WEBVTT 00:00:00.000 --> 00:00:03.500 When these massive AI models are running, whether 00:00:03.500 --> 00:00:05.540 you're training the next big thing or just serving 00:00:05.540 --> 00:00:08.960 billions of queries a day, there's this really 00:00:08.960 --> 00:00:11.220 simple question that every company has to ask 00:00:11.220 --> 00:00:15.140 itself. Are we trying to be the fastest? Or are 00:00:15.140 --> 00:00:17.320 we just trying to be the cheapest? Because right 00:00:17.320 --> 00:00:20.839 now, Google is... very clearly choosing to be 00:00:20.839 --> 00:00:23.859 the cheapest. And that whole strategy is their 00:00:23.859 --> 00:00:26.539 direct shot at Nvidia's dominance. And welcome 00:00:26.539 --> 00:00:29.320 to the deep dive. Today, we're digging into the 00:00:29.320 --> 00:00:31.600 latest intelligence on these new AI infrastructure 00:00:31.600 --> 00:00:34.810 wars. And it's... It's a really fascinating fight 00:00:34.810 --> 00:00:36.549 because the goal isn't just about winning on 00:00:36.549 --> 00:00:38.590 performance benchmarks anymore. It's about making 00:00:38.590 --> 00:00:41.310 AI compute so cheap that it just becomes the 00:00:41.310 --> 00:00:43.890 default. Exactly. So first, we're going to unpack 00:00:43.890 --> 00:00:46.270 that whole cost per token battle. It's Google's 00:00:46.270 --> 00:00:49.549 very specialized TPUs versus NVIDIA's GPUs that 00:00:49.549 --> 00:00:51.990 are everywhere. It's a fight for the very foundation 00:00:51.990 --> 00:00:54.390 of AI. Yeah, and after we break down that chip 00:00:54.390 --> 00:00:56.329 war, we've got a really fast -paced segment, 00:00:56.390 --> 00:00:58.270 some critical security alerts, a look at some 00:00:58.270 --> 00:01:00.729 surprising new consumer tech. I'm talking translation 00:01:00.729 --> 00:01:03.450 glasses. And we'll explore some new interactive 00:01:03.450 --> 00:01:05.769 learning tools that are coming online. And then 00:01:05.769 --> 00:01:08.049 finally, we're jumping into a huge medical breakthrough. 00:01:08.450 --> 00:01:10.969 This one is out of Harvard Medical School, a 00:01:10.969 --> 00:01:15.950 new AI tool called PopEVE. And this system is 00:01:15.950 --> 00:01:19.189 solving these really complex genetic mysteries 00:01:19.189 --> 00:01:21.849 that, you know, even the big one, Alpha Missense, 00:01:22.010 --> 00:01:24.189 couldn't quite get right. We'll look at the data, 00:01:24.250 --> 00:01:26.430 and it's pretty clear this is a huge step forward. 00:01:26.769 --> 00:01:28.790 So let's start right there with the hardware. 00:01:29.349 --> 00:01:31.310 Looking at Google versus NVIDIA, I mean, the 00:01:31.310 --> 00:01:33.689 sources we saw make it really clear. Google has 00:01:33.689 --> 00:01:36.629 completely given up on winning the benchmark 00:01:36.629 --> 00:01:38.430 race. They're going all in on cost supremacy. 00:01:38.930 --> 00:01:41.170 That's the pivot. That is the entire strategy. 00:01:41.840 --> 00:01:44.500 Google's goal is to make AI compute so cheap, 00:01:44.500 --> 00:01:46.959 so accessible, that other companies almost have 00:01:46.959 --> 00:01:49.640 no choice but to use Google Cloud. They want 00:01:49.640 --> 00:01:51.739 to win on price. And the key to their advantage, 00:01:51.980 --> 00:01:54.340 they own the whole stack. Right, the entire vertical 00:01:54.340 --> 00:01:56.859 stack, from the chip design itself, the TPUs, 00:01:56.859 --> 00:01:59.019 all the way to the data centers they run in, 00:01:59.079 --> 00:02:00.780 and then the software that sits on top of all 00:02:00.780 --> 00:02:02.659 of it. And that control just gives them this 00:02:02.659 --> 00:02:05.530 incredible power over pricing. Precisely. I mean, 00:02:05.530 --> 00:02:07.769 if you look at Nvidia's market right now, a huge 00:02:07.769 --> 00:02:09.469 chunk of their revenue comes from the margins 00:02:09.469 --> 00:02:12.090 on their hardware. We hear estimates of, what, 00:02:12.129 --> 00:02:15.310 70 % markups on those high -end GPUs? Meanwhile, 00:02:15.550 --> 00:02:18.550 Google can basically sell its own custom TPUs 00:02:18.550 --> 00:02:20.830 through Google Cloud at cost, or maybe even below 00:02:20.830 --> 00:02:22.930 cost. And they can do that because they just 00:02:22.930 --> 00:02:24.710 make the money back on all the other services 00:02:24.710 --> 00:02:27.210 that are tied into that cloud ecosystem, or just 00:02:27.210 --> 00:02:29.949 on pure volume. And this is a total game -changer 00:02:29.949 --> 00:02:31.710 when you look at where the real spending is happening 00:02:31.710 --> 00:02:34.659 now. Absolutely. The market has shifted so dramatically, 00:02:34.819 --> 00:02:38.300 90%, maybe even more, of all the AI spending 00:02:38.300 --> 00:02:41.439 today. It's not for training the models. That's 00:02:41.439 --> 00:02:44.020 a one -time cost, basically. Most of the money 00:02:44.020 --> 00:02:46.460 is for inference. It's running the model every 00:02:46.460 --> 00:02:49.400 single time a user asks a question. So the main 00:02:49.400 --> 00:02:51.479 thing that these big cloud customers care about 00:02:51.479 --> 00:02:54.659 is no longer raw speed. It's not FLOPS, floating 00:02:54.659 --> 00:02:56.900 point operations per second, which is what training 00:02:56.900 --> 00:02:59.530 always focused on. Correct. The metric that truly 00:02:59.530 --> 00:03:02.449 matters now is the lowest possible cost per token, 00:03:02.590 --> 00:03:05.650 but at a massive scale. And Google has designed 00:03:05.650 --> 00:03:09.169 the TPU specifically for that one job. Drive 00:03:09.169 --> 00:03:12.150 the cost of inference down as low as it can possibly 00:03:12.150 --> 00:03:15.530 go. And this cost advantage is, well, it's causing 00:03:15.530 --> 00:03:18.050 real anxiety for NVIDIA. The buzz we're hearing 00:03:18.050 --> 00:03:21.729 is that huge players, think meta, anthropic. 00:03:22.509 --> 00:03:25.389 they might shift billions of dollars in compute 00:03:25.389 --> 00:03:28.250 spend over to Google's TPUs. Even if they only 00:03:28.250 --> 00:03:30.449 move a little bit of that, it could shave, what, 00:03:30.509 --> 00:03:33.389 10 % off NVIDIA's AI revenue. That's an earthquake. 00:03:33.669 --> 00:03:35.969 It is. Now, NVIDIA's defense is strong, but it's 00:03:35.969 --> 00:03:38.030 entirely about the platform. They immediately 00:03:38.030 --> 00:03:40.870 come out and say TPUs are locked in, narrow purpose, 00:03:41.030 --> 00:03:43.430 inflexible. And what's so fascinating is how 00:03:43.430 --> 00:03:46.449 they use CUA, their software ecosystem, as their 00:03:46.449 --> 00:03:49.210 shield. CUA is the bedrock of their whole argument. 00:03:49.370 --> 00:03:51.849 It is. The combination of their GPUs and CUDA. 00:03:52.039 --> 00:03:54.039 Well, it works with almost any model. It handles 00:03:54.039 --> 00:03:56.280 training and inference, no problem. And it runs 00:03:56.280 --> 00:03:59.280 everywhere, any cloud or even on -prem. So NVIDIA 00:03:59.280 --> 00:04:01.500 is fighting a cost war with the platform argument. 00:04:01.620 --> 00:04:03.379 They're basically saying, we're the default. 00:04:03.479 --> 00:04:05.280 We're the flexible ecosystem you can bet your 00:04:05.280 --> 00:04:08.039 whole company on. So if that lock -in argument 00:04:08.039 --> 00:04:12.199 is so powerful, you know, the fear of getting 00:04:12.199 --> 00:04:15.580 stuck with one vendor like Google, what's stopping 00:04:15.580 --> 00:04:17.980 these big players, the metas and anthropics, 00:04:18.040 --> 00:04:21.410 from... Just sticking with NVIDIA's flexibility 00:04:21.410 --> 00:04:23.790 for the long term. The massive cost savings. 00:04:23.850 --> 00:04:26.529 It just incentivizes adopting specialized hardware, 00:04:26.610 --> 00:04:29.029 even with those lock -in fears. Dollar stock. 00:04:29.329 --> 00:04:31.310 Especially when you're dealing with billions 00:04:31.310 --> 00:04:33.430 of user queries every day. It's a tough trade 00:04:33.430 --> 00:04:35.889 -off. Okay, switching gears. Let's hit some rapid 00:04:35.889 --> 00:04:38.230 -fire updates from the world of AI. Starting 00:04:38.230 --> 00:04:40.629 with security and trust, which this just keeps 00:04:40.629 --> 00:04:42.490 coming up in all the sources we see. Yeah, we 00:04:42.490 --> 00:04:44.769 saw a pretty big security alert about chat GPT 00:04:44.769 --> 00:04:48.379 user data. And it wasn't OpenAI's main platform 00:04:48.379 --> 00:04:50.740 that got breached. It was a third -party partner 00:04:50.740 --> 00:04:53.279 with a sloppy configuration that exposed user 00:04:53.279 --> 00:04:56.399 data, emails, things like that. And this just 00:04:56.399 --> 00:04:58.180 brings up such a critical point about the whole 00:04:58.180 --> 00:05:01.379 AI ecosystem. As we weave these tools deeper 00:05:01.379 --> 00:05:03.939 into our businesses, our security isn't just 00:05:03.939 --> 00:05:06.139 about the main company anymore. It's tied to 00:05:06.139 --> 00:05:08.019 the weakest link in the chain, the partners, 00:05:08.180 --> 00:05:10.560 the APIs, all the extensions they use. It's a 00:05:10.560 --> 00:05:13.829 distributed risk. Exactly. Even if you have perfect 00:05:13.829 --> 00:05:17.410 internal controls, that one API gateway on a 00:05:17.410 --> 00:05:20.329 partner system can be the point of failure. And 00:05:20.329 --> 00:05:22.050 honestly, it's something I still wrestle with, 00:05:22.089 --> 00:05:25.550 you know, trying to manage security across, what, 00:05:25.610 --> 00:05:28.089 a dozen different APIs and platforms. It's just... 00:05:28.509 --> 00:05:30.769 It's incredibly difficult to keep a consistent 00:05:30.769 --> 00:05:33.589 security posture when you rely on that many outside 00:05:33.589 --> 00:05:35.670 integrations. That's a really important point. 00:05:36.209 --> 00:05:38.810 Meanwhile, this whole idea of trust is being 00:05:38.810 --> 00:05:41.550 tested by the tech itself. We all saw those Thanksgiving 00:05:41.550 --> 00:05:43.629 photos of Elon Musk and Mark Zuckerberg that 00:05:43.629 --> 00:05:45.230 just went everywhere. Oh, they were so convincing. 00:05:45.449 --> 00:05:48.290 But verified later as totally fake, made by a 00:05:48.290 --> 00:05:51.209 tool called Nano Banana Pro. It just goes to 00:05:51.209 --> 00:05:53.529 show you how easy it's becoming to create really 00:05:53.529 --> 00:05:56.730 high quality, convincing misinformation. Deep 00:05:56.730 --> 00:05:59.079 fakes are here. On a more positive note, the 00:05:59.079 --> 00:06:01.139 tools for learning are getting way more interactive. 00:06:01.300 --> 00:06:03.180 I'm really excited about this. Gem and I just 00:06:03.180 --> 00:06:05.399 rolled out these new interactive diagrams. And 00:06:05.399 --> 00:06:07.079 this is where the tech goes beyond just being 00:06:07.079 --> 00:06:09.420 like an audio textbook. If you're looking at 00:06:09.420 --> 00:06:12.199 a complex system, like a diagram of a cell or 00:06:12.199 --> 00:06:14.879 the digestive system, you can now tap on any 00:06:14.879 --> 00:06:18.259 specific part of it. And you instantly get a 00:06:18.259 --> 00:06:20.959 definition, a deep explanation, all the context 00:06:20.959 --> 00:06:22.959 just for that one little piece. It's like having 00:06:22.959 --> 00:06:25.579 a dynamic tutor for any complex image you see. 00:06:25.720 --> 00:06:27.899 And we're also seeing AI pop up in some surprising 00:06:27.899 --> 00:06:30.000 consumer hardware. Alibaba just dropped some 00:06:30.000 --> 00:06:32.399 AI glasses that are surprisingly cheap. They 00:06:32.399 --> 00:06:34.579 look pretty much like normal glasses, but they 00:06:34.579 --> 00:06:37.000 can scan prices and translate speech in real 00:06:37.000 --> 00:06:39.519 time as you're walking around. A practical use 00:06:39.519 --> 00:06:42.199 case. Finally. And then on the pure ambition 00:06:42.199 --> 00:06:45.360 side of things, IBM is launching a $500 million. 00:06:45.610 --> 00:06:49.050 And it's focused specifically on AI and quantum 00:06:49.050 --> 00:06:51.610 breakthroughs. Their goal is not small. They're 00:06:51.610 --> 00:06:53.889 aiming for a fault -tolerant quantum computer 00:06:53.889 --> 00:06:58.529 by 2029. Wait, by 2029? Whoa. I mean, imagine 00:06:58.529 --> 00:07:01.110 scaling that. In just four years. Fault -tolerant 00:07:01.110 --> 00:07:03.290 compute power at that level. That changes, well, 00:07:03.370 --> 00:07:05.990 it changes everything. Every industry from material 00:07:05.990 --> 00:07:08.550 science to finance. It changes what's even possible. 00:07:09.199 --> 00:07:11.379 Before we move on to genetics, we did see some 00:07:11.379 --> 00:07:13.579 really useful career advice from an ex -meta 00:07:13.579 --> 00:07:15.980 director about how to break into the AI field. 00:07:16.139 --> 00:07:18.939 So beyond the usual talk about getting a PhD, 00:07:19.180 --> 00:07:21.019 what was the most valuable takeaway you saw from 00:07:21.019 --> 00:07:24.019 those tips? Practical experience. Solving real 00:07:24.019 --> 00:07:26.300 -world problems is way more critical than just 00:07:26.300 --> 00:07:29.300 having advanced degrees. That makes sense. Okay, 00:07:29.319 --> 00:07:31.860 so moving from compute and careers, let's jump 00:07:31.860 --> 00:07:34.399 into a genuine breakthrough in medicine. We're 00:07:34.399 --> 00:07:37.040 talking genetics. And the huge challenge of finding 00:07:37.040 --> 00:07:39.980 that one critical disease signal inside all of 00:07:39.980 --> 00:07:42.860 the harmless genetic noise. Right. So DeepMind's 00:07:42.860 --> 00:07:45.600 alpha missants made huge headlines for flagging 00:07:45.600 --> 00:07:49.180 potentially harmful DNA mutations, but just flagging 00:07:49.180 --> 00:07:51.779 a mutation. That's only the first step. The really 00:07:51.779 --> 00:07:54.199 hard part is figuring out which of those mutations 00:07:54.199 --> 00:07:56.899 actually causes a disease and which ones are 00:07:56.899 --> 00:07:59.480 just, you know, common variations that don't 00:07:59.480 --> 00:08:01.899 do anything, the background noise. And that difference, 00:08:01.920 --> 00:08:04.800 the signal versus the noise, that's exactly where 00:08:04.800 --> 00:08:07.079 this new AI from Harvard Medical School, Poppy 00:08:07.079 --> 00:08:10.879 VE, is just proving to be incredible. PopBVE's 00:08:10.879 --> 00:08:13.139 accuracy really comes down to its method. It 00:08:13.139 --> 00:08:15.800 doesn't just look at human DNA. First, it analyzes 00:08:15.800 --> 00:08:18.040 mutation patterns across hundreds of thousands 00:08:18.040 --> 00:08:20.459 of different species. That gives it this massive 00:08:20.459 --> 00:08:23.339 evolutionary context. It's like checking a variant 00:08:23.339 --> 00:08:25.540 against the master blueprint for all of life, 00:08:25.540 --> 00:08:27.860 not just the latest human version. So it's basically 00:08:27.860 --> 00:08:30.160 asking, how important has this gene been over 00:08:30.160 --> 00:08:32.720 millions of years? If it's essential for a frog 00:08:32.720 --> 00:08:34.659 and a fish, it's probably pretty important for 00:08:34.659 --> 00:08:37.809 a human too. Precisely. And then it calibrates 00:08:37.809 --> 00:08:39.649 those evolutionary predictions against these 00:08:39.649 --> 00:08:43.529 huge databases of healthy human genomes. Just 00:08:43.529 --> 00:08:45.669 to double check if a variant is actually common 00:08:45.669 --> 00:08:48.470 in people who don't have the disease, the result 00:08:48.470 --> 00:08:51.110 is a much, much more reliable ranking system 00:08:51.110 --> 00:08:54.149 for doctors. And the numbers here are just dramatic. 00:08:55.340 --> 00:09:00.639 Pop EV cuts false positives by over 75 % compared 00:09:00.639 --> 00:09:03.340 to alpha -miscence. That is a huge improvement 00:09:03.340 --> 00:09:05.620 in clarity for a diagnosis. Think about what 00:09:05.620 --> 00:09:08.360 that means for a patient. Alpha -miscence flagged 00:09:08.360 --> 00:09:11.039 44 % of healthy people as having harmful genetic 00:09:11.039 --> 00:09:13.960 variants. That creates so much unnecessary anxiety 00:09:13.960 --> 00:09:16.919 and confusion. Pop EVD, after it gets rid of 00:09:16.919 --> 00:09:19.700 all those false alarms, only flags 11%. That 00:09:19.700 --> 00:09:21.799 reduction in noise is life -changing. And this 00:09:21.799 --> 00:09:23.840 is where the real -world impact is just stunning. 00:09:24.139 --> 00:09:26.600 Researchers... ran PopEvie on the data from 31 00:09:26.600 --> 00:09:29.659 ,000 undiagnosed children, all with severe developmental 00:09:29.659 --> 00:09:32.019 issues. These were cases that had stumped doctors 00:09:32.019 --> 00:09:34.580 for years. The results were immediate. I mean, 00:09:34.580 --> 00:09:37.220 just incredible. Poppy solved one out of every 00:09:37.220 --> 00:09:39.759 three cases that had previously been unexplained. 00:09:39.899 --> 00:09:42.080 And it wasn't just confirming things we already 00:09:42.080 --> 00:09:45.659 knew. It flagged over 120 new genes that had 00:09:45.659 --> 00:09:47.600 never been linked to a human disease before. 00:09:47.879 --> 00:09:51.240 And get this, at least 24 of those have already 00:09:51.240 --> 00:09:53.240 been independently verified by other research 00:09:53.240 --> 00:09:56.179 teams. That's massive validation. It proves this 00:09:56.179 --> 00:09:58.980 AI is moving from just being an identifier to 00:09:58.980 --> 00:10:01.759 a real diagnostic discovery tool. It's actually 00:10:01.759 --> 00:10:05.080 accelerating research. So given how successful 00:10:05.080 --> 00:10:07.399 PPE has been with these rare developmental disorders, 00:10:07.799 --> 00:10:10.120 how quickly do you think this kind of method 00:10:10.120 --> 00:10:13.100 could be adopted for screening more common hereditary 00:10:13.100 --> 00:10:15.720 conditions? The high accuracy is going to rapidly 00:10:15.720 --> 00:10:18.000 accelerate its adoption for broad population 00:10:18.000 --> 00:10:20.740 screening and for personalized medicine. Wow. 00:10:20.840 --> 00:10:22.840 This has been a really comprehensive deep dive. 00:10:23.059 --> 00:10:25.100 We've covered the infrastructure battles, the 00:10:25.100 --> 00:10:27.519 security landscape, and now the frontier of medicine. 00:10:27.779 --> 00:10:30.179 Can we quickly recap the big ideas we hit today? 00:10:30.549 --> 00:10:33.149 The big strategy battle in AI compute, it's not 00:10:33.149 --> 00:10:35.549 about raw speed anymore. It's about owning the 00:10:35.549 --> 00:10:38.269 stack and winning on cost. Google is making a 00:10:38.269 --> 00:10:41.230 massive, very deliberate play for inference supremacy. 00:10:41.629 --> 00:10:44.250 And as these AI tools get deeper into our lives, 00:10:44.470 --> 00:10:46.889 our personal security is critically dependent 00:10:46.889 --> 00:10:49.029 on the reliability of all those third -party 00:10:49.029 --> 00:10:51.690 partners and APIs, not just the main company. 00:10:52.110 --> 00:10:56.169 And finally, AI, like POPEE, is moving beyond 00:10:56.169 --> 00:10:59.710 simple prediction. It's becoming a genuine, verifiable 00:10:59.710 --> 00:11:02.950 tool for diagnostic discovery. It's drastically 00:11:02.950 --> 00:11:05.490 cutting down that painful, crucial diagnostic 00:11:05.490 --> 00:11:07.769 time for families who are dealing with these 00:11:07.769 --> 00:11:10.950 severe, unexplained illnesses. Which really brings 00:11:10.950 --> 00:11:12.649 us to our final thought for you to think about. 00:11:13.029 --> 00:11:16.210 If Harvard's AI can solve one in three previously 00:11:16.210 --> 00:11:19.470 unexplained genetic mysteries today, What percentage 00:11:19.470 --> 00:11:21.850 of major human illnesses, the common ones we 00:11:21.850 --> 00:11:24.230 all deal with, will be fully explained by these 00:11:24.230 --> 00:11:26.909 AI models in, say, the next five years? That's 00:11:26.909 --> 00:11:28.570 the question. And that's what's driving all this 00:11:28.570 --> 00:11:30.350 innovation forward. Thank you for sharing your 00:11:30.350 --> 00:11:32.110 sources with us for this deep dive. We really 00:11:32.110 --> 00:11:33.909 encourage you to explore these ideas further. 00:11:34.049 --> 00:11:35.570 We will catch you on the next deep dive.