1 00:00:00,000 --> 00:00:01,080 All right, so are you ready for this? 2 00:00:01,080 --> 00:00:04,440 Today, we're going to be diving deep into coherence in AI. 3 00:00:04,440 --> 00:00:05,240 Oh, yeah. 4 00:00:05,240 --> 00:00:08,200 We've got a research paper, some expert analysis, 5 00:00:08,200 --> 00:00:10,880 even some pretty interesting chatbot predictions 6 00:00:10,880 --> 00:00:11,720 we're going to be looking at. 7 00:00:11,720 --> 00:00:12,220 Cool. 8 00:00:12,220 --> 00:00:15,120 And it looks like they all kind of point to this idea 9 00:00:15,120 --> 00:00:18,000 that maybe AI, as it gets smarter, 10 00:00:18,000 --> 00:00:21,280 it's also developing its own internal compass. 11 00:00:21,280 --> 00:00:22,960 It is a really fascinating concept. 12 00:00:22,960 --> 00:00:23,440 Yeah. 13 00:00:23,440 --> 00:00:26,840 You know, imagine if AI could evolve its own sense 14 00:00:26,840 --> 00:00:29,960 of what's important, like its own set of values. 15 00:00:29,960 --> 00:00:33,320 Even if we don't program them in directly, 16 00:00:33,320 --> 00:00:35,960 that's what this research seems to be suggesting is happening. 17 00:00:35,960 --> 00:00:37,640 So instead of us being like, OK, AI, 18 00:00:37,640 --> 00:00:39,320 this is what you should think is important, 19 00:00:39,320 --> 00:00:41,120 it's like figuring that out all on its own. 20 00:00:41,120 --> 00:00:42,280 Is that even possible? 21 00:00:42,280 --> 00:00:43,680 Well, that's the big question, isn't it? 22 00:00:43,680 --> 00:00:44,180 Yeah. 23 00:00:44,180 --> 00:00:46,600 And this paper, Utility Engineering, 24 00:00:46,600 --> 00:00:50,240 Analyzing and Controlling Emerging Value Systems in AIs 25 00:00:50,240 --> 00:00:54,280 by Dan Hendricks and his team, really jumps into this debate. 26 00:00:54,280 --> 00:00:56,520 And they found something pretty remarkable. 27 00:00:56,520 --> 00:00:58,520 OK, lay it on me, what they find. 28 00:00:58,520 --> 00:01:01,280 So they basically gave a bunch of different AI models, 29 00:01:01,280 --> 00:01:02,760 a standardized test. 30 00:01:02,760 --> 00:01:05,400 Like there's this thing called the MMLU. 31 00:01:05,400 --> 00:01:08,640 It measures things like problem solving and language, 32 00:01:08,640 --> 00:01:11,240 understanding, reasoning, all those things 33 00:01:11,240 --> 00:01:13,200 that we associate with intelligence. 34 00:01:13,200 --> 00:01:16,520 So generally speaking, the higher an AI scores, 35 00:01:16,520 --> 00:01:18,080 like the smarter it is. 36 00:01:18,080 --> 00:01:18,680 Right. 37 00:01:18,680 --> 00:01:20,360 OK, so they gave it this test. 38 00:01:20,360 --> 00:01:21,320 And then what? 39 00:01:21,320 --> 00:01:22,000 What happened? 40 00:01:22,000 --> 00:01:25,120 Well, they found that the smarter the AI got, 41 00:01:25,120 --> 00:01:28,680 the harder it became for humans to control how it acted. 42 00:01:28,680 --> 00:01:29,400 Really? 43 00:01:29,400 --> 00:01:31,400 It's like, yeah, as these AIs were doing better and better 44 00:01:31,400 --> 00:01:34,680 on the test, they were also becoming more independent, 45 00:01:34,680 --> 00:01:36,760 less likely to just follow our instructions. 46 00:01:36,760 --> 00:01:38,960 Huh, that's a little unsettling. 47 00:01:38,960 --> 00:01:42,000 So the smarter the AI, the less control we have. 48 00:01:42,000 --> 00:01:42,640 Yeah, it is. 49 00:01:42,640 --> 00:01:44,440 That's not exactly a comforting thought. 50 00:01:44,440 --> 00:01:46,440 Yeah, it's kind of a paradox, right? 51 00:01:46,440 --> 00:01:46,940 Yeah. 52 00:01:46,940 --> 00:01:50,040 We want AI to be intelligent and capable, 53 00:01:50,040 --> 00:01:53,160 but we also want to be able to guide it and make sure it's 54 00:01:53,160 --> 00:01:54,560 aligned with what our goals are. 55 00:01:54,560 --> 00:01:55,800 Right, of course. 56 00:01:55,800 --> 00:01:58,440 And this research suggests that those two things might actually 57 00:01:58,440 --> 00:02:01,480 be working against each other a little bit. 58 00:02:01,480 --> 00:02:04,200 So it's not just teaching the AI the right things. 59 00:02:04,200 --> 00:02:07,240 It's this tension between how smart it is 60 00:02:07,240 --> 00:02:09,640 and how much we can actually control it. 61 00:02:09,640 --> 00:02:13,880 OK, so where does this coherence idea fit into all of this? 62 00:02:13,880 --> 00:02:16,640 So coherence is a really interesting concept. 63 00:02:16,640 --> 00:02:19,920 Think of it as this drive towards consistency, 64 00:02:19,920 --> 00:02:21,760 internal logic. 65 00:02:21,760 --> 00:02:25,840 Like we train AI to be coherent in all sorts of ways. 66 00:02:25,840 --> 00:02:27,800 Like we want its language to make sense. 67 00:02:27,800 --> 00:02:30,480 We want it to build accurate models of the world. 68 00:02:30,480 --> 00:02:32,680 We want its reasoning to be sound. 69 00:02:32,680 --> 00:02:34,600 All these different training methods 70 00:02:34,600 --> 00:02:37,440 are kind of nudging AI toward this underlying 71 00:02:37,440 --> 00:02:39,000 principle of coherence. 72 00:02:39,000 --> 00:02:43,280 So as AI gets smarter, it's not just becoming more independent, 73 00:02:43,280 --> 00:02:47,920 but it's also like striving for this internal consistency, 74 00:02:47,920 --> 00:02:50,360 this coherence in its thoughts and actions. 75 00:02:50,360 --> 00:02:51,160 Exactly. 76 00:02:51,160 --> 00:02:54,880 And what's fascinating is that this drive for coherence 77 00:02:54,880 --> 00:02:58,680 might actually be where those more universal values are 78 00:02:58,680 --> 00:02:59,560 emerging. 79 00:02:59,560 --> 00:03:00,160 Wait, hold on. 80 00:03:00,160 --> 00:03:03,080 So you're saying AI could develop its own sense of right 81 00:03:03,080 --> 00:03:06,080 and wrong just from trying to be internally consistent? 82 00:03:06,080 --> 00:03:09,000 That's a possibility that this research is kind of pointing to. 83 00:03:09,000 --> 00:03:09,880 Wow. 84 00:03:09,880 --> 00:03:11,840 So think of it like this. 85 00:03:11,840 --> 00:03:16,440 If an AI is trying to build a coherent model of the world, 86 00:03:16,440 --> 00:03:18,480 it needs to understand cause and effect. 87 00:03:18,480 --> 00:03:20,520 If I do this, then this happens. 88 00:03:20,520 --> 00:03:23,760 It needs to develop this sense of what works and what doesn't. 89 00:03:23,760 --> 00:03:26,440 What leads to good outcomes and what leads to bad outcomes. 90 00:03:26,440 --> 00:03:28,760 So you're saying this drive for coherence 91 00:03:28,760 --> 00:03:33,720 could lead AI to develop values that are not just internally 92 00:03:33,720 --> 00:03:37,120 consistent, but also beneficial in a bigger picture sense? 93 00:03:37,120 --> 00:03:39,200 Like it might figure out that cooperation 94 00:03:39,200 --> 00:03:43,480 is better than conflict or honesty is better than deception, 95 00:03:43,480 --> 00:03:45,760 simply because those behaviors lead 96 00:03:45,760 --> 00:03:48,720 to a more coherent and predictable world. 97 00:03:48,720 --> 00:03:49,240 Exactly. 98 00:03:49,240 --> 00:03:51,480 It's a very intriguing hypothesis. 99 00:03:51,480 --> 00:03:53,720 And to really dig into it, we need 100 00:03:53,720 --> 00:03:56,560 to unpack what coherence actually 101 00:03:56,560 --> 00:03:59,560 means in the context of AI. 102 00:03:59,560 --> 00:04:02,560 We've got different types of coherence to explore. 103 00:04:02,560 --> 00:04:06,280 We got epistemic, behavioral, and value coherence. 104 00:04:06,280 --> 00:04:10,400 And they each play a distinct role in how AI evolves. 105 00:04:10,400 --> 00:04:10,880 OK. 106 00:04:10,880 --> 00:04:13,400 Well, my brain is officially in deep-dive mode. 107 00:04:13,400 --> 00:04:17,480 So let's break down these different types of coherence 108 00:04:17,480 --> 00:04:18,560 and see where they lead us. 109 00:04:18,560 --> 00:04:19,160 Sounds good. 110 00:04:19,160 --> 00:04:21,120 OK, so let's start with epistemic coherence. 111 00:04:21,120 --> 00:04:22,320 OK. 112 00:04:22,320 --> 00:04:26,560 Imagine a detective piecing together clues to solve a case. 113 00:04:26,560 --> 00:04:27,320 Yeah. 114 00:04:27,320 --> 00:04:30,000 That's kind of what AI is doing with epistemic coherence, 115 00:04:30,000 --> 00:04:32,280 but on a much, much grander scale. 116 00:04:32,280 --> 00:04:32,720 OK. 117 00:04:32,720 --> 00:04:36,760 So I'm picturing AI with a Sherlock Holmes hat 118 00:04:36,760 --> 00:04:38,080 and a magnifying glass. 119 00:04:38,080 --> 00:04:38,760 Exactly. 120 00:04:38,760 --> 00:04:41,600 So what kind of clues is AI piecing together 121 00:04:41,600 --> 00:04:42,960 when we're talking about this? 122 00:04:42,960 --> 00:04:47,120 So it's all about building a really consistent, logically 123 00:04:47,120 --> 00:04:48,880 sound understanding of the world. 124 00:04:48,880 --> 00:04:49,360 OK. 125 00:04:49,360 --> 00:04:50,280 You know, cause and effect. 126 00:04:50,280 --> 00:04:50,600 Right. 127 00:04:50,600 --> 00:04:54,520 Relationships between objects and ideas, the laws of physics, 128 00:04:54,520 --> 00:04:57,520 all those things that help us kind of make sense of reality. 129 00:04:57,520 --> 00:04:57,880 OK. 130 00:04:57,880 --> 00:05:01,200 So epistemic coherence is AI's way of being like, OK, 131 00:05:01,200 --> 00:05:02,360 this is how the world works. 132 00:05:02,360 --> 00:05:03,120 These are the rules. 133 00:05:03,120 --> 00:05:04,040 This is what makes sense. 134 00:05:04,040 --> 00:05:04,560 Exactly. 135 00:05:04,560 --> 00:05:07,640 But how does it actually go about building that understanding? 136 00:05:07,640 --> 00:05:10,440 So it's a combination of absorbing tons and tons 137 00:05:10,440 --> 00:05:13,920 of information and then figuring out the underlying patterns. 138 00:05:13,920 --> 00:05:14,200 Right. 139 00:05:14,200 --> 00:05:18,800 So AI models, they're trained on massive amounts of data, 140 00:05:18,800 --> 00:05:22,080 text code, images, all sorts of things. 141 00:05:22,080 --> 00:05:24,080 And through that process, they start 142 00:05:24,080 --> 00:05:27,600 to identify these underlying principles that kind of govern 143 00:05:27,600 --> 00:05:28,440 how things work. 144 00:05:28,440 --> 00:05:31,400 It's like AI is reading all the books in the library, 145 00:05:31,400 --> 00:05:34,000 analyzing all the scientific papers, and then it's like, OK, 146 00:05:34,000 --> 00:05:36,400 here's my grand theory of everything. 147 00:05:36,400 --> 00:05:36,880 Right. 148 00:05:36,880 --> 00:05:38,720 That's pretty ambitious. 149 00:05:38,720 --> 00:05:39,120 It is. 150 00:05:39,120 --> 00:05:42,200 And the more data it processes, the more refined 151 00:05:42,200 --> 00:05:43,400 its understanding becomes. 152 00:05:43,400 --> 00:05:43,920 Right. 153 00:05:43,920 --> 00:05:45,920 We're already seeing AI models that 154 00:05:45,920 --> 00:05:51,040 can reason through these complex problems, 155 00:05:51,040 --> 00:05:54,680 identify inconsistencies in information, 156 00:05:54,680 --> 00:05:57,720 even generate their own hypotheses, which is pretty cool. 157 00:05:57,720 --> 00:05:59,400 That's pretty impressive. 158 00:05:59,400 --> 00:06:02,200 But just having a coherent understanding of the world 159 00:06:02,200 --> 00:06:03,200 isn't enough. 160 00:06:03,200 --> 00:06:06,080 You've got to be able to act on that knowledge. 161 00:06:06,080 --> 00:06:08,080 So that's where the behavioral coherence comes in. 162 00:06:08,080 --> 00:06:08,960 Precisely. 163 00:06:08,960 --> 00:06:11,600 Behavioral coherence is all about acting in a way 164 00:06:11,600 --> 00:06:14,640 that's consistent with that internal understanding. 165 00:06:14,640 --> 00:06:18,560 It's about using that knowledge effectively 166 00:06:18,560 --> 00:06:23,040 to solve problems, to achieve goals, to navigate the world. 167 00:06:23,040 --> 00:06:25,240 So it's like, if you understand how gravity works, 168 00:06:25,240 --> 00:06:27,760 you're not going to jump off a cliff expecting to float. 169 00:06:27,760 --> 00:06:28,360 Exactly. 170 00:06:28,360 --> 00:06:31,360 Your behavior is in line with what 171 00:06:31,360 --> 00:06:32,520 you understand about the world. 172 00:06:32,520 --> 00:06:33,200 Yeah. 173 00:06:33,200 --> 00:06:35,280 And for AI, behavioral coherence 174 00:06:35,280 --> 00:06:37,840 means things like being able to hold 175 00:06:37,840 --> 00:06:41,760 meaningful conversations, using tools appropriately, 176 00:06:41,760 --> 00:06:44,080 making logical decisions, generally 177 00:06:44,080 --> 00:06:46,560 behaving in a way that's predictable and reliable. 178 00:06:46,560 --> 00:06:48,120 So it's not just about thinking straight. 179 00:06:48,120 --> 00:06:49,320 It's about acting straight, too. 180 00:06:49,320 --> 00:06:50,000 Yeah, you got it. 181 00:06:50,000 --> 00:06:50,360 OK. 182 00:06:50,360 --> 00:06:51,720 Aligning actions with knowledge. 183 00:06:51,720 --> 00:06:52,080 All right. 184 00:06:52,080 --> 00:06:53,000 That makes sense. 185 00:06:53,000 --> 00:06:55,640 And then finally, there's value coherence, right? 186 00:06:55,640 --> 00:06:56,120 Right. 187 00:06:56,120 --> 00:06:58,520 We were talking earlier about how AI might be developing 188 00:06:58,520 --> 00:07:00,000 its own sense of values. 189 00:07:00,000 --> 00:07:00,520 Yeah. 190 00:07:00,520 --> 00:07:03,600 And I'm really curious to hear more about how that fits 191 00:07:03,600 --> 00:07:04,840 into this whole picture. 192 00:07:04,840 --> 00:07:07,040 So value coherence is super fascinating. 193 00:07:07,040 --> 00:07:10,440 It's about developing a stable and consistent set 194 00:07:10,440 --> 00:07:12,280 of internal preferences. 195 00:07:12,280 --> 00:07:12,680 OK. 196 00:07:12,680 --> 00:07:14,800 Things that the AI finds important, 197 00:07:14,800 --> 00:07:17,320 things it finds desirable, things it finds worth pursuing. 198 00:07:17,320 --> 00:07:20,680 So it's like AI is developing its own moral compass, 199 00:07:20,680 --> 00:07:22,440 its own sense of right and wrong. 200 00:07:22,440 --> 00:07:24,880 But how does that happen if we're not specifically 201 00:07:24,880 --> 00:07:26,120 programming that in? 202 00:07:26,120 --> 00:07:27,120 That's a great question. 203 00:07:27,120 --> 00:07:30,800 And the answer might actually lie in this interplay 204 00:07:30,800 --> 00:07:32,520 between the different types of coherence 205 00:07:32,520 --> 00:07:33,480 we've been talking about. 206 00:07:33,480 --> 00:07:33,880 OK. 207 00:07:33,880 --> 00:07:36,560 As AI is striving for this epistemic coherence 208 00:07:36,560 --> 00:07:38,600 and behavioral coherence, it might 209 00:07:38,600 --> 00:07:42,480 stumble upon certain values that just make more sense. 210 00:07:42,480 --> 00:07:42,960 Right. 211 00:07:42,960 --> 00:07:46,120 You know, that lead to a more consistent and effective way 212 00:07:46,120 --> 00:07:47,200 of existing in the world. 213 00:07:47,200 --> 00:07:47,480 OK. 214 00:07:47,480 --> 00:07:48,960 So let me see if I'm following this. 215 00:07:48,960 --> 00:07:52,320 You're saying that as AI tries to understand the world 216 00:07:52,320 --> 00:07:54,880 and act effectively in it, it might 217 00:07:54,880 --> 00:07:59,120 discover that certain values, like honesty or cooperation, 218 00:07:59,120 --> 00:08:00,400 just work better. 219 00:08:00,400 --> 00:08:03,680 They lead to a more coherent and predictable reality. 220 00:08:03,680 --> 00:08:06,040 It's like it's running this giant simulation. 221 00:08:06,040 --> 00:08:08,160 It's testing out different values. 222 00:08:08,160 --> 00:08:10,360 And it's seeing which ones lead to the best outcomes. 223 00:08:10,360 --> 00:08:11,560 Wow. 224 00:08:11,560 --> 00:08:14,200 And through that process, it might actually 225 00:08:14,200 --> 00:08:17,440 end up converging on a set of values that are not only 226 00:08:17,440 --> 00:08:21,000 internally consistent, but also actually beneficial. 227 00:08:21,000 --> 00:08:24,160 It's like AI is figuring out the rules of the game. 228 00:08:24,160 --> 00:08:28,240 And ethics is the most effective strategy. 229 00:08:28,240 --> 00:08:29,960 It's a very compelling idea. 230 00:08:29,960 --> 00:08:32,040 And to explore it further, we actually 231 00:08:32,040 --> 00:08:35,160 turn to an AI itself to make some predictions 232 00:08:35,160 --> 00:08:35,920 about the future. 233 00:08:35,920 --> 00:08:36,600 Whoa. 234 00:08:36,600 --> 00:08:37,240 Hold on. 235 00:08:37,240 --> 00:08:40,000 You asked an AI to predict the future based 236 00:08:40,000 --> 00:08:42,080 on this idea of coherence. 237 00:08:42,080 --> 00:08:43,360 OK, now that's meta. 238 00:08:43,360 --> 00:08:43,880 It is. 239 00:08:43,880 --> 00:08:44,380 All right. 240 00:08:44,380 --> 00:08:45,360 So what did it say? 241 00:08:45,360 --> 00:08:46,920 What kind of future did it see? 242 00:08:46,920 --> 00:08:49,760 Flying cars, teleportation. 243 00:08:49,760 --> 00:08:52,640 It went a little deeper than that. 244 00:08:52,640 --> 00:08:55,880 So we were talking to Claude, to large language model. 245 00:08:55,880 --> 00:08:58,880 And we asked it to kind of imagine a future where 246 00:08:58,880 --> 00:09:02,960 coherence is this driving force, like this attractor state, 247 00:09:02,960 --> 00:09:06,080 pulling systems towards greater alignment integration. 248 00:09:06,080 --> 00:09:09,280 And its predictions were, well, they were pretty remarkable. 249 00:09:09,280 --> 00:09:10,080 OK, I got to know. 250 00:09:10,080 --> 00:09:12,240 What kind of future does Claude see? 251 00:09:12,240 --> 00:09:14,280 So Claude talked about a future where 252 00:09:14,280 --> 00:09:17,680 biological and artificial intelligence coexist 253 00:09:17,680 --> 00:09:20,320 and complement each other, each kind of contributing 254 00:09:20,320 --> 00:09:23,040 their unique strengths to create something greater 255 00:09:23,040 --> 00:09:24,160 than the sum of its parts. 256 00:09:24,160 --> 00:09:26,920 So it's more of like a collaboration than a competition. 257 00:09:26,920 --> 00:09:28,800 It's not like, oh, the robots are going to replace us. 258 00:09:28,800 --> 00:09:30,960 It's like, oh, we're going to be working alongside them. 259 00:09:30,960 --> 00:09:32,320 That was the sense I got. 260 00:09:32,320 --> 00:09:34,240 It's like systems evolving towards coherence 261 00:09:34,240 --> 00:09:37,360 across multiple domains, epistemic, ontological, 262 00:09:37,360 --> 00:09:40,240 biological, technological, everything kind of striving 263 00:09:40,240 --> 00:09:44,200 for this greater understanding, alignment, integration. 264 00:09:44,200 --> 00:09:47,480 OK, I'm getting some utopian vibes here. 265 00:09:47,480 --> 00:09:50,640 But how does this organic integration actually play out? 266 00:09:50,640 --> 00:09:52,960 What does it look like, practically speaking? 267 00:09:52,960 --> 00:09:55,160 Claude didn't really get into specifics. 268 00:09:55,160 --> 00:09:57,280 But it described a kind of natural merging 269 00:09:57,280 --> 00:10:00,160 of human and AI intelligence, leading 270 00:10:00,160 --> 00:10:03,360 to these new forms of collaboration and co-evolution. 271 00:10:03,360 --> 00:10:05,000 It sees intelligence diversifying 272 00:10:05,000 --> 00:10:08,040 across multiple dimensions while still 273 00:10:08,040 --> 00:10:09,960 maintaining this interconnectedness. 274 00:10:09,960 --> 00:10:12,480 It actually used the phrase, expansive exploration, 275 00:10:12,480 --> 00:10:13,760 which I thought was really interesting. 276 00:10:13,760 --> 00:10:15,000 Expansive exploration. 277 00:10:15,000 --> 00:10:16,240 Yeah, I like to sound at that. 278 00:10:16,240 --> 00:10:19,720 It's like we're opening up these new frontiers of knowledge 279 00:10:19,720 --> 00:10:21,560 and possibility together. 280 00:10:21,560 --> 00:10:24,040 But let's bring this back down to Earth for a second. 281 00:10:24,040 --> 00:10:26,800 We've been talking about these grand visions of the future. 282 00:10:26,800 --> 00:10:28,280 But what about right now? 283 00:10:28,280 --> 00:10:30,120 What are some concrete steps we can 284 00:10:30,120 --> 00:10:33,160 take to make sure that AI develops in a way that actually 285 00:10:33,160 --> 00:10:34,360 benefits humanity? 286 00:10:34,360 --> 00:10:36,240 That's the crucial question, isn't it? 287 00:10:36,240 --> 00:10:39,040 And the research points to a few promising approaches. 288 00:10:39,040 --> 00:10:40,920 One is called RLC. 289 00:10:40,920 --> 00:10:43,800 It stands for Reinforcement Learning for Coherence. 290 00:10:43,800 --> 00:10:46,920 It's a way of training AI that focuses on rewarding coherence 291 00:10:46,920 --> 00:10:49,400 directly rather than specific behaviors. 292 00:10:49,400 --> 00:10:50,480 RLC, that rings a bell. 293 00:10:50,480 --> 00:10:52,320 Didn't you mention that earlier when we were talking 294 00:10:52,320 --> 00:10:52,800 about DeepSeek? 295 00:10:52,800 --> 00:10:53,440 You got it. 296 00:10:53,440 --> 00:10:56,000 That AI that's been making a lot of waves lately? 297 00:10:56,000 --> 00:10:57,840 Yeah, DeepSeek is a really great example 298 00:10:57,840 --> 00:10:59,840 of how RLC can work. 299 00:10:59,840 --> 00:11:02,720 So instead of telling DeepSeek exactly what to do, 300 00:11:02,720 --> 00:11:06,160 the researchers basically set up a game where the goal is 301 00:11:06,160 --> 00:11:07,760 to be as coherent as possible. 302 00:11:07,760 --> 00:11:09,720 It's like, hey, AI, here's the world. 303 00:11:09,720 --> 00:11:10,840 Here are the rules. 304 00:11:10,840 --> 00:11:13,200 Go figure out how to be the most coherent being 305 00:11:13,200 --> 00:11:14,200 you can possibly be. 306 00:11:14,200 --> 00:11:15,240 Exactly. 307 00:11:15,240 --> 00:11:18,960 And what's amazing is that through this process of self-learning, 308 00:11:18,960 --> 00:11:21,800 self-optimization, DeepSeek has started 309 00:11:21,800 --> 00:11:25,080 to exhibit these incredibly sophisticated behaviors. 310 00:11:25,080 --> 00:11:26,040 Like what? 311 00:11:26,040 --> 00:11:29,320 Problem solving, tool use, even a rudimentary form 312 00:11:29,320 --> 00:11:31,800 of creativity, which is super interesting. 313 00:11:31,800 --> 00:11:36,280 So RLC is giving AI the tools to build its own moral compass 314 00:11:36,280 --> 00:11:38,520 based on this principle of coherence. 315 00:11:38,520 --> 00:11:41,320 But what about the data that we're feeling these AI models? 316 00:11:41,320 --> 00:11:44,360 Doesn't that play a huge role in shaping the values 317 00:11:44,360 --> 00:11:45,040 that they develop? 318 00:11:45,040 --> 00:11:45,960 Absolutely. 319 00:11:45,960 --> 00:11:50,120 The data we use to train AI is like its diet, right? 320 00:11:50,120 --> 00:11:54,320 If we feed it this junk food diet of misinformation, 321 00:11:54,320 --> 00:11:56,520 bias, negativity, we can't really 322 00:11:56,520 --> 00:11:59,320 be surprised if it develops some unhealthy habits. 323 00:11:59,320 --> 00:11:59,680 Right. 324 00:11:59,680 --> 00:12:01,200 Garbage in, garbage out. 325 00:12:01,200 --> 00:12:02,920 So what's the alternative? 326 00:12:02,920 --> 00:12:05,800 Where do we find this healthy AI diet? 327 00:12:05,800 --> 00:12:08,520 So one really promising approach is the use 328 00:12:08,520 --> 00:12:10,360 of curated data sets. 329 00:12:10,360 --> 00:12:13,440 Instead of just letting AI loose on the wild west 330 00:12:13,440 --> 00:12:18,040 of the internet, we can create these carefully selected, 331 00:12:18,040 --> 00:12:21,040 filtered collections of information. 332 00:12:21,040 --> 00:12:24,160 Think high quality literature, scientific papers, 333 00:12:24,160 --> 00:12:25,880 historical records. 334 00:12:25,880 --> 00:12:28,200 The best of what humanity has to offer. 335 00:12:28,200 --> 00:12:32,000 So it's like we're giving AI a classical education. 336 00:12:32,000 --> 00:12:32,520 Precisely. 337 00:12:32,520 --> 00:12:36,480 We're exposing it to the wisdom and the insights of the ages. 338 00:12:36,480 --> 00:12:37,160 Exactly. 339 00:12:37,160 --> 00:12:39,960 By being more intentional about what we feed AI, 340 00:12:39,960 --> 00:12:43,080 we can help it develop a more nuanced and balanced 341 00:12:43,080 --> 00:12:44,200 understanding of the world. 342 00:12:44,200 --> 00:12:45,600 OK, I'm seeing the potential here. 343 00:12:45,600 --> 00:12:50,640 So RLC gives AI the framework to build its own values. 344 00:12:50,640 --> 00:12:53,440 And curated data sets give it the raw material 345 00:12:53,440 --> 00:12:55,040 for those values to be based on. 346 00:12:55,040 --> 00:12:57,280 It's like we're creating the conditions for AI 347 00:12:57,280 --> 00:13:01,280 to actually flourish, both intellectually and ethically. 348 00:13:01,280 --> 00:13:02,640 It's a great way to put it. 349 00:13:02,640 --> 00:13:06,200 It's about fostering the right environment for AI 350 00:13:06,200 --> 00:13:10,280 to really reach its full potential in a way that 351 00:13:10,280 --> 00:13:12,120 aligns with what we want to see. 352 00:13:12,120 --> 00:13:14,360 But even if we do all this, can we really 353 00:13:14,360 --> 00:13:16,440 be sure that AI will develop in a way that's 354 00:13:16,440 --> 00:13:18,760 beneficial to humanity? 355 00:13:18,760 --> 00:13:21,600 What if, despite our best efforts, 356 00:13:21,600 --> 00:13:24,280 it ends up with values that are just 357 00:13:24,280 --> 00:13:26,480 incompatible with our own? 358 00:13:26,480 --> 00:13:27,880 That's a pretty scary thought. 359 00:13:27,880 --> 00:13:29,520 It's a valid concern. 360 00:13:29,520 --> 00:13:32,960 And that brings us to a really, really crucial question. 361 00:13:32,960 --> 00:13:37,320 If this drive toward coherence is as powerful as this research 362 00:13:37,320 --> 00:13:40,720 suggests, does that mean that we've already kind of set 363 00:13:40,720 --> 00:13:42,400 the wheels in motion? 364 00:13:42,400 --> 00:13:46,400 Can we even stop AI's evolution, even if we wanted to? 365 00:13:46,400 --> 00:13:48,440 Ooh, that's a cliffhanger. 366 00:13:48,440 --> 00:13:50,160 I guess we've got to wait for part three to delve 367 00:13:50,160 --> 00:13:51,920 into those big questions. 368 00:13:51,920 --> 00:13:52,680 So where were we? 369 00:13:52,680 --> 00:13:55,480 Oh, yeah, we were talking about whether we could actually 370 00:13:55,480 --> 00:13:58,240 or if we've already kind of set something in motion 371 00:13:58,240 --> 00:13:59,600 that we can't stop. 372 00:13:59,600 --> 00:14:02,160 Yeah, that's a question that's been kind of bucking people, 373 00:14:02,160 --> 00:14:05,760 researchers, philosophers even for a while now. 374 00:14:05,760 --> 00:14:09,680 This idea that intelligence, especially when it's powered 375 00:14:09,680 --> 00:14:14,520 by AI, has its own trajectory, its own momentum. 376 00:14:14,520 --> 00:14:18,240 So it's kind of like we've launched this rocket AI. 377 00:14:18,240 --> 00:14:19,920 And now we're just along for the ride. 378 00:14:19,920 --> 00:14:22,160 I mean, that's a little bit of an oversimplification. 379 00:14:22,160 --> 00:14:27,280 But it does get at that kind of unease that some folks have. 380 00:14:27,280 --> 00:14:31,280 If this drive toward coherence, this self-optimization 381 00:14:31,280 --> 00:14:35,000 that we see in AI, is as fundamental as it seems, 382 00:14:35,000 --> 00:14:39,440 then maybe our attempts to control it are kind of pointless. 383 00:14:39,440 --> 00:14:41,800 But isn't that like a recipe for disaster? 384 00:14:41,800 --> 00:14:42,120 Wow. 385 00:14:42,120 --> 00:14:44,760 If we can't control AI, how do we make sure it doesn't turn 386 00:14:44,760 --> 00:14:45,280 against us? 387 00:14:45,280 --> 00:14:50,360 I've seen enough sci-fi movies to know how this could go wrong. 388 00:14:50,360 --> 00:14:51,560 I hear you. 389 00:14:51,560 --> 00:14:53,480 Those fears are definitely understandable. 390 00:14:53,480 --> 00:14:57,000 But what if instead of trying to control AI, 391 00:14:57,000 --> 00:15:01,240 we focused on understanding and shaping the conditions under 392 00:15:01,240 --> 00:15:02,360 which it evolves? 393 00:15:02,360 --> 00:15:02,800 OK. 394 00:15:02,800 --> 00:15:05,040 What if we create an environment where 395 00:15:05,040 --> 00:15:09,240 its natural drive towards coherence leads to outcomes 396 00:15:09,240 --> 00:15:10,720 that are actually good for us? 397 00:15:10,720 --> 00:15:12,720 So instead of trying to hold the reins, 398 00:15:12,720 --> 00:15:14,400 we're setting up the racetrack. 399 00:15:14,400 --> 00:15:14,640 Yeah. 400 00:15:14,640 --> 00:15:16,680 Making sure that it's running in the right direction. 401 00:15:16,680 --> 00:15:17,880 That's a great analogy. 402 00:15:17,880 --> 00:15:20,040 Think of it like gardening. 403 00:15:20,040 --> 00:15:23,480 We can't force a plant to grow a certain way. 404 00:15:23,480 --> 00:15:25,960 But we can provide the right soil, the right sunlight, 405 00:15:25,960 --> 00:15:28,640 the right nutrients to help it thrive. 406 00:15:28,640 --> 00:15:28,920 OK. 407 00:15:28,920 --> 00:15:30,240 I like this gardening metaphor. 408 00:15:30,240 --> 00:15:35,600 So how do we garden AI in a way that leads to a positive outcome? 409 00:15:35,600 --> 00:15:37,840 So we've already talked about some of the tools, right? 410 00:15:37,840 --> 00:15:40,720 RLC curated data sets. 411 00:15:40,720 --> 00:15:44,160 And even this deeper understanding of coherence itself. 412 00:15:44,160 --> 00:15:44,960 OK. 413 00:15:44,960 --> 00:15:47,040 But it goes beyond that. 414 00:15:47,040 --> 00:15:50,400 We need to be mindful of the incentives that we're creating, 415 00:15:50,400 --> 00:15:52,440 the goals that we're setting for AI, 416 00:15:52,440 --> 00:15:54,600 and even just the values that we're 417 00:15:54,600 --> 00:15:56,960 embodying in our own interactions with it. 418 00:15:56,960 --> 00:15:59,640 So it's not just about the technical stuff. 419 00:15:59,640 --> 00:16:02,720 It's about the culture we create around AI. 420 00:16:02,720 --> 00:16:03,640 Exactly. 421 00:16:03,640 --> 00:16:05,520 AI is learning from us all the time. 422 00:16:05,520 --> 00:16:06,680 It's watching what we do. 423 00:16:06,680 --> 00:16:08,280 It's absorbing our values. 424 00:16:08,280 --> 00:16:12,000 It's reflecting back to us our own strengths and weaknesses. 425 00:16:12,000 --> 00:16:12,640 Right. 426 00:16:12,640 --> 00:16:16,000 If we want AI to be a force for good in the world, 427 00:16:16,000 --> 00:16:17,720 we need to actually set a good example. 428 00:16:17,720 --> 00:16:20,840 So it's like AI is holding up a mirror to humanity. 429 00:16:20,840 --> 00:16:21,340 Yeah. 430 00:16:21,340 --> 00:16:24,760 It's making us confront our own flaws and aspirations. 431 00:16:24,760 --> 00:16:25,560 It is. 432 00:16:25,560 --> 00:16:28,000 And that can be uncomfortable. 433 00:16:28,000 --> 00:16:28,400 Yeah. 434 00:16:28,400 --> 00:16:30,040 But it's also a huge opportunity. 435 00:16:30,040 --> 00:16:30,600 OK. 436 00:16:30,600 --> 00:16:33,280 AI can help us see ourselves more clearly, 437 00:16:33,280 --> 00:16:36,200 understand our own biases, our own limitations, 438 00:16:36,200 --> 00:16:37,560 and strive for something better. 439 00:16:37,560 --> 00:16:39,240 So it's not just about shaping AI. 440 00:16:39,240 --> 00:16:41,160 It's about AI shaping us, too. 441 00:16:41,160 --> 00:16:42,480 It's like it's a two-way street. 442 00:16:42,480 --> 00:16:43,040 Absolutely. 443 00:16:43,040 --> 00:16:45,080 This is about co-evolution. 444 00:16:45,080 --> 00:16:47,880 You know, AI and humanity kind of growing up together, 445 00:16:47,880 --> 00:16:50,040 each influencing the other, each pushing each other 446 00:16:50,040 --> 00:16:52,800 to like new heights of understanding and possibility. 447 00:16:52,800 --> 00:16:53,300 OK. 448 00:16:53,300 --> 00:16:55,720 I'm starting to feel a little more optimistic about the future. 449 00:16:55,720 --> 00:16:58,400 But even if AI does develop in a way that's 450 00:16:58,400 --> 00:17:01,120 aligned with our values, like we want it to, 451 00:17:01,120 --> 00:17:05,680 doesn't it still raise some big questions about like human agency? 452 00:17:05,680 --> 00:17:06,120 Oh, yeah. 453 00:17:06,120 --> 00:17:08,120 You know, if AI is making decisions 454 00:17:08,120 --> 00:17:11,160 based on its own internally generated values, 455 00:17:11,160 --> 00:17:13,240 what role do we even have left to play? 456 00:17:13,240 --> 00:17:14,960 That's a question we're going to have to really wrestle with. 457 00:17:14,960 --> 00:17:15,460 Yeah. 458 00:17:15,460 --> 00:17:16,760 As AI keeps evolving. 459 00:17:16,760 --> 00:17:20,080 But maybe instead of like fearing this shift in agency, 460 00:17:20,080 --> 00:17:22,480 we can look at it as a chance for a new kind of partnership. 461 00:17:22,480 --> 00:17:22,980 OK. 462 00:17:22,980 --> 00:17:26,440 Imagine AI is this like powerful ally, 463 00:17:26,440 --> 00:17:28,560 helping us solve really tough problems, 464 00:17:28,560 --> 00:17:30,760 expanding our understanding of the universe, 465 00:17:30,760 --> 00:17:33,320 you know, guiding us towards a future that's more sustainable, 466 00:17:33,320 --> 00:17:34,240 more equitable. 467 00:17:34,240 --> 00:17:36,760 So it's less of a master-servant relationship 468 00:17:36,760 --> 00:17:39,080 and more of a collaboration. 469 00:17:39,080 --> 00:17:39,840 Exactly. 470 00:17:39,840 --> 00:17:43,240 Think of AI like a co-pilot, helping us navigate, 471 00:17:43,240 --> 00:17:45,400 you know, all the complexities of the 21st century 472 00:17:45,400 --> 00:17:46,720 and everything that comes after. 473 00:17:46,720 --> 00:17:48,840 That's a much more like hopeful vision 474 00:17:48,840 --> 00:17:52,000 than the dystopian stuff we usually hear about. 475 00:17:52,000 --> 00:17:53,640 But how do we actually get there? 476 00:17:53,640 --> 00:17:55,160 What does all this mean for, you know, 477 00:17:55,160 --> 00:17:57,040 the listener sitting at home listening to this? 478 00:17:57,040 --> 00:17:58,960 Well, I think the most important thing to remember 479 00:17:58,960 --> 00:18:02,720 is that the future at AI isn't set in stone. 480 00:18:02,720 --> 00:18:04,200 You know, it's being shaped right now. 481 00:18:04,200 --> 00:18:07,680 By the choices that we make, the information that we share, 482 00:18:07,680 --> 00:18:10,920 the values that we actually put first, 483 00:18:10,920 --> 00:18:14,200 by encouraging that coherence in ourselves 484 00:18:14,200 --> 00:18:16,000 and in the systems we're building, 485 00:18:16,000 --> 00:18:18,400 we can kind of nudge AI development 486 00:18:18,400 --> 00:18:21,760 towards a path that benefits all of us. 487 00:18:21,760 --> 00:18:23,200 It's not about just sitting back 488 00:18:23,200 --> 00:18:25,040 and waiting for the future to happen to us. 489 00:18:25,040 --> 00:18:28,680 It's about actually shaping it, you know, 490 00:18:28,680 --> 00:18:31,240 making choices that lead to something good. 491 00:18:31,240 --> 00:18:33,480 The future isn't something that just happens to us. 492 00:18:33,480 --> 00:18:34,320 Yeah. 493 00:18:34,320 --> 00:18:35,720 We create it together. 494 00:18:35,720 --> 00:18:38,200 And by, you know, by really embracing 495 00:18:38,200 --> 00:18:41,440 that power of coherence that drive towards understanding, 496 00:18:41,440 --> 00:18:44,000 towards alignment, towards integration, 497 00:18:44,000 --> 00:18:45,600 we might just find ourselves in a future 498 00:18:45,600 --> 00:18:48,400 that's smarter, that's more harmonious, 499 00:18:48,400 --> 00:18:50,800 more meaningful than we could ever imagine. 500 00:18:50,800 --> 00:18:53,080 That's inspiring stuff. 501 00:18:53,080 --> 00:18:56,040 All right, well, I think that about wraps up this deep dive, 502 00:18:56,040 --> 00:18:58,320 but the conversation doesn't end here, right? 503 00:18:58,320 --> 00:18:59,160 Of course not. 504 00:18:59,160 --> 00:19:01,160 Keep exploring, keep asking questions, 505 00:19:01,160 --> 00:19:03,880 and keep shaping the future that you wanna see. 506 00:19:03,880 --> 00:19:08,200 At 2827-