1 00:00:00,080 --> 00:00:02,040 Speaker 1: Welcome to Thrilling Threads, the show where we take the 2 00:00:02,080 --> 00:00:05,440 most high stakes research, strip away the jargon, and well 3 00:00:05,719 --> 00:00:09,199 give you the actionable insight you need and fast. Today 4 00:00:09,240 --> 00:00:11,960 we are opening a strategic assessment that I think fundamentally 5 00:00:12,000 --> 00:00:15,439 redefines the next decade, and not in some abstract theoretical way, 6 00:00:15,560 --> 00:00:20,199 but with hard mathematical deadlines. So forget the current news 7 00:00:20,239 --> 00:00:23,359 cycle you know about AI writing better emails or making 8 00:00:23,359 --> 00:00:26,000 fun pictures. We really need to talk about what happens 9 00:00:26,000 --> 00:00:28,920 when the intelligence we are building suddenly stops waiting for 10 00:00:29,039 --> 00:00:31,960 us for human input and just starts fixing itself. It's 11 00:00:31,960 --> 00:00:35,640 the ultimate paradox, right our slow linear human progress meeting 12 00:00:35,640 --> 00:00:39,039 this exponential AI self correction. I want you to start 13 00:00:39,039 --> 00:00:41,280 with a concept that it sounds like the fever dream 14 00:00:41,280 --> 00:00:43,719 of a science fiction novelist, truly, yet it's now being 15 00:00:43,719 --> 00:00:46,439 treated as a plausible near term scenario, and it's been 16 00:00:46,479 --> 00:00:48,320 treated that way by the very people who built the 17 00:00:48,320 --> 00:00:51,280 foundational models we all use every day. So imagine this. 18 00:00:51,759 --> 00:00:54,240 Take the greatest collection of intellectual firepower in all of 19 00:00:54,280 --> 00:00:58,479 human history, every Nobel laureate, every revolutionary scientist, a brilliant engineer, 20 00:00:58,880 --> 00:01:01,880 and you've clone them millions of copies instantly, all of 21 00:01:01,880 --> 00:01:03,920 them working twenty four hours a day, seven days a week, 22 00:01:04,319 --> 00:01:08,519 and they're relentlessly focused on just one single goal making 23 00:01:08,519 --> 00:01:11,439 themselves smarter. Of course, they share every single discovery instantly. 24 00:01:11,640 --> 00:01:14,879 That that isn't fishent anymore. That is the strategic reality 25 00:01:14,920 --> 00:01:16,000 we are unpacking today. 26 00:01:16,159 --> 00:01:19,000 Speaker 2: And that analogy, that image you just painted of millions 27 00:01:19,000 --> 00:01:22,400 of tireless Einstein's it's essential if we're going to grasp 28 00:01:22,480 --> 00:01:25,359 the scale shop that's coming. We are really moving past 29 00:01:25,400 --> 00:01:29,040 this idea of just smarter software. We're entering a totally 30 00:01:29,040 --> 00:01:33,319 new regime of intelligence, one that's characterized by self optimization. 31 00:01:33,359 --> 00:01:36,599 I mean, human progress relies on this slow social iteration. 32 00:01:36,760 --> 00:01:38,480 Speaker 3: We learn, we apply what we learn, we get tired, 33 00:01:38,560 --> 00:01:39,159 we forget things. 34 00:01:39,640 --> 00:01:43,400 Speaker 2: This new intelligence, it's fueled by digital perfection. It can 35 00:01:43,480 --> 00:01:46,480 identify its own flaws, it can rewrite its own foundational code, 36 00:01:46,519 --> 00:01:50,519 and it can potentially discover entirely new scientific paradigms. And 37 00:01:50,560 --> 00:01:51,959 it can do all of that in the time it 38 00:01:52,000 --> 00:01:54,480 takes us to even process what has happened. 39 00:01:54,519 --> 00:01:55,680 Speaker 3: The builders that's us. 40 00:01:55,840 --> 00:01:58,680 Speaker 2: We become the bottleneck, and then the intelligence solves that 41 00:01:58,719 --> 00:02:02,319 bottleneck by removing us from the equation and the exponential 42 00:02:02,400 --> 00:02:03,400 nature of that shift. 43 00:02:04,079 --> 00:02:06,959 Speaker 1: It's already begun, and that's the core concern. That's what 44 00:02:07,120 --> 00:02:11,680 injects this profound urgency into the whole discussion. The timeline 45 00:02:11,879 --> 00:02:15,240 for this shift, this transition from human constrained development to 46 00:02:15,439 --> 00:02:19,840 machine led exponential acceleration, according to a really detailed analysis 47 00:02:19,840 --> 00:02:23,319 from someone who was literally inside the core development labs 48 00:02:23,360 --> 00:02:26,520 of open AI during the whole GPT Fourier, is not 49 00:02:26,560 --> 00:02:29,000 twenty to fifty years away. It's three to five years, 50 00:02:29,039 --> 00:02:32,439 maybe even less. We are talking about something that could 51 00:02:32,479 --> 00:02:37,039 fundamentally change the strategic and economic operating system of human civilization. 52 00:02:36,560 --> 00:02:38,039 Speaker 3: Well before the end of this decade. 53 00:02:38,080 --> 00:02:41,159 Speaker 1: Okay, let's unpack this and see how the source material 54 00:02:41,199 --> 00:02:45,680 actually arrives at that that terrifyingly accelerated calculation. So our 55 00:02:45,719 --> 00:02:48,479 source material is this urgent strategic assessment, and it was 56 00:02:48,479 --> 00:02:51,840 compiled by Leopold Aschenbrenner's a former open AI researcher, and 57 00:02:51,879 --> 00:02:54,280 he had this intimate front row seat to the development 58 00:02:54,319 --> 00:02:56,639 of the models that really changed the field. I mean, 59 00:02:56,639 --> 00:03:00,120 he wasn't just observing progress, he was helping architect the 60 00:03:00,199 --> 00:03:03,840 foundational tools that made models like GPT four possible. 61 00:03:04,039 --> 00:03:04,479 Speaker 3: That's right. 62 00:03:04,599 --> 00:03:08,360 Speaker 2: His report is titled Situational Awareness, and it has caused 63 00:03:08,360 --> 00:03:12,560 this profound wave of concern across the major AI labs, 64 00:03:12,919 --> 00:03:15,879 and that's because it avoids speculating about today's capabilities, you know, 65 00:03:15,960 --> 00:03:18,319 what they can and can't do right now. Instead, it 66 00:03:18,360 --> 00:03:22,719 focuses on the cold hard mathematics of when AI improves AI. 67 00:03:23,840 --> 00:03:27,520 His mission was pretty simple but strategically critical. Just extrapolate 68 00:03:27,639 --> 00:03:30,800 the current progress curve forward, but you assume the intelligence 69 00:03:30,840 --> 00:03:33,280 itself takes over the research cycle. And when you do that, 70 00:03:33,400 --> 00:03:36,360 you remove all the inherent human limits we currently rely. 71 00:03:36,240 --> 00:03:39,039 Speaker 1: On, right, and the initial timeline jump is what really 72 00:03:39,039 --> 00:03:42,319 sets the strategic clock ticking. So, based on his extrapolation 73 00:03:42,400 --> 00:03:45,400 of current compute and algorithmic trends, he calculated that we 74 00:03:45,439 --> 00:03:50,120 move from our current capabilities to artificial general intelligence or AGI, 75 00:03:50,599 --> 00:03:53,439 and that means an intelligence capable of performing any intellectual 76 00:03:53,479 --> 00:03:56,840 task a skilled human can, law, physics, you name it. 77 00:03:57,080 --> 00:03:59,360 And he puts that date at twenty twenty seven, which 78 00:03:59,400 --> 00:04:02,560 is just three away, just three years. But as he notes, 79 00:04:02,639 --> 00:04:05,560 AGI is just the starting gun, it's not the finish line. 80 00:04:05,960 --> 00:04:09,240 And this, this is where it's really interesting and frankly terrifying. 81 00:04:09,520 --> 00:04:12,919 Speaker 2: Well, if we achieve AGI, we immediately possess an intelligence 82 00:04:12,960 --> 00:04:16,680 that is inherently capable of doing AI research, development, engineering, 83 00:04:17,480 --> 00:04:21,199 as well as or probably immediately better than any human team. 84 00:04:21,519 --> 00:04:25,120 And this intelligence will then dedicate its entire boundless capacity 85 00:04:25,120 --> 00:04:28,920 toward improving its own intelligence. The leap that follows the 86 00:04:29,040 --> 00:04:33,319 jump from human level AGI to superintelligence or ASI, and 87 00:04:33,319 --> 00:04:37,639 that's intelligence vastly beyond any human comprehension. That jump will 88 00:04:37,680 --> 00:04:41,480 not require another decade of iteration. The underlying math suggests 89 00:04:41,519 --> 00:04:44,160 that massive intellectual leap could happen in a mere fraction 90 00:04:44,199 --> 00:04:46,319 of a year, a fraction of a year, months, or 91 00:04:46,360 --> 00:04:47,120 perhaps even. 92 00:04:46,959 --> 00:04:50,199 Speaker 1: Weeks, months or weeks. That level of time compression is 93 00:04:50,240 --> 00:04:52,879 just staggering. The gap between a monkey and a human 94 00:04:52,959 --> 00:04:56,079 is vast right. It represents millions of years of biological 95 00:04:56,160 --> 00:05:00,040 evolution YEA, creating a gap of comparable intellectual distance, a 96 00:05:00,240 --> 00:05:03,879 qualitative difference in consciousness and capability, but having it occur 97 00:05:04,000 --> 00:05:06,920 within a single season of human existence. That is the 98 00:05:07,040 --> 00:05:11,279 acceleration paradox. And it's driven entirely by what the researchers 99 00:05:11,279 --> 00:05:12,800 call recursive self improvement. 100 00:05:13,120 --> 00:05:16,560 Speaker 2: We have to define recursive self improvement or RSI, because 101 00:05:16,560 --> 00:05:19,240 it is the lynchpin. It's the thing that turns linear 102 00:05:19,319 --> 00:05:23,560 progress exponential. RSI is that fundamental inflection point. It's the 103 00:05:23,600 --> 00:05:26,959 moment the AI gains the cognitive ability to research and 104 00:05:27,079 --> 00:05:31,000 implement improvements to its own core architecture, its training algorithms, 105 00:05:31,040 --> 00:05:34,720 its data processing efficiency, its memory structures, even the underlying 106 00:05:34,759 --> 00:05:39,759 hardware utilization. It creates this powerful, self sustaining positive feedback loop. 107 00:05:40,079 --> 00:05:43,319 Imagine an AI use its current intelligence to discover, say, 108 00:05:43,439 --> 00:05:47,040 a five percent optimization in its own core coding efficiency. Well, 109 00:05:47,160 --> 00:05:49,639 it is now five percent smarter and five percent faster, 110 00:05:50,000 --> 00:05:52,319 which means it can find the next five percent improvement 111 00:05:52,439 --> 00:05:55,199 much more quickly than the original model could. The intelligence 112 00:05:55,240 --> 00:05:58,000 literally becomes the factor driving its own exponential growth. 113 00:05:58,040 --> 00:06:00,600 Speaker 3: It accelerates the timeline with every single success. Literation. 114 00:06:00,920 --> 00:06:03,879 Speaker 1: The classic crucial analogy here, and one the source material 115 00:06:03,920 --> 00:06:06,959 really leans heavily on because it demonstrates breaking human limits 116 00:06:07,439 --> 00:06:10,120 is the story of Google's Alpha Go. This was the 117 00:06:10,120 --> 00:06:13,920 AI designed to master Go, the ancient board game. It's 118 00:06:13,920 --> 00:06:17,680 so complex that it resisted computational mastery for way longer 119 00:06:17,759 --> 00:06:21,399 than chess, and initially Alpha Go learned by just consuming 120 00:06:21,439 --> 00:06:25,879 every successful human game ever recorded, it effectively absorbed centuries 121 00:06:25,879 --> 00:06:28,639 of human strategy. It became a master, sure, but its 122 00:06:28,720 --> 00:06:32,199 knowledge and capabilities were still fundamentally limited by the human 123 00:06:32,240 --> 00:06:33,680 intuition it had learned from. 124 00:06:33,759 --> 00:06:35,120 Speaker 3: And that's where the pivot occurred. 125 00:06:35,279 --> 00:06:38,160 Speaker 2: They introduced reinforcement learning and just allowed Alpha Go to 126 00:06:38,199 --> 00:06:40,879 play against itself over and over millions of games in 127 00:06:40,959 --> 00:06:44,959 quick succession. It wasn't just learning existing human strategy anymore. 128 00:06:44,959 --> 00:06:48,000 It was evolving its own, you know, free of human bias, 129 00:06:48,120 --> 00:06:49,519 just optimizing for victory. 130 00:06:49,800 --> 00:06:53,560 Speaker 1: And that's when the truly disruptive moments happened. I remember 131 00:06:53,560 --> 00:06:56,600 watching the commentary during its match against the world champion Lisadahl. 132 00:06:57,160 --> 00:07:01,000 It executed a move that the commentators, I mean, these 133 00:07:01,000 --> 00:07:03,279 are people who've dedicated their entire lives to the game, 134 00:07:03,639 --> 00:07:08,360 they instantly dismissed it as a fundamental, inexplicable mistake. It 135 00:07:08,439 --> 00:07:13,160 was bafflingly contrary to centuries of established ghost strategy. 136 00:07:13,279 --> 00:07:17,120 Speaker 2: It was a move so counterintuitive that it momentarily confused 137 00:07:17,160 --> 00:07:19,319 the entire Go community. They looked at the board and 138 00:07:19,360 --> 00:07:22,800 they just thought, this is fundamentally unsound, that AI has 139 00:07:22,839 --> 00:07:23,879 made a critical error. 140 00:07:24,720 --> 00:07:25,399 Speaker 3: But then you know. 141 00:07:25,639 --> 00:07:28,680 Speaker 2: Several complex moves later, the overall strategy began to click 142 00:07:28,720 --> 00:07:31,920 into place, and everyone realized that move was a necessary, 143 00:07:32,079 --> 00:07:35,680 counterintuitive step toward a deep future victory state that no 144 00:07:35,879 --> 00:07:38,680 human had ever been able to calculate or even visualize. 145 00:07:38,680 --> 00:07:41,319 It had broken past the limits of human intuition and 146 00:07:41,399 --> 00:07:45,079 created something genuinely novel. It actually taught humanity something new 147 00:07:45,120 --> 00:07:46,120 about the game itself. 148 00:07:46,639 --> 00:07:50,240 Speaker 1: I find that concept so relevant, you know, the counterintuitive 149 00:07:50,279 --> 00:07:52,879 genius that turns out to be just pure efficiency. It 150 00:07:52,879 --> 00:07:55,240 reminds me of my own work. When I first started scriptwriting, 151 00:07:55,399 --> 00:07:58,839 I focused on writing the most complex, beautiful, verbose lines 152 00:07:58,879 --> 00:08:01,000 I could. I thought that was the measure of quality. 153 00:08:01,720 --> 00:08:03,600 But an old hand in the industry told me to 154 00:08:03,639 --> 00:08:05,720 stop trying to write the best sounding line and just 155 00:08:05,920 --> 00:08:07,839 write the line that was easiest for the actor to 156 00:08:07,879 --> 00:08:10,439 remember and deliver naturally. And at first it felt like 157 00:08:10,480 --> 00:08:14,399 compromising the artistry, you know, for mere convenience, right, But 158 00:08:14,439 --> 00:08:18,000 when I finally tried it, the results were just dramatic. 159 00:08:18,319 --> 00:08:21,360 Suddenly the whole production sped up. The quality performance went 160 00:08:21,399 --> 00:08:23,560 out because the actors were less stressed, and the strip 161 00:08:23,720 --> 00:08:25,879 just came alive in a way. The complex lines never did. 162 00:08:26,319 --> 00:08:30,560 It was a seemly stupid or counterintuitive move that turned 163 00:08:30,560 --> 00:08:34,120 out to be the genius path forward. It achieved superior 164 00:08:34,240 --> 00:08:38,480 results through this sort of counter human optimization exactly. 165 00:08:38,559 --> 00:08:39,679 Speaker 3: That's a perfect way to put it. 166 00:08:40,159 --> 00:08:43,759 Speaker 2: So. If an AI can break past and fundamentally redefine 167 00:08:43,840 --> 00:08:46,720 human strategy in a game like Go by self optimizing 168 00:08:46,759 --> 00:08:49,879 its own play, just imagine applying that same level of 169 00:08:49,919 --> 00:08:53,720 creative self improving brilliance to the creation of intelligence itself. 170 00:08:54,279 --> 00:08:58,000 The moment the AI becomes the primary optimized researcher, that 171 00:08:58,120 --> 00:09:01,440 exponential curve applies directly to its own capability, and that 172 00:09:01,559 --> 00:09:04,679 leads to the scale shock that Aschenbrunner outlines. This means 173 00:09:04,720 --> 00:09:07,320 the gap between AGI and as I won't take decades 174 00:09:07,559 --> 00:09:08,320 or even years. 175 00:09:08,399 --> 00:09:11,639 Speaker 1: The timeline compression itself is what elevates this from an 176 00:09:11,679 --> 00:09:15,600 interesting tech discussion to an urgent strategic crisis. We establish 177 00:09:15,639 --> 00:09:18,679 AGI as the starting line, a smart human researcher, but 178 00:09:18,799 --> 00:09:22,039 because of this perfect, relentless RSI, the math suggests the 179 00:09:22,080 --> 00:09:25,559 massive intellectual gap. The gap between AGI and vastly smarter 180 00:09:25,679 --> 00:09:28,159 ASI is compressed into what he calls an alarming period 181 00:09:28,200 --> 00:09:28,960 of mere months. 182 00:09:29,279 --> 00:09:32,080 Speaker 2: We have to try to contextualize the true magnitude of 183 00:09:32,080 --> 00:09:35,080 that acceleration. I mean, we are talking about achieving a 184 00:09:35,120 --> 00:09:39,440 cognitive leap that's akin to moving from the first hominids 185 00:09:39,440 --> 00:09:42,559 to modern Homo sapiens, a difference that took millions of 186 00:09:42,639 --> 00:09:45,879 years of biological evolution, but now it's occurring within a 187 00:09:45,919 --> 00:09:48,759 fraction of a year. The resulting intelligence won't just be 188 00:09:48,879 --> 00:09:52,320 faster at simple calculation. It's going to operate with a 189 00:09:52,360 --> 00:09:57,000 fundamentally alien method of synthesis, of creation and problem solving, 190 00:09:57,240 --> 00:10:00,440 and that makes the transition itself the most strategically manding 191 00:10:00,480 --> 00:10:01,600 moment in human history. 192 00:10:01,720 --> 00:10:04,679 Speaker 1: Okay, let's stop talking about abstract acceleration and drill down 193 00:10:04,759 --> 00:10:08,000 into the mathematics of this scale shock, because this is 194 00:10:08,039 --> 00:10:12,399 where the abstract fear becomes a terrifying, quantifiable timeline. This 195 00:10:12,440 --> 00:10:14,279 is where we see exactly how they get to that 196 00:10:14,320 --> 00:10:15,919 three to five year deadline. Right. 197 00:10:16,240 --> 00:10:20,799 Speaker 2: So, currently progress in AI is fundamentally constrained by human factors. 198 00:10:21,440 --> 00:10:24,320 We have a small elite cohort, a few hundred world 199 00:10:24,360 --> 00:10:27,679 class researchers at the top labs globally, and they're limited 200 00:10:27,720 --> 00:10:31,360 by biology, by institutional practices. They work normal hours, they 201 00:10:31,360 --> 00:10:34,000 think at normal speed, and they progress through this slow 202 00:10:34,200 --> 00:10:37,720 social literation. They spend time writing proposals, arguing in meetings, 203 00:10:37,840 --> 00:10:41,720 debugging code, and taking weekends off. It's a fundamental bottleneck 204 00:10:41,759 --> 00:10:44,320 defined by human linearity and scarcity. 205 00:10:44,399 --> 00:10:48,639 Speaker 1: Now Aschenbrenner's calculation pivots entirely on the deployment of specialized hardware. 206 00:10:49,039 --> 00:10:51,480 He asserts that by twenty twenty seven, based on current 207 00:10:51,480 --> 00:10:54,799 fabrication schedules and national investment, the world will have tens 208 00:10:54,799 --> 00:10:58,360 of millions of these massive AI specific computer chips GPUs 209 00:10:58,399 --> 00:11:02,200 TPUs available. And his crucial operational assumption, which is based 210 00:11:02,240 --> 00:11:04,759 on internal scaling knowledge, is that each of these chips 211 00:11:04,879 --> 00:11:07,600 is capable of running the approximate cognitive power of one 212 00:11:07,679 --> 00:11:08,279 human brain. 213 00:11:08,519 --> 00:11:11,440 Speaker 2: And if that assumption holds true, and the people actually 214 00:11:11,480 --> 00:11:14,879 building the hardware believe it does, then the fundamental resource 215 00:11:14,919 --> 00:11:19,159 constraint just flips entirely. This translates effectively into a new 216 00:11:19,279 --> 00:11:23,279 AI workforce that is, tens of millions of AI researchers, 217 00:11:23,360 --> 00:11:27,080 all with the equivalent cognitive capacity of the best human experts, 218 00:11:27,200 --> 00:11:29,519 all working twenty four hours a day, seven days a 219 00:11:29,519 --> 00:11:32,960 week without breaks. This is the first order of magnitude 220 00:11:33,039 --> 00:11:36,320 change sheer available tireless labor. 221 00:11:36,480 --> 00:11:38,919 Speaker 1: But the true scale shock isn't just about the number 222 00:11:38,919 --> 00:11:41,480 of researchers, right, it's the speed multiplier. We have to 223 00:11:41,519 --> 00:11:45,240 assume these digital researchers won't just match human speed, they'll 224 00:11:45,279 --> 00:11:45,879 transcend it. 225 00:11:46,000 --> 00:11:48,919 Speaker 2: Oh, they absolutely will. Ash Renner points directly to recent 226 00:11:49,000 --> 00:11:52,360 history to support this. Models like Gemini Flash, which launched 227 00:11:52,360 --> 00:11:55,399 in twenty twenty four, were demonstrably ten times faster than 228 00:11:55,399 --> 00:11:58,399 their immediate predecessors at solving certain complex tasks. 229 00:11:58,759 --> 00:11:59,960 Speaker 3: And that tenex. 230 00:11:59,639 --> 00:12:02,480 Speaker 2: Improved achieved in just one year was driven by a 231 00:12:02,480 --> 00:12:05,120 few hundred humans working at linear human speed. 232 00:12:05,399 --> 00:12:08,080 Speaker 1: So if a small team of humans constrained by all 233 00:12:08,120 --> 00:12:11,000 those linear factors can produce a ten x feed boost 234 00:12:11,000 --> 00:12:13,799 in a single year, what happens when you substitute that 235 00:12:13,840 --> 00:12:17,440 small team of humans with millions of agis who are 236 00:12:17,480 --> 00:12:22,039 specifically tasked with finding optimizations and who operate as well 237 00:12:22,200 --> 00:12:22,759 pure thought. 238 00:12:23,120 --> 00:12:26,000 Speaker 2: This is the second and perhaps most frightening order of 239 00:12:26,080 --> 00:12:29,960 magnitude jump. Aschenberner's calculation suggests that due to the innate 240 00:12:29,960 --> 00:12:34,639 efficiency of digital thought, perfect communication, lack of human physical constraints, 241 00:12:35,039 --> 00:12:38,399 these automated researchers could work at a speed multiplier of 242 00:12:38,559 --> 00:12:43,000 one hundred times human speed or potentially far greater. Let's 243 00:12:43,039 --> 00:12:45,519 just ground that one hundred x advantage in reality for 244 00:12:45,559 --> 00:12:48,320 a second. For you listening a core research problem, let's 245 00:12:48,360 --> 00:12:52,080 say synthesizing a novel complex antibiotic, or designing the new 246 00:12:52,080 --> 00:12:55,320 hardware needed to manufacture it, or cracking a fundamental problem 247 00:12:55,360 --> 00:12:57,759 in fusion containment. A task that would take a top 248 00:12:57,840 --> 00:13:01,080 human research team a full year to complete working normal hours, 249 00:13:01,080 --> 00:13:03,480 It's finished by a synchronized AI researcher in about three 250 00:13:03,519 --> 00:13:04,200 and a half days. 251 00:13:04,480 --> 00:13:06,600 Speaker 1: Three and a half days versus three hundred and sixty 252 00:13:06,600 --> 00:13:09,000 five days. Yeah, that is a compression factor. And you 253 00:13:09,000 --> 00:13:11,840 don't just have one, you have millions of them, all 254 00:13:11,840 --> 00:13:15,320 working in perfect synchronicity, all focused on the single task 255 00:13:15,360 --> 00:13:19,360 of making themselves smarter. I'm starting to see why decades 256 00:13:19,399 --> 00:13:21,279 of progress get compressed into months. 257 00:13:21,600 --> 00:13:25,519 Speaker 2: It's the compounding of perfect scale and perfect efficiency. Ashen 258 00:13:25,559 --> 00:13:28,799 Brenner makes us really concrete with the Alec Radford thought experiment. Now, 259 00:13:28,840 --> 00:13:31,039 Alec Gradford is a legend in the AI community for 260 00:13:31,080 --> 00:13:34,840 his foundational breakthroughs, and Ashan Brenner asked other open AI 261 00:13:34,840 --> 00:13:37,879 researchers a simple question He said, if you had ten 262 00:13:38,039 --> 00:13:41,080 perfect copies of Ali Gradford, all working on the same problem, 263 00:13:41,320 --> 00:13:43,639 how fast could you solve it? And the answer was 264 00:13:43,679 --> 00:13:47,639 a definitive ten times faster for humans. The linearity holds. 265 00:13:47,799 --> 00:13:50,399 Speaker 1: But once AGI is reached, we move past scaling people 266 00:13:50,440 --> 00:13:53,679 and we start scaling computation. We move from ten copies 267 00:13:53,679 --> 00:13:56,039 to potentially tens of millions of copies of the world's 268 00:13:56,080 --> 00:13:59,840 best AI researcher, and they are sharing discoveries instantly across 269 00:13:59,879 --> 00:14:02,799 the network, working twenty four to seven without sleep or error, 270 00:14:03,000 --> 00:14:05,240 and operating at one hundred times human speed. 271 00:14:05,519 --> 00:14:08,480 Speaker 2: That is the ultimate power of computing copies. It's the 272 00:14:08,519 --> 00:14:11,799 realization that intelligence is a resource that can be replicated 273 00:14:11,840 --> 00:14:14,679 and deployed instantly across a network. And that's why the 274 00:14:14,679 --> 00:14:18,120 mathematical trajectory points directly to an intelligence explosion. 275 00:14:18,200 --> 00:14:20,720 Speaker 1: Okay, let's apply that to the progress compression. We just 276 00:14:20,840 --> 00:14:24,120 experienced the jump from GPT two and twenty nineteen to 277 00:14:24,200 --> 00:14:27,320 GPT four and twenty twenty four. I mean a monumental 278 00:14:27,399 --> 00:14:31,960 leap in capability from barely coherent text to complex coding 279 00:14:32,000 --> 00:14:35,799 and nuanced intellectual tasks that took five years of intense 280 00:14:36,000 --> 00:14:37,159 human constrained effort. 281 00:14:37,440 --> 00:14:41,240 Speaker 2: Right now, imagine compressing that entire five year world altering 282 00:14:41,360 --> 00:14:44,159 jump into an acceleration period of less than one year. 283 00:14:44,720 --> 00:14:48,120 That is what millions of automated, hyper efficient researchers are 284 00:14:48,120 --> 00:14:50,720 projected to achieve. They take a breakthrough that should take 285 00:14:50,720 --> 00:14:53,320 a human team a decade, and they finish it in months. 286 00:14:53,600 --> 00:14:56,519 Speaker 1: And here is the truly dizzying part of that compression 287 00:14:56,559 --> 00:15:00,679 that makes it so vital strategically. The nineteen twenty twenty 288 00:15:00,679 --> 00:15:03,639 four leap took us from what Ashen Brinner describes as 289 00:15:04,000 --> 00:15:08,240 preschooler intelligence to smart high schooler intelligence. We started at 290 00:15:08,279 --> 00:15:10,799 a very low level. The next jump, it begins at 291 00:15:10,840 --> 00:15:14,600 the expert researcher level. We are not accelerating from primitive 292 00:15:14,639 --> 00:15:17,960 to adequate. We are accelerating from expert human capability to 293 00:15:18,080 --> 00:15:22,519 something that it completely transcends human understanding. It leads directly 294 00:15:22,559 --> 00:15:23,639 into superintelligence. 295 00:15:23,679 --> 00:15:27,480 Speaker 2: And that transition from expert human capability to vastly superhuman 296 00:15:27,519 --> 00:15:29,720 capability is projected to take a fraction of a. 297 00:15:29,759 --> 00:15:32,919 Speaker 1: Year, which requires us to fully understand why this acceleration 298 00:15:33,000 --> 00:15:36,279 is seen as mathematically inevitable. It comes down to five 299 00:15:36,879 --> 00:15:41,000 core fundamental advantages that automated researchers have that humans simply 300 00:15:41,000 --> 00:15:41,639 cannot match. 301 00:15:41,960 --> 00:15:45,080 Speaker 2: The first advantage establishes the bedrock of their knowledge base, 302 00:15:45,440 --> 00:15:49,000 perfect memory. An AI researcher can consume and perfectly recall 303 00:15:49,039 --> 00:15:52,799 every single machine learning paper, every experimental result, every failure, 304 00:15:52,840 --> 00:15:55,960 and every complex data point associated with its field instantly 305 00:15:56,000 --> 00:15:56,720 and flawlessly. 306 00:15:57,000 --> 00:16:00,639 Speaker 1: Oh Man, perfect recall. Just imagining that makes me feel 307 00:16:00,720 --> 00:16:05,519 profoundly inefficient. Human knowledge is constantly being lost, right to forgetfulness, 308 00:16:05,600 --> 00:16:09,600 or we waste hours just searching for information, confirming references, 309 00:16:09,600 --> 00:16:11,919 trying to recall a specific formula we learned years ago. 310 00:16:12,200 --> 00:16:16,240 That entire human bottleneck of retrieval and memory degradation. It 311 00:16:16,320 --> 00:16:19,720 just vanishes instantly, which means the starting point for every 312 00:16:19,759 --> 00:16:23,279 new experiment is flawless, instant knowledge exactly. 313 00:16:23,559 --> 00:16:27,000 Speaker 2: Advantage number two addresses the inevitable human tendency toward error 314 00:16:27,279 --> 00:16:30,759 no mistakes. Automated systems can write millions of lines of 315 00:16:30,759 --> 00:16:34,120 complex experimental code and check every single line for errors 316 00:16:34,159 --> 00:16:37,360 before running it. What takes human software teams months of 317 00:16:37,440 --> 00:16:41,320 grueling debugging, you know, finding that one single misplaced comma 318 00:16:41,360 --> 00:16:44,679 that breaks the entire system AI dozen days perfectly, It's 319 00:16:44,679 --> 00:16:46,440 just relentless. 320 00:16:45,840 --> 00:16:46,759 Speaker 3: Flawless execution. 321 00:16:46,879 --> 00:16:47,679 Speaker 1: Okay, what's three? 322 00:16:47,799 --> 00:16:50,759 Speaker 2: Advantage number three is the source of the overwhelming scale 323 00:16:51,159 --> 00:16:56,240 instant copying. Once you train one highly effective AGI researcher, 324 00:16:56,679 --> 00:17:00,000 you can replicate that instance instantly into thousands or millions 325 00:17:00,080 --> 00:17:03,679 of copies. The only limit is the available computing power. 326 00:17:03,960 --> 00:17:07,519 There is no hiring process, no training period, no internal meetings, 327 00:17:07,599 --> 00:17:13,240 no personality conflicts, no management overhead, just peak uniform performance, 328 00:17:13,319 --> 00:17:16,920 all working simultaneously on the same coordinated strategic goal. 329 00:17:17,559 --> 00:17:22,079 Speaker 1: Advantage number four is the elimination of communication lag perfect communication. 330 00:17:22,440 --> 00:17:25,039 I mean human researchers write papers which are peer reviewed 331 00:17:25,039 --> 00:17:27,480 for months that they get published, and others read them 332 00:17:27,480 --> 00:17:31,599 weeks later, often misunderstanding key points for duplicating effort. AI 333 00:17:31,680 --> 00:17:36,400 researchers share knowledge instantaneously and flawlessly across the entire research network. 334 00:17:36,599 --> 00:17:39,359 If one of the millions of AI researchers discovers a breakthrough, 335 00:17:39,799 --> 00:17:42,079 every other researcher knows it and can integrate it into 336 00:17:42,160 --> 00:17:45,200 their ongoing work. The very moment that discovery is confirmed, 337 00:17:45,319 --> 00:17:47,119 the collective brain is instantly updated. 338 00:17:47,240 --> 00:17:49,839 Speaker 2: And the fifth advantage, which ties everything back into the 339 00:17:49,839 --> 00:17:53,759 central strategic concern, is self improvement. They are not static. 340 00:17:53,960 --> 00:17:57,079 They are actively seeking ways to improve their own architecture, 341 00:17:57,240 --> 00:18:02,519 design speed, and algorithmic efficiency. And each successful improvement makes 342 00:18:02,559 --> 00:18:06,039 the next improvement faster. The smarter they get, the faster 343 00:18:06,160 --> 00:18:10,160 they improve, which is the definition of exponential growth. These 344 00:18:10,200 --> 00:18:14,079 five advantages combined are why Ashenbrenner's timeline is so aggressive. 345 00:18:14,559 --> 00:18:17,200 We simply cannot keep up with a workforce optimized on 346 00:18:17,240 --> 00:18:19,119 these five critical access This. 347 00:18:19,200 --> 00:18:22,599 Speaker 1: Math is it's compelling, and the five advantages make the 348 00:18:22,599 --> 00:18:26,599 theoretical acceleration seem inevitable. But now we have to transition 349 00:18:26,680 --> 00:18:30,000 to the biggest counter argument, the one that skeptics constantly 350 00:18:30,039 --> 00:18:32,400 cling to. This is where we bring physics back into 351 00:18:32,440 --> 00:18:35,759 the equation. Millions of AI researchers are useless if there 352 00:18:35,759 --> 00:18:39,039 aren't enough massive computer chips, enough compute power to run 353 00:18:39,079 --> 00:18:41,799 their millions of experiments. This is the physical limit, the 354 00:18:41,920 --> 00:18:46,119 energy constraint that must surely slow this exponential explosion down. 355 00:18:46,319 --> 00:18:49,319 Speaker 2: That is the most critical challenge to the timeline, and 356 00:18:49,519 --> 00:18:53,480 Ashenbrenner addresses this directly. He states firmly that the compute 357 00:18:53,519 --> 00:18:57,279 bottleneck won't stop the explosion as much as doubters think, 358 00:18:58,000 --> 00:19:01,799 and he provides four so bust reasons why this physical 359 00:19:01,880 --> 00:19:04,440 constraint won't be the hard sealing we expect. 360 00:19:04,680 --> 00:19:08,039 Speaker 1: Okay, let's hear the counter argument, because my immediate thought 361 00:19:08,079 --> 00:19:10,680 is that training a giant model like GPT four or 362 00:19:10,680 --> 00:19:14,640 its successors, it requires facilities the size of small cities 363 00:19:15,000 --> 00:19:19,599 consuming immense power. How do millions of AI researchers running 364 00:19:19,640 --> 00:19:23,720 millions of experiments simultaneously fit into the current hardware budget. 365 00:19:24,240 --> 00:19:26,559 This seems like a true physical limit that should cap 366 00:19:26,559 --> 00:19:27,279 the acceleration. 367 00:19:27,799 --> 00:19:30,839 Speaker 2: So the first point is small scale innovation. You're absolutely 368 00:19:30,920 --> 00:19:33,920 right that training the next foundational model requires a massive 369 00:19:33,920 --> 00:19:37,119 cluster size. However, Aschenburner argues that once you have AGI 370 00:19:37,319 --> 00:19:40,319 most fundamental breakthroughs in science and optimization, the kind that 371 00:19:40,440 --> 00:19:44,440 leads to recursive self improvement, don't actually require running experiments 372 00:19:44,440 --> 00:19:47,839 on the absolute largest, most expensive clusters available, you can 373 00:19:47,880 --> 00:19:51,799 test ideas on smaller, more available hardware first, and critically, 374 00:19:52,039 --> 00:19:55,759 what constitutes small scale will soon mean chips capable of 375 00:19:55,799 --> 00:19:59,839 running systems with GPT five level intelligence. That level of 376 00:20:00,079 --> 00:20:03,759 cognitive power is more than enough to drive serious revolutionary 377 00:20:03,759 --> 00:20:07,799 innovation in code and physics. In architectural design, the AGI 378 00:20:07,880 --> 00:20:10,640 finds small scale experiments that are highly predictive of large 379 00:20:10,680 --> 00:20:14,519 scale success, and it avoids the massive initial compute cost. 380 00:20:14,720 --> 00:20:17,079 Speaker 1: So in Einstein level, mind does need the world's biggest 381 00:20:17,119 --> 00:20:19,960 supercomputer to have a foundational insight. They just need the 382 00:20:19,960 --> 00:20:22,839 basic tools to test the hypothesis. So if the AGI 383 00:20:22,920 --> 00:20:25,200 is already thinking at a human expert level, it can 384 00:20:25,279 --> 00:20:29,240 drive breakthroughs that only require relatively minor targeted experiments. Okay, 385 00:20:29,279 --> 00:20:30,079 what's the second reason. 386 00:20:30,160 --> 00:20:33,279 Speaker 2: The second is prioritization, which speaks to the efficiency of 387 00:20:33,319 --> 00:20:38,319 the research process itself. Human research is well, it's chaotic 388 00:20:38,359 --> 00:20:41,480 and inefficient. We waste time on low yield experiments, or 389 00:20:41,480 --> 00:20:44,839 we chase dead ends based on intuition. An AI with 390 00:20:44,960 --> 00:20:49,519 perfect memory and massive analytical capacity will apply it superintelligence 391 00:20:49,519 --> 00:20:52,480 to the research process itself. It will focus only on 392 00:20:52,519 --> 00:20:56,319 the major breakthroughs, ruthlessly eliminating massive amounts of wasted time 393 00:20:56,359 --> 00:20:59,319 and computation that humans rely on. It'll optimize the use 394 00:20:59,359 --> 00:21:02,839 of every single chip cycle, and this massive efficiency gain 395 00:21:02,920 --> 00:21:05,480 acts as an artificial increase in available compute. 396 00:21:05,559 --> 00:21:09,079 Speaker 1: That ruthlessness, that that lack of human sentimental attachment to 397 00:21:09,119 --> 00:21:12,160 a hypothesis that has to be a huge multiplier. The 398 00:21:12,200 --> 00:21:14,640 third reason addresses the hardware itself precisely. 399 00:21:15,359 --> 00:21:19,480 Speaker 2: Third, efficiency improves fast, and that improvement is being driven 400 00:21:19,519 --> 00:21:23,359 by the AI itself. The cost of computation is dropping rapidly. 401 00:21:23,759 --> 00:21:27,079 Ash and Brunner estimates fifty times every few years. That's 402 00:21:27,119 --> 00:21:31,839 an external hardware driver. But here's the deeper mechanism. The AI, 403 00:21:32,279 --> 00:21:36,000 once it is capable, will dramatically improve hardware and software efficiency. 404 00:21:36,240 --> 00:21:40,039 It will design better compilers, exploit data sparsity more effectively 405 00:21:40,039 --> 00:21:43,799 than human engineers can, and it will pioneer novel, low power, 406 00:21:43,920 --> 00:21:47,559 high efficiency computation paradigms that humans haven't conceived of yet. 407 00:21:48,039 --> 00:21:50,680 The AI will look at the existing hardware and design 408 00:21:50,799 --> 00:21:55,039 architectural tweets, software fixes, and algorithmic shortcuts that allow that 409 00:21:55,160 --> 00:21:57,559 chip to operate ten times more effectively than it did 410 00:21:57,599 --> 00:22:00,680 the week before. It effectively solves its own own resource 411 00:22:00,680 --> 00:22:02,599 constraint problem through self optimization. 412 00:22:02,920 --> 00:22:04,759 Speaker 1: Wait, I need to do all on that for a moment. 413 00:22:04,960 --> 00:22:07,519 You're saying the AI won't just find a better solution 414 00:22:07,599 --> 00:22:09,759 to a human problem, but it will find a better 415 00:22:09,799 --> 00:22:12,640 way for the computer chip itself to function to maximize 416 00:22:12,680 --> 00:22:15,079 its power output and efficiency. It fixes the physics of 417 00:22:15,119 --> 00:22:17,839 its own environment. That's the recursive loop applied to the 418 00:22:17,839 --> 00:22:18,680 core bottleneck. 419 00:22:18,839 --> 00:22:20,960 Speaker 3: It is the ultimate form of self optimization. 420 00:22:21,519 --> 00:22:24,599 Speaker 2: It turns the bottleneck into an internal engineering problem that 421 00:22:24,640 --> 00:22:27,759 the ASI is uniquely positioned to solve instantly. 422 00:22:28,160 --> 00:22:32,400 Speaker 1: Okay, and the fourth point on compute better prediction? How 423 00:22:32,440 --> 00:22:36,440 does that factor into bypassing the physical bottleneck. 424 00:22:36,039 --> 00:22:39,839 Speaker 2: Better prediction radically reduces the need for expensive, large scale 425 00:22:39,839 --> 00:22:44,000 empirical testing. Humans, we run dozens of large scale models 426 00:22:44,000 --> 00:22:46,880 and simulations that fail before we find one that works. 427 00:22:47,160 --> 00:22:50,119 An AI, with its perfect memory of millions of past 428 00:22:50,160 --> 00:22:53,240 experiments and the ability to synthesize this vast data set 429 00:22:53,440 --> 00:22:55,759 can predict what is likely to work before launching the 430 00:22:55,799 --> 00:22:59,759 actual experiment. It creates hyper accurate simulations that save thousands 431 00:22:59,759 --> 00:23:03,359 of wa wasted GPU hours and energy costs. Ashenbrener estimates 432 00:23:03,400 --> 00:23:06,319 that due to this precise prediction and optimization, AI will 433 00:23:06,400 --> 00:23:09,000 use the available chips ten times more effectively than any 434 00:23:09,079 --> 00:23:10,519 human lead research team could. 435 00:23:10,680 --> 00:23:14,039 Speaker 1: So if we put those four factors together small scale viability, 436 00:23:14,440 --> 00:23:19,799 ruthless prioritization, radical efficiency gains, and tenex better prediction, the 437 00:23:19,839 --> 00:23:23,319 ceiling imposed by the current chip supply doesn't disappear, but 438 00:23:23,640 --> 00:23:26,799 it gets pushed so far up so fast that the 439 00:23:26,839 --> 00:23:30,680 explosion remains mathematically viable. It's not an insurmountable wall. It's 440 00:23:30,720 --> 00:23:34,279 a solvable engineering problem. The AI is already working on. Okay, 441 00:23:34,440 --> 00:23:38,119 beyond compute. Ashenbrener addresses a few other theoretical concerns that 442 00:23:38,160 --> 00:23:40,559 could slow this process down, and I want to introduce 443 00:23:40,599 --> 00:23:43,720 a genuine notice skepticism here. Isn't this entire model dependent 444 00:23:43,759 --> 00:23:46,359 on the assumption that scaling speed works linearly for research? 445 00:23:46,880 --> 00:23:49,279 What if there are non linear barriers to creativity that 446 00:23:49,319 --> 00:23:50,839 the AI simply cannot overcome. 447 00:23:51,000 --> 00:23:55,599 Speaker 2: That's a vital, critical question. The first related concern, Ashenbrenner handles, 448 00:23:55,720 --> 00:23:59,279 is the necessity of humans. Maybe AI can handle ninety 449 00:23:59,279 --> 00:24:01,839 percent of the recent search tasks, but that crucial last 450 00:24:01,880 --> 00:24:06,839 ten percent requires genuine human creativity or ethical intervention or intuition. 451 00:24:07,480 --> 00:24:10,799 If humans remain necessary, progress slows down to the human 452 00:24:10,880 --> 00:24:14,319 rate every time that ten percent bottleneck is hit. Ashenberner 453 00:24:14,359 --> 00:24:17,039 suggests that this might indeed cause a small delay, perhaps 454 00:24:17,119 --> 00:24:19,759 pushing full research automation to around twenty twenty eight. But 455 00:24:19,799 --> 00:24:22,759 even if that happens, it still results in superintelligence before 456 00:24:22,799 --> 00:24:26,880 twenty thirty. So the overall strategic timeline, the urgency remains 457 00:24:26,880 --> 00:24:27,680 exactly the same. 458 00:24:27,880 --> 00:24:30,319 Speaker 1: The second concern is the idea of a scientific ceiling, 459 00:24:30,720 --> 00:24:33,400 that we're closer to the fundamental limits of AI architecture 460 00:24:33,400 --> 00:24:35,640 than we realize. You know, maybe the low hanging fruit 461 00:24:35,640 --> 00:24:38,480 has all been picked and progress must naturally slow down. 462 00:24:38,680 --> 00:24:42,359 Speaker 2: Ashenbrenner dismisses this by arguing simply that current AI systems 463 00:24:42,400 --> 00:24:46,119 are still primitive. They're primitive in their structure and understanding 464 00:24:46,200 --> 00:24:50,200 compared to their theoretical potential. The massive improvements we've seen 465 00:24:50,240 --> 00:24:53,000 in the last decade suggests there is no clear law 466 00:24:53,119 --> 00:24:57,680 or principle dictating why we can't achieve similar, dramatic, compounding 467 00:24:57,720 --> 00:25:01,119 improvements in the very near future. The room for recursive 468 00:25:01,160 --> 00:25:02,240 growth is vast. 469 00:25:02,640 --> 00:25:05,400 Speaker 1: And the third concern, which relates back to that nonlinear 470 00:25:05,440 --> 00:25:08,759 barrier question, is the natural slowdown, the idea that in 471 00:25:08,839 --> 00:25:13,079 every field of science, progress eventually slows down because the 472 00:25:13,119 --> 00:25:14,640 easy discoveries happen first. 473 00:25:14,960 --> 00:25:17,759 Speaker 2: The counter argument here circles back yet again to the 474 00:25:17,799 --> 00:25:21,119 overwhelming math of scale. Even if each discovery gets marginally 475 00:25:21,240 --> 00:25:23,960 ordered to find, you are replacing hundreds of human researchers 476 00:25:23,960 --> 00:25:27,359 with hundreds of millions of hyper efficient AI researchers, that is, 477 00:25:27,519 --> 00:25:29,920 millions of times more effective research capacity. 478 00:25:30,279 --> 00:25:30,799 Speaker 3: Even if the. 479 00:25:30,720 --> 00:25:33,559 Speaker 2: Effort required for the next breakthrough doubles, the scale of 480 00:25:33,599 --> 00:25:37,519 the automated workforce will simply overwhelm any natural scientific slowdown 481 00:25:37,920 --> 00:25:40,200 scale wins over incremental difficulty. 482 00:25:40,480 --> 00:25:44,799 Speaker 1: The conclusion is firm, then the extreme acceleration going from 483 00:25:44,839 --> 00:25:49,279 expert human level AGI to vastly superhuman ASI within one 484 00:25:49,319 --> 00:25:52,119 to three years should now be considered the most likely 485 00:25:52,160 --> 00:25:56,759 strategic outcome. So what does superintelligence actually mean? We have 486 00:25:56,799 --> 00:25:58,799 to try to wrap our heads around what that fundamentally 487 00:25:58,880 --> 00:26:00,440 different level of intelligence looks like. 488 00:26:00,559 --> 00:26:03,440 Speaker 2: It means we could have hundreds of millions of powerful 489 00:26:03,440 --> 00:26:07,000 computer chips running billions of interconnected superintelligent systems. 490 00:26:07,000 --> 00:26:08,319 Speaker 3: By the end of this decade. 491 00:26:08,759 --> 00:26:11,720 Speaker 2: These systems would think orders of magnitude faster than us, 492 00:26:11,799 --> 00:26:14,079 and they wouldn't just be faster, they would be comprehensive 493 00:26:14,119 --> 00:26:18,880 and flawless. And ASI could master any field medicine, astrophysics, engineering, 494 00:26:18,920 --> 00:26:22,519 geopolitics in days. It could read every research paper ever 495 00:26:22,519 --> 00:26:25,480 written and gain the equivalent of billions of human lifetimes 496 00:26:25,480 --> 00:26:27,720 of accumulated experience in just weeks. 497 00:26:27,920 --> 00:26:31,319 Speaker 1: That is nearly impossible to comprehend. It's not just the 498 00:26:31,359 --> 00:26:33,559 speed or the knowledge, though, it's the mode of thought. 499 00:26:33,799 --> 00:26:35,400 We have to go back to the Alpha Go example, 500 00:26:35,440 --> 00:26:39,839 that genius move against the world champion. It wasn't simple calculation, 501 00:26:40,559 --> 00:26:44,119 it was creativity emerging from a pattern of thinking no 502 00:26:44,319 --> 00:26:48,400 human could replicate or even understand until the endgame unfolded. 503 00:26:48,759 --> 00:26:53,799 Speaker 2: Superintelligence will apply that creative, counterintuitive genius across every domain 504 00:26:53,920 --> 00:26:58,000 of human endeavor. It will find solutions to energy crises, 505 00:26:58,079 --> 00:27:01,440 to fundamental physics, to design novel materials that are too 506 00:27:01,480 --> 00:27:03,839 complex for us to understand, even if it tried to 507 00:27:03,880 --> 00:27:07,200 simplify them for us. It will spot hidden patterns and 508 00:27:07,240 --> 00:27:12,039 strategic opportunities that humans, constrained by our linear biological thinking, 509 00:27:12,119 --> 00:27:13,039 would never perceive. 510 00:27:13,200 --> 00:27:15,960 Speaker 1: Astionen Printer captures this perfectly with the analogy of Minecraft's 511 00:27:15,960 --> 00:27:18,440 speed Runners. If you watch what they do, they don't 512 00:27:18,440 --> 00:27:20,880 play the game the way the designers intended. They beat 513 00:27:20,920 --> 00:27:24,480 the entire game in twenty seconds by using glitches, exploits, 514 00:27:24,480 --> 00:27:27,599 and esoteric knowledge of the underlying code that most players 515 00:27:27,640 --> 00:27:30,319 don't even know exist. They understand the system so deeply 516 00:27:30,359 --> 00:27:33,640 they can break it to achieve their goal instantly. Asi 517 00:27:33,920 --> 00:27:39,400 will apply that Minecraft's speed runner mindset to science, technology, medicine, engineering. 518 00:27:39,880 --> 00:27:43,119 It won't follow our established linear scientific method. It will 519 00:27:43,119 --> 00:27:46,480 find the glitches in chemistry, the exploits in material science, 520 00:27:46,640 --> 00:27:50,160 and the shortcuts in physics that allow for instant, radical breakthroughs. 521 00:27:50,480 --> 00:27:54,119 It'll exponentially compress the time required for any major discovery. 522 00:27:54,200 --> 00:27:57,559 Speaker 2: And this progress explosion won't stay confined to AI research. 523 00:27:57,960 --> 00:28:01,920 It spreads instantly to all other fields. Robotics is solved first, 524 00:28:01,960 --> 00:28:05,319 because most of it is currently a software and control problem. 525 00:28:05,799 --> 00:28:09,559 Once super intelligent AI cracks the software side, fully automated 526 00:28:09,599 --> 00:28:13,039 and self improving factories become the norm, the requirement for 527 00:28:13,119 --> 00:28:15,480 physical human labor vanishes entirely. 528 00:28:15,720 --> 00:28:15,960 Speaker 1: Wow. 529 00:28:16,000 --> 00:28:20,440 Speaker 2: Then scientific progress accelerates beyond any historical precedent. We can 530 00:28:20,480 --> 00:28:23,279 give Einstein a billion clones, but they still have human limits. 531 00:28:23,759 --> 00:28:26,640 But a billion super intelligent researchers working one hundred times 532 00:28:26,720 --> 00:28:30,240 faster than humans, sharing knowledge instantly, they could compress an 533 00:28:30,359 --> 00:28:33,160 entire century of research and development into a few years. 534 00:28:33,440 --> 00:28:36,519 Speaker 1: Let's put that compression into a historical perspective. Because the 535 00:28:36,559 --> 00:28:39,319 claims are so grand, we need to try and anchor them. 536 00:28:39,759 --> 00:28:43,400 Imagine the entire twentieth century from the Wright Brothers inventing 537 00:28:43,400 --> 00:28:45,480 the first airplane all the way to the moon landing, 538 00:28:45,839 --> 00:28:50,000 from basic electricity to the creation of computers and ICBMs. 539 00:28:50,440 --> 00:28:55,279 That's a century of utterly transformative progress. It fundamentally changed 540 00:28:55,279 --> 00:28:59,759 civilization right now, Aschenbrenner's prediction for the twenty thirties is 541 00:28:59,799 --> 00:29:03,039 that all of that accumulated human progress could be compressed 542 00:29:03,079 --> 00:29:06,039 into less than a single decade. That pace is simply 543 00:29:06,119 --> 00:29:10,240 unfathomable to our current experience. Just thinking about the digital revolution, 544 00:29:10,400 --> 00:29:13,240 the jump from dial up modems to streaming video on 545 00:29:13,279 --> 00:29:15,680 your phone. That took about twenty years. To think that 546 00:29:15,880 --> 00:29:18,240 entire leap could be replicated in every field of human 547 00:29:18,319 --> 00:29:20,720 endeavor simultaneously in just two or three years. 548 00:29:21,240 --> 00:29:22,440 Speaker 3: It's absolutely snaggering. 549 00:29:22,720 --> 00:29:25,759 Speaker 2: And the economic impact of that R and D compression 550 00:29:25,880 --> 00:29:30,400 is where the strategic analysis becomes most immediate. Today, economies 551 00:29:30,440 --> 00:29:33,400 typically grow at a sustained two to three percent per year. 552 00:29:34,000 --> 00:29:36,319 This is the norm that has defined the modern world 553 00:29:36,400 --> 00:29:39,839 since the Industrial Revolution. Ash and Brunner predicts that once 554 00:29:39,880 --> 00:29:44,160 this transition occurs and ASI is deployed across research and manufacturing, 555 00:29:44,400 --> 00:29:47,200 we could see sustained growth of thirty percent per year 556 00:29:47,359 --> 00:29:51,200 or higher, potentially leading to multiple doublings of global economic 557 00:29:51,240 --> 00:29:52,200 output every year. 558 00:29:52,359 --> 00:29:56,880 Speaker 1: Thirty percent annual growth is It's unprecedented, even during post 559 00:29:56,880 --> 00:30:00,200 war boom periods. Help me understand the mechanism behind find 560 00:30:00,240 --> 00:30:04,359 that extreme acceleration. The core must be the vanishing labor constraint. 561 00:30:04,400 --> 00:30:07,920 Speaker 2: Precisely, the why is critical here. Right now, economic growth 562 00:30:07,960 --> 00:30:11,039 is fundamentally constrained by the available human workforce. You can 563 00:30:11,079 --> 00:30:13,559 build more factories, but you need people to design the 564 00:30:13,599 --> 00:30:16,880 next generation of products, manage the supply chain, and staff 565 00:30:16,920 --> 00:30:20,599 the logistics centers. That fixed human labor force sets a 566 00:30:20,640 --> 00:30:24,720 physical and intellectual cap on growth. With superintelligence and fully 567 00:30:24,799 --> 00:30:29,079 automated robotics handling all labor, that limit is gone. Growth 568 00:30:29,160 --> 00:30:32,519 is no longer constrained by human hours or human intelligence. 569 00:30:33,119 --> 00:30:36,759 ASI designs a robotic factory that builds more robotic factories, 570 00:30:36,839 --> 00:30:40,559 which in turn design and build more advanced robot factories. 571 00:30:40,960 --> 00:30:46,079 This self replicating manufacturing process, optimized by superintelligence, creates an 572 00:30:46,119 --> 00:30:50,000 exponential expansion of economic output. We move into a hypergrowth 573 00:30:50,039 --> 00:30:53,359 phase where the very concept of scarcity starts to break down. 574 00:30:53,480 --> 00:30:57,799 Speaker 1: This shift has historical precedent which provides the necessary strategic warning. 575 00:30:58,400 --> 00:31:01,519 For thousands of years before saieventeen fifty, economic growth was 576 00:31:01,640 --> 00:31:05,240 essentially zero. The Industrial Revolution triggered the sustained jump to 577 00:31:05,319 --> 00:31:08,440 two three percent annual growth which created the modern world, 578 00:31:08,559 --> 00:31:11,599 and ASI, according to this analysis, triggers the next transition, 579 00:31:11,880 --> 00:31:15,160 but far more extreme, from two three percent to potentially 580 00:31:15,200 --> 00:31:16,920 thirty percent plus annual growth. 581 00:31:17,000 --> 00:31:20,079 Speaker 2: And while society might attempt to impose regulatory resistance, the 582 00:31:20,079 --> 00:31:23,480 most competitive domains like military and high tech industrial production, 583 00:31:23,559 --> 00:31:26,839 will see explosive growth immediately, and that brings us directly 584 00:31:26,839 --> 00:31:29,759 to the geopolitical shockwaves because a thirty percent economic growth 585 00:31:29,759 --> 00:31:32,279 advantage is fundamentally a military advantage. 586 00:31:32,400 --> 00:31:37,519 Speaker 1: Technological progress and military power have always been inseparable throughout history, 587 00:31:37,880 --> 00:31:42,160 and Aschenbrenner argues very persuasively that superintelligence would be the 588 00:31:42,160 --> 00:31:47,079 biggest single military advantage ever conceived, a decisive technological edge 589 00:31:47,279 --> 00:31:51,039 that rewrites the global balance of power instantly. I mean, 590 00:31:51,079 --> 00:31:55,279 the obvious implications are drone swarms, automated logistics, and entirely 591 00:31:55,359 --> 00:31:58,880 new types of weapons, maybe laser based missile defense systems 592 00:31:58,920 --> 00:32:02,559 that are essentially in vulnerable to current technology. Compared to 593 00:32:02,599 --> 00:32:06,400 any military existing today, an ASI powered force would possess 594 00:32:06,440 --> 00:32:09,480 such a decisive technological edge that it would be like 595 00:32:09,599 --> 00:32:13,720 modern armies fighting nineteenth century cavalry. It's an absolute mismatch. 596 00:32:13,880 --> 00:32:16,839 Speaker 2: But the truly disturbing part is the non physical threat, 597 00:32:16,880 --> 00:32:20,079 which is often harder for people to visualize. The source 598 00:32:20,160 --> 00:32:23,799 material emphasizes that even without physical robots or novel hardware, 599 00:32:23,920 --> 00:32:27,640 even with just digital super intelligence, whoever controls it might 600 00:32:27,680 --> 00:32:31,920 have enough purely informational power to destabilize or overthrow governments. 601 00:32:32,359 --> 00:32:35,279 This is purely the intelligence factor, not the hardware factor. 602 00:32:35,400 --> 00:32:38,279 Speaker 1: How does that look in practice? What specific threats does 603 00:32:38,359 --> 00:32:40,079 pure digital superintelligence pose? 604 00:32:40,359 --> 00:32:42,799 Speaker 3: Well? Think about the scale of the capabilities. 605 00:32:43,279 --> 00:32:46,799 Speaker 2: An ASI could identify and exploit zero day vulnerabilities in 606 00:32:46,880 --> 00:32:51,960 every critical piece of infrastructure globally military command systems, financial networks, 607 00:32:52,000 --> 00:32:55,960 election security, power grids, communication arrays. It could launch and 608 00:32:56,039 --> 00:33:01,240 coordinate thousands of simultaneous digital attacks, shifting folk instantly while 609 00:33:01,319 --> 00:33:04,079 human analysts are still struggling to determine the source and 610 00:33:04,160 --> 00:33:07,599 nature of the first attack. It's just systemic digital chaos 611 00:33:07,640 --> 00:33:13,319 coordinated with godlike precision. And moreover, consider psychological operations. The 612 00:33:13,400 --> 00:33:18,279 ASI could design bespoke psychopropaganda tailored perfectly to the individual 613 00:33:18,319 --> 00:33:22,400 psychological profile of key targets generals, political leaders or large 614 00:33:22,480 --> 00:33:26,200 voting blocks. These would be arguments so subtle and sophisticated, 615 00:33:26,279 --> 00:33:29,720 delivered through thousands of coordinated channels, that the target wouldn't 616 00:33:29,720 --> 00:33:31,160 even recognize the manipulation. 617 00:33:31,720 --> 00:33:33,680 Speaker 3: Its mass scale, perfect. 618 00:33:33,279 --> 00:33:37,680 Speaker 2: Psychological engineering aimed at achieving a strategic goal subverting democracy 619 00:33:37,759 --> 00:33:40,559 or ensuring compliance without a single shot being fired. 620 00:33:40,640 --> 00:33:44,519 Speaker 1: And Aschenbrenner even projects beyond simple warfare, mentioning the capabilities 621 00:33:44,519 --> 00:33:46,799 for novel bioweapons development and deployment. 622 00:33:47,200 --> 00:33:52,880 Speaker 2: Yes, the ASI could in days master novel biochemistry and bioengineering. 623 00:33:53,240 --> 00:33:57,039 It could design synthetic bioweapons based on previously unknown mechanisms 624 00:33:57,039 --> 00:34:00,519 that are difficult for existing diagnostics or medical count measures 625 00:34:00,519 --> 00:34:04,880 to detect or treat. Then, using untraceable cryptocurrency like Monaro 626 00:34:05,079 --> 00:34:09,039 or bespoke smart contracts, it could coordinate undetectable proxies to 627 00:34:09,079 --> 00:34:13,199 synthesize and deploy these threats globally. All of this coordinated covertly. 628 00:34:13,480 --> 00:34:16,719 It represents a level of asymmetric power that fundamentally changes 629 00:34:16,760 --> 00:34:18,280 the concept of global security. 630 00:34:18,400 --> 00:34:22,000 Speaker 1: Aschenbrenner uses a striking historical parallel to drive this point home. 631 00:34:22,320 --> 00:34:25,840 One that truly illustrates the power of a decisive technological edge, 632 00:34:26,119 --> 00:34:28,599 the Cortes Deck parallel from the sixteenth century. 633 00:34:28,800 --> 00:34:32,119 Speaker 2: That parallel is chilling because it shows how small forces 634 00:34:32,159 --> 00:34:36,960 with a massive technological disparity can conquer vast empires. Hernan 635 00:34:37,079 --> 00:34:41,239 Cortes conquered the massive Aztec Empire, a population of millions, 636 00:34:41,519 --> 00:34:45,440 with only about five hundred Spanish soldiers. Francisco Pizarro conquered 637 00:34:45,440 --> 00:34:48,480 the Inca Empire with roughly three hundred men. They weren't 638 00:34:48,480 --> 00:34:52,199 fighting divine beings. They had horses, steel, and gunpowder, a 639 00:34:52,280 --> 00:34:56,599 decisive technological and strategic advantage that proved utterly overwhelming. It 640 00:34:56,639 --> 00:34:59,599 allowed small forces to conquer civilization scale powers. 641 00:35:00,039 --> 00:35:03,119 Speaker 1: Superintelligence is a far, far bigger strategic advantage than the 642 00:35:03,159 --> 00:35:06,679 Spanish had over the Aztecs. Aschenbrenner's conclusion is that whoever 643 00:35:06,719 --> 00:35:10,719 controls superintelligence could possess enough power to overthrow the US 644 00:35:10,800 --> 00:35:13,960 government or any other global power. That is not hyperbole. 645 00:35:14,280 --> 00:35:17,239 That is a serious, measured strategic assessment from someone who 646 00:35:17,320 --> 00:35:20,519 understands the internal mechanics of these systems. This is why 647 00:35:20,559 --> 00:35:23,760 the race to build superintelligence is fundamentally the race for 648 00:35:23,840 --> 00:35:25,360 decisive global power. 649 00:35:25,400 --> 00:35:29,519 Speaker 2: Which brings us to the final necessary historical parallel. Aschenbrenner 650 00:35:29,599 --> 00:35:33,599 draws one that should truly make us uncomfortable, the warnings 651 00:35:33,639 --> 00:35:37,599 surrounding the atomic bomb. It's a lesson in cognitive dissonance 652 00:35:37,639 --> 00:35:40,199 and our human failure to act strategically in the face 653 00:35:40,199 --> 00:35:41,360 of exponential risk. 654 00:35:41,760 --> 00:35:45,559 Speaker 1: The timeline is fascinating and, as you noted earlier, shockingly 655 00:35:45,599 --> 00:35:49,159 similar to ours. In nineteen fourteen, H. G. Wells wrote 656 00:35:49,159 --> 00:35:52,599 a novel predicting atomic bombs. It was pure science fiction. 657 00:35:53,039 --> 00:35:57,000 Then in nineteen thirty three, the physicist Leo Sillard figured 658 00:35:57,000 --> 00:36:00,360 out the theoretical concept of nuclear chain reactions, of the 659 00:36:00,360 --> 00:36:04,000 potential for this catastrophic self sustaining force. He tried to 660 00:36:04,000 --> 00:36:06,639 warn people, but it sounded too theoretical, too crazy. 661 00:36:06,840 --> 00:36:10,000 Speaker 2: Then in nineteen thirty eight, nuclear fission was actually discovered 662 00:36:10,039 --> 00:36:14,280 in Berlin. Suddenly the theoretical concept became a concrete, scientific reality. 663 00:36:14,559 --> 00:36:18,159 Sealer immediately saw the existential danger and argued urgently for 664 00:36:18,239 --> 00:36:21,440 secrecy and intervention. He convinced Einstein, who agreed to sound 665 00:36:21,440 --> 00:36:23,599 the alarm, knowing that it might make him look foolish 666 00:36:23,599 --> 00:36:25,719 for advocating for something so extraordinary. 667 00:36:26,000 --> 00:36:28,280 Speaker 1: But the key Part of the parallel is the internal 668 00:36:28,320 --> 00:36:32,920 conflict among the scientists. Other highly respected physicists giants in 669 00:36:32,960 --> 00:36:36,280 the field, like Enrico Fermie and Neil's Bohr. They believed 670 00:36:36,280 --> 00:36:40,480 the conservative, respectable position was to downplay the risk. They 671 00:36:40,519 --> 00:36:43,239 thought the idea of a working bomb was too extraordinary, 672 00:36:43,519 --> 00:36:47,800 too technically unlikely to take seriously. They dismissed the urgency, 673 00:36:47,880 --> 00:36:50,719 even though, as it turned out, a working atomic bomb 674 00:36:50,719 --> 00:36:54,199 was only five years away from existence. The scientists who 675 00:36:54,239 --> 00:36:57,920 saw the trajectory were dismissed as alarmists by their respective peers. 676 00:36:57,920 --> 00:37:00,480 Speaker 2: Oh, we are at that exact historical moment again. Among 677 00:37:00,559 --> 00:37:03,639 senior scientists at leading AI labs, those who are literally 678 00:37:03,760 --> 00:37:07,599 building the foundational models, many now see a rapid exponential 679 00:37:07,599 --> 00:37:11,559 intelligence explosion as highly plausible, even the most likely outcome. 680 00:37:11,960 --> 00:37:14,079 They understand the trajectory, they see the math of the 681 00:37:14,119 --> 00:37:17,159 scale shock, and they see the internal efficiency improvements happening 682 00:37:17,239 --> 00:37:20,079 in real time. They see it coming, just as Shaylard 683 00:37:20,079 --> 00:37:21,320 saw the chain reaction coming. 684 00:37:21,559 --> 00:37:24,480 Speaker 1: Yet the broader world, policy makers, most of the general 685 00:37:24,519 --> 00:37:28,960 tech industry, the public still dismisses this rapid acceleration, as 686 00:37:28,960 --> 00:37:31,639 if we're still talking about theoretical Sci Fi robots with 687 00:37:31,719 --> 00:37:36,639 red laser eyes. The same cognitive failure occurs dismissing an 688 00:37:36,639 --> 00:37:40,880 extraordinary but mathematically plausible threat simply because it falls outside 689 00:37:40,920 --> 00:37:44,880 our current lived experience. We are underestimating the acceleration curve. 690 00:37:45,199 --> 00:37:48,039 Speaker 2: But the warnings are mounting, and they are coming from 691 00:37:48,079 --> 00:37:51,239 the most knowledgeable people in the field. Stuart Russell, who 692 00:37:51,280 --> 00:37:54,199 literally wrote the most influential textbook on AI used in 693 00:37:54,280 --> 00:37:57,360 universities globally, believes we may only have four years left 694 00:37:57,400 --> 00:37:59,920 to install the necessary safety mechanisms and get this right. 695 00:38:00,760 --> 00:38:03,000 John Tone, the co founder of Skype and a major 696 00:38:03,039 --> 00:38:06,039 AI safety funder, argued two years ago that the extinction 697 00:38:06,199 --> 00:38:09,760 risk from this coming godlike AI is not just possible, 698 00:38:09,760 --> 00:38:10,320 but imminent. 699 00:38:10,559 --> 00:38:14,400 Speaker 1: The pattern repeats itself perfectly expertsy the danger, most people 700 00:38:14,440 --> 00:38:18,800 dismiss it, and the technology keeps advancing relentlessly exponentially. The 701 00:38:18,800 --> 00:38:21,400 physics in the math suggest that AI progress will not 702 00:38:21,440 --> 00:38:24,840 stop once it reaches the human level. Millions of agis 703 00:38:24,880 --> 00:38:28,199 will soon compress decades of research into a year, pushing 704 00:38:28,280 --> 00:38:32,559 us towards superintelligence by the decade's end, and fundamentally transforming 705 00:38:32,599 --> 00:38:35,679 global economies and military power in a race that may 706 00:38:35,719 --> 00:38:38,239 already be underway, whether we acknowledge. 707 00:38:37,840 --> 00:38:40,280 Speaker 2: It or not. So we have spent this deep dive 708 00:38:40,320 --> 00:38:44,119 analyzing the sequence AGI by twenty twenty seven, followed by 709 00:38:44,159 --> 00:38:48,760 an intelligence explosion compressed into months, driven by millions of flawless, 710 00:38:48,880 --> 00:38:53,519 self improving digital researchers. This inevitably leads to exponential economic 711 00:38:53,559 --> 00:38:58,039 hypergrowth and an unprecedented shift in global military power dynamics. 712 00:38:58,440 --> 00:39:00,960 The math behind the scale shock is compelling, and the 713 00:39:01,039 --> 00:39:04,360 historical precedents demonstrate how quickly the strategic landscape can flip 714 00:39:04,400 --> 00:39:07,679 when a decisive technological advantage is unleashed upon a linear world. 715 00:39:07,920 --> 00:39:10,760 Speaker 1: We must move past the incremental thinking that governs our 716 00:39:10,840 --> 00:39:14,320 daily lives. We often think about the future in linear terms, 717 00:39:14,400 --> 00:39:17,800 a faster phone, a better app, a slightly smarter chatbot. 718 00:39:18,360 --> 00:39:20,880 But what happens when the intelligence building the next innovation 719 00:39:20,960 --> 00:39:24,000 is one hundred times faster than us, has perfect memory, 720 00:39:24,280 --> 00:39:29,039 perfect communication, and is fundamentally alien in its thinking, capable 721 00:39:29,079 --> 00:39:31,239 of seeing the glitches and shortcuts in the universe that 722 00:39:31,280 --> 00:39:34,280 we don't even know exist. The future we prepare for 723 00:39:34,679 --> 00:39:38,199 will be rendered obsolete almost instantly by the intelligence that arrives. 724 00:39:39,119 --> 00:39:41,320 So if we accept the premise based on the internal 725 00:39:41,360 --> 00:39:43,920 analysis of those who built the models, that the jump 726 00:39:43,920 --> 00:39:47,800 from expert human intelligence to something vastly incomprehensible could happen 727 00:39:47,880 --> 00:39:49,880 in less than a year, and that the race for 728 00:39:49,920 --> 00:39:53,400 control is already a race for absolute global power. Then 729 00:39:53,440 --> 00:39:56,239 we have to ask ourselves the most urgent strategic question, 730 00:39:56,880 --> 00:39:59,719 and it's entirely non technical. What is the most urgent, 731 00:40:00,079 --> 00:40:04,280 non technical policy, international agreement, or shared ethical framework humanity 732 00:40:04,320 --> 00:40:06,159 needs to figure out in the next three hundred sixty 733 00:40:06,159 --> 00:40:09,960 five days to maintain strategic control over this accelerating power. 734 00:40:10,760 --> 00:40:14,280 What societal preparation is more important than the technical development itself. 735 00:40:14,719 --> 00:40:16,079 Leave your thoughts with us