:46.990 application is where the real value is being 00:13:46.990 --> 00:13:49.509 found today. Knowledge is most valuable when 00:13:49.509 --> 00:13:52.370 it's understood and applied. And all these changes 00:13:52.370 --> 00:13:54.929 show that specialized application is where the 00:13:54.929 --> 00:13:57.830 market is placing its bets now. Whether that's 00:13:57.830 --> 00:14:00.690 extracting complex medical data from a PDF or 00:14:00.690 --> 00:14:03.789 reverse engineering a viral YouTube hook. So 00:14:03.789 --> 00:14:06.789 if AI can now reliably parse complex medical 00:14:06.789 --> 00:14:09.769 scans and massive amounts of unstructured lab 00:14:09.769 --> 00:14:12.330 reports and perform memory -based trend analysis 00:14:12.330 --> 00:14:15.289 over time, what previously impossible corporate 00:14:15.289 --> 00:14:17.669 task is just waiting for the right specialized 00:14:17.669 --> 00:14:20.129 open source model. That's something to mull over 00:14:20.129 --> 00:14:21.950 as you look at your own workflow. Thank you for 00:14:21.950 --> 00:14:23.870 sharing your sources and participating in this 00:14:23.870 --> 00:14:25.570 deep dive with us. We'll catch you next time.