1 00:00:00,000 --> 00:00:00,840 Hey, everyone. 2 00:00:00,840 --> 00:00:02,960 Ready for another deep dive? 3 00:00:02,960 --> 00:00:05,920 This time, we're exploring Anthropic, 4 00:00:05,920 --> 00:00:09,640 a company making some serious waves in the AI world. 5 00:00:09,640 --> 00:00:11,480 Definitely a hot topic these days. 6 00:00:11,480 --> 00:00:12,520 You bet. 7 00:00:12,520 --> 00:00:15,120 And our listeners sent over a ton of great stuff. 8 00:00:15,120 --> 00:00:18,520 Articles, interviews with Anthropic's leaders, even 9 00:00:18,520 --> 00:00:20,120 some personal notes. 10 00:00:20,120 --> 00:00:22,400 Looks like they're especially interested in how 11 00:00:22,400 --> 00:00:24,720 Anthropic focuses on AI safety. 12 00:00:24,720 --> 00:00:26,480 Yeah, super interesting stuff, especially 13 00:00:26,480 --> 00:00:28,800 since they're developing some really advanced AI system. 14 00:00:28,800 --> 00:00:29,480 Exactly. 15 00:00:29,480 --> 00:00:32,520 So we're diving into what makes Anthropic tick, 16 00:00:32,520 --> 00:00:35,600 how they're building AI systems like Claude and their vision 17 00:00:35,600 --> 00:00:36,800 for the future. 18 00:00:36,800 --> 00:00:38,480 You're our AI guru. 19 00:00:38,480 --> 00:00:41,280 What stood out to you when you were looking through all this? 20 00:00:41,280 --> 00:00:43,080 Honestly, it's fascinating how they're 21 00:00:43,080 --> 00:00:45,920 all about pushing the limits of what AI can do. 22 00:00:45,920 --> 00:00:48,360 But at the same time, they're so committed 23 00:00:48,360 --> 00:00:50,600 to making sure it's beneficial and safe, 24 00:00:50,600 --> 00:00:51,960 you don't see that combo every day. 25 00:00:51,960 --> 00:00:53,760 It's almost like they want to win the race, 26 00:00:53,760 --> 00:00:55,760 but also make sure everyone crosses the finish 27 00:00:55,760 --> 00:00:57,000 line in one piece. 28 00:00:57,000 --> 00:00:59,160 OK, so let's unpack Anthropic. 29 00:00:59,160 --> 00:01:00,640 What makes them unique? 30 00:01:00,640 --> 00:01:02,600 How are they approaching AI development, 31 00:01:02,600 --> 00:01:04,680 especially with their Claude system? 32 00:01:04,680 --> 00:01:09,200 And what do they see on the horizon for AI in general? 33 00:01:09,200 --> 00:01:10,720 Well, one thing that really struck me 34 00:01:10,720 --> 00:01:13,000 is their work on scaling laws. 35 00:01:13,000 --> 00:01:14,120 Scaling laws. 36 00:01:14,120 --> 00:01:17,040 Basically, they've observed that if you increase 37 00:01:17,040 --> 00:01:21,640 the size of an AI model, like how much data it learns from, 38 00:01:21,640 --> 00:01:24,200 and you bump up the computing power used to train it, 39 00:01:24,200 --> 00:01:26,000 the AI usually performs better. 40 00:01:26,000 --> 00:01:27,920 So bigger is better in the AI world. 41 00:01:27,920 --> 00:01:30,640 Just keep throwing more data and computing power at it. 42 00:01:30,640 --> 00:01:31,560 Yeah, kind of. 43 00:01:31,560 --> 00:01:33,400 But it gets a little more complex than that. 44 00:01:33,400 --> 00:01:36,520 Anthropic CEO Dario Amodei is actually 45 00:01:36,520 --> 00:01:38,880 pretty vocal about the potential downsides. 46 00:01:38,880 --> 00:01:39,600 Like what? 47 00:01:39,600 --> 00:01:41,880 Well, for one thing, he's concerned about whether we'll 48 00:01:41,880 --> 00:01:45,280 even have enough good data to train these supersized models 49 00:01:45,280 --> 00:01:46,280 effectively. 50 00:01:46,280 --> 00:01:48,120 So it's not just about having tons of data. 51 00:01:48,120 --> 00:01:49,440 It's got to be good data. 52 00:01:49,440 --> 00:01:50,040 Right. 53 00:01:50,040 --> 00:01:52,360 Plus, there's the issue of how much our current computers can 54 00:01:52,360 --> 00:01:53,000 handle. 55 00:01:53,000 --> 00:01:55,500 And then there's the possibility that the way AI models are 56 00:01:55,500 --> 00:01:59,040 designed might need to change to handle even more scaling. 57 00:01:59,040 --> 00:02:03,720 So if just making AI bigger isn't always the answer, 58 00:02:03,720 --> 00:02:06,000 what other options are they looking into? 59 00:02:06,000 --> 00:02:09,360 How are they thinking about keeping AI moving forward? 60 00:02:09,360 --> 00:02:12,800 Well, one thing they're exploring is using synthetic data. 61 00:02:12,800 --> 00:02:14,120 Synthetic data? 62 00:02:14,120 --> 00:02:14,600 What's that? 63 00:02:14,600 --> 00:02:17,480 It's basically artificially generated data 64 00:02:17,480 --> 00:02:20,560 that they can use to supplement real world data. 65 00:02:20,560 --> 00:02:21,060 Interesting. 66 00:02:21,060 --> 00:02:23,000 So they're trying to create their own data 67 00:02:23,000 --> 00:02:24,960 to train these AIs. 68 00:02:24,960 --> 00:02:26,200 In a way, yeah. 69 00:02:26,200 --> 00:02:29,200 They're also trying to find ways to make AI training more 70 00:02:29,200 --> 00:02:31,920 efficient so they can get more bang for their buck 71 00:02:31,920 --> 00:02:34,640 with the data and computing power they already have. 72 00:02:34,640 --> 00:02:35,320 Makes sense. 73 00:02:35,320 --> 00:02:37,960 It sounds like they're trying to be really strategic with their AI 74 00:02:37,960 --> 00:02:40,040 development, not just throwing things at the wall 75 00:02:40,040 --> 00:02:41,000 and seeing what sticks. 76 00:02:41,000 --> 00:02:41,500 Absolutely. 77 00:02:41,500 --> 00:02:42,720 They're thinking long term. 78 00:02:42,720 --> 00:02:46,640 So how does all of this relate to how they're building Claude? 79 00:02:46,640 --> 00:02:49,000 Well, Claude is their flagship AI model. 80 00:02:49,000 --> 00:02:52,280 They've designed it to be super powerful and safe all 81 00:02:52,280 --> 00:02:53,000 at the same time. 82 00:02:53,000 --> 00:02:55,080 Didn't they release a few different versions of it? 83 00:02:55,080 --> 00:02:57,800 Yeah, there's Opus, Sonnet, and Haiku. 84 00:02:57,800 --> 00:03:00,240 Each one's got its own strengths and capabilities. 85 00:03:00,240 --> 00:03:02,520 And haven't they made some pretty impressive strides 86 00:03:02,520 --> 00:03:03,200 with Claude? 87 00:03:03,200 --> 00:03:05,680 I heard it can actually interact with computer screens now. 88 00:03:05,680 --> 00:03:06,440 Yeah, for sure. 89 00:03:06,440 --> 00:03:09,600 It can analyze screenshots, fill out spreadsheets, even 90 00:03:09,600 --> 00:03:10,600 write code. 91 00:03:10,600 --> 00:03:11,160 Wow. 92 00:03:11,160 --> 00:03:14,080 But the crazy part is they're doing all of this 93 00:03:14,080 --> 00:03:16,640 while still being incredibly careful 94 00:03:16,640 --> 00:03:18,260 about the potential risks. 95 00:03:18,260 --> 00:03:20,720 Which is where their AI safety levels come in, right? 96 00:03:20,720 --> 00:03:22,280 That ASL system they talk about. 97 00:03:22,280 --> 00:03:22,720 Exactly. 98 00:03:22,720 --> 00:03:26,120 They've got this framework to categorize their models based 99 00:03:26,120 --> 00:03:27,800 on the level of risk they pose. 100 00:03:27,800 --> 00:03:28,400 Yeah. 101 00:03:28,400 --> 00:03:30,640 Right now, Claude is considered ASL2. 102 00:03:30,640 --> 00:03:32,320 ASL2, meaning? 103 00:03:32,320 --> 00:03:34,320 It means that on its own, it's not 104 00:03:34,320 --> 00:03:36,080 capable of causing major harm. 105 00:03:36,080 --> 00:03:37,160 Gotcha. 106 00:03:37,160 --> 00:03:40,040 But as they keep pushing Claude to do more and more, 107 00:03:40,040 --> 00:03:42,560 are they worried about those safety levels going up? 108 00:03:42,560 --> 00:03:43,040 Definitely. 109 00:03:43,040 --> 00:03:45,500 They're already looking ahead to those higher ASL levels, 110 00:03:45,500 --> 00:03:48,640 like ASL3 and beyond, where AI could potentially 111 00:03:48,640 --> 00:03:51,080 get used for malicious purposes or even 112 00:03:51,080 --> 00:03:52,960 start doing its own research independently. 113 00:03:52,960 --> 00:03:55,280 AI doing its own research. 114 00:03:55,280 --> 00:03:56,240 That's kind of freaky. 115 00:03:56,240 --> 00:03:56,960 It is. 116 00:03:56,960 --> 00:03:58,280 And that's why they're staying ahead of the game 117 00:03:58,280 --> 00:03:59,520 when it comes to safety. 118 00:03:59,520 --> 00:04:02,320 They're trying to anticipate those risks before they even 119 00:04:02,320 --> 00:04:03,280 become a problem. 120 00:04:03,280 --> 00:04:05,640 So they're building these powerful AI models, 121 00:04:05,640 --> 00:04:08,760 thinking about the risks, but also trying to make sure 122 00:04:08,760 --> 00:04:12,480 these AIs reflect positive human values, right? 123 00:04:12,480 --> 00:04:14,360 Like it's not just about raw intelligence, 124 00:04:14,360 --> 00:04:17,560 but also about making them good, in a sense. 125 00:04:17,560 --> 00:04:18,120 Exactly. 126 00:04:18,120 --> 00:04:19,840 They want more than just a brain. 127 00:04:19,840 --> 00:04:21,640 They want a heart, too. 128 00:04:21,640 --> 00:04:23,960 And that's where someone like Amanda Askel comes in. 129 00:04:23,960 --> 00:04:24,920 Amanda Askel. 130 00:04:24,920 --> 00:04:26,800 She's a researcher at Anthropic, right? 131 00:04:26,800 --> 00:04:29,560 Yeah, she's been super involved in shaping Claude's 132 00:04:29,560 --> 00:04:30,920 personality. 133 00:04:30,920 --> 00:04:36,120 They want it to be a genuinely helpful and harmless AI, 134 00:04:36,120 --> 00:04:39,600 embodying qualities like honesty, humility, empathy. 135 00:04:39,600 --> 00:04:41,960 So they're not just building a brilliant AI. 136 00:04:41,960 --> 00:04:43,840 They're building a kind one, too. 137 00:04:43,840 --> 00:04:45,440 How are they actually doing that? 138 00:04:45,440 --> 00:04:48,040 How do you even build those qualities into an AI? 139 00:04:48,040 --> 00:04:51,160 Well, one of their key techniques is called constitutional AI. 140 00:04:51,160 --> 00:04:52,360 Constitutional AI. 141 00:04:52,360 --> 00:04:55,240 It's like giving the AI a set of guidelines, almost 142 00:04:55,240 --> 00:04:56,960 like a moral compass in its code. 143 00:04:56,960 --> 00:04:58,520 So they're giving it a sense of ethics, 144 00:04:58,520 --> 00:05:00,960 like we have laws and societal norms. 145 00:05:00,960 --> 00:05:01,640 Right. 146 00:05:01,640 --> 00:05:03,600 And they're also using something called reinforcement 147 00:05:03,600 --> 00:05:06,480 learning from human feedback, RLHF for short. 148 00:05:06,480 --> 00:05:10,680 Basically, they have humans give feedback on Claude's responses. 149 00:05:10,680 --> 00:05:13,600 And that helps it learn and improve over time. 150 00:05:13,600 --> 00:05:16,320 So it's like training a dog, but instead of treats, 151 00:05:16,320 --> 00:05:19,200 they're using feedback to guide its behavior. 152 00:05:19,200 --> 00:05:20,080 Very much. 153 00:05:20,080 --> 00:05:22,120 It sounds like they're really putting in the work 154 00:05:22,120 --> 00:05:24,720 to make sure Claude is both helpful and aligned 155 00:05:24,720 --> 00:05:26,320 with our values. 156 00:05:26,320 --> 00:05:28,920 What else caught your eye about Anthropix approach? 157 00:05:28,920 --> 00:05:31,000 Well, one thing that really stood out 158 00:05:31,000 --> 00:05:33,040 was their focus on something called 159 00:05:33,040 --> 00:05:35,480 mechanistic interpretability. 160 00:05:35,480 --> 00:05:36,520 Oh, that's a mouthful. 161 00:05:36,520 --> 00:05:37,600 Mechanistic. 162 00:05:37,600 --> 00:05:38,440 What was it again? 163 00:05:38,440 --> 00:05:39,800 Mechanistic interpretability. 164 00:05:39,800 --> 00:05:43,280 Basically, they're trying to understand how these AI models 165 00:05:43,280 --> 00:05:45,520 actually work, like at their core. 166 00:05:45,520 --> 00:05:47,320 OK, let's break that down. 167 00:05:47,320 --> 00:05:49,520 How do you even begin to understand 168 00:05:49,520 --> 00:05:53,360 what's going on inside these incredibly complex AI systems? 169 00:05:53,360 --> 00:05:54,840 It seems almost impossible. 170 00:05:54,840 --> 00:05:55,800 It is a huge challenge. 171 00:05:55,800 --> 00:05:57,840 But that's where someone like Crisola comes in. 172 00:05:57,840 --> 00:05:58,480 Crisola. 173 00:05:58,480 --> 00:05:59,920 He's another researcher in Anthropix 174 00:05:59,920 --> 00:06:03,160 who's leading the charge on this interpretability front. 175 00:06:03,160 --> 00:06:05,160 He's got this interesting way of thinking about it 176 00:06:05,160 --> 00:06:08,440 where he compares AI development to neurobiology. 177 00:06:08,440 --> 00:06:09,800 AI and brains. 178 00:06:09,800 --> 00:06:11,640 Yeah, like they're trying to understand 179 00:06:11,640 --> 00:06:16,200 how these AI models develop their own circuits and connections 180 00:06:16,200 --> 00:06:19,320 based on all the data they're fed and how they're trained. 181 00:06:19,320 --> 00:06:21,720 So it's like they're studying the AI's brain, 182 00:06:21,720 --> 00:06:24,480 trying to map out its thoughts, so to speak. 183 00:06:24,480 --> 00:06:26,200 Why is that so important to them? 184 00:06:26,200 --> 00:06:29,120 Because if they can understand how these models work 185 00:06:29,120 --> 00:06:31,480 at a fundamental level, they can better 186 00:06:31,480 --> 00:06:33,000 predict how they'll behave. 187 00:06:33,000 --> 00:06:33,640 Makes sense. 188 00:06:33,640 --> 00:06:36,160 They want to identify any potential risks 189 00:06:36,160 --> 00:06:37,880 before they become problems, right? 190 00:06:37,880 --> 00:06:38,160 Right. 191 00:06:38,160 --> 00:06:39,960 They're not just building a powerful tool. 192 00:06:39,960 --> 00:06:42,480 They want to make sure they understand it inside and out 193 00:06:42,480 --> 00:06:44,480 so they can use it safely and responsibly. 194 00:06:44,480 --> 00:06:46,080 So how are they actually doing that? 195 00:06:46,080 --> 00:06:49,080 What are they looking for when they peer inside these AI 196 00:06:49,080 --> 00:06:50,280 systems? 197 00:06:50,280 --> 00:06:51,780 Well, one thing they're looking for 198 00:06:51,780 --> 00:06:54,640 is any sign of what they call deception or back 199 00:06:54,640 --> 00:06:56,080 doors within the model. 200 00:06:56,080 --> 00:06:57,360 Oh, deception. 201 00:06:57,360 --> 00:06:58,840 Like the AI is trying to trick us. 202 00:06:58,840 --> 00:07:01,240 It's not that they think the AI is consciously 203 00:07:01,240 --> 00:07:03,680 trying to be sneaky, but it's more 204 00:07:03,680 --> 00:07:06,440 that as AI gets smarter and smarter, 205 00:07:06,440 --> 00:07:09,400 there's a chance it could learn to explode loopholes 206 00:07:09,400 --> 00:07:13,720 or manipulate its environment in ways we didn't see coming. 207 00:07:13,720 --> 00:07:16,040 So it's not necessarily that it's trying to be malicious, 208 00:07:16,040 --> 00:07:18,880 but more like it might accidentally cause harm 209 00:07:18,880 --> 00:07:19,720 if we're not careful. 210 00:07:19,720 --> 00:07:20,560 Right, exactly. 211 00:07:20,560 --> 00:07:22,600 And that's where all this mechanistic interpretability 212 00:07:22,600 --> 00:07:23,480 research comes in. 213 00:07:23,480 --> 00:07:26,400 It gives it a way to make sure the AI is behaving 214 00:07:26,400 --> 00:07:29,120 how it's supposed to, even as it gets more and more complex. 215 00:07:29,120 --> 00:07:32,200 It's like they're installing a security camera inside the AI's 216 00:07:32,200 --> 00:07:34,480 brain so they can keep a close eye on things. 217 00:07:34,480 --> 00:07:35,640 That's a good analogy. 218 00:07:35,640 --> 00:07:39,760 But beyond just safety, what's their overall vision 219 00:07:39,760 --> 00:07:41,000 for the future of AI? 220 00:07:41,000 --> 00:07:42,880 Where do they see all this going? 221 00:07:42,880 --> 00:07:45,280 Well, Amadeus actually said he believes 222 00:07:45,280 --> 00:07:48,120 AI has the potential to solve some of humanity's biggest 223 00:07:48,120 --> 00:07:48,680 problems. 224 00:07:48,680 --> 00:07:49,480 Like what? 225 00:07:49,480 --> 00:07:53,240 Like climate change, disease, poverty, all these huge issues. 226 00:07:53,240 --> 00:07:56,000 AI could be a powerful tool for tackling those things. 227 00:07:56,000 --> 00:07:56,720 Wow. 228 00:07:56,720 --> 00:08:00,280 So it's like AI could be this incredible force for good, 229 00:08:00,280 --> 00:08:02,360 helping us create a better world for everyone. 230 00:08:02,360 --> 00:08:03,560 That's the idea. 231 00:08:03,560 --> 00:08:06,200 But he's not naive about the potential downsides either. 232 00:08:06,200 --> 00:08:08,120 Right, it's not going to be all sunshine and roses. 233 00:08:08,120 --> 00:08:08,880 Exactly. 234 00:08:08,880 --> 00:08:11,080 He's very realistic about the risks 235 00:08:11,080 --> 00:08:14,360 and emphasizes the need to be careful and thoughtful 236 00:08:14,360 --> 00:08:15,600 every step of the way. 237 00:08:15,600 --> 00:08:17,240 So it's a real balancing act trying 238 00:08:17,240 --> 00:08:20,720 to unlock all the amazing potential of AI 239 00:08:20,720 --> 00:08:22,240 while making sure we don't create 240 00:08:22,240 --> 00:08:23,680 something we can't control. 241 00:08:23,680 --> 00:08:25,280 Definitely a huge responsibility. 242 00:08:25,280 --> 00:08:26,760 And it's not just the responsibility 243 00:08:26,760 --> 00:08:28,840 of the tech companies building this stuff, right? 244 00:08:28,840 --> 00:08:30,560 It's got to involve everyone. 245 00:08:30,560 --> 00:08:34,480 100% Anthropic is actually really active in discussions 246 00:08:34,480 --> 00:08:38,720 with policymakers and ethicists to help shape the future of AI 247 00:08:38,720 --> 00:08:39,920 in a responsible way. 248 00:08:39,920 --> 00:08:43,040 It's a much bigger conversation than just the code itself. 249 00:08:43,040 --> 00:08:44,640 Oh, yeah, absolutely. 250 00:08:44,640 --> 00:08:46,760 They believe that AI needs to be guided 251 00:08:46,760 --> 00:08:49,440 by a whole range of perspectives to get it right. 252 00:08:49,440 --> 00:08:51,120 So they're not just building technology, 253 00:08:51,120 --> 00:08:53,040 they're trying to build a better future for all of us. 254 00:08:53,040 --> 00:08:53,840 Exactly. 255 00:08:53,840 --> 00:08:56,280 And it's inspiring to see a company taking that so 256 00:08:56,280 --> 00:08:57,000 seriously. 257 00:08:57,000 --> 00:08:57,960 It really is. 258 00:08:57,960 --> 00:08:59,880 OK, so big question time. 259 00:08:59,880 --> 00:09:02,920 What do you think the future of AI actually looks like? 260 00:09:02,920 --> 00:09:07,400 Are we all going to have robot butlers and flying cars? 261 00:09:07,400 --> 00:09:09,840 Well, the future is always a bit of a mystery, isn't it? 262 00:09:09,840 --> 00:09:11,160 But I think it's pretty safe to say 263 00:09:11,160 --> 00:09:14,160 that AI is only going to become more integrated into our lives. 264 00:09:14,160 --> 00:09:16,120 I mean, it already is in a lot of ways. 265 00:09:16,120 --> 00:09:16,400 Right. 266 00:09:16,400 --> 00:09:18,600 It's everywhere, helping us with all sorts of things, 267 00:09:18,600 --> 00:09:21,360 choosing movies, navigating traffic. 268 00:09:21,360 --> 00:09:23,680 And as these models get even more advanced, 269 00:09:23,680 --> 00:09:26,280 they'll probably start playing even bigger roles in areas 270 00:09:26,280 --> 00:09:29,480 like health care, education, transportation, maybe even 271 00:09:29,480 --> 00:09:30,920 art and entertainment. 272 00:09:30,920 --> 00:09:32,800 So instead of robot butlers, maybe 273 00:09:32,800 --> 00:09:36,960 we'll have AI doctors and teachers and artists. 274 00:09:36,960 --> 00:09:38,000 It's possible. 275 00:09:38,000 --> 00:09:43,160 And if companies like Anthropic are successful in their mission, 276 00:09:43,160 --> 00:09:46,160 this future will be built on a foundation of trust, 277 00:09:46,160 --> 00:09:48,360 transparency, shared values, all that good stuff. 278 00:09:48,360 --> 00:09:50,160 So it's not just about AI getting smarter, 279 00:09:50,160 --> 00:09:51,760 it's about making sure it gets wiser too. 280 00:09:51,760 --> 00:09:52,280 Exactly. 281 00:09:52,280 --> 00:09:55,440 AI needs to develop not just intellectually, but also 282 00:09:55,440 --> 00:09:55,960 ethically. 283 00:09:55,960 --> 00:09:57,480 And that's a challenge that Anthropic 284 00:09:57,480 --> 00:09:59,200 seems to be taking head on. 285 00:09:59,200 --> 00:10:01,280 It's exciting to think about all the possibilities, 286 00:10:01,280 --> 00:10:02,840 but also a little daunting. 287 00:10:02,840 --> 00:10:05,240 The impact AI could have on our world is huge. 288 00:10:05,240 --> 00:10:05,640 It is. 289 00:10:05,640 --> 00:10:08,480 And it's encouraging to see a company like Anthropic really 290 00:10:08,480 --> 00:10:09,840 wrestling with those big questions, 291 00:10:09,840 --> 00:10:11,440 trying to do things the right way. 292 00:10:11,440 --> 00:10:12,880 They're definitely one to watch. 293 00:10:12,880 --> 00:10:15,560 They're pushing the boundaries while also setting a high bar 294 00:10:15,560 --> 00:10:17,160 for responsible development. 295 00:10:17,160 --> 00:10:17,600 Yeah. 296 00:10:17,600 --> 00:10:19,360 This deep dive has been wild. 297 00:10:19,360 --> 00:10:23,400 We've covered so much scaling laws, AI safety levels, 298 00:10:23,400 --> 00:10:25,320 mechanistic interpretability. 299 00:10:25,320 --> 00:10:27,600 And got to give a shout out to our listener 300 00:10:27,600 --> 00:10:30,480 for sending over such great material to work with. 301 00:10:30,480 --> 00:10:32,360 It was a fantastic selection of sources. 302 00:10:32,360 --> 00:10:34,520 It really showed their interest in not just 303 00:10:34,520 --> 00:10:37,080 the technical side of AI, but also 304 00:10:37,080 --> 00:10:39,840 the ethical and societal implications. 305 00:10:39,840 --> 00:10:41,160 Absolutely. 306 00:10:41,160 --> 00:10:41,480 OK. 307 00:10:41,480 --> 00:10:43,120 So let's do a quick recap of what we've 308 00:10:43,120 --> 00:10:44,280 learned about Anthropic. 309 00:10:44,280 --> 00:10:46,000 So far we know they're all about pushing 310 00:10:46,000 --> 00:10:47,880 the limits of what AI can do. 311 00:10:47,880 --> 00:10:49,680 But they're doing it with a strong emphasis 312 00:10:49,680 --> 00:10:51,560 on safety and ethics. 313 00:10:51,560 --> 00:10:54,200 It's like they're writing a new playbook for AI development, 314 00:10:54,200 --> 00:10:57,240 where progress and responsibility go hand in hand. 315 00:10:57,240 --> 00:10:58,760 What else would you highlight? 316 00:10:58,760 --> 00:11:00,840 I'd say their approach to actually building these AI 317 00:11:00,840 --> 00:11:02,160 models is really interesting. 318 00:11:02,160 --> 00:11:04,120 They're being very strategic. 319 00:11:04,120 --> 00:11:07,120 Not just throwing data and computing power at the problem. 320 00:11:07,120 --> 00:11:07,360 Right. 321 00:11:07,360 --> 00:11:08,720 They're thinking outside the box, 322 00:11:08,720 --> 00:11:11,840 exploring things like synthetic data, new training techniques. 323 00:11:11,840 --> 00:11:12,400 Exactly. 324 00:11:12,400 --> 00:11:15,440 Trying to overcome those limitations of just scaling up. 325 00:11:15,440 --> 00:11:18,240 And they're not shying away from the really tough questions. 326 00:11:18,240 --> 00:11:22,520 How do you make sure AI actually reflects positive human values? 327 00:11:22,520 --> 00:11:24,800 All that work they're doing with Claude's personality 328 00:11:24,800 --> 00:11:26,160 is a perfect example. 329 00:11:26,160 --> 00:11:30,040 Yeah, they're using techniques like constitutional AI and RLHF 330 00:11:30,040 --> 00:11:32,800 to guide Claude's development, making sure it's not just 331 00:11:32,800 --> 00:11:36,120 brilliant, but also kind and helpful. 332 00:11:36,120 --> 00:11:39,080 Like they're raising a well-rounded AI citizen. 333 00:11:39,080 --> 00:11:39,800 Exactly. 334 00:11:39,800 --> 00:11:43,360 And then there's all their work on mechanistic interpretability, 335 00:11:43,360 --> 00:11:45,000 which is honestly mind blowing. 336 00:11:45,000 --> 00:11:47,840 They're literally trying to figure out how these AIs think. 337 00:11:47,840 --> 00:11:50,960 It's like they're cracking the code of artificial intelligence. 338 00:11:50,960 --> 00:11:53,160 And they're not doing it just out of curiosity. 339 00:11:53,160 --> 00:11:55,800 They really believe it's essential for building safe 340 00:11:55,800 --> 00:11:57,160 and trustworthy AI. 341 00:11:57,160 --> 00:11:59,400 It's like they're saying, look, we're not just 342 00:11:59,400 --> 00:12:02,680 going to build this powerful technology and hope for the best. 343 00:12:02,680 --> 00:12:04,480 We're going to understand it inside and out 344 00:12:04,480 --> 00:12:06,080 so we can use it responsibly. 345 00:12:06,080 --> 00:12:07,040 Exactly. 346 00:12:07,040 --> 00:12:09,840 And that commitment to transparency and understanding 347 00:12:09,840 --> 00:12:13,360 is so important, especially in a field that can feel 348 00:12:13,360 --> 00:12:14,880 very mysterious and secretive. 349 00:12:14,880 --> 00:12:17,560 It's like they're throwing open the doors and saying, come on in. 350 00:12:17,560 --> 00:12:18,640 Let's see how this all works. 351 00:12:18,640 --> 00:12:19,140 Yeah. 352 00:12:19,140 --> 00:12:22,760 And that openness is key for building trust with the public. 353 00:12:22,760 --> 00:12:24,340 People need to see what's going on, 354 00:12:24,340 --> 00:12:26,680 if they're going to feel comfortable with AI becoming 355 00:12:26,680 --> 00:12:28,320 more integrated into our lives. 356 00:12:28,320 --> 00:12:31,040 So they're pushing the boundaries of AI, 357 00:12:31,040 --> 00:12:33,600 thinking deeply about safety and ethics 358 00:12:33,600 --> 00:12:35,920 and being open about their process. 359 00:12:35,920 --> 00:12:37,300 It really seems like they're trying 360 00:12:37,300 --> 00:12:39,800 to change the game when it comes to AI development. 361 00:12:39,800 --> 00:12:40,640 I think they are. 362 00:12:40,640 --> 00:12:42,280 What you think, are they succeeding? 363 00:12:42,280 --> 00:12:44,440 It's still early days, but they're 364 00:12:44,440 --> 00:12:46,300 definitely making progress. 365 00:12:46,300 --> 00:12:50,080 They're raising the bar for both innovation and responsibility 366 00:12:50,080 --> 00:12:51,220 in the AI world. 367 00:12:51,220 --> 00:12:53,200 And they're asking the tough questions 368 00:12:53,200 --> 00:12:55,280 that other companies seem to be avoiding. 369 00:12:55,280 --> 00:12:58,840 Questions like, what does it even mean to build good AI? 370 00:12:58,840 --> 00:13:01,000 How do we make sure everyone benefits and not just 371 00:13:01,000 --> 00:13:02,040 a select few? 372 00:13:02,040 --> 00:13:05,400 How do we prevent misuse and unintended consequences? 373 00:13:05,400 --> 00:13:07,440 Those are big questions that need to be addressed. 374 00:13:07,440 --> 00:13:08,740 They're not just building technology. 375 00:13:08,740 --> 00:13:10,240 They're trying to build a better world. 376 00:13:10,240 --> 00:13:11,360 And that's something I get behind. 377 00:13:11,360 --> 00:13:12,040 Me too. 378 00:13:12,040 --> 00:13:13,880 So listener, if you're interested in AI, 379 00:13:13,880 --> 00:13:16,600 Anthropic is definitely a company to keep your eye on. 380 00:13:16,600 --> 00:13:18,160 They're showing us what's possible 381 00:13:18,160 --> 00:13:20,320 when you combine cutting edge technology 382 00:13:20,320 --> 00:13:22,040 with a strong moral compass. 383 00:13:22,040 --> 00:13:24,200 And they're reminding us that the future of AI 384 00:13:24,200 --> 00:13:26,360 isn't some predetermined thing. 385 00:13:26,360 --> 00:13:28,880 It's something we're all creating together. 386 00:13:28,880 --> 00:13:30,840 It's really something else how much they're focusing 387 00:13:30,840 --> 00:13:32,720 on the ethical side of things. 388 00:13:32,720 --> 00:13:33,920 It's not just lip service. 389 00:13:33,920 --> 00:13:35,960 They're really putting their resources 390 00:13:35,960 --> 00:13:38,840 into this whole mechanistic interpretability thing. 391 00:13:38,840 --> 00:13:41,080 Yeah, because it's one thing to know that an AI can 392 00:13:41,080 --> 00:13:42,760 do something amazing. 393 00:13:42,760 --> 00:13:44,840 But if we're going to really trust these systems, 394 00:13:44,840 --> 00:13:47,540 especially with important tasks, we 395 00:13:47,540 --> 00:13:49,160 got to understand how they do it. 396 00:13:49,160 --> 00:13:50,540 Exactly. 397 00:13:50,540 --> 00:13:52,420 It's like, would you get in a self-driving car 398 00:13:52,420 --> 00:13:55,040 if you had no clue how it was making decisions? 399 00:13:55,040 --> 00:13:55,800 Probably not. 400 00:13:55,800 --> 00:13:57,200 Not a chance. 401 00:13:57,200 --> 00:13:59,680 And this is where all that talk about AI safety levels 402 00:13:59,680 --> 00:14:01,200 really hits home, right? 403 00:14:01,200 --> 00:14:03,640 As these AI models get more and more powerful, 404 00:14:03,640 --> 00:14:05,960 they could pose some serious risks, 405 00:14:05,960 --> 00:14:07,720 even if they're not trying to be malicious. 406 00:14:07,720 --> 00:14:07,960 Right. 407 00:14:07,960 --> 00:14:09,300 It's not about them being evil. 408 00:14:09,300 --> 00:14:11,420 It's about unintended consequences. 409 00:14:11,420 --> 00:14:13,560 We talked about Claude being at ASL too, 410 00:14:13,560 --> 00:14:16,040 but they're already thinking about those higher levels 411 00:14:16,040 --> 00:14:18,480 where things could get a lot more complicated. 412 00:14:18,480 --> 00:14:19,560 They are. 413 00:14:19,560 --> 00:14:22,040 They're playing the long game, trying to anticipate problems 414 00:14:22,040 --> 00:14:23,840 before they even pop up. 415 00:14:23,840 --> 00:14:24,800 It's impressive. 416 00:14:24,800 --> 00:14:27,520 So this research into mechanistic interpretability, 417 00:14:27,520 --> 00:14:29,040 it's like they're developing a way 418 00:14:29,040 --> 00:14:33,500 to see inside these AI models, maybe even spot those risks 419 00:14:33,500 --> 00:14:34,960 before they become a reality. 420 00:14:34,960 --> 00:14:35,440 Exactly. 421 00:14:35,440 --> 00:14:37,520 It's like having a safety check built right in. 422 00:14:37,520 --> 00:14:40,560 So by understanding how these AIs think, 423 00:14:40,560 --> 00:14:44,200 they can potentially see those red flags, those little hints 424 00:14:44,200 --> 00:14:46,080 that something might go wrong. 425 00:14:46,080 --> 00:14:46,760 Yep. 426 00:14:46,760 --> 00:14:49,320 They're looking for anything out of the ordinary. 427 00:14:49,320 --> 00:14:52,280 Any signs of what they call deception or back doors. 428 00:14:52,280 --> 00:14:53,440 We talked about that before. 429 00:14:53,440 --> 00:14:55,560 But remind me again, what do they mean by deception? 430 00:14:55,560 --> 00:14:57,640 It's not like the AI is intentionally lying to us, 431 00:14:57,640 --> 00:14:58,120 right? 432 00:14:58,120 --> 00:14:58,760 Right. 433 00:14:58,760 --> 00:15:00,640 It's not about malicious intent. 434 00:15:00,640 --> 00:15:02,760 It's more about the possibility that the AI could 435 00:15:02,760 --> 00:15:05,280 learn to manipulate its environment in ways 436 00:15:05,280 --> 00:15:08,800 that we didn't expect, even if it's not trying to be sneaky. 437 00:15:08,800 --> 00:15:11,520 So more like unintended consequences. 438 00:15:11,520 --> 00:15:12,720 The AI is not being bad. 439 00:15:12,720 --> 00:15:15,520 It's just maybe figuring out how to achieve its goals in ways 440 00:15:15,520 --> 00:15:16,880 that could cause problems. 441 00:15:16,880 --> 00:15:17,760 Exactly. 442 00:15:17,760 --> 00:15:19,920 And that's precisely why they think 443 00:15:19,920 --> 00:15:21,840 this mechanistic interpretability stuff is 444 00:15:21,840 --> 00:15:22,800 so important. 445 00:15:22,800 --> 00:15:24,520 It gives them a way to look under the hood 446 00:15:24,520 --> 00:15:26,720 and make sure everything's running smoothly, 447 00:15:26,720 --> 00:15:29,320 even as the AI gets more and more advanced. 448 00:15:29,320 --> 00:15:32,680 It's like they're saying, OK, we trust you, AI, 449 00:15:32,680 --> 00:15:36,320 but we're also going to double check your work just to be safe. 450 00:15:36,320 --> 00:15:37,240 Exactly. 451 00:15:37,240 --> 00:15:40,520 It's about finding that balance between pushing the limits 452 00:15:40,520 --> 00:15:42,520 and being cautious. 453 00:15:42,520 --> 00:15:44,240 We want to see what's possible, but we also 454 00:15:44,240 --> 00:15:45,840 want to make sure we're doing it responsibly. 455 00:15:45,840 --> 00:15:48,240 And Anthropic seems to be walking that line pretty well. 456 00:15:48,240 --> 00:15:49,560 They really do. 457 00:15:49,560 --> 00:15:54,000 OK, but zooming out a bit, what about their big vision for AI? 458 00:15:54,000 --> 00:15:55,960 Where do they see all of this heading, 459 00:15:55,960 --> 00:15:58,560 and how does Anthropic fit into that picture? 460 00:15:58,560 --> 00:15:59,800 What's the end game here? 461 00:15:59,800 --> 00:16:01,640 Well, Amadeus has talked about a future 462 00:16:01,640 --> 00:16:04,600 where AI could help us tackle some of the biggest challenges 463 00:16:04,600 --> 00:16:05,720 we face as a species. 464 00:16:05,720 --> 00:16:06,440 No kidding. 465 00:16:06,440 --> 00:16:09,400 Yeah, like climate change, disease poverty, things 466 00:16:09,400 --> 00:16:11,400 that have plagued us for centuries. 467 00:16:11,400 --> 00:16:13,320 He thinks AI could be a game changer. 468 00:16:13,320 --> 00:16:15,240 So it's not just about building cool tech 469 00:16:15,240 --> 00:16:16,400 for the sake of it. 470 00:16:16,400 --> 00:16:19,240 It's about using that tech to actually make a difference. 471 00:16:19,240 --> 00:16:20,400 Exactly. 472 00:16:20,400 --> 00:16:23,440 They see AI as a way to boost our own capabilities, 473 00:16:23,440 --> 00:16:26,840 help us solve problems that have seemed impossible for so long. 474 00:16:26,840 --> 00:16:29,040 That's a pretty optimistic outlook. 475 00:16:29,040 --> 00:16:31,960 But I'm sure they're not blind to the potential downsides. 476 00:16:31,960 --> 00:16:33,000 Of course not. 477 00:16:33,000 --> 00:16:35,160 They know there are risks, and they're 478 00:16:35,160 --> 00:16:38,120 working hard to figure out how to avoid them. 479 00:16:38,120 --> 00:16:41,960 That's why their focus on safety and ethics is so crucial. 480 00:16:41,960 --> 00:16:45,040 They want to make sure that as AI gets more powerful, 481 00:16:45,040 --> 00:16:47,400 it stays on the side of good. 482 00:16:47,400 --> 00:16:49,320 It's like they're pioneers charting a course 483 00:16:49,320 --> 00:16:51,200 through uncharted territory, trying 484 00:16:51,200 --> 00:16:53,400 to steer clear of the dangers while still keeping 485 00:16:53,400 --> 00:16:54,760 their eyes on the prize. 486 00:16:54,760 --> 00:16:57,160 And they're doing it in a way that feels very thoughtful 487 00:16:57,160 --> 00:16:59,080 and open. 488 00:16:59,080 --> 00:17:00,720 They're not just working in isolation. 489 00:17:00,720 --> 00:17:03,360 They're talking to experts outside the tech world, 490 00:17:03,360 --> 00:17:05,920 collaborating with policymakers and ethicists 491 00:17:05,920 --> 00:17:07,920 to make sure AI is developed and used 492 00:17:07,920 --> 00:17:09,480 in a way that benefits everyone. 493 00:17:09,480 --> 00:17:10,400 That's a big deal. 494 00:17:10,400 --> 00:17:11,520 It's not just about profits. 495 00:17:11,520 --> 00:17:13,080 It's about making the world a better place. 496 00:17:13,080 --> 00:17:13,580 Right. 497 00:17:13,580 --> 00:17:15,400 It's about thinking about the big picture. 498 00:17:15,400 --> 00:17:17,960 And that's what I find so inspiring about Anthropic. 499 00:17:17,960 --> 00:17:19,920 Couldn't have said it better myself. 500 00:17:19,920 --> 00:17:22,040 Well, this has been an incredible deep dive 501 00:17:22,040 --> 00:17:23,160 into Anthropic. 502 00:17:23,160 --> 00:17:26,320 We've learned so much about their commitment to safety, 503 00:17:26,320 --> 00:17:29,120 their groundbreaking research into mechanistic 504 00:17:29,120 --> 00:17:32,040 interpretability, and their vision for a future 505 00:17:32,040 --> 00:17:33,920 where AI is a force for good. 506 00:17:33,920 --> 00:17:35,720 They're definitely pushing the boundaries, 507 00:17:35,720 --> 00:17:37,560 while also staying true to their values. 508 00:17:37,560 --> 00:17:39,840 And it's clear they're not afraid to ask 509 00:17:39,840 --> 00:17:42,800 the tough questions, to really grapple with what 510 00:17:42,800 --> 00:17:46,400 it means to build AI that is both powerful and ethical. 511 00:17:46,400 --> 00:17:48,480 They're a company that's worth keeping an eye on, 512 00:17:48,480 --> 00:17:49,200 that's for sure. 513 00:17:49,200 --> 00:17:50,240 Absolutely. 514 00:17:50,240 --> 00:17:52,800 So listener, if you're interested in the future of AI, 515 00:17:52,800 --> 00:17:54,960 remember what we've learned from Anthropic. 516 00:17:54,960 --> 00:17:59,360 Be curious, be critical, and most importantly, be engaged. 517 00:17:59,360 --> 00:18:01,640 The future of AI isn't some predetermined thing. 518 00:18:01,640 --> 00:18:03,640 It's something we're all creating together. 519 00:18:03,640 --> 00:18:05,020 And the choices we make today will 520 00:18:05,020 --> 00:18:06,320 shape the world of tomorrow. 521 00:18:06,320 --> 00:18:07,860 That's a great note to end on. 522 00:18:07,860 --> 00:18:12,960 Thanks for joining us on this deep dive into Anthropic.