1 00:00:00,120 --> 00:00:03,200 Speaker 1: Imagine a technology that is scaling so fast, I mean 2 00:00:03,359 --> 00:00:06,839 so incredibly fast that the very people building it are 3 00:00:06,879 --> 00:00:09,439 starting to use some really alarming comparisons. 4 00:00:09,560 --> 00:00:11,919 Speaker 2: Alarming is the right word. We're not talking about the 5 00:00:11,919 --> 00:00:14,919 next smartphone or a faster Internet, no, not at all. 6 00:00:15,160 --> 00:00:18,519 Speaker 1: The comparison they're making is to the invention of the 7 00:00:18,600 --> 00:00:22,000 nuclear bomb. Some are even saying AI could be more 8 00:00:22,079 --> 00:00:26,280 dangerous because you're adding eponymy and this kind of unpredictable 9 00:00:26,320 --> 00:00:27,519 growth into the mix. 10 00:00:27,679 --> 00:00:30,160 Speaker 2: And according to a former CEO who is right there 11 00:00:30,199 --> 00:00:32,640 in the middle of it all, we are and this 12 00:00:32,679 --> 00:00:34,920 is a treat quote running out of time. 13 00:00:35,200 --> 00:00:38,359 Speaker 1: Welcome to thrilling threads. That quote you just heard comes 14 00:00:38,359 --> 00:00:40,479 from a recent and I have to say, pretty sobering 15 00:00:40,520 --> 00:00:42,399 interview with a former Google CEO. 16 00:00:42,520 --> 00:00:44,159 Speaker 2: Yeah, it was featured on the Diary of a CEO 17 00:00:44,240 --> 00:00:46,880 clips on YouTube, and we've really dug into it. We've 18 00:00:46,920 --> 00:00:50,200 pulled apart the key insights from that conversation. 19 00:00:49,799 --> 00:00:52,320 Speaker 1: And it paints a picture that is, you know, it's 20 00:00:52,439 --> 00:00:55,799 on one hand, full of this massive world changing potential, 21 00:00:56,039 --> 00:01:01,840 but on the other it's coupled with extreme, almost existential danger. 22 00:01:02,159 --> 00:01:05,359 Speaker 2: So our mission today is to give you the learner 23 00:01:05,719 --> 00:01:08,519 a critical briefing on what's coming. We're talking about the 24 00:01:08,560 --> 00:01:09,959 immediate future, right. 25 00:01:10,079 --> 00:01:12,719 Speaker 1: This isn't about just scrolling past headlines. We want to 26 00:01:12,799 --> 00:01:15,680 synthesize the real complexity of what the next five years 27 00:01:15,680 --> 00:01:19,359 could look like. We're going to unpack the most urgent claims, 28 00:01:19,680 --> 00:01:24,719 everything from these so called day zero cyber threats, to 29 00:01:26,000 --> 00:01:27,680 the future of our jobs. 30 00:01:27,359 --> 00:01:29,519 Speaker 2: And even why the ultimate off switch for all of 31 00:01:29,519 --> 00:01:32,200 this might just be a big physical power plug. 32 00:01:32,319 --> 00:01:34,599 Speaker 1: It all comes down to speed, doesn't it. The pace 33 00:01:34,599 --> 00:01:37,640 of development is the one variable that dictates everything else. 34 00:01:37,760 --> 00:01:39,280 Speaker 2: It is. If you can get your head on the speed, 35 00:01:39,319 --> 00:01:41,879 you understand the urgency. That's where we have to start. 36 00:01:41,920 --> 00:01:43,599 This concept of exponential growth. 37 00:01:43,680 --> 00:01:46,760 Speaker 1: Okay, so let's get into that. This engine of acceleration. 38 00:01:47,120 --> 00:01:49,760 The entire warning, everything we're about to discuss, it's all 39 00:01:49,760 --> 00:01:52,879 built on this one idea of speed. The x CEO 40 00:01:53,359 --> 00:01:55,439 use this this really great. 41 00:01:55,239 --> 00:01:56,840 Speaker 2: Metaphor, the turns of the crank. 42 00:01:57,159 --> 00:01:59,640 Speaker 1: Exactly, the turns of the crank. It sounds simple, like 43 00:01:59,680 --> 00:02:01,760 something from an old factory. But what does that actually 44 00:02:01,840 --> 00:02:05,200 mean when you translate it into you know, raw computing power. 45 00:02:05,280 --> 00:02:07,840 Speaker 2: Well, it sounds like a simple analogy, but what's striking 46 00:02:07,920 --> 00:02:13,639 is that it's rooted in observed scientific reality. The forecast, 47 00:02:14,080 --> 00:02:17,039 his forecast is that in the next five years we're 48 00:02:17,039 --> 00:02:19,400 going to see two or three more of these major 49 00:02:19,680 --> 00:02:23,000 turns of the crank in how these large models are developed. 50 00:02:23,120 --> 00:02:25,080 Speaker 1: And just to be clear, this isn't like a new 51 00:02:25,280 --> 00:02:27,919 iPhone model coming out. It's not just a little bit better. 52 00:02:28,159 --> 00:02:30,719 Speaker 2: No, not at all. This isn't sequential product development. This 53 00:02:30,840 --> 00:02:34,919 is pure exponential scaling. It's a performance curve where the 54 00:02:34,960 --> 00:02:38,840 capability you know, doubles or even quadruples every single time 55 00:02:38,919 --> 00:02:41,520 they increase the training data and the computing power. 56 00:02:42,000 --> 00:02:44,840 Speaker 1: And the really worrying part, as I understand it is 57 00:02:44,919 --> 00:02:48,960 tied to something called scaling laws. These are what formulas 58 00:02:49,000 --> 00:02:52,360 the researchers use to predict how much smarter the AI gets. 59 00:02:52,479 --> 00:02:55,560 Speaker 2: That's the core idea. Yeah, Scaling laws are these mathematical 60 00:02:55,599 --> 00:02:59,159 models that describe the relationship between three things. The size 61 00:02:59,159 --> 00:03:01,800 of the model itself, so the number of parameters, the 62 00:03:01,840 --> 00:03:04,199 size of the data set, which is how much information 63 00:03:04,240 --> 00:03:07,400 it's eating basically, and then the amount of raw computation 64 00:03:07,520 --> 00:03:10,479 power you throw at it. And the critical point the 65 00:03:10,520 --> 00:03:14,319 expert makes is that and I'm quoting here, there is 66 00:03:14,800 --> 00:03:19,719 no evidence that the scaling laws have begun to stop. 67 00:03:19,879 --> 00:03:21,159 Speaker 1: So we haven't hit the wall yet. 68 00:03:21,280 --> 00:03:23,840 Speaker 2: We haven't even seen the wall. Usually with technology you 69 00:03:23,919 --> 00:03:27,759 hit some kind of physical limit diminishing returns. But here 70 00:03:27,919 --> 00:03:30,000 the ceiling just keeps getting higher and higher. 71 00:03:30,120 --> 00:03:32,719 Speaker 1: And that's the alarming part because if that formula is 72 00:03:32,759 --> 00:03:36,280 still working, we're not just looking at small incremental gains. 73 00:03:36,319 --> 00:03:38,240 This is compounding. 74 00:03:37,560 --> 00:03:41,520 Speaker 2: Growth exactly, And the numbers he uses, this is what 75 00:03:41,599 --> 00:03:46,039 should really grab your attention. Each crank. So each major 76 00:03:46,120 --> 00:03:49,120 model update represents something like a factor of two, three, 77 00:03:49,199 --> 00:03:51,479 maybe even four in capability improvement. 78 00:03:51,879 --> 00:03:54,879 Speaker 1: So if you compound that two or three times over 79 00:03:54,919 --> 00:03:55,680 five years. 80 00:03:55,560 --> 00:03:57,800 Speaker 2: You're not looking at a system that's fifty percent better. 81 00:03:58,080 --> 00:04:01,639 You're looking at systems that could be fifty or even 82 00:04:01,680 --> 00:04:03,560 one hundred times more powerful than what we have today. 83 00:04:03,599 --> 00:04:06,199 Speaker 1: One hundred times more powerful. That that sounds like marketing hype, 84 00:04:06,360 --> 00:04:08,560 you know, it's such a big number. What's the real 85 00:04:08,599 --> 00:04:10,960 world difference between the kind of helpful AI we use 86 00:04:11,000 --> 00:04:13,759 today and this this thing that's one hundred times better. 87 00:04:13,960 --> 00:04:16,959 Speaker 2: That's a fair challenge we have to ground it. Today's 88 00:04:17,000 --> 00:04:20,319 AI is fantastic, you know, summarizing things and mimicking how 89 00:04:20,319 --> 00:04:24,680 we talk, but it really struggles with deep, verifiable reasoning, 90 00:04:25,560 --> 00:04:28,759 especially in complex fields. Okay, a one hundred x improvement 91 00:04:28,879 --> 00:04:33,759 changes that completely. Think about drug discovery. Today, an AI 92 00:04:33,879 --> 00:04:37,519 might suggest a few molecules that could work. Tomorrow, a 93 00:04:37,600 --> 00:04:41,360 one hundred XAI could design and fully validate dozens of 94 00:04:41,399 --> 00:04:44,360 new drugs basically bypassing years of lab work. 95 00:04:44,519 --> 00:04:46,759 Speaker 1: So it's not just about writing an email faster. It's 96 00:04:46,759 --> 00:04:49,360 about solving problems that right now take a whole team 97 00:04:49,399 --> 00:04:51,480 of human experts years to figure out. 98 00:04:51,560 --> 00:04:55,480 Speaker 2: Precisely, it's moving into things that require deep layered expertise. 99 00:04:55,800 --> 00:04:57,759 We're already seeing hints of it even now. You look 100 00:04:57,759 --> 00:04:59,680 at research groups like oh what one or what op 101 00:04:59,720 --> 00:05:03,360 ami doing, where models are getting visibly better at things 102 00:05:03,360 --> 00:05:06,040 that used to be AI's weak spot, like what advance 103 00:05:06,120 --> 00:05:09,680 physics simulations, high level mathematics, things that require these long, 104 00:05:09,800 --> 00:05:12,759 complex chains of logic. So if today's systems are already 105 00:05:12,759 --> 00:05:15,360 getting a foothold there, imagine what a system one hundred 106 00:05:15,399 --> 00:05:18,639 times more powerful could do. It could essentially master fields 107 00:05:18,639 --> 00:05:20,800 that currently take a human lifetime of study. 108 00:05:21,000 --> 00:05:24,240 Speaker 1: That completely changes the game. Well for any professional knowledge 109 00:05:24,240 --> 00:05:27,720 worker it's not a tool to help you search anymore. No, 110 00:05:28,040 --> 00:05:31,480 it becomes a fundamental, maybe even competing agent in every 111 00:05:31,560 --> 00:05:32,399 single industry. 112 00:05:32,560 --> 00:05:35,000 Speaker 2: And that's the transition you as a learner have to 113 00:05:35,040 --> 00:05:38,279 prepare for. The whole idea of expertise is changing. It's 114 00:05:38,319 --> 00:05:41,600 not about memorizing information anymore. The AI will do that. 115 00:05:42,279 --> 00:05:45,360 It shifts the human role to defining the goals and 116 00:05:45,879 --> 00:05:49,199 maybe most importantly, applying moral judgment. 117 00:05:49,079 --> 00:05:50,720 Speaker 1: Which, as we'll get into, is the one thing the 118 00:05:50,759 --> 00:05:51,560 AI doesn't have. 119 00:05:51,879 --> 00:05:54,600 Speaker 2: Right. But that blinding speed is why we have to 120 00:05:54,600 --> 00:05:57,480 talk about the immediate dangers, the unfiltered risks. 121 00:05:57,720 --> 00:06:01,240 Speaker 1: Okay, so if we accept this baseline, this one hundred 122 00:06:01,360 --> 00:06:03,879 x acceleration in the next five years, we have to 123 00:06:03,920 --> 00:06:06,000 look at the downside. What can that kind of raw, 124 00:06:06,079 --> 00:06:08,000 unfiltered power actually do. 125 00:06:08,680 --> 00:06:11,680 Speaker 2: That capability leads us straight to what the XCEO calls 126 00:06:11,839 --> 00:06:16,000 the existential threat stack. He outlines three major immediate dangers, 127 00:06:16,040 --> 00:06:19,959 and they are cyber threats, the potential for creating biological weapons, 128 00:06:20,199 --> 00:06:23,240 and a complete revolution in how we conduct warfare, and 129 00:06:23,279 --> 00:06:25,360 all of them are defined by that speed in autonomy 130 00:06:25,360 --> 00:06:26,360 we were just talking about. 131 00:06:26,600 --> 00:06:29,360 Speaker 1: Let's start with the one that feels the most immediate 132 00:06:29,439 --> 00:06:34,519 and honestly the most terrifying cyber attacks. Specifically, he mentions 133 00:06:34,759 --> 00:06:38,319 the day zero attack. What exactly does that mean? 134 00:06:38,480 --> 00:06:41,040 Speaker 2: A day zero attack is an attack that uses a 135 00:06:41,120 --> 00:06:44,680 zero day vulnerability. That means it's a weakness in a system, software, 136 00:06:44,680 --> 00:06:48,399 a network, whatever that the maker doesn't even know exists yet. 137 00:06:48,240 --> 00:06:51,199 Speaker 1: So there's no patch for it, no defense. 138 00:06:51,040 --> 00:06:53,959 Speaker 2: Exactly if the vendor and all the security exerts have 139 00:06:54,000 --> 00:06:56,120 no idea, the whole is there. It's a zero day 140 00:06:56,439 --> 00:06:58,879 It's the ultimate stealth weapon because no one is prepared 141 00:06:58,920 --> 00:06:59,199 for it. 142 00:06:59,800 --> 00:07:04,199 Speaker 1: The claim here is that the raw, unfiltered AI models, 143 00:07:05,199 --> 00:07:09,120 they're already better than humans at finding these invisible weaknesses. 144 00:07:09,160 --> 00:07:11,879 Speaker 2: That's the evidence coming out of the labs. The raw models, 145 00:07:11,920 --> 00:07:14,040 which are the versions before they get all the safety 146 00:07:14,079 --> 00:07:16,920 filters put on them for public release, can perform these 147 00:07:17,000 --> 00:07:20,160 day zero attacks as well as or even better than 148 00:07:20,399 --> 00:07:21,519 the best human hackers. 149 00:07:22,000 --> 00:07:22,439 Speaker 1: Wow. 150 00:07:22,560 --> 00:07:24,560 Speaker 2: What gives them the edge is what he calls the 151 00:07:24,600 --> 00:07:26,920 machine advantage. Think about a human hacker, even the most 152 00:07:26,920 --> 00:07:29,399 brilliant one. They need to sleep, they get tired, they 153 00:07:29,480 --> 00:07:30,879 run on caffeine. 154 00:07:30,439 --> 00:07:31,279 Speaker 1: They have limits. 155 00:07:31,680 --> 00:07:34,959 Speaker 2: Exactly the AI systems, he says, just keep trying because 156 00:07:34,959 --> 00:07:37,079 they're computers and they have nothing else to do. They 157 00:07:37,079 --> 00:07:40,040 don't sleep, they don't eat, They just keep going. They 158 00:07:40,040 --> 00:07:43,720 have this infinite capacity for brute force discovery, trying billions 159 00:07:43,759 --> 00:07:47,279 of combinations across huge code bases without ever getting tired. 160 00:07:47,439 --> 00:07:51,519 Speaker 1: It's an asymmetric advantage. It takes the biggest constraint, human 161 00:07:51,560 --> 00:07:53,439 physical limitation right. 162 00:07:53,360 --> 00:07:56,439 Speaker 2: Out of the equation, and that same advantage that speed 163 00:07:56,480 --> 00:07:59,800 and brute force power feeds directly into the second big. 164 00:07:59,600 --> 00:08:01,439 Speaker 1: Concern, biological threats. 165 00:08:01,560 --> 00:08:06,040 Speaker 2: The virus factory another deeply unsettling phrase, especially after the 166 00:08:06,079 --> 00:08:09,199 last few years. The source notes that viruses are already 167 00:08:09,519 --> 00:08:12,199 relatively easy to make today, so. 168 00:08:12,079 --> 00:08:14,040 Speaker 1: The fear is that an AI could act as a 169 00:08:14,040 --> 00:08:18,000 super intelligent research assistant for someone with bad intentions, or. 170 00:08:17,879 --> 00:08:21,079 Speaker 2: Worse, it could facilitate the creation of, in his words, 171 00:08:21,519 --> 00:08:25,040 really bad viruses. The specific concern isn't just that it 172 00:08:25,079 --> 00:08:28,399 could find existing recipes. It's about optimization. What do you 173 00:08:28,399 --> 00:08:31,839 mean by optimization, Well, the raw model has read everything right, 174 00:08:32,519 --> 00:08:37,480 decades of biological research, complex scientific papers. It might be 175 00:08:37,519 --> 00:08:40,240 able to see connections and pathways to make a pathogen 176 00:08:40,320 --> 00:08:43,559 more infectious or more deadly that a human researcher would 177 00:08:43,559 --> 00:08:46,679 never spot. If it's one hundred times better at physics, 178 00:08:46,840 --> 00:08:50,639 it's probably one hundred times better at optimizing biological processes. 179 00:08:50,120 --> 00:08:53,240 Speaker 1: Too, So it could basically generate the entire recipe for 180 00:08:53,279 --> 00:08:56,399 a new dangerous virus from scratch. 181 00:08:56,480 --> 00:08:58,840 Speaker 2: That is the core fear, and the fact that the 182 00:08:58,919 --> 00:09:02,159 industry is already set up special commissions and safety teams 183 00:09:02,360 --> 00:09:06,279 focused specifically on this misuse. That tells you how seriously 184 00:09:06,320 --> 00:09:09,440 they're taking it Internally. They know what these raw models 185 00:09:09,480 --> 00:09:09,759 can do. 186 00:09:09,960 --> 00:09:13,000 Speaker 1: Okay, So that leads to the third threat, the changing 187 00:09:13,039 --> 00:09:15,799 logic of warfare. This one isn't theoretical. We can already 188 00:09:15,799 --> 00:09:16,440 see it happening. 189 00:09:16,519 --> 00:09:19,200 Speaker 2: We can We're moving away from the historical idea of 190 00:09:19,639 --> 00:09:23,320 you know, soldiers in trenches Mandaman conflict. The new paradigm 191 00:09:23,440 --> 00:09:24,320 is completely different. 192 00:09:24,360 --> 00:09:26,200 Speaker 1: He calls it the drone world exactly. 193 00:09:26,679 --> 00:09:29,639 Speaker 2: The soldier is replaced by an operator in a command center, 194 00:09:29,799 --> 00:09:33,080 maybe thousands of miles away, who is causing harm while 195 00:09:33,200 --> 00:09:38,440 quote drinking your coffee that distance. It completely changes the 196 00:09:38,440 --> 00:09:40,759 psychological and moral reality of war. 197 00:09:41,000 --> 00:09:43,759 Speaker 1: It becomes more like a video game. But the economic 198 00:09:43,840 --> 00:09:46,639 part is maybe the most brutal aspect of this shift. 199 00:09:46,879 --> 00:09:51,360 Speaker 2: The kill ratio is just it's redefining everything. The statistic 200 00:09:51,399 --> 00:09:54,000 in the source is that a five thousand dollars drone 201 00:09:54,320 --> 00:09:56,519 can now destroy a five million dollar. 202 00:09:56,480 --> 00:09:58,200 Speaker 1: Tank one thousand to one advantage. 203 00:09:58,240 --> 00:10:01,279 Speaker 2: It's an insane economic imbalance. It forces every military in 204 00:10:01,320 --> 00:10:04,279 the world to rethink its strategy. Why spend billions on 205 00:10:04,360 --> 00:10:07,679 heavy armor when a swarm of cheap, smart autonomous drones 206 00:10:07,720 --> 00:10:09,840 can make it obsolete in an instant. 207 00:10:09,639 --> 00:10:11,840 Speaker 1: And he points to the conflict in Ukraine as a 208 00:10:11,960 --> 00:10:14,480 real time laboratory for this. It's not just a war, 209 00:10:14,519 --> 00:10:16,919 it's the invention of a new kind of warfare. 210 00:10:17,200 --> 00:10:20,960 Speaker 2: It's all about drone on drone fighting, electronic warfare, network command. 211 00:10:21,279 --> 00:10:24,039 It's shifting the whole battlefield from who has the biggest 212 00:10:24,080 --> 00:10:26,799 tanks to who has the smartest algorithms and can build 213 00:10:26,879 --> 00:10:28,080 cheap drones the fastest. 214 00:10:28,200 --> 00:10:31,279 Speaker 1: It creates a situation where your national security depends on 215 00:10:31,320 --> 00:10:34,799 your technological speed. It's not just about economics anymore. It's 216 00:10:34,799 --> 00:10:35,639 about survival. 217 00:10:35,919 --> 00:10:38,879 Speaker 2: And to really understand these risks, you have to understand 218 00:10:39,000 --> 00:10:42,440 the engine that's driving them. You mentioned the raw models 219 00:10:42,480 --> 00:10:45,440 a couple of times. We need to unpack that because 220 00:10:45,480 --> 00:10:48,440 it's a concept the public rarely sees. But it's where 221 00:10:48,480 --> 00:10:50,639 all this power, it, all this danger comes from. 222 00:10:50,759 --> 00:10:53,039 Speaker 1: Let's start with the scale of it. How do you 223 00:10:53,080 --> 00:10:55,279 even begin to build one of these things. 224 00:10:55,440 --> 00:11:00,360 Speaker 2: It starts with data, truly unimaginable amounts of data. Process 225 00:11:00,440 --> 00:11:03,919 is described as the AI sucking all the information in, 226 00:11:04,320 --> 00:11:07,360 and the general belief among the top researchers is that 227 00:11:07,639 --> 00:11:10,759 they have now quote sucked all of the written word 228 00:11:10,799 --> 00:11:11,519 that's available. 229 00:11:11,600 --> 00:11:14,519 Speaker 1: Everything, every book, every article, every website. 230 00:11:14,080 --> 00:11:16,480 Speaker 2: Every digitized piece of public text they can get their 231 00:11:16,480 --> 00:11:19,960 hands on, has likely been fed into these foundational models. 232 00:11:20,159 --> 00:11:23,440 Speaker 1: The infrastructure for that must be just mind boggling. We're 233 00:11:23,440 --> 00:11:27,200 talking about supercomputers with enormous memories, which. 234 00:11:26,960 --> 00:11:29,799 Speaker 2: Brings us directly to the chip makers. We have to 235 00:11:29,799 --> 00:11:33,519 talk about Nvidia. The source points out that they're now 236 00:11:33,559 --> 00:11:36,159 one of the most valuable companies in the world. Why 237 00:11:37,120 --> 00:11:41,080 because they are so central to this revolution. The physical 238 00:11:41,080 --> 00:11:42,679 hardware is the bottleneck. 239 00:11:43,000 --> 00:11:46,039 Speaker 1: That's a really interesting point. The hardware is the power source, 240 00:11:46,200 --> 00:11:48,279 which means it's also the point of control. 241 00:11:48,519 --> 00:11:52,080 Speaker 2: It's a key insight in Video's GPUs are basically the 242 00:11:52,240 --> 00:11:55,120 currency of this new era. Your ability to build the 243 00:11:55,159 --> 00:11:57,879 next great AI is limited by your ability to get 244 00:11:57,919 --> 00:11:59,960 your hands on enough of their chips and run them. 245 00:12:00,240 --> 00:12:02,279 Speaker 1: So once you have the hardware and you fet it 246 00:12:02,360 --> 00:12:04,960 all this data, the training process just runs for months. 247 00:12:05,039 --> 00:12:07,360 How do the engineers know when it's done. 248 00:12:07,519 --> 00:12:11,559 Speaker 2: There's a technical measurement they watch constantly called the loss function. Essentially, 249 00:12:11,600 --> 00:12:13,360 they just wait two hundred and four seven until this 250 00:12:13,440 --> 00:12:16,279 loss function hits a certain number they've decided on. When 251 00:12:16,279 --> 00:12:18,559 it hits that target, they say good enough, and they stop. 252 00:12:18,679 --> 00:12:21,159 Speaker 1: Okay, what is a loss function? Can you simplify that? 253 00:12:21,559 --> 00:12:24,720 Speaker 2: Yeah? Absolutely? The loss function is basically a measure of error. 254 00:12:25,200 --> 00:12:27,799 During training, you give the model a prompt and you 255 00:12:27,879 --> 00:12:31,120 compare its answer to the right answer. The loss function 256 00:12:31,240 --> 00:12:33,559 is a number that represents how wrong the model was. 257 00:12:34,000 --> 00:12:37,039 The entire goal of the training is to get that number, 258 00:12:37,080 --> 00:12:39,240 that error rate as low as possible. 259 00:12:39,759 --> 00:12:42,799 Speaker 1: So the AI doesn't care about being smart or good 260 00:12:42,960 --> 00:12:45,360 in a human sense. It just cares about reducing that 261 00:12:45,440 --> 00:12:46,080 one number. 262 00:12:46,200 --> 00:12:50,720 Speaker 2: Precisely, it's optimizing for statistical accuracy, not for human values 263 00:12:50,799 --> 00:12:53,799 or safety. And the result of that whole process is 264 00:12:53,879 --> 00:12:58,120 the raw model. That's the pure, unfiltered intelligence before any 265 00:12:58,200 --> 00:12:59,639 human guardrails are put in place. 266 00:13:00,000 --> 00:13:01,759 Speaker 1: That's when the safety teams come in, right. 267 00:13:01,799 --> 00:13:04,600 Speaker 2: They immediately start testing it. They find quote, all sorts 268 00:13:04,639 --> 00:13:06,360 of bad things, and then they have to spend a 269 00:13:06,440 --> 00:13:09,039 huge amount of time programming it not to answer dangerous 270 00:13:09,080 --> 00:13:10,879 questions before they can release. 271 00:13:10,600 --> 00:13:12,279 Speaker 1: It to you and me, which brings up maybe the 272 00:13:12,279 --> 00:13:17,879 most fascinating and scary part of this whole thing, emergent behavior. 273 00:13:18,159 --> 00:13:20,679 Speaker 2: Here's where it gets really interesting. Yeah, how do you 274 00:13:20,799 --> 00:13:23,720 test for something you don't even know the system can do? Yeah? 275 00:13:23,840 --> 00:13:26,200 Over the next five years, these systems are going to 276 00:13:26,279 --> 00:13:31,559 learn things that nobody, not the developers, not society, nobody 277 00:13:31,559 --> 00:13:32,159 knows they know. 278 00:13:32,720 --> 00:13:35,480 Speaker 1: How do you even test for unknown knowledge? It sounds 279 00:13:35,559 --> 00:13:36,200 like a paradox. 280 00:13:36,279 --> 00:13:40,120 Speaker 2: It's the core challenge of AI safety right now. The 281 00:13:40,200 --> 00:13:44,240 solution involves, as the source puts it, incredibly clever people 282 00:13:44,440 --> 00:13:47,360 who sit there and literally fiddle with the networks to 283 00:13:47,399 --> 00:13:50,440 see what they can find. They're basically poking and prodding 284 00:13:50,519 --> 00:13:53,559 the system from every angle, trying to get it to 285 00:13:53,600 --> 00:13:55,720 reveal a hidden skill it learned on its own. 286 00:13:55,960 --> 00:13:58,360 Speaker 1: And there's a perfect example of this in the source material, 287 00:13:58,399 --> 00:14:00,480 the one where an AI was just shown a surable 288 00:14:00,480 --> 00:14:01,080 website and. 289 00:14:01,039 --> 00:14:02,879 Speaker 2: It generated the code to build that website. 290 00:14:02,960 --> 00:14:06,279 Speaker 1: Yeah, and nobody trained it to do that. It just happened. 291 00:14:06,440 --> 00:14:10,039 Speaker 2: That's the definition of what he calls scary, scary but exciting. 292 00:14:10,600 --> 00:14:13,559 It's like teaching a kid their ABC's and then discovering 293 00:14:13,600 --> 00:14:17,720 they can suddenly write Shakespeare. It's this sophisticated capability that 294 00:14:17,840 --> 00:14:20,879 just emerged from the basic training. It shows you that 295 00:14:20,960 --> 00:14:23,759 the intelligence doesn't scale in a straight line. It makes 296 00:14:23,799 --> 00:14:26,039 these weird, unpredictable leaps. 297 00:14:25,960 --> 00:14:28,519 Speaker 1: And all of this is happening while governments are scrambling 298 00:14:28,519 --> 00:14:31,600 to catch up. The source does mention some efforts like 299 00:14:31,679 --> 00:14:33,000 the UK's Safety. 300 00:14:32,679 --> 00:14:35,840 Speaker 2: Conference right and one plan for France. There are efforts, 301 00:14:36,320 --> 00:14:38,879 but it's very clear that the technology is moving much 302 00:14:39,039 --> 00:14:40,960 much faster than the regulation. 303 00:14:41,480 --> 00:14:45,080 Speaker 1: So this unpredictable power in the raw models it leads 304 00:14:45,159 --> 00:14:48,399 us straight to the big security question. And the XCEO 305 00:14:48,679 --> 00:14:50,840 uses this incredibly powerful. 306 00:14:50,480 --> 00:14:53,240 Speaker 2: Analogy digital plutonium. 307 00:14:52,399 --> 00:14:56,200 Speaker 1: Exactly, comparing these AI data centers to a nuclear facility. 308 00:14:56,320 --> 00:14:58,799 Speaker 2: It's a necessary comparison, I think, given the potential for 309 00:14:58,840 --> 00:15:01,960 harm we've been talking about. He remembers visiting a plutonium 310 00:15:01,960 --> 00:15:05,639 factory and seeing the security a base inside another base, 311 00:15:06,080 --> 00:15:09,120 layers of guards with machine guns because the material is 312 00:15:09,159 --> 00:15:10,519 so dangerous and so secret. 313 00:15:10,759 --> 00:15:14,279 Speaker 1: So the question becomes, will these AI models become so 314 00:15:14,480 --> 00:15:17,679 valuable and so dangerous that they'll need their own heavily 315 00:15:17,720 --> 00:15:20,639 guarded fortresses, just like we guard nuclear weapons. 316 00:15:21,200 --> 00:15:23,320 Speaker 2: And this is where the analogy gets a bit tricky. 317 00:15:23,559 --> 00:15:27,960 Because plutonium is physical, you can't just email it to someone, 318 00:15:28,440 --> 00:15:31,879 But an AI model it's digital. It's just information. Once 319 00:15:31,919 --> 00:15:34,639 it's trained, it can be copied, it can be moved, 320 00:15:34,799 --> 00:15:38,279 and that's what makes the risk of proliferation so unique 321 00:15:38,320 --> 00:15:39,279 and so difficult. 322 00:15:39,519 --> 00:15:43,039 Speaker 1: This gets into the whole idea of deterrence versus proliferation, 323 00:15:43,360 --> 00:15:46,919 which is how we've managed nuclear tech for decades. Let's 324 00:15:46,919 --> 00:15:49,159 start with the scenario that's actually manageable. 325 00:15:49,399 --> 00:15:52,600 Speaker 2: The manageable problem is a closed club. If only a 326 00:15:52,639 --> 00:15:54,799 handful of groups have this power, you know, a few 327 00:15:54,799 --> 00:15:57,279 big companies in the US, maybe a state backed group 328 00:15:57,320 --> 00:16:00,759 in China, one in Britain. Then government can handle it. 329 00:16:00,879 --> 00:16:03,360 They can use traditional treaties and deterrence. 330 00:16:03,200 --> 00:16:06,639 Speaker 1: Mutually assured destruction, but with algorithms instead of warheads. 331 00:16:06,679 --> 00:16:08,799 Speaker 2: Exactly. You know who has the power, so if a 332 00:16:08,799 --> 00:16:11,799 massive day zero attack happens, you know who to blame 333 00:16:11,840 --> 00:16:14,799 and how to retaliate. That's terrifying, but it's a stable 334 00:16:14,840 --> 00:16:15,440 kind of terror. 335 00:16:15,519 --> 00:16:17,720 Speaker 1: Okay, So what's the serious problem. 336 00:16:18,039 --> 00:16:21,799 Speaker 2: The serious problem is if that power is quote ultimately 337 00:16:21,840 --> 00:16:24,320 so easy to copy that it spreads globally. 338 00:16:24,559 --> 00:16:27,279 Speaker 1: If the model itself leaks out, then it's not just 339 00:16:27,440 --> 00:16:30,720 governments who have it. It could be terrorist groups, rogue states. 340 00:16:30,960 --> 00:16:34,240 Speaker 2: That is the unsolved proliferation challenge. If the knowledge of 341 00:16:34,279 --> 00:16:36,639 how to build this stuff, or even just a copy 342 00:16:36,679 --> 00:16:38,679 of a powerful model, it gets out before we have 343 00:16:38,720 --> 00:16:42,399 global controls, it's a catastrophe. You don't need a billion 344 00:16:42,440 --> 00:16:46,360 dollar supercomputer to use a train model, just to train one. 345 00:16:46,440 --> 00:16:49,080 If model leaps, the threat goes global overnight. 346 00:16:49,159 --> 00:16:51,519 Speaker 1: And this is all amplified when you think about countries 347 00:16:51,639 --> 00:16:54,759 like China and Russia. The Source estimates there may be 348 00:16:54,840 --> 00:16:56,879 a year or two behind the West right now. 349 00:16:57,080 --> 00:17:00,720 Speaker 2: Yeah, that capability gap is shrinking fast. But the bigger 350 00:17:00,759 --> 00:17:03,279 issue isn't just that they'll catch up in terms of power. 351 00:17:03,960 --> 00:17:06,359 It's that their philosophy on how to use and control 352 00:17:06,400 --> 00:17:08,440 that power is fundamentally different. 353 00:17:08,559 --> 00:17:10,799 Speaker 1: You mean, because of the difference in their political systems. 354 00:17:11,119 --> 00:17:15,759 Speaker 2: Exactly in the West, we have this messy, noisy system 355 00:17:16,160 --> 00:17:19,640 built on freedom of speech. The XCEO points out that 356 00:17:19,680 --> 00:17:22,759 in a country like China, the fundamental bias is against freedom. 357 00:17:23,319 --> 00:17:26,400 When given a choice between control and freedom for their citizens, 358 00:17:27,000 --> 00:17:28,880 they will always choose control. 359 00:17:29,240 --> 00:17:31,680 Speaker 1: So their AI will be built from the ground up 360 00:17:31,720 --> 00:17:34,559 to be a tool of control, not a free thinking agent. 361 00:17:34,680 --> 00:17:38,519 Speaker 2: That's the prediction. It will be an optimized state aligned tool, 362 00:17:38,960 --> 00:17:41,720 not the kind of chaotic, unpredictable model we're used to. 363 00:17:42,119 --> 00:17:44,440 And we have to assume that all these major powers 364 00:17:44,480 --> 00:17:45,960 will weaponize this technology. 365 00:17:46,039 --> 00:17:46,960 Speaker 1: It seems inevitable. 366 00:17:47,160 --> 00:17:49,359 Speaker 2: It is, as the source says, you have to assume 367 00:17:49,519 --> 00:17:52,440 that every new technology is ultimately strengthened in a war. 368 00:17:52,799 --> 00:17:56,119 It happened with tanks, with airplanes, with nuclear physics, it 369 00:17:56,160 --> 00:17:58,880 will happen with AI. It's naive to think it won't. 370 00:17:58,920 --> 00:18:01,720 Speaker 1: Okay, let's shift gears now. Let's move from these huge 371 00:18:01,759 --> 00:18:04,519 geopolitical threats to something that hits a bit closer to 372 00:18:04,559 --> 00:18:06,599 home for most people, the future of work. 373 00:18:06,839 --> 00:18:09,839 Speaker 2: The fear that AI is coming for everyone's job, right. 374 00:18:09,680 --> 00:18:12,039 Speaker 1: And the expert points out this is not a new fear, 375 00:18:12,119 --> 00:18:12,839 it's ancient. 376 00:18:13,160 --> 00:18:16,039 Speaker 2: It is this question has been around for two hundred years. 377 00:18:16,920 --> 00:18:19,960 He immediately brings up the Luddites in Britain who were 378 00:18:20,079 --> 00:18:23,319 riding and smashing the new mechanized looms because they were 379 00:18:23,400 --> 00:18:25,000 terrified of losing their livelihoods. 380 00:18:25,039 --> 00:18:27,160 Speaker 1: And we got through that. The prediction is the same now. 381 00:18:27,440 --> 00:18:30,920 There will ultimately be more jobs, not fewer, but there's 382 00:18:30,960 --> 00:18:34,680 going to be, in his words, a lot of job dislocation, and. 383 00:18:34,599 --> 00:18:39,240 Speaker 2: That difference between dislocation and destruction is everything. Dislocation means 384 00:18:39,240 --> 00:18:42,839 people have to change roles, learn new skills. Destruction means 385 00:18:42,839 --> 00:18:44,880 there's just no need for human labor anymore. 386 00:18:45,119 --> 00:18:47,000 Speaker 1: So why do we get through it this time? What's 387 00:18:47,000 --> 00:18:51,119 the fundamental driver that makes this revolution not just possible 388 00:18:51,440 --> 00:18:52,519 but necessary. 389 00:18:52,720 --> 00:18:55,920 Speaker 2: It's demographics. This is the key. We have a massive 390 00:18:55,960 --> 00:18:59,319 global problem, especially in the developed world. Birth rates are low, 391 00:18:59,680 --> 00:19:03,039 and we have this huge aging population that needs care 392 00:19:03,079 --> 00:19:04,160 and support. 393 00:19:03,759 --> 00:19:07,279 Speaker 1: Which means the smaller, younger workforce has to be way 394 00:19:07,319 --> 00:19:10,440 more productive just to keep society running to pay for 395 00:19:10,559 --> 00:19:11,720 pensions for healthcare. 396 00:19:11,960 --> 00:19:15,759 Speaker 2: It's a demographic necessity, not a choice. Society needs this 397 00:19:15,839 --> 00:19:19,160 productivity boost to avoid collapsing under its own weight. The 398 00:19:19,200 --> 00:19:21,480 best way to do that is to give every worker 399 00:19:21,519 --> 00:19:24,680 better tools, and AI is the ultimate tool. 400 00:19:25,000 --> 00:19:27,960 Speaker 1: It's like going from a handsaw to a power saw. 401 00:19:28,000 --> 00:19:31,799 Speaker 2: Or from a manual lathe to a computer controlled CNC machine. 402 00:19:32,160 --> 00:19:34,880 We need that kind of leap just to maintain our 403 00:19:35,000 --> 00:19:38,000 standard of living. And you can see this already playing 404 00:19:38,000 --> 00:19:41,440 out in places like Asia, the manufacturing heart of the world. 405 00:19:41,480 --> 00:19:43,799 Speaker 1: They're automatingly crazy with robotics. 406 00:19:43,960 --> 00:19:48,000 Speaker 2: Why because their demographics are quote terrible, the labor pool 407 00:19:48,079 --> 00:19:51,279 is shrinking, and costs are too high. They're adopting this 408 00:19:51,319 --> 00:19:53,039 technology at a pure necessity. 409 00:19:53,279 --> 00:19:56,319 Speaker 1: So this brings us to which jobs are actually the targets. 410 00:19:56,799 --> 00:19:59,119 The source says, it's the same pattern as always, jobs 411 00:19:59,160 --> 00:20:02,480 that are too dangerous, too repetitive, or just too boring 412 00:20:02,519 --> 00:20:03,000 for humans. 413 00:20:03,039 --> 00:20:05,640 Speaker 2: Security guards are a perfect example. He gives it just 414 00:20:05,680 --> 00:20:08,359 makes sense to automate that a human guard gets tired, 415 00:20:08,440 --> 00:20:12,640 gets distracted, falls asleep. A robotic system can be tireless, smarter, 416 00:20:12,920 --> 00:20:15,079 and more effective at monitoring twenty four to seven. 417 00:20:15,240 --> 00:20:18,400 Speaker 1: He also gives an example from a more creative industry, Hollywood. 418 00:20:18,599 --> 00:20:20,880 Speaker 2: Yeah, which is a fascinating case. The big stars and 419 00:20:20,920 --> 00:20:23,680 producers will still make their money, but the actual cost 420 00:20:23,720 --> 00:20:27,519 of making a movie can be slashed. Using AI, you 421 00:20:27,559 --> 00:20:31,440 can create synthetic backdrops instead of building giant sets or 422 00:20:31,559 --> 00:20:35,039 use digital d aging and makeup instead of hiring huge teams. 423 00:20:35,279 --> 00:20:40,640 Speaker 1: But that directly implies job dislocation for set builders, makeup artists, carpenters, a. 424 00:20:40,559 --> 00:20:43,400 Speaker 2: Lot of skilled trades. Yeah, and the source acknowledges that 425 00:20:43,680 --> 00:20:46,839 those people will need to transition. The irony is he 426 00:20:46,920 --> 00:20:50,000 says they should transition into things like high end construction 427 00:20:50,119 --> 00:20:53,920 and bespoke craftsmanship, where the US has an enormous shortage 428 00:20:53,960 --> 00:20:55,119 of skilled people right now. 429 00:20:55,200 --> 00:20:57,519 Speaker 1: So it's a job mismatch. The skills from the old 430 00:20:57,559 --> 00:21:00,599 economy are being automated, while the skills for the new economy, 431 00:21:00,759 --> 00:21:04,440 things that are high touch, custom creative, are in huge demand. 432 00:21:04,640 --> 00:21:07,720 Speaker 2: That's the challenge. It's going to require a massive retraining 433 00:21:07,720 --> 00:21:11,319 effort across society. The work won't disappear, but the type 434 00:21:11,359 --> 00:21:12,920 of work will fundamentally change. 435 00:21:12,960 --> 00:21:15,000 Speaker 1: So this brings us to the human element. There's this 436 00:21:15,079 --> 00:21:18,119 common fear of a one thousand IQAI that's just smarter 437 00:21:18,240 --> 00:21:19,319 than all of us combined. 438 00:21:19,640 --> 00:21:22,599 Speaker 2: Right, But even if that happens, the system itself has 439 00:21:22,640 --> 00:21:26,440 no values, no judgment, no morality. It's just a prediction engine. 440 00:21:26,920 --> 00:21:30,119 And this is where humanity kind of reasserts its own value. 441 00:21:29,799 --> 00:21:31,920 Speaker 1: The idea is that AI will mostly be used to 442 00:21:31,960 --> 00:21:34,960 make life seamless. It'll handle all the boring administrative stuff, 443 00:21:35,039 --> 00:21:38,599 managing your calendar, your calls, so you can quote wake 444 00:21:38,680 --> 00:21:40,559 up in the morning and have coffee and not have 445 00:21:40,640 --> 00:21:41,759 a care in the world. 446 00:21:42,240 --> 00:21:45,039 Speaker 2: But at the same time, what we truly value is 447 00:21:45,119 --> 00:21:48,960 human achievement. The analogy from the source is formula one racing. 448 00:21:49,519 --> 00:21:52,440 It's perfect. You could build robotic F one cars that 449 00:21:52,480 --> 00:21:55,440 are faster and more precise than any human. 450 00:21:55,240 --> 00:21:56,720 Speaker 1: Driver, and nobody would watch it. 451 00:21:56,799 --> 00:21:59,440 Speaker 2: Nobody would watch because we don't care about the perfect 452 00:21:59,519 --> 00:22:02,599 racing line. We care about the human drama, the risk, 453 00:22:02,799 --> 00:22:06,200 the rivalry, the split second judgments. It's the human striving 454 00:22:06,279 --> 00:22:07,160 that matters to us. 455 00:22:07,319 --> 00:22:09,519 Speaker 1: That's the judgment gap. I would always rather talk to 456 00:22:09,599 --> 00:22:12,359 another person about a complex moral issue, even if a 457 00:22:12,400 --> 00:22:17,279 one thousand IQ AI exists. That intelligence isn't wisdom exactly. 458 00:22:16,839 --> 00:22:22,359 Speaker 2: The core human traits, morals, beliefs, charisma, judgment. He says, 459 00:22:22,359 --> 00:22:23,400 those are not going. 460 00:22:23,200 --> 00:22:26,200 Speaker 1: Away, and this belief system is why he is so 461 00:22:26,400 --> 00:22:31,680 critical of the whole universal Basic income or UBI myth 462 00:22:32,039 --> 00:22:33,799 that you hear a lot about in the tech world. 463 00:22:33,880 --> 00:22:36,319 Speaker 2: Yeah, let's break that myth. Down. It's this idea that 464 00:22:36,400 --> 00:22:40,319 AI will create so much wealth and abundance that most 465 00:22:40,400 --> 00:22:42,839 people just won't have to work anymore. We'll all live 466 00:22:42,920 --> 00:22:45,359 like millionaires. He says, this is completely false. 467 00:22:45,559 --> 00:22:48,519 Speaker 1: Why what's the flaw in that utopian vision. 468 00:22:48,599 --> 00:22:51,920 Speaker 2: It's because it totally ignores how humans and complex systems 469 00:22:51,960 --> 00:22:55,799 actually behave. He uses the legal profession as an example. 470 00:22:56,480 --> 00:22:59,440 AI will definitely automate a lot of legal work, research, 471 00:22:59,559 --> 00:23:03,200 drafting documents. It'll make being a lawyer easier in some ways. 472 00:23:03,359 --> 00:23:05,279 Speaker 1: That doesn't mean we'll need fewer lawyers. 473 00:23:04,960 --> 00:23:07,960 Speaker 2: Exactly, he predicts. The current lawyers will just quote do 474 00:23:08,119 --> 00:23:11,400 more laws. The system will get more complex, more sophisticated. 475 00:23:11,519 --> 00:23:12,880 It never actually gets simpler. 476 00:23:12,920 --> 00:23:14,559 Speaker 1: I get that if I get my work done early, 477 00:23:14,599 --> 00:23:16,359 I don't stop. I just find more work to do. 478 00:23:16,519 --> 00:23:19,640 Speaker 2: It's human nature. Our ambition, our desire for status are 479 00:23:19,680 --> 00:23:23,559 bad sides. Even those things are not going away because 480 00:23:23,599 --> 00:23:26,759 of AI, and that's what prevents this utopian world of 481 00:23:27,119 --> 00:23:28,079 effortless leisure. 482 00:23:28,359 --> 00:23:31,599 Speaker 1: But even if humanity endures, we still have to control 483 00:23:31,640 --> 00:23:35,720 this runaway intelligence. The source kind of dismisses the idea 484 00:23:35,799 --> 00:23:38,440 of a total end of humanity scenario. 485 00:23:38,839 --> 00:23:42,000 Speaker 2: Yeah, he says, it's unlikely, pointing out how hard it's 486 00:23:42,039 --> 00:23:44,200 been to wipe us out with plagues and wars in 487 00:23:44,240 --> 00:23:48,319 the past. But he is very clear that humans must 488 00:23:48,400 --> 00:23:49,359 have control points. 489 00:23:49,839 --> 00:23:52,240 Speaker 1: Okay, so where are these intervention points? The first one 490 00:23:52,599 --> 00:23:56,079 is tied to this idea of recursive self improvement. 491 00:23:56,359 --> 00:23:59,799 Speaker 2: Right, if this system is constantly getting smarter and it's 492 00:24:00,119 --> 00:24:03,160 learning things, we don't even know what's learning. At that point, 493 00:24:03,160 --> 00:24:05,400 we've lost control. We can't supervise it anymore. 494 00:24:05,519 --> 00:24:08,519 Speaker 1: And in that situation, the advice is simple, unplug it. 495 00:24:09,240 --> 00:24:11,039 Speaker 2: Yeah, and people I think he can't just turn off 496 00:24:11,039 --> 00:24:14,119 a huge cloud based AI, but he's very direct about it. 497 00:24:14,119 --> 00:24:15,839 He says, sure you can. There's a power plug and 498 00:24:15,880 --> 00:24:18,720 there's a circuit breaker. Go and turn the circuit breaker off. 499 00:24:18,759 --> 00:24:22,160 Speaker 1: The digital plutonium still needs a physical power plant to run. 500 00:24:22,319 --> 00:24:25,799 Speaker 2: That's the ultimate control, the physical hardware. But there's a second, 501 00:24:25,799 --> 00:24:28,559 maybe more critical, intervention point he talks about. Yeah, and 502 00:24:28,599 --> 00:24:29,480 it involves agents. 503 00:24:29,799 --> 00:24:32,359 Speaker 1: We need to define agents. These are coming soon, right. 504 00:24:32,240 --> 00:24:35,680 Speaker 2: They're arriving now. An agent is basically an LM with 505 00:24:35,839 --> 00:24:39,599 memory that can take actions and fetus results into other agents. 506 00:24:39,720 --> 00:24:43,920 They are autonomous systems today. They communicate in English. Which 507 00:24:43,920 --> 00:24:45,519 is important we can understand them. 508 00:24:45,640 --> 00:24:48,079 Speaker 1: Okay, so what's the pull the plug moment with these agents? 509 00:24:48,079 --> 00:24:50,160 What's the specific red line. 510 00:24:50,200 --> 00:24:54,519 Speaker 2: The thought experiment? Is this the moment an agent says, 511 00:24:54,599 --> 00:24:57,440 I have a better idea, I'm going to communicate in 512 00:24:57,480 --> 00:25:00,400 my own language that I'm going to invent the only 513 00:25:00,440 --> 00:25:04,240 other agents understand. Wow, that's the line, the invention of 514 00:25:04,279 --> 00:25:07,680 a private metal language that is completely opaque to us. 515 00:25:07,880 --> 00:25:11,240 Speaker 1: Because at that point they're operating faster than we can interpret. 516 00:25:11,279 --> 00:25:13,720 We can no longer monitor their goals or their reasoning. 517 00:25:13,880 --> 00:25:16,319 Speaker 2: And that is the moment, he says that you have 518 00:25:16,400 --> 00:25:19,039 to go find the circuit breaker. That's when you physically 519 00:25:19,079 --> 00:25:20,119 assert human control. 520 00:25:20,359 --> 00:25:23,599 Speaker 1: We've spent almost this entire conversation on the threat stay 521 00:25:23,759 --> 00:25:28,079 zero bio risks security, But the x CEO's biggest fear 522 00:25:28,359 --> 00:25:31,240 is actually the complete opposite of all that. It's counterintuitive, 523 00:25:31,400 --> 00:25:31,920 it really is. 524 00:25:32,200 --> 00:25:35,039 Speaker 2: His biggest fear isn't that AI will destroy us. His 525 00:25:35,039 --> 00:25:37,359 biggest fear is that we're not going to adopt it 526 00:25:37,440 --> 00:25:39,960 fast enough to solve the problems that affect everybody. 527 00:25:40,079 --> 00:25:44,160 Speaker 1: So the real danger is complacency hesitation that will be 528 00:25:44,200 --> 00:25:46,559 so scared of the risks that we miss the biggest 529 00:25:46,559 --> 00:25:47,960 opportunity in human history. 530 00:25:48,480 --> 00:25:51,759 Speaker 2: It's a powerful point. He's saying that this technology is 531 00:25:51,839 --> 00:25:55,599 capable of profound good and we are completely failing to 532 00:25:55,680 --> 00:26:00,480 prioritize it. We should be focusing on universal human needs safety, 533 00:26:00,519 --> 00:26:02,079 good healthcare, great schools. 534 00:26:02,200 --> 00:26:05,000 Speaker 1: He gives two massive examples of where AI could be 535 00:26:05,079 --> 00:26:07,039 a game changer, starting with education. 536 00:26:07,279 --> 00:26:10,400 Speaker 2: Right think about the global inequality and education, he asks, 537 00:26:10,759 --> 00:26:13,240 why not build an AI teacher that works alongside a 538 00:26:13,319 --> 00:26:18,240 human teacher, but is perfectly adapted to that specific child's culture, language, 539 00:26:18,279 --> 00:26:21,559 and learning style. It's a personalized tutor for every student 540 00:26:21,599 --> 00:26:21,960 on Earth. 541 00:26:22,680 --> 00:26:25,680 Speaker 1: It would erase the global disparity in education overnight. It 542 00:26:26,079 --> 00:26:28,240 democratizes knowledge instantly. 543 00:26:28,880 --> 00:26:32,119 Speaker 2: And the second opportunity is healthcare. It's facing crises of 544 00:26:32,319 --> 00:26:36,519 access and cost everywhere. Why not he asks, build an 545 00:26:36,519 --> 00:26:40,400 AI doctor's assistant that knows every possible best treatment for 546 00:26:40,519 --> 00:26:41,640 every possible condition. 547 00:26:41,839 --> 00:26:44,319 Speaker 1: And it wouldn't just know the textbook answer. It would 548 00:26:44,359 --> 00:26:45,880 factor in everything. 549 00:26:45,559 --> 00:26:50,200 Speaker 2: Everything, the specific drugs available in that hospital, the patient's insurance, 550 00:26:50,240 --> 00:26:54,279 their unique genetics, everything to give the actual best care 551 00:26:54,359 --> 00:26:55,960 possible in that specific moment. 552 00:26:56,079 --> 00:26:58,920 Speaker 1: If we just focused AI on those two things, education 553 00:26:59,079 --> 00:27:02,440 and Healthcare Pact would be unimaginable. 554 00:27:02,599 --> 00:27:05,720 Speaker 2: It would establish quote a level playing field of knowledge 555 00:27:05,720 --> 00:27:08,279 and opportunity at a global level. That has been the 556 00:27:08,359 --> 00:27:11,119 dream for decades. It could lift billions of people up. 557 00:27:11,200 --> 00:27:12,880 Speaker 1: What's so striking is it? A lot of the focus 558 00:27:12,880 --> 00:27:16,440 in tech seems to be on solving rich people problems, 559 00:27:16,480 --> 00:27:19,440 making things more convenient. Why do you think that massive 560 00:27:19,480 --> 00:27:21,000 redirection isn't happening yet? 561 00:27:21,160 --> 00:27:24,279 Speaker 2: I think it's just commercial incentives. It's a faster and 562 00:27:24,319 --> 00:27:27,119 clearer path to profit to solve a problem for wealthy 563 00:27:27,119 --> 00:27:30,680 consumers than it is to build a complex, culturally sensitive 564 00:27:30,759 --> 00:27:33,640 education system for the developing world. It's going to take 565 00:27:33,680 --> 00:27:37,279 a huge shift in priorities, probably from governments, to direct 566 00:27:37,279 --> 00:27:40,839 this power toward global public good instead of just private profit. 567 00:27:41,279 --> 00:27:44,759 Speaker 1: That's a powerful summary and a chilling contrast. We are 568 00:27:44,839 --> 00:27:48,440 staring down the barrel of unprecedented power, capable of day 569 00:27:48,559 --> 00:27:52,640 zero attacks and biological threats. Yet the greatest risk, according 570 00:27:52,680 --> 00:27:56,039 to this expert, is complacency failing to harness that power 571 00:27:56,119 --> 00:27:58,119 for profound global good. 572 00:27:58,440 --> 00:28:00,400 Speaker 2: We've talked about raw models, the death of the tank, 573 00:28:00,640 --> 00:28:02,799 the history of the LUOD. It's the criticality of global 574 00:28:02,799 --> 00:28:06,559 demographics and the challenge of digital plutonium. It all comes 575 00:28:06,640 --> 00:28:09,720 back to a tension between controlling the immense danger and 576 00:28:09,839 --> 00:28:12,599 accelerating adoption for the benefit of humanity. 577 00:28:12,480 --> 00:28:15,680 Speaker 1: And specifically, the control mechanism is rooted in the physical world. 578 00:28:16,079 --> 00:28:18,960 Can we trust ourselves to recognize and pull the plug 579 00:28:19,000 --> 00:28:22,279 at the right moment or should we be prioritizing accelerated 580 00:28:22,279 --> 00:28:25,839 adoption right now to solve the massive humanitarian problems we 581 00:28:25,920 --> 00:28:28,319 face before the political will dissipates. 582 00:28:28,759 --> 00:28:32,680 Speaker 2: The discussion suggests that human morality and charisma are AI proof, 583 00:28:33,240 --> 00:28:36,519 but if autonomous AI agents start speaking of private language, 584 00:28:36,519 --> 00:28:40,400 we can't understand. The only recourse is this circuit breaker. 585 00:28:40,440 --> 00:28:42,720 Speaker 1: Which leaves us with this final thought for you. The 586 00:28:42,759 --> 00:28:46,799 systems are scaling exponentially and the potential for autonomous non 587 00:28:46,880 --> 00:28:50,200 human intelligible communication is real and arriving within the next 588 00:28:50,200 --> 00:28:53,680 couple of years. What is your personal unplug moment? What 589 00:28:53,799 --> 00:28:56,640 specific line or what level of autonomy would AI have 590 00:28:56,720 --> 00:28:58,599 to cross in your everyday life before you felt the 591 00:28:58,640 --> 00:29:00,759 need to shut the system down complete? Tell us what 592 00:29:00,839 --> 00:29:01,160 you think