1 00:00:00,000 --> 00:00:02,400 Hey, everyone, and welcome back. 2 00:00:02,400 --> 00:00:05,840 It's wild to see just how fast the world of AI is moving. 3 00:00:05,840 --> 00:00:06,920 Yeah, it really is. 4 00:00:06,920 --> 00:00:08,920 It seems like there are new models every week 5 00:00:08,920 --> 00:00:13,760 and new announcements from companies like Open AI and Google, 6 00:00:13,760 --> 00:00:16,160 and even governments are getting involved now. 7 00:00:16,160 --> 00:00:17,680 It's fascinating to see. 8 00:00:17,680 --> 00:00:19,960 So are you ready to unpack all of this with me? 9 00:00:19,960 --> 00:00:20,840 Absolutely, I am. 10 00:00:20,840 --> 00:00:21,720 OK, great. 11 00:00:21,720 --> 00:00:25,440 Well, I guess a good place to start would be Open AI. 12 00:00:25,440 --> 00:00:28,480 They just released their newest model, O1, 13 00:00:28,480 --> 00:00:30,600 and it's causing quite a stir. 14 00:00:30,600 --> 00:00:31,200 It is. 15 00:00:31,200 --> 00:00:33,560 They're claiming it's a huge step forward in terms 16 00:00:33,560 --> 00:00:37,240 of speed and accuracy and what it can actually do. 17 00:00:37,240 --> 00:00:40,160 Yeah, it's designed for those really complex real world 18 00:00:40,160 --> 00:00:44,200 tasks, which is a key shift from those previous models. 19 00:00:44,200 --> 00:00:45,360 Yeah, that's right. 20 00:00:45,360 --> 00:00:47,680 And they've added features like function calling, 21 00:00:47,680 --> 00:00:50,080 structured outputs, and even the ability 22 00:00:50,080 --> 00:00:51,240 to work with images. 23 00:00:51,240 --> 00:00:52,560 Oh, wow. 24 00:00:52,560 --> 00:00:54,120 So it sounds like they're really trying 25 00:00:54,120 --> 00:00:57,560 to make AI way more practical, moving beyond just 26 00:00:57,560 --> 00:01:00,000 like text generation and into things 27 00:01:00,000 --> 00:01:01,120 that we can use every day. 28 00:01:01,120 --> 00:01:01,480 Exactly. 29 00:01:01,480 --> 00:01:02,400 Yeah. 30 00:01:02,400 --> 00:01:04,640 I'm especially interested in those vision capabilities. 31 00:01:04,640 --> 00:01:05,960 Yeah, think about the possibilities 32 00:01:05,960 --> 00:01:10,240 like analyzing images for medical diagnoses, 33 00:01:10,240 --> 00:01:13,320 assisting with design tasks, even understanding 34 00:01:13,320 --> 00:01:15,000 the visual elements in code. 35 00:01:15,000 --> 00:01:15,680 OK. 36 00:01:15,680 --> 00:01:19,120 It opens up a whole new level of AI applications. 37 00:01:19,120 --> 00:01:21,360 Yeah, that's mind blowing. 38 00:01:21,360 --> 00:01:24,280 But while everyone is kind of buzzing about Open AI's 39 00:01:24,280 --> 00:01:28,560 video generation model, Sora, there's no API access yet. 40 00:01:28,560 --> 00:01:29,320 Yeah, that's right. 41 00:01:29,320 --> 00:01:31,600 And this puts Open AI a little bit 42 00:01:31,600 --> 00:01:35,200 behind competitors like Google and AWS, 43 00:01:35,200 --> 00:01:37,240 who are already letting developers kind of play 44 00:01:37,240 --> 00:01:39,160 with their video AI tools. 45 00:01:39,160 --> 00:01:42,200 It seems like Open AI is being extra cautious with Sora. 46 00:01:42,200 --> 00:01:42,880 OK. 47 00:01:42,880 --> 00:01:45,200 You know, generating realistic videos obviously 48 00:01:45,200 --> 00:01:48,360 raises a lot of questions about responsible use. 49 00:01:48,360 --> 00:01:49,000 Sure. 50 00:01:49,000 --> 00:01:51,000 But it could also mean that Open AI just 51 00:01:51,000 --> 00:01:53,680 doesn't have the capacity yet to support really 52 00:01:53,680 --> 00:01:54,840 widespread access. 53 00:01:54,840 --> 00:01:55,400 Interesting. 54 00:01:55,400 --> 00:01:57,760 So we might have to wait a little bit to get our hands on Sora. 55 00:01:57,760 --> 00:01:58,880 Yeah, it seems like it. 56 00:01:58,880 --> 00:02:00,800 But in the meantime, the US government 57 00:02:00,800 --> 00:02:02,680 is stepping into the AI ring. 58 00:02:02,680 --> 00:02:03,320 They are. 59 00:02:03,320 --> 00:02:06,160 A Heist AI task force just released this report, 60 00:02:06,160 --> 00:02:09,120 pushing for a balanced approach to regulation. 61 00:02:09,120 --> 00:02:10,640 You know, they want to support innovation, 62 00:02:10,640 --> 00:02:13,400 but they're also really concerned about national security. 63 00:02:13,400 --> 00:02:16,120 Well, it's a sign that governments are realizing 64 00:02:16,120 --> 00:02:17,520 they can't ignore AI. 65 00:02:17,520 --> 00:02:18,120 Yeah. 66 00:02:18,120 --> 00:02:19,680 It's exploding. 67 00:02:19,680 --> 00:02:23,880 And this report really highlights the need for adaptable 68 00:02:23,880 --> 00:02:27,120 regulations given how fast AI is evolving. 69 00:02:27,120 --> 00:02:27,560 Right. 70 00:02:27,560 --> 00:02:30,240 And it seems like they're also aware of the potential pitfalls, 71 00:02:30,240 --> 00:02:33,600 like biased decision making, and are pushing for human 72 00:02:33,600 --> 00:02:36,120 oversight, especially in critical areas like health care. 73 00:02:36,120 --> 00:02:36,720 Absolutely. 74 00:02:36,720 --> 00:02:38,960 AI should be a tool to help humans not 75 00:02:38,960 --> 00:02:40,440 to replace human judgment. 76 00:02:40,440 --> 00:02:41,280 Right. 77 00:02:41,280 --> 00:02:44,400 Especially in those areas with really complex ethical 78 00:02:44,400 --> 00:02:45,000 considerations. 79 00:02:45,000 --> 00:02:46,760 Yeah, absolutely. 80 00:02:46,760 --> 00:02:48,920 Speaking of complexity, across the pond, 81 00:02:48,920 --> 00:02:51,400 the UK is tackling a big question. 82 00:02:51,400 --> 00:02:52,040 Oh, yeah. 83 00:02:52,040 --> 00:02:54,640 How should AI be allowed to use copyrighted material 84 00:02:54,640 --> 00:02:55,360 for training? 85 00:02:55,360 --> 00:02:56,480 Oh, that's a good question. 86 00:02:56,480 --> 00:02:56,960 Yeah. 87 00:02:56,960 --> 00:02:58,320 It's sparking huge debates. 88 00:02:58,320 --> 00:02:58,920 It is. 89 00:02:58,920 --> 00:03:00,920 Some proposals say AI companies should 90 00:03:00,920 --> 00:03:02,160 be able to use anything. 91 00:03:02,160 --> 00:03:02,680 Wow. 92 00:03:02,680 --> 00:03:05,360 But others want strict licensing agreements. 93 00:03:05,360 --> 00:03:08,360 So it's really a tug of war between the AI industry 94 00:03:08,360 --> 00:03:10,120 and artists' rights. 95 00:03:10,120 --> 00:03:10,640 Yeah. 96 00:03:10,640 --> 00:03:12,160 You can definitely see both sides. 97 00:03:12,160 --> 00:03:12,280 Yeah. 98 00:03:12,280 --> 00:03:14,840 Artists are understandably worried about their work 99 00:03:14,840 --> 00:03:16,000 being exploited. 100 00:03:16,000 --> 00:03:16,760 Right. 101 00:03:16,760 --> 00:03:19,160 Overly restrictive regulations could also 102 00:03:19,160 --> 00:03:20,560 hinder AI development. 103 00:03:20,560 --> 00:03:21,040 Exactly. 104 00:03:21,040 --> 00:03:23,040 And finding that balance is going to be crucial. 105 00:03:23,040 --> 00:03:23,440 Yeah. 106 00:03:23,440 --> 00:03:26,720 We need to protect creators while also making sure 107 00:03:26,720 --> 00:03:30,040 that AI has the data it needs to learn and evolve. 108 00:03:30,040 --> 00:03:30,560 Yeah. 109 00:03:30,560 --> 00:03:32,320 This debate is definitely not over. 110 00:03:32,320 --> 00:03:33,240 No, it's not. 111 00:03:33,240 --> 00:03:36,960 Meanwhile, Google is not letting open AI steal all the thunder. 112 00:03:36,960 --> 00:03:38,160 Oh, what are they doing? 113 00:03:38,160 --> 00:03:40,440 They're rolling out a new experimental version 114 00:03:40,440 --> 00:03:41,760 of their Gemini model. 115 00:03:41,760 --> 00:03:42,240 OK. 116 00:03:42,240 --> 00:03:45,040 Specifically for Gemini advanced subscribers. 117 00:03:45,040 --> 00:03:46,680 And what's this one supposed to be good at? 118 00:03:46,680 --> 00:03:51,440 Well, this one is supposed to be a rock star at coding, math, 119 00:03:51,440 --> 00:03:55,040 reasoning, those tasks that require serious logic. 120 00:03:55,040 --> 00:03:55,640 Oh, wow. 121 00:03:55,640 --> 00:03:59,040 It's clear that Google is going head to head with open AI 122 00:03:59,040 --> 00:04:01,160 to build the most capable AI. 123 00:04:01,160 --> 00:04:02,320 It seems like it, yeah. 124 00:04:02,320 --> 00:04:03,880 So I guess the question is, are we 125 00:04:03,880 --> 00:04:07,880 getting closer to AI that can truly understand and solve 126 00:04:07,880 --> 00:04:09,040 those complex problems? 127 00:04:09,040 --> 00:04:10,600 That's the goal, right? 128 00:04:10,600 --> 00:04:11,000 Yeah. 129 00:04:11,000 --> 00:04:13,440 I mean, it would revolutionize so many fields. 130 00:04:13,440 --> 00:04:14,120 It would. 131 00:04:14,120 --> 00:04:15,000 We are getting there. 132 00:04:15,000 --> 00:04:15,520 OK. 133 00:04:15,520 --> 00:04:17,680 But there's still this big hurdle. 134 00:04:17,680 --> 00:04:19,760 And that's AI hallucination. 135 00:04:19,760 --> 00:04:20,520 OK. 136 00:04:20,520 --> 00:04:22,680 And Google is trying to tackle this head 137 00:04:22,680 --> 00:04:26,000 on with something called FCTS grounding. 138 00:04:26,000 --> 00:04:28,040 OK, this sounds intriguing. 139 00:04:28,040 --> 00:04:32,200 Is this the key to finally getting AI to stick to the facts? 140 00:04:32,200 --> 00:04:33,400 It's a big step in that direction. 141 00:04:33,400 --> 00:04:34,280 OK. 142 00:04:34,280 --> 00:04:38,560 If FCTS grounding is a way to measure how well an AI model 143 00:04:38,560 --> 00:04:41,600 actually grounds its answers in real information 144 00:04:41,600 --> 00:04:42,960 instead of making things up. 145 00:04:42,960 --> 00:04:44,000 OK. 146 00:04:44,000 --> 00:04:46,000 And they even have a leaderboard on Kaggle 147 00:04:46,000 --> 00:04:48,200 so you can see how all the different models stack up. 148 00:04:48,200 --> 00:04:48,640 Oh, wow. 149 00:04:48,640 --> 00:04:50,600 So it's like a fact checking system for AI. 150 00:04:50,600 --> 00:04:51,200 Pretty much. 151 00:04:51,200 --> 00:04:52,040 OK. 152 00:04:52,040 --> 00:04:54,360 That's essential if we want to rely on AI for things 153 00:04:54,360 --> 00:04:56,880 like research or making really important decisions. 154 00:04:56,880 --> 00:04:57,200 It is. 155 00:04:57,200 --> 00:04:59,240 Building trust in AI is crucial. 156 00:04:59,240 --> 00:05:01,560 And knowing that it's not just going to spit out some random 157 00:05:01,560 --> 00:05:03,560 falsehood is a big part of that. 158 00:05:03,560 --> 00:05:04,280 Right. 159 00:05:04,280 --> 00:05:06,280 Now, while those big players are really focused 160 00:05:06,280 --> 00:05:08,160 on building these massive models, 161 00:05:08,160 --> 00:05:10,160 there's also this growing movement 162 00:05:10,160 --> 00:05:12,800 towards smaller, more efficient AI. 163 00:05:12,800 --> 00:05:13,680 There is, yeah. 164 00:05:13,680 --> 00:05:16,640 The Technology Innovation Institute just released Falcon 165 00:05:16,640 --> 00:05:19,800 3, a family of open source small language models. 166 00:05:19,800 --> 00:05:20,880 It's interesting. 167 00:05:20,880 --> 00:05:23,120 So does this mean you don't need a supercomputer 168 00:05:23,120 --> 00:05:24,600 to start playing with AI? 169 00:05:24,600 --> 00:05:25,520 That's the idea. 170 00:05:25,520 --> 00:05:26,080 OK. 171 00:05:26,080 --> 00:05:27,800 And it opens up a lot of opportunities 172 00:05:27,800 --> 00:05:31,000 for researchers, startups, and just individuals 173 00:05:31,000 --> 00:05:32,680 who want to see what AI can do. 174 00:05:32,680 --> 00:05:33,240 Right. 175 00:05:33,240 --> 00:05:36,280 And Falcon 3 is already proving that you don't always 176 00:05:36,280 --> 00:05:38,800 need to sacrifice performance for efficiency. 177 00:05:38,800 --> 00:05:39,280 OK. 178 00:05:39,280 --> 00:05:41,080 It's actually already outperforming some 179 00:05:41,080 --> 00:05:42,720 of those bigger models on certain tasks. 180 00:05:42,720 --> 00:05:44,200 Oh, wow. 181 00:05:44,200 --> 00:05:45,400 That's impressive. 182 00:05:45,400 --> 00:05:49,320 It seems like we're seeing this shift in the AI world moving 183 00:05:49,320 --> 00:05:52,040 away from that bigger is always better mindset 184 00:05:52,040 --> 00:05:54,680 towards something that's more efficient and accessible. 185 00:05:54,680 --> 00:05:56,720 There's definitely a growing recognition 186 00:05:56,720 --> 00:06:00,760 that AI needs to be practical and accessible if it's really 187 00:06:00,760 --> 00:06:01,800 going to have an impact on everyone. 188 00:06:01,800 --> 00:06:02,280 Yeah. 189 00:06:02,280 --> 00:06:03,880 And speaking of impact, researchers 190 00:06:03,880 --> 00:06:06,760 are also finding ways to improve performance 191 00:06:06,760 --> 00:06:09,880 without just throwing more computing power at a problem. 192 00:06:09,880 --> 00:06:11,280 I've heard whispers about this. 193 00:06:11,280 --> 00:06:11,760 Yeah. 194 00:06:11,760 --> 00:06:14,880 Something about letting AI think longer? 195 00:06:14,880 --> 00:06:16,600 It sounds kind of counterintuitive. 196 00:06:16,600 --> 00:06:18,600 It does, but it's a fascinating approach. 197 00:06:18,600 --> 00:06:21,000 It's called scaling test time compute. 198 00:06:21,000 --> 00:06:21,880 OK. 199 00:06:21,880 --> 00:06:27,200 And it essentially lets AI spend more time 200 00:06:27,200 --> 00:06:30,920 processing information, which can allow smaller models 201 00:06:30,920 --> 00:06:32,680 to compete with much bigger models. 202 00:06:32,680 --> 00:06:35,520 So instead of constantly building bigger and bigger models, 203 00:06:35,520 --> 00:06:39,280 we can unlock the potential of the existing models 204 00:06:39,280 --> 00:06:41,880 just by giving them a little more time to process 205 00:06:41,880 --> 00:06:42,880 the information. 206 00:06:42,880 --> 00:06:45,400 It's a really promising area of research. 207 00:06:45,400 --> 00:06:46,000 Yeah. 208 00:06:46,000 --> 00:06:48,880 And it could lead to a more sustainable and democratic 209 00:06:48,880 --> 00:06:51,040 approach to AI development, where 210 00:06:51,040 --> 00:06:53,280 you don't need massive supercomputers 211 00:06:53,280 --> 00:06:55,040 to have access to powerful AI. 212 00:06:55,040 --> 00:06:55,760 That's incredible. 213 00:06:55,760 --> 00:06:57,520 We've covered so much of ground already. 214 00:06:57,520 --> 00:07:01,200 Everything from security concerns to groundbreaking research. 215 00:07:01,200 --> 00:07:01,840 We have. 216 00:07:01,840 --> 00:07:05,560 And it's clear that the AI landscape is changing so rapidly. 217 00:07:05,560 --> 00:07:06,240 It really is. 218 00:07:06,240 --> 00:07:07,520 And this is just a snapshot. 219 00:07:07,520 --> 00:07:07,920 I know. 220 00:07:07,920 --> 00:07:09,120 Things will be different tomorrow. 221 00:07:09,120 --> 00:07:11,160 Maybe even by the time you finish listening to this. 222 00:07:11,160 --> 00:07:12,160 Exactly. 223 00:07:12,160 --> 00:07:14,840 So what stands out to you in all of this? 224 00:07:14,840 --> 00:07:17,440 You know, for me, it's this push to make AI 225 00:07:17,440 --> 00:07:19,160 more practical and accessible. 226 00:07:19,160 --> 00:07:21,360 We're seeing it with these smaller models, 227 00:07:21,360 --> 00:07:23,360 the focus on efficient computing, 228 00:07:23,360 --> 00:07:28,120 and this real drive to create AI that anyone can use. 229 00:07:28,120 --> 00:07:31,080 Not just those with access to these massive computational 230 00:07:31,080 --> 00:07:31,800 resources. 231 00:07:31,800 --> 00:07:32,360 Yeah. 232 00:07:32,360 --> 00:07:36,440 It's a sign that AI is really moving from science fiction 233 00:07:36,440 --> 00:07:38,040 into our everyday lives. 234 00:07:38,040 --> 00:07:38,960 I love that. 235 00:07:38,960 --> 00:07:39,460 Yeah. 236 00:07:39,460 --> 00:07:41,080 It is exciting to think about all the ways 237 00:07:41,080 --> 00:07:43,240 that AI is going to be woven into our lives, 238 00:07:43,240 --> 00:07:45,840 especially as it becomes more user friendly. 239 00:07:45,840 --> 00:07:47,640 But before we get too carried away 240 00:07:47,640 --> 00:07:49,640 with all these possibilities, we really 241 00:07:49,640 --> 00:07:50,960 need to kind of circle back to something 242 00:07:50,960 --> 00:07:52,360 that you mentioned earlier. 243 00:07:52,360 --> 00:07:54,440 Those AI hallucinations. 244 00:07:54,440 --> 00:07:55,680 It sounds a bit funny. 245 00:07:55,680 --> 00:07:56,680 It does. 246 00:07:56,680 --> 00:07:58,560 But it is a serious concern, right? 247 00:07:58,560 --> 00:07:58,960 It is. 248 00:07:58,960 --> 00:08:02,200 As we start to really rely more and more on AI 249 00:08:02,200 --> 00:08:04,920 for information and decision making, 250 00:08:04,920 --> 00:08:07,640 we have to make sure that it's not just making things up. 251 00:08:07,640 --> 00:08:09,800 OK, so explain this hallucination thing to me 252 00:08:09,800 --> 00:08:10,520 in plain English. 253 00:08:10,520 --> 00:08:11,000 OK. 254 00:08:11,000 --> 00:08:13,440 Does this mean the AI is intentionally lying? 255 00:08:13,440 --> 00:08:15,040 Well, not necessarily lying. 256 00:08:15,040 --> 00:08:15,540 OK. 257 00:08:15,540 --> 00:08:18,520 But it's when an AI generates outputs 258 00:08:18,520 --> 00:08:19,840 that are factually wrong. 259 00:08:19,840 --> 00:08:22,920 Even if it presents those like a high degree of confidence. 260 00:08:22,920 --> 00:08:23,520 OK. 261 00:08:23,520 --> 00:08:26,680 It's like the AI is confidently stating something 262 00:08:26,680 --> 00:08:28,080 that's just plain false. 263 00:08:28,080 --> 00:08:29,560 OK, that's a little unsettling. 264 00:08:29,560 --> 00:08:30,320 It is a little bit. 265 00:08:30,320 --> 00:08:32,120 It makes you question everything it tells you. 266 00:08:32,120 --> 00:08:35,200 So what causes these hallucinations? 267 00:08:35,200 --> 00:08:37,600 Is it just a lack of data? 268 00:08:37,600 --> 00:08:39,240 It's a little more complicated than that. 269 00:08:39,240 --> 00:08:39,740 OK. 270 00:08:39,740 --> 00:08:44,200 One factor is definitely the way AI models are trained. 271 00:08:44,200 --> 00:08:44,720 OK. 272 00:08:44,720 --> 00:08:47,080 They learn from massive amounts of data. 273 00:08:47,080 --> 00:08:53,080 But if that data is biased or incomplete or contradictory, 274 00:08:53,080 --> 00:08:56,320 the AI might learn to make these incorrect connections. 275 00:08:56,320 --> 00:08:58,600 So it's kind of like that old saying, 276 00:08:58,600 --> 00:09:00,000 garbage in, garbage out. 277 00:09:00,000 --> 00:09:00,440 Exactly. 278 00:09:00,440 --> 00:09:04,400 Even with a massive data set, if the data itself is flawed, 279 00:09:04,400 --> 00:09:07,240 the AI's output is going to reflect those flaws. 280 00:09:07,240 --> 00:09:07,880 Exactly. 281 00:09:07,880 --> 00:09:09,840 Another factor is the AI's understanding 282 00:09:09,840 --> 00:09:11,080 of the real world. 283 00:09:11,080 --> 00:09:11,920 OK. 284 00:09:11,920 --> 00:09:15,280 You know, current models, even as advanced as they are, 285 00:09:15,280 --> 00:09:19,520 they still lack that deep nuanced understanding 286 00:09:19,520 --> 00:09:20,600 that humans have. 287 00:09:20,600 --> 00:09:21,000 Right. 288 00:09:21,000 --> 00:09:24,640 You don't really grasp things like common sense or ethics. 289 00:09:24,640 --> 00:09:25,480 Uh-huh. 290 00:09:25,480 --> 00:09:27,960 Or the complexity of human language and behavior. 291 00:09:27,960 --> 00:09:28,460 Yeah. 292 00:09:28,460 --> 00:09:32,960 So even if the AI has access to all the facts in the world, 293 00:09:32,960 --> 00:09:35,600 it might not understand how those facts fit together 294 00:09:35,600 --> 00:09:36,560 in the real world. 295 00:09:36,560 --> 00:09:37,040 Right. 296 00:09:37,040 --> 00:09:38,840 It's missing that human context. 297 00:09:38,840 --> 00:09:39,560 Precisely. 298 00:09:39,560 --> 00:09:40,040 Yeah. 299 00:09:40,040 --> 00:09:41,600 And that lack of context can lead 300 00:09:41,600 --> 00:09:45,520 to all sorts of misinterpretations and illogical leaps 301 00:09:45,520 --> 00:09:47,600 and ultimately hallucinations. 302 00:09:47,600 --> 00:09:47,800 OK. 303 00:09:47,800 --> 00:09:51,080 So this is where Google's FCPS grounding comes in. 304 00:09:51,080 --> 00:09:54,800 It's a way to measure how well the AI sticks to the facts. 305 00:09:54,800 --> 00:09:55,300 It is. 306 00:09:55,300 --> 00:10:00,120 It's essentially a benchmark for evaluating how well AI models 307 00:10:00,120 --> 00:10:02,920 ground their answers in real information 308 00:10:02,920 --> 00:10:05,720 and avoid making things up, especially 309 00:10:05,720 --> 00:10:09,840 in those longer form responses that require pulling together 310 00:10:09,840 --> 00:10:11,400 information from different sources. 311 00:10:11,400 --> 00:10:11,900 OK. 312 00:10:11,900 --> 00:10:15,120 So instead of just like asking the AI a simple question, 313 00:10:15,120 --> 00:10:16,920 you give it like a whole document to work with. 314 00:10:16,920 --> 00:10:17,400 Exactly. 315 00:10:17,400 --> 00:10:19,360 And then see if it can accurately summarize 316 00:10:19,360 --> 00:10:21,880 that information without adding its own embellishment. 317 00:10:21,880 --> 00:10:22,380 Exactly. 318 00:10:22,380 --> 00:10:25,480 It's a much more challenging test for the AI. 319 00:10:25,480 --> 00:10:28,880 And all the results are put on a leaderboard on Kaggle. 320 00:10:28,880 --> 00:10:29,480 Oh, cool. 321 00:10:29,480 --> 00:10:32,040 So you can see how all the different models are performing. 322 00:10:32,040 --> 00:10:34,480 So it's like a public test of AI honesty. 323 00:10:34,480 --> 00:10:35,400 Yeah, pretty much. 324 00:10:35,400 --> 00:10:36,000 I love that. 325 00:10:36,000 --> 00:10:37,600 It's great to see that transparency. 326 00:10:37,600 --> 00:10:41,960 But is FASTS grounding like the ultimate solution? 327 00:10:41,960 --> 00:10:43,920 Well, it's definitely a valuable tool. 328 00:10:43,920 --> 00:10:44,320 OK. 329 00:10:44,320 --> 00:10:46,200 But it's only one piece of the puzzle. 330 00:10:46,200 --> 00:10:46,680 OK. 331 00:10:46,680 --> 00:10:49,200 There are a lot of different approaches being explored. 332 00:10:49,200 --> 00:10:50,200 OK, like what? 333 00:10:50,200 --> 00:10:52,240 Well, one is reinforcement learning, 334 00:10:52,240 --> 00:10:58,200 where you essentially reward the AI for giving accurate answers 335 00:10:58,200 --> 00:11:01,160 and then penalize it for hallucinations. 336 00:11:01,160 --> 00:11:03,880 So it's like training a pet but for AI. 337 00:11:03,880 --> 00:11:04,640 Yeah, kind of. 338 00:11:04,640 --> 00:11:07,120 Reward the good behavior, discourage the bad. 339 00:11:07,120 --> 00:11:08,200 Exactly. 340 00:11:08,200 --> 00:11:10,280 Another approach is giving the AI access 341 00:11:10,280 --> 00:11:14,400 to external knowledge sources, like fact checking websites, 342 00:11:14,400 --> 00:11:16,000 so it can double check its claims. 343 00:11:16,000 --> 00:11:17,800 It's like giving it a research assistant, 344 00:11:17,800 --> 00:11:20,000 being like, hey, before you tell me something, 345 00:11:20,000 --> 00:11:21,680 maybe make sure it's actually true. 346 00:11:21,680 --> 00:11:22,800 Exactly. 347 00:11:22,800 --> 00:11:25,920 There's also a push to develop specialized AI models 348 00:11:25,920 --> 00:11:30,160 that are designed for tasks where accuracy is super important, 349 00:11:30,160 --> 00:11:32,080 like summarizing scientific research 350 00:11:32,080 --> 00:11:33,920 or generating legal documents. 351 00:11:33,920 --> 00:11:35,840 It's like having different AI specialists, 352 00:11:35,840 --> 00:11:37,960 each trained to be really careful with the facts 353 00:11:37,960 --> 00:11:39,640 in their own specific field. 354 00:11:39,640 --> 00:11:40,800 Exactly. 355 00:11:40,800 --> 00:11:42,760 And then there's a lot of ongoing research 356 00:11:42,760 --> 00:11:46,680 into making AI models more interpretable, 357 00:11:46,680 --> 00:11:50,400 meaning we can understand how the AI arrives at its conclusion. 358 00:11:50,400 --> 00:11:52,240 So it's not just about getting the right answer. 359 00:11:52,240 --> 00:11:55,280 It's about understanding the AI thought process, 360 00:11:55,280 --> 00:11:58,920 so we can identify any potential biases or errors. 361 00:11:58,920 --> 00:11:59,560 Exactly. 362 00:11:59,560 --> 00:12:02,680 If we can understand how the AI thinks, 363 00:12:02,680 --> 00:12:05,560 we can better assess its reliability 364 00:12:05,560 --> 00:12:09,160 and spot those areas where it might be prone to hallucinations. 365 00:12:09,160 --> 00:12:10,800 This is all very encouraging. 366 00:12:10,800 --> 00:12:11,400 It is. 367 00:12:11,400 --> 00:12:13,560 It seems like researchers are really working hard 368 00:12:13,560 --> 00:12:15,520 to tackle this hallucination problem. 369 00:12:15,520 --> 00:12:16,000 They are. 370 00:12:16,000 --> 00:12:20,080 Which is crucial if we want to trust and rely on these AI systems. 371 00:12:20,080 --> 00:12:22,920 Addressing this problem is really essential 372 00:12:22,920 --> 00:12:25,840 if AI is going to reach its full potential. 373 00:12:25,840 --> 00:12:27,760 We have to be able to confidently rely 374 00:12:27,760 --> 00:12:29,120 on the information it provides. 375 00:12:29,120 --> 00:12:30,840 Yeah, absolutely. 376 00:12:30,840 --> 00:12:33,160 Speaking of trust, let's talk a little bit 377 00:12:33,160 --> 00:12:37,880 about how governments are getting involved in the AI landscape, 378 00:12:37,880 --> 00:12:39,440 especially in the US and the UK. 379 00:12:39,440 --> 00:12:41,240 We touched on this briefly before. 380 00:12:41,240 --> 00:12:41,640 We did. 381 00:12:41,640 --> 00:12:45,720 But given the potential impact of AI on everything, 382 00:12:45,720 --> 00:12:49,000 like national security to the creative industries, 383 00:12:49,000 --> 00:12:50,800 I think it deserves a deeper look. 384 00:12:50,800 --> 00:12:51,760 I think so too. 385 00:12:51,760 --> 00:12:54,120 So governments are realizing that AI 386 00:12:54,120 --> 00:12:57,280 isn't just some futuristic concept. 387 00:12:57,280 --> 00:13:00,480 It's a technology that's already changing the world. 388 00:13:00,480 --> 00:13:03,840 And they need to figure out how to regulate it responsibly. 389 00:13:03,840 --> 00:13:04,920 Yeah, they do. 390 00:13:04,920 --> 00:13:06,520 So let's start with the US. 391 00:13:06,520 --> 00:13:11,160 We talked about the House AI Task Force Report. 392 00:13:11,160 --> 00:13:13,920 What are some of the key takeaways from that? 393 00:13:13,920 --> 00:13:17,640 Well, their main goal seems to be to find that balance 394 00:13:17,640 --> 00:13:21,240 between encouraging innovation and protecting 395 00:13:21,240 --> 00:13:22,800 national interests. 396 00:13:22,800 --> 00:13:24,720 They don't want to stifle progress. 397 00:13:24,720 --> 00:13:27,240 But they also want to make sure that AI is developed and used 398 00:13:27,240 --> 00:13:29,880 in a way that benefits society. 399 00:13:29,880 --> 00:13:32,400 So it's not about stopping AI. 400 00:13:32,400 --> 00:13:34,720 It's about steering it in the right direction. 401 00:13:34,720 --> 00:13:36,080 Exactly. 402 00:13:36,080 --> 00:13:38,840 And one of their big concerns is that AI 403 00:13:38,840 --> 00:13:41,600 could be used to threaten national security. 404 00:13:41,600 --> 00:13:44,160 They're also worried about bias decision making, 405 00:13:44,160 --> 00:13:47,120 especially in areas like law enforcement or defense, 406 00:13:47,120 --> 00:13:48,720 where the stakes are really high. 407 00:13:48,720 --> 00:13:50,960 Yeah, it goes back to what we were talking about earlier 408 00:13:50,960 --> 00:13:54,040 about addressing bias in the data that's 409 00:13:54,040 --> 00:13:56,880 used to train these AI models. 410 00:13:56,880 --> 00:14:00,760 If the data is flawed, the AI's decisions 411 00:14:00,760 --> 00:14:02,400 will reflect those flaws. 412 00:14:02,400 --> 00:14:02,960 They will. 413 00:14:02,960 --> 00:14:05,440 And that could have really serious consequences. 414 00:14:05,440 --> 00:14:06,600 Absolutely. 415 00:14:06,600 --> 00:14:08,520 The Task Force is really advocating 416 00:14:08,520 --> 00:14:12,800 for transparency and accountability in AI systems, 417 00:14:12,800 --> 00:14:14,680 especially those used by the government. 418 00:14:14,680 --> 00:14:17,880 They want to be able to understand how these systems work, 419 00:14:17,880 --> 00:14:20,440 what data they're using, and how they're making decisions. 420 00:14:20,440 --> 00:14:22,920 So no more black boxes when it comes to AI. 421 00:14:22,920 --> 00:14:23,520 Yeah, pretty much. 422 00:14:23,520 --> 00:14:26,760 We need to be able to understand how the sausage is made. 423 00:14:26,760 --> 00:14:27,960 So to speak. 424 00:14:27,960 --> 00:14:29,600 To make sure this being made responsible. 425 00:14:29,600 --> 00:14:30,520 Exactly. 426 00:14:30,520 --> 00:14:33,600 But they're not just focused on preventing misuse. 427 00:14:33,600 --> 00:14:37,240 They also see AI as this really powerful tool 428 00:14:37,240 --> 00:14:39,920 that could be used to improve government efficiency 429 00:14:39,920 --> 00:14:41,320 and effectiveness. 430 00:14:41,320 --> 00:14:45,280 So they're encouraging agencies to explore how AI could 431 00:14:45,280 --> 00:14:49,680 be used to streamline processes, reduce costs, 432 00:14:49,680 --> 00:14:52,440 and provide better services to citizens. 433 00:14:52,440 --> 00:14:55,360 So they recognize that AI can be both a challenge 434 00:14:55,360 --> 00:14:56,520 and an opportunity. 435 00:14:56,520 --> 00:14:57,560 It can. 436 00:14:57,560 --> 00:15:00,600 It's about finding ways to mitigate the risks 437 00:15:00,600 --> 00:15:01,960 while maximizing the benefit. 438 00:15:01,960 --> 00:15:02,440 Exactly. 439 00:15:02,440 --> 00:15:04,880 And they think education is a key part of that. 440 00:15:04,880 --> 00:15:08,840 The Task Force is calling for more investment in AI literacy, 441 00:15:08,840 --> 00:15:11,200 starting in kindergarten all the way through high school. 442 00:15:11,200 --> 00:15:11,760 That's great. 443 00:15:11,760 --> 00:15:15,200 It's about making sure that future generations have 444 00:15:15,200 --> 00:15:20,240 the knowledge and skills to navigate this AI-driven world. 445 00:15:20,240 --> 00:15:23,040 They want to make sure that people are prepared to engage 446 00:15:23,040 --> 00:15:26,200 with AI, understand its potential, 447 00:15:26,200 --> 00:15:29,280 and make informed decisions about its use. 448 00:15:29,280 --> 00:15:32,400 All right, now let's shift our focus to the UK. 449 00:15:32,400 --> 00:15:35,080 They're grappling with a really interesting issue. 450 00:15:35,080 --> 00:15:39,000 How should AI be allowed to use copyrighted material 451 00:15:39,000 --> 00:15:40,200 for training? 452 00:15:40,200 --> 00:15:41,400 That's a great question. 453 00:15:41,400 --> 00:15:43,960 Yeah, it's a question that's causing a lot of tension 454 00:15:43,960 --> 00:15:46,480 between the AI industry and the creative community. 455 00:15:46,480 --> 00:15:47,440 That's right. 456 00:15:47,440 --> 00:15:51,360 AI models learn from massive amounts of data. 457 00:15:51,360 --> 00:15:52,120 They do. 458 00:15:52,120 --> 00:15:55,400 And that data often includes things like text, images, 459 00:15:55,400 --> 00:15:59,880 music, and even code that are all protected by copyright. 460 00:15:59,880 --> 00:16:00,280 It is. 461 00:16:00,280 --> 00:16:02,160 It's a tough balancing act. 462 00:16:02,160 --> 00:16:05,240 On one hand, you have these AI companies saying, 463 00:16:05,240 --> 00:16:07,800 they need access to this data to develop 464 00:16:07,800 --> 00:16:09,680 their sophisticated models. 465 00:16:09,680 --> 00:16:11,520 But then you have artists and creators 466 00:16:11,520 --> 00:16:15,040 who are understandably worried about their work being 467 00:16:15,040 --> 00:16:17,760 used without their permission or compensation. 468 00:16:17,760 --> 00:16:19,560 So it's really hard to find a solution that 469 00:16:19,560 --> 00:16:20,480 works for everybody. 470 00:16:20,480 --> 00:16:21,320 It is. 471 00:16:21,320 --> 00:16:24,040 The UK government is exploring a lot of different policy 472 00:16:24,040 --> 00:16:25,000 options. 473 00:16:25,000 --> 00:16:27,040 Ranging from a more permissive approach, 474 00:16:27,040 --> 00:16:30,160 where AI companies would have more freedom 475 00:16:30,160 --> 00:16:33,880 to use that copyrighted material to a more restrictive 476 00:16:33,880 --> 00:16:36,880 approach, where they would need to obtain licenses 477 00:16:36,880 --> 00:16:37,760 and pay royalties. 478 00:16:37,760 --> 00:16:39,760 So it's a whole spectrum of possibilities. 479 00:16:39,760 --> 00:16:40,200 It is. 480 00:16:40,200 --> 00:16:42,440 Each with its own pros and cons. 481 00:16:42,440 --> 00:16:45,400 So what are some of the arguments for and against 482 00:16:45,400 --> 00:16:47,040 each approach? 483 00:16:47,040 --> 00:16:50,720 Well, those who favor a more permissive approach 484 00:16:50,720 --> 00:16:54,840 argue that it would encourage innovation in the AI industry. 485 00:16:54,840 --> 00:16:55,400 Yeah. 486 00:16:55,400 --> 00:16:58,040 And allow the UK to stay competitive. 487 00:16:58,040 --> 00:16:58,520 Right. 488 00:16:58,520 --> 00:17:00,040 In the global AI race. 489 00:17:00,040 --> 00:17:00,960 Neckhens. 490 00:17:00,960 --> 00:17:03,160 They say that they're worried that strict regulations 491 00:17:03,160 --> 00:17:04,880 would stifle progress. 492 00:17:04,880 --> 00:17:08,840 So their argument is, don't tie the hands of our AI developers. 493 00:17:08,840 --> 00:17:09,160 Right. 494 00:17:09,160 --> 00:17:11,600 We need to give them the freedom to experiment 495 00:17:11,600 --> 00:17:12,520 and push the boundaries. 496 00:17:12,520 --> 00:17:13,560 Exactly. 497 00:17:13,560 --> 00:17:16,080 But then the creative community is pushing back 498 00:17:16,080 --> 00:17:16,920 hard against that. 499 00:17:16,920 --> 00:17:17,320 OK. 500 00:17:17,320 --> 00:17:19,880 They say it would basically be legalizing 501 00:17:19,880 --> 00:17:21,880 theft of their intellectual property. 502 00:17:21,880 --> 00:17:22,240 Right. 503 00:17:22,240 --> 00:17:24,760 They're saying, our work has value. 504 00:17:24,760 --> 00:17:25,260 Yeah. 505 00:17:25,260 --> 00:17:28,960 And we deserve to be compensated when it's used even by AI. 506 00:17:28,960 --> 00:17:31,440 And they're also concerned that their work will 507 00:17:31,440 --> 00:17:33,000 be used without their consent. 508 00:17:33,000 --> 00:17:33,640 Yeah. 509 00:17:33,640 --> 00:17:36,000 And that they won't see any of the financial benefits from it. 510 00:17:36,000 --> 00:17:36,440 Yeah. 511 00:17:36,440 --> 00:17:38,600 I mean, there's also this worry that AI 512 00:17:38,600 --> 00:17:42,320 will be used to create derivative works that 513 00:17:42,320 --> 00:17:44,520 are so similar to the original creations 514 00:17:44,520 --> 00:17:46,400 that it devalues their work. 515 00:17:46,400 --> 00:17:46,960 Exactly. 516 00:17:46,960 --> 00:17:48,920 It makes it harder for them to earn a living. 517 00:17:48,920 --> 00:17:49,400 Right. 518 00:17:49,400 --> 00:17:51,160 It's a valid concern. 519 00:17:51,160 --> 00:17:55,000 It is, if AI can just churn out imitations of art, music, 520 00:17:55,000 --> 00:17:59,040 and literature, it could really flood the market 521 00:17:59,040 --> 00:18:01,480 and make it difficult for human creators to stand out. 522 00:18:01,480 --> 00:18:02,080 Yeah. 523 00:18:02,080 --> 00:18:04,680 And beyond the economic impact, there's also 524 00:18:04,680 --> 00:18:07,320 that question of artistic integrity 525 00:18:07,320 --> 00:18:10,760 and the right of creators to control how their work is 526 00:18:10,760 --> 00:18:11,440 being used. 527 00:18:11,440 --> 00:18:12,280 Exactly. 528 00:18:12,280 --> 00:18:14,080 A lot of artists feel like their creations 529 00:18:14,080 --> 00:18:15,480 are an extension of themselves. 530 00:18:15,480 --> 00:18:16,160 Yeah. 531 00:18:16,160 --> 00:18:19,480 And they don't want to see them exploited or misused 532 00:18:19,480 --> 00:18:20,520 even by AI. 533 00:18:20,520 --> 00:18:22,600 So for them, it's not just about the money. 534 00:18:22,600 --> 00:18:22,920 Right. 535 00:18:22,920 --> 00:18:25,160 It's about respect for their creative work 536 00:18:25,160 --> 00:18:26,880 and the effort they put into it. 537 00:18:26,880 --> 00:18:27,920 Absolutely. 538 00:18:27,920 --> 00:18:31,920 And then those who favor the more restrictive approach, 539 00:18:31,920 --> 00:18:35,160 they argue that requiring AI companies 540 00:18:35,160 --> 00:18:38,280 to obtain licenses and pay royalties 541 00:18:38,280 --> 00:18:40,160 would be a much fairer balance. 542 00:18:40,160 --> 00:18:41,760 So they're looking for a solution 543 00:18:41,760 --> 00:18:44,920 where AI can still access that data, 544 00:18:44,920 --> 00:18:46,880 but creators are fairly compensated. 545 00:18:46,880 --> 00:18:47,480 Exactly. 546 00:18:47,480 --> 00:18:49,920 They think a well-defined licensing system 547 00:18:49,920 --> 00:18:51,560 would be a win-win situation. 548 00:18:51,560 --> 00:18:53,160 It sounds good in theory. 549 00:18:53,160 --> 00:18:53,800 It does. 550 00:18:53,800 --> 00:18:55,080 But is it practical? 551 00:18:55,080 --> 00:18:55,960 Well, I've got the question. 552 00:18:55,960 --> 00:18:56,400 Yeah. 553 00:18:56,400 --> 00:18:59,160 I mean, managing a system like that 554 00:18:59,160 --> 00:19:00,440 seems incredibly complex. 555 00:19:00,440 --> 00:19:01,880 It would definitely not be easy. 556 00:19:01,880 --> 00:19:04,560 It would require clear guidelines, 557 00:19:04,560 --> 00:19:06,520 streamlined licensing agreements, 558 00:19:06,520 --> 00:19:09,520 and robust mechanisms for tracking and distributing 559 00:19:09,520 --> 00:19:10,200 royalties. 560 00:19:10,200 --> 00:19:10,840 Yeah. 561 00:19:10,840 --> 00:19:14,160 But there are a lot of different organizations 562 00:19:14,160 --> 00:19:16,200 working to develop these kinds of systems. 563 00:19:16,200 --> 00:19:18,240 So there is hope that we can find a solution. 564 00:19:18,240 --> 00:19:18,800 There is. 565 00:19:18,800 --> 00:19:23,000 That protects creators' rights without stifling AI innovation. 566 00:19:23,000 --> 00:19:23,880 That's the goal. 567 00:19:23,880 --> 00:19:24,160 Yeah. 568 00:19:24,160 --> 00:19:25,760 It really highlights something important. 569 00:19:25,760 --> 00:19:26,560 What's that? 570 00:19:26,560 --> 00:19:28,400 AI isn't just about the technology. 571 00:19:28,400 --> 00:19:30,840 It's about the impact that it has on our lives, 572 00:19:30,840 --> 00:19:34,240 our livelihoods, and even our creative expressions. 573 00:19:34,240 --> 00:19:35,040 Absolutely. 574 00:19:35,040 --> 00:19:39,760 And as AI continues to evolve, these ethical and societal 575 00:19:39,760 --> 00:19:43,240 considerations are going to become even more crucial. 576 00:19:43,240 --> 00:19:44,240 I think you're right. 577 00:19:44,240 --> 00:19:46,440 So this has been a fantastic discussion. 578 00:19:46,440 --> 00:19:47,000 It has. 579 00:19:47,000 --> 00:19:48,240 We've covered a lot of ground. 580 00:19:48,240 --> 00:19:48,760 We have. 581 00:19:48,760 --> 00:19:51,440 We've talked about the latest AI models, 582 00:19:51,440 --> 00:19:54,280 the challenges of these AI hallucinations, 583 00:19:54,280 --> 00:19:57,720 and even these really complex policy debates surrounding 584 00:19:57,720 --> 00:19:59,720 AI use of copyrighted material. 585 00:19:59,720 --> 00:20:00,720 It's a lot to think about. 586 00:20:00,720 --> 00:20:01,200 Yeah. 587 00:20:01,200 --> 00:20:02,800 And we're just scratching the surface. 588 00:20:02,800 --> 00:20:03,080 I know. 589 00:20:03,080 --> 00:20:05,160 The world of AI is constantly changing, 590 00:20:05,160 --> 00:20:05,640 Right. 591 00:20:05,640 --> 00:20:06,840 Which is what makes it so exciting. 592 00:20:06,840 --> 00:20:07,840 It does. 593 00:20:07,840 --> 00:20:09,280 But also a little overwhelming. 594 00:20:09,280 --> 00:20:09,600 Yeah. 595 00:20:09,600 --> 00:20:12,000 It definitely feels like we're standing on the precipice 596 00:20:12,000 --> 00:20:13,880 of a technological revolution. 597 00:20:13,880 --> 00:20:14,600 It does. 598 00:20:14,600 --> 00:20:15,040 Yeah. 599 00:20:15,040 --> 00:20:18,120 It really is amazing to think about how quickly things 600 00:20:18,120 --> 00:20:20,520 are changing in the AI world. 601 00:20:20,520 --> 00:20:21,120 It is. 602 00:20:21,120 --> 00:20:25,000 It seems like what was science fiction just a few years ago 603 00:20:25,000 --> 00:20:27,160 is now becoming part of our reality. 604 00:20:27,160 --> 00:20:29,760 It really is an incredible time to be a part of this field. 605 00:20:29,760 --> 00:20:30,120 Yeah. 606 00:20:30,120 --> 00:20:30,760 It is. 607 00:20:30,760 --> 00:20:33,360 And with all these advancements, 608 00:20:33,360 --> 00:20:36,200 one trend that I find particularly interesting 609 00:20:36,200 --> 00:20:39,680 is this move toward these smaller, more efficient AI 610 00:20:39,680 --> 00:20:40,320 models. 611 00:20:40,320 --> 00:20:42,680 Oh, you're talking about things like those Falcon 3 models 612 00:20:42,680 --> 00:20:43,400 we discussed. 613 00:20:43,400 --> 00:20:43,800 Exactly. 614 00:20:43,800 --> 00:20:46,640 The ones that are designed to run on less powerful hardware. 615 00:20:46,640 --> 00:20:49,200 Yeah, there's this growing realization 616 00:20:49,200 --> 00:20:52,560 that bigger isn't always better when it comes to AI. 617 00:20:52,560 --> 00:20:53,120 Right. 618 00:20:53,120 --> 00:20:56,320 Sometimes a smaller, more focused model 619 00:20:56,320 --> 00:20:59,440 can be just as effective, if not more so, 620 00:20:59,440 --> 00:21:02,360 than these massive resource-intensive models. 621 00:21:02,360 --> 00:21:05,560 So it's kind of like finding that perfect tool for the job. 622 00:21:05,560 --> 00:21:06,120 Exactly. 623 00:21:06,120 --> 00:21:08,080 You don't always need a sledgehammer. 624 00:21:08,080 --> 00:21:08,480 Right. 625 00:21:08,480 --> 00:21:11,360 When a well-placed tap with a hammer will do. 626 00:21:11,360 --> 00:21:12,640 That's a great analogy. 627 00:21:12,640 --> 00:21:14,600 And this shift towards smaller models 628 00:21:14,600 --> 00:21:16,440 is being driven by a few different things. 629 00:21:16,440 --> 00:21:16,920 OK. 630 00:21:16,920 --> 00:21:17,880 Like what? 631 00:21:17,880 --> 00:21:20,280 Well, the increase in cost of energy, 632 00:21:20,280 --> 00:21:23,760 a growing awareness of the environmental impact 633 00:21:23,760 --> 00:21:28,480 of these massive data centers, and just this desire 634 00:21:28,480 --> 00:21:32,160 to make AI more accessible to a wider range of users. 635 00:21:32,160 --> 00:21:32,960 Yeah, it makes sense. 636 00:21:32,960 --> 00:21:35,560 I mean, not everyone has access to a supercomputer. 637 00:21:35,560 --> 00:21:35,880 Right. 638 00:21:35,880 --> 00:21:39,320 And if we want AI to benefit society, 639 00:21:39,320 --> 00:21:41,600 we have to make sure that it's available to everyone. 640 00:21:41,600 --> 00:21:42,720 Exactly. 641 00:21:42,720 --> 00:21:45,000 And that brings us to another really interesting development. 642 00:21:45,000 --> 00:21:45,680 OK. 643 00:21:45,680 --> 00:21:48,880 The idea of scaling test time compute. 644 00:21:48,880 --> 00:21:49,280 Oh, yeah. 645 00:21:49,280 --> 00:21:50,320 Let's revisit that. 646 00:21:50,320 --> 00:21:50,960 OK. 647 00:21:50,960 --> 00:21:53,560 You mentioned that it's about giving the AI more time 648 00:21:53,560 --> 00:21:56,320 to think, which sounds a little counterintuitive. 649 00:21:56,320 --> 00:21:57,360 It does a little bit. 650 00:21:57,360 --> 00:22:00,320 But the idea is that instead of just giving a problem 651 00:22:00,320 --> 00:22:03,560 to the AI and expecting an instant answer, 652 00:22:03,560 --> 00:22:07,360 we let it spend more time processing that information, 653 00:22:07,360 --> 00:22:09,000 exploring different possibilities, 654 00:22:09,000 --> 00:22:10,560 and refining its response. 655 00:22:10,560 --> 00:22:12,000 So you're giving the AI the chance 656 00:22:12,000 --> 00:22:15,120 to really think things through rather than just 657 00:22:15,120 --> 00:22:16,720 jumping to the first conclusion. 658 00:22:16,720 --> 00:22:17,640 Exactly. 659 00:22:17,640 --> 00:22:20,680 And what's really remarkable is that this approach is allowing 660 00:22:20,680 --> 00:22:26,080 these smaller, less powerful AI models 661 00:22:26,080 --> 00:22:28,160 to get results that we thought could only 662 00:22:28,160 --> 00:22:31,560 be achieved by these massive computationally expensive models. 663 00:22:31,560 --> 00:22:32,840 That's what makes it so exciting. 664 00:22:32,840 --> 00:22:33,600 It is. 665 00:22:33,600 --> 00:22:35,720 It opens up a whole world of possibilities 666 00:22:35,720 --> 00:22:39,320 for people who don't have access to those giant supercomputers. 667 00:22:39,320 --> 00:22:39,800 Exactly. 668 00:22:39,800 --> 00:22:45,960 Imagine students being able to experiment with cutting-edge 669 00:22:45,960 --> 00:22:51,000 AI models on their laptops or small startups developing 670 00:22:51,000 --> 00:22:53,800 these innovative AI applications without having 671 00:22:53,800 --> 00:22:55,600 to invest in a ton of infrastructure. 672 00:22:55,600 --> 00:22:56,040 Yeah. 673 00:22:56,040 --> 00:22:58,280 It really democratizes access to AI. 674 00:22:58,280 --> 00:22:59,200 It does. 675 00:22:59,200 --> 00:23:03,560 But how exactly does this thinking longer process work? 676 00:23:03,560 --> 00:23:04,040 Well. 677 00:23:04,040 --> 00:23:06,200 Does the AI just sit there and ponder 678 00:23:06,200 --> 00:23:08,040 until it has some kind of breakthrough? 679 00:23:08,040 --> 00:23:09,880 Well, it's a little more structured than that. 680 00:23:09,880 --> 00:23:10,320 OK. 681 00:23:10,320 --> 00:23:13,520 Researchers are developing specific techniques 682 00:23:13,520 --> 00:23:16,160 to really guide that AI's thought process 683 00:23:16,160 --> 00:23:17,360 and make it more efficient. 684 00:23:17,360 --> 00:23:17,720 OK. 685 00:23:17,720 --> 00:23:21,400 And a key element is this thing called a process reward model. 686 00:23:21,400 --> 00:23:21,760 OK. 687 00:23:21,760 --> 00:23:23,160 Remind me what that is again. 688 00:23:23,160 --> 00:23:25,880 Think of it like a coach or mentor for the AI. 689 00:23:25,880 --> 00:23:26,360 OK. 690 00:23:26,360 --> 00:23:29,600 Instead of just looking at the AI's final answer, 691 00:23:29,600 --> 00:23:33,160 a process reward model provides feedback at each step 692 00:23:33,160 --> 00:23:34,840 of the AI's reasoning process. 693 00:23:34,840 --> 00:23:36,600 So it's kind of like having a teacher who 694 00:23:36,600 --> 00:23:38,760 doesn't just grade your final exam, 695 00:23:38,760 --> 00:23:40,800 but actually gives you feedback along the way. 696 00:23:40,800 --> 00:23:41,360 Exactly. 697 00:23:41,360 --> 00:23:42,960 So you can improve your approach. 698 00:23:42,960 --> 00:23:44,840 And one way this guidance is used 699 00:23:44,840 --> 00:23:47,040 is through something called Beam Search. 700 00:23:47,040 --> 00:23:48,160 Beam Search. 701 00:23:48,160 --> 00:23:49,440 That sounds familiar. 702 00:23:49,440 --> 00:23:49,720 Yeah. 703 00:23:49,720 --> 00:23:51,120 We talked about it briefly before. 704 00:23:51,120 --> 00:23:55,600 It's a way for the AI to explore multiple possible solution 705 00:23:55,600 --> 00:23:57,040 paths at the same time. 706 00:23:57,040 --> 00:23:57,720 OK. 707 00:23:57,720 --> 00:24:00,760 Kind of like branching out in different directions 708 00:24:00,760 --> 00:24:02,880 to find the best route through a maze. 709 00:24:02,880 --> 00:24:06,000 So instead of just sticking to one line of reasoning. 710 00:24:06,000 --> 00:24:06,680 Right. 711 00:24:06,680 --> 00:24:09,600 The AI is exploring different possibilities 712 00:24:09,600 --> 00:24:10,640 simultaneously. 713 00:24:10,640 --> 00:24:11,360 Exactly. 714 00:24:11,360 --> 00:24:13,320 And the process reward model helps 715 00:24:13,320 --> 00:24:14,960 to prioritize those paths. 716 00:24:14,960 --> 00:24:15,720 OK. 717 00:24:15,720 --> 00:24:18,160 Guiding the AI toward the most promising solutions. 718 00:24:18,160 --> 00:24:18,840 OK. 719 00:24:18,840 --> 00:24:20,600 And pruning away the less likely ones. 720 00:24:20,600 --> 00:24:21,200 Got it. 721 00:24:21,200 --> 00:24:23,080 Another technique that they're using 722 00:24:23,080 --> 00:24:25,600 is something called Diverse Verifier Tree Search. 723 00:24:25,600 --> 00:24:27,560 Diverse Verifier Tree Search. 724 00:24:27,560 --> 00:24:28,640 I know it's a mouthful. 725 00:24:28,640 --> 00:24:30,400 That one sounds even more complex. 726 00:24:30,400 --> 00:24:31,440 It is a little bit more involved. 727 00:24:31,440 --> 00:24:31,840 OK. 728 00:24:31,840 --> 00:24:35,000 But the basic idea is to encourage the AI 729 00:24:35,000 --> 00:24:38,760 to explore a wider range of possible solutions. 730 00:24:38,760 --> 00:24:41,600 Even those that might seem unconventional or less 731 00:24:41,600 --> 00:24:42,960 obvious at first. 732 00:24:42,960 --> 00:24:46,160 So it's kind of like telling the AI to think outside the box. 733 00:24:46,160 --> 00:24:46,800 Exactly. 734 00:24:46,800 --> 00:24:48,840 And this emphasis on diversity can often 735 00:24:48,840 --> 00:24:51,400 lead to more creative and innovative solutions. 736 00:24:51,400 --> 00:24:52,000 Interesting. 737 00:24:52,000 --> 00:24:54,440 Solutions that a more conventional AI might never 738 00:24:54,440 --> 00:24:55,600 even consider. 739 00:24:55,600 --> 00:24:57,880 It's like the difference between a group of people 740 00:24:57,880 --> 00:25:01,880 brainstorming ideas versus just one person trying 741 00:25:01,880 --> 00:25:03,400 to solve a problem alone. 742 00:25:03,400 --> 00:25:03,880 Right. 743 00:25:03,880 --> 00:25:05,760 And you know, the group is more likely to come up 744 00:25:05,760 --> 00:25:08,080 with more ideas because they have all those different 745 00:25:08,080 --> 00:25:09,560 perspectives and experiences. 746 00:25:09,560 --> 00:25:09,920 Right. 747 00:25:09,920 --> 00:25:12,640 And by encouraging this diversity 748 00:25:12,640 --> 00:25:15,560 in the AI's thought process, we're essentially 749 00:25:15,560 --> 00:25:18,840 tapping into that same power of collective intelligence. 750 00:25:18,840 --> 00:25:19,400 OK. 751 00:25:19,400 --> 00:25:21,400 Even though the collective in this case 752 00:25:21,400 --> 00:25:24,320 is just these different parts of the AI model itself. 753 00:25:24,320 --> 00:25:25,760 That's amazing. 754 00:25:25,760 --> 00:25:27,160 This is all so fascinating. 755 00:25:27,160 --> 00:25:32,120 So it seems like scaling test time compute 756 00:25:32,120 --> 00:25:35,360 combined with these innovative search techniques 757 00:25:35,360 --> 00:25:38,960 and the guidance of those process reward models 758 00:25:38,960 --> 00:25:42,840 is opening up a whole new world of possibilities for AI. 759 00:25:42,840 --> 00:25:43,440 I agree. 760 00:25:43,440 --> 00:25:45,680 It's a very exciting time to be working in this field. 761 00:25:45,680 --> 00:25:46,440 Yeah, it is. 762 00:25:46,440 --> 00:25:49,360 And it's a reminder that the future of AI 763 00:25:49,360 --> 00:25:50,560 isn't predetermined. 764 00:25:50,560 --> 00:25:51,000 Right. 765 00:25:51,000 --> 00:25:53,000 You know, it's something that we are actively 766 00:25:53,000 --> 00:25:55,800 shaping through our choices, our research, 767 00:25:55,800 --> 00:25:57,320 and our collective imagination. 768 00:25:57,320 --> 00:25:58,680 That's a great point. 769 00:25:58,680 --> 00:26:01,160 Well, on that inspiring note, I think 770 00:26:01,160 --> 00:26:04,200 it's time to wrap up our deep dive into the world of AI. 771 00:26:04,200 --> 00:26:05,200 Yeah, I think so too. 772 00:26:05,200 --> 00:26:06,640 We've covered so much. 773 00:26:06,640 --> 00:26:10,520 We've gone from the latest advancements in these AI 774 00:26:10,520 --> 00:26:14,920 models to the challenges of those hallucinations, 775 00:26:14,920 --> 00:26:17,200 the ethical considerations of AI, 776 00:26:17,200 --> 00:26:19,920 and its use of copyrighted material, 777 00:26:19,920 --> 00:26:22,120 and this really exciting potential 778 00:26:22,120 --> 00:26:25,000 of those smaller, more efficient AI models. 779 00:26:25,000 --> 00:26:26,640 Yeah, it's been a fantastic discussion. 780 00:26:26,640 --> 00:26:27,360 It has been. 781 00:26:27,360 --> 00:26:30,240 And I hope that our listeners have gained some valuable insights 782 00:26:30,240 --> 00:26:31,960 into this rapidly evolving field. 783 00:26:31,960 --> 00:26:33,080 Yeah, me too. 784 00:26:33,080 --> 00:26:35,360 And for all of you who are eager to keep exploring 785 00:26:35,360 --> 00:26:37,560 this really fascinating world, we 786 00:26:37,560 --> 00:26:40,440 encourage you to stay curious, stay engaged, 787 00:26:40,440 --> 00:26:42,360 and never stop asking questions. 788 00:26:42,360 --> 00:26:43,720 Yeah, that's great advice. 789 00:26:43,720 --> 00:26:46,600 You know, the future of AI is in our hands, 790 00:26:46,600 --> 00:26:49,440 and it's up to all of us to shape it responsibly. 791 00:26:49,440 --> 00:26:50,280 Well said. 792 00:26:50,280 --> 00:26:52,240 Thanks for joining us on this deep dive, everyone. 793 00:26:52,240 --> 00:27:01,880 Yeah, until next time.