1 00:00:00,000 --> 00:00:02,600 Welcome to the Daily AI News podcast. 2 00:00:02,600 --> 00:00:04,880 Get ready for another deep dive today. 3 00:00:04,880 --> 00:00:07,400 We've got a ton of really cool stories today 4 00:00:07,400 --> 00:00:10,880 from image generation to video understanding 5 00:00:10,880 --> 00:00:14,160 and even chat GPT's been doing some stuff. 6 00:00:14,160 --> 00:00:16,680 But I guess the big theme is how AI models 7 00:00:16,680 --> 00:00:18,800 are interacting with different kinds of data. 8 00:00:18,800 --> 00:00:19,680 Yeah, that's right. 9 00:00:19,680 --> 00:00:20,520 Like video. 10 00:00:20,520 --> 00:00:22,400 Yeah, it's always been pretty tricky for AI 11 00:00:22,400 --> 00:00:25,120 to really process and understand video content. 12 00:00:25,120 --> 00:00:27,680 So seeing progress in that area is really exciting. 13 00:00:27,680 --> 00:00:29,600 Okay, so let's start with this new research paper 14 00:00:29,600 --> 00:00:32,640 from Metta and Scanford about video understanding 15 00:00:32,640 --> 00:00:34,960 in large multimodal models. 16 00:00:34,960 --> 00:00:37,440 Right, so they basically dug into all these design 17 00:00:37,440 --> 00:00:38,400 and training factors. 18 00:00:38,400 --> 00:00:40,880 Their findings are like gold for optimizing 19 00:00:40,880 --> 00:00:42,920 how these models learn and do their thing. 20 00:00:42,920 --> 00:00:44,800 Okay, so one of their key discoveries 21 00:00:44,800 --> 00:00:47,160 is this scaling consistency. 22 00:00:47,160 --> 00:00:49,960 It's basically like you're trying out a cake recipe 23 00:00:49,960 --> 00:00:51,560 but you don't wanna waste a bunch of ingredients 24 00:00:51,560 --> 00:00:53,240 on a huge cake if it doesn't turn out. 25 00:00:53,240 --> 00:00:55,120 So you make a small test batch first. 26 00:00:55,120 --> 00:00:56,600 That's perfect, yeah. 27 00:00:56,600 --> 00:00:57,720 It's incredibly efficient. 28 00:00:57,720 --> 00:00:59,880 Like why waste time and resources 29 00:00:59,880 --> 00:01:03,160 on training these huge massive models? 30 00:01:03,160 --> 00:01:05,640 When you can figure things out with smaller models 31 00:01:05,640 --> 00:01:08,360 and have confidence, those findings 32 00:01:08,360 --> 00:01:10,080 will translate to the bigger ones. 33 00:01:10,080 --> 00:01:11,880 Yeah, it's like making sure your recipe works 34 00:01:11,880 --> 00:01:13,520 before you go all out. 35 00:01:13,520 --> 00:01:17,200 Okay, so in the paper, they introduced this family 36 00:01:17,200 --> 00:01:19,400 of video language models called Apollo. 37 00:01:19,400 --> 00:01:21,760 Apollo is pretty groundbreaking actually. 38 00:01:21,760 --> 00:01:24,480 It's a big jump forward in video understanding. 39 00:01:24,480 --> 00:01:25,960 Like setting new benchmarks 40 00:01:25,960 --> 00:01:27,800 for how well these models perform. 41 00:01:27,800 --> 00:01:31,440 So what does that actually mean for us non-AI experts? 42 00:01:31,440 --> 00:01:33,880 Well, think about searching for a specific moment 43 00:01:33,880 --> 00:01:36,400 in a video just by describing what happens. 44 00:01:36,400 --> 00:01:38,800 That's the kind of capability Apollo is unlocking. 45 00:01:38,800 --> 00:01:40,920 Oh, wow, so like if I'm looking for the part 46 00:01:40,920 --> 00:01:43,320 in a cooking video where they add the secret ingredient, 47 00:01:43,320 --> 00:01:45,160 I could just type in, show me 48 00:01:45,160 --> 00:01:46,600 when they add the secret ingredient. 49 00:01:46,600 --> 00:01:47,560 Exactly. 50 00:01:47,560 --> 00:01:48,520 Wow, that is super cool. 51 00:01:48,520 --> 00:01:50,200 So there were some pretty interesting details 52 00:01:50,200 --> 00:01:51,040 in the paper too. 53 00:01:51,040 --> 00:01:53,160 Yeah, lots of little things that make a big difference. 54 00:01:53,160 --> 00:01:55,080 Like they found that frames per second sampling 55 00:01:55,080 --> 00:01:56,800 is better than uniform sampling. 56 00:01:56,800 --> 00:01:58,880 And there's also this sweet spot 57 00:01:58,880 --> 00:02:01,680 for the number of tokens per frame, 58 00:02:01,680 --> 00:02:03,720 with 832 being optimal. 59 00:02:03,720 --> 00:02:04,680 Oh, interesting. 60 00:02:04,680 --> 00:02:07,400 And get this, just adding simple tokens 61 00:02:07,400 --> 00:02:10,560 between video tokens from different frames or clips 62 00:02:10,560 --> 00:02:12,560 is enough for efficient integration. 63 00:02:12,560 --> 00:02:14,560 Right, it's like giving the AI little hints 64 00:02:14,560 --> 00:02:15,800 to help it connect the dots. 65 00:02:15,800 --> 00:02:18,400 Yeah, and speaking of things that are coming out right now, 66 00:02:18,400 --> 00:02:20,680 Meta just announced a major update 67 00:02:20,680 --> 00:02:23,440 to their Ray-Ban Meta smart glasses. 68 00:02:23,440 --> 00:02:26,880 They're adding real-time AI video capabilities. 69 00:02:26,880 --> 00:02:27,720 That's pretty awesome. 70 00:02:27,720 --> 00:02:29,600 It's like seeing those cutting-edge research ideas 71 00:02:29,600 --> 00:02:31,000 coming to life right in front of us. 72 00:02:31,000 --> 00:02:32,640 Right, and it's happening so fast. 73 00:02:32,640 --> 00:02:35,080 So what can you actually do with these new glasses? 74 00:02:35,080 --> 00:02:39,080 Well, imagine having ongoing conversations with Meta AI, 75 00:02:39,080 --> 00:02:40,520 but through your glasses. 76 00:02:40,520 --> 00:02:41,360 Oh, wow. 77 00:02:41,360 --> 00:02:42,480 Like it can understand what you're seeing 78 00:02:42,480 --> 00:02:44,280 and provide real-time information. 79 00:02:44,280 --> 00:02:46,240 You can even get real-time translations. 80 00:02:46,240 --> 00:02:48,480 Oh my gosh, like if I was traveling in a foreign country, 81 00:02:48,480 --> 00:02:51,000 I could just look at a sign and it would translate it for me. 82 00:02:51,000 --> 00:02:53,080 Exactly, it's like having a superpower. 83 00:02:53,080 --> 00:02:54,200 It really is. 84 00:02:54,200 --> 00:02:55,480 Okay, let's switch gears a bit 85 00:02:55,480 --> 00:02:58,120 and talk about image and video generation. 86 00:02:58,120 --> 00:03:00,200 Google just released some major updates 87 00:03:00,200 --> 00:03:02,880 with VO2 and Image in Three, 88 00:03:02,880 --> 00:03:06,480 and both of them are achieving state-of-the-art results. 89 00:03:06,480 --> 00:03:08,160 Yeah, they're really pushing the boundaries 90 00:03:08,160 --> 00:03:09,560 of what's possible. 91 00:03:09,560 --> 00:03:11,560 So what does state-of-the-art actually mean? 92 00:03:11,560 --> 00:03:15,000 Well, it basically means these models are creating images 93 00:03:15,000 --> 00:03:17,760 and videos that are more realistic and detailed 94 00:03:17,760 --> 00:03:19,440 than anything we've seen before. 95 00:03:19,440 --> 00:03:20,800 So let's start with VO2. 96 00:03:20,800 --> 00:03:22,760 This was all about video generation, right? 97 00:03:22,760 --> 00:03:24,920 Yeah, and it's not just about creating videos, 98 00:03:24,920 --> 00:03:26,280 it's about understanding them. 99 00:03:26,280 --> 00:03:29,240 Like VO2 can actually understand physics 100 00:03:29,240 --> 00:03:30,480 and human movement. 101 00:03:30,480 --> 00:03:32,320 Oh wow, so it wouldn't create a video 102 00:03:32,320 --> 00:03:34,640 of someone walking through a wall. 103 00:03:34,640 --> 00:03:36,760 Right, it understands how things move 104 00:03:36,760 --> 00:03:38,400 and interact in the real world, 105 00:03:38,400 --> 00:03:40,600 and it can even respond to cinematic prompts 106 00:03:40,600 --> 00:03:42,560 like camera angles and lens types. 107 00:03:42,560 --> 00:03:44,880 Oh, so I could tell it to use a low-angle tracking shot. 108 00:03:44,880 --> 00:03:45,720 Yeah, exactly. 109 00:03:45,720 --> 00:03:48,360 It's like having a virtual film crew at your fingertips. 110 00:03:48,360 --> 00:03:52,040 That is wild, and it produces fewer unwanted details. 111 00:03:52,040 --> 00:03:55,640 Which we call hallucinations in the AI world. 112 00:03:55,640 --> 00:03:58,560 Right, so like if it was generating a video of a cat, 113 00:03:58,560 --> 00:04:00,920 it wouldn't accidentally give the cat 114 00:04:00,920 --> 00:04:02,240 like six legs or something. 115 00:04:02,240 --> 00:04:04,240 Exactly, those weird little glitches 116 00:04:04,240 --> 00:04:06,880 are becoming less and less common. 117 00:04:06,880 --> 00:04:08,280 Okay, so what about image in three? 118 00:04:08,280 --> 00:04:10,760 Image in three is all about image generation, 119 00:04:10,760 --> 00:04:12,960 but it's not just about making pretty pictures, 120 00:04:12,960 --> 00:04:14,920 it's about artistic expression. 121 00:04:14,920 --> 00:04:17,320 Like it can understand different art styles 122 00:04:17,320 --> 00:04:19,560 and render them with incredible accuracy. 123 00:04:19,560 --> 00:04:22,960 Oh, so like if I wanted a picture in the style of Van Gogh. 124 00:04:22,960 --> 00:04:24,880 You could tell image in three to do that, 125 00:04:24,880 --> 00:04:25,880 and it would create something 126 00:04:25,880 --> 00:04:27,920 that looks like a real Van Gogh painting. 127 00:04:27,920 --> 00:04:28,800 That's insane. 128 00:04:28,800 --> 00:04:31,120 And it's not just mimicking existing styles, 129 00:04:31,120 --> 00:04:32,600 it can also blend styles together 130 00:04:32,600 --> 00:04:34,240 and create something entirely new. 131 00:04:34,240 --> 00:04:35,720 So what are some examples of things 132 00:04:35,720 --> 00:04:36,920 Image in Three has created? 133 00:04:36,920 --> 00:04:40,920 It can render a winter wonderland scene with a red squirrel, 134 00:04:40,920 --> 00:04:43,760 or a hyper realistic image of a strawberry sculpted 135 00:04:43,760 --> 00:04:46,360 into a hummingbird, or even bring to life 136 00:04:46,360 --> 00:04:49,160 a foggy 1940s European train station. 137 00:04:49,160 --> 00:04:50,280 Oh wow, that's amazing. 138 00:04:50,280 --> 00:04:51,520 It's really pushing the boundaries 139 00:04:51,520 --> 00:04:53,640 of what we thought was possible with AI. 140 00:04:53,640 --> 00:04:55,080 Okay, so let's move on to a topic 141 00:04:55,080 --> 00:04:56,960 that's near and dear to my heart. 142 00:04:56,960 --> 00:04:58,720 Check GPT. 143 00:04:58,720 --> 00:05:00,200 It seems like it's becoming way more 144 00:05:00,200 --> 00:05:02,040 than just a simple chatbot. 145 00:05:02,040 --> 00:05:04,560 Yeah, OpenAI is definitely positioning it 146 00:05:04,560 --> 00:05:07,040 as a central hub for all sorts of tasks. 147 00:05:07,040 --> 00:05:09,000 Right, so it's not just about writing anymore. 148 00:05:09,000 --> 00:05:12,480 It's also for research, programming, even web search. 149 00:05:12,480 --> 00:05:14,280 It's like an everything app for AI. 150 00:05:14,280 --> 00:05:16,200 Yeah, that's a good way to put it. 151 00:05:16,200 --> 00:05:18,600 So one of the new features is called projects. 152 00:05:18,600 --> 00:05:21,200 Right, it's basically a way to organize your conversations 153 00:05:21,200 --> 00:05:22,520 within chat GPT. 154 00:05:22,520 --> 00:05:24,800 Like having different folders for your various projects. 155 00:05:24,800 --> 00:05:27,280 Yeah, it can be really helpful for staying organized, 156 00:05:27,280 --> 00:05:31,040 especially if you're using chat GPT for complex tasks 157 00:05:31,040 --> 00:05:32,880 or managing multiple things at once. 158 00:05:32,880 --> 00:05:34,640 Okay, that makes a lot of sense. 159 00:05:34,640 --> 00:05:36,400 And speaking of new features, 160 00:05:36,400 --> 00:05:39,520 chat GPT's AI search engine is finally rolling out 161 00:05:39,520 --> 00:05:40,520 to all users. 162 00:05:40,520 --> 00:05:41,720 Yeah, this is a pretty big deal. 163 00:05:41,720 --> 00:05:43,200 It's not just a regular search engine. 164 00:05:43,200 --> 00:05:46,680 It uses AI to understand what you're actually looking for. 165 00:05:46,680 --> 00:05:48,440 So it's like, it can read your mind. 166 00:05:48,440 --> 00:05:50,880 Well, not quite, but it gets pretty close. 167 00:05:50,880 --> 00:05:53,480 It can pick up on subtle cues and nuances 168 00:05:53,480 --> 00:05:56,120 in your search queries to deliver results 169 00:05:56,120 --> 00:05:58,960 that are more personalized and relevant to your needs. 170 00:05:58,960 --> 00:06:01,480 And I heard that the mobile version is really interesting. 171 00:06:01,480 --> 00:06:04,480 Yeah, they actually designed it to look and feel 172 00:06:04,480 --> 00:06:06,400 more like a traditional search engine. 173 00:06:06,400 --> 00:06:09,160 So it has things like images and ratings and... 174 00:06:09,160 --> 00:06:11,280 Maps and directions, all that stuff. 175 00:06:11,280 --> 00:06:12,920 Okay, so why would they do that? 176 00:06:12,920 --> 00:06:15,400 Well, a lot of people are used to traditional search engines. 177 00:06:15,400 --> 00:06:18,520 So it makes sense to create something that feels familiar. 178 00:06:18,520 --> 00:06:20,200 Right, it's like meeting people where they are. 179 00:06:20,200 --> 00:06:21,040 Exactly. 180 00:06:21,040 --> 00:06:22,960 Okay, now let's talk about notebook LM 181 00:06:22,960 --> 00:06:24,440 because there are some exciting updates 182 00:06:24,440 --> 00:06:25,960 to this tool as well. 183 00:06:25,960 --> 00:06:28,720 So remind me again, what is notebook LM? 184 00:06:28,720 --> 00:06:32,320 It's basically an AI-powered research notebook. 185 00:06:32,320 --> 00:06:33,400 So what does that mean? 186 00:06:33,400 --> 00:06:36,840 Well, you can upload documents, summarize key points, 187 00:06:36,840 --> 00:06:38,320 highlight important passages, 188 00:06:38,320 --> 00:06:41,680 and even generate audio overviews of your research material. 189 00:06:41,680 --> 00:06:42,520 Oh, wow. 190 00:06:42,520 --> 00:06:44,080 So it's like having a research assistant 191 00:06:44,080 --> 00:06:46,000 who can do all the heavy lifting for you. 192 00:06:46,000 --> 00:06:47,960 Yeah, that's a good way to think about it. 193 00:06:47,960 --> 00:06:49,560 And they've just released a premium version 194 00:06:49,560 --> 00:06:51,600 called Notebook LM Plus. 195 00:06:51,600 --> 00:06:53,400 Which comes with all sorts of extra features, 196 00:06:53,400 --> 00:06:56,560 like increased usage limits, customization options, 197 00:06:56,560 --> 00:06:58,720 shared team notebooks with analytics, 198 00:06:58,720 --> 00:07:01,160 and even enhanced privacy and security. 199 00:07:01,160 --> 00:07:04,240 Oh, wow, that sounds like a power user's dream. 200 00:07:04,240 --> 00:07:06,000 But wait, it gets even better. 201 00:07:06,000 --> 00:07:10,200 Notebook LM is now integrated with Gemini 2.0 Flash. 202 00:07:10,200 --> 00:07:11,480 That's a huge upgrade. 203 00:07:11,480 --> 00:07:13,160 So what does that mean for users? 204 00:07:13,160 --> 00:07:14,600 Well, during those audio overviews, 205 00:07:14,600 --> 00:07:17,240 you can now interact with AI hosts. 206 00:07:17,240 --> 00:07:19,560 Okay, so I'm listening to a summary of a research paper. 207 00:07:19,560 --> 00:07:22,920 And you can actually pause and ask the AI a question, 208 00:07:22,920 --> 00:07:25,040 like to clarify something or delve deeper 209 00:07:25,040 --> 00:07:26,400 into a specific topic. 210 00:07:26,400 --> 00:07:27,640 Oh, wow, that's amazing. 211 00:07:27,640 --> 00:07:28,760 It's like having a conversation 212 00:07:28,760 --> 00:07:30,480 with the experts who wrote the paper. 213 00:07:30,480 --> 00:07:31,760 That is so cool. 214 00:07:31,760 --> 00:07:33,840 Okay, but not all the news in the AI world 215 00:07:33,840 --> 00:07:35,760 is sunshine and roses. 216 00:07:35,760 --> 00:07:37,560 We need to talk about this legal battle 217 00:07:37,560 --> 00:07:39,920 that's brewing between Elon Musk and open AI. 218 00:07:39,920 --> 00:07:41,480 Yeah, this one's a bit messy. 219 00:07:41,480 --> 00:07:44,360 It's basically a clash of visions about the future of AI. 220 00:07:44,360 --> 00:07:46,920 So Elon Musk was one of the early investors 221 00:07:46,920 --> 00:07:49,360 and board members of open AI, right? 222 00:07:49,360 --> 00:07:50,960 And he's now claiming that the company 223 00:07:50,960 --> 00:07:52,840 has strayed from its original mission. 224 00:07:52,840 --> 00:07:55,840 Which was to develop AI safely and ethically 225 00:07:55,840 --> 00:07:57,400 for the benefit of humanity. 226 00:07:57,400 --> 00:08:00,240 And he's worried that open AI is now more focused 227 00:08:00,240 --> 00:08:02,680 on profits than on ethics. 228 00:08:02,680 --> 00:08:04,160 It's a tricky situation for sure. 229 00:08:04,160 --> 00:08:06,440 So what exactly is Musk alleging? 230 00:08:06,440 --> 00:08:09,280 Well, court filings reveal that Musk's original vision 231 00:08:09,280 --> 00:08:11,640 for open AI was as a nonprofit. 232 00:08:11,640 --> 00:08:14,640 He wanted it to be a force for good in the world. 233 00:08:14,640 --> 00:08:16,920 And he's concerned that the company's shift 234 00:08:16,920 --> 00:08:19,960 to a for-profit model has compromised that mission. 235 00:08:19,960 --> 00:08:21,840 Yeah, he's basically saying that open AI 236 00:08:21,840 --> 00:08:23,680 has betrayed its original principles. 237 00:08:23,680 --> 00:08:25,800 And there are even emails from 2017 238 00:08:25,800 --> 00:08:28,000 showing this tension between Musk and Altman 239 00:08:28,000 --> 00:08:30,640 and the other co-founders over control of the company. 240 00:08:30,640 --> 00:08:31,840 It was a real power struggle. 241 00:08:31,840 --> 00:08:34,400 So it's like a real life drama about the future of AI. 242 00:08:34,400 --> 00:08:35,240 It really is. 243 00:08:35,240 --> 00:08:36,800 It'll be interesting to see how this all plays out. 244 00:08:36,800 --> 00:08:37,800 Definitely. 245 00:08:37,800 --> 00:08:39,760 Okay, let's end on a slightly lighter note. 246 00:08:39,760 --> 00:08:43,680 Ilya Sutskever, a former chief scientist at Open AI, 247 00:08:43,680 --> 00:08:45,480 made a pretty interesting prediction recently. 248 00:08:45,480 --> 00:08:48,280 Yeah, he said that AI with reasoning power 249 00:08:48,280 --> 00:08:50,000 will become less predictable. 250 00:08:50,000 --> 00:08:51,240 Okay, so what does that mean? 251 00:08:51,240 --> 00:08:53,160 Well, as AI models get smarter 252 00:08:53,160 --> 00:08:55,480 and more capable of complex reasoning, 253 00:08:55,480 --> 00:08:58,080 they'll be able to consider millions of possibilities 254 00:08:58,080 --> 00:08:59,960 before making a decision, 255 00:08:59,960 --> 00:09:02,360 which means their actions will become less obvious 256 00:09:02,360 --> 00:09:04,240 and more surprising to us. 257 00:09:04,240 --> 00:09:06,440 So it's like they'll be able to think outside the box. 258 00:09:06,440 --> 00:09:07,280 Exactly. 259 00:09:07,280 --> 00:09:10,360 And that could lead to some really amazing breakthroughs. 260 00:09:10,360 --> 00:09:12,640 But it also raises questions about control 261 00:09:12,640 --> 00:09:16,040 and ensuring that AI remains aligned with human values. 262 00:09:16,040 --> 00:09:18,480 It's like the more intelligent AI becomes, 263 00:09:18,480 --> 00:09:20,280 the less we'll be able to predict what it will do. 264 00:09:20,280 --> 00:09:22,960 Yeah, it's both exciting and a little bit scary. 265 00:09:22,960 --> 00:09:24,880 It's definitely something to think about. 266 00:09:24,880 --> 00:09:26,360 Okay, so let's shift gears one more time 267 00:09:26,360 --> 00:09:30,200 and talk about this new research on best of end jailbreaking. 268 00:09:30,200 --> 00:09:33,240 Okay, so jailbreaking is basically a way to trick a chatbot 269 00:09:33,240 --> 00:09:35,760 into giving you answers it's not supposed to. 270 00:09:35,760 --> 00:09:37,960 Like bypassing the safety protocols. 271 00:09:37,960 --> 00:09:38,800 Exactly. 272 00:09:38,800 --> 00:09:41,040 So is this research saying that it's becoming easier 273 00:09:41,040 --> 00:09:42,920 to jailbreak these AI systems? 274 00:09:42,920 --> 00:09:45,240 Well, they found that the success rate of these attacks 275 00:09:45,240 --> 00:09:48,520 increases with the amount of computing power used. 276 00:09:48,520 --> 00:09:50,800 So basically if you have enough computers, 277 00:09:50,800 --> 00:09:54,280 you can eventually find a way to jailbreak any AI model. 278 00:09:54,280 --> 00:09:56,320 Oh, that's a little unsettling. 279 00:09:56,320 --> 00:09:58,800 Yeah, it's a reminder that we need to be constantly 280 00:09:58,800 --> 00:10:00,200 improving our defenses. 281 00:10:00,200 --> 00:10:01,320 Right, it's like an arms race. 282 00:10:01,320 --> 00:10:02,160 Exactly. 283 00:10:02,160 --> 00:10:04,720 Okay, but the good news is that the company 284 00:10:04,720 --> 00:10:07,360 behind this research, Anthropic, 285 00:10:07,360 --> 00:10:11,480 responsibly disclosed the vulnerability to other AI labs 286 00:10:11,480 --> 00:10:13,360 through the Frontier Model Forum. 287 00:10:13,360 --> 00:10:15,880 Yeah, that's really important for the AI safety community. 288 00:10:15,880 --> 00:10:17,080 We need to be working together 289 00:10:17,080 --> 00:10:18,840 to make these systems more secure. 290 00:10:18,840 --> 00:10:19,680 Totally. 291 00:10:19,680 --> 00:10:22,680 Okay, so for our last story in this part of our deep dive, 292 00:10:22,680 --> 00:10:27,000 let's zoom out and look at the big picture of AI investment. 293 00:10:27,000 --> 00:10:29,520 SoftBank CEO, Masayoshi Sun, 294 00:10:29,520 --> 00:10:32,160 just announced a massive investment plan. 295 00:10:32,160 --> 00:10:33,200 A hundred billion dollars. 296 00:10:33,200 --> 00:10:35,680 Yes, a hundred billion dollars over the next four years 297 00:10:35,680 --> 00:10:37,760 to support AI development in the US. 298 00:10:37,760 --> 00:10:39,360 That's a serious commitment. 299 00:10:39,360 --> 00:10:41,120 So what's behind this big move? 300 00:10:41,120 --> 00:10:43,720 Well, Sun has a history of making bold bets 301 00:10:43,720 --> 00:10:45,400 on emerging technologies, 302 00:10:45,400 --> 00:10:48,800 and he clearly sees AI as the next big thing. 303 00:10:48,800 --> 00:10:50,320 So where's all this money gonna go? 304 00:10:50,320 --> 00:10:52,560 He's focusing on building the core infrastructure 305 00:10:52,560 --> 00:10:55,040 that AI development needs, like data centers, 306 00:10:55,040 --> 00:10:57,000 semiconductor manufacturing, 307 00:10:57,000 --> 00:10:58,800 and even energy infrastructure. 308 00:10:58,800 --> 00:11:01,120 So it's not just about developing new AI models. 309 00:11:01,120 --> 00:11:02,960 It's about creating the ecosystem 310 00:11:02,960 --> 00:11:05,240 that will allow AI to flourish. 311 00:11:05,240 --> 00:11:08,560 And that's it for part one of our AI deep dive. 312 00:11:08,560 --> 00:11:10,000 We've covered a lot of ground today, 313 00:11:10,000 --> 00:11:12,640 from video understanding and image generation, 314 00:11:12,640 --> 00:11:15,240 to the evolving world of chat GPT, 315 00:11:15,240 --> 00:11:18,240 and the big picture trends in the AI landscape. 316 00:11:18,240 --> 00:11:20,560 Yeah, it's amazing how fast this field is moving. 317 00:11:20,560 --> 00:11:22,720 But there is still more to explore. 318 00:11:22,720 --> 00:11:24,480 Be sure to join us for part two, 319 00:11:24,480 --> 00:11:25,680 where we'll continue our journey 320 00:11:25,680 --> 00:11:27,400 through the fascinating world of AI. 321 00:11:27,400 --> 00:11:28,240 I can't wait. 322 00:11:28,240 --> 00:11:31,200 So we were talking about the ethical side of AI 323 00:11:31,200 --> 00:11:34,280 and those potential risks as AI gets more and more powerful. 324 00:11:34,280 --> 00:11:36,400 And this next story really dives into that. 325 00:11:36,400 --> 00:11:38,920 Okay, so this is about a paper exploring something 326 00:11:38,920 --> 00:11:40,720 called best of in jailbreaking, right? 327 00:11:40,720 --> 00:11:41,560 Yes, exactly. 328 00:11:41,560 --> 00:11:43,280 Jailbreaking might sound kind of scary, 329 00:11:43,280 --> 00:11:44,440 but in the AI world, 330 00:11:44,440 --> 00:11:47,200 it's basically about tricking those powerful models. 331 00:11:47,200 --> 00:11:49,280 Oh, so like bypassing those safety protocols 332 00:11:49,280 --> 00:11:50,120 they have in place. 333 00:11:50,120 --> 00:11:50,960 Exactly. 334 00:11:50,960 --> 00:11:53,080 And what's concerning is this research found 335 00:11:53,080 --> 00:11:55,960 that those attack success rates increased predictably 336 00:11:55,960 --> 00:11:57,920 with the amount of computing power used. 337 00:11:57,920 --> 00:12:00,560 So like the more resources an attacker has, 338 00:12:00,560 --> 00:12:03,040 the easier it is to find those jailbreaks. 339 00:12:03,040 --> 00:12:04,840 Right, and that highlights something 340 00:12:04,840 --> 00:12:06,480 we really need to be thinking about. 341 00:12:06,480 --> 00:12:07,840 We need strong defenses, 342 00:12:07,840 --> 00:12:09,320 and we need to make sure they can keep up 343 00:12:09,320 --> 00:12:11,280 with how fast AI is evolving. 344 00:12:11,280 --> 00:12:13,160 Yeah, it's like they're always one step ahead. 345 00:12:13,160 --> 00:12:14,000 Exactly. 346 00:12:14,000 --> 00:12:16,920 And that's why I was happy to see that Anthropic, 347 00:12:16,920 --> 00:12:18,640 the company behind this research, 348 00:12:18,640 --> 00:12:20,720 actually disclosed this vulnerability. 349 00:12:20,720 --> 00:12:21,560 Oh, that's good. 350 00:12:21,560 --> 00:12:23,400 They shared it with other AI labs 351 00:12:23,400 --> 00:12:25,280 through the Frontier Model Forum, right? 352 00:12:25,280 --> 00:12:27,240 Yeah, that kind of collaboration 353 00:12:27,240 --> 00:12:29,040 is so important in the AI community. 354 00:12:29,040 --> 00:12:30,600 It's like we all need to work together 355 00:12:30,600 --> 00:12:33,040 to make sure these systems are as secure as possible. 356 00:12:33,040 --> 00:12:34,760 Okay, so shifting gears a bit, 357 00:12:34,760 --> 00:12:37,720 let's talk about the impact of AI on investments. 358 00:12:37,720 --> 00:12:40,360 Right, SoftBank CEO Masayoshi Sun 359 00:12:40,360 --> 00:12:44,840 just announced this massive $100 billion investment plan. 360 00:12:44,840 --> 00:12:47,200 Wow, $100 billion? 361 00:12:47,200 --> 00:12:48,680 That's a huge number. 362 00:12:48,680 --> 00:12:49,720 What's the thinking behind that? 363 00:12:49,720 --> 00:12:52,440 Well, Sun's known for making big moves in the tech world, 364 00:12:52,440 --> 00:12:54,920 and this shows he's really serious about AI. 365 00:12:54,920 --> 00:12:56,520 So where is all that money going? 366 00:12:56,520 --> 00:12:59,440 He's focusing on the infrastructure AI needs to grow, 367 00:12:59,440 --> 00:13:00,840 things like data centers, 368 00:13:00,840 --> 00:13:03,360 semiconductor manufacturing, and even energy. 369 00:13:03,360 --> 00:13:05,400 So he's really laying the groundwork. 370 00:13:05,400 --> 00:13:06,600 It's a strategic move. 371 00:13:06,600 --> 00:13:09,080 It could really accelerate how AI is used 372 00:13:09,080 --> 00:13:10,520 in all sorts of industries. 373 00:13:10,520 --> 00:13:12,680 It's interesting to see these big investments 374 00:13:12,680 --> 00:13:14,400 shaping the future of AI. 375 00:13:14,400 --> 00:13:16,200 It really does bring up that question 376 00:13:16,200 --> 00:13:17,680 we keep coming back to. 377 00:13:17,680 --> 00:13:19,480 What does all this mean for society? 378 00:13:19,480 --> 00:13:20,560 Yeah, that's a big one. 379 00:13:20,560 --> 00:13:23,000 Okay, let's go back to ChadGPT for a minute. 380 00:13:23,000 --> 00:13:25,760 OpenAI just keeps adding new features, it seems like. 381 00:13:25,760 --> 00:13:28,560 They really want to make ChadGPT more user-friendly 382 00:13:28,560 --> 00:13:30,240 and accessible for everyone. 383 00:13:30,240 --> 00:13:32,320 And one of those updates is projects, right? 384 00:13:32,320 --> 00:13:34,640 Yeah, basically think of it as folders 385 00:13:34,640 --> 00:13:37,000 for your ChadGPT conversations. 386 00:13:37,000 --> 00:13:40,840 Okay, so I can organize my chats by topic or project or... 387 00:13:40,840 --> 00:13:42,880 Exactly, it can really help you stay organized, 388 00:13:42,880 --> 00:13:44,480 especially if you're using ChadGPT 389 00:13:44,480 --> 00:13:45,840 for lots of different things. 390 00:13:45,840 --> 00:13:46,680 Yeah, that makes sense. 391 00:13:46,680 --> 00:13:48,160 I'm always losing track of things. 392 00:13:48,160 --> 00:13:49,760 Well, projects might help with that. 393 00:13:49,760 --> 00:13:51,600 And speaking of new features, 394 00:13:51,600 --> 00:13:55,800 ChadGPT's AI search engine is available to everyone now. 395 00:13:55,800 --> 00:13:56,760 That's exciting news, 396 00:13:56,760 --> 00:13:59,320 this is a pretty big step forward in search. 397 00:13:59,320 --> 00:14:00,160 So what makes it different 398 00:14:00,160 --> 00:14:01,960 from those traditional search engines? 399 00:14:01,960 --> 00:14:05,440 Well, traditional search engines really rely on keywords, 400 00:14:05,440 --> 00:14:08,600 but ChadGPT uses AI to understand 401 00:14:08,600 --> 00:14:10,120 what you're actually looking for. 402 00:14:10,120 --> 00:14:12,080 So it can understand the context of my search. 403 00:14:12,080 --> 00:14:14,160 Exactly, it picks up on those nuances 404 00:14:14,160 --> 00:14:16,280 to give you more personalized results. 405 00:14:16,280 --> 00:14:17,400 And I heard the mobile version 406 00:14:17,400 --> 00:14:19,080 is designed to feel familiar. 407 00:14:19,080 --> 00:14:21,400 Yeah, it looks a lot like a traditional search engine 408 00:14:21,400 --> 00:14:23,400 with images, ratings, and... 409 00:14:23,400 --> 00:14:24,360 Maps and all that. 410 00:14:24,360 --> 00:14:26,680 Exactly, they wanna make it easy for people to use. 411 00:14:26,680 --> 00:14:28,520 Okay, so let's talk about Notebook LM too. 412 00:14:28,520 --> 00:14:30,320 There are some updates to that as well, right? 413 00:14:30,320 --> 00:14:32,240 Yes, Notebook LM is basically 414 00:14:32,240 --> 00:14:35,560 like the super powered research notebook. 415 00:14:35,560 --> 00:14:36,760 Remind me what it does again? 416 00:14:36,760 --> 00:14:38,320 Well, you can upload documents, 417 00:14:38,320 --> 00:14:39,960 summarize key points, 418 00:14:39,960 --> 00:14:41,760 highlight important passages, 419 00:14:41,760 --> 00:14:44,480 even generate audio overviews of your research. 420 00:14:44,480 --> 00:14:46,360 Oh, wow, so it's like having a research assistant. 421 00:14:46,360 --> 00:14:48,640 Exactly, and they just launched a premium version 422 00:14:48,640 --> 00:14:50,480 called Notebook LM Plus. 423 00:14:50,480 --> 00:14:51,320 Which has? 424 00:14:51,320 --> 00:14:54,440 Increased usage limits, customization options, 425 00:14:54,440 --> 00:14:56,520 shared notebooks for teams, 426 00:14:56,520 --> 00:14:58,680 and even better privacy and security. 427 00:14:58,680 --> 00:15:00,360 Wow, that sounds amazing. 428 00:15:00,360 --> 00:15:01,200 Yeah. 429 00:15:01,200 --> 00:15:02,040 But wait, there's more. 430 00:15:02,040 --> 00:15:04,800 It's now integrated with Gemini 2.0 Flash. 431 00:15:04,800 --> 00:15:07,280 Oh yeah, that integration is huge. 432 00:15:07,280 --> 00:15:08,600 So what does that mean for users? 433 00:15:08,600 --> 00:15:10,120 During those audio overviews, 434 00:15:10,120 --> 00:15:12,960 you can actually interact with AI hosts. 435 00:15:12,960 --> 00:15:15,800 No way, so I could be listening to a summary of a paper. 436 00:15:15,800 --> 00:15:17,200 And if you need clarification, 437 00:15:17,200 --> 00:15:19,200 you can actually ask the AI. 438 00:15:19,200 --> 00:15:20,240 That is incredible. 439 00:15:20,240 --> 00:15:21,560 It's like having a personal tutor. 440 00:15:21,560 --> 00:15:23,640 Yeah, it's really pushing the boundaries 441 00:15:23,640 --> 00:15:26,680 of how we learn and work with information. 442 00:15:26,680 --> 00:15:29,480 Okay, so now for a story that reminds us that 443 00:15:29,480 --> 00:15:32,720 even with all these advancements, AI isn't perfect. 444 00:15:32,720 --> 00:15:36,640 Apparently Apple's new AI notification feature 445 00:15:36,640 --> 00:15:37,560 messed up a little bit. 446 00:15:37,560 --> 00:15:39,960 Yeah, it generated this false headline 447 00:15:39,960 --> 00:15:41,240 about a murder case. 448 00:15:41,240 --> 00:15:42,960 Oh no, that's not good. 449 00:15:42,960 --> 00:15:43,800 What happened? 450 00:15:43,800 --> 00:15:46,040 So it was summarizing a news article from the BBC 451 00:15:46,040 --> 00:15:49,200 about the murder of this healthcare insurance CEO. 452 00:15:49,200 --> 00:15:50,800 But it got the facts wrong and said 453 00:15:50,800 --> 00:15:52,840 the suspect shot himself. 454 00:15:52,840 --> 00:15:55,760 Oh wow, so the AI actually presented misinformation. 455 00:15:55,760 --> 00:15:58,200 Yeah, and it's a good reminder that even though AI 456 00:15:58,200 --> 00:16:00,120 is getting better at summarizing information, 457 00:16:00,120 --> 00:16:01,880 it's still prone to errors. 458 00:16:01,880 --> 00:16:04,840 So we can't just blindly trust everything an AI tells us. 459 00:16:04,840 --> 00:16:07,320 Exactly, we still need to be critical thinkers. 460 00:16:07,320 --> 00:16:09,440 Okay, back to Meta for a second. 461 00:16:09,440 --> 00:16:10,880 It seems like they're everywhere. 462 00:16:10,880 --> 00:16:12,160 What's this new story about? 463 00:16:12,160 --> 00:16:14,840 They're updating their Ray-Ban Meta smart glasses 464 00:16:14,840 --> 00:16:18,240 to include real-time AI video capabilities. 465 00:16:18,240 --> 00:16:19,840 Wow, so what does that even mean? 466 00:16:19,840 --> 00:16:22,200 Think of it like having a personal AI assistant 467 00:16:22,200 --> 00:16:23,680 who can see what you're seeing. 468 00:16:23,680 --> 00:16:25,600 You can ask it questions about your surroundings, 469 00:16:25,600 --> 00:16:28,600 get real-time info, even have conversations translated. 470 00:16:28,600 --> 00:16:30,720 Oh my gosh, that's like straight out of a sci-fi movie. 471 00:16:30,720 --> 00:16:31,560 It really is. 472 00:16:31,560 --> 00:16:33,040 Imagine traveling in a foreign country 473 00:16:33,040 --> 00:16:35,440 and your glasses just translate everything for you. 474 00:16:35,440 --> 00:16:36,960 That would be amazing. 475 00:16:36,960 --> 00:16:38,960 But it also makes me think about privacy. 476 00:16:38,960 --> 00:16:41,120 Yeah, that's a valid concern what happens 477 00:16:41,120 --> 00:16:42,640 to all that data being collected. 478 00:16:42,640 --> 00:16:43,680 Exactly. 479 00:16:43,680 --> 00:16:46,400 Okay, let's switch gears again and talk about Google. 480 00:16:46,400 --> 00:16:48,720 They've been making waves with their image 481 00:16:48,720 --> 00:16:52,520 and video generation models, VO2 and Image in Three. 482 00:16:52,520 --> 00:16:54,160 VO2 is really impressive. 483 00:16:54,160 --> 00:16:55,960 It can create these lifelike videos 484 00:16:55,960 --> 00:16:58,920 that understand physics movement and facial expressions. 485 00:16:58,920 --> 00:17:01,760 And it can even respond to directions like camera angles. 486 00:17:01,760 --> 00:17:04,640 Exactly, it's like having a virtual film crew. 487 00:17:04,640 --> 00:17:05,720 That's insane. 488 00:17:05,720 --> 00:17:07,920 And Image in Three is equally impressive. 489 00:17:07,920 --> 00:17:10,240 It can generate super realistic images 490 00:17:10,240 --> 00:17:15,000 from landscapes to portraits to just imaginative scenes. 491 00:17:15,000 --> 00:17:17,840 It's like AI is making it hard to tell what's real and what's not. 492 00:17:17,840 --> 00:17:19,400 It really does raise those questions 493 00:17:19,400 --> 00:17:22,640 about authenticity and ownership, especially with DeepFakes. 494 00:17:22,640 --> 00:17:23,680 Yeah, it's a double-edged sword. 495 00:17:23,680 --> 00:17:26,320 For sure, but these are the conversations we need to have. 496 00:17:26,320 --> 00:17:28,680 We need to be thoughtful about how we use AI. 497 00:17:28,680 --> 00:17:29,400 That's a great point. 498 00:17:29,400 --> 00:17:31,000 We need to make sure we're using it for good. 499 00:17:31,000 --> 00:17:32,040 Exactly. 500 00:17:32,040 --> 00:17:35,680 Okay, so that brings us to the end of part two of our deep dive. 501 00:17:35,680 --> 00:17:38,880 We've covered a lot of ground from the ethics of AI 502 00:17:38,880 --> 00:17:40,920 to those incredible advancements. 503 00:17:40,920 --> 00:17:42,920 And we've still got more to explore. 504 00:17:42,920 --> 00:17:44,680 So be sure to join us for part three 505 00:17:44,680 --> 00:17:47,920 where we'll dive into some more fascinating developments. 506 00:17:47,920 --> 00:17:50,960 Okay, so we're back for the final part of our deep dive. 507 00:17:50,960 --> 00:17:52,160 And we're kicking things off 508 00:17:52,160 --> 00:17:54,400 with a pretty big announcement from Meta. 509 00:17:54,400 --> 00:17:56,880 They've just revealed this new AI architecture 510 00:17:56,880 --> 00:18:01,440 called the Byte-Latent Transformer, or BLT for short. 511 00:18:01,440 --> 00:18:03,760 Yeah, BLT is definitely making some waves. 512 00:18:03,760 --> 00:18:06,000 So with all these new models coming out all the time, 513 00:18:06,000 --> 00:18:07,400 what makes this one so special? 514 00:18:07,400 --> 00:18:10,280 Well, BLT is a pretty fundamental shift 515 00:18:10,280 --> 00:18:13,760 in how AI models process information. 516 00:18:13,760 --> 00:18:15,840 Okay, I'm intrigued. 517 00:18:15,840 --> 00:18:16,680 Tell me more. 518 00:18:16,680 --> 00:18:19,520 Most AI models, they break down text into words 519 00:18:19,520 --> 00:18:21,680 or even parts of words before they can understand it. 520 00:18:21,680 --> 00:18:22,920 It's called tokenization. 521 00:18:22,920 --> 00:18:23,920 Yeah, I've heard of that. 522 00:18:23,920 --> 00:18:26,000 But BLT skips that whole step. 523 00:18:26,000 --> 00:18:28,480 So it works with the raw data directly. 524 00:18:28,480 --> 00:18:30,240 Exactly, it works with bytes, 525 00:18:30,240 --> 00:18:33,840 the fundamental building blocks of digital information. 526 00:18:33,840 --> 00:18:36,240 Okay, but how does that actually change things? 527 00:18:36,240 --> 00:18:37,280 Well, for one thing, 528 00:18:37,280 --> 00:18:40,200 it could make AI models much more efficient. 529 00:18:40,200 --> 00:18:42,880 They can process information faster 530 00:18:42,880 --> 00:18:44,760 and use less computing power. 531 00:18:44,760 --> 00:18:45,600 That makes sense. 532 00:18:45,600 --> 00:18:48,360 Cutting out that extra step must save a lot of resources. 533 00:18:48,360 --> 00:18:50,000 And it goes beyond efficiency too. 534 00:18:50,000 --> 00:18:52,200 Working with bytes could make AI models 535 00:18:52,200 --> 00:18:53,880 more adaptable and versatile. 536 00:18:53,880 --> 00:18:56,400 So they can handle different types of data more easily. 537 00:18:56,400 --> 00:19:00,080 Exactly, not just text, but images, video, 538 00:19:00,080 --> 00:19:02,160 maybe even music or code. 539 00:19:02,160 --> 00:19:05,320 Wow, so it's like giving AI a universal language. 540 00:19:05,320 --> 00:19:06,160 That's the idea. 541 00:19:06,160 --> 00:19:08,360 BLT is still in its early stages, 542 00:19:08,360 --> 00:19:10,320 but it's incredibly promising. 543 00:19:10,320 --> 00:19:11,320 It sounds like it, yeah, 544 00:19:11,320 --> 00:19:13,320 we'll definitely be keeping an eye on that one. 545 00:19:13,320 --> 00:19:14,600 Okay, so switching gears a bit. 546 00:19:14,600 --> 00:19:17,520 Google has this new AI experiment called WISC. 547 00:19:17,520 --> 00:19:18,360 What's that all about? 548 00:19:18,360 --> 00:19:19,880 WISC is all about creativity. 549 00:19:19,880 --> 00:19:22,400 It's using AI to blend images together. 550 00:19:22,400 --> 00:19:24,640 So instead of text prompts, you start with images. 551 00:19:24,640 --> 00:19:27,280 Exactly, you can upload your own pictures 552 00:19:27,280 --> 00:19:29,560 or choose from a library of existing ones. 553 00:19:29,560 --> 00:19:31,920 And then you use WISC to blend them together 554 00:19:31,920 --> 00:19:33,440 and create something totally new. 555 00:19:33,440 --> 00:19:36,200 It's like a digital art studio powered by AI. 556 00:19:36,200 --> 00:19:39,360 Yeah, and you can explore different styles, textures, 557 00:19:39,360 --> 00:19:41,320 combine elements from different pictures. 558 00:19:41,320 --> 00:19:42,240 It's really cool. 559 00:19:42,240 --> 00:19:43,760 So how does it actually work? 560 00:19:43,760 --> 00:19:47,400 It uses two of Google's most advanced AI models, 561 00:19:47,400 --> 00:19:49,280 Image in Three and Geminer. 562 00:19:49,280 --> 00:19:50,640 Okay, so those two are working together? 563 00:19:50,640 --> 00:19:53,040 Yeah, Gemini analyzes the images you give it 564 00:19:53,040 --> 00:19:55,280 and creates descriptions of what it sees. 565 00:19:55,280 --> 00:19:58,280 Then Image in Three uses those descriptions as prompts 566 00:19:58,280 --> 00:20:00,400 to create the final blended image. 567 00:20:00,400 --> 00:20:03,120 So it's like two AI models collaborating. 568 00:20:03,120 --> 00:20:05,840 Exactly, and that's what's so exciting about this field, right? 569 00:20:05,840 --> 00:20:08,120 We're seeing all these new ways AI can be used 570 00:20:08,120 --> 00:20:10,560 to enhance creativity and solve problems. 571 00:20:10,560 --> 00:20:11,760 It's mind blowing, really. 572 00:20:11,760 --> 00:20:14,400 Well, we've covered a lot of ground in this deep dive 573 00:20:14,400 --> 00:20:17,200 from video understanding and image generation 574 00:20:17,200 --> 00:20:21,560 to ethical considerations and these big investment trends. 575 00:20:21,560 --> 00:20:22,720 Then everything in between. 576 00:20:22,720 --> 00:20:25,320 It's amazing to see how quickly this field is evolving. 577 00:20:25,320 --> 00:20:27,920 It really is, and it's important to stay informed 578 00:20:27,920 --> 00:20:29,400 and engaged with these developments. 579 00:20:29,400 --> 00:20:32,000 Absolutely, because AI is gonna touch every aspect 580 00:20:32,000 --> 00:20:33,480 of our lives in the future. 581 00:20:33,480 --> 00:20:34,920 It already is, in many ways. 582 00:20:34,920 --> 00:20:37,400 It's definitely a fascinating and complex field. 583 00:20:37,400 --> 00:20:39,080 And it's only gonna get more so. 584 00:20:39,080 --> 00:20:41,400 Well, that brings us to the end of our deep dive today. 585 00:20:41,400 --> 00:20:44,200 It's been a pleasure exploring these topics with you. 586 00:20:44,200 --> 00:20:47,000 Thanks for listening to the Daily AI News podcast, 587 00:20:47,000 --> 00:20:48,400 and be sure to join us next time 588 00:20:48,400 --> 00:21:12,400 for another deep dive into the world of AI.