WEBVTT 00:00:00.000 --> 00:00:03.259 The contrast in the AI world this week is, well, 00:00:03.319 --> 00:00:05.740 it's staggering. Yeah. On one side, you have 00:00:05.740 --> 00:00:08.900 these large language models suddenly cutting 00:00:08.900 --> 00:00:11.800 their operational costs in half, just like that. 00:00:11.839 --> 00:00:14.000 Right. And at the exact same time, those same 00:00:14.000 --> 00:00:16.379 models, now more efficient, are helping solve 00:00:16.379 --> 00:00:19.820 conceptual problems, problems that stumped, you 00:00:19.820 --> 00:00:22.219 know, the world's leading quantum computing minds. 00:00:22.399 --> 00:00:25.059 It's kind of wild. It really is. Welcome to the 00:00:25.059 --> 00:00:27.519 Deep Dive. You've shared a stack of sources this 00:00:27.519 --> 00:00:30.399 week, and they really prove AI is getting radically 00:00:30.399 --> 00:00:34.520 smarter and much more economically viable, both 00:00:34.520 --> 00:00:36.619 at once. This feels like the moment the ceiling 00:00:36.619 --> 00:00:39.460 gets raised, you know, on both capacity and capability. 00:00:40.020 --> 00:00:42.280 Our mission today is pretty clear, then. We're 00:00:42.280 --> 00:00:44.420 going to unpack DeepSeek's cost -cutting secret, 00:00:44.640 --> 00:00:47.020 this sparse attention thing, and figure out why 00:00:47.020 --> 00:00:49.840 it actually matters for, say, API prices. Okay. 00:00:50.280 --> 00:00:51.840 Then we're diving into the conceptual frontier. 00:00:51.960 --> 00:00:53.939 We'll look at how GPT -5 actually provided a 00:00:53.939 --> 00:00:56.000 genuine breakthrough in quantum mechanics. That's 00:00:56.000 --> 00:00:58.640 the Aronson story, right? Fascinating stuff. 00:00:58.960 --> 00:01:01.880 Exactly. And finally, we'll hit those critical 00:01:01.880 --> 00:01:03.920 shifts happening across the industry. Copyright 00:01:03.920 --> 00:01:07.200 fights, talent moving around, lots going on. 00:01:07.359 --> 00:01:09.079 All right, let's do it. Where do we start? The 00:01:09.079 --> 00:01:10.700 efficiency angle. Yeah, let's start with the 00:01:10.700 --> 00:01:13.480 money problem. It's fundamental to scaling AI, 00:01:13.599 --> 00:01:16.000 isn't it? Training, running these huge models. 00:01:16.819 --> 00:01:19.430 It's hitting a wall. Totally. And in the U .S., 00:01:19.430 --> 00:01:21.810 the major labs, you know, they often just throw 00:01:21.810 --> 00:01:24.409 more hardware at it, more NVIDIA GPUs, especially 00:01:24.409 --> 00:01:28.049 when dealing with long prompts, long conversations. 00:01:28.209 --> 00:01:30.329 That's the brute force method. Yeah. Yeah. More 00:01:30.329 --> 00:01:33.170 compute, more cost. Simple as that. Pretty much. 00:01:33.469 --> 00:01:36.450 But DeepSeek, being based in China, operates 00:01:36.450 --> 00:01:39.230 under different... let's say, resource constraints. 00:01:39.450 --> 00:01:42.209 They were kind of forced to find a smarter way 00:01:42.209 --> 00:01:44.310 to scale up. A different path. And they found 00:01:44.310 --> 00:01:46.650 a massive one. They just dropped this model, 00:01:46.810 --> 00:01:50.709 DeepSeek V3 .2 XP, and the headline here is huge. 00:01:50.930 --> 00:01:55.290 It cuts our operational costs by 50%. 50 % without 00:01:55.290 --> 00:01:57.709 losing performance. That's what they claim. No 00:01:57.709 --> 00:01:59.790 loss in quality compared to their previous models, 00:01:59.930 --> 00:02:02.569 which is, frankly, enormous news for the whole 00:02:02.569 --> 00:02:06.290 field. Okay, that is genuinely massive. So what's 00:02:06.290 --> 00:02:08.310 the technical trick? What lever do they pull 00:02:08.310 --> 00:02:11.449 for that kind of efficiency gain? The secret, 00:02:11.610 --> 00:02:13.590 or maybe the not -so -secret -anymore secret, 00:02:13.770 --> 00:02:16.129 is sparse attention. Sparse attention, okay. 00:02:16.210 --> 00:02:18.729 Yeah. And to get why it's such a big deal, you've 00:02:18.729 --> 00:02:20.810 got to remember how the original LLM architecture 00:02:20.810 --> 00:02:23.710 works, that 2017 transformer model. Right, the 00:02:23.710 --> 00:02:26.120 foundation for most of this stuff. Exactly. It 00:02:26.120 --> 00:02:29.319 uses this process where basically every single 00:02:29.319 --> 00:02:31.800 word looks at every other single word in the 00:02:31.800 --> 00:02:34.199 whole sequence. It compares everything to everything. 00:02:34.460 --> 00:02:36.819 Like that analogy of Lego blocks. Yeah. Every 00:02:36.819 --> 00:02:38.699 block has to talk to every other block to figure 00:02:38.699 --> 00:02:40.360 out the structure. That's a good way to put it. 00:02:40.379 --> 00:02:42.879 And it works fine for short bits of text. But 00:02:42.879 --> 00:02:45.819 imagine a document with, say, 10 ,000 words. 00:02:46.060 --> 00:02:48.860 Uh -oh. Yeah, that comparison process just explodes. 00:02:49.080 --> 00:02:52.039 It hits what's called quadratic staling. Basically, 00:02:52.060 --> 00:02:54.360 if you double the length of the input, the computation 00:02:54.360 --> 00:02:58.099 cost goes up by four times. It gets exponentially 00:02:58.099 --> 00:03:00.800 slower and way more expensive as your inputs 00:03:00.800 --> 00:03:03.699 get longer. Which explains why trying to summarize 00:03:03.699 --> 00:03:07.080 a really long document with AI can feel sluggish 00:03:07.080 --> 00:03:09.819 and why it often costs more credits or tokens. 00:03:09.979 --> 00:03:13.159 That quadratic thing. is the bottleneck it's 00:03:13.159 --> 00:03:15.340 the economic and technical bottleneck absolutely 00:03:15.340 --> 00:03:18.599 now sparse attention which deep seek is using 00:03:18.599 --> 00:03:21.199 here it completely changes that process how so 00:03:21.199 --> 00:03:23.400 well instead of comparing every word to every 00:03:23.400 --> 00:03:26.599 other word it's smarter it's an llm process that 00:03:26.599 --> 00:03:29.039 basically picks out only the keywords that actually 00:03:29.039 --> 00:03:31.120 matter for the comparison at that moment so it 00:03:31.120 --> 00:03:33.620 skips the noise yeah focuses on the relevant 00:03:33.620 --> 00:03:36.759 connections precisely it intelligently ignores 00:03:36.759 --> 00:03:39.409 the irrelevant stuff And how did they engineer 00:03:39.409 --> 00:03:41.770 that? How do you make the model know which words 00:03:41.770 --> 00:03:44.990 to skip? They built this tiny specialized system. 00:03:45.030 --> 00:03:46.530 They call it the lightning indexer. Lightning 00:03:46.530 --> 00:03:49.150 indexer, okay. This little indexer helps the 00:03:49.150 --> 00:03:51.530 main model prioritize which connections are important, 00:03:51.650 --> 00:03:53.430 which words need to talk to which other words, 00:03:53.550 --> 00:03:56.050 and that shifts the scaling away from that awful 00:03:56.050 --> 00:03:58.550 dollar. Towards something more manageable, like 00:03:58.550 --> 00:04:02.009 linear. Closer to linear, yeah. Much, much more 00:04:02.009 --> 00:04:03.810 manageable. And that's what suddenly makes really 00:04:03.810 --> 00:04:06.210 long context economically viable. And they published 00:04:06.210 --> 00:04:08.750 benchmarks showing it holds up like performance 00:04:08.750 --> 00:04:11.550 is still good compared to their old dense model 00:04:11.550 --> 00:04:14.990 V3 .1 Terminus. Yep. They claim no performance 00:04:14.990 --> 00:04:18.970 drop, but half the compute cost. OK, but you 00:04:18.970 --> 00:04:20.709 mentioned the original Transformers from 2017. 00:04:21.069 --> 00:04:23.759 Is sparse attention brand new science? That's 00:04:23.759 --> 00:04:26.319 the interesting part. Not really. OpenAI was 00:04:26.319 --> 00:04:28.759 actually pioneering sparse transformers back 00:04:28.759 --> 00:04:32.019 in 2019. And Google released something similar 00:04:32.019 --> 00:04:34.600 called Reformer in 2020. So the idea has been 00:04:34.600 --> 00:04:37.160 around for years. Exactly. But DeepSeek seems 00:04:37.160 --> 00:04:40.379 to be the first major lab to really openly publish 00:04:40.379 --> 00:04:42.819 their specific implementation, the results, the 00:04:42.819 --> 00:04:45.259 cost savings. They kind of put it all out there, 00:04:45.319 --> 00:04:47.399 made the tech public in a way. And the impact 00:04:47.399 --> 00:04:50.170 is already happening. Oh, yeah. API prices for 00:04:50.170 --> 00:04:52.269 handling long inputs, they've already been slashed 00:04:52.269 --> 00:04:55.370 up to 50 % in some cases. And this matters for 00:04:55.370 --> 00:04:57.810 you, the listener, because systems like, say, 00:04:57.870 --> 00:05:00.949 ChatGPT, they often still reprocess all the previous 00:05:00.949 --> 00:05:02.850 words in your conversation every time you add 00:05:02.850 --> 00:05:05.149 something new. Right, which is why those chat 00:05:05.149 --> 00:05:07.009 sessions can feel like they're slowing down the 00:05:07.009 --> 00:05:09.329 longer they go on. It's that silent tax on interaction 00:05:09.329 --> 00:05:12.170 length. Sparse attention helps fix that. Okay, 00:05:12.230 --> 00:05:14.569 so here's the probing question then. If this 00:05:14.569 --> 00:05:16.870 tech is so good at cutting costs, and it's not 00:05:16.870 --> 00:05:20.589 exactly brand new, Why has OpenAI been so quiet 00:05:20.589 --> 00:05:23.610 about using it, or if they're using it, in GPT 00:05:23.610 --> 00:05:26.189 -4 or GPT -5? That's a really good question. 00:05:26.230 --> 00:05:29.509 And it hints at maybe some economic friction, 00:05:29.629 --> 00:05:32.470 you could say. While the tech is proven, fully 00:05:32.470 --> 00:05:35.310 adopting it might mean rethinking or even partially 00:05:35.310 --> 00:05:38.310 abandoning massive investments already made in 00:05:38.310 --> 00:05:40.889 GPU clusters designed for the old, dense way 00:05:40.889 --> 00:05:43.009 of doing things. So yeah, efficiency remains 00:05:43.009 --> 00:05:45.149 this kind of hidden scaling constraint for all 00:05:45.149 --> 00:05:47.129 the big players. Okay, so we've tackled the money 00:05:47.129 --> 00:05:49.490 problem, the efficiency leap. But what about 00:05:49.490 --> 00:05:52.149 the actual brainpower, the capability side? Right. 00:05:52.230 --> 00:05:54.449 Let's talk about Scott Aronson, the quantum computing 00:05:54.449 --> 00:05:56.790 legend. We're shifting gears now from saving 00:05:56.790 --> 00:06:00.069 money to actual conceptual breakthroughs. Yeah, 00:06:00.110 --> 00:06:01.589 this is where it gets really, really interesting 00:06:01.589 --> 00:06:03.990 for me. Aronson was deep into this notoriously 00:06:03.990 --> 00:06:06.970 tricky quantum proof. It concerns something called 00:06:06.970 --> 00:06:12.569 QMA. QMA. OK, for those of us not deep in complexity 00:06:12.569 --> 00:06:16.100 theory. What is that in simple terms? Simple 00:06:16.100 --> 00:06:19.339 terms. OK, think of it like this. You know, MP 00:06:19.339 --> 00:06:21.879 problems, problems where if someone gives you 00:06:21.879 --> 00:06:23.920 a solution, you can check it quickly on a regular 00:06:23.920 --> 00:06:27.240 computer. QMA or quantum Merlin Arthur is basically 00:06:27.240 --> 00:06:30.040 the quantum version of that. It deals with proofs 00:06:30.040 --> 00:06:31.939 that need a quantum computer to verify quickly. 00:06:32.139 --> 00:06:35.180 Got it. Quantum MP. Kind of. Kind of. Yeah. And 00:06:35.180 --> 00:06:38.769 Aronson was stuck on a specific part. He was 00:06:38.769 --> 00:06:40.769 trying to prove that a new method he found for 00:06:40.769 --> 00:06:42.949 amplifying the certainty of these QMA proofs 00:06:42.949 --> 00:06:45.910 was truly optimal. He couldn't quite nail down 00:06:45.910 --> 00:06:48.310 the perfect mathematical verifier function. So 00:06:48.310 --> 00:06:50.230 a technical roadblock in his own mathematical 00:06:50.230 --> 00:06:54.069 work. A human stuck point. Exactly. And so he 00:06:54.069 --> 00:06:57.029 turned to GPT -5 thinking for help. Just ask 00:06:57.029 --> 00:06:59.670 the AI. Yep. And what happened next is pretty 00:06:59.670 --> 00:07:01.769 remarkable, especially given who Aronson is. 00:07:01.930 --> 00:07:04.209 Apparently, the model gave a couple of unhelpful 00:07:04.209 --> 00:07:06.930 ideas at first. Okay. Typical AI sometimes. Right. 00:07:07.009 --> 00:07:09.470 But Aronson gave it some course correction, nudged 00:07:09.470 --> 00:07:12.389 it a bit, and then GPT -5 suggested a specific 00:07:12.389 --> 00:07:14.350 mathematical function. And this function? Mm 00:07:14.350 --> 00:07:17.649 -hmm. It worked. It broke his mental block completely. 00:07:17.829 --> 00:07:20.230 And the key thing here is Aronson's reaction, 00:07:20.350 --> 00:07:22.389 right? What did he say about it? He called the 00:07:22.389 --> 00:07:25.769 suggestion non -obvious and genuinely useful. 00:07:25.970 --> 00:07:28.529 Not obvious. Yeah. And he even added this quote, 00:07:28.610 --> 00:07:30.930 which I think just says it all. If a grad student 00:07:30.930 --> 00:07:33.009 had given it to me, I'd have called it clever. 00:07:33.269 --> 00:07:36.939 Wow. OK, clever. Yeah. From Scott Aronson about 00:07:36.939 --> 00:07:39.560 an AI suggestion on quantum complexity theory. 00:07:39.740 --> 00:07:42.180 Right. That means the AI didn't just compute 00:07:42.180 --> 00:07:44.560 something he asked for. It didn't just run through 00:07:44.560 --> 00:07:48.279 known patterns. It actually found a novel step, 00:07:48.399 --> 00:07:51.420 an elegant shortcut, maybe something a top human 00:07:51.420 --> 00:07:53.379 mind working on the problem had missed. Whoa. 00:07:53.839 --> 00:07:57.240 I mean, just imagine that AI actually contributing 00:07:57.240 --> 00:08:00.040 a genuinely clever step to advanced scientific 00:08:00.040 --> 00:08:02.990 theory. That feels. different. It feels very 00:08:02.990 --> 00:08:04.269 different. This is probably the first really 00:08:04.269 --> 00:08:06.149 clear example we've seen of this kind of thing, 00:08:06.189 --> 00:08:07.550 but you can bet it's going to be the first of 00:08:07.550 --> 00:08:11.370 thousands. It points towards true AI co -creation 00:08:11.370 --> 00:08:14.449 and science. Beat. So does a story like this 00:08:14.449 --> 00:08:17.569 confirm that AI is moving beyond just simulation, 00:08:17.889 --> 00:08:21.129 beyond pattern matching, towards genuine co -creation 00:08:21.129 --> 00:08:24.029 in really high level research? I think it strongly 00:08:24.029 --> 00:08:26.649 suggests that, yeah, AI is becoming an invaluable 00:08:26.649 --> 00:08:29.829 partner for generating novel, non -obvious research 00:08:29.829 --> 00:08:33.129 insights. Okay, let's pivot then. Broader industry 00:08:33.129 --> 00:08:37.370 shift. Application adoption. Maybe some friction 00:08:37.370 --> 00:08:40.870 points. Quickfire. Let's do it. First up, institutional 00:08:40.870 --> 00:08:43.330 adoption seems to be accelerating like crazy. 00:08:43.570 --> 00:08:47.370 USC just gave full access to ChatGPT to all students, 00:08:47.509 --> 00:08:50.870 staff, faculty. Big deal. Apparently $1 .5 million 00:08:50.870 --> 00:08:54.159 for one year. That's scale. That is huge scale. 00:08:54.340 --> 00:08:57.000 And while adoption like that speeds up, the legal 00:08:57.000 --> 00:08:59.460 fights are also heating up. No surprise there, 00:08:59.539 --> 00:09:02.500 really, especially around content. OpenAI launched 00:09:02.500 --> 00:09:05.080 Sora 2. Yeah, the video app. Yeah. Makes those 00:09:05.080 --> 00:09:07.299 short. like 10 second clips. They look pretty 00:09:07.299 --> 00:09:09.440 amazing. They do. But here's the controversy, 00:09:09.539 --> 00:09:11.659 the big friction point. Sources are saying it 00:09:11.659 --> 00:09:14.360 uses copyrighted material unless the owners actively 00:09:14.360 --> 00:09:17.419 opt out. Ah, the opt out model. That flips the 00:09:17.419 --> 00:09:19.480 burden entirely, doesn't it? Completely. It puts 00:09:19.480 --> 00:09:21.820 the massive job of protecting content onto the 00:09:21.820 --> 00:09:24.080 creators, not the AI company scraping the data. 00:09:24.159 --> 00:09:26.340 You're kind of presumed in unless you fight to 00:09:26.340 --> 00:09:28.519 get out. Guilty until proven innocent, almost. 00:09:29.259 --> 00:09:32.419 You can see it that way. And it forces big media 00:09:32.419 --> 00:09:34.980 companies to constantly police what the AI is 00:09:34.980 --> 00:09:37.519 learning from. We already saw Disney apparently 00:09:37.519 --> 00:09:40.059 bail on that kind of arrangement. Yeah, I saw 00:09:40.059 --> 00:09:42.519 that. This feels like it's going to define copyright 00:09:42.519 --> 00:09:45.340 law for the next decade. I think so, too. OK, 00:09:45.500 --> 00:09:48.559 shifting from consumption. to creation tools. 00:09:49.019 --> 00:09:52.000 Anthropic put out an important paper making a 00:09:52.000 --> 00:09:54.919 distinction between context engineering and prompt 00:09:54.919 --> 00:09:57.419 engineering. Okay. That sounds useful. We hear 00:09:57.419 --> 00:09:59.840 prompt engineering all the time, but context 00:09:59.840 --> 00:10:01.639 engineering, what's the difference they're highlighting? 00:10:01.860 --> 00:10:05.059 Good question. Because honestly, I still wrestle 00:10:05.059 --> 00:10:07.610 with prompt drift myself sometimes. Yeah, me 00:10:07.610 --> 00:10:11.049 too. You spend ages crafting this perfect persona 00:10:11.049 --> 00:10:14.409 or instruction set for the AI. And then like 00:10:14.409 --> 00:10:16.090 two turns later, it's completely forgotten and 00:10:16.090 --> 00:10:18.750 gone off track. Is context engineering meant 00:10:18.750 --> 00:10:21.230 to fix that? That's basically the pain point 00:10:21.230 --> 00:10:23.009 it addresses, yeah. Yeah. So prompt engineering 00:10:23.009 --> 00:10:25.889 is just what you ask the model, the actual question 00:10:25.889 --> 00:10:28.409 or command. Context engineering is about building 00:10:28.409 --> 00:10:30.970 the stuff around the prompt before you even ask. 00:10:31.389 --> 00:10:33.350 It's like setting up the instruction manual, 00:10:33.470 --> 00:10:37.389 the environment, the guardrails for the AI. Ensuring 00:10:37.389 --> 00:10:40.509 the model really understands its role, its constraints, 00:10:40.730 --> 00:10:43.429 the background, before it even starts thinking 00:10:43.429 --> 00:10:45.470 about your specific question. So it's like setting 00:10:45.470 --> 00:10:48.190 the stage properly, not just delivering the lines. 00:10:48.389 --> 00:10:51.309 Exactly. And Anthropic argues pretty convincingly 00:10:51.309 --> 00:10:54.090 that it leads to much more consistent, reliable 00:10:54.090 --> 00:10:59.350 AI behavior. Less drift. Makes sense. Okay, what 00:10:59.350 --> 00:11:02.029 else? Infrastructure. Yeah, quick one. Google 00:11:02.029 --> 00:11:05.009 Drive is now using AI, apparently, to spot ransomware 00:11:05.009 --> 00:11:07.509 attacks and help users quickly restore files 00:11:07.509 --> 00:11:09.950 that got scrambled. Oh, that's practical. Security 00:11:09.950 --> 00:11:12.330 becoming a core AI application. Good to see. 00:11:12.529 --> 00:11:14.370 Definitely. And then there's the talent shift. 00:11:14.590 --> 00:11:16.590 This was noteworthy. Sources mentioned about 00:11:16.590 --> 00:11:19.450 20 top AI researchers have left the big established 00:11:19.450 --> 00:11:23.889 labs, OpenAI, Google, Meta. 20? Wow, where'd 00:11:23.889 --> 00:11:26.200 they go? To start a new company together. A new 00:11:26.200 --> 00:11:29.100 AI venture. That is a massive brain drain from 00:11:29.100 --> 00:11:32.100 the incumbents. 20 top people. Yeah. It shows 00:11:32.100 --> 00:11:34.120 incredible confidence in the market for specialized 00:11:34.120 --> 00:11:36.740 new ventures, right? Especially considering how 00:11:36.740 --> 00:11:39.080 insanely expensive it is to start a frontier 00:11:39.080 --> 00:11:42.299 AI lab from scratch these days. For sure. Funding 00:11:42.299 --> 00:11:44.659 spotlight Assort Health. They secured $76 million 00:11:44.659 --> 00:11:47.159 for their voice AI platform. The scale numbers 00:11:47.159 --> 00:11:50.059 were impressive. 14 languages handled 42 million 00:11:50.059 --> 00:11:52.980 patients. Eightfold revenue growth. Shows health 00:11:52.980 --> 00:11:55.960 AI is scaling fast. Voice AI and healthcare, 00:11:56.120 --> 00:11:59.759 big area. And finally, tool update. Right. Cursor, 00:11:59.759 --> 00:12:02.379 the AI -first code editor, now supports controlling 00:12:02.379 --> 00:12:04.399 your browser, grabbing screenshots, debugging 00:12:04.399 --> 00:12:06.980 client -side issues, all integrated with Cloud 00:12:06.980 --> 00:12:09.580 Sonnet 4 .5, apparently, which is making the 00:12:09.580 --> 00:12:11.740 developer workflow smoother. Interesting. Tools 00:12:11.740 --> 00:12:14.159 getting more integrated. So that talent exodus. 00:12:14.399 --> 00:12:16.519 Yeah. It does raise a question, doesn't it? What 00:12:16.519 --> 00:12:18.919 does this movement of top talent away from the 00:12:18.919 --> 00:12:21.759 big labs really suggest about where future AI 00:12:21.759 --> 00:12:23.980 innovation might be heading? Yeah, that's a good 00:12:23.980 --> 00:12:25.899 point. I guess it suggests that frontier innovation 00:12:25.899 --> 00:12:29.500 is increasingly driven by these focused, specialized 00:12:29.500 --> 00:12:32.620 startups, maybe. More agility, perhaps. Seems 00:12:32.620 --> 00:12:34.700 plausible. Yeah. Okay, let's try and recap the 00:12:34.700 --> 00:12:37.639 big ideas from this deep dive. Sounds good. So 00:12:37.639 --> 00:12:40.720 we saw two really massive threads changing the 00:12:40.720 --> 00:12:44.389 AI. landscape just this week it feels like first 00:12:44.389 --> 00:12:47.029 the whole economic reality shifted deep seek 00:12:47.029 --> 00:12:49.070 basically proved you could potentially have your 00:12:49.070 --> 00:12:52.009 operational costs using these sparse attention 00:12:52.009 --> 00:12:54.669 techniques like their lightning indexer it tackles 00:12:54.669 --> 00:12:56.830 that critical quadratic scaling problem right 00:12:56.830 --> 00:12:59.250 making ai cheaper and feasible for much larger 00:12:59.250 --> 00:13:02.690 tasks and second the conceptual ceiling got pushed 00:13:02.690 --> 00:13:06.470 way higher. Yeah, the Aronson story. GPT -5 providing 00:13:06.470 --> 00:13:09.490 that genuinely clever, non -obvious step for 00:13:09.490 --> 00:13:12.129 a quantum proof. That confirms we're really moving 00:13:12.129 --> 00:13:15.070 into an era of AI co -authorship, even in really 00:13:15.070 --> 00:13:17.509 hard science. Definitely. And then layered on 00:13:17.509 --> 00:13:19.720 top of that, you have all these... critical industry 00:13:19.720 --> 00:13:21.940 shifts happening fast. Yeah, the rapid adoption, 00:13:22.059 --> 00:13:24.600 like at USC, the looming copyright battles, especially 00:13:24.600 --> 00:13:28.039 over that Sora 2 opt -out model. Right, and that 00:13:28.039 --> 00:13:30.620 talent fragmentation, top researchers leaving 00:13:30.620 --> 00:13:33.679 big labs for nimbler startups. So the whole AI 00:13:33.679 --> 00:13:35.820 landscape, it feels like it's simultaneously 00:13:35.820 --> 00:13:38.240 maturing, becoming more efficient, more stable, 00:13:38.320 --> 00:13:41.620 and accelerating, getting dramatically more capable, 00:13:41.919 --> 00:13:44.340 tackling harder problems. It's kind of doing 00:13:44.340 --> 00:13:46.559 both at once. Yeah, that's a good way to put 00:13:46.559 --> 00:13:49.419 it, maturing and accelerating. So for you listening, 00:13:49.580 --> 00:13:51.919 maybe explore one of those quick hits we mentioned. 00:13:52.019 --> 00:13:54.679 Good idea. You can look into the details of Anthropic's 00:13:54.679 --> 00:13:56.820 context engineering paper, see if it helps your 00:13:56.820 --> 00:14:00.159 own AI interactions. Or maybe dig into the implications 00:14:00.159 --> 00:14:03.419 of that SORA2 opt -out model for copyright and 00:14:03.419 --> 00:14:06.559 creative work. Yeah, lots to dig into. But here's 00:14:06.559 --> 00:14:08.600 the final, maybe provocative thought we want 00:14:08.600 --> 00:14:12.120 to leave you with. Okay. If AI can demonstrably 00:14:12.120 --> 00:14:15.120 help break conceptual roadblocks in something 00:14:15.120 --> 00:14:19.110 as complex as quantum theory today. What previously 00:14:19.110 --> 00:14:21.870 intractable scientific problem, maybe in medicine 00:14:21.870 --> 00:14:25.210 or material science or fundamental physics, what's 00:14:25.210 --> 00:14:27.809 it going to tackle next year? That is something 00:14:27.809 --> 00:14:29.909 to think about. Where does this capability actually 00:14:29.909 --> 00:14:31.870 lead us? A big question. Well, thank you for 00:14:31.870 --> 00:14:33.870 sharing your sources with us for this deep dive. 00:14:34.090 --> 00:14:34.769 Always fascinating.