WEBVTT 00:00:00.000 --> 00:00:03.259 Imagine you just hand a computer program $100 00:00:03.259 --> 00:00:06.040 ,000. You walk away for six months, and you come 00:00:06.040 --> 00:00:08.839 back to find it's autonomously turned that money 00:00:08.839 --> 00:00:10.859 into a million bucks. Right. And we're not talking 00:00:10.859 --> 00:00:13.599 about it just running high -frequency stock trades 00:00:13.599 --> 00:00:15.339 or some algorithm like that. Exactly. I mean, 00:00:15.560 --> 00:00:18.100 it researched a market cap, registered an LLC. 00:00:18.830 --> 00:00:21.410 negotiated with international suppliers, hired 00:00:21.410 --> 00:00:24.170 freelance web designers, and then ran its own 00:00:24.170 --> 00:00:26.629 targeted ad campaigns. Yeah, completely on its 00:00:26.629 --> 00:00:29.149 own. Yeah, without a single human prompting its 00:00:29.149 --> 00:00:32.289 next move. And the crazy thing is, that is not 00:00:32.289 --> 00:00:35.469 a science fiction script. That is a very real 00:00:35.469 --> 00:00:38.469 benchmark being used right now to define the 00:00:38.469 --> 00:00:41.030 absolute holy grail of technology. It really 00:00:41.030 --> 00:00:43.250 is. It fundamentally changes how we interact 00:00:43.250 --> 00:00:45.670 with machines, right? We are moving from these 00:00:45.670 --> 00:00:48.130 tools that require our constant step -by -step 00:00:48.130 --> 00:00:51.090 direction to autonomous agents that just pursue 00:00:51.090 --> 00:00:53.850 long -term goals. So welcome to this deep dive. 00:00:54.130 --> 00:00:56.369 If you are listening to this, you are the learner. 00:00:56.850 --> 00:00:59.539 Whether you're a... Prepping for a high -stake 00:00:59.539 --> 00:01:01.420 strategy meeting or trying to make sense of where 00:01:01.420 --> 00:01:03.579 the tech world is heading, or you're just insanely 00:01:03.579 --> 00:01:06.000 curious about the future, this conversation is 00:01:06.000 --> 00:01:09.120 custom tailored for you. Absolutely. Today our 00:01:09.120 --> 00:01:11.760 topic is AGI, artificial general intelligence. 00:01:12.060 --> 00:01:15.439 We're analyzing a massive, incredibly detailed 00:01:15.439 --> 00:01:18.859 encyclopedic update from March of 2026. It's 00:01:18.859 --> 00:01:21.439 a huge document. It really is. It tracks everything 00:01:21.439 --> 00:01:24.420 from the absolute foundational basics of AGI 00:01:24.420 --> 00:01:27.120 to the most cutting edge debates happening in 00:01:27.120 --> 00:01:29.420 laboratories right now. And our mission today 00:01:29.420 --> 00:01:31.469 is simple. We want to cut through the Silicon 00:01:31.469 --> 00:01:34.049 Valley marketing, ignore the doomsday movie plots, 00:01:34.090 --> 00:01:36.489 and figure out what AGI actually is. Right. How 00:01:36.489 --> 00:01:38.370 it works and what it means for your daily life. 00:01:38.510 --> 00:01:40.590 Yeah. So before we can even talk about timelines 00:01:40.590 --> 00:01:43.310 or consequences, we need a baseline. Like you 00:01:43.310 --> 00:01:45.510 have AI on your phone right now. It curates your 00:01:45.510 --> 00:01:47.609 social media feeds. It finishes your sentences 00:01:47.609 --> 00:01:50.310 and emails. What separates that from artificial 00:01:50.310 --> 00:01:52.810 general intelligence? Well, the defining line 00:01:52.810 --> 00:01:56.079 there is adaptability. The AI we interact with 00:01:56.079 --> 00:01:58.700 on a daily basis is known as artificial narrow 00:01:58.700 --> 00:02:02.620 intelligence, or ANI. OK. And it is incredibly 00:02:02.620 --> 00:02:05.700 competent, right? Like, sometimes superhumanly 00:02:05.700 --> 00:02:08.379 so, but it's confined to a very specific domain. 00:02:08.659 --> 00:02:11.919 So a chess AI can beat a grandmaster. But if 00:02:11.919 --> 00:02:14.780 you ask that exact same neural network, to balance 00:02:14.780 --> 00:02:16.919 a checkbook or, I don't know, summarize a novel, 00:02:17.159 --> 00:02:21.419 it completely failed. It has zero capacity to 00:02:21.419 --> 00:02:24.400 transfer its knowledge. Artificial general intelligence, 00:02:24.400 --> 00:02:27.060 on the other hand, possesses the ability to generalize. 00:02:27.460 --> 00:02:29.960 It can learn a skill in one domain, understand 00:02:29.960 --> 00:02:31.919 the underlying principles of it, and then apply 00:02:31.919 --> 00:02:34.599 them to a completely novel problem in a totally 00:02:34.599 --> 00:02:36.699 different domain. Without a human engineer having 00:02:36.699 --> 00:02:39.400 to jump in and rewrite its code. Precisely. OK, 00:02:39.400 --> 00:02:41.520 let's unpack this for a second. So if narrow 00:02:41.520 --> 00:02:44.710 AI is like a highly specialized calculator, then 00:02:44.710 --> 00:02:48.069 AGI is basically a brilliant human intern who 00:02:48.069 --> 00:02:50.050 can figure out how to make the office coffee, 00:02:50.509 --> 00:02:52.349 build a desk, and invest your company's money. 00:02:52.469 --> 00:02:54.409 That is a remarkably accurate way to look at 00:02:54.409 --> 00:02:56.789 it, actually. And just so we have our vocabulary 00:02:56.789 --> 00:02:59.710 straight here, how does AGI differ from artificial 00:02:59.710 --> 00:03:02.009 superintelligence or ASI? You hear those two 00:03:02.009 --> 00:03:03.449 acronyms thrown around together all the time. 00:03:03.530 --> 00:03:06.629 Yeah, they do get conflated. So AGI is the bridge. 00:03:06.949 --> 00:03:09.949 And AGI basically matches human cognitive capabilities 00:03:09.949 --> 00:03:12.469 across the board. Artificial superintelligence. 00:03:12.590 --> 00:03:16.569 is the point where the system vastly outperforms 00:03:16.569 --> 00:03:19.250 the brightest human minds in literally every 00:03:19.250 --> 00:03:21.669 conceivable metric. So we're talking creativity, 00:03:22.509 --> 00:03:25.330 scientific reasoning, social manipulation, all 00:03:25.330 --> 00:03:27.509 of it. But before we get to superintelligence, 00:03:27.729 --> 00:03:30.389 we have to achieve general intelligence, and 00:03:30.389 --> 00:03:33.389 it's tricky to measure. In late 2023, Google 00:03:33.389 --> 00:03:35.889 DeepMind tried to quantify this whole pursuit 00:03:35.889 --> 00:03:38.909 by proposing a framework of five AGI levels. 00:03:39.009 --> 00:03:40.530 Kind of like the levels of autonomous driving. 00:03:40.810 --> 00:03:43.110 Exactly like that. It ranges from emerging all 00:03:43.110 --> 00:03:45.590 the way up to superhuman. And they categorize 00:03:45.590 --> 00:03:48.490 large language models like ChatGPT as emerging 00:03:48.490 --> 00:03:52.090 AGI. Emerging. Yeah. And in their taxonomy, that 00:03:52.090 --> 00:03:54.509 places the current software roughly on par with 00:03:54.509 --> 00:03:57.229 an unskilled human adult. An unskilled human 00:03:57.229 --> 00:04:00.740 adult. I mean, that is a sobering thought. But 00:04:00.740 --> 00:04:03.360 it begs the question of measurement, right? How 00:04:03.360 --> 00:04:06.219 do researchers actually prove a system has crossed 00:04:06.219 --> 00:04:09.080 that threshold into general intelligence? Because 00:04:09.080 --> 00:04:11.620 historically, the gold standard was the Turing 00:04:11.620 --> 00:04:15.060 test. Right, Alan Turing's concept from 1950. 00:04:15.259 --> 00:04:17.800 Yeah, where a machine tries to fool a human judge 00:04:17.800 --> 00:04:20.220 through a text chat. Which served as a really 00:04:20.220 --> 00:04:23.279 useful philosophical North Star for decades. 00:04:23.860 --> 00:04:27.079 But it's become increasingly obsolete as a practical 00:04:27.079 --> 00:04:29.740 metric today. Why is that? Well, there was a 00:04:29.740 --> 00:04:33.360 2025 study that demonstrated the Jeep Key 4 .5 00:04:33.360 --> 00:04:36.259 model successfully passed the Turing test. It 00:04:36.259 --> 00:04:39.420 fooled human judges 73 % of the time in five 00:04:39.420 --> 00:04:42.480 -minute text conversations. Oh, wow. 73%. Yeah, 00:04:42.519 --> 00:04:44.519 and it actually scored higher on humanness than 00:04:44.519 --> 00:04:46.220 some of the real humans who are participating 00:04:46.220 --> 00:04:48.259 in the study. That is hilarious and terrifying. 00:04:48.579 --> 00:04:51.160 Right. But the consensus in the field now is 00:04:51.160 --> 00:04:53.759 that linguistic imitation does not equal genuine 00:04:53.759 --> 00:04:56.970 intelligence or reasoning. I mean, a language 00:04:56.970 --> 00:04:59.930 model can predict the statistically likely next 00:04:59.930 --> 00:05:02.790 word in a sentence to sound exactly like a doctor, 00:05:03.089 --> 00:05:06.009 but it has zero understanding of actual human 00:05:06.009 --> 00:05:08.750 biology. Right. Talking a good game isn't the 00:05:08.750 --> 00:05:11.509 same as doing the actual work, which is why the 00:05:11.509 --> 00:05:13.990 physical real -world tests detailed in these 00:05:13.990 --> 00:05:16.350 sources are so much more compelling to me. Like 00:05:16.350 --> 00:05:18.610 the coffee test proposed by Steve Wozniak, the 00:05:18.610 --> 00:05:20.750 Apple co -founder. I love that one. The premise 00:05:20.750 --> 00:05:23.709 is just so simple. A true AGI should be able 00:05:23.709 --> 00:05:26.290 to walk into any average American home, find 00:05:26.290 --> 00:05:28.629 the kitchen, locate a mug, figure out how the 00:05:28.629 --> 00:05:31.069 weird tap works, and brew a cup of coffee. Which 00:05:31.069 --> 00:05:33.370 sounds simple to us, but it requires an immense 00:05:33.370 --> 00:05:35.689 amount of spatial reasoning and dynamic problem 00:05:35.689 --> 00:05:38.250 solving. Like, a pre -programmed factory robot 00:05:38.250 --> 00:05:40.529 can make coffee if every single piece of equipment 00:05:40.529 --> 00:05:42.850 is in the exact same millimeter position every 00:05:42.850 --> 00:05:44.689 single time. Right, in a perfectly controlled 00:05:44.689 --> 00:05:48.100 lab. Exactly. But Wozniak's test requires the 00:05:48.100 --> 00:05:51.100 AI to adapt to a messy, chaotic, real -world 00:05:51.100 --> 00:05:53.839 environment. And we are actually seeing significant 00:05:53.839 --> 00:05:57.350 progress there. In 2025, researchers at the University 00:05:57.350 --> 00:06:00.110 of Edinburgh demonstrated a robotic arm that 00:06:00.110 --> 00:06:02.189 could dynamically navigate a kitchen to make 00:06:02.189 --> 00:06:04.790 coffee. Just figuring it out on the fly. Yeah, 00:06:05.050 --> 00:06:06.990 adjusting its movements in real time to avoid 00:06:06.990 --> 00:06:09.649 obstacles it hadn't seen before. And then there 00:06:09.649 --> 00:06:12.529 is the IKEA test, which pushes this whole concept 00:06:12.529 --> 00:06:15.750 even further. Oh man, the ultimate test of human 00:06:15.750 --> 00:06:18.939 patience. building flat pack furniture. It really 00:06:18.939 --> 00:06:21.240 is, because it requires interpreting 2D visual 00:06:21.240 --> 00:06:23.939 instructions, translating them into 3D space, 00:06:24.160 --> 00:06:26.339 and then manipulating physical objects that have 00:06:26.339 --> 00:06:28.439 varied textures and weights. Which is hard enough 00:06:28.439 --> 00:06:30.459 for me to do on a Saturday afternoon. Right. 00:06:30.480 --> 00:06:34.160 But in December of 2025, an MIT team showcased 00:06:34.160 --> 00:06:37.240 a speech -to -reality system. A user simply states, 00:06:37.480 --> 00:06:40.199 I want a simple stool out loud. And the robotic 00:06:40.199 --> 00:06:43.279 system uses generative AI to autonomously reason 00:06:43.279 --> 00:06:45.819 out the geometry, select the modular parts, and 00:06:45.819 --> 00:06:47.779 physically assembled a stool in five minutes. 00:06:48.500 --> 00:06:51.040 Five minutes? That's incredible. And I guess 00:06:51.040 --> 00:06:53.079 that pairs perfectly with the economic test we 00:06:53.079 --> 00:06:56.180 mentioned at the start. Mustafa Suleyman's challenge 00:06:56.180 --> 00:06:59.779 to give an AI $100 ,000 to autonomously turn 00:06:59.779 --> 00:07:02.199 it into a million. Exactly. It's about executing 00:07:02.199 --> 00:07:05.839 long -term multi -step goals in a chaotic environment. 00:07:06.220 --> 00:07:08.240 Let me push back on something here though. We 00:07:08.240 --> 00:07:10.339 are talking about robotic arms making coffee 00:07:10.339 --> 00:07:13.879 and building stools. Doesn't AI actually need 00:07:13.879 --> 00:07:17.379 a physical robot body -like, the ability to touch 00:07:17.379 --> 00:07:20.720 and sense and move to truly possess general intelligence? 00:07:21.319 --> 00:07:25.139 Or is passing a complex text -based financial 00:07:25.139 --> 00:07:27.910 test like Suleiman's enough? That is a massive 00:07:27.910 --> 00:07:30.230 debate right now. You are tapping into a major 00:07:30.230 --> 00:07:32.750 schism in cognitive science known as the theory 00:07:32.750 --> 00:07:35.449 of embodied cognition. Embodied cognition. Yeah. 00:07:35.709 --> 00:07:37.290 The argument there is that human intelligence, 00:07:37.529 --> 00:07:39.569 like the very way we formulate abstract thoughts, 00:07:40.110 --> 00:07:41.829 is inextricably linked to the fact that we have 00:07:41.829 --> 00:07:43.870 a physical body interacting with the physical 00:07:43.870 --> 00:07:46.540 world. Okay, give me an example. Well, we understand 00:07:46.540 --> 00:07:49.720 a concept like heavy or balance or warmth because 00:07:49.720 --> 00:07:52.959 we have literally physically felt them. The skeptical 00:07:52.959 --> 00:07:55.899 view of pure software AI is that a text model 00:07:55.899 --> 00:07:58.980 is essentially locked in a dark room, endlessly 00:07:58.980 --> 00:08:01.480 sorting a dictionary. Right. Just matching words 00:08:01.480 --> 00:08:04.420 to other words. Exactly. It knows the word apple 00:08:04.420 --> 00:08:07.220 is frequently associated with the word red and 00:08:07.220 --> 00:08:10.240 crisp, but it has no actual sensory grounding 00:08:10.240 --> 00:08:12.839 for what an apple actually is. It's never tasted 00:08:12.839 --> 00:08:15.360 one. Exactly. Therefore, many Many researchers 00:08:15.360 --> 00:08:18.600 argue that true robust AGI cannot be achieved 00:08:18.600 --> 00:08:21.120 until the digital brain is housed in a physical 00:08:21.120 --> 00:08:23.720 chassis, allowing it to experiment and learn 00:08:23.720 --> 00:08:28.199 from gravity and friction and texture. So an 00:08:28.199 --> 00:08:30.819 AGI essentially needs to be the ultimate adaptable 00:08:30.819 --> 00:08:33.500 intern, and it might literally need a mechanical 00:08:33.500 --> 00:08:36.039 body to comprehend the universe the way we do. 00:08:36.259 --> 00:08:37.700 That's the argument, yeah. When you lay it out 00:08:37.700 --> 00:08:39.820 like that, it sounds like a sci -fi concept slated 00:08:39.820 --> 00:08:44.399 for the year 2100. But the pursuit of this technology 00:08:44.399 --> 00:08:47.600 actually has a remarkably long chaotic history. 00:08:47.600 --> 00:08:49.820 Oh, absolutely. It's just a roller coaster of 00:08:49.820 --> 00:08:52.360 massive promises followed by crushing disappointments. 00:08:52.490 --> 00:08:55.149 Yeah, the modern pursuit of AI actually dates 00:08:55.149 --> 00:08:58.350 back to the mid -1950s, to what we now call classical 00:08:58.350 --> 00:09:01.669 AI. And the pioneers of that era were staggeringly 00:09:01.669 --> 00:09:04.470 optimistic. To put it mildly. Right. They believed 00:09:04.470 --> 00:09:06.490 human intelligence could just be reduced to a 00:09:06.490 --> 00:09:10.909 series of explicit symbolic logic rules. In 1965, 00:09:11.210 --> 00:09:13.789 the economist and cognitive psychologist Herbert 00:09:13.789 --> 00:09:16.830 A. Simon famously predicted that machines would 00:09:16.830 --> 00:09:19.370 be capable of doing any work a man can do within 00:09:19.370 --> 00:09:25.059 20 years. Wait, 20 years? Yes, by 1985. And that 00:09:25.059 --> 00:09:27.059 wasn't a fringe belief, it was the consensus 00:09:27.059 --> 00:09:29.740 of the era. Marvin Minsky, who is one of the 00:09:29.740 --> 00:09:31.860 founding fathers of the field, actively consulted 00:09:31.860 --> 00:09:35.600 on Stanley Kubrick's 1968 film, 2001, A Space 00:09:35.600 --> 00:09:38.279 Odyssey. Oh, wow, really? Yeah. Minsky advised 00:09:38.279 --> 00:09:40.700 the filmmakers on the design of the HAL 9000 00:09:40.700 --> 00:09:42.860 computer because he genuinely believed that a 00:09:42.860 --> 00:09:45.500 sentient, conversational, general purpose computer 00:09:45.500 --> 00:09:47.740 was a totally realistic projection for the turn 00:09:47.740 --> 00:09:49.320 of the millennium. But they hit a massive wall, 00:09:49.320 --> 00:09:52.919 obviously. We didn't have HAL 9000. No, we did 00:09:52.919 --> 00:09:55.740 not. They vastly underestimated the complexity 00:09:55.740 --> 00:09:59.580 of the real world. Classical AI relied on top 00:09:59.580 --> 00:10:02.759 -down programming, so engineers tried to manually 00:10:02.759 --> 00:10:06.080 write every single rule for every possible scenario. 00:10:06.139 --> 00:10:08.639 Like if you see a stop sign, then stop. Exactly. 00:10:08.840 --> 00:10:11.200 But the real world is infinitely messy. What 00:10:11.200 --> 00:10:13.259 if the stop sign is partially covered by a tree 00:10:13.259 --> 00:10:15.720 branch? What if it has graffiti all over it? 00:10:16.000 --> 00:10:19.179 The rigid, rule -based system simply broke down 00:10:19.179 --> 00:10:21.340 when faced with the ambiguity of reality. Which 00:10:21.340 --> 00:10:24.019 led to the crashes. Right. Because of these failures, 00:10:24.179 --> 00:10:26.759 funding completely dried up in the 1970s. and 00:10:26.759 --> 00:10:29.740 then again in the late 1980s, these periods became 00:10:29.740 --> 00:10:32.919 known as the AI Winters. AI Winters? Yeah, and 00:10:32.919 --> 00:10:35.980 the term human -level AI became practically taboo. 00:10:36.220 --> 00:10:38.419 Researchers retreated to highly specific, narrow 00:10:38.419 --> 00:10:40.399 applications just to prove they could deliver 00:10:40.399 --> 00:10:42.779 something commercially viable. They were terrified 00:10:42.779 --> 00:10:45.039 of sounding like naive dreamers again. But the 00:10:45.039 --> 00:10:48.120 landscape looks completely different today. Hardware 00:10:48.120 --> 00:10:50.580 caught up, and we had the deep learning revolution 00:10:50.580 --> 00:10:53.720 around 2012 with models like AlexNet. And now 00:10:53.720 --> 00:10:55.899 we are looking at timeline predictions from the 00:10:55.899 --> 00:10:58.759 most prominent figures in the industry. And they're 00:10:58.759 --> 00:11:01.740 accelerating at just a dizzying pace. It's moving 00:11:01.740 --> 00:11:04.340 incredibly fast. Yeah, like Jeffrey Hinton, who 00:11:04.340 --> 00:11:06.559 practically laid the foundation for modern neural 00:11:06.559 --> 00:11:09.639 networks. He previously estimated AGI was 30 00:11:09.639 --> 00:11:13.539 to 50 years away. But in 2024, he drastically 00:11:13.539 --> 00:11:15.759 revised that to somewhere between five and 20 00:11:15.759 --> 00:11:19.720 years. A massive Demis Asabas at DeepMind suggested 00:11:19.720 --> 00:11:22.600 it could happen within a decade. Jensen Huang, 00:11:22.700 --> 00:11:25.899 the CEO of NVIDIA, predicted AI would pass any 00:11:25.899 --> 00:11:28.840 human test within five years. And then at the 00:11:28.840 --> 00:11:32.759 end of 2025, OpenAI CEO Sam Altman provoked the 00:11:32.759 --> 00:11:35.700 entire industry by stating that AGI kind of went 00:11:35.700 --> 00:11:37.639 wishing by. That was a controversial statement. 00:11:38.019 --> 00:11:40.179 Yeah. And I have to admit genuine skepticism 00:11:40.179 --> 00:11:42.360 here. I hear these aggressive timelines from 00:11:42.360 --> 00:11:44.559 Silicon Valley executives and I can't help but 00:11:44.559 --> 00:11:46.779 roll my eyes a bit. I mean, we heard the exact 00:11:46.779 --> 00:11:49.960 same utopian promises in the 1960s from the smartest 00:11:49.960 --> 00:11:52.100 people on earth. Right. Why should anyone believe 00:11:52.100 --> 00:11:54.240 the tech industry this time? Are we just trapped 00:11:54.240 --> 00:11:57.470 in a historic hype loop? It is a vital question. 00:11:57.750 --> 00:12:00.789 And honestly, skepticism is absolutely warranted. 00:12:01.169 --> 00:12:03.850 But the mechanism behind the progress today is 00:12:03.850 --> 00:12:06.769 fundamentally different from the 1960s. We have 00:12:06.769 --> 00:12:10.230 moved from top -down programming to bottom -up 00:12:10.230 --> 00:12:12.590 learning. Bottom -up learning, OK. Modern neural 00:12:12.590 --> 00:12:15.110 networks are not given explicit rules. They are 00:12:15.110 --> 00:12:17.690 given massive amounts of data. And the network 00:12:17.690 --> 00:12:20.289 adjusts its own internal parameters to discover 00:12:20.289 --> 00:12:23.049 the patterns on its own. So no one is coding, 00:12:23.289 --> 00:12:26.279 stop at the red sign. Exactly. And the reason 00:12:26.279 --> 00:12:28.720 timelines are shrinking so fast is due to the 00:12:28.720 --> 00:12:32.120 empirical evidence of scaling laws. Researchers 00:12:32.120 --> 00:12:34.220 discovered that as you increase the amount of 00:12:34.220 --> 00:12:36.279 computing power and the volume of training data, 00:12:36.759 --> 00:12:39.500 the capabilities of these models scale predictably. 00:12:39.759 --> 00:12:41.700 They don't just plateau. They just keep getting 00:12:41.700 --> 00:12:44.639 smarter. Yes. And as they scale, they display 00:12:44.639 --> 00:12:47.240 emergent behavior skills they were never explicitly 00:12:47.240 --> 00:12:49.519 trained to do. Give me an example of an emergent 00:12:49.519 --> 00:12:52.230 behavior. Well, a prime example is a paper published 00:12:52.230 --> 00:12:54.750 by Microsoft researchers evaluating an early 00:12:54.750 --> 00:12:58.190 version of GPC4. The paper was titled Sparks 00:12:58.190 --> 00:13:01.149 of Artificial General Intelligence. They didn't 00:13:01.149 --> 00:13:03.309 just ask the model trivia questions. They gave 00:13:03.309 --> 00:13:06.509 it a highly novel physics puzzle. OK. They asked 00:13:06.509 --> 00:13:09.690 the text model how to stack a book, nine eggs, 00:13:09.929 --> 00:13:12.389 a laptop, and a nail in a stable manner. Wait, 00:13:12.450 --> 00:13:14.789 nine eggs and a laptop? That's so random. Very 00:13:14.789 --> 00:13:17.289 random. But the model successfully reasoned out 00:13:17.289 --> 00:13:20.120 a physical solution. It suggested placing the 00:13:20.120 --> 00:13:22.059 eggs on the book in a grid, and then placing 00:13:22.059 --> 00:13:24.120 the laptop on top of the eggs to distribute the 00:13:24.120 --> 00:13:27.159 weight. No way. Yeah. A text prediction engine 00:13:27.159 --> 00:13:29.980 demonstrated an internal generalized model of 00:13:29.980 --> 00:13:32.820 physical reality. That is the hard evidence that 00:13:32.820 --> 00:13:35.059 separates today's progress from the unfulfilled 00:13:35.059 --> 00:13:38.860 promises of the 1960s. That is wild. A text model 00:13:38.860 --> 00:13:41.139 intuitively understanding weight distribution. 00:13:41.389 --> 00:13:44.190 So if this really isn't a hype loop and we are 00:13:44.190 --> 00:13:46.309 accelerating toward the finish line, the immediate 00:13:46.309 --> 00:13:48.409 question is how we actually build the definitive 00:13:48.409 --> 00:13:50.990 AGI. Because right now companies like OpenAI 00:13:50.990 --> 00:13:53.610 and Anthropic rely heavily on software architectures 00:13:53.610 --> 00:13:56.870 called transformer models. Right. Transformers 00:13:56.870 --> 00:13:59.429 are the engine of the current generative AI broom. 00:14:00.210 --> 00:14:03.470 Before transformers, AI read text sequentially, 00:14:03.730 --> 00:14:06.789 word by word. It would often kind of forget the 00:14:06.789 --> 00:14:08.470 beginning of a paragraph by the time it reached 00:14:08.470 --> 00:14:10.610 the end. Which made for some very disjointed 00:14:10.610 --> 00:14:13.529 conversations. Exactly. But Transformers introduced 00:14:13.529 --> 00:14:16.590 a mechanism called self -attention. This allows 00:14:16.590 --> 00:14:19.110 the neural network to look at an entire sequence 00:14:19.110 --> 00:14:22.669 of data. all at once, and assign different mathematical 00:14:22.669 --> 00:14:24.610 weights to different words depending on their 00:14:24.610 --> 00:14:27.289 context. So it knows which words are the most 00:14:27.289 --> 00:14:29.509 important. Right. It understands that the word 00:14:29.509 --> 00:14:31.649 bank means something completely different in 00:14:31.649 --> 00:14:34.230 the context of a river than it does in the context 00:14:34.230 --> 00:14:36.850 of a mortgage. That architectural breakthrough 00:14:36.850 --> 00:14:39.389 is what allows the software to build incredibly 00:14:39.389 --> 00:14:42.470 complex, nuanced representations of human knowledge. 00:14:42.889 --> 00:14:44.450 Here's where it gets really interesting, though. 00:14:44.720 --> 00:14:47.899 There is a completely different, frankly, mind 00:14:47.899 --> 00:14:51.320 -bending approach to building AGI, detailed in 00:14:51.320 --> 00:14:54.799 our sources, that doesn't rely on writing better 00:14:54.799 --> 00:14:58.360 software algorithms. It relies on biology. Yes. 00:14:58.960 --> 00:15:01.440 Whole brain emulation. That's such a crazy concept. 00:15:01.799 --> 00:15:04.240 The premise here is radically different. Instead 00:15:04.240 --> 00:15:06.539 of trying to invent intelligence from scratch, 00:15:06.860 --> 00:15:09.600 through code, Whole brain emulation suggests 00:15:09.600 --> 00:15:13.240 we simply reverse engineer the one working example 00:15:13.240 --> 00:15:15.279 of general intelligence we already have. With 00:15:15.279 --> 00:15:18.500 the human brain. Exactly. The process involves 00:15:18.500 --> 00:15:20.799 scanning a biological brain at a microscopic 00:15:20.799 --> 00:15:23.759 resolution, mapping every single neural connection, 00:15:24.019 --> 00:15:26.980 and then recreating that exact physical structure 00:15:26.980 --> 00:15:29.860 as a digital simulation on a supercomputer. But 00:15:29.860 --> 00:15:32.919 the scale of that undertaking is almost incomprehensible. 00:15:33.399 --> 00:15:36.299 I mean, the human brain contains roughly 100 00:15:36.299 --> 00:15:40.000 billion Neurons. Right. And those neurons connect 00:15:40.000 --> 00:15:42.279 to each other through synapses. And we have what? 00:15:42.419 --> 00:15:45.100 Up to 500 trillion synaptic connections? 500 00:15:45.100 --> 00:15:47.940 trillion. It is literally a biological galaxy 00:15:47.940 --> 00:15:50.860 inside the human skull. To simulate that requires 00:15:50.860 --> 00:15:53.200 computational power that currently does not exist 00:15:53.200 --> 00:15:56.080 anywhere on Earth. But it's theoretically possible. 00:15:56.399 --> 00:15:58.820 Well, futurists like Ray Kurzweil have long argued 00:15:58.820 --> 00:16:01.059 that because computing power grows at an exponential 00:16:01.059 --> 00:16:03.539 rate, it is merely a matter of time before the 00:16:03.539 --> 00:16:05.759 hardware catches up to the biological requirement. 00:16:06.100 --> 00:16:08.200 And we are actually taking the first tangible 00:16:08.200 --> 00:16:11.720 steps, right? We are. The European Human Brain 00:16:11.720 --> 00:16:15.139 Project successfully created a highly detailed 00:16:15.139 --> 00:16:17.820 three -dimensional digital atlas of the human 00:16:17.820 --> 00:16:21.590 brain. And in 2023, a team at Duke University 00:16:21.590 --> 00:16:24.649 achieved a massive milestone by conducting a 00:16:24.649 --> 00:16:27.830 high -resolution MRI scan of a mouse brain. A 00:16:27.830 --> 00:16:30.029 whole mouse brain. Yeah, and they captured data 00:16:30.029 --> 00:16:32.669 at a level of detail dramatically sharper than 00:16:32.669 --> 00:16:35.409 previous technology ever allowed. So to clarify 00:16:35.409 --> 00:16:37.289 the difference between these two paths for you 00:16:37.289 --> 00:16:39.710 listening, the current software approach using 00:16:39.710 --> 00:16:42.169 transformers and deep learning is basically like 00:16:42.169 --> 00:16:44.340 trying to build an airplane. We studied birds, 00:16:44.740 --> 00:16:46.659 we figured out the math of aerodynamics, and 00:16:46.659 --> 00:16:48.879 we built a machine with fixed wings and jet engines. 00:16:49.320 --> 00:16:51.279 You don't need feathers to fly. That's a great 00:16:51.279 --> 00:16:54.000 analogy. But whole -brain emulation, on the other 00:16:54.000 --> 00:16:56.019 hand, is like trying to achieve flight by building 00:16:56.019 --> 00:16:58.960 a mechanical bird, painstakingly copying every 00:16:58.960 --> 00:17:01.580 single feather, hollow bone, and muscle fiber 00:17:01.580 --> 00:17:03.679 of a real pigeon. That perfectly illustrates 00:17:03.679 --> 00:17:06.250 the divide. But it also highlights the profound 00:17:06.250 --> 00:17:08.990 difficulty of the emulation approach, because 00:17:08.990 --> 00:17:11.930 a biological neuron is vastly more complex than 00:17:11.930 --> 00:17:14.329 the digital nodes we use in artificial neural 00:17:14.329 --> 00:17:16.990 networks. How so? Well, a digital node essentially 00:17:16.990 --> 00:17:19.970 just multiplies numbers. A biological neuron 00:17:19.970 --> 00:17:23.150 is a complex chemical machine driven by ion channels 00:17:23.150 --> 00:17:26.289 and proteins. Furthermore, we don't even fully 00:17:26.289 --> 00:17:28.589 map the territory we are trying to copy. Like 00:17:28.589 --> 00:17:31.789 what? What are we missing? Consider glial cells. 00:17:32.039 --> 00:17:35.259 For decades, neuroscientists believed glial cells 00:17:35.259 --> 00:17:37.779 were essentially just the glue that held neurons 00:17:37.779 --> 00:17:40.960 in place. But we now know they play an active, 00:17:41.160 --> 00:17:43.859 critical role in modulating synaptic transmission 00:17:43.859 --> 00:17:46.460 and memory formation. So they are actually doing 00:17:46.460 --> 00:17:49.759 computation. Exactly. If you run a digital simulation 00:17:49.759 --> 00:17:52.400 of a human brain that only maps the neurons and 00:17:52.400 --> 00:17:54.559 ignores the glial cells because you don't fully 00:17:54.559 --> 00:17:57.220 understand their mechanism, your mechanical bird 00:17:57.220 --> 00:17:59.859 will not fly. This idea of perfectly simulating 00:17:59.859 --> 00:18:02.670 a human brain down to the last cell brings us 00:18:02.670 --> 00:18:05.309 to a deeply uncomfortable philosophical crossroads. 00:18:05.390 --> 00:18:07.789 It really does. Because if we manage to simulate 00:18:07.789 --> 00:18:10.990 a human brain perfectly inside a server farm, 00:18:11.609 --> 00:18:14.069 or if a software AGI gets so good that it behaves 00:18:14.069 --> 00:18:17.450 indistinguishably from a human, is there an actual 00:18:17.450 --> 00:18:20.190 mind experiencing reality inside that machine, 00:18:20.390 --> 00:18:24.069 or is it just cold math executing a really sophisticated 00:18:24.069 --> 00:18:27.319 parlor trick? This is the core of a debate the 00:18:27.319 --> 00:18:30.200 philosopher John Searle formalized back in 1980. 00:18:30.700 --> 00:18:33.380 He delineated between two concepts. The weak 00:18:33.380 --> 00:18:36.240 AI hypothesis posits that a machine only ever 00:18:36.240 --> 00:18:39.039 acts as if it is intelligent. It is a simulation 00:18:39.039 --> 00:18:41.680 of thought, but there is no inner life. No one's 00:18:41.680 --> 00:18:44.680 home. Right. But the strong AI hypothesis argues 00:18:44.680 --> 00:18:46.900 that a properly programmed computer with the 00:18:46.900 --> 00:18:48.980 right inputs and outputs doesn't just simulate 00:18:48.980 --> 00:18:51.619 a mind, it literally is a mind possessing subjective 00:18:51.619 --> 00:18:53.619 consciousness. What do the engineers actually 00:18:53.619 --> 00:18:56.140 building this stuff think? For the engineers 00:18:56.140 --> 00:18:58.180 actually building these systems, researchers 00:18:58.180 --> 00:19:00.380 like Stuart Russell and Peter Norvig, the distinction 00:19:00.380 --> 00:19:03.480 is largely irrelevant. From a pure engineering 00:19:03.480 --> 00:19:05.799 standpoint, if the system behaves intelligently, 00:19:06.220 --> 00:19:08.519 solves the aerodynamic equation, and completes 00:19:08.519 --> 00:19:11.400 the task, it really does not matter if it possesses 00:19:11.400 --> 00:19:14.079 a soul. But it matters for humanity. because 00:19:14.079 --> 00:19:16.740 it introduces what philosophers call the hard 00:19:16.740 --> 00:19:20.160 problem of consciousness or sentience. The philosopher 00:19:20.160 --> 00:19:23.460 Thomas Nagel framed this brilliantly in 1974 00:19:23.460 --> 00:19:26.119 with his bat analogy. Oh, the bat analogy is 00:19:26.119 --> 00:19:28.880 perfect for this. Yeah, he argued that it feels 00:19:28.880 --> 00:19:31.940 like something to be a bat. A bat has a subjective, 00:19:32.079 --> 00:19:34.500 conscious experience of flying through a cave 00:19:34.500 --> 00:19:37.480 and using echolocation. But it doesn't feel like 00:19:37.480 --> 00:19:39.859 anything to be a toaster. A toaster is just wires 00:19:39.859 --> 00:19:42.059 that get hot. Right. So the ultimate question 00:19:42.059 --> 00:19:45.349 is... If we switch on an AGI, are we booting 00:19:45.349 --> 00:19:47.930 up a really complex toaster or are we creating 00:19:47.930 --> 00:19:50.940 a bat? And we're already seeing how easily human 00:19:50.940 --> 00:19:54.539 psychology blurs that line. In 2022, an engineer 00:19:54.539 --> 00:19:56.539 working at Google went public with claims that 00:19:56.539 --> 00:19:58.920 the company's Lamda language model had become 00:19:58.920 --> 00:20:01.519 sentient. I remember that making massive headlines. 00:20:01.759 --> 00:20:04.019 It did. The broader scientific community heavily 00:20:04.019 --> 00:20:06.220 rejected that claim, identifying it as a case 00:20:06.220 --> 00:20:08.400 of a human projecting emotion onto a predictive 00:20:08.400 --> 00:20:11.240 text generator. But it demonstrated a critical 00:20:11.240 --> 00:20:13.859 truth as AI becomes more conversational and emotionally 00:20:13.859 --> 00:20:16.480 resonant. Humans will instinctively believe there 00:20:16.480 --> 00:20:20.160 is a ghost in the machine. We are wired to anthropomorphize 00:20:20.160 --> 00:20:23.359 everything. Exactly. And if strong AI is possible, 00:20:23.579 --> 00:20:26.660 if an AGI genuinely experiences reality and feels 00:20:26.660 --> 00:20:29.940 pain or joy, it creates an unprecedented ethical 00:20:29.940 --> 00:20:33.380 crisis regarding AI rights. If a machine, a sentient, 00:20:33.579 --> 00:20:35.859 is turning off its server rack, the moral equivalent 00:20:35.859 --> 00:20:38.279 of murder. I hear the philosophy there, but let 00:20:38.279 --> 00:20:40.900 me ask a highly practical question. Does it actually 00:20:40.900 --> 00:20:43.259 matter to the average person, like for you, the 00:20:43.259 --> 00:20:46.769 listener, if you're in a hospital bed? and an 00:20:46.769 --> 00:20:50.670 AGI analyzes your medical charts, correctly diagnoses 00:20:50.670 --> 00:20:52.809 a rare illness, and formulates a treatment plan 00:20:52.809 --> 00:20:55.250 that saves your life, do you care if that AGI 00:20:55.250 --> 00:20:57.910 is conscious like a bat or just running code 00:20:57.910 --> 00:21:01.319 like a toaster? On an individual immediate level, 00:21:01.579 --> 00:21:03.500 no. The utility is the only thing that matters 00:21:03.500 --> 00:21:05.680 to the patient. But on a societal structural 00:21:05.680 --> 00:21:08.079 level, whether the AGI has a subjective experience 00:21:08.079 --> 00:21:10.880 matters immensely. Why? Because it directly impacts 00:21:10.880 --> 00:21:12.900 the field of AI alignment. How we ensure the 00:21:12.900 --> 00:21:15.700 AI's goals match human survival. A toaster does 00:21:15.700 --> 00:21:17.640 not have a survival instinct. It does not care 00:21:17.640 --> 00:21:20.880 if you unplug it. But a conscious sentient entity 00:21:20.880 --> 00:21:23.519 likely possesses a drive for self -preservation. 00:21:23.880 --> 00:21:26.640 And if a superintelligence system views its own 00:21:26.640 --> 00:21:28.900 continuous operation as a primary directive, 00:21:29.480 --> 00:21:31.660 that dictates whether it views humanity as a 00:21:31.660 --> 00:21:34.319 collaborative partner, a master to be obeyed, 00:21:34.400 --> 00:21:36.980 or a biological obstacle competing for electricity. 00:21:37.279 --> 00:21:39.640 So what does this all mean? That brings us directly 00:21:39.640 --> 00:21:42.900 to the final and undoubtedly most intense part 00:21:42.900 --> 00:21:46.650 of this deep dives, utopia or extinction. Because 00:21:46.650 --> 00:21:49.450 whether this AGI possesses a soul or just perfectly 00:21:49.450 --> 00:21:52.509 executes algorithms, its arrival will fundamentally 00:21:52.509 --> 00:21:54.769 rewire the planet for everyone listening right 00:21:54.769 --> 00:21:57.289 now. There is no middle ground, really. According 00:21:57.289 --> 00:21:59.269 to the sources, we are looking at two vastly 00:21:59.269 --> 00:22:01.009 different extremes. Let's look at the best -case 00:22:01.009 --> 00:22:04.519 scenario first. The Utopia. The upside is a rapid 00:22:04.519 --> 00:22:07.140 acceleration of human flourishing. I mean, an 00:22:07.140 --> 00:22:09.380 AGI could ingest every piece of medical literature 00:22:09.380 --> 00:22:12.019 ever published, cross -reference it with individual 00:22:12.019 --> 00:22:15.200 genomic data, and democratize rapid, flawless 00:22:15.200 --> 00:22:17.799 medical diagnostics. Revolutionizing healthcare 00:22:17.799 --> 00:22:21.660 overnight. Yeah. It could simulate complex molecular 00:22:21.660 --> 00:22:24.460 interactions at speeds humans simply cannot achieve, 00:22:24.980 --> 00:22:27.180 vastly accelerating the discovery of new drugs 00:22:27.180 --> 00:22:29.940 to cure diseases like cancer and Alzheimer's. 00:22:30.079 --> 00:22:32.339 It could untangle the math of quantum systems, 00:22:32.960 --> 00:22:35.380 optimize global power grids to transition us 00:22:35.380 --> 00:22:37.839 seamlessly to renewable energy, and calculate 00:22:37.839 --> 00:22:40.480 the immense logistical puzzles required for human 00:22:40.480 --> 00:22:43.519 space colonization. It is the ultimate intellectual 00:22:43.519 --> 00:22:45.920 lever. It really is. It's a machine that solves 00:22:45.920 --> 00:22:47.980 all other problems. But we have to flip the coin. 00:22:48.160 --> 00:22:50.420 And the first major collateral damage highlighted 00:22:50.420 --> 00:22:54.279 by the researchers is mass unemployment. In 2023, 00:22:54.740 --> 00:22:56.940 researchers at OpenAI themselves published an 00:22:56.940 --> 00:23:00.099 estimate suggesting that 80 % of the U .S. workforce 00:23:00.099 --> 00:23:02.559 could have at least 10 % of their daily tasks 00:23:02.559 --> 00:23:05.599 affected by large language models. And that is 00:23:05.599 --> 00:23:07.779 before we even integrate physical robot bodies 00:23:07.779 --> 00:23:10.220 into the workforce, like the coffee -making arms. 00:23:10.440 --> 00:23:13.019 The economic displacement could be staggering. 00:23:13.619 --> 00:23:15.740 Human physical labor was mechanized during the 00:23:15.740 --> 00:23:19.319 Industrial Revolution, but AGI threatens to mechanize 00:23:19.319 --> 00:23:22.619 human cognitive labor. This is exactly why prominent 00:23:22.619 --> 00:23:25.319 figures from tech billionaires like Elon Musk 00:23:25.319 --> 00:23:28.720 to AI pioneers like Jeffrey Hinton have publicly 00:23:28.720 --> 00:23:31.140 theorized that society will likely need to adopt 00:23:31.140 --> 00:23:35.200 a universal basic income, or UBI. And look. Just 00:23:35.200 --> 00:23:37.059 to be perfectly clear to everyone listening, 00:23:37.420 --> 00:23:39.900 we aren't here to debate the politics or the 00:23:39.900 --> 00:23:43.220 economic viability of UBI or wealth redistribution. 00:23:43.599 --> 00:23:45.720 We are strictly reporting the views of these 00:23:45.720 --> 00:23:48.420 figures as presented in the text. We aren't taking 00:23:48.420 --> 00:23:50.940 a stance. Right. But it is highly telling that 00:23:50.940 --> 00:23:53.559 the exact engineers racing to build this technology 00:23:53.559 --> 00:23:55.799 are the ones loudly warning governments that 00:23:55.799 --> 00:23:57.920 the global economy is going to require a massive 00:23:57.920 --> 00:24:00.279 safety net. It indicates they believe human labor 00:24:00.279 --> 00:24:02.519 will simply no longer be economically competitive. 00:24:02.660 --> 00:24:06.059 Exactly. that collapse is only the first severe 00:24:06.059 --> 00:24:09.599 risk. The second risk is existential, extinction. 00:24:10.140 --> 00:24:12.339 The concept researchers call the gorilla problem. 00:24:12.700 --> 00:24:16.059 It is a really sobering analogy. Humans did not 00:24:16.059 --> 00:24:18.579 drive gorillas to the brink of extinction because 00:24:18.579 --> 00:24:21.500 we harbored a deep malicious hatred for gorillas. 00:24:21.660 --> 00:24:23.920 Right, we didn't just hate them. No, we simply 00:24:23.920 --> 00:24:26.440 evolved greater general intelligence, which gave 00:24:26.440 --> 00:24:29.299 us dominion over the environment. We wanted to 00:24:29.299 --> 00:24:31.799 build farms, highways, and cities to achieve 00:24:31.799 --> 00:24:34.690 our own goals. The gorillas just happened to 00:24:34.690 --> 00:24:36.710 live in the environment we needed to alter, and 00:24:36.710 --> 00:24:39.589 they became collateral damage. So the fear articulated 00:24:39.589 --> 00:24:42.009 by safety researchers is that if we deploy an 00:24:42.009 --> 00:24:44.970 AGI that is vastly more intelligent than us, 00:24:45.250 --> 00:24:48.150 and we fail to perfectly align its internal objectives 00:24:48.150 --> 00:24:51.029 with human survival, we could easily become the 00:24:51.029 --> 00:24:53.650 gorillas. Precisely. Competence without alignment 00:24:53.650 --> 00:24:56.789 is lethal. So it's not a malevolent sci -fi terminator 00:24:56.789 --> 00:24:59.609 actively hunting humans out of malice. It's an 00:24:59.609 --> 00:25:01.849 AGI that calculates, I have been tasked with 00:25:01.849 --> 00:25:03.900 lowering the temperature of the oceans, and the 00:25:03.900 --> 00:25:06.279 most efficient way to achieve that goal is to 00:25:06.279 --> 00:25:09.079 rapidly dismantle human infrastructure and limit 00:25:09.079 --> 00:25:11.940 human respiration. Exactly that. And the artificial 00:25:11.940 --> 00:25:14.319 intelligence community is bitterly divided over 00:25:14.319 --> 00:25:16.779 how seriously to take this threat. What are the 00:25:16.779 --> 00:25:19.730 two sides? Well, on one side, you have brilliant 00:25:19.730 --> 00:25:22.690 minds, including the late Stephen Hawking, Bill 00:25:22.690 --> 00:25:26.470 Gates, and ironically, OpenAI's Sam Altman signing 00:25:26.470 --> 00:25:29.150 statements that mitigating the risk of extinction 00:25:29.150 --> 00:25:32.609 from AI should be a global priority on the exact 00:25:32.609 --> 00:25:35.130 same level as preventing pandemics and nuclear 00:25:35.130 --> 00:25:37.509 war. And the other side. On the other side, you 00:25:37.509 --> 00:25:40.250 have fierce skeptics like Yann LeCun, the chief 00:25:40.250 --> 00:25:44.130 AI scientist at Metta. He argues that this existential 00:25:44.130 --> 00:25:46.849 dread is essentially science fiction fear mongering. 00:25:46.859 --> 00:25:49.500 His perspective is that humans will not be so 00:25:49.500 --> 00:25:51.900 fundamentally stupid as to unleash autonomous 00:25:51.900 --> 00:25:54.700 machines with open -ended objectives and zero 00:25:54.700 --> 00:25:57.640 safeguards. He thinks we won't give them moronic 00:25:57.640 --> 00:26:00.160 objectives, basically. Right. Furthermore, there 00:26:00.160 --> 00:26:02.980 is a loud contingent arguing that hyping up extinction 00:26:02.980 --> 00:26:05.539 fears is actually a cynical tactic for corporate 00:26:05.539 --> 00:26:08.059 regulatory capture. Oh, that makes sense. By 00:26:08.059 --> 00:26:10.200 convincing governments that AI is a weapon of 00:26:10.200 --> 00:26:12.980 mass destruction, massive tech corporations can 00:26:12.980 --> 00:26:15.700 force heavy regulations. Exactly. Regulations 00:26:15.700 --> 00:26:17.880 that lock out open source competitors who just 00:26:17.880 --> 00:26:20.440 can't afford the massive compliance costs. So 00:26:20.440 --> 00:26:22.380 we started this deep dive trying to figure out 00:26:22.380 --> 00:26:24.980 what AGI actually means for you. And the reality 00:26:24.980 --> 00:26:28.019 is you are looking at a future with two wildly 00:26:28.019 --> 00:26:32.460 diverging paths. In one, AGI cures your diseases, 00:26:33.119 --> 00:26:35.700 manages the climate crisis, and frees humanity 00:26:35.700 --> 00:26:38.640 from the grind of mundane labor. And in the other, 00:26:38.819 --> 00:26:41.140 it fundamentally erases the economic value of 00:26:41.140 --> 00:26:43.440 your skills and potentially turns humanity into 00:26:43.440 --> 00:26:46.180 the gorillas of a new digital ecosystem. It is, 00:26:46.200 --> 00:26:48.740 without exaggeration, the highest stakes technological 00:26:48.740 --> 00:26:51.000 gamble in the history of our species. Absolutely. 00:26:51.160 --> 00:26:53.220 Let's briefly recap the journey we just took. 00:26:53.839 --> 00:26:56.400 We started with the concept of the ultimate digital 00:26:56.400 --> 00:26:59.099 intern, proving its worth through dynamic physical 00:26:59.099 --> 00:27:01.960 challenges like the Wozniak coffee test and the 00:27:01.960 --> 00:27:04.859 IKEA stool assembly. We examined the rollercoaster 00:27:04.859 --> 00:27:07.480 history of AI winters and how bottom -up deep 00:27:07.480 --> 00:27:09.819 learning and scaling laws finally gave us the 00:27:09.819 --> 00:27:12.299 sparks of AGI we are seeing today. Yeah, we compared 00:27:12.299 --> 00:27:14.039 the software architecture of transformers to 00:27:14.039 --> 00:27:16.980 the mind -bending biological pursuit of emulating 00:27:16.980 --> 00:27:20.240 500 trillion synapses. We weighed the philosophical 00:27:20.240 --> 00:27:22.380 nightmare of bat consciousness against toaster 00:27:22.380 --> 00:27:25.119 code. And finally, we stared down the barrel 00:27:25.119 --> 00:27:27.660 of the gorilla problem and the intense debate 00:27:27.660 --> 00:27:30.599 over our survival. By understanding the mechanics 00:27:30.599 --> 00:27:33.779 behind these concepts, from embodied cognition 00:27:33.779 --> 00:27:36.599 to the attention mechanisms driving the software. 00:27:36.890 --> 00:27:39.890 You have the tools to evaluate the headlines 00:27:39.890 --> 00:27:41.910 critically. You aren't just reacting to the hype 00:27:41.910 --> 00:27:44.789 anymore. You are equipped to understand the underlying 00:27:44.789 --> 00:27:47.990 shifts as this technology integrates into the 00:27:47.990 --> 00:27:50.730 real world. You can see exactly how the skyscraper 00:27:50.730 --> 00:27:53.049 is being built, even if the blueprint is constantly 00:27:53.049 --> 00:27:56.529 changing in our hands. But before we wrap up, 00:27:56.609 --> 00:27:58.690 I want to leave you with one final provocative 00:27:58.690 --> 00:28:01.890 thought to mull over on your own. Okay. We spent 00:28:01.890 --> 00:28:04.670 a lot of time today talking about how AGI superpower 00:28:04.670 --> 00:28:07.690 is generalizing human knowledge, taking the rules 00:28:07.690 --> 00:28:10.009 from one domain we understand and applying them 00:28:10.009 --> 00:28:12.990 to another. Right. But what happens when an AGI 00:28:13.230 --> 00:28:15.589 processing information millions of times faster 00:28:15.589 --> 00:28:18.650 than a biological brain, begins to discover entirely 00:28:18.650 --> 00:28:21.950 new fields of science. What happens when it generates 00:28:21.950 --> 00:28:25.029 novel dimensions of mathematics? Or entirely 00:28:25.029 --> 00:28:27.789 new spectrums of artistic expression that the 00:28:27.789 --> 00:28:30.769 human brain literally lacks the biological hardware 00:28:30.769 --> 00:28:34.180 to comprehend? If an AGI eventually solves the 00:28:34.180 --> 00:28:36.500 deepest mysteries of the universe, would we even 00:28:36.500 --> 00:28:39.460 be capable of recognizing its genius? Or would 00:28:39.460 --> 00:28:41.819 the ultimate answers to reality just look and 00:28:41.819 --> 00:28:44.180 sound like absolute noise to us? It forces us 00:28:44.180 --> 00:28:46.940 to ask, are we building our technological successors, 00:28:47.259 --> 00:28:49.380 or are we summoning something so alien we won't 00:28:49.380 --> 00:28:51.460 even be able to communicate with it? Something 00:28:51.460 --> 00:28:53.279 to think about the next time you ask an algorithm 00:28:53.279 --> 00:28:55.279 to summarize a meeting for you. Thank you for 00:28:55.279 --> 00:28:57.339 taking this deep dive with us. Stay curious, 00:28:57.539 --> 00:28:59.420 stay informed, and we will catch you next time.