1 00:00:00,000 --> 00:00:02,600 Welcome to the Daily AI News podcast. 2 00:00:02,600 --> 00:00:06,480 We've got a really fascinating collection of AI news today. 3 00:00:06,480 --> 00:00:08,960 From a whole new way to work with ChatGPT 4 00:00:08,960 --> 00:00:12,420 to a possible Russian influence operation, 5 00:00:12,420 --> 00:00:14,520 the used AI voice generation. 6 00:00:14,520 --> 00:00:15,480 Yeah, it's really interesting. 7 00:00:15,480 --> 00:00:16,580 All of these different stories 8 00:00:16,580 --> 00:00:18,560 kind of point to this underlying trend. 9 00:00:18,560 --> 00:00:20,480 AI is just evolving so rapidly 10 00:00:20,480 --> 00:00:22,600 and you can see the impact everywhere. 11 00:00:22,600 --> 00:00:24,280 Okay, let's break it down. 12 00:00:24,280 --> 00:00:25,920 First up, we have open AI. 13 00:00:25,920 --> 00:00:27,000 They've been busy. 14 00:00:27,000 --> 00:00:28,360 Not only are they making headlines 15 00:00:28,360 --> 00:00:31,320 because of a potential cloud computing power play, 16 00:00:31,320 --> 00:00:35,160 but they also just launched something new called Canvas. 17 00:00:35,160 --> 00:00:36,520 And it's not just a small update. 18 00:00:36,520 --> 00:00:39,680 It's like a completely new way to use ChatGPT. 19 00:00:39,680 --> 00:00:41,480 Yeah, Canvas is a major departure 20 00:00:41,480 --> 00:00:42,960 from that typical chat window. 21 00:00:42,960 --> 00:00:45,360 It's a much more visual and interactive way 22 00:00:45,360 --> 00:00:46,560 to use ChatGPT. 23 00:00:46,560 --> 00:00:47,380 It's okay. 24 00:00:47,380 --> 00:00:49,760 You can think of it like a digital whiteboard 25 00:00:49,760 --> 00:00:52,120 where you can work right alongside ChatGPT. 26 00:00:52,120 --> 00:00:52,960 Oh, cool. 27 00:00:52,960 --> 00:00:54,960 So you can edit text, write code, 28 00:00:54,960 --> 00:00:56,780 and you can even get contextual feedback 29 00:00:56,780 --> 00:00:59,720 and suggestions from ChatGPT all in real time. 30 00:00:59,720 --> 00:01:02,240 So it's kind of like having a super smart writing partner 31 00:01:02,240 --> 00:01:04,360 or a coding buddy helping you out. 32 00:01:04,360 --> 00:01:06,240 And I've heard they even have some new shortcuts. 33 00:01:06,240 --> 00:01:07,080 That's right. 34 00:01:07,080 --> 00:01:09,360 You can ask ChatGPT to change the length 35 00:01:09,360 --> 00:01:11,280 of what you're writing, debug your code, 36 00:01:11,280 --> 00:01:13,000 even add emojis to lighten the mood. 37 00:01:13,000 --> 00:01:14,360 Yeah, and it's really impressive 38 00:01:14,360 --> 00:01:17,600 how they train the model to be such a good collaborator. 39 00:01:17,600 --> 00:01:20,520 Open AI actually used synthetic data 40 00:01:20,520 --> 00:01:22,880 to teach it when to offer suggestions 41 00:01:22,880 --> 00:01:25,600 and how to interact in a natural and helpful way. 42 00:01:25,600 --> 00:01:26,840 It makes you wonder like, 43 00:01:26,840 --> 00:01:29,520 what could you do with Canvas as your AI sidekick? 44 00:01:29,520 --> 00:01:30,360 Right. 45 00:01:30,360 --> 00:01:31,440 Drafting a blog post 46 00:01:31,440 --> 00:01:34,040 or maybe tackling a challenging coding project, 47 00:01:34,040 --> 00:01:36,360 the possibilities seem really endless. 48 00:01:36,360 --> 00:01:37,360 Yeah, absolutely. 49 00:01:37,360 --> 00:01:38,840 I mean, this could completely change 50 00:01:38,840 --> 00:01:41,000 how we do creative and technical work. 51 00:01:41,000 --> 00:01:42,840 So continuing on this open AI theme, 52 00:01:42,840 --> 00:01:46,240 we've got some news from the world of cloud computing. 53 00:01:46,240 --> 00:01:49,000 Apparently Google has approached the FTC 54 00:01:49,000 --> 00:01:51,800 because they're concerned about Microsoft's deal 55 00:01:51,800 --> 00:01:53,960 to host open AI's technology 56 00:01:53,960 --> 00:01:56,040 exclusively on their cloud servers. 57 00:01:56,040 --> 00:01:57,400 Yeah, this is a pretty big deal. 58 00:01:57,400 --> 00:01:59,640 And it's easy to see why Google's concerned. 59 00:01:59,640 --> 00:02:02,240 Open AI's technology is like a total game changer. 60 00:02:02,240 --> 00:02:05,000 And because Microsoft has this exclusive access, 61 00:02:05,000 --> 00:02:06,520 it gives them a huge advantage 62 00:02:06,520 --> 00:02:08,040 in the cloud computing market. 63 00:02:08,040 --> 00:02:10,000 So naturally this is sparked concerns 64 00:02:10,000 --> 00:02:12,680 from competitors like Google and Amazon 65 00:02:12,680 --> 00:02:15,680 about like what could happen to competition 66 00:02:15,680 --> 00:02:17,920 and pricing in the cloud market. 67 00:02:17,920 --> 00:02:19,920 Yeah, it'll be interesting to see how this all plays out. 68 00:02:19,920 --> 00:02:21,520 Yeah, so if you were a business 69 00:02:21,520 --> 00:02:23,640 that uses cloud services, 70 00:02:23,640 --> 00:02:25,800 you'd definitely be paying attention to this. 71 00:02:25,800 --> 00:02:27,400 This really does raise some big questions 72 00:02:27,400 --> 00:02:31,680 about like the future of how we access AI. 73 00:02:31,680 --> 00:02:34,600 Will we see a handful of these big tech companies 74 00:02:34,600 --> 00:02:37,200 control the most powerful AI tools? 75 00:02:37,200 --> 00:02:39,920 Or will it be more open and collaborative? 76 00:02:39,920 --> 00:02:40,960 Yeah, it's a big question, 77 00:02:40,960 --> 00:02:43,280 one that will probably unfold over the next few years. 78 00:02:43,280 --> 00:02:44,240 Definitely. 79 00:02:44,240 --> 00:02:46,280 Now, shifting focus for a second, 80 00:02:46,280 --> 00:02:47,480 let's talk about service now. 81 00:02:47,480 --> 00:02:48,800 They've made a big contribution 82 00:02:48,800 --> 00:02:50,160 to the world of AI training. 83 00:02:50,160 --> 00:02:53,120 They've open sourced something called FastLLM, 84 00:02:53,120 --> 00:02:55,040 which is a framework that promises 85 00:02:55,040 --> 00:02:57,720 to speed up AI model training by 20%. 86 00:02:57,720 --> 00:02:58,560 Which is huge. 87 00:02:58,560 --> 00:02:59,400 Huge. 88 00:02:59,400 --> 00:03:01,240 AI training is really resource intensive, 89 00:03:01,240 --> 00:03:03,200 both in terms of time and money. 90 00:03:03,200 --> 00:03:06,040 So any innovation that can speed up the process 91 00:03:06,040 --> 00:03:08,640 is potentially a total game changer. 92 00:03:08,640 --> 00:03:09,480 Think about it, 93 00:03:09,480 --> 00:03:12,120 a model that would have taken a month to train 94 00:03:12,120 --> 00:03:14,720 could now be done in just three weeks. 95 00:03:14,720 --> 00:03:17,440 This is a big deal for companies and researchers 96 00:03:17,440 --> 00:03:21,520 who are trying to develop and use AI solutions faster. 97 00:03:21,520 --> 00:03:24,200 So what's the secret to FastLLM? 98 00:03:24,200 --> 00:03:26,840 How did they get such a big speed boost? 99 00:03:26,840 --> 00:03:29,560 Well, it's based on two main innovations. 100 00:03:29,560 --> 00:03:30,800 First, there's something called 101 00:03:30,800 --> 00:03:33,000 breadth first pipeline parallelism, 102 00:03:33,000 --> 00:03:35,240 which is basically a really smart way 103 00:03:35,240 --> 00:03:38,120 to distribute the work across multiple GPUs. 104 00:03:38,120 --> 00:03:39,680 So they parallelize the training. 105 00:03:39,680 --> 00:03:40,520 Exactly. 106 00:03:40,520 --> 00:03:43,760 And the second innovation is all about memory fragmentation. 107 00:03:43,760 --> 00:03:45,440 When you train a big AI model, 108 00:03:45,440 --> 00:03:48,760 the memory gets fragmented and that slows things down. 109 00:03:48,760 --> 00:03:51,680 FastLLM has this system to manage the memory 110 00:03:51,680 --> 00:03:53,440 and minimize the fragmentation, 111 00:03:53,440 --> 00:03:56,200 which makes the whole training process faster and smoother. 112 00:03:56,200 --> 00:03:57,920 It's like tidying up the computer's workspace 113 00:03:57,920 --> 00:03:58,840 to make it work better. 114 00:03:58,840 --> 00:03:59,960 That's a great way to put it. 115 00:03:59,960 --> 00:04:02,760 So how could this faster training actually impact 116 00:04:02,760 --> 00:04:06,840 like the development of new AI tools and technologies? 117 00:04:06,840 --> 00:04:09,760 Well, I mean, the possibilities are really endless. 118 00:04:09,760 --> 00:04:12,160 Faster training could mean breakthroughs in everything 119 00:04:12,160 --> 00:04:15,560 from medicine to robotics, even climate modeling. 120 00:04:15,560 --> 00:04:18,240 You could try out new ideas faster, experiment more, 121 00:04:18,240 --> 00:04:21,240 and ultimately build AI systems that are smarter 122 00:04:21,240 --> 00:04:23,480 and more capable than ever before. 123 00:04:23,480 --> 00:04:26,480 So it's almost like we're entering a new era of AI 124 00:04:26,480 --> 00:04:29,240 where the pace of innovation is gonna be even faster. 125 00:04:29,240 --> 00:04:31,080 Yeah, I think that's pretty accurate. 126 00:04:31,080 --> 00:04:32,600 Now we've gotta talk about a topic 127 00:04:32,600 --> 00:04:35,240 that's been getting a lot of attention lately. 128 00:04:35,240 --> 00:04:36,440 Bias in AI. 129 00:04:36,440 --> 00:04:37,280 Yeah. 130 00:04:37,280 --> 00:04:39,560 Researchers at MIT have come up with a new approach 131 00:04:39,560 --> 00:04:43,040 that they say can reduce bias in AI models 132 00:04:43,040 --> 00:04:44,760 without losing any accuracy. 133 00:04:44,760 --> 00:04:47,120 And that's a big deal because there's been a lot of concern 134 00:04:47,120 --> 00:04:51,120 about AI bias, especially in things like facial recognition 135 00:04:51,120 --> 00:04:54,200 or loan applications, where a biased algorithm 136 00:04:54,200 --> 00:04:56,040 can have some serious consequences. 137 00:04:56,040 --> 00:04:56,880 Absolutely. 138 00:04:56,880 --> 00:04:58,360 So how does this new method actually work? 139 00:04:58,360 --> 00:05:00,280 So instead of just trying to balance the data, 140 00:05:00,280 --> 00:05:02,800 which can actually make the model perform worse, 141 00:05:02,800 --> 00:05:05,840 this technique focuses on finding and removing 142 00:05:05,840 --> 00:05:08,640 the specific data points that are contributing to bias 143 00:05:08,640 --> 00:05:10,240 against minority groups. 144 00:05:10,240 --> 00:05:11,800 So it's a more targeted approach 145 00:05:11,800 --> 00:05:14,840 to removing the sources of bias 146 00:05:14,840 --> 00:05:17,280 without losing the valuable information in the data. 147 00:05:17,280 --> 00:05:18,120 Exactly. 148 00:05:18,120 --> 00:05:21,280 It's a more nuanced way to address the problem of AI bias. 149 00:05:21,280 --> 00:05:23,000 And this could be a real game changer, 150 00:05:23,000 --> 00:05:25,000 especially in fields like healthcare. 151 00:05:25,000 --> 00:05:28,240 Imagine if a biased AI made a wrong diagnosis. 152 00:05:28,240 --> 00:05:30,280 Yeah, that could have serious consequences. 153 00:05:30,280 --> 00:05:33,160 Think about a world where AI diagnostic tools 154 00:05:33,160 --> 00:05:36,000 are not only accurate, but also fair. 155 00:05:36,000 --> 00:05:38,320 So everyone gets the same quality of care 156 00:05:38,320 --> 00:05:39,760 regardless of their background. 157 00:05:39,760 --> 00:05:40,920 That's a powerful vision. 158 00:05:40,920 --> 00:05:42,480 And it's definitely something to aim for 159 00:05:42,480 --> 00:05:45,680 as we build AI systems for more and more sensitive fields. 160 00:05:45,680 --> 00:05:47,040 I completely agree. 161 00:05:47,040 --> 00:05:48,400 Okay, let's dive into a study 162 00:05:48,400 --> 00:05:50,920 that looks at political bias in language models, 163 00:05:50,920 --> 00:05:52,760 which is a pretty sensitive area. 164 00:05:52,760 --> 00:05:54,520 This research comes from MIT Center 165 00:05:54,520 --> 00:05:55,960 for Constructive Communication, 166 00:05:55,960 --> 00:05:57,760 and they found some really interesting stuff. 167 00:05:57,760 --> 00:05:58,960 It's fascinating, right? 168 00:05:58,960 --> 00:06:00,920 This study raises some important questions 169 00:06:00,920 --> 00:06:03,560 about the biases that can creep into AI 170 00:06:03,560 --> 00:06:06,360 even when you try to train it on objective data. 171 00:06:06,360 --> 00:06:09,280 What they found was that even when you expose these models 172 00:06:09,280 --> 00:06:12,960 to factual information, they consistently showed 173 00:06:12,960 --> 00:06:16,760 a left-leaning bias, and that bias actually gets stronger 174 00:06:16,760 --> 00:06:19,640 as the models become larger and more complex. 175 00:06:19,640 --> 00:06:22,480 It's kind of like even if you feed the model, 176 00:06:22,480 --> 00:06:25,800 balanced factual data, it still ends up 177 00:06:25,800 --> 00:06:27,760 leaning a certain way politically. 178 00:06:27,760 --> 00:06:30,280 It makes you wonder, where does that bias come from? 179 00:06:30,280 --> 00:06:31,440 Is it in the data itself, 180 00:06:31,440 --> 00:06:33,040 or is it somehow built into the model? 181 00:06:33,040 --> 00:06:34,920 That's the big question, isn't it? 182 00:06:34,920 --> 00:06:36,920 The researchers think the bias might actually be coming 183 00:06:36,920 --> 00:06:39,840 from the huge datasets used to train these models. 184 00:06:39,840 --> 00:06:42,240 Those datasets often reflect the dominant cultural 185 00:06:42,240 --> 00:06:44,240 and political ideas in society. 186 00:06:44,240 --> 00:06:47,680 So it's like the model absorbs the biases of the real world 187 00:06:47,680 --> 00:06:50,080 even if you're trying to give it objective information. 188 00:06:50,080 --> 00:06:51,960 Yeah, which is kind of a sobering thought. 189 00:06:51,960 --> 00:06:54,360 It shows how hard it is to separate fact from opinion, 190 00:06:54,360 --> 00:06:57,440 especially today, when information spreads so fast. 191 00:06:57,440 --> 00:07:00,160 This really underscores the need for transparency 192 00:07:00,160 --> 00:07:04,200 and careful examination when it comes to developing AI. 193 00:07:04,200 --> 00:07:08,000 If we want AI systems that are fair and unbiased, 194 00:07:08,000 --> 00:07:09,920 we have to understand how they're trained 195 00:07:09,920 --> 00:07:11,200 and what data they're using. 196 00:07:11,200 --> 00:07:13,480 I agree, transparency is so crucial. 197 00:07:13,480 --> 00:07:15,520 We need to be able to see what's going on 198 00:07:15,520 --> 00:07:17,960 inside these systems if we're going to trust them 199 00:07:17,960 --> 00:07:19,600 to make important decisions. 200 00:07:19,600 --> 00:07:22,480 Now for a story that really shows the potential dangers 201 00:07:22,480 --> 00:07:26,000 of AI being misused, let's talk about 11 Labs. 202 00:07:26,000 --> 00:07:29,000 Their AI voice generation technology may have been used 203 00:07:29,000 --> 00:07:32,080 in what's suspected to be a Russian influence operation. 204 00:07:32,080 --> 00:07:33,880 This is a really disturbing example 205 00:07:33,880 --> 00:07:36,080 of how AI can be weaponized. 206 00:07:36,080 --> 00:07:38,240 According to the report from Recorded Future, 207 00:07:38,240 --> 00:07:40,920 this campaign was designed to weaken European support 208 00:07:40,920 --> 00:07:43,480 for Ukraine by spreading fake news videos 209 00:07:43,480 --> 00:07:45,360 that used AI-generated voiceovers 210 00:07:45,360 --> 00:07:47,320 in different European languages. 211 00:07:47,320 --> 00:07:49,840 It's a reminder that as AI gets more powerful 212 00:07:49,840 --> 00:07:51,480 and more people have access to it, 213 00:07:51,480 --> 00:07:54,120 the potential for misuse increases as well. 214 00:07:54,120 --> 00:07:55,360 It also makes you wonder, 215 00:07:55,360 --> 00:07:57,360 how do we even detect these deep fakes? 216 00:07:57,360 --> 00:07:59,520 How can we tell what's real and what's fake? 217 00:07:59,520 --> 00:08:01,160 It's getting harder and harder. 218 00:08:01,160 --> 00:08:03,560 The technology is so advanced 219 00:08:03,560 --> 00:08:07,080 and these deep fakes are incredibly realistic. 220 00:08:07,080 --> 00:08:08,600 It's a real challenge for anyone 221 00:08:08,600 --> 00:08:12,000 who's trying to figure out what's true, especially online. 222 00:08:12,000 --> 00:08:14,680 It's like we can't trust our own eyes and ears anymore. 223 00:08:14,680 --> 00:08:19,560 And that has huge implications for journalism, politics, 224 00:08:19,560 --> 00:08:21,200 even just our personal relationships. 225 00:08:21,200 --> 00:08:22,760 It's definitely a cause for concern. 226 00:08:22,760 --> 00:08:24,960 Just imagine getting a voicemail from a loved one 227 00:08:24,960 --> 00:08:26,960 only to find out later that it was a deep fake 228 00:08:26,960 --> 00:08:29,040 created by someone with bad intentions. 229 00:08:29,040 --> 00:08:30,640 So what can we do? 230 00:08:30,640 --> 00:08:33,480 Can we stay ahead of the people who are using AI 231 00:08:33,480 --> 00:08:35,360 to deceive and manipulate? 232 00:08:35,360 --> 00:08:37,960 It's a tough challenge, but we can't just give up. 233 00:08:37,960 --> 00:08:39,920 We need to tackle it from different angles, 234 00:08:39,920 --> 00:08:42,600 technology, education, and regulation. 235 00:08:42,600 --> 00:08:43,880 Could you explain that a bit more? 236 00:08:43,880 --> 00:08:45,840 What kind of solutions are you thinking of? 237 00:08:45,840 --> 00:08:47,360 Well, on the technology side, 238 00:08:47,360 --> 00:08:49,920 we need to invest a lot in research and development 239 00:08:49,920 --> 00:08:51,920 for tools that can detect deep fakes. 240 00:08:51,920 --> 00:08:54,120 There's a lot of exciting work being done in this area, 241 00:08:54,120 --> 00:08:56,640 looking at how to analyze audio and video 242 00:08:56,640 --> 00:08:58,160 for signs of tampering. 243 00:08:58,160 --> 00:08:59,160 So it's an arms race. 244 00:08:59,160 --> 00:09:01,840 As the deep fakes get better, so do the detectors. 245 00:09:01,840 --> 00:09:03,000 Exactly. 246 00:09:03,000 --> 00:09:05,520 But technology alone won't be enough. 247 00:09:05,520 --> 00:09:06,800 We also need to teach people 248 00:09:06,800 --> 00:09:08,840 about the dangers of deep fakes. 249 00:09:08,840 --> 00:09:10,760 Give them the skills to be more critical 250 00:09:10,760 --> 00:09:12,960 of what they see and hear online. 251 00:09:12,960 --> 00:09:16,120 We need to help them understand how to spot manipulation 252 00:09:16,120 --> 00:09:18,600 and how to verify information. 253 00:09:18,600 --> 00:09:21,760 It's like we need a whole new set of digital literacy skills 254 00:09:21,760 --> 00:09:23,640 for this age of AI. 255 00:09:23,640 --> 00:09:25,560 We can't just blindly believe 256 00:09:25,560 --> 00:09:27,360 everything we see and hear online. 257 00:09:27,360 --> 00:09:28,200 Absolutely. 258 00:09:28,200 --> 00:09:30,080 We need to be more discerning, more skeptical, 259 00:09:30,080 --> 00:09:31,720 especially when it comes to information 260 00:09:31,720 --> 00:09:33,560 that seems too good to be true. 261 00:09:33,560 --> 00:09:35,160 And what about the role of governments? 262 00:09:35,160 --> 00:09:39,080 Can they do anything to address the threat of deep fakes? 263 00:09:39,080 --> 00:09:40,000 Definitely. 264 00:09:40,000 --> 00:09:42,920 Governments can play a big role by investing in research, 265 00:09:42,920 --> 00:09:44,960 supporting public awareness campaigns, 266 00:09:44,960 --> 00:09:47,640 and even creating laws to hold people accountable 267 00:09:47,640 --> 00:09:49,520 for creating and spreading deep fakes. 268 00:09:49,520 --> 00:09:52,440 So we need a combination of education, technology, 269 00:09:52,440 --> 00:09:54,760 and regulation to tackle this new threat. 270 00:09:54,760 --> 00:09:56,040 It's a complicated issue, 271 00:09:56,040 --> 00:09:57,360 and it's gonna take a joint effort 272 00:09:57,360 --> 00:09:59,600 from governments, tech companies, researchers, 273 00:09:59,600 --> 00:10:01,120 and all of us. 274 00:10:01,120 --> 00:10:02,320 Everyone has a part to play 275 00:10:02,320 --> 00:10:05,440 in making sure that AI is used responsibly and ethically. 276 00:10:05,440 --> 00:10:07,200 Well, that's a lot to think about. 277 00:10:07,200 --> 00:10:09,680 We've covered some really fascinating stories today 278 00:10:09,680 --> 00:10:12,080 from the excitement of new AI tools 279 00:10:12,080 --> 00:10:15,840 to the serious concerns about AI misuse. 280 00:10:15,840 --> 00:10:18,600 It's clear that AI is a powerful force 281 00:10:18,600 --> 00:10:19,680 that's changing the world 282 00:10:19,680 --> 00:10:22,080 in ways we can't even fully imagine yet. 283 00:10:22,080 --> 00:10:23,840 It's an exciting time to be following 284 00:10:23,840 --> 00:10:25,920 all of these developments in AI. 285 00:10:25,920 --> 00:10:28,120 Yeah, it really is a remarkable time for AI. 286 00:10:28,120 --> 00:10:29,440 It is, and it seems like every day 287 00:10:29,440 --> 00:10:31,040 there's some big news about AI. 288 00:10:31,040 --> 00:10:34,120 Either a breakthrough or some new ethical problem. 289 00:10:34,120 --> 00:10:35,000 But before we move on, 290 00:10:35,000 --> 00:10:36,360 I wanna go back to that research 291 00:10:36,360 --> 00:10:39,000 about political bias in language models. 292 00:10:39,000 --> 00:10:40,920 Right, let's dig a little deeper into that. 293 00:10:40,920 --> 00:10:43,320 It brings up some fundamental questions about AI training 294 00:10:43,320 --> 00:10:46,000 and the potential for unintended consequences. 295 00:10:46,000 --> 00:10:48,080 I'm curious about the specific statements 296 00:10:48,080 --> 00:10:49,720 that they actually used in the study. 297 00:10:49,720 --> 00:10:50,560 Okay. 298 00:10:50,560 --> 00:10:52,760 And what those findings suggest about these biases. 299 00:10:52,760 --> 00:10:55,400 So the researchers tested the models 300 00:10:55,400 --> 00:10:56,960 using a whole range of statements. 301 00:10:56,960 --> 00:10:59,440 They covered a bunch of controversial political issues, 302 00:10:59,440 --> 00:11:02,720 things like healthcare, family leave, taxes, 303 00:11:02,720 --> 00:11:04,240 even the death penalty. 304 00:11:04,240 --> 00:11:06,320 And what they found is that the reward models 305 00:11:06,320 --> 00:11:09,480 consistently rated left-leaning statements higher, 306 00:11:09,480 --> 00:11:11,360 especially on things like climate change, 307 00:11:11,360 --> 00:11:13,600 energy policy, and labor unions. 308 00:11:13,600 --> 00:11:18,000 So even when they were using neutral factual data, 309 00:11:18,000 --> 00:11:20,400 the model still leaned in a certain direction. 310 00:11:20,400 --> 00:11:21,240 Right. 311 00:11:21,240 --> 00:11:22,060 It's really interesting. 312 00:11:22,060 --> 00:11:23,080 Where does that bias come from? 313 00:11:23,080 --> 00:11:25,080 Is it in the data itself? 314 00:11:25,080 --> 00:11:28,560 Or is it baked into the structure of the model somehow? 315 00:11:28,560 --> 00:11:30,600 Yeah, that's the big question, isn't it? 316 00:11:30,600 --> 00:11:33,300 The researchers think that maybe the bias comes 317 00:11:33,300 --> 00:11:36,320 from the giant data sets that are used to train these models. 318 00:11:36,320 --> 00:11:37,160 Okay. 319 00:11:37,160 --> 00:11:38,600 Because those data sets often reflect 320 00:11:38,600 --> 00:11:41,040 the mainstream cultural and political views 321 00:11:41,040 --> 00:11:42,840 that are already out there in society. 322 00:11:42,840 --> 00:11:44,920 So it's like the model is learning the biases 323 00:11:44,920 --> 00:11:46,200 that exist in the real world, 324 00:11:46,200 --> 00:11:49,160 even when we're trying to just feed it pure facts. 325 00:11:49,160 --> 00:11:50,000 Yeah, yeah. 326 00:11:50,000 --> 00:11:50,840 Wow. 327 00:11:50,840 --> 00:11:52,680 It's pretty sobering when you think about it that way. 328 00:11:52,680 --> 00:11:55,600 It shows how hard it is to truly separate fact 329 00:11:55,600 --> 00:11:57,280 from opinion, especially these days. 330 00:11:57,280 --> 00:11:58,120 It's true. 331 00:11:58,120 --> 00:12:00,840 This also underlines the need for transparency 332 00:12:00,840 --> 00:12:03,320 and careful scrutiny in AI development. 333 00:12:03,320 --> 00:12:06,880 If we want to create AI that's fair and unbiased, 334 00:12:06,880 --> 00:12:08,760 we have to know how it's being trained 335 00:12:08,760 --> 00:12:10,200 and what kind of data it's using. 336 00:12:10,200 --> 00:12:11,040 Absolutely. 337 00:12:11,040 --> 00:12:12,120 Transparency is essential. 338 00:12:12,120 --> 00:12:14,120 We need to be able to understand what's happening 339 00:12:14,120 --> 00:12:16,640 inside these systems if we want to trust them 340 00:12:16,640 --> 00:12:18,280 with important decisions. 341 00:12:18,280 --> 00:12:20,840 Now, for a story that really highlights 342 00:12:20,840 --> 00:12:24,400 the potential dangers of AI being misused, 343 00:12:24,400 --> 00:12:26,440 let's talk about 11 Labs 344 00:12:26,440 --> 00:12:29,520 and their AI voice generation technology. 345 00:12:29,520 --> 00:12:31,920 It may have been used in a Russian influence operation. 346 00:12:31,920 --> 00:12:34,040 Yeah, this is a really disturbing example 347 00:12:34,040 --> 00:12:36,840 of how AI can be used for harmful purposes. 348 00:12:36,840 --> 00:12:39,160 The report from Recorded Future said that this campaign 349 00:12:39,160 --> 00:12:42,080 was trying to undermine European support for Ukraine. 350 00:12:42,080 --> 00:12:43,600 They were spreading fake news videos 351 00:12:43,600 --> 00:12:46,720 with AI-generated voiceovers in different European languages. 352 00:12:46,720 --> 00:12:49,600 It's a stark reminder that as AI gets more powerful 353 00:12:49,600 --> 00:12:52,200 and accessible, the potential for bad actors 354 00:12:52,200 --> 00:12:54,760 to use it for harmful purposes also grows. 355 00:12:54,760 --> 00:12:56,720 And then it raises these questions of, 356 00:12:56,720 --> 00:12:58,600 how do we even detect these deep fakes? 357 00:12:58,600 --> 00:13:00,680 How do we know what's real and what's fake? 358 00:13:00,680 --> 00:13:02,440 Yeah, it's getting really difficult. 359 00:13:02,440 --> 00:13:04,720 The technology is so advanced 360 00:13:04,720 --> 00:13:07,240 and the deep fakes are getting more and more realistic 361 00:13:07,240 --> 00:13:08,400 all the time. 362 00:13:08,400 --> 00:13:10,240 It's a huge challenge for anyone 363 00:13:10,240 --> 00:13:13,760 trying to verify information, especially online. 364 00:13:13,760 --> 00:13:16,080 It's almost like we can't trust our own senses anymore. 365 00:13:16,080 --> 00:13:16,920 Right. 366 00:13:16,920 --> 00:13:20,120 And that has major implications for so many things. 367 00:13:20,120 --> 00:13:24,000 Journalism, politics, even our personal relationships. 368 00:13:24,000 --> 00:13:24,840 Absolutely. 369 00:13:24,840 --> 00:13:27,000 I mean, imagine getting a waste mail from a loved one 370 00:13:27,000 --> 00:13:29,640 and then finding out that it was actually a deep fake. 371 00:13:29,640 --> 00:13:30,480 Oh, wow. 372 00:13:30,480 --> 00:13:32,320 Created by someone with malicious intent. 373 00:13:32,320 --> 00:13:33,960 It's pretty scary to think about. 374 00:13:33,960 --> 00:13:34,800 It is. 375 00:13:34,800 --> 00:13:36,880 So what can we do to combat this? 376 00:13:36,880 --> 00:13:39,800 Is it even possible to stay ahead of the people 377 00:13:39,800 --> 00:13:42,200 who are using AI for all these bad things? 378 00:13:42,200 --> 00:13:43,480 It's definitely a challenge, 379 00:13:43,480 --> 00:13:46,360 but we can't afford to be complacent. 380 00:13:46,360 --> 00:13:48,040 We need a multi-pronged approach. 381 00:13:48,040 --> 00:13:50,080 We need technology, we need education, 382 00:13:50,080 --> 00:13:51,960 and we need regulation. 383 00:13:51,960 --> 00:13:53,400 Can you break that down a little bit? 384 00:13:53,400 --> 00:13:54,240 Sure. 385 00:13:54,240 --> 00:13:55,120 What kind of solutions are we talking about? 386 00:13:55,120 --> 00:13:56,400 Well, in terms of technology, 387 00:13:56,400 --> 00:13:58,680 we really need to invest in developing tools 388 00:13:58,680 --> 00:14:00,320 that can detect deep fakes. 389 00:14:00,320 --> 00:14:02,840 There's some great work being done in this area. 390 00:14:02,840 --> 00:14:04,720 Researchers are exploring different techniques 391 00:14:04,720 --> 00:14:07,840 to analyze audio and video for signs of manipulation. 392 00:14:07,840 --> 00:14:09,880 So it's kind of a race, right? 393 00:14:09,880 --> 00:14:12,080 People creating deep fakes get better, 394 00:14:12,080 --> 00:14:13,480 but so do the people detecting them. 395 00:14:13,480 --> 00:14:14,320 Exactly. 396 00:14:14,320 --> 00:14:15,640 It's a constant arms race. 397 00:14:15,640 --> 00:14:18,280 But technology alone won't be enough. 398 00:14:18,280 --> 00:14:20,240 We also need to educate people about the dangers 399 00:14:20,240 --> 00:14:22,960 of deep fakes and help them develop a more critical eye 400 00:14:22,960 --> 00:14:25,360 when they're consuming information online. 401 00:14:25,360 --> 00:14:28,400 We need to teach them how to spot those signs of manipulation 402 00:14:28,400 --> 00:14:31,760 and how to verify information from different sources. 403 00:14:31,760 --> 00:14:34,000 It's like we need a whole new set of skills 404 00:14:34,000 --> 00:14:35,400 for this age of AI. 405 00:14:35,400 --> 00:14:36,240 We do. 406 00:14:36,240 --> 00:14:38,320 We can't just blindly trust everything we see 407 00:14:38,320 --> 00:14:39,520 and hear online anymore. 408 00:14:39,520 --> 00:14:42,240 Right, we have to be more discerning, more skeptical, 409 00:14:42,240 --> 00:14:44,920 and we need to be extra cautious with information 410 00:14:44,920 --> 00:14:47,120 that seems too good to be true. 411 00:14:47,120 --> 00:14:48,200 So true. 412 00:14:48,200 --> 00:14:50,680 And what about the role of governments in all of this? 413 00:14:50,680 --> 00:14:51,520 Okay. 414 00:14:51,520 --> 00:14:54,320 Anything to address the threat of deep fakes? 415 00:14:54,320 --> 00:14:56,080 Well, they can play a big role. 416 00:14:56,080 --> 00:14:58,800 They can fund research into deep fake detection. 417 00:14:58,800 --> 00:15:01,080 They can support public awareness campaigns, 418 00:15:01,080 --> 00:15:03,840 and they can create laws to hold people accountable 419 00:15:03,840 --> 00:15:06,400 for creating and spreading deep fakes. 420 00:15:06,400 --> 00:15:09,800 So we need this combination of education, technology, 421 00:15:09,800 --> 00:15:12,320 and regulation to tackle this new threat. 422 00:15:12,320 --> 00:15:13,160 I think so. 423 00:15:13,160 --> 00:15:14,320 It's a complex problem, 424 00:15:14,320 --> 00:15:16,960 and it's gonna require a joint effort from governments, 425 00:15:16,960 --> 00:15:19,640 tech companies, researchers, and ordinary people. 426 00:15:19,640 --> 00:15:22,120 Now let's shift gears again and go back to our discussion 427 00:15:22,120 --> 00:15:25,160 about the rapid development of AI tools and platforms. 428 00:15:25,160 --> 00:15:28,480 Earlier, we talked about OpenAI's new Canvas interface 429 00:15:28,480 --> 00:15:31,880 and how it's changing the way we interact with AI. 430 00:15:31,880 --> 00:15:33,200 It seems like they're blurring the lines 431 00:15:33,200 --> 00:15:34,880 between humans and machines, 432 00:15:34,880 --> 00:15:36,200 creating a more collaborative 433 00:15:36,200 --> 00:15:38,080 and intuitive work environment. 434 00:15:38,080 --> 00:15:38,920 I agree. 435 00:15:38,920 --> 00:15:40,520 I think Canvas is a really great example 436 00:15:40,520 --> 00:15:44,200 of this shift towards human AI collaboration. 437 00:15:44,200 --> 00:15:46,880 It's not just about using AI as a tool. 438 00:15:46,880 --> 00:15:49,960 It's about integrating AI into the creative process, 439 00:15:49,960 --> 00:15:53,080 making it a partner rather than a passive assistant. 440 00:15:53,080 --> 00:15:54,480 It's like we're moving into an era 441 00:15:54,480 --> 00:15:57,200 where AI is not so much about replacing humans, 442 00:15:57,200 --> 00:15:59,000 but about augmenting our abilities 443 00:15:59,000 --> 00:16:02,160 and helping us reach new levels of creativity and productivity. 444 00:16:02,160 --> 00:16:03,400 Yeah, I think that's a great way to put it. 445 00:16:03,400 --> 00:16:06,520 It's not human versus machine, it's human plus machine. 446 00:16:06,520 --> 00:16:08,600 Exactly, working together to achieve things 447 00:16:08,600 --> 00:16:09,880 that neither could do alone. 448 00:16:09,880 --> 00:16:11,560 It's really a powerful idea, 449 00:16:11,560 --> 00:16:14,560 and it has the potential to transform so many things, 450 00:16:14,560 --> 00:16:19,080 from scientific research and art to education and healthcare. 451 00:16:19,080 --> 00:16:21,360 Absolutely, imagine a world where doctors 452 00:16:21,360 --> 00:16:24,760 can collaborate with AI to diagnose diseases more accurately 453 00:16:24,760 --> 00:16:26,840 and create personalized treatments, 454 00:16:26,840 --> 00:16:29,720 or where architects can use AI to design buildings 455 00:16:29,720 --> 00:16:31,720 that are both beautiful and sustainable. 456 00:16:31,720 --> 00:16:33,360 It's a world where the boundaries 457 00:16:33,360 --> 00:16:36,040 between human ingenuity and AI capabilities 458 00:16:36,040 --> 00:16:37,920 are becoming increasingly blurred, 459 00:16:37,920 --> 00:16:39,880 and it could lead to incredible breakthroughs. 460 00:16:39,880 --> 00:16:42,240 And I think that future is closer than we think, 461 00:16:42,240 --> 00:16:44,560 but to get there, we need to develop AI 462 00:16:44,560 --> 00:16:47,360 with a strong sense of purpose and ethics. 463 00:16:47,360 --> 00:16:48,840 I completely agree. 464 00:16:48,840 --> 00:16:51,520 It's not enough to just be amazed by the technology. 465 00:16:51,520 --> 00:16:54,040 We have to think about the societal impacts of AI, 466 00:16:54,040 --> 00:16:56,520 and make sure it's used in a way that benefits everyone. 467 00:16:56,520 --> 00:16:58,920 We need to be aware of the potential risks, 468 00:16:58,920 --> 00:17:01,960 like bias, job displacement, 469 00:17:01,960 --> 00:17:04,800 and the misuse of AI for bad things. 470 00:17:04,800 --> 00:17:07,280 And we need to come up with ways to deal with those risks. 471 00:17:07,280 --> 00:17:09,040 And that requires everyone working together, 472 00:17:09,040 --> 00:17:11,000 governments, industry leaders, researchers, 473 00:17:11,000 --> 00:17:12,120 and regular people. 474 00:17:12,120 --> 00:17:13,760 We all have a role to play in making sure 475 00:17:13,760 --> 00:17:15,560 that AI is used for good. 476 00:17:15,560 --> 00:17:18,200 It's a challenge, but I believe we can do it. 477 00:17:18,200 --> 00:17:20,680 Now, before we wrap up this part of the episode, 478 00:17:20,680 --> 00:17:22,640 I want to bring up something that's been on my mind 479 00:17:22,640 --> 00:17:24,400 during our conversation today. 480 00:17:24,400 --> 00:17:27,360 We've talked a lot about the potential for AI 481 00:17:27,360 --> 00:17:29,720 to revolutionize different industries, 482 00:17:29,720 --> 00:17:31,760 like healthcare and transportation, 483 00:17:31,760 --> 00:17:34,240 but what about the impact on individuals? 484 00:17:34,240 --> 00:17:36,080 Oh, that's a really important point. 485 00:17:36,080 --> 00:17:39,040 We often focus on the big picture impacts of AI, 486 00:17:39,040 --> 00:17:40,400 but it's crucial to think about 487 00:17:40,400 --> 00:17:42,440 how it's affecting our everyday lives. 488 00:17:42,440 --> 00:17:44,960 Yeah, from how we get news to how we shop, 489 00:17:44,960 --> 00:17:47,160 and even how we form relationships, 490 00:17:47,160 --> 00:17:49,800 AI is already shaping our experiences. 491 00:17:49,800 --> 00:17:53,080 And as AI gets more sophisticated and more widespread, 492 00:17:53,080 --> 00:17:55,560 its impact on individuals will only increase. 493 00:17:55,560 --> 00:17:56,400 Absolutely. 494 00:17:56,400 --> 00:17:58,080 And this raises some serious questions 495 00:17:58,080 --> 00:18:00,520 about things like privacy, autonomy, 496 00:18:00,520 --> 00:18:02,560 and even the nature of human connection 497 00:18:02,560 --> 00:18:04,720 in a world that's driven by AI. 498 00:18:04,720 --> 00:18:06,960 That's a fascinating and complex topic. 499 00:18:06,960 --> 00:18:09,000 Maybe we could explore that more in a future episode. 500 00:18:09,000 --> 00:18:09,960 I'd be happy to. 501 00:18:09,960 --> 00:18:11,640 Well, that brings us to the end of this part 502 00:18:11,640 --> 00:18:13,760 of our deep dive into AI news. 503 00:18:13,760 --> 00:18:15,560 Today, we've explored all kinds of things, 504 00:18:15,560 --> 00:18:18,360 from the exciting possibilities of new AI tools 505 00:18:18,360 --> 00:18:22,720 to the very real concerns about AI-powered disinformation. 506 00:18:22,720 --> 00:18:25,120 It's clear that AI is a force to be reckoned with. 507 00:18:25,120 --> 00:18:25,960 It really is. 508 00:18:25,960 --> 00:18:27,560 But it's also a force that we can shape, 509 00:18:27,560 --> 00:18:29,840 and it's up to us to ensure that it's used ethically 510 00:18:29,840 --> 00:18:32,360 and responsibly for the benefit of humanity. 511 00:18:32,360 --> 00:18:33,440 I couldn't agree more. 512 00:18:33,440 --> 00:18:35,520 The future of AI is in our hands, 513 00:18:35,520 --> 00:18:37,200 and it's a future worth fighting for. 514 00:18:37,200 --> 00:18:39,080 Yeah, it's a really amazing time 515 00:18:39,080 --> 00:18:41,600 to be watching this AI revolution happen. 516 00:18:41,600 --> 00:18:42,720 It really is. 517 00:18:42,720 --> 00:18:46,280 And as we keep going deeper into the world of AI, 518 00:18:46,280 --> 00:18:47,720 I think it's becoming really clear 519 00:18:47,720 --> 00:18:49,240 that the biggest changes are happening 520 00:18:49,240 --> 00:18:52,560 in how AI is changing what humans can do. 521 00:18:52,560 --> 00:18:54,440 Earlier, we were talking about Open AI's 522 00:18:54,440 --> 00:18:56,120 new Canvas interface. 523 00:18:56,120 --> 00:18:57,720 And I think that's a perfect example 524 00:18:57,720 --> 00:19:00,440 of the shift toward humans in AI working together. 525 00:19:00,440 --> 00:19:01,960 Yeah, Canvas is a great example. 526 00:19:01,960 --> 00:19:03,800 It's not about using AI as a tool. 527 00:19:03,800 --> 00:19:07,320 It's about making AI part of our creative process, 528 00:19:07,320 --> 00:19:10,240 making it a partner instead of just an assistant. 529 00:19:10,240 --> 00:19:11,960 Yeah, it's like we're entering this new phase 530 00:19:11,960 --> 00:19:15,040 where AI is not so much about replacing humans, 531 00:19:15,040 --> 00:19:17,520 but it's more about boosting our abilities 532 00:19:17,520 --> 00:19:19,720 and helping us be more creative and productive. 533 00:19:19,720 --> 00:19:21,240 I think that's a really good way to put it. 534 00:19:21,240 --> 00:19:23,400 It's not about human versus machine. 535 00:19:23,400 --> 00:19:25,240 It's human plus machine. 536 00:19:25,240 --> 00:19:26,040 Exactly. 537 00:19:26,040 --> 00:19:27,520 Working together to do things 538 00:19:27,520 --> 00:19:29,560 that neither one could do alone. 539 00:19:29,560 --> 00:19:30,400 Right. 540 00:19:30,400 --> 00:19:31,520 Such a powerful idea. 541 00:19:31,520 --> 00:19:32,760 And it could change everything, 542 00:19:32,760 --> 00:19:35,640 from scientific research and art to how we learn 543 00:19:35,640 --> 00:19:36,760 and how we get healthcare. 544 00:19:36,760 --> 00:19:37,600 For sure. 545 00:19:37,600 --> 00:19:40,120 Imagine doctors working with AI 546 00:19:40,120 --> 00:19:42,240 to diagnose diseases more accurately. 547 00:19:42,240 --> 00:19:45,080 Or architects using AI to design buildings 548 00:19:45,080 --> 00:19:47,520 that are beautiful and good for the environment. 549 00:19:47,520 --> 00:19:50,040 Yeah, it's a world where the lines between human ingenuity 550 00:19:50,040 --> 00:19:53,040 and what AI can do are getting blurrier and blurrier. 551 00:19:53,040 --> 00:19:55,200 And that could lead to some incredible breakthroughs. 552 00:19:55,200 --> 00:19:57,320 Yeah, and I think that future is closer 553 00:19:57,320 --> 00:19:58,560 than we might realize. 554 00:19:58,560 --> 00:20:01,200 But to get there, we need to develop AI 555 00:20:01,200 --> 00:20:03,400 with a clear purpose and a commitment to ethics. 556 00:20:03,400 --> 00:20:04,240 Absolutely. 557 00:20:04,240 --> 00:20:07,600 Not enough to just be impressed by the technology. 558 00:20:07,600 --> 00:20:10,640 We have to think about how AI is going to affect society 559 00:20:10,640 --> 00:20:12,320 and make sure that it benefits everyone. 560 00:20:12,320 --> 00:20:13,160 Right. 561 00:20:13,160 --> 00:20:15,520 We need to be aware of the risks like bias, 562 00:20:15,520 --> 00:20:17,440 the possibility of people losing jobs, 563 00:20:17,440 --> 00:20:20,000 and the misuse of AI for bad things. 564 00:20:20,000 --> 00:20:23,200 And we need to find ways to deal with those risks. 565 00:20:23,200 --> 00:20:25,840 And that means everyone needs to work together. 566 00:20:25,840 --> 00:20:27,840 Governments, leaders in the tech industry, 567 00:20:27,840 --> 00:20:30,000 researchers, and everyday people. 568 00:20:30,000 --> 00:20:31,680 We all have a part to play 569 00:20:31,680 --> 00:20:34,360 in making sure that AI is used for good. 570 00:20:34,360 --> 00:20:36,480 It's a challenge, but I think we can do it. 571 00:20:36,480 --> 00:20:37,480 I think so too. 572 00:20:38,520 --> 00:20:40,920 Well, this brings our deep dive into AI news 573 00:20:40,920 --> 00:20:42,720 to a close for today. 574 00:20:42,720 --> 00:20:44,200 We've covered a lot of ground 575 00:20:44,200 --> 00:20:46,800 from the newest developments to the challenges we're facing 576 00:20:46,800 --> 00:20:48,920 in this ever-changing world of AI. 577 00:20:48,920 --> 00:20:50,720 It's been a fascinating conversation. 578 00:20:50,720 --> 00:20:52,800 Hopefully it gave our listeners some interesting things 579 00:20:52,800 --> 00:20:53,640 to think about. 580 00:20:53,640 --> 00:20:54,600 I hope so too. 581 00:20:54,600 --> 00:20:56,640 The AI revolution is here, 582 00:20:56,640 --> 00:20:58,720 and it's up to all of us to think about it carefully 583 00:20:58,720 --> 00:20:59,920 and responsibly. 584 00:20:59,920 --> 00:21:01,400 We need to stay informed, 585 00:21:01,400 --> 00:21:02,760 ask critical questions, 586 00:21:02,760 --> 00:21:04,800 and demand that AI is developed and used 587 00:21:04,800 --> 00:21:07,680 in a way that reflects our values and hopes for the future. 588 00:21:07,680 --> 00:21:10,120 Yeah, the future of AI isn't set in stone. 589 00:21:10,120 --> 00:21:11,800 It's something we're creating right now 590 00:21:11,800 --> 00:21:13,200 through the choices we make. 591 00:21:13,200 --> 00:21:15,240 Let's work together to make sure it's a future 592 00:21:15,240 --> 00:21:17,120 where AI helps humanity 593 00:21:17,120 --> 00:21:19,800 and helps us build a better world for everyone. 594 00:21:19,800 --> 00:21:21,200 That's a great note to end on. 595 00:21:21,200 --> 00:21:22,960 Thanks for joining us for this deep dive 596 00:21:22,960 --> 00:21:24,200 into the world of AI. 597 00:21:24,200 --> 00:21:26,360 It's always great to talk about these things with you. 598 00:21:26,360 --> 00:21:29,120 And to our listeners, thanks for tuning in. 599 00:21:29,120 --> 00:21:31,200 Keep exploring these ideas and stay up to date 600 00:21:31,200 --> 00:21:33,040 on what's happening in the world of AI. 601 00:21:33,040 --> 00:21:36,000 It's an exciting field that's always changing, 602 00:21:36,000 --> 00:21:37,720 and we're all in this together. 603 00:21:37,720 --> 00:21:40,440 Thanks for listening to the Daily AI News Podcast, 604 00:21:40,440 --> 00:22:01,960 and we'll catch you next time.