WEBVTT 00:00:00.000 --> 00:00:03.740 Probably the biggest plot twist in AI this year, 00:00:03.899 --> 00:00:07.099 Apple. You know, the company famous for total 00:00:07.099 --> 00:00:10.400 control is paying its biggest rival, Google, 00:00:10.619 --> 00:00:13.240 a billion dollars. Yeah. It's really the ultimate 00:00:13.240 --> 00:00:15.320 sign, isn't it? That trying to keep everything 00:00:15.320 --> 00:00:18.379 proprietary, building your own intelligence completely 00:00:18.379 --> 00:00:20.879 in -house, it's just too slow right now. The 00:00:20.879 --> 00:00:24.500 pace is incredible. Welcome to the Deep Dive. 00:00:24.500 --> 00:00:26.600 We're unpacking a really fascinating stack of 00:00:26.600 --> 00:00:28.879 sources this week. We're looking at who is truly 00:00:28.879 --> 00:00:32.000 building, or maybe more accurately, who is buying 00:00:32.000 --> 00:00:34.670 AI power right now. Exactly. We're going to dive 00:00:34.670 --> 00:00:37.710 into the, frankly, shocking details of that Apple 00:00:37.710 --> 00:00:39.869 -Gemini partnership. We'll also look at some 00:00:39.869 --> 00:00:43.090 dramatic breakups and bailouts shaking up the 00:00:43.090 --> 00:00:46.030 big AI giants. And then we'll examine a pretty 00:00:46.030 --> 00:00:48.710 crucial technical paper that's basically democratizing 00:00:48.710 --> 00:00:51.590 access to those huge trillion parameter models, 00:00:51.770 --> 00:00:53.829 making them more accessible. Okay, right. So 00:00:53.829 --> 00:00:56.109 let's start with that core conflict. Apple, you 00:00:56.109 --> 00:00:57.729 know, the champion of the closed garden, the 00:00:57.729 --> 00:01:00.509 controlled ecosystem, they're licensing Google's 00:01:00.509 --> 00:01:02.670 Gemini model for their main Apple intelligence 00:01:02.670 --> 00:01:05.939 stuff. And that price tag, a reported $1 billion 00:01:05.939 --> 00:01:09.760 a year. It just shows the scale of this compromise, 00:01:09.980 --> 00:01:11.879 doesn't it? It's definitely more than just irony. 00:01:12.060 --> 00:01:14.680 It feels like a temporary necessity. I mean, 00:01:14.719 --> 00:01:17.400 our sources point out what a lot of users already 00:01:17.400 --> 00:01:21.739 feel. Siri has been, and this is a quote, embarrassingly 00:01:21.739 --> 00:01:24.980 bad for years, especially when you compare it 00:01:24.980 --> 00:01:27.719 to modern AIs. Right. Apple is racing to build 00:01:27.719 --> 00:01:30.099 its own massive model, like a one trillion parameter 00:01:30.099 --> 00:01:33.530 one. But the timeline seems to be. Well, maybe 00:01:33.530 --> 00:01:36.489 late 2026 at the earliest. And they just couldn't 00:01:36.489 --> 00:01:39.049 wait. So we should clarify, Gemini isn't handling 00:01:39.049 --> 00:01:41.730 the actual voice part of Siri. No, exactly. It's 00:01:41.730 --> 00:01:43.329 powering the really heavy lifting underneath, 00:01:43.569 --> 00:01:46.370 the complex stuff, specifically the summarizer 00:01:46.370 --> 00:01:48.709 tool and those sophisticated planner features 00:01:48.709 --> 00:01:51.030 in the new OS updates. They needed that advanced 00:01:51.030 --> 00:01:53.310 reasoning like yesterday. Because their internal 00:01:53.310 --> 00:01:56.010 models weren't cutting it. Apparently not. That 00:01:56.010 --> 00:01:59.189 need became pretty urgent. because their internal 00:01:59.189 --> 00:02:01.730 efforts were struggling. We heard Apple recently 00:02:01.730 --> 00:02:04.670 lost, what, four to seven key researchers, people 00:02:04.670 --> 00:02:07.769 working on those foundational models. Precisely 00:02:07.769 --> 00:02:10.349 because the in -house solutions weren't performing 00:02:10.349 --> 00:02:12.750 well enough. So they kind of had to go shopping. 00:02:12.909 --> 00:02:15.129 And what's really fascinating here, especially 00:02:15.129 --> 00:02:17.490 given Apple's reputation, is how they've set 00:02:17.490 --> 00:02:20.009 up the security. Google gets the billion dollars, 00:02:20.250 --> 00:02:23.689 sure, but the AI itself runs entirely on Apple's 00:02:23.689 --> 00:02:25.870 own system, the private cloud compute architecture. 00:02:26.069 --> 00:02:29.129 Right. which means Google doesn't touch the user 00:02:29.129 --> 00:02:32.270 data. That lets Apple keep that critical control 00:02:32.270 --> 00:02:35.669 layer they value so much, even if the core brain 00:02:35.669 --> 00:02:39.509 is, well, outsourced. Beat. And it wasn't like 00:02:39.509 --> 00:02:41.830 Google was the only option they looked at. No, 00:02:41.870 --> 00:02:43.750 definitely not. We know they actively tested 00:02:43.750 --> 00:02:46.569 competitors. OpenAI's models were in the mix. 00:02:46.669 --> 00:02:50.210 Anthropix clawed. But Gemini won out. Why? Seems 00:02:50.210 --> 00:02:52.569 like it demonstrated superior instruction following. 00:02:52.919 --> 00:02:55.840 so it handles multi -step commands better, and 00:02:55.840 --> 00:02:58.180 it apparently has a longer context memory, which 00:02:58.180 --> 00:03:00.340 is super important for planning complex tasks 00:03:00.340 --> 00:03:03.680 or summarizing long documents. OK, so it was 00:03:03.680 --> 00:03:05.919 really about the technical fit for the specific 00:03:05.919 --> 00:03:08.539 job. Makes sense. Yeah. But the complexity just 00:03:08.539 --> 00:03:10.500 keeps growing, doesn't it? Because Google is 00:03:10.500 --> 00:03:13.319 banned in mainland China. Apple's now having 00:03:13.319 --> 00:03:16.479 to cut separate deals, local deals with giants 00:03:16.479 --> 00:03:19.580 like Alibaba and Baidu just for the version of 00:03:19.580 --> 00:03:21.919 Siri that runs there. Which creates this sort 00:03:21.919 --> 00:03:23.960 of fragmented intelligence stack, doesn't it? 00:03:24.000 --> 00:03:25.860 Yeah. Different brains, depending on where you 00:03:25.860 --> 00:03:27.699 are. We're thinking about what this means for 00:03:27.699 --> 00:03:29.780 Apple long term. OK, let me just ask directly 00:03:29.780 --> 00:03:33.840 then. Beyond the huge price tag, what specific 00:03:33.840 --> 00:03:38.539 technical capability really made Gemini the unavoidable 00:03:38.539 --> 00:03:40.840 winner for Apple's needs, especially for those 00:03:40.840 --> 00:03:42.840 planning tools? It showed better instruction 00:03:42.840 --> 00:03:45.620 following and could maintain context over longer 00:03:45.620 --> 00:03:48.780 interactions. Simple as that. Okay, now shifting 00:03:48.780 --> 00:03:51.400 gears a bit, let's talk about the corporate tectonic 00:03:51.400 --> 00:03:54.500 plates. Because we're seeing some major realignments 00:03:54.500 --> 00:03:56.939 happening. There's this big shift, maybe even 00:03:56.939 --> 00:03:59.879 a split, happening between Microsoft and OpenAI. 00:04:00.530 --> 00:04:02.810 Microsoft is now officially breaking away, setting 00:04:02.810 --> 00:04:05.110 up its own separate superintelligence team. Their 00:04:05.110 --> 00:04:08.449 goal, build AGI independently. Wow. So they're 00:04:08.449 --> 00:04:11.310 still partners, but strategically. Strategically, 00:04:11.310 --> 00:04:13.729 the priority seems to be shifting towards Microsoft 00:04:13.729 --> 00:04:17.089 having its own proprietary AGI development effort. 00:04:17.290 --> 00:04:21.370 Still working with OpenAI, but... Also hedging 00:04:21.370 --> 00:04:23.410 their bets, maybe building their own thing. And 00:04:23.410 --> 00:04:25.709 this probably all ties back to the sheer cost, 00:04:25.810 --> 00:04:28.490 right? We saw that little drama where OpenAI's 00:04:28.490 --> 00:04:31.129 CFO hinted they might need a, what was it, a 00:04:31.129 --> 00:04:34.129 $1 .4 trillion chip bailout? Yeah. A number so 00:04:34.129 --> 00:04:36.569 big it sounds like a typo. Sam Altman immediately 00:04:36.569 --> 00:04:39.569 denied it, of course. Right. But whether that 00:04:39.569 --> 00:04:41.310 specific number is real or not, it definitely 00:04:41.310 --> 00:04:43.829 highlights the absolutely staggering amount of 00:04:43.829 --> 00:04:47.810 capital needed to seriously chase AGI. Like nation 00:04:47.810 --> 00:04:50.810 state level spending. Only a tiny handful of 00:04:50.810 --> 00:04:52.670 entities on the planet can even think about playing 00:04:52.670 --> 00:04:55.209 at that level. It's mind boggling. But, you know, 00:04:55.209 --> 00:04:56.589 looking at the other side of that investment 00:04:56.589 --> 00:04:58.449 coin, it's not all just about buying more chips. 00:04:58.790 --> 00:05:01.350 The OpenAI Foundation is also investing heavily, 00:05:01.410 --> 00:05:04.610 about $25 billion into things like health care 00:05:04.610 --> 00:05:07.810 applications and AI resilience. And they're partnering 00:05:07.810 --> 00:05:10.050 strongly with Microsoft and SoftBank on that. 00:05:10.110 --> 00:05:12.009 So it seems like a kind of two pronged strategy. 00:05:12.490 --> 00:05:15.709 Chase AGI, you know, at almost any cost, but 00:05:15.709 --> 00:05:17.470 also fund these critical sector applications. 00:05:17.810 --> 00:05:20.769 Meanwhile. We also saw a pretty vulnerable admission 00:05:20.769 --> 00:05:24.370 that really illustrates the maybe the ethical 00:05:24.370 --> 00:05:27.149 debt the scaling race can create. You're talking 00:05:27.149 --> 00:05:29.689 about Meta. Yeah. Sources revealed that Meta, 00:05:29.689 --> 00:05:32.930 to help fund its massive AI expansion, was bankrolled 00:05:32.930 --> 00:05:36.569 by roughly $16 billion. And part of that came 00:05:36.569 --> 00:05:39.949 from deliberately allowing scam ads, ads targeting 00:05:39.949 --> 00:05:43.050 users daily. It's pretty stark. They apparently. 00:05:44.120 --> 00:05:46.279 tolerated these profitable scams because, well, 00:05:46.300 --> 00:05:48.019 they generated too much revenue to just shut 00:05:48.019 --> 00:05:50.379 them down easily. Ouch. And that pursuit of short 00:05:50.379 --> 00:05:52.180 -term profit, it creates a long -term problem, 00:05:52.319 --> 00:05:54.779 doesn't it? If your revenue relies on bad actors, 00:05:54.959 --> 00:05:57.399 it inevitably poisons the data you're using to 00:05:57.399 --> 00:05:59.439 train your models. And it just fundamentally 00:05:59.439 --> 00:06:01.620 damages user trust. Yeah, that tension between 00:06:01.620 --> 00:06:05.850 integrity and just sheer scale. It's not just 00:06:05.850 --> 00:06:07.649 social media either. Look at Google. They've 00:06:07.649 --> 00:06:10.490 aggressively pushed into becoming what sources 00:06:10.490 --> 00:06:13.930 call a full -on financial AI researcher. They're 00:06:13.930 --> 00:06:16.129 rolling out features, answering complex market 00:06:16.129 --> 00:06:18.970 questions, providing live earnings transcripts 00:06:18.970 --> 00:06:21.029 for traders in real time. They're pushing into 00:06:21.029 --> 00:06:23.370 every potentially profitable sector they can 00:06:23.370 --> 00:06:26.670 find with AI. It's an aggressive expansion. So 00:06:26.670 --> 00:06:28.430 thinking about that meta -admission specifically. 00:06:29.629 --> 00:06:32.170 What does that pursuit of high revenue, even 00:06:32.170 --> 00:06:34.350 when it comes from bad actors, tell us about 00:06:34.350 --> 00:06:37.449 the maybe the foundational integrity of the AI 00:06:37.449 --> 00:06:40.790 models being built on that data? High profit 00:06:40.790 --> 00:06:43.930 from bad actors risks data quality and user trust, 00:06:44.069 --> 00:06:46.819 creating ethical debt. OK. Let's dive into the 00:06:46.819 --> 00:06:48.560 tech side now, because this is where it gets 00:06:48.560 --> 00:06:50.540 really interesting for me. We need to talk about 00:06:50.540 --> 00:06:52.579 the physical barriers, the hardware barriers 00:06:52.579 --> 00:06:55.279 that have mostly kept these super advanced AI 00:06:55.279 --> 00:06:57.339 models in the hands of just a few giants like 00:06:57.339 --> 00:06:59.839 Google or Nvidia. Right. We're talking about 00:06:59.839 --> 00:07:02.629 the huge models, the ones with like. a trillion 00:07:02.629 --> 00:07:05.910 parameters they usually need specialized incredibly 00:07:05.910 --> 00:07:09.189 expensive hardware setups exactly and critically 00:07:09.189 --> 00:07:13.350 most ai teams have kind of avoided using standard 00:07:13.350 --> 00:07:16.170 cloud infrastructure like aws for these really 00:07:16.170 --> 00:07:18.709 massive models and there's a specific reason 00:07:18.709 --> 00:07:22.490 why which is aws's networking tech called efa 00:07:22.490 --> 00:07:25.329 lacks a key feature something called gpu direct 00:07:25.329 --> 00:07:28.069 async okay hold on gpu direct async we need to 00:07:28.069 --> 00:07:30.269 break that down what does that missing feature 00:07:30.269 --> 00:07:37.790 actually mean in Okay, think of it like this. 00:07:37.829 --> 00:07:40.430 Imagine your giant AI model is spread across, 00:07:40.529 --> 00:07:42.870 say, 10 different computers, each packed with 00:07:42.870 --> 00:07:45.829 GPUs. Got it. Those 10 machines need to talk 00:07:45.829 --> 00:07:47.709 to each other constantly, sharing data back and 00:07:47.709 --> 00:07:49.699 forth, like... instantaneously for the model 00:07:49.699 --> 00:07:53.560 to work. GPU direct async lets the GPUs on different 00:07:53.560 --> 00:07:55.620 machines talk directly to each other super fast. 00:07:55.759 --> 00:07:57.860 It cuts out the middleman, which is usually the 00:07:57.860 --> 00:08:00.620 main computer brain, the CPU. Without it, the 00:08:00.620 --> 00:08:02.920 GPUs have to kind of wait for the CPU on each 00:08:02.920 --> 00:08:05.639 machine to manage all that data traffic. It creates 00:08:05.639 --> 00:08:08.300 bottlenecks. The communication slows down so 00:08:08.300 --> 00:08:11.660 much that these giant complex models often just... 00:08:11.930 --> 00:08:14.370 crash or fail. Yeah. That's been the big hurdle 00:08:14.370 --> 00:08:17.370 on standard cloud setups like AWS. Right. The 00:08:17.370 --> 00:08:19.290 communication highway wasn't fast enough between 00:08:19.290 --> 00:08:23.069 the different GPU workers. Exactly. But here's 00:08:23.069 --> 00:08:25.550 the breakthrough. It comes from perplexity. They 00:08:25.550 --> 00:08:27.189 published some really groundbreaking research 00:08:27.189 --> 00:08:29.730 showing how you can actually run these trillion 00:08:29.730 --> 00:08:32.789 parameter Mogi models, models like Kimi K2 and 00:08:32.789 --> 00:08:36.590 DeepSeek V3 on regular off -the -shelf AWS cloud 00:08:36.590 --> 00:08:39.330 machines. How? If the hardware feature is missing. 00:08:39.799 --> 00:08:42.000 They basically found a clever software workaround. 00:08:42.120 --> 00:08:44.879 Instead of needing new expensive hardware, they 00:08:44.879 --> 00:08:47.419 built an intelligent system using software. The 00:08:47.419 --> 00:08:49.740 CPU still helps coordinate, but the key is how 00:08:49.740 --> 00:08:51.960 they move the data. Okay. They pack and shuttle 00:08:51.960 --> 00:08:54.179 the data really smartly using a different technology 00:08:54.179 --> 00:08:57.639 called RDMA. RDMA. Remote Direct Memory Access. 00:08:57.720 --> 00:08:59.539 That's basically a way for computers to swap 00:08:59.539 --> 00:09:01.700 data directly between their memories super fast, 00:09:01.759 --> 00:09:04.990 right? Precisely. It's like... Imagine stacking 00:09:04.990 --> 00:09:07.970 Lego blocks of data between the machines incredibly 00:09:07.970 --> 00:09:10.549 fast and doing lots of stacks at the same time 00:09:10.549 --> 00:09:13.149 concurrently. They figured out how to orchestrate 00:09:13.149 --> 00:09:15.960 this data flow efficiently. using software and 00:09:15.960 --> 00:09:19.759 rdma even without that specific gpu direct async 00:09:19.759 --> 00:09:22.120 hardware feature so they built a software pipeline 00:09:22.120 --> 00:09:24.019 yeah essentially a high -speed software pipeline 00:09:24.019 --> 00:09:26.100 that makes the standard cloud function almost 00:09:26.100 --> 00:09:28.220 like one of those specialized super expensive 00:09:28.220 --> 00:09:30.779 supercomputers at least for this kind of workload 00:09:30.779 --> 00:09:33.419 whoa okay just pause there for a second imagine 00:09:33.419 --> 00:09:37.120 scaling your ai to handle like a billion queries 00:09:37.120 --> 00:09:40.919 without needing custom -built proprietary hardware 00:09:40.919 --> 00:09:43.809 that costs millions and millions This is like 00:09:43.809 --> 00:09:46.610 figuring out how to run a Formula One car and 00:09:46.610 --> 00:09:49.570 run it well on regular city streets using standard 00:09:49.570 --> 00:09:52.330 infrastructure. That feels really profound, a 00:09:52.330 --> 00:09:54.769 huge shift in accessibility. It absolutely is. 00:09:54.870 --> 00:09:57.950 That's the core impact, democratization. Suddenly, 00:09:57.950 --> 00:10:01.009 any development team, any startup with the decent 00:10:01.009 --> 00:10:05.179 cloud budget and the necessary brainpower. they 00:10:05.179 --> 00:10:07.019 can potentially join the Trillium Parameter League. 00:10:07.120 --> 00:10:09.740 It kind of dissolves that hardware barrier. It 00:10:09.740 --> 00:10:12.120 proves that really smart software engineering, 00:10:12.299 --> 00:10:15.639 efficient code, can actually beat massive capital 00:10:15.639 --> 00:10:18.190 spending. at least in some cases. So for the 00:10:18.190 --> 00:10:20.370 developers, for the companies out there listening, 00:10:20.629 --> 00:10:24.409 what's the key technical lesson from Perplexity's 00:10:24.409 --> 00:10:26.809 work here about getting around these hardware 00:10:26.809 --> 00:10:29.809 constraints? Smart software orchestration can 00:10:29.809 --> 00:10:33.409 often bypass assumed limits of commodity cloud 00:10:33.409 --> 00:10:36.269 hardware. Clever software beats brute force hardware. 00:10:36.509 --> 00:10:39.259 Love it. Okay, just a quick detour to highlight 00:10:39.259 --> 00:10:41.080 a couple of interesting new tools we're seeing 00:10:41.080 --> 00:10:43.860 pop up. There's one called Maya One, which is 00:10:43.860 --> 00:10:46.360 generating highly expressive speech. We're talking 00:10:46.360 --> 00:10:49.200 over 20 different emotions. So moving way beyond 00:10:49.200 --> 00:10:52.279 that typical kind of flat robotic AI voice, much 00:10:52.279 --> 00:10:53.899 more natural sound. That's cool. And the other 00:10:53.899 --> 00:10:57.139 one? Llama .cpp. This is really interesting for 00:10:57.139 --> 00:10:59.879 developers and tinkerers. Okay, define Llama 00:10:59.879 --> 00:11:03.039 .cpp for us simply. It's basically a lean framework 00:11:03.039 --> 00:11:06.720 letting you run large open source AI models directly 00:11:06.720 --> 00:11:09.539 on your own computer, even a decent laptop sometimes. 00:11:10.159 --> 00:11:12.879 And the significance of that, you know, it really 00:11:12.879 --> 00:11:14.879 shouldn't be understated. It means you don't 00:11:14.879 --> 00:11:17.340 have to send your sensitive data off to Google 00:11:17.340 --> 00:11:21.679 or OpenAI or whoever to get really good AI performance. 00:11:21.919 --> 00:11:23.820 Right. It ties back to that control theme we 00:11:23.820 --> 00:11:26.990 started with. Apple's struggle. Exactly. It potentially 00:11:26.990 --> 00:11:29.470 gives sovereignty back to the user or the small 00:11:29.470 --> 00:11:31.909 developer. You can run powerful models locally. 00:11:32.129 --> 00:11:34.289 And sticking with actually using these models, 00:11:34.529 --> 00:11:37.409 what about advanced prompting? Any new tricks? 00:11:37.929 --> 00:11:40.409 Well, sources continue to highlight the effectiveness 00:11:40.409 --> 00:11:43.129 of persona prompting. You know, telling the model 00:11:43.129 --> 00:11:46.860 to ultra -think like Steve Jobs. or act as an 00:11:46.860 --> 00:11:49.500 expert physicist explaining quantum entanglement. 00:11:49.580 --> 00:11:52.000 Does that really work? It tends to get much sharper, 00:11:52.019 --> 00:11:54.200 more focused answers from models like Claude 00:11:54.200 --> 00:11:56.840 or GPT -4. It works, I think, because you're 00:11:56.840 --> 00:11:59.259 essentially forcing the model into a specific 00:11:59.259 --> 00:12:01.919 defined set of constraints, a particular style 00:12:01.919 --> 00:12:04.080 or knowledge base, rather than letting it give 00:12:04.080 --> 00:12:07.379 a generic average response. Yeah, I have to admit, 00:12:07.440 --> 00:12:09.480 I still wrestle with prompt drift myself sometimes. 00:12:10.440 --> 00:12:13.399 Getting a model to consistently sound like a 00:12:13.399 --> 00:12:16.740 specific leader or maintain a complex persona 00:12:16.740 --> 00:12:21.019 over a long conversation, it takes constant tweaking 00:12:21.019 --> 00:12:23.860 and refinement. It's never quite as simple as 00:12:23.860 --> 00:12:26.720 those quick online guides make it seem. It's 00:12:26.720 --> 00:12:29.340 definitely an iterative process. Lots of back 00:12:29.340 --> 00:12:31.580 and forth. For sure. It's an art as much as a 00:12:31.580 --> 00:12:34.120 science right now. Okay. So if we pull way back 00:12:34.120 --> 00:12:35.779 now, look at everything I've discussed. Yeah. 00:12:35.879 --> 00:12:38.399 The big idea emerging today seems pretty clear. 00:12:39.059 --> 00:12:42.009 Control. That old model of total proprietary 00:12:42.009 --> 00:12:45.330 control is rapidly becoming obsolete, or at least 00:12:45.330 --> 00:12:48.690 incredibly difficult to maintain. You have Apple, 00:12:48.850 --> 00:12:52.190 the absolute champion of proprietary power, forced 00:12:52.190 --> 00:12:54.549 to essentially rent its core intelligence from 00:12:54.549 --> 00:12:57.269 its biggest rival, Google. And at the exact same 00:12:57.269 --> 00:13:01.070 time, you have this grassroots technical brilliance, 00:13:01.070 --> 00:13:03.590 like what Perplexity demonstrated, that's dissolving 00:13:03.590 --> 00:13:05.750 the hardware barriers that used to enforce that 00:13:05.750 --> 00:13:08.570 control. It sends a powerful message, doesn't 00:13:08.570 --> 00:13:11.409 it? That really efficient code, smart software 00:13:11.409 --> 00:13:14.129 can fundamentally challenge the financial might 00:13:14.129 --> 00:13:15.950 of the world's biggest tech companies when it 00:13:15.950 --> 00:13:18.990 comes to AI development. The material we explore 00:13:18.990 --> 00:13:20.950 today really shows that the whole balance of 00:13:20.950 --> 00:13:24.389 power in this AI race. It's unstable. It's constantly 00:13:24.389 --> 00:13:27.490 shifting based on a new partnership deal one 00:13:27.490 --> 00:13:30.289 week or a new algorithmic breakthrough the next. 00:13:30.490 --> 00:13:32.750 Which leads to maybe a provocative thought to 00:13:32.750 --> 00:13:35.649 leave folks with. If the world's most secretive, 00:13:35.649 --> 00:13:38.590 most control -obsessed company, Apple, finds 00:13:38.590 --> 00:13:41.909 itself forced to outsource core AI intelligence, 00:13:42.370 --> 00:13:45.809 how much longer can any major company realistically 00:13:45.809 --> 00:13:49.610 maintain true proprietary control over the most 00:13:49.610 --> 00:13:53.320 advanced forms of AGI when they emerge? Is that 00:13:53.320 --> 00:13:55.980 even the right model anymore? That's a great 00:13:55.980 --> 00:13:57.899 question to ponder. And for you, the learner 00:13:57.899 --> 00:13:59.659 listening today, if you want to dig deeper into 00:13:59.659 --> 00:14:01.720 the tech enabling some of this shift, we really 00:14:01.720 --> 00:14:03.820 encourage you to look into the concept of RDMA. 00:14:03.879 --> 00:14:06.259 That's remote direct memory access. It's kind 00:14:06.259 --> 00:14:08.759 of the invisible engine fueling this new era 00:14:08.759 --> 00:14:11.580 of cloud efficiency for AI and understanding 00:14:11.580 --> 00:14:13.580 it might be key to understanding where a large 00:14:13.580 --> 00:14:16.120 scale AI is heading next. Thank you for joining 00:14:16.120 --> 00:14:18.399 us on this deep dive. Outro music.