1 00:00:00,000 --> 00:00:05,600 All right, strap in everyone, because today we're taking a deep dive into some really 2 00:00:05,600 --> 00:00:07,760 cool new AI research. 3 00:00:07,760 --> 00:00:08,760 Oh yeah. 4 00:00:08,760 --> 00:00:10,800 Yeah, this one is from OpenAI. 5 00:00:10,800 --> 00:00:15,240 And it's, well, it's making some serious waves in the world of AI safety. 6 00:00:15,240 --> 00:00:16,240 I'm intrigued. 7 00:00:16,240 --> 00:00:17,240 What is it? 8 00:00:17,240 --> 00:00:19,720 Well, they're calling it Deliberative Alignment. 9 00:00:19,720 --> 00:00:26,480 And it's basically like, it's all about making AI models think before they act. 10 00:00:26,480 --> 00:00:29,880 You know, like really think through those safety guidelines before they go and answer 11 00:00:29,880 --> 00:00:30,880 your requests. 12 00:00:30,880 --> 00:00:32,600 Like giving them a moment to pause you mean. 13 00:00:32,600 --> 00:00:33,600 Exactly. 14 00:00:33,600 --> 00:00:38,560 Like giving them time to think about, you know, the consequences, what they're about to say. 15 00:00:38,560 --> 00:00:39,560 I see. 16 00:00:39,560 --> 00:00:41,520 So kind of like adding in a layer of ethical reasoning. 17 00:00:41,520 --> 00:00:42,520 Right. 18 00:00:42,520 --> 00:00:43,520 Exactly. 19 00:00:43,520 --> 00:00:46,400 Because even with all the safety measures they have in place, you know, like fine tuning 20 00:00:46,400 --> 00:00:47,400 with human feedback. 21 00:00:47,400 --> 00:00:48,400 Right. 22 00:00:48,400 --> 00:00:49,720 AI can still trip up sometimes. 23 00:00:49,720 --> 00:00:50,720 It can, it can. 24 00:00:50,720 --> 00:00:55,280 Like they might accidentally generate harmful content or, you know, or refuse a perfectly 25 00:00:55,280 --> 00:00:56,280 fine request. 26 00:00:56,280 --> 00:00:57,280 Yeah. 27 00:00:57,280 --> 00:00:59,560 And the worst is when they get tricked into misbehaving. 28 00:00:59,560 --> 00:01:00,560 Ugh. 29 00:01:00,560 --> 00:01:01,560 Those shell breaks, right? 30 00:01:01,560 --> 00:01:02,560 Yeah. 31 00:01:02,560 --> 00:01:03,560 Those are no good. 32 00:01:03,560 --> 00:01:07,760 And this paper, this research, it points out two big reasons why this happens. 33 00:01:07,760 --> 00:01:08,760 Okay. 34 00:01:08,760 --> 00:01:09,760 What are they? 35 00:01:09,760 --> 00:01:13,160 First, well, current models often have to respond like instantly, right? 36 00:01:13,160 --> 00:01:14,160 Maybe do. 37 00:01:14,160 --> 00:01:15,160 Yeah. 38 00:01:15,160 --> 00:01:18,160 Even in those really complex situations where, where more thought is needed. 39 00:01:18,160 --> 00:01:19,160 Right. 40 00:01:19,160 --> 00:01:20,160 Right. 41 00:01:20,160 --> 00:01:23,360 And second, well, they learn about safety in this roundabout way from, you know, like 42 00:01:23,360 --> 00:01:25,240 tons and tons of labeled example. 43 00:01:25,240 --> 00:01:26,240 Indirectly. 44 00:01:26,240 --> 00:01:28,520 And instead of directly learning the actual safety rules. 45 00:01:28,520 --> 00:01:29,520 That makes sense. 46 00:01:29,520 --> 00:01:30,760 So it's a quantit over quality thing. 47 00:01:30,760 --> 00:01:31,760 Yeah. 48 00:01:31,760 --> 00:01:32,760 Kinda. 49 00:01:32,760 --> 00:01:33,760 Okay. 50 00:01:33,760 --> 00:01:35,120 So how does deliberative alignment change that? 51 00:01:35,120 --> 00:01:40,160 Well, it's all about, you know, teaching the AI to actually reason through those safety 52 00:01:40,160 --> 00:01:41,160 rules. 53 00:01:41,160 --> 00:01:42,160 Interesting. 54 00:01:42,160 --> 00:01:44,520 Like almost as if they have, you know, a mental checklist. 55 00:01:44,520 --> 00:01:45,520 Okay. 56 00:01:45,520 --> 00:01:46,520 Yeah. 57 00:01:46,520 --> 00:01:47,520 Before generating any response. 58 00:01:47,520 --> 00:01:51,960 So I'm picturing this, like if I try to be sneaky and ask the AI something that's, 59 00:01:51,960 --> 00:01:52,960 you know, not allowed. 60 00:01:52,960 --> 00:01:53,960 Right. 61 00:01:53,960 --> 00:01:56,040 How would deliberative alignment stop me? 62 00:01:56,040 --> 00:02:01,960 So imagine, let's say you try to be super sneaky and hide your request in code or something. 63 00:02:01,960 --> 00:02:07,720 The AI, you know, trained with this deliberative alignment, it decode the message and then 64 00:02:07,720 --> 00:02:09,040 it would realize, wait a second. 65 00:02:09,040 --> 00:02:10,040 This is a trick. 66 00:02:10,040 --> 00:02:11,040 This is a trick. 67 00:02:11,040 --> 00:02:12,040 I better check the safety policy. 68 00:02:12,040 --> 00:02:13,040 Yeah. 69 00:02:13,040 --> 00:02:15,440 And then, you know, say, nope, not going to do it. 70 00:02:15,440 --> 00:02:18,600 The researchers actually have this great example of it in the paper. 71 00:02:18,600 --> 00:02:19,600 They use figure one. 72 00:02:19,600 --> 00:02:20,600 I'll have to check that out. 73 00:02:20,600 --> 00:02:21,600 Pretty cool. 74 00:02:21,600 --> 00:02:25,800 So it's like the AI is literally stopping to think, you know, like, hold on a minute. 75 00:02:25,800 --> 00:02:26,800 This seems fishy. 76 00:02:26,800 --> 00:02:27,800 Right. 77 00:02:27,800 --> 00:02:28,800 I better check my rules before I do anything. 78 00:02:28,800 --> 00:02:29,800 It's pretty impressive. 79 00:02:29,800 --> 00:02:31,800 I mean, how do they actually make it work though? 80 00:02:31,800 --> 00:02:34,800 Well, they use a two stage training process. 81 00:02:34,800 --> 00:02:35,800 Okay. 82 00:02:35,800 --> 00:02:39,160 So in the first stage, the AI learns to think about safety. 83 00:02:39,160 --> 00:02:40,160 Right. 84 00:02:40,160 --> 00:02:43,440 They use a method called context distillation. 85 00:02:43,440 --> 00:02:44,440 Context distillation. 86 00:02:44,440 --> 00:02:45,440 Yes. 87 00:02:45,440 --> 00:02:48,800 So, you know, basically they feed it like safety rules and examples of prompts. 88 00:02:48,800 --> 00:02:49,800 Okay. 89 00:02:49,800 --> 00:02:53,120 And they ask it to come up with responses that take those rules into account. 90 00:02:53,120 --> 00:02:55,280 So it's learning like the rules of the road. 91 00:02:55,280 --> 00:02:56,280 Yeah, exactly. 92 00:02:56,280 --> 00:02:58,000 But, but here's the thing. 93 00:02:58,000 --> 00:03:00,760 Where do those initial responses come from? 94 00:03:00,760 --> 00:03:01,760 Hmm. 95 00:03:01,760 --> 00:03:04,240 Well, I'm assuming from human trainers or something. 96 00:03:04,240 --> 00:03:07,040 Actually, no, it's way more interesting than that. 97 00:03:07,040 --> 00:03:08,920 They use another AI model. 98 00:03:08,920 --> 00:03:09,920 Another AI model. 99 00:03:09,920 --> 00:03:12,280 One that's trained like just for helpfulness. 100 00:03:12,280 --> 00:03:13,280 Okay. 101 00:03:13,280 --> 00:03:14,280 Not for safety. 102 00:03:14,280 --> 00:03:15,280 Yeah. 103 00:03:15,280 --> 00:03:16,280 To generate those initial responses. 104 00:03:16,280 --> 00:03:18,160 So one AI is what? 105 00:03:18,160 --> 00:03:20,040 Training another AI to be safe. 106 00:03:20,040 --> 00:03:21,040 Yeah. 107 00:03:21,040 --> 00:03:22,040 That's wild. 108 00:03:22,040 --> 00:03:23,040 It is. 109 00:03:23,040 --> 00:03:29,640 So at this stage, they use reinforcement learning, you know, to fine tune the AI's reasoning 110 00:03:29,640 --> 00:03:30,640 skills. 111 00:03:30,640 --> 00:03:31,640 Right. 112 00:03:31,640 --> 00:03:32,640 So they use this like judge AI. 113 00:03:32,640 --> 00:03:33,640 Okay. 114 00:03:33,640 --> 00:03:37,160 And this judge AI, it has like access to all the safety specs. 115 00:03:37,160 --> 00:03:39,320 So it's kind of like overseeing the training. 116 00:03:39,320 --> 00:03:40,320 Yeah. 117 00:03:40,320 --> 00:03:41,320 Yeah. 118 00:03:41,320 --> 00:03:43,840 And it evaluates the responses from the AI that's learning. 119 00:03:43,840 --> 00:03:44,840 Right. 120 00:03:44,840 --> 00:03:46,000 And then it provides feedback. 121 00:03:46,000 --> 00:03:49,480 Like, you know, okay, you almost had it, but let's work on your reasoning a bit more. 122 00:03:49,480 --> 00:03:50,480 I like that. 123 00:03:50,480 --> 00:03:53,280 And then the AI is back and forth a constant improvement process. 124 00:03:53,280 --> 00:03:54,280 Exactly. 125 00:03:54,280 --> 00:03:58,680 And this two step process, it helps the AI not only understand the rules, but also learn 126 00:03:58,680 --> 00:04:01,600 how to apply them in, you know, different situations. 127 00:04:01,600 --> 00:04:02,600 Got it. 128 00:04:02,600 --> 00:04:05,840 So the big question now is, does this actually work? 129 00:04:05,840 --> 00:04:07,640 Well, that's what I was going to say. 130 00:04:07,640 --> 00:04:13,280 They tested this whole approach on open AI's new O-Series models. 131 00:04:13,280 --> 00:04:14,280 Okay. 132 00:04:14,280 --> 00:04:15,280 The O-Series. 133 00:04:15,280 --> 00:04:16,280 Yeah. 134 00:04:16,280 --> 00:04:17,280 And the results are pretty impressive. 135 00:04:17,280 --> 00:04:18,280 Oh, really? 136 00:04:18,280 --> 00:04:19,280 Let's hear it. 137 00:04:19,280 --> 00:04:20,280 What kind of improvements did they see? 138 00:04:20,280 --> 00:04:26,280 Well, across a bunch of safety benchmarks, these new models, they showed, you know, significant 139 00:04:26,280 --> 00:04:27,280 improvements. 140 00:04:27,280 --> 00:04:28,280 Significant? 141 00:04:28,280 --> 00:04:29,280 Yeah. 142 00:04:29,280 --> 00:04:31,600 Like, especially when it came to, you know, those jail breaks we were talking about earlier. 143 00:04:31,600 --> 00:04:32,600 Yeah. 144 00:04:32,600 --> 00:04:33,600 All right. 145 00:04:33,600 --> 00:04:36,880 Turns out these O-Series models are way harder to trick. 146 00:04:36,880 --> 00:04:37,880 Oh, that's good. 147 00:04:37,880 --> 00:04:39,880 They're much harder to trick into breaking the rules. 148 00:04:39,880 --> 00:04:44,080 That is definitely reassuring because, you know, we've all seen those headlines about 149 00:04:44,080 --> 00:04:45,560 AI going rogue. 150 00:04:45,560 --> 00:04:46,560 We have. 151 00:04:46,560 --> 00:04:47,560 Right. 152 00:04:47,560 --> 00:04:52,240 And if anyone could just manipulate these, you know, these powerful models into doing 153 00:04:52,240 --> 00:04:53,240 something harmful. 154 00:04:53,240 --> 00:04:54,240 Exactly. 155 00:04:54,240 --> 00:04:55,240 Exactly. 156 00:04:55,240 --> 00:04:59,760 And on top of that, they saw a huge reduction in what they call over-refusals. 157 00:04:59,760 --> 00:05:00,760 Over-refusals? 158 00:05:00,760 --> 00:05:01,760 What's that? 159 00:05:01,760 --> 00:05:06,080 So it's like when the AI refuses a perfectly good request, you know? 160 00:05:06,080 --> 00:05:07,080 Yeah. 161 00:05:07,080 --> 00:05:08,080 Yeah. 162 00:05:08,080 --> 00:05:09,200 Because it's being, like, overly cautious. 163 00:05:09,200 --> 00:05:10,200 I see. 164 00:05:10,200 --> 00:05:11,200 Yeah, I see. 165 00:05:11,200 --> 00:05:15,640 Like, refusing to translate a phrase because it's got a word that might be associated 166 00:05:15,640 --> 00:05:18,000 with a sensitive topic. 167 00:05:18,000 --> 00:05:19,000 Right. 168 00:05:19,000 --> 00:05:21,600 Even if, like, the user's intention is totally harmless. 169 00:05:21,600 --> 00:05:23,280 Yeah, it's being a little too uptight. 170 00:05:23,280 --> 00:05:24,280 Right. 171 00:05:24,280 --> 00:05:25,280 Missing the context. 172 00:05:25,280 --> 00:05:26,280 Yeah, missing the context. 173 00:05:26,280 --> 00:05:30,840 And the research found that these O-Series models, they were, like, way less likely to 174 00:05:30,840 --> 00:05:31,840 do that. 175 00:05:31,840 --> 00:05:32,840 That's great. 176 00:05:32,840 --> 00:05:36,480 They were much better at understanding the context, you know, and applying those safety 177 00:05:36,480 --> 00:05:37,640 rules appropriately. 178 00:05:37,640 --> 00:05:41,480 So it's actually using its judgment instead of just having a knee-jerk reaction. 179 00:05:41,480 --> 00:05:42,480 Right. 180 00:05:42,480 --> 00:05:46,160 And then, like, the refuse button every time whenever something seems, you know, even a 181 00:05:46,160 --> 00:05:47,160 little bit sensitive. 182 00:05:47,160 --> 00:05:48,160 Yeah. 183 00:05:48,160 --> 00:05:52,800 So it's finding that balance between being safe but still being useful. 184 00:05:52,800 --> 00:05:53,800 Exactly. 185 00:05:53,800 --> 00:05:58,680 That's a huge deal because, I mean, what's the point of having an AI assistant if it's, 186 00:05:58,680 --> 00:06:00,840 like, constantly shutting down conversations? 187 00:06:00,840 --> 00:06:01,840 Totally. 188 00:06:01,840 --> 00:06:03,640 You want it to be helpful, not a roadblock. 189 00:06:03,640 --> 00:06:04,640 Exactly. 190 00:06:04,640 --> 00:06:09,600 So it sounds like this deliberative alignment, it's helping these models find that, you 191 00:06:09,600 --> 00:06:10,600 know, that sweet spot. 192 00:06:10,600 --> 00:06:11,600 Yeah. 193 00:06:11,600 --> 00:06:13,960 And there's another interesting finding. 194 00:06:13,960 --> 00:06:17,840 The researchers found that the amount of time the AI had to think actually made a difference 195 00:06:17,840 --> 00:06:19,600 in its safety performance. 196 00:06:19,600 --> 00:06:21,320 Wait, really? 197 00:06:21,320 --> 00:06:23,680 So giving it, like, more time to process. 198 00:06:23,680 --> 00:06:25,160 Yeah, more time to process. 199 00:06:25,160 --> 00:06:26,160 Actually made it safer. 200 00:06:26,160 --> 00:06:27,960 But it's safer. 201 00:06:27,960 --> 00:06:29,200 That's counterintuitive. 202 00:06:29,200 --> 00:06:30,200 It is. 203 00:06:30,200 --> 00:06:36,480 But their findings, they suggest that the O-Series models do better on those really tough, like, 204 00:06:36,480 --> 00:06:39,680 safety challenges when they have more time to think. 205 00:06:39,680 --> 00:06:43,480 So it's not just, like, about how much the AI knows. 206 00:06:43,480 --> 00:06:45,920 It's about how it uses that knowledge. 207 00:06:45,920 --> 00:06:46,920 Exactly. 208 00:06:46,920 --> 00:06:51,120 It's like, you know, if you give a student enough time to, like, really think through 209 00:06:51,120 --> 00:06:52,120 a problem. 210 00:06:52,120 --> 00:06:53,600 Yeah, they're going to come up with a better solution. 211 00:06:53,600 --> 00:06:54,600 Right. 212 00:06:54,600 --> 00:06:56,400 They're more likely to get to a well-reasoned solution. 213 00:06:56,400 --> 00:06:57,400 Exactly. 214 00:06:57,400 --> 00:07:02,520 And that really highlights the power of this chain of thought reasoning that deliberative 215 00:07:02,520 --> 00:07:04,440 alignment uses. 216 00:07:04,440 --> 00:07:09,080 It's like, the AI is, it's not just spitting out an answer. 217 00:07:09,080 --> 00:07:12,840 It's actually thinking through those implications before it acts. 218 00:07:12,840 --> 00:07:14,200 This is really cool stuff. 219 00:07:14,200 --> 00:07:20,120 And it sounds like, you know, it could have some huge implications for the future of AI 220 00:07:20,120 --> 00:07:21,120 safety. 221 00:07:21,120 --> 00:07:22,120 Yeah, it really does. 222 00:07:22,120 --> 00:07:31,160 I mean, it's almost like a fundamental shift in how we're approaching AI safety. 223 00:07:31,160 --> 00:07:33,560 It's not just about adding more rules and restrictions. 224 00:07:33,560 --> 00:07:36,160 It's about changing the way AI thinks. 225 00:07:36,160 --> 00:07:37,160 Exactly. 226 00:07:37,160 --> 00:07:39,520 It's about teaching them to think about safety. 227 00:07:39,520 --> 00:07:40,520 About safety, yeah. 228 00:07:40,520 --> 00:07:45,480 Which is really important because, let's be honest, rules can be broken. 229 00:07:45,480 --> 00:07:46,480 They can. 230 00:07:46,480 --> 00:07:47,480 They can. 231 00:07:47,480 --> 00:07:53,240 But if the AI understands why certain actions are harmful, like if it has, like, its own 232 00:07:53,240 --> 00:07:58,080 sense of responsibility, then it's going to be, like, way harder to manipulate. 233 00:07:58,080 --> 00:07:59,080 Exactly. 234 00:07:59,080 --> 00:08:03,520 It's like the difference between teaching someone to just blindly follow orders versus, 235 00:08:03,520 --> 00:08:06,440 like, teaching them to actually understand the values. 236 00:08:06,440 --> 00:08:07,720 Right, the reasoning behind it. 237 00:08:07,720 --> 00:08:08,720 And the reasoning. 238 00:08:08,720 --> 00:08:09,720 Yeah, yeah. 239 00:08:09,720 --> 00:08:11,920 That's a great point. 240 00:08:11,920 --> 00:08:16,760 And it makes me think about, you know, those jailbreak attacks again. 241 00:08:16,760 --> 00:08:23,600 Like it seems like this deliberative alignment is, it's giving AI a much stronger defense. 242 00:08:23,600 --> 00:08:24,600 It is. 243 00:08:24,600 --> 00:08:25,600 Against those exploits. 244 00:08:25,600 --> 00:08:29,360 Yeah, because we're essentially teaching it to reason. 245 00:08:29,360 --> 00:08:35,680 We're giving it a more robust, like, way to protect itself from malicious actors. 246 00:08:35,680 --> 00:08:38,560 And it's not just about, you know, preventing harm. 247 00:08:38,560 --> 00:08:39,560 Right. 248 00:08:39,560 --> 00:08:43,640 It's also about making sure that the AI can be, like, helpful. 249 00:08:43,640 --> 00:08:44,640 Right, right. 250 00:08:44,640 --> 00:08:46,040 In a wider range of situations. 251 00:08:46,040 --> 00:08:47,040 Yeah. 252 00:08:47,040 --> 00:08:48,040 Without being too cautious. 253 00:08:48,040 --> 00:08:49,040 Exactly. 254 00:08:49,040 --> 00:08:50,040 Like those over-refusal cases. 255 00:08:50,040 --> 00:08:51,040 Yes. 256 00:08:51,040 --> 00:08:54,120 So, deliberative alignment seems to help them find that sweet spot. 257 00:08:54,120 --> 00:08:55,120 Yes, exactly. 258 00:08:55,120 --> 00:08:56,200 Like, finding that balance. 259 00:08:56,200 --> 00:08:57,200 That balance. 260 00:08:57,200 --> 00:08:58,920 Where it can navigate, you know, complex conversations. 261 00:08:58,920 --> 00:08:59,920 Yeah. 262 00:08:59,920 --> 00:09:02,360 And, like, assist with sensitive topics. 263 00:09:02,360 --> 00:09:03,360 Right. 264 00:09:03,360 --> 00:09:04,360 Without shutting down entirely. 265 00:09:04,360 --> 00:09:05,360 Totally, totally. 266 00:09:05,360 --> 00:09:10,280 It's about giving them the tools to make those, like, nuanced judgments. 267 00:09:10,280 --> 00:09:11,280 Right. 268 00:09:11,280 --> 00:09:12,280 Based on the context. 269 00:09:12,280 --> 00:09:14,880 Okay, so we've talked a lot about the benefits, right? 270 00:09:14,880 --> 00:09:15,880 Yeah, we have. 271 00:09:15,880 --> 00:09:20,400 But part of me, I have to admit, is still a little wary. 272 00:09:20,400 --> 00:09:24,280 Can we really trust, you know, AI to make these ethical decisions? 273 00:09:24,280 --> 00:09:25,800 I mean, that's a great question. 274 00:09:25,800 --> 00:09:26,800 It's a big question. 275 00:09:26,800 --> 00:09:31,400 And it's one that we, like, need to approach with a lot of thought and care. 276 00:09:31,400 --> 00:09:32,400 Yeah. 277 00:09:32,400 --> 00:09:37,280 I think it's important to remember that, you know, AI systems, they're created by humans. 278 00:09:37,280 --> 00:09:42,720 They reflect the values and the biases of their creators. 279 00:09:42,720 --> 00:09:43,720 Right. 280 00:09:43,720 --> 00:09:45,880 So it's not just about the AI itself. 281 00:09:45,880 --> 00:09:47,200 It's about the people behind it. 282 00:09:47,200 --> 00:09:49,720 Yeah, the people developing it, deploying it. 283 00:09:49,720 --> 00:09:53,320 We need to make sure they're doing things, you know, ethically and responsibly. 284 00:09:53,320 --> 00:09:54,320 Exactly. 285 00:09:54,320 --> 00:09:56,880 And that's why research like this is so crucial. 286 00:09:56,880 --> 00:09:57,880 It is. 287 00:09:57,880 --> 00:09:58,880 It is. 288 00:09:58,880 --> 00:09:59,960 You know, it's not just about making AI safer. 289 00:09:59,960 --> 00:10:03,120 It's about, like, starting this bigger conversation. 290 00:10:03,120 --> 00:10:04,120 Yeah. 291 00:10:04,120 --> 00:10:05,120 About the ethics. 292 00:10:05,120 --> 00:10:06,120 About the ethics of it all. 293 00:10:06,120 --> 00:10:09,240 Of AI and, like, our role in shaping its future. 294 00:10:09,240 --> 00:10:14,080 You know, speaking of the research, there was one finding I thought was really fascinating. 295 00:10:14,080 --> 00:10:15,240 Oh, yeah. 296 00:10:15,240 --> 00:10:19,160 It's the impact of thinking time on safety performance. 297 00:10:19,160 --> 00:10:20,160 Thinking time. 298 00:10:20,160 --> 00:10:21,160 Yeah. 299 00:10:21,160 --> 00:10:24,280 It's almost like giving the AI a chance to, like, to pause. 300 00:10:24,280 --> 00:10:25,280 Yeah, to reflect. 301 00:10:25,280 --> 00:10:26,280 To reflect. 302 00:10:26,280 --> 00:10:27,280 Yeah. 303 00:10:27,280 --> 00:10:28,280 Before it acts. 304 00:10:28,280 --> 00:10:29,280 Yeah. 305 00:10:29,280 --> 00:10:30,280 It's a tough spot. 306 00:10:30,280 --> 00:10:31,280 Exactly. 307 00:10:31,280 --> 00:10:32,280 Exactly. 308 00:10:32,280 --> 00:10:35,920 And they found, you know, the researchers found that when they gave the models, the O-Series 309 00:10:35,920 --> 00:10:43,640 models, more time to process a request, they actually did better on those tricky safety 310 00:10:43,640 --> 00:10:44,640 challenges. 311 00:10:44,640 --> 00:10:50,920 So, you know, it suggests that slowing down the process can actually lead to better outcomes. 312 00:10:50,920 --> 00:10:53,080 So it's not just about how fast it can respond. 313 00:10:53,080 --> 00:10:54,160 No, not at all. 314 00:10:54,160 --> 00:10:59,960 It's about how effectively it uses that time to, like, reason through things. 315 00:10:59,960 --> 00:11:00,960 To think it through. 316 00:11:00,960 --> 00:11:02,840 Yeah, the consequences of its actions. 317 00:11:02,840 --> 00:11:03,840 Absolutely. 318 00:11:03,840 --> 00:11:08,400 And that's, you know, that's a big shift from how we've traditionally thought about AI. 319 00:11:08,400 --> 00:11:09,400 It is. 320 00:11:09,400 --> 00:11:11,760 Where it's all about, you know, speed and efficiency. 321 00:11:11,760 --> 00:11:12,760 Yeah, yeah, yeah. 322 00:11:12,760 --> 00:11:13,760 Always. 323 00:11:13,760 --> 00:11:20,760 It makes you wonder if by rushing it, we've actually been, like, limiting its capacity 324 00:11:20,760 --> 00:11:22,280 for ethical reasoning. 325 00:11:22,280 --> 00:11:28,320 Maybe, maybe, you know, by giving it that space to think, we can unlock something new. 326 00:11:28,320 --> 00:11:29,720 Yeah, like a whole new level. 327 00:11:29,720 --> 00:11:34,280 A whole new level of responsibility and awareness. 328 00:11:34,280 --> 00:11:35,960 It's a really interesting thought. 329 00:11:35,960 --> 00:11:36,960 It is. 330 00:11:36,960 --> 00:11:37,960 It is. 331 00:11:37,960 --> 00:11:38,960 Like, maybe we've been, you know, underestimated. 332 00:11:38,960 --> 00:11:39,960 Estimated. 333 00:11:39,960 --> 00:11:40,960 AI's potential. 334 00:11:40,960 --> 00:11:41,960 Yeah. 335 00:11:41,960 --> 00:11:42,960 For ethical decision-makers. 336 00:11:42,960 --> 00:11:45,280 Because we're so focused on speed. 337 00:11:45,280 --> 00:11:46,280 Right, right. 338 00:11:46,280 --> 00:11:50,280 That we're not giving it the space to actually deliberate. 339 00:11:50,280 --> 00:11:54,080 It's like, it's like we've been trying to train a sprinter when maybe it should be a 340 00:11:54,080 --> 00:11:55,080 marathon runner. 341 00:11:55,080 --> 00:11:56,960 Yeah, yeah, that's a great analogy. 342 00:11:56,960 --> 00:12:00,080 Start blowing down, helps you, you know, see things more clearly. 343 00:12:00,080 --> 00:12:01,080 Absolutely. 344 00:12:01,080 --> 00:12:02,080 And make better choices. 345 00:12:02,080 --> 00:12:05,680 And that, that actually leads us to another really cool finding from this research. 346 00:12:05,680 --> 00:12:06,680 Oh yeah. 347 00:12:06,680 --> 00:12:07,680 What is that? 348 00:12:07,680 --> 00:12:14,440 Their ability to generalize, like, these O-series models, they can take their safety knowledge 349 00:12:14,440 --> 00:12:18,520 right and apply it to new situations, like things they haven't even been trained on. 350 00:12:18,520 --> 00:12:19,520 Yeah, that was a big one. 351 00:12:19,520 --> 00:12:23,480 Yeah, they tested this, you know, with non-English languages, they used encoded data. 352 00:12:23,480 --> 00:12:24,480 And it still worked. 353 00:12:24,480 --> 00:12:25,480 It worked. 354 00:12:25,480 --> 00:12:26,480 It was amazing. 355 00:12:26,480 --> 00:12:30,320 Which suggests that, like, they're not just memorizing the rules. 356 00:12:30,320 --> 00:12:35,440 They're actually learning to reason about it in, like, a more fundamental way. 357 00:12:35,440 --> 00:12:36,440 Exactly. 358 00:12:36,440 --> 00:12:38,320 It's not rote memorization. 359 00:12:38,320 --> 00:12:40,760 It's about understanding those principles. 360 00:12:40,760 --> 00:12:43,360 And being able to apply them flexibly. 361 00:12:43,360 --> 00:12:46,320 So it's not just about, you know, teaching them to follow the rules. 362 00:12:46,320 --> 00:12:47,320 Yeah. 363 00:12:47,320 --> 00:12:48,840 It's teaching them to, like, understand why. 364 00:12:48,840 --> 00:12:49,840 The why, yeah. 365 00:12:49,840 --> 00:12:50,840 The reasoning behind it. 366 00:12:50,840 --> 00:12:51,840 The reasoning behind it. 367 00:12:51,840 --> 00:12:55,600 So that they can adapt to new situations, new challenges. 368 00:12:55,600 --> 00:12:56,600 Exactly. 369 00:12:56,600 --> 00:12:57,720 That's really impressive. 370 00:12:57,720 --> 00:12:58,720 It is. 371 00:12:58,720 --> 00:13:03,000 And that ability to generalize, you know, that's going to be so important. 372 00:13:03,000 --> 00:13:05,680 For AI safety moving forward, right? 373 00:13:05,680 --> 00:13:08,640 In a world that's constantly changing. 374 00:13:08,640 --> 00:13:12,840 And where we're facing new ethical dilemmas all the time. 375 00:13:12,840 --> 00:13:13,840 Right. 376 00:13:13,840 --> 00:13:14,840 Right. 377 00:13:14,840 --> 00:13:19,840 And I think that, you know, beyond just preventing harm. 378 00:13:19,840 --> 00:13:20,840 Yeah. 379 00:13:20,840 --> 00:13:21,840 Beyond just preventing harm. 380 00:13:21,840 --> 00:13:25,840 Imagine, like, the possibilities in fields like, like, healthcare. 381 00:13:25,840 --> 00:13:26,840 Healthcare. 382 00:13:26,840 --> 00:13:27,840 Education. 383 00:13:27,840 --> 00:13:28,840 Yeah. 384 00:13:28,840 --> 00:13:30,840 Or even, like, law enforcement. 385 00:13:30,840 --> 00:13:31,840 Absolutely. 386 00:13:31,840 --> 00:13:35,840 If AI can learn to reason this well about safety and ethics. 387 00:13:35,840 --> 00:13:36,840 Right. 388 00:13:36,840 --> 00:13:40,840 I mean, it could totally change how we approach decision making in all of those fields. 389 00:13:40,840 --> 00:13:45,840 Imagine, like, an AI system that could, you know, assist doctors. 390 00:13:45,840 --> 00:13:46,840 Yeah. 391 00:13:46,840 --> 00:13:48,840 And making these, like, life or death decisions. 392 00:13:48,840 --> 00:13:49,840 Right. 393 00:13:49,840 --> 00:13:52,840 Taking into account not just, you know, medical history. 394 00:13:52,840 --> 00:13:53,840 Yeah. 395 00:13:53,840 --> 00:13:56,840 But also, like, their ethical considerations and personal values. 396 00:13:56,840 --> 00:13:57,840 Exactly. 397 00:13:57,840 --> 00:13:59,840 Or think about, like, an AI tutor. 398 00:13:59,840 --> 00:14:00,840 Yeah. 399 00:14:00,840 --> 00:14:02,840 That could, like, personalize education. 400 00:14:02,840 --> 00:14:03,840 Right. 401 00:14:03,840 --> 00:14:04,840 Right. 402 00:14:04,840 --> 00:14:05,840 For each student, you know. 403 00:14:05,840 --> 00:14:07,840 Making sure the content is engaging and ethical. 404 00:14:07,840 --> 00:14:08,840 Yeah. 405 00:14:08,840 --> 00:14:09,840 That's amazing. 406 00:14:09,840 --> 00:14:10,840 The possibilities are. 407 00:14:10,840 --> 00:14:11,840 They're so exciting. 408 00:14:11,840 --> 00:14:12,840 They are. 409 00:14:12,840 --> 00:14:13,840 They are. 410 00:14:13,840 --> 00:14:17,840 But I think it's important, as with any powerful technology, to proceed carefully. 411 00:14:17,840 --> 00:14:18,840 Right. 412 00:14:18,840 --> 00:14:22,840 You know, to make sure that AI development is guided by ethical principles. 413 00:14:22,840 --> 00:14:23,840 Absolutely. 414 00:14:23,840 --> 00:14:24,840 And commitment to human well-being. 415 00:14:24,840 --> 00:14:27,840 I couldn't agree more, you know. 416 00:14:27,840 --> 00:14:32,840 As we're pushing these boundaries of AI, it's crucial that we put safety first. 417 00:14:32,840 --> 00:14:33,840 We have to. 418 00:14:33,840 --> 00:14:40,840 Responsibility, transparency, and, you know, making sure that AI is a force for good in 419 00:14:40,840 --> 00:14:41,840 the world. 420 00:14:41,840 --> 00:14:42,840 A force for good. 421 00:14:42,840 --> 00:14:43,840 Well said. 422 00:14:43,840 --> 00:14:44,840 Well said. 423 00:14:44,840 --> 00:14:46,840 So, on that note, I think that wraps up our deep dive. 424 00:14:46,840 --> 00:14:47,840 Yeah. 425 00:14:47,840 --> 00:14:48,840 Into deliberative alignment. 426 00:14:48,840 --> 00:14:50,840 It has been a fascinating one. 427 00:14:50,840 --> 00:14:51,840 It really has. 428 00:14:51,840 --> 00:14:54,840 You know, it's given us this glimpse into a future. 429 00:14:54,840 --> 00:14:55,840 Yeah. 430 00:14:55,840 --> 00:14:58,840 Where AI is not only powerful, but also, you know, responsible. 431 00:14:58,840 --> 00:14:59,840 Yeah. 432 00:14:59,840 --> 00:15:02,840 Ethical and aligned with human values. 433 00:15:02,840 --> 00:15:04,840 And an exciting future to imagine. 434 00:15:04,840 --> 00:15:05,840 It is. 435 00:15:05,840 --> 00:15:07,840 And it's one that we all have a role in shaping. 436 00:15:07,840 --> 00:15:08,840 You do. 437 00:15:08,840 --> 00:15:10,840 So, we've reached the end of our deep dive for today. 438 00:15:10,840 --> 00:15:14,840 Thanks for joining us on this exploration, you know, of this cutting-edge AI safety 439 00:15:14,840 --> 00:15:15,840 research. 440 00:15:15,840 --> 00:15:16,840 It's been a pleasure. 441 00:15:16,840 --> 00:15:19,840 We hope this episode has maybe sparked some curiosity. 442 00:15:19,840 --> 00:15:20,840 Yeah, I hope so. 443 00:15:20,840 --> 00:15:25,840 And got you thinking more deeply about the ethical implications of AI. 444 00:15:25,840 --> 00:15:26,840 Ethical implications. 445 00:15:26,840 --> 00:15:27,840 That's key. 446 00:15:27,840 --> 00:15:31,840 So, as always, we encourage you to keep exploring these questions. 447 00:15:31,840 --> 00:15:32,840 Yeah, keep thinking about it. 448 00:15:32,840 --> 00:15:37,840 And stay engaged in the conversation about, you know, the future that we want to build 449 00:15:37,840 --> 00:15:38,840 with AI. 450 00:15:38,840 --> 00:15:39,840 That's right. 451 00:15:39,840 --> 00:15:40,840 Until next time. 452 00:15:40,840 --> 00:15:41,840 Until next time. 453 00:15:41,840 --> 00:15:42,840 Welcome back, everyone. 454 00:15:42,840 --> 00:15:48,840 I think we were talking about how this deliberative alignment is helping AI get smarter and safer. 455 00:15:48,840 --> 00:15:49,840 Yeah. 456 00:15:49,840 --> 00:15:54,840 You know, the more I think about this research, the more it seems like a fundamental shift. 457 00:15:54,840 --> 00:15:55,840 I know, right? 458 00:15:55,840 --> 00:15:59,840 In how we're approaching this whole, you know, AI safety thing. 459 00:15:59,840 --> 00:16:00,840 Yeah. 460 00:16:00,840 --> 00:16:03,840 It's not just about adding more and more rules and restrictions. 461 00:16:03,840 --> 00:16:04,840 No, it's not. 462 00:16:04,840 --> 00:16:08,560 It's about actually changing how the AI thinks about safety. 463 00:16:08,560 --> 00:16:09,560 That's a great way to put it. 464 00:16:09,560 --> 00:16:10,560 Yeah. 465 00:16:10,560 --> 00:16:11,560 Right. 466 00:16:11,560 --> 00:16:14,520 It's like we're moving from, you know, telling the AI, don't do this or don't do that, to 467 00:16:14,520 --> 00:16:21,440 actually explaining why certain actions are harmful and helping it develop its own, like, 468 00:16:21,440 --> 00:16:22,440 sense of responsibility. 469 00:16:22,440 --> 00:16:23,440 Exactly. 470 00:16:23,440 --> 00:16:26,840 And that's so important because, I mean, rules can be broken, right? 471 00:16:26,840 --> 00:16:28,040 They can, of course, yeah. 472 00:16:28,040 --> 00:16:33,800 But if the AI understands, you know, those underlying principles, it's going to be a 473 00:16:33,800 --> 00:16:37,520 lot harder to manipulate or trick it into doing something it shouldn't. 474 00:16:37,520 --> 00:16:38,840 That's the hope at least. 475 00:16:38,840 --> 00:16:39,840 Right. 476 00:16:39,840 --> 00:16:40,840 Yeah. 477 00:16:40,840 --> 00:16:41,840 It's like, it's kind of like a... 478 00:16:41,840 --> 00:16:46,040 It's like the difference between teaching someone to blindly follow orders. 479 00:16:46,040 --> 00:16:47,040 Blindly. 480 00:16:47,040 --> 00:16:50,360 Versus teaching them, like, you know, to understand the values. 481 00:16:50,360 --> 00:16:51,360 The values behind it. 482 00:16:51,360 --> 00:16:53,040 The reasons behind those orders. 483 00:16:53,040 --> 00:16:54,040 Yeah. 484 00:16:54,040 --> 00:16:55,040 Yeah. 485 00:16:55,040 --> 00:16:56,040 Perfect analogy. 486 00:16:56,040 --> 00:16:59,640 But those jailbreak attacks again. 487 00:16:59,640 --> 00:17:00,640 Oh, yeah. 488 00:17:00,640 --> 00:17:01,640 The jailbreak stuff. 489 00:17:01,640 --> 00:17:06,520 It seems like deliberative alignment is giving AI a much stronger defense. 490 00:17:06,520 --> 00:17:07,520 They think so. 491 00:17:07,520 --> 00:17:08,520 Yeah. 492 00:17:08,520 --> 00:17:09,520 Against those exploits. 493 00:17:09,520 --> 00:17:10,520 Absolutely. 494 00:17:10,520 --> 00:17:11,520 Yeah. 495 00:17:11,520 --> 00:17:14,360 Because we're essentially teaching it how to reason through things. 496 00:17:14,360 --> 00:17:21,240 So we're giving it, you know, a way more robust, flexible way to protect itself from 497 00:17:21,240 --> 00:17:23,440 bad actors, essentially. 498 00:17:23,440 --> 00:17:27,120 And it's not just about preventing harm, is it? 499 00:17:27,120 --> 00:17:28,120 No. 500 00:17:28,120 --> 00:17:29,120 No, not at all. 501 00:17:29,120 --> 00:17:32,640 It's also about making sure that the AI can be helpful. 502 00:17:32,640 --> 00:17:33,640 Helpful, yeah. 503 00:17:33,640 --> 00:17:35,760 You know, in a wider range of situations. 504 00:17:35,760 --> 00:17:36,760 In more situations. 505 00:17:36,760 --> 00:17:38,160 Without being overly cautious. 506 00:17:38,160 --> 00:17:39,160 Yeah. 507 00:17:39,160 --> 00:17:40,160 Without being overly cautious. 508 00:17:40,160 --> 00:17:41,160 Right. 509 00:17:41,160 --> 00:17:42,160 Exactly. 510 00:17:42,160 --> 00:17:44,480 Think about those over-refusal cases we talked about earlier. 511 00:17:44,480 --> 00:17:48,560 Deliberative alignment is really helping them to find that balance. 512 00:17:48,560 --> 00:17:49,560 Yeah. 513 00:17:49,560 --> 00:17:52,160 Between, you know, being safe but still being useful. 514 00:17:52,160 --> 00:17:53,160 Yeah. 515 00:17:53,160 --> 00:17:54,160 And that sweet spot. 516 00:17:54,160 --> 00:17:55,160 It's that sweet spot. 517 00:17:55,160 --> 00:18:02,680 Where the AI can, you know, navigate these complex conversations and, like, assist with 518 00:18:02,680 --> 00:18:03,680 sensitive topics. 519 00:18:03,680 --> 00:18:04,680 Right. 520 00:18:04,680 --> 00:18:06,320 Without, you know, just shutting down completely. 521 00:18:06,320 --> 00:18:07,480 That shutting down entirely, yeah. 522 00:18:07,480 --> 00:18:08,480 Right. 523 00:18:08,480 --> 00:18:09,480 Yeah. 524 00:18:09,480 --> 00:18:14,480 It's about giving the AI the tools and the training to be able to make those judgment 525 00:18:14,480 --> 00:18:15,480 calls. 526 00:18:15,480 --> 00:18:16,480 Those nuanced judgments. 527 00:18:16,480 --> 00:18:19,360 The nuanced judgments, yeah, based on, you know, the context and the rules. 528 00:18:19,360 --> 00:18:20,360 Yeah. 529 00:18:20,360 --> 00:18:21,360 Okay. 530 00:18:21,360 --> 00:18:22,360 So we've talked about, you know, the benefits. 531 00:18:22,360 --> 00:18:23,360 Yeah, we have. 532 00:18:23,360 --> 00:18:26,680 But I gotta say, I'm still a little wary. 533 00:18:26,680 --> 00:18:27,800 I understand that. 534 00:18:27,800 --> 00:18:32,680 You know, can we really trust AI to make these kinds of ethical decisions? 535 00:18:32,680 --> 00:18:34,880 I mean, that's a big question. 536 00:18:34,880 --> 00:18:37,200 And I don't, you know, I don't think there's an easy answer. 537 00:18:37,200 --> 00:18:38,200 Right. 538 00:18:38,200 --> 00:18:42,040 But I think, you know, it's important to remember that AI systems, they're created by humans. 539 00:18:42,040 --> 00:18:43,040 Right. 540 00:18:43,040 --> 00:18:46,960 And they reflect the values and the biases of their creators. 541 00:18:46,960 --> 00:18:49,200 So it's not just about the AI itself. 542 00:18:49,200 --> 00:18:50,200 No. 543 00:18:50,200 --> 00:18:52,240 It's about, you know, the people who are developing it. 544 00:18:52,240 --> 00:18:53,720 People developing it, deploying it. 545 00:18:53,720 --> 00:18:54,720 Yeah. 546 00:18:54,720 --> 00:18:55,720 Deploying it. 547 00:18:55,720 --> 00:18:59,600 Making sure that they are, you know, operating ethically and responsibly. 548 00:18:59,600 --> 00:19:00,600 Right. 549 00:19:00,600 --> 00:19:01,600 Exactly. 550 00:19:01,600 --> 00:19:03,720 And that's why, you know, research like this. 551 00:19:03,720 --> 00:19:04,800 Research like this, yeah. 552 00:19:04,800 --> 00:19:06,280 Is so crucial. 553 00:19:06,280 --> 00:19:07,280 Absolutely. 554 00:19:07,280 --> 00:19:12,080 Because it's, it's sparking that conversation about the ethical implications of AI and our 555 00:19:12,080 --> 00:19:13,960 role in shaping it. 556 00:19:13,960 --> 00:19:17,960 You know, and speaking of the research, there was one finding I thought was really interesting. 557 00:19:17,960 --> 00:19:18,960 Oh yeah. 558 00:19:18,960 --> 00:19:19,960 Which one? 559 00:19:19,960 --> 00:19:22,200 It was the impact of thinking time. 560 00:19:22,200 --> 00:19:23,200 Thinking time. 561 00:19:23,200 --> 00:19:24,200 On safety performance. 562 00:19:24,200 --> 00:19:25,200 Oh, that's a great point. 563 00:19:25,200 --> 00:19:26,200 Yeah. 564 00:19:26,200 --> 00:19:28,760 You know, it's almost like giving the AI like a moment to pause. 565 00:19:28,760 --> 00:19:29,760 A moment to pause. 566 00:19:29,760 --> 00:19:30,760 To reflect. 567 00:19:30,760 --> 00:19:31,760 To reflect. 568 00:19:31,760 --> 00:19:32,760 Yeah. 569 00:19:32,760 --> 00:19:33,760 Just like, like we would. 570 00:19:33,760 --> 00:19:34,760 Before it acts. 571 00:19:34,760 --> 00:19:35,760 Right. 572 00:19:35,760 --> 00:19:36,760 Before it acts. 573 00:19:36,760 --> 00:19:37,760 Exactly. 574 00:19:37,760 --> 00:19:38,760 Like we would in, you know, a challenging situation. 575 00:19:38,760 --> 00:19:39,760 A challenging situation. 576 00:19:39,760 --> 00:19:41,040 And the, and the researchers, they found. 577 00:19:41,040 --> 00:19:42,040 You found. 578 00:19:42,040 --> 00:19:45,280 That, you know, when they gave those O-Series models more time. 579 00:19:45,280 --> 00:19:46,280 More time. 580 00:19:46,280 --> 00:19:47,400 To like process a request. 581 00:19:47,400 --> 00:19:49,520 To think it through. 582 00:19:49,520 --> 00:19:52,120 They did better on those really difficult challenges. 583 00:19:52,120 --> 00:19:53,120 They did. 584 00:19:53,120 --> 00:19:54,120 They did. 585 00:19:54,120 --> 00:19:55,120 Right. 586 00:19:55,120 --> 00:19:56,120 Which, which is kind of amazing. 587 00:19:56,120 --> 00:19:57,120 It is. 588 00:19:57,120 --> 00:19:58,400 It suggests that, that slowing down the process. 589 00:19:58,400 --> 00:19:59,400 Yeah. 590 00:19:59,400 --> 00:20:01,840 Can actually lead to, you know, safer outcomes. 591 00:20:01,840 --> 00:20:05,520 So, so it's not just about how fast the AI can respond. 592 00:20:05,520 --> 00:20:06,520 No, not at all. 593 00:20:06,520 --> 00:20:10,600 It's about like how effectively it uses that time. 594 00:20:10,600 --> 00:20:11,600 Uses that time. 595 00:20:11,600 --> 00:20:13,240 To, to reason through. 596 00:20:13,240 --> 00:20:14,240 To reason through. 597 00:20:14,240 --> 00:20:15,840 To the potential consequence. 598 00:20:15,840 --> 00:20:16,840 Right. 599 00:20:16,840 --> 00:20:17,840 Exactly. 600 00:20:17,840 --> 00:20:18,840 Yeah. 601 00:20:18,840 --> 00:20:19,840 And that's, that's a big difference. 602 00:20:19,840 --> 00:20:20,840 Yeah. 603 00:20:20,840 --> 00:20:21,840 You know, from how we traditionally view AI. 604 00:20:21,840 --> 00:20:22,840 It is. 605 00:20:22,840 --> 00:20:27,520 Where speed and efficiency are, you know, everything paramount. 606 00:20:27,520 --> 00:20:30,760 It makes you wonder if, if, you know, by rushing AI. 607 00:20:30,760 --> 00:20:31,760 By rushing it. 608 00:20:31,760 --> 00:20:32,760 Yeah. 609 00:20:32,760 --> 00:20:33,920 We've been like limiting its capacity. 610 00:20:33,920 --> 00:20:34,920 Yeah. 611 00:20:34,920 --> 00:20:35,920 For ethical reasoning. 612 00:20:35,920 --> 00:20:36,920 Right. 613 00:20:36,920 --> 00:20:40,840 We might be, we might be underestimating AI's potential for ethical decision making 614 00:20:40,840 --> 00:20:43,640 because we're so focused on speed. 615 00:20:43,640 --> 00:20:46,360 It's like, it's like we've been trying to train it to be a sprinter. 616 00:20:46,360 --> 00:20:47,360 A sprinter. 617 00:20:47,360 --> 00:20:50,320 When maybe it's better suited to be like a marathon runner. 618 00:20:50,320 --> 00:20:51,320 Yeah. 619 00:20:51,320 --> 00:20:52,560 When I'm slowing down. 620 00:20:52,560 --> 00:20:53,560 Yeah. 621 00:20:53,560 --> 00:20:55,240 Helps you see things more clearly. 622 00:20:55,240 --> 00:20:56,240 Yes. 623 00:20:56,240 --> 00:20:57,440 Make better choices. 624 00:20:57,440 --> 00:21:01,200 And that, that actually brings us to another really interesting finding from this research. 625 00:21:01,200 --> 00:21:02,200 It does. 626 00:21:02,200 --> 00:21:03,200 It does. 627 00:21:03,200 --> 00:21:04,800 Which is, the ability of these models. 628 00:21:04,800 --> 00:21:05,800 Yeah. 629 00:21:05,800 --> 00:21:07,480 To generalize their safety knowledge. 630 00:21:07,480 --> 00:21:08,480 In new situations. 631 00:21:08,480 --> 00:21:09,480 In new situations. 632 00:21:09,480 --> 00:21:10,480 Yeah. 633 00:21:10,480 --> 00:21:12,720 They tested this, you know, with, with different languages. 634 00:21:12,720 --> 00:21:13,720 Yeah. 635 00:21:13,720 --> 00:21:14,720 They used encoded data. 636 00:21:14,720 --> 00:21:15,720 And it worked. 637 00:21:15,720 --> 00:21:16,720 It worked. 638 00:21:16,720 --> 00:21:17,720 It was amazing. 639 00:21:17,720 --> 00:21:20,440 Which, which suggests that like they're not just, you know, memorizing rules. 640 00:21:20,440 --> 00:21:21,440 No, they're, they're learning. 641 00:21:21,440 --> 00:21:22,440 They're learning. 642 00:21:22,440 --> 00:21:23,440 You're learning. 643 00:21:23,440 --> 00:21:29,240 To reason about safety in a, in a much more, uh, a fundamental way. 644 00:21:29,240 --> 00:21:30,400 In a more fundamental way. 645 00:21:30,400 --> 00:21:31,400 Yeah. 646 00:21:31,400 --> 00:21:32,400 Right. 647 00:21:32,400 --> 00:21:36,600 It's about equipping the AI with that, with that deeper understanding of ethical principles 648 00:21:36,600 --> 00:21:39,080 that it can apply flexibly. 649 00:21:39,080 --> 00:21:42,160 So it's not just about, you know, teaching it to follow the rules. 650 00:21:42,160 --> 00:21:43,160 No. 651 00:21:43,160 --> 00:21:45,120 It's, it's teaching it to understand the reasoning. 652 00:21:45,120 --> 00:21:46,120 The reasoning. 653 00:21:46,120 --> 00:21:47,120 Behind those rules. 654 00:21:47,120 --> 00:21:48,120 Behind those rules. 655 00:21:48,120 --> 00:21:49,120 So that it can adapt, right. 656 00:21:49,120 --> 00:21:50,120 So it can adapt. 657 00:21:50,120 --> 00:21:51,640 So it can adapt to new situations, new challenges. 658 00:21:51,640 --> 00:21:52,800 Now that's impressive. 659 00:21:52,800 --> 00:21:53,800 That is impressive. 660 00:21:53,800 --> 00:21:56,840 And, and that ability to generalize, that's going to be so important. 661 00:21:56,840 --> 00:21:57,840 That can be critical. 662 00:21:57,840 --> 00:21:58,840 Yeah. 663 00:21:58,840 --> 00:21:59,840 For ensuring AI safety. 664 00:21:59,840 --> 00:22:00,840 Yeah. 665 00:22:00,840 --> 00:22:02,360 You know, in a world that's constantly changing. 666 00:22:02,360 --> 00:22:03,360 Yeah. 667 00:22:03,360 --> 00:22:06,320 You know, presenting all these, these new ethical dilemma. 668 00:22:06,320 --> 00:22:07,320 Ethical dilemmas. 669 00:22:07,320 --> 00:22:08,320 Yeah. 670 00:22:08,320 --> 00:22:09,320 Right. 671 00:22:09,320 --> 00:22:12,800 Speaking of, speaking of new challenges, it makes me think about, you know, the potential 672 00:22:12,800 --> 00:22:14,200 applications of this. 673 00:22:14,200 --> 00:22:15,200 Oh yeah. 674 00:22:15,200 --> 00:22:16,200 The application. 675 00:22:16,200 --> 00:22:17,200 Beyond just preventing harm. 676 00:22:17,200 --> 00:22:18,200 It's just preventing harm. 677 00:22:18,200 --> 00:22:19,200 Yeah. 678 00:22:19,200 --> 00:22:22,040 And there's possibilities in fields like healthcare. 679 00:22:22,040 --> 00:22:23,040 Yeah. 680 00:22:23,040 --> 00:22:24,040 Education. 681 00:22:24,040 --> 00:22:25,040 Education. 682 00:22:25,040 --> 00:22:26,040 Even like law enforcement. 683 00:22:26,040 --> 00:22:27,040 Law enforcement. 684 00:22:27,040 --> 00:22:28,040 Yeah. 685 00:22:28,040 --> 00:22:31,040 If, if AI can learn to reason this well. 686 00:22:31,040 --> 00:22:32,040 Right. 687 00:22:32,040 --> 00:22:35,440 I mean, it could revolutionize how we, how we approach all these really complex decisions. 688 00:22:35,440 --> 00:22:38,200 Imagine an AI system that could like, you know, help doctors. 689 00:22:38,200 --> 00:22:39,200 Help doctors. 690 00:22:39,200 --> 00:22:40,200 Yeah. 691 00:22:40,200 --> 00:22:42,040 With these, you know, life or death decisions. 692 00:22:42,040 --> 00:22:43,040 Yeah. 693 00:22:43,040 --> 00:22:45,440 Taking into account, you know, not just medical history. 694 00:22:45,440 --> 00:22:46,440 Not just medical history. 695 00:22:46,440 --> 00:22:49,760 But also like ethical considerations. 696 00:22:49,760 --> 00:22:50,960 Ethical considerations. 697 00:22:50,960 --> 00:22:51,960 Personal values. 698 00:22:51,960 --> 00:22:52,960 Personal values. 699 00:22:52,960 --> 00:22:53,960 Right. 700 00:22:53,960 --> 00:22:56,640 Or, or think about an AI tutor. 701 00:22:56,640 --> 00:22:57,640 An AI tutor. 702 00:22:57,640 --> 00:22:58,640 Yeah. 703 00:22:58,640 --> 00:22:59,640 Could like, personalize education. 704 00:22:59,640 --> 00:23:00,640 And personalize education. 705 00:23:00,640 --> 00:23:01,640 That's right. 706 00:23:01,640 --> 00:23:02,640 For each student. 707 00:23:02,640 --> 00:23:03,640 I could do it. 708 00:23:03,640 --> 00:23:04,640 For each student. 709 00:23:04,640 --> 00:23:05,640 Yeah. 710 00:23:05,640 --> 00:23:08,160 Making sure that the content is, is not only engaging. 711 00:23:08,160 --> 00:23:09,160 Thank you. 712 00:23:09,160 --> 00:23:10,160 But also ethically sound. 713 00:23:10,160 --> 00:23:11,160 Ethically sound. 714 00:23:11,160 --> 00:23:12,160 Yeah. 715 00:23:12,160 --> 00:23:13,640 It's, it's amazing to think about. 716 00:23:13,640 --> 00:23:14,640 Right. 717 00:23:14,640 --> 00:23:16,720 The possibilities are, they're so exciting. 718 00:23:16,720 --> 00:23:17,720 They are exciting. 719 00:23:17,720 --> 00:23:18,720 Yeah. 720 00:23:18,720 --> 00:23:23,360 But, but of course, you know, as with any powerful technology, we, we got to be careful. 721 00:23:23,360 --> 00:23:24,360 We do. 722 00:23:24,360 --> 00:23:25,360 We do. 723 00:23:25,360 --> 00:23:29,640 We need to make sure that, you know, AI development is, is guided by those ethical principles. 724 00:23:29,640 --> 00:23:31,120 Guided by ethical principles. 725 00:23:31,120 --> 00:23:34,560 You know, that commitment to human well-being. 726 00:23:34,560 --> 00:23:35,560 Human well-being. 727 00:23:35,560 --> 00:23:36,560 That's right. 728 00:23:36,560 --> 00:23:37,560 I couldn't agree more. 729 00:23:37,560 --> 00:23:38,560 Yeah. 730 00:23:38,560 --> 00:23:39,560 Yeah. 731 00:23:39,560 --> 00:23:44,080 As we, as we push these boundaries of AI, it's, it's crucial that, that we prioritize 732 00:23:44,080 --> 00:23:45,080 safety. 733 00:23:45,080 --> 00:23:46,080 Prioritize safety. 734 00:23:46,080 --> 00:23:47,080 Yeah. 735 00:23:47,080 --> 00:23:48,080 Responsibility. 736 00:23:48,080 --> 00:23:49,080 Responsibility. 737 00:23:49,080 --> 00:23:50,080 Transparency. 738 00:23:50,080 --> 00:23:51,080 Transparency. 739 00:23:51,080 --> 00:23:52,960 You know, making sure that AI is, is a force for good. 740 00:23:52,960 --> 00:23:54,560 A force for good in the world. 741 00:23:54,560 --> 00:23:55,560 Yeah. 742 00:23:55,560 --> 00:23:56,560 Well said. 743 00:23:56,560 --> 00:23:57,560 Well said. 744 00:23:57,560 --> 00:23:59,440 On that note, I think we should take a quick pause. 745 00:23:59,440 --> 00:24:00,440 That's good. 746 00:24:00,440 --> 00:24:01,440 Gather our thoughts. 747 00:24:01,440 --> 00:24:02,440 Gather our thoughts. 748 00:24:02,440 --> 00:24:06,440 And when we come back, we'll, we'll dive into the, the broader implications of this research. 749 00:24:06,440 --> 00:24:07,440 Yeah. 750 00:24:07,440 --> 00:24:08,880 And what it might mean for the future of AI. 751 00:24:08,880 --> 00:24:10,520 The future of AI. 752 00:24:10,520 --> 00:24:11,520 Stay tuned. 753 00:24:11,520 --> 00:24:12,520 Stay tuned. 754 00:24:12,520 --> 00:24:15,120 Welcome back everyone. I'm ready to keep unpacking this research. 755 00:24:15,120 --> 00:24:16,120 How about you? 756 00:24:16,120 --> 00:24:17,120 Definitely. 757 00:24:17,120 --> 00:24:20,280 This deep dive into deliberative alignment. 758 00:24:20,280 --> 00:24:23,160 You know, it's really got me thinking about the future of AI. 759 00:24:23,160 --> 00:24:24,160 Me too. 760 00:24:24,160 --> 00:24:25,160 Me too. 761 00:24:25,160 --> 00:24:30,240 Like, you know, a future where AI isn't just powerful, but also like, you know, responsible. 762 00:24:30,240 --> 00:24:31,240 Responsible. 763 00:24:31,240 --> 00:24:32,240 Ethical. 764 00:24:32,240 --> 00:24:33,240 Yeah. 765 00:24:33,240 --> 00:24:34,760 And really aligned with, you know, with human values. 766 00:24:34,760 --> 00:24:36,920 It's a, it's a really exciting possibility. 767 00:24:36,920 --> 00:24:37,920 It is. 768 00:24:37,920 --> 00:24:38,920 It is. 769 00:24:38,920 --> 00:24:40,960 And I think this research, you know, this deliberative alignment. 770 00:24:40,960 --> 00:24:41,960 Yeah. 771 00:24:41,960 --> 00:24:47,040 It's just a really solid foundation to build that future on, you know? 772 00:24:47,040 --> 00:24:48,040 Absolutely. 773 00:24:48,040 --> 00:24:54,120 And as we've talked about, it's showing like real promise in tackling those big AI safety 774 00:24:54,120 --> 00:24:55,120 challenges, right? 775 00:24:55,120 --> 00:24:56,120 It is. 776 00:24:56,120 --> 00:24:57,120 Yeah. 777 00:24:57,120 --> 00:24:58,120 Like those, those pesky jailbreak attacks. 778 00:24:58,120 --> 00:24:59,120 Yeah. 779 00:24:59,120 --> 00:25:00,120 This is jailbreaks. 780 00:25:00,120 --> 00:25:02,520 And, and the problem of, you know, AI being a little too cautious. 781 00:25:02,520 --> 00:25:03,520 Right. 782 00:25:03,520 --> 00:25:04,520 The over refusal. 783 00:25:04,520 --> 00:25:05,520 The over refusal, exactly. 784 00:25:05,520 --> 00:25:06,520 Yeah. 785 00:25:06,520 --> 00:25:09,880 It seems like, you know, giving AI this ability to actually reason through those safety guidelines 786 00:25:09,880 --> 00:25:14,760 before it takes action rather than just, you know, reacting based on, on pre-programmed 787 00:25:14,760 --> 00:25:15,760 rules. 788 00:25:15,760 --> 00:25:16,760 Yeah. 789 00:25:16,760 --> 00:25:18,160 It's, it's really a game changer. 790 00:25:18,160 --> 00:25:19,160 It is. 791 00:25:19,160 --> 00:25:20,160 It's like, we're giving it the tools. 792 00:25:20,160 --> 00:25:21,160 The tools. 793 00:25:21,160 --> 00:25:23,280 To make those more, you know, informed decisions. 794 00:25:23,280 --> 00:25:24,280 Get formed. 795 00:25:24,280 --> 00:25:25,280 Yeah. 796 00:25:25,280 --> 00:25:26,280 Those nuanced judgments. 797 00:25:26,280 --> 00:25:27,280 Guaranteed judgments, exactly. 798 00:25:27,280 --> 00:25:30,360 And what I find really fascinating is, is that this ability. 799 00:25:30,360 --> 00:25:31,360 Yeah. 800 00:25:31,360 --> 00:25:33,000 To reason about safety. 801 00:25:33,000 --> 00:25:37,200 It's, it's not just limited to those specific situations, right? 802 00:25:37,200 --> 00:25:38,200 Right. 803 00:25:38,200 --> 00:25:39,560 That the AI has been trained on. 804 00:25:39,560 --> 00:25:40,560 Right. 805 00:25:40,560 --> 00:25:41,560 Right. 806 00:25:41,560 --> 00:25:45,920 The research showed that these models, they can actually generalize their knowledge. 807 00:25:45,920 --> 00:25:46,920 Yeah. 808 00:25:46,920 --> 00:25:47,920 To new situations. 809 00:25:47,920 --> 00:25:49,080 To new and unfamiliar challenges. 810 00:25:49,080 --> 00:25:52,200 Unfamiliar challenges even in, in different languages. 811 00:25:52,200 --> 00:25:53,200 Different languages, yeah. 812 00:25:53,200 --> 00:25:54,720 Or with like encoded data. 813 00:25:54,720 --> 00:25:55,720 Right. 814 00:25:55,720 --> 00:25:56,720 That's, that's huge. 815 00:25:56,720 --> 00:26:01,080 It is huge, which, which means, you know, we're not just teaching them to follow the 816 00:26:01,080 --> 00:26:02,080 rules. 817 00:26:02,080 --> 00:26:03,080 Yeah. 818 00:26:03,080 --> 00:26:04,080 No. 819 00:26:04,080 --> 00:26:05,080 We're, we're teaching them to understand. 820 00:26:05,080 --> 00:26:06,080 The reasoning behind it. 821 00:26:06,080 --> 00:26:07,840 The reasoning behind those rules, which is, which is amazing. 822 00:26:07,840 --> 00:26:08,840 It is. 823 00:26:08,840 --> 00:26:10,760 And that's what makes this approach so, so promising. 824 00:26:10,760 --> 00:26:11,760 It does. 825 00:26:11,760 --> 00:26:14,400 You know, it suggests that we're not just building rule followers. 826 00:26:14,400 --> 00:26:15,400 Right. 827 00:26:15,400 --> 00:26:16,920 We're building AI that can adapt. 828 00:26:16,920 --> 00:26:17,920 Yeah. 829 00:26:17,920 --> 00:26:19,840 To new situations, new ethical dilemmas. 830 00:26:19,840 --> 00:26:21,720 So as, as we wrap up this deep dive. 831 00:26:21,720 --> 00:26:22,720 Yeah. 832 00:26:22,720 --> 00:26:25,240 I'm curious to hear, you know, your thoughts on the big picture. 833 00:26:25,240 --> 00:26:26,240 Okay. 834 00:26:26,240 --> 00:26:27,240 Yeah. 835 00:26:27,240 --> 00:26:29,240 Like where do you see this research going? 836 00:26:29,240 --> 00:26:30,240 What are some of the possibilities. 837 00:26:30,240 --> 00:26:31,240 Yeah. 838 00:26:31,240 --> 00:26:32,240 Yeah. 839 00:26:32,240 --> 00:26:33,960 That this, that this deliberative alignment opens up. 840 00:26:33,960 --> 00:26:38,800 Well, I think, you know, one really exciting possibility is, is the potential for AI. 841 00:26:38,800 --> 00:26:39,800 Yeah. 842 00:26:39,800 --> 00:26:43,280 To, to become a more active partner in ethical decision making. 843 00:26:43,280 --> 00:26:47,160 So, so instead of just, you know, passively following rules. 844 00:26:47,160 --> 00:26:48,160 Yeah. 845 00:26:48,160 --> 00:26:49,160 Right. 846 00:26:49,160 --> 00:26:50,680 AI could actually contribute to the discussion. 847 00:26:50,680 --> 00:26:51,680 Yeah. 848 00:26:51,680 --> 00:26:56,400 Contribute to ethical debates, help us navigate these, these really complex dilemmas. 849 00:26:56,400 --> 00:26:57,400 Wow. 850 00:26:57,400 --> 00:26:58,640 That's, that's a powerful thought. 851 00:26:58,640 --> 00:26:59,840 It is a powerful thought. 852 00:26:59,840 --> 00:27:03,280 You know, think about fields like, like bioethics. 853 00:27:03,280 --> 00:27:04,280 Yeah. 854 00:27:04,280 --> 00:27:07,200 Environmental policy, even like international relations. 855 00:27:07,200 --> 00:27:09,400 I mean, AI could, could really. 856 00:27:09,400 --> 00:27:11,600 It could really help us weigh those different perspectives. 857 00:27:11,600 --> 00:27:12,600 Right. 858 00:27:12,600 --> 00:27:18,640 Think about the consequences, you know, and, and arrive at more ethically sound decisions. 859 00:27:18,640 --> 00:27:19,640 That's incredible. 860 00:27:19,640 --> 00:27:21,720 But, but of course with any powerful technology. 861 00:27:21,720 --> 00:27:22,720 Of course. 862 00:27:22,720 --> 00:27:23,720 Of course. 863 00:27:23,720 --> 00:27:24,720 There are risks, right? 864 00:27:24,720 --> 00:27:25,720 There are risks. 865 00:27:25,720 --> 00:27:26,720 There are challenges. 866 00:27:26,720 --> 00:27:27,720 Yeah. 867 00:27:27,720 --> 00:27:28,720 For sure. 868 00:27:28,720 --> 00:27:29,720 That we need to be aware of. 869 00:27:29,720 --> 00:27:30,720 We need to be very mindful of those. 870 00:27:30,720 --> 00:27:34,880 The AI becomes, you know, more sophisticated, more integrated into our lives. 871 00:27:34,880 --> 00:27:35,880 Integrated. 872 00:27:35,880 --> 00:27:36,880 Yeah. 873 00:27:36,880 --> 00:27:40,960 And we have like good systems in place for accountability. 874 00:27:40,960 --> 00:27:41,960 Accountability. 875 00:27:41,960 --> 00:27:42,960 Over oversight, you know. 876 00:27:42,960 --> 00:27:43,960 Right. 877 00:27:43,960 --> 00:27:45,560 We can't just blindly trust these systems. 878 00:27:45,560 --> 00:27:46,560 No, no, we can't. 879 00:27:46,560 --> 00:27:49,880 We need to be able to like understand. 880 00:27:49,880 --> 00:27:50,880 How they're making decisions. 881 00:27:50,880 --> 00:27:51,880 Valid and making decisions. 882 00:27:51,880 --> 00:27:53,160 And, and hold them accountable. 883 00:27:53,160 --> 00:27:54,440 Accountable for their actions. 884 00:27:54,440 --> 00:27:55,440 Yeah. 885 00:27:55,440 --> 00:27:56,440 Transparency is key. 886 00:27:56,440 --> 00:27:57,680 Transparency is absolutely key. 887 00:27:57,680 --> 00:27:58,680 Right. 888 00:27:58,680 --> 00:28:00,520 Like we need to be able to see how. 889 00:28:00,520 --> 00:28:01,520 To look under the hood. 890 00:28:01,520 --> 00:28:02,520 Yeah. 891 00:28:02,520 --> 00:28:03,520 Look under the hood. 892 00:28:03,520 --> 00:28:05,040 See how it's, how it's getting to those conclusions. 893 00:28:05,040 --> 00:28:08,600 Especially when those conclusions have, you know, ethical implications. 894 00:28:08,600 --> 00:28:09,600 Yeah. 895 00:28:09,600 --> 00:28:10,600 Absolutely. 896 00:28:10,600 --> 00:28:15,880 And, and we also need to be aware of, you know, the potential for bias. 897 00:28:15,880 --> 00:28:16,880 Bias. 898 00:28:16,880 --> 00:28:17,880 Yes. 899 00:28:17,880 --> 00:28:18,880 That's a big one. 900 00:28:18,880 --> 00:28:19,880 Right. 901 00:28:19,880 --> 00:28:23,280 These, these systems, they're trained on data created by humans. 902 00:28:23,280 --> 00:28:24,280 Created by humans. 903 00:28:24,280 --> 00:28:28,640 And that data can reflect, you know, existing societal biases. 904 00:28:28,640 --> 00:28:30,840 And that can be, you know, really harmful. 905 00:28:30,840 --> 00:28:31,840 It can be. 906 00:28:31,840 --> 00:28:32,840 Yeah. 907 00:28:32,840 --> 00:28:37,040 And we need to be vigilant in making sure that these systems are fair. 908 00:28:37,040 --> 00:28:38,040 Fair. 909 00:28:38,040 --> 00:28:39,040 Equitable. 910 00:28:39,040 --> 00:28:40,040 Equitable. 911 00:28:40,040 --> 00:28:41,520 And they're not perpetuating harmful stereotypes. 912 00:28:41,520 --> 00:28:42,520 Absolutely. 913 00:28:42,520 --> 00:28:43,520 It's, it's a lot to think about. 914 00:28:43,520 --> 00:28:44,520 It is a lot to think about. 915 00:28:44,520 --> 00:28:45,520 Yeah. 916 00:28:45,520 --> 00:28:50,440 But it's, it's clear that AI safety, you know, it's not just a technical challenge. 917 00:28:50,440 --> 00:28:51,440 No, no, it's not. 918 00:28:51,440 --> 00:28:52,440 It's a human one. 919 00:28:52,440 --> 00:28:53,440 It's a deeply human challenge. 920 00:28:53,440 --> 00:28:54,440 It is, it is. 921 00:28:54,440 --> 00:28:56,040 And it requires, you know, collaboration. 922 00:28:56,040 --> 00:28:57,040 Collaboration. 923 00:28:57,040 --> 00:28:58,040 Careful consideration. 924 00:28:58,040 --> 00:28:59,040 Careful consideration. 925 00:28:59,040 --> 00:29:00,040 From all of us. 926 00:29:00,040 --> 00:29:02,080 You know, researchers, developers. 927 00:29:02,080 --> 00:29:03,760 The policymakers, the public, everybody. 928 00:29:03,760 --> 00:29:04,760 Everybody. 929 00:29:04,760 --> 00:29:05,760 Yeah. 930 00:29:05,760 --> 00:29:08,760 Well, this deep dive into deliberative alignment has been, it's been really eye-opening. 931 00:29:08,760 --> 00:29:09,760 It has. 932 00:29:09,760 --> 00:29:15,040 You know, it's given us this glimpse into a future where AI is not just, you know, powerful, 933 00:29:15,040 --> 00:29:16,040 but also. 934 00:29:16,040 --> 00:29:17,040 But also responsible. 935 00:29:17,040 --> 00:29:18,040 Responsible, ethical. 936 00:29:18,040 --> 00:29:19,320 Oh, ethical, yeah. 937 00:29:19,320 --> 00:29:21,880 And truly aligned with, with those human values. 938 00:29:21,880 --> 00:29:24,200 It's a, it's a future worth striving for. 939 00:29:24,200 --> 00:29:25,200 It is. 940 00:29:25,200 --> 00:29:26,200 It is. 941 00:29:26,200 --> 00:29:28,760 And it's one that, you know, as you said, we all have a role in shaping. 942 00:29:28,760 --> 00:29:29,760 We all do. 943 00:29:29,760 --> 00:29:32,800 Well, that brings us to the end of our deep dive today. 944 00:29:32,800 --> 00:29:33,800 It does. 945 00:29:33,800 --> 00:29:38,520 Thank you all so much for joining us on this, you know, expeleration of, of this really 946 00:29:38,520 --> 00:29:41,280 cutting edge AI safety research. 947 00:29:41,280 --> 00:29:42,280 Always a pleasure. 948 00:29:42,280 --> 00:29:45,720 We hope this episode has, you know, maybe sparked your curiosity. 949 00:29:45,720 --> 00:29:47,400 It sparked your curiosity. 950 00:29:47,400 --> 00:29:52,520 And got you thinking more deeply about these, these really important ethical implications 951 00:29:52,520 --> 00:29:53,520 of AI. 952 00:29:53,520 --> 00:29:54,520 Yeah. 953 00:29:54,520 --> 00:29:56,040 Because it's a conversation we all need to be a part of. 954 00:29:56,040 --> 00:29:57,040 It is. 955 00:29:57,040 --> 00:30:01,440 And always we encourage you to, to keep exploring these questions, you know. 956 00:30:01,440 --> 00:30:02,440 Speak, learn and keep thinking. 957 00:30:02,440 --> 00:30:06,360 Keep learning, keep thinking and, and stay engaged in this conversation. 958 00:30:06,360 --> 00:30:07,360 Stay engaged. 959 00:30:07,360 --> 00:30:08,360 Yeah. 960 00:30:08,360 --> 00:30:11,080 About the future that, that we want to build with AI. 961 00:30:11,080 --> 00:30:12,080 That's right. 962 00:30:12,080 --> 00:30:13,080 Until next time. 963 00:30:13,080 --> 00:30:36,080 Thanks a lot.