WEBVTT 00:00:00.000 --> 00:00:02.839 Can the soaring promises of AI hype actually 00:00:02.839 --> 00:00:06.139 meet the physical reality of the universe? That's 00:00:06.139 --> 00:00:07.820 really the tension at the center of our discussion 00:00:07.820 --> 00:00:10.279 today, because if you want advanced AI chips, 00:00:10.400 --> 00:00:13.660 the literal hardware that powers models like 00:00:13.660 --> 00:00:17.280 ChatGPT, the path runs through one machine, a 00:00:17.280 --> 00:00:19.140 single piece of equipment that costs a quarter 00:00:19.140 --> 00:00:21.679 of a billion dollars. it really is the ultimate 00:00:21.679 --> 00:00:24.019 physical constraint isn't it we're talking about 00:00:24.019 --> 00:00:28.059 extreme ultraviolet lithography or euv it's the 00:00:28.059 --> 00:00:30.719 very edge of what's possible in engineering almost 00:00:30.719 --> 00:00:32.840 impossible to build and without it you know the 00:00:32.840 --> 00:00:35.299 dreams of silicon valley just They can't happen. 00:00:35.439 --> 00:00:37.700 And what's fascinating is this brutal hardware 00:00:37.700 --> 00:00:40.299 race is happening just as Stanford experts are 00:00:40.299 --> 00:00:43.000 predicting a massive reality check for AI. Welcome 00:00:43.000 --> 00:00:44.759 to the deep dive. You sent over a fascinating 00:00:44.759 --> 00:00:47.799 stack of sources this week, everything from top 00:00:47.799 --> 00:00:49.820 secret Chinese prototypes being tested to the 00:00:49.820 --> 00:00:51.979 very skeptical view from Stanford's Institute 00:00:51.979 --> 00:00:54.570 for Human Centered AI. Right. Our mission is 00:00:54.570 --> 00:00:57.270 to take these two worlds, the cold, hard limits 00:00:57.270 --> 00:00:59.630 of manufacturing and the shifting expectations 00:00:59.630 --> 00:01:02.789 for AI and connect them for you. So we've broken 00:01:02.789 --> 00:01:05.489 the dive into three areas. First, this pretty 00:01:05.489 --> 00:01:08.310 incredible breakthrough in high stakes, super 00:01:08.310 --> 00:01:10.670 secret chip manufacturing. Then we'll do a quick 00:01:10.670 --> 00:01:13.090 run through of the ecosystem updates, you know, 00:01:13.150 --> 00:01:16.069 app stores, new models, huge investments. And 00:01:16.069 --> 00:01:18.849 finally, we'll spend some real time on why 2026 00:01:18.849 --> 00:01:22.170 is shaping up to be the year AI has to stop demoing 00:01:22.170 --> 00:01:25.069 and start. proving it actually works. Yeah. And 00:01:25.069 --> 00:01:26.750 that's according to the very researchers who 00:01:26.750 --> 00:01:29.310 helped start this whole LLM boom. Exactly. It's 00:01:29.310 --> 00:01:31.769 the journey from theory to accountability. Okay. 00:01:31.829 --> 00:01:34.419 Let's start with that foundational piece. EUV 00:01:34.419 --> 00:01:36.700 lithography, the quarter billion dollar machine. 00:01:36.760 --> 00:01:38.719 You said it's the bottleneck. It is the single 00:01:38.719 --> 00:01:40.920 most critical choke point, to put it simply for 00:01:40.920 --> 00:01:42.780 everyone listening. Advanced chips, the kind 00:01:42.780 --> 00:01:45.140 you need for AI, are made by etching microscopic 00:01:45.140 --> 00:01:49.060 circuits onto silicon. EUV tech uses this incredibly 00:01:49.060 --> 00:01:51.519 precise high energy light to print lines that 00:01:51.519 --> 00:01:53.599 are, well, they're measured in nanometers. And 00:01:53.599 --> 00:01:56.299 ASML, a company in the Netherlands, is the only 00:01:56.299 --> 00:01:58.140 place in the world that makes these machines. 00:01:58.400 --> 00:02:00.560 The only one. That really says something about 00:02:00.560 --> 00:02:03.140 the complexity. But the huge news in your sources 00:02:03.140 --> 00:02:06.099 is that China has apparently built a prototype 00:02:06.099 --> 00:02:09.819 EUV machine. That is the headline. Reuters is 00:02:09.819 --> 00:02:12.479 reporting it's now being tested in Shenzhen. 00:02:12.960 --> 00:02:15.500 Until now, this kind of manufacturing capability 00:02:15.500 --> 00:02:18.240 was just completely off limits to them because 00:02:18.240 --> 00:02:20.939 of geopolitical restrictions. So this prototype, 00:02:21.060 --> 00:02:23.819 it's a massive leap for their own tech independence. 00:02:24.319 --> 00:02:26.080 Let's talk about the technical difficulty for 00:02:26.080 --> 00:02:27.780 a second, because it's not just another machine. 00:02:28.240 --> 00:02:30.819 What's the big achievement here? Your sources 00:02:30.819 --> 00:02:32.379 said the hardest part is just generating the 00:02:32.379 --> 00:02:34.620 EUV light. Oh, absolutely. Think about what that 00:02:34.620 --> 00:02:37.599 means. To generate light waves short enough to 00:02:37.599 --> 00:02:39.780 etch at an atomic scale. They're taking molten 00:02:39.780 --> 00:02:42.479 tin, heating it to a plasma, then... hitting 00:02:42.479 --> 00:02:44.780 it with specialized lasers to create the light, 00:02:44.900 --> 00:02:47.120 and then focusing it with mirrors that are almost 00:02:47.120 --> 00:02:49.400 impossibly sensitive. The fact that the prototype 00:02:49.400 --> 00:02:51.280 successfully generates the light, that's the 00:02:51.280 --> 00:02:53.860 huge win. And what about reverse engineering? 00:02:53.979 --> 00:02:56.139 How much did that play into this? It was crucial. 00:02:56.439 --> 00:02:59.039 The report suggests that former engineers from 00:02:59.039 --> 00:03:02.120 ASML were involved, helping to reverse engineer 00:03:02.120 --> 00:03:04.979 key parts and providing that deep institutional 00:03:04.979 --> 00:03:07.750 knowledge. It's a classic pattern, you know, 00:03:07.789 --> 00:03:10.409 when a technology is heavily restricted, nations 00:03:10.409 --> 00:03:12.650 are just forced to pour everything into building 00:03:12.650 --> 00:03:14.810 it themselves. So connecting this back to the 00:03:14.810 --> 00:03:17.189 bigger picture, the urgency was there because 00:03:17.189 --> 00:03:18.949 they couldn't just keep buying older equipment. 00:03:19.090 --> 00:03:21.169 They couldn't. The license restrictions, mostly 00:03:21.169 --> 00:03:23.030 from the U .S. and the Netherlands, got tighter 00:03:23.030 --> 00:03:25.870 and tighter. China was basically blocked from 00:03:25.870 --> 00:03:29.090 buying, fixing or upgrading any of the top tier 00:03:29.090 --> 00:03:32.370 ASML gear. Just maintaining their existing machines 00:03:32.370 --> 00:03:35.009 became a nightmare. It forced them to go all 00:03:35.009 --> 00:03:36.900 in on this. They had no other choice if they 00:03:36.900 --> 00:03:39.240 wanted to build their own next -gen AI models. 00:03:39.479 --> 00:03:42.460 And the timeline. I mean, even with a successful 00:03:42.460 --> 00:03:44.800 prototype, we're not talking about mass production 00:03:44.800 --> 00:03:47.139 tomorrow. Not at all. The target for actual production 00:03:47.139 --> 00:03:50.219 is still like 2028 to 2030. Building the prototype 00:03:50.219 --> 00:03:52.740 is just step one of a marathon. It buys the West 00:03:52.740 --> 00:03:55.659 time. But the fundamental problem, generating 00:03:55.659 --> 00:03:58.460 the light, seems to be solved. I think that one 00:03:58.460 --> 00:04:01.229 anecdote you found really... grounds how hard 00:04:01.229 --> 00:04:03.229 this is. The one about the Chinese engineers 00:04:03.229 --> 00:04:06.069 with the deep UV machine. Oh, that story is perfect. 00:04:06.530 --> 00:04:09.030 So they tried to study one of ASML's slightly 00:04:09.030 --> 00:04:12.030 older machines, a DUV machine, by taking it completely 00:04:12.030 --> 00:04:14.930 apart. When they tried to reassemble it, they 00:04:14.930 --> 00:04:16.930 couldn't. The alignment was so critical. The 00:04:16.930 --> 00:04:19.769 tolerance is so fine. The machine just wouldn't 00:04:19.769 --> 00:04:21.689 work. They couldn't get it back together. Nope. 00:04:21.769 --> 00:04:24.769 They actually had to call ASML to fly in a team 00:04:24.769 --> 00:04:27.629 just to fix the misalignment they created. That 00:04:27.629 --> 00:04:30.769 just highlights the, the... astonishing precision 00:04:30.769 --> 00:04:33.410 needed. We're talking about a process where a 00:04:33.410 --> 00:04:35.790 single bump, a stray particle, or an atom -sized 00:04:35.790 --> 00:04:38.730 misalignment can kill a $250 million machine. 00:04:39.110 --> 00:04:41.629 It's more like physics research than it is manufacturing. 00:04:42.009 --> 00:04:44.949 Whoa. Just imagine trying to scale that kind 00:04:44.949 --> 00:04:47.410 of precision, that complexity, to handle a billion 00:04:47.410 --> 00:04:50.750 user queries for an AI. That level of control 00:04:50.750 --> 00:04:53.730 over matter is just... It's almost unbelievable. 00:04:53.990 --> 00:04:55.990 It really is. It defines the floor of global 00:04:55.990 --> 00:04:58.529 AI ambition. So given this immense difficulty, 00:04:58.829 --> 00:05:01.829 how critical is success in this EUV breakthrough 00:05:01.829 --> 00:05:05.490 for China's long -term AI independence? It determines 00:05:05.490 --> 00:05:07.550 whether they can truly build the chips needed 00:05:07.550 --> 00:05:10.430 for future advanced AI models. Okay, that sets 00:05:10.430 --> 00:05:12.889 the hardware stage perfectly. Now let's pivot 00:05:12.889 --> 00:05:15.970 from the physical foundation to the digital superstructure. 00:05:16.500 --> 00:05:19.240 because things are moving incredibly fast at 00:05:19.240 --> 00:05:21.139 the application layer. Right. This is where we 00:05:21.139 --> 00:05:23.139 see the development every day. The software ecosystem 00:05:23.139 --> 00:05:25.839 is basically trying to outrun the physical limits. 00:05:26.060 --> 00:05:28.699 Just look at the updates you found. On the learning 00:05:28.699 --> 00:05:32.600 side, Google's AI Agents course got 1 .5 million 00:05:32.600 --> 00:05:35.060 learners in five days. One and a half million 00:05:35.060 --> 00:05:38.410 in less than a week. That's a huge sign of global 00:05:38.410 --> 00:05:40.689 demand for this knowledge. People want to be 00:05:40.689 --> 00:05:42.769 builders, not just users. And now that course 00:05:42.769 --> 00:05:45.290 is completely free, which just accelerates things. 00:05:45.569 --> 00:05:48.310 Then on the deployment side, you have OpenAI 00:05:48.310 --> 00:05:51.290 launching the ChatGPT App Store. That's a major 00:05:51.290 --> 00:05:53.509 move. It's not just a chatbot anymore. It's becoming 00:05:53.509 --> 00:05:55.949 an operating system. Exactly. You can use apps 00:05:55.949 --> 00:05:59.610 like Apple Music, Photoshop, Spotify, inside 00:05:59.610 --> 00:06:02.829 ChatGPT. It's a platform. And here's the strategic 00:06:02.829 --> 00:06:05.670 part. They are already accepting app submissions 00:06:05.670 --> 00:06:08.209 for stuff that won't even roll out until 2026. 00:06:08.730 --> 00:06:10.990 They're locking developers into their ecosystem 00:06:10.990 --> 00:06:15.170 way, way in advance. And on a lighter note, the 00:06:15.170 --> 00:06:18.509 model names are getting interesting. Meta is 00:06:18.509 --> 00:06:21.269 sticking with the fruit theme. Yeah, after Avocado 00:06:21.269 --> 00:06:23.649 for Text, now they've got image and video models 00:06:23.649 --> 00:06:26.629 named Mango. It's a bit whimsical, but it shows 00:06:26.629 --> 00:06:29.480 they're diversifying fast. Moving way beyond 00:06:29.480 --> 00:06:32.180 just text models into creative tools. Meanwhile, 00:06:32.300 --> 00:06:34.399 competitors like Claude are stepping up their 00:06:34.399 --> 00:06:36.860 game. Yep, Claude Code just dropped four bid 00:06:36.860 --> 00:06:39.399 updates, including prompt suggestions and a plug 00:06:39.399 --> 00:06:41.939 -in marketplace. They're focusing on the developer 00:06:41.939 --> 00:06:44.279 experience, trying to make it easier to build 00:06:44.279 --> 00:06:46.600 on their models to gain market share. Okay, let's 00:06:46.600 --> 00:06:49.199 talk about the money. The investments keep circling 00:06:49.199 --> 00:06:51.100 back to that hardware problem we started with. 00:06:51.279 --> 00:06:53.740 Your source mentioned Amazon is in talks to invest 00:06:53.740 --> 00:06:57.920 $10 billion in OpenAI. A textbook example of 00:06:57.920 --> 00:06:59.980 how this race is structured now, it's a circular 00:06:59.980 --> 00:07:03.060 deal. Amazon gives OpenAI capital, which they 00:07:03.060 --> 00:07:05.839 desperately need for training. In return, OpenAI 00:07:05.839 --> 00:07:09.800 agrees to use Amazon's custom AI chips. It guarantees 00:07:09.800 --> 00:07:12.540 Amazon a massive customer for their hardware. 00:07:12.680 --> 00:07:14.540 And why is that circularity so important right 00:07:14.540 --> 00:07:17.439 now? Because of the chip crunch. Securing access 00:07:17.439 --> 00:07:19.839 to hardware is more valuable than just cash. 00:07:20.350 --> 00:07:23.250 This way, OpenAI gets the training capacity it 00:07:23.250 --> 00:07:25.730 needs and relies less on the traditional supply 00:07:25.730 --> 00:07:27.649 chains that are facing those EUV constraints. 00:07:28.230 --> 00:07:31.350 These deals are really strategic hardware partnerships. 00:07:31.670 --> 00:07:33.629 And beyond the big players, there are some really 00:07:33.629 --> 00:07:36.470 interesting new tools popping up. For sure. Look 00:07:36.470 --> 00:07:39.899 at Meta's SAM Audio. It can accurately separate 00:07:39.899 --> 00:07:42.620 any single sound like a voice or an instrument 00:07:42.620 --> 00:07:45.660 from a complex audio source. Which is a huge 00:07:45.660 --> 00:07:48.220 deal for creators. Absolutely. Then you have 00:07:48.220 --> 00:07:50.560 something like Bitrig, which can turn an idea 00:07:50.560 --> 00:07:53.019 into a real app on your phone. No coding needed. 00:07:53.279 --> 00:07:55.579 And then there's Cluely, which takes meeting 00:07:55.579 --> 00:07:58.100 notes and can give you real time undetectable 00:07:58.100 --> 00:08:00.319 answers during a meeting. That one is fascinating 00:08:00.319 --> 00:08:02.360 and maybe a little unsettling. It just speaks 00:08:02.360 --> 00:08:04.920 to how fast AI is integrating into our daily 00:08:04.920 --> 00:08:07.519 work. So with these huge investments like the 00:08:07.519 --> 00:08:10.509 Amazon. deal. Does this signal a slowdown or 00:08:10.509 --> 00:08:13.350 an acceleration in the AI race? This shows major 00:08:13.350 --> 00:08:16.689 tech giants are doubling down on strategic interlocked 00:08:16.689 --> 00:08:19.129 hardware partnerships. Okay, we've covered the 00:08:19.129 --> 00:08:21.370 physical constraints and the rapid expansion 00:08:21.370 --> 00:08:24.009 of the application layer. Now for that reality 00:08:24.009 --> 00:08:26.810 check. Let's pivot to the shifting expectations 00:08:26.810 --> 00:08:30.050 for 2026. Right, and this comes directly from 00:08:30.050 --> 00:08:32.429 the Stanford Institute for Human -Centered AI, 00:08:32.570 --> 00:08:36.330 or HAI. Their take is that 2025 has been the 00:08:36.330 --> 00:08:40.470 year of hype and demos. But 2026 is when AI has 00:08:40.470 --> 00:08:42.850 to prove it actually works. And this critique 00:08:42.850 --> 00:08:44.809 carries so much weight coming from Stanford, 00:08:44.929 --> 00:08:46.409 doesn't it? I mean, this is one of the universities 00:08:46.409 --> 00:08:49.830 that fueled the early LLM research boom. Now 00:08:49.830 --> 00:08:52.350 they're questioning if all this investment, over 00:08:52.350 --> 00:08:55.870 $10 billion in 2025 alone, is smoke or fire. 00:08:56.190 --> 00:08:58.129 That's the core of it. And four of their top 00:08:58.129 --> 00:09:00.350 researchers made some specific predictions. First 00:09:00.350 --> 00:09:03.460 one. No AGI. They think the industry will finally 00:09:03.460 --> 00:09:05.620 have to admit that there will be no AGI this 00:09:05.620 --> 00:09:08.299 year. AGI being Artificial General Intelligence, 00:09:08.539 --> 00:09:10.679 that sci -fi AI that can do any intellectual 00:09:10.679 --> 00:09:13.440 task a human can. Exactly. They expect companies 00:09:13.440 --> 00:09:15.460 to admit that AI hasn't really delivered big 00:09:15.460 --> 00:09:17.860 wins outside of very narrow areas, like code 00:09:17.860 --> 00:09:19.879 generation or call centers. The all -purpose 00:09:19.879 --> 00:09:21.720 AI is still a long way off. And we're going to 00:09:21.720 --> 00:09:23.419 start measuring this stuff, not just talking 00:09:23.419 --> 00:09:27.690 about it. Yes. They predict the rise of AI dashboards. 00:09:28.169 --> 00:09:30.809 Enterprises will start tracking AI's impact job 00:09:30.809 --> 00:09:33.509 changes, productivity shifts on a monthly basis. 00:09:33.710 --> 00:09:36.190 It'll be tracked like cash flow, like tracking 00:09:36.190 --> 00:09:38.590 steps on your watch. They also pointed to one 00:09:38.590 --> 00:09:41.250 specific sector for a breakout moment. A chat 00:09:41.250 --> 00:09:44.029 GPT moment for health care, meaning we'll finally 00:09:44.029 --> 00:09:47.009 see breakout real world deployments in clinical 00:09:47.009 --> 00:09:50.370 settings that get wide usage. And crucially, 00:09:50.509 --> 00:09:52.370 they'll have to prove they're safe and effective. 00:09:52.649 --> 00:09:55.509 That would be huge. And for law. It seems like 00:09:55.509 --> 00:09:58.289 the novelty is wearing off. Completely. The demo 00:09:58.289 --> 00:10:00.909 of an AI writing a simple contract is over. In 00:10:00.909 --> 00:10:05.129 2026, law AI gets serious. The focus shifts from 00:10:05.129 --> 00:10:07.190 what it can do to what the liability is. How 00:10:07.190 --> 00:10:09.289 accurate is it? How risky is it? Is it legally 00:10:09.289 --> 00:10:11.450 safe to use this in court? You know, I still 00:10:11.450 --> 00:10:13.490 wrestle with prompt drift myself, just using 00:10:13.490 --> 00:10:15.610 these tools daily. Small changes in my input 00:10:15.610 --> 00:10:17.830 can give me these wildly different outputs. So 00:10:17.830 --> 00:10:20.210 I get why Stanford is pushing for verifiable, 00:10:20.309 --> 00:10:22.110 measurable outcomes instead of just flashy demos. 00:10:22.350 --> 00:10:24.169 That's a really important point. It highlights. 00:10:24.360 --> 00:10:27.740 the core problem, reliability. When you think 00:10:27.740 --> 00:10:29.980 about the billions invested, the scrutiny makes 00:10:29.980 --> 00:10:32.580 perfect sense. The industry has to mature from 00:10:32.580 --> 00:10:34.840 research project to reliable infrastructure. 00:10:35.279 --> 00:10:37.500 So if general AI performance is slowing down, 00:10:37.720 --> 00:10:41.059 where does enterprise focus shift in 2026? The 00:10:41.059 --> 00:10:43.419 focus will move toward measuring AI's impact 00:10:43.419 --> 00:10:46.879 on jobs and verifiable productivity gains. So 00:10:46.879 --> 00:10:49.559 tying this all together, we started with the 00:10:49.559 --> 00:10:52.159 monumental engineering required for AI, that 00:10:52.159 --> 00:10:56.049 $250 million EUV machine. And we end with this 00:10:56.049 --> 00:10:58.769 inevitable shift to scrutinizing AI's actual 00:10:58.769 --> 00:11:01.830 utility Stanford's call for accountability. The 00:11:01.830 --> 00:11:03.610 main takeaway for you is that the future of AI 00:11:03.610 --> 00:11:06.350 isn't just about faster models. It's about solving 00:11:06.350 --> 00:11:08.809 nearly impossible engineering problems and delivering 00:11:08.809 --> 00:11:11.769 measurable, legally safe results. That tension, 00:11:12.049 --> 00:11:14.549 the fight to build a physical foundation while 00:11:14.549 --> 00:11:16.590 the digital promises are being held to the fire, 00:11:16.809 --> 00:11:19.389 that defines this moment. And it brings us to 00:11:19.389 --> 00:11:21.529 a final thought from your source material. If 00:11:21.529 --> 00:11:23.230 the cost to make advanced chips starts with a 00:11:23.230 --> 00:11:25.370 quarter billion dollar machine, What happens 00:11:25.370 --> 00:11:27.950 to the accessibility of AI power globally? It 00:11:27.950 --> 00:11:30.330 creates a tremendous divide. Is only the wealthiest 00:11:30.330 --> 00:11:32.309 corner of the world allowed to participate in 00:11:32.309 --> 00:11:35.370 building the future of AI? That's a fundamental 00:11:35.370 --> 00:11:37.610 question of equity we need to keep exploring. 00:11:37.850 --> 00:11:40.049 Thank you for sharing your sources and joining 00:11:40.049 --> 00:11:42.210 this deep dive with us today. It was a fascinating 00:11:42.210 --> 00:11:45.049 look at the limits of engineering and the limits 00:11:45.049 --> 00:11:48.289 of hype. Keep learning, keep exploring, and we'll 00:11:48.289 --> 00:11:49.710 catch you on the next deep dive.