WEBVTT 00:00:00.000 --> 00:00:02.279 Let's start right at the top. The ultimate stakes. 00:00:03.120 --> 00:00:06.360 Some of the sources we looked at, they cite the 00:00:06.360 --> 00:00:09.939 the absolute worst case scenario for where A 00:00:09.939 --> 00:00:12.240 .I. could go. Yeah. And it's not just, you know, 00:00:12.240 --> 00:00:15.160 job displacement. It's way beyond that human 00:00:15.160 --> 00:00:18.359 obsolescence, potentially losing civil liberties 00:00:18.359 --> 00:00:22.210 and even. A complete extinction. Heavy stuff. 00:00:22.390 --> 00:00:24.489 It really is. And that tension, that's what defines 00:00:24.489 --> 00:00:27.129 this whole moment, doesn't it? Are we building 00:00:27.129 --> 00:00:29.390 these incredible new tools that will lift us 00:00:29.390 --> 00:00:33.390 up? Or are we just racing towards something uncontrollable, 00:00:33.429 --> 00:00:35.149 something that makes us irrelevant? That's the 00:00:35.149 --> 00:00:37.149 core question. Absolutely. And it frames what 00:00:37.149 --> 00:00:38.950 we're trying to do in this deep dive perfectly. 00:00:39.310 --> 00:00:41.229 We've gone through that stack of articles and 00:00:41.229 --> 00:00:44.189 research you gathered. Our mission is basically 00:00:44.189 --> 00:00:47.229 to pull out the key pieces, the nuggets of insight. 00:00:47.710 --> 00:00:50.030 from this paradox. So first, we'll dig into that 00:00:50.030 --> 00:00:52.590 really deep philosophical fight, why some top, 00:00:52.590 --> 00:00:55.189 top experts are pushing for a global ban on artificial 00:00:55.189 --> 00:00:57.829 superintelligence. Then second, we'll pivot to 00:00:57.829 --> 00:01:00.469 the speed, just the blistering pace of innovation 00:01:00.469 --> 00:01:03.549 right now, especially the hardware leaps, massive 00:01:03.549 --> 00:01:06.090 jumps. And finally, we'll bring it back down 00:01:06.090 --> 00:01:09.469 to earth a bit, a necessary reality check, using 00:01:09.469 --> 00:01:12.730 a pretty tough medical exam to show where AI 00:01:12.730 --> 00:01:15.569 still has very real limits today. Okay, let's 00:01:15.569 --> 00:01:17.700 start with that call for prohibition. It really 00:01:17.700 --> 00:01:20.200 feels like a flashpoint in the whole AI safety 00:01:20.200 --> 00:01:22.280 discussion. Definitely. There's this open letter, 00:01:22.359 --> 00:01:24.760 and it's specifically asking governments worldwide 00:01:24.760 --> 00:01:29.219 to actually prohibit developing artificial superintelligence, 00:01:29.439 --> 00:01:32.040 ASI. Yeah, and maybe we should quickly define 00:01:32.040 --> 00:01:36.620 ASI for everyone. Good idea. So ASI, it's theoretical 00:01:36.620 --> 00:01:40.379 intelligence. The idea is it would far, far surpass 00:01:40.379 --> 00:01:43.280 any human ability, like across every possible 00:01:43.280 --> 00:01:45.739 field. Exactly. And the conditions these folks 00:01:45.739 --> 00:01:47.959 are demanding before anyone should even think 00:01:47.959 --> 00:01:50.099 about proceeding, they're incredibly strict. 00:01:50.400 --> 00:01:52.219 Extremely. The Future of Life Institute, they 00:01:52.219 --> 00:01:54.340 published a letter and they want two big things. 00:01:54.519 --> 00:01:59.180 First, 100 % verifiable certainty that any ASI 00:01:59.180 --> 00:02:01.739 created would be controllable. 100 % certainty. 00:02:01.739 --> 00:02:04.640 Wow. Right. And second, total societal consensus 00:02:04.640 --> 00:02:07.159 before moving forward. Imagine achieving that. 00:02:07.500 --> 00:02:10.099 The risks they point to, well, catastrophic, 00:02:10.379 --> 00:02:12.740 permanent human obsolescence in the economy, 00:02:12.879 --> 00:02:16.379 losing civil liberties systemically, and that 00:02:16.379 --> 00:02:18.259 extinction risk you mentioned earlier. Yeah. 00:02:18.419 --> 00:02:20.599 And what really gives this whole debate, you 00:02:20.599 --> 00:02:22.439 know, moral weight is who signed it. We're not 00:02:22.439 --> 00:02:24.500 talking about just anybody. No, these are heavy 00:02:24.500 --> 00:02:26.400 hitters. You've got Yoshua Bengio and Jeffrey 00:02:26.400 --> 00:02:28.719 Hinton, often called the godfathers of AI, Steve 00:02:28.719 --> 00:02:33.000 Wozniak, Apple co -founder, even a current OpenAI 00:02:33.000 --> 00:02:36.639 staffer, Leo Gao. Their names alone force you 00:02:36.639 --> 00:02:38.900 to take the potential threat seriously. Absolutely. 00:02:39.039 --> 00:02:42.020 But, and this is a huge but, highlighted in the 00:02:42.020 --> 00:02:44.599 sources. Yeah, here's the catch. Not one single 00:02:44.599 --> 00:02:47.939 major commercial AI lab signed on. No OpenAI, 00:02:48.159 --> 00:02:51.219 no Google DeepMind, Anthropic, Meta, XAI, none 00:02:51.219 --> 00:02:53.800 of them. Right. Beat. And without their cooperation, 00:02:54.280 --> 00:02:56.379 I mean, these are the companies with the resources, 00:02:56.639 --> 00:02:59.659 the talent, the data. The ban just feels kind 00:02:59.659 --> 00:03:02.000 of symbolic, doesn't it? More than actually effective. 00:03:02.509 --> 00:03:05.030 It exposes that core tension, you know, safety 00:03:05.030 --> 00:03:07.689 versus the drive to innovate and compete commercially. 00:03:07.949 --> 00:03:11.030 And the sources also really hammer home the ambiguity 00:03:11.030 --> 00:03:14.090 problem. There just isn't a clear, agreed upon 00:03:14.090 --> 00:03:16.610 definition of what superintelligence even is. 00:03:16.729 --> 00:03:18.830 Right. So why would these labs sign a ban against 00:03:18.830 --> 00:03:21.449 something that isn't even properly defined, especially 00:03:21.449 --> 00:03:24.150 if it might limit their market edge down the 00:03:24.150 --> 00:03:26.949 line? Exactly. So, OK, given that definition 00:03:26.949 --> 00:03:29.110 problem and the fact that the big labs doing 00:03:29.110 --> 00:03:32.879 the work aren't on board. How does having the 00:03:32.879 --> 00:03:35.539 godfathers involved stop this whole safety push 00:03:35.539 --> 00:03:38.139 from just being ignored? I think it comes down 00:03:38.139 --> 00:03:42.439 to this. The founder's moral alarm kind of outweighs 00:03:42.439 --> 00:03:45.120 the corporate push for sheer speed. It forces 00:03:45.120 --> 00:03:47.639 a conversation, even if it doesn't force immediate 00:03:47.639 --> 00:03:50.150 action. Okay, that makes sense. Their reputations 00:03:50.150 --> 00:03:51.909 count for a lot. And that brings us perfectly 00:03:51.909 --> 00:03:54.330 to the speed factor, actually, that lack of definition 00:03:54.330 --> 00:03:56.569 we just talked about. It feels almost directly 00:03:56.569 --> 00:03:59.610 related to the incredible pace of development 00:03:59.610 --> 00:04:02.349 we're seeing. How so? Well, the market competition 00:04:02.349 --> 00:04:05.229 is just hyper intense now. It seems to be revealing 00:04:05.229 --> 00:04:08.330 some deep anxieties. Like Microsoft, they put 00:04:08.330 --> 00:04:11.490 $13 billion into open AI, right? Right. Huge 00:04:11.490 --> 00:04:14.240 investment. Yet they rushed out a nearly identical 00:04:14.240 --> 00:04:17.879 AI browser just 48 hours after OpenAI announced 00:04:17.879 --> 00:04:22.040 its new Atlas model. 48 hours. That sounds less 00:04:22.040 --> 00:04:24.980 like thoughtful innovation and more like, well, 00:04:25.060 --> 00:04:27.839 commercial paranoia. Exactly. The sources suggest 00:04:27.839 --> 00:04:30.879 this kind of aggressive, often overlapping product 00:04:30.879 --> 00:04:33.939 launch cycle isn't really driven by deep breakthroughs 00:04:33.939 --> 00:04:35.879 sometimes, but just the fear of falling behind. 00:04:36.389 --> 00:04:38.389 And it's not just the big players either. The 00:04:38.389 --> 00:04:40.670 sources mentioned a new browser, Strawberry. 00:04:41.410 --> 00:04:44.110 Apparently it outperformed the big four AI browsers 00:04:44.110 --> 00:04:46.430 in some head -to -head tests. Yeah, Strawberry. 00:04:46.689 --> 00:04:48.790 That probably comes down to some smart architectural 00:04:48.790 --> 00:04:52.189 choices, maybe using sophisticated agents, you 00:04:52.189 --> 00:04:54.329 know, AI systems that act autonomously to achieve 00:04:54.329 --> 00:04:57.509 goals, or maybe really optimized retrieval methods. 00:04:57.769 --> 00:04:59.310 Interesting. And we're not just tweaking browsers. 00:04:59.470 --> 00:05:02.089 We're building worlds now. Tencent released Hunyuan 00:05:02.089 --> 00:05:04.689 World. It's open source, and it takes regular 00:05:04.689 --> 00:05:06.970 videos or multi -view images and turns them into 00:05:06.970 --> 00:05:10.250 these incredibly detailed 3D worlds. Wow. Models 00:05:10.250 --> 00:05:12.470 like that. They need enormous speed. Yeah. Which 00:05:12.470 --> 00:05:15.790 brings us to hardware. Ah, yes. The Google announcement. 00:05:16.810 --> 00:05:18.769 This felt like a really big deal in the sources. 00:05:18.910 --> 00:05:22.290 A major quantum chip milestone. It's staggering. 00:05:22.449 --> 00:05:25.550 It apparently runs 13 ,000 times faster than 00:05:25.550 --> 00:05:28.589 the current top supercomputers. 13 ,000. Whoa. 00:05:29.009 --> 00:05:31.639 Just take a second to imagine scaling that. That 00:05:31.639 --> 00:05:36.500 13 ,000x speed applied to, say, a billion queries 00:05:36.500 --> 00:05:38.939 across the globe every single day. Yeah, that 00:05:38.939 --> 00:05:41.759 kind of leap in acceleration. Yeah. It just fundamentally 00:05:41.759 --> 00:05:43.699 changes the timeline for everything, solving 00:05:43.699 --> 00:05:47.300 huge problems maybe or potentially getting to 00:05:47.300 --> 00:05:49.560 ASI much faster than we thought. Right. That 00:05:49.560 --> 00:05:52.540 level of raw computational speed really underpins 00:05:52.540 --> 00:05:54.420 this whole race, doesn't it? It absolutely does. 00:05:54.519 --> 00:05:56.379 And the money flooding in reflects that speed, 00:05:56.459 --> 00:05:59.079 too. Found AI, a company that just hosts these 00:05:59.079 --> 00:06:01.480 big models, they just raised $250 million. The 00:06:01.480 --> 00:06:03.620 company is valued at over $4 billion now. Incredible. 00:06:03.660 --> 00:06:05.879 And the human effort matches the capital, apparently. 00:06:06.160 --> 00:06:08.620 Top researchers at all the big places, working 00:06:08.620 --> 00:06:11.259 80, even 100 hours a week. It's an insane pace, 00:06:11.439 --> 00:06:14.279 which leads to a question, right? If you look 00:06:14.279 --> 00:06:17.300 at this speed, the funding, these overlapping 00:06:17.300 --> 00:06:21.300 launches. Are these rapid releases a sign of 00:06:21.300 --> 00:06:23.819 genuine, deep innovation happening constantly? 00:06:24.279 --> 00:06:27.160 Or is it more about, you know, unsustainable 00:06:27.160 --> 00:06:28.819 market pressure just to keep up appearances? 00:06:29.160 --> 00:06:31.540 Based on the examples like that Microsoft browser 00:06:31.540 --> 00:06:34.180 relaunch, the evidence seems to lean heavily 00:06:34.180 --> 00:06:36.800 towards hyper -competitive market pressure driving 00:06:36.800 --> 00:06:39.939 these quick, sometimes redundant orgesses. Yeah, 00:06:39.959 --> 00:06:42.459 that feels right. Mid -role sponsor, red placeholder. 00:06:42.699 --> 00:06:45.149 Okay, let's ground ourselves again. Contrast 00:06:45.149 --> 00:06:47.209 all that speed and theoretical potential with 00:06:47.209 --> 00:06:49.750 the practical limits of AI as it exists right 00:06:49.750 --> 00:06:52.600 now. The sources had some fascinating data on 00:06:52.600 --> 00:06:55.540 this. Researchers tested CHAT -GPT against actual 00:06:55.540 --> 00:06:57.839 orthopedic residents. Yeah, on the OITE core 00:06:57.839 --> 00:07:00.139 exam. It's a tough, specialized test. And the 00:07:00.139 --> 00:07:02.600 results. Pretty sobering, actually. A real reality 00:07:02.600 --> 00:07:05.459 check. Definitely. CHAT -GPT barely kept up with 00:07:05.459 --> 00:07:08.000 the first -year residents. That's PGY -1, PGY 00:07:08.000 --> 00:07:10.879 -2 level folks. Right. And it significantly underperformed 00:07:10.879 --> 00:07:12.639 the national averages for all resident levels 00:07:12.639 --> 00:07:14.680 they tested. The difference was statistically 00:07:14.680 --> 00:07:17.350 significant, too. And there were a couple of 00:07:17.350 --> 00:07:19.829 key limitations causing that, right? First, the 00:07:19.829 --> 00:07:24.370 exam uses a lot of images, x -rays, scans. Which 00:07:24.370 --> 00:07:28.250 text -based models like ChatGPT still really 00:07:28.250 --> 00:07:30.490 struggle to interpret in context. They're not 00:07:30.490 --> 00:07:33.629 visual systems at their core. Exactly. And second, 00:07:33.709 --> 00:07:35.509 there's that really crucial failure mechanism 00:07:35.509 --> 00:07:37.910 that tells us a lot about how these large language 00:07:37.910 --> 00:07:40.569 models actually work. Yeah, explain that bit. 00:07:40.670 --> 00:07:43.889 It seemed important. So the sources explain they're 00:07:43.889 --> 00:07:47.019 autoregressive. Fancy word. But it just means 00:07:47.019 --> 00:07:48.939 they predict the next word based on the previous 00:07:48.939 --> 00:07:51.800 ones, sequentially. Okay. So if the model starts 00:07:51.800 --> 00:07:54.579 down the wrong path, makes a conceptual mistake 00:07:54.579 --> 00:07:58.079 early on, it can't really go back and rethink 00:07:58.079 --> 00:08:00.420 its whole approach like a human doctor might. 00:08:00.660 --> 00:08:02.980 Ah, I see. It just keeps building on the initial 00:08:02.980 --> 00:08:05.720 error. Pretty much. It compounds the mistake 00:08:05.720 --> 00:08:07.939 instead of correcting it. That's a really clear 00:08:07.939 --> 00:08:10.540 boundary of the current tech. It's not how human 00:08:10.540 --> 00:08:13.279 diagnostic reasoning works. That failure mechanism, 00:08:13.399 --> 00:08:15.439 that feels like essential knowledge for anyone 00:08:15.439 --> 00:08:17.160 using these tools, doesn't it? You can't just 00:08:17.160 --> 00:08:19.420 blindly trust the output. Absolutely critical. 00:08:19.660 --> 00:08:22.500 But, you know, balancing those limitations, AI 00:08:22.500 --> 00:08:25.959 still offers huge potential for learning, especially 00:08:25.959 --> 00:08:28.680 as a teaching tool, a didactic tool. Like that 00:08:28.680 --> 00:08:31.379 hack mentioned in the sources. Yeah. Using AI 00:08:31.379 --> 00:08:34.120 to organize massive amounts of information, like 00:08:34.120 --> 00:08:36.340 sorting through endless YouTube videos to create 00:08:36.340 --> 00:08:39.200 a clear curriculum, or even personalized audio 00:08:39.200 --> 00:08:42.350 lessons. It provides structure, which simplifies 00:08:42.350 --> 00:08:44.830 learning complex topics. You know, I have to 00:08:44.830 --> 00:08:46.929 admit, I still wrestle with prompt drift myself 00:08:46.929 --> 00:08:48.909 sometimes, especially when I'm trying to learn 00:08:48.909 --> 00:08:51.490 about a totally new field. So I really appreciate 00:08:51.490 --> 00:08:55.250 tools or resources that help simplify how these 00:08:55.250 --> 00:08:57.850 underlying systems actually function. Me too. 00:08:58.070 --> 00:09:01.159 It's easy to get lost. And to get the most out 00:09:01.159 --> 00:09:03.700 of AI, understanding those foundations is key. 00:09:03.980 --> 00:09:06.460 The source has actually highlighted three core 00:09:06.460 --> 00:09:08.799 concepts that underpin a lot of the most effective 00:09:08.799 --> 00:09:10.960 tools now. Okay, let's define those clearly. 00:09:11.159 --> 00:09:14.559 Sure. First is R -RAG. That's Retrieval Augmented 00:09:14.559 --> 00:09:17.000 Generation, where the AI searches trusted knowledge 00:09:17.000 --> 00:09:19.360 first to find answers. Okay, so it's not just 00:09:19.360 --> 00:09:21.799 making things up. Right. Then there's LORAO. 00:09:21.879 --> 00:09:23.899 That stands for Low Rank Adaptations. Basically 00:09:23.899 --> 00:09:26.679 a way to make fine -tuning models or personalizing 00:09:26.679 --> 00:09:28.879 them much faster and more efficient. Got it. 00:09:28.919 --> 00:09:32.340 And the third? Agents. We mentioned them briefly 00:09:32.340 --> 00:09:35.940 before. AI systems that act autonomously to achieve 00:09:35.940 --> 00:09:39.120 goals. They can take actions, not just generate 00:09:39.120 --> 00:09:43.269 text. Okay. Rag. LoRa, agents. Knowing those 00:09:43.269 --> 00:09:45.509 makes some of the practical applications seem 00:09:45.509 --> 00:09:48.070 less like magic, like that finance guide we saw. 00:09:48.210 --> 00:09:50.149 Yeah, the one showing five beginner -friendly 00:09:50.149 --> 00:09:52.950 ways to use free chat GPT for pretty sophisticated 00:09:52.950 --> 00:09:55.330 stuff like trading analysis or figuring out position 00:09:55.330 --> 00:09:58.590 sizes. Understanding R or how agents might work 00:09:58.590 --> 00:10:01.139 there makes it more transparent. So given those 00:10:01.139 --> 00:10:04.159 clear performance limits we saw in, say, orthopedics, 00:10:04.159 --> 00:10:07.139 a specialized high stakes field, how important 00:10:07.139 --> 00:10:10.279 is it really for the average, you know, infirmed 00:10:10.279 --> 00:10:12.899 user to grasp fundamentals like RAG or LoRa? 00:10:13.179 --> 00:10:14.620 I'd say it's pretty important. Understanding 00:10:14.620 --> 00:10:16.639 those foundations helps you critically evaluate 00:10:16.639 --> 00:10:19.340 what the AI gives you and use the tools effectively 00:10:19.340 --> 00:10:21.440 rather than just blindly trusting whatever answer 00:10:21.440 --> 00:10:23.740 pops out. Right. Knowing the how helps you judge 00:10:23.740 --> 00:10:26.299 the what. Exactly. So if we try to synthesize 00:10:26.299 --> 00:10:28.480 this whole paradox we've explored today. Let's 00:10:28.480 --> 00:10:30.700 recap the big picture. On one hand, you have 00:10:30.700 --> 00:10:34.639 the AI pioneers, world leaders, raising serious 00:10:34.639 --> 00:10:37.220 alarms about a potential superintelligence, calling 00:10:37.220 --> 00:10:41.019 for bans, citing legitimate extinction level 00:10:41.019 --> 00:10:43.080 risks. That's the theoretical high stakes end. 00:10:43.240 --> 00:10:46.620 But then the current operational reality, it's 00:10:46.620 --> 00:10:49.539 defined by these incredible quantum leaps in 00:10:49.539 --> 00:10:52.740 speed hardware running 13 ,000 times faster. 00:10:54.889 --> 00:10:57.789 Very specific, very real technical limits proven 00:10:57.789 --> 00:11:00.570 by things like that. AI failing a specialized 00:11:00.570 --> 00:11:03.289 medical test. So this huge gap between the theoretical 00:11:03.289 --> 00:11:06.230 danger and the immediate practical hurdles. Exactly. 00:11:06.230 --> 00:11:08.970 The tension there is just palpable. You can really 00:11:08.970 --> 00:11:11.820 feel it. Which leaves us and you. with a final 00:11:11.820 --> 00:11:13.879 kind of provocative thought to chew on after 00:11:13.879 --> 00:11:16.240 this deep dive. If the biggest labs, the ones 00:11:16.240 --> 00:11:19.039 with all the power and resources, refuse to even 00:11:19.039 --> 00:11:21.100 properly define artificial superintelligence, 00:11:21.139 --> 00:11:24.100 let alone agree to ban it, does that 13 ,000x 00:11:24.100 --> 00:11:27.080 speed increase from Google's quantum ship just 00:11:27.080 --> 00:11:30.000 push us closer to some unknown, potentially uncontrollable 00:11:30.000 --> 00:11:32.139 threat, faster than we can even agree on what 00:11:32.139 --> 00:11:34.779 the word superintelligence actually means? Beat. 00:11:34.960 --> 00:11:36.840 Something to think about. Thanks for digging 00:11:36.840 --> 00:11:39.080 into this critical landscape with us today. Out 00:11:39.080 --> 00:11:39.700 to your own music.