1 00:00:03,440 --> 00:00:10,360 Speaker 1: From Bloomberg News and iHeartRadio. It's the big take. I'm Westkasova. 2 00:00:10,600 --> 00:00:16,000 Speaker 1: Today chat GPT keeps getting better, but is that a 3 00:00:16,040 --> 00:00:28,240 Speaker 1: good thing? Chat GPT is what's called a large language model. 4 00:00:28,880 --> 00:00:32,080 Speaker 1: It has the ability to generate human like responses to 5 00:00:32,159 --> 00:00:36,320 Speaker 1: a wide range of questions and prompts, and its capabilities 6 00:00:36,320 --> 00:00:39,199 Speaker 1: have been praised by many as a major breakthrough in 7 00:00:39,240 --> 00:00:44,080 Speaker 1: the field of artificial intelligence. However, there are also those 8 00:00:44,200 --> 00:00:48,159 Speaker 1: who have raised concerns about the potential risks and drawbacks 9 00:00:48,200 --> 00:00:51,200 Speaker 1: of chat GPT. Some worry that it could be used 10 00:00:51,200 --> 00:00:54,840 Speaker 1: to spread misinformation or propaganda, or that it could be 11 00:00:55,000 --> 00:01:01,480 Speaker 1: used to manipulate public opinion in dangerous ways. It's Ava, 12 00:01:01,680 --> 00:01:04,680 Speaker 1: Oh hey, Ava, what's up? You should tell listeners you 13 00:01:04,760 --> 00:01:07,840 Speaker 1: had help with that intro. Hey, I was just getting 14 00:01:07,840 --> 00:01:11,360 Speaker 1: to that. Usually I write the intros to the show, 15 00:01:11,400 --> 00:01:15,160 Speaker 1: but today everything I said just now was written by 16 00:01:15,319 --> 00:01:18,880 Speaker 1: chat GPT. I asked the bot to write an intro 17 00:01:19,040 --> 00:01:23,319 Speaker 1: for a podcast episode about chat GBT, and that's what 18 00:01:23,480 --> 00:01:25,959 Speaker 1: it spit out. And to be honest, it kind of 19 00:01:25,959 --> 00:01:28,440 Speaker 1: freaked me out a little bit. I mean, it's a 20 00:01:28,480 --> 00:01:32,160 Speaker 1: little stiff, not as sparkling as my usual prose if 21 00:01:32,160 --> 00:01:35,680 Speaker 1: I do say so, but it still made me fear 22 00:01:35,800 --> 00:01:40,600 Speaker 1: for my job. But Ava, yes, Wes, you should tell 23 00:01:40,640 --> 00:01:44,640 Speaker 1: listeners that you're not a person, are you. Well, you've 24 00:01:44,680 --> 00:01:47,400 Speaker 1: got me there, Ava, as you might have guessed, is 25 00:01:47,440 --> 00:01:51,800 Speaker 1: herself an AI generated voice, and that technology too, is 26 00:01:51,840 --> 00:01:56,200 Speaker 1: getting closer and closer to sounding like us. The point, 27 00:01:56,240 --> 00:01:58,400 Speaker 1: of course, is that it's becoming a lot harder these 28 00:01:58,520 --> 00:02:01,840 Speaker 1: days to tell what's real and not and what's a 29 00:02:02,080 --> 00:02:06,240 Speaker 1: person and what's a machine. Fortunately, we have some real, 30 00:02:06,320 --> 00:02:10,480 Speaker 1: live humans here to dig into all these questions. Bloomberg 31 00:02:10,520 --> 00:02:13,480 Speaker 1: Tech reporters Dina Bass in Seattle and Rachel Mets in 32 00:02:13,520 --> 00:02:18,560 Speaker 1: San Francisco, and Permi Olson, a Bloomberg opinion columnist in London, 33 00:02:18,600 --> 00:02:24,600 Speaker 1: they've all been covering the rapid rise of chat GPT. Rachel, 34 00:02:24,639 --> 00:02:27,800 Speaker 1: We're all hearing a lot about chat GPT and the 35 00:02:27,840 --> 00:02:32,760 Speaker 1: company open Ai that created it. Microsoft is using open 36 00:02:32,800 --> 00:02:35,600 Speaker 1: AIS technology and its products, and other companies like Google 37 00:02:36,120 --> 00:02:40,560 Speaker 1: are scrambling to catch up. Can you tell us what 38 00:02:40,800 --> 00:02:43,760 Speaker 1: is chat GPT and how does it fit into the 39 00:02:43,800 --> 00:02:49,200 Speaker 1: world of artificial intelligence or AI. Well, it's most basic. 40 00:02:49,280 --> 00:02:52,120 Speaker 1: Chat GPT is a chat bought you ask it a 41 00:02:52,200 --> 00:02:55,480 Speaker 1: question you are asking it something and it's going to 42 00:02:55,520 --> 00:02:58,760 Speaker 1: give you something back. We've seen that before, right, either 43 00:02:58,840 --> 00:03:01,639 Speaker 1: you talk to something like a XA or you can 44 00:03:01,720 --> 00:03:04,840 Speaker 1: type to like a customer service chatpod on a website. 45 00:03:05,000 --> 00:03:08,600 Speaker 1: But this is different in that it's generative AI. So 46 00:03:08,960 --> 00:03:12,040 Speaker 1: it's not working quite the same way as the chat 47 00:03:12,080 --> 00:03:14,800 Speaker 1: boss of your so to speak, and it's in partless 48 00:03:14,800 --> 00:03:17,320 Speaker 1: because it's been trained on just a massive amount of 49 00:03:17,400 --> 00:03:20,480 Speaker 1: data from the Internet. When you say generative AI, what 50 00:03:20,520 --> 00:03:24,240 Speaker 1: do you mean by that? Basically you're giving it some input, 51 00:03:24,360 --> 00:03:28,120 Speaker 1: which might be a question or command like write me 52 00:03:28,200 --> 00:03:32,639 Speaker 1: a poem about Hello Kitty, and it's going to give 53 00:03:32,680 --> 00:03:36,040 Speaker 1: you a response, and it's not pulling it from a 54 00:03:36,120 --> 00:03:38,760 Speaker 1: database or something like that. It's sort of coming up 55 00:03:38,760 --> 00:03:41,280 Speaker 1: with it wholly anew and each which also means each 56 00:03:41,320 --> 00:03:43,200 Speaker 1: time you ask that kind of question, you'll get a 57 00:03:43,200 --> 00:03:46,200 Speaker 1: different answer, and they might be a little weird. They 58 00:03:46,280 --> 00:03:49,040 Speaker 1: might sound very factually correct but be like a little 59 00:03:49,040 --> 00:03:52,240 Speaker 1: bit skewed. But in a lot of cases it's actually 60 00:03:52,760 --> 00:03:58,440 Speaker 1: interesting information that could be usefuldina, Where is the bot 61 00:03:58,480 --> 00:04:02,720 Speaker 1: actually drawing this information from in order to form an 62 00:04:02,760 --> 00:04:07,160 Speaker 1: answer that reads like a person rot it? So basically, 63 00:04:07,160 --> 00:04:10,240 Speaker 1: what chat gypt is doing is drawing from the whole Internet. 64 00:04:10,640 --> 00:04:15,200 Speaker 1: Open AI basically scan large volumes of Internet content from 65 00:04:15,240 --> 00:04:20,920 Speaker 1: all sorts of places, including social media, Reddit, various web pages, Wikipedia, 66 00:04:21,040 --> 00:04:23,040 Speaker 1: and as a result, it gets a lot of different 67 00:04:23,040 --> 00:04:27,760 Speaker 1: points of view. It uses that text in an artificial 68 00:04:27,800 --> 00:04:30,720 Speaker 1: intelligence model that it's created, and it draws on what 69 00:04:30,760 --> 00:04:33,919 Speaker 1: it is quote unquote learned but not really to create 70 00:04:34,000 --> 00:04:36,559 Speaker 1: sort of a mimicry of human speech that it learned 71 00:04:36,560 --> 00:04:39,200 Speaker 1: from how people talk on the Internet. You can see 72 00:04:39,240 --> 00:04:41,760 Speaker 1: immediately what the benefits of that and the problems of 73 00:04:41,760 --> 00:04:44,880 Speaker 1: that might be. So we're getting really interesting things like 74 00:04:45,120 --> 00:04:48,719 Speaker 1: people who asked it to generate Seinfeld scripts or cocktail recipes. 75 00:04:49,000 --> 00:04:53,320 Speaker 1: But we also get misinformation, things that are incorrect, things 76 00:04:53,360 --> 00:04:55,960 Speaker 1: that can be abusive or not very nice, things that 77 00:04:56,279 --> 00:04:58,680 Speaker 1: can be creepy. One of the problems is when you 78 00:04:58,760 --> 00:05:02,080 Speaker 1: get something that's wrong, chat ept tends to state it 79 00:05:02,120 --> 00:05:04,680 Speaker 1: with the utmost confidence, so it's not going to flag 80 00:05:04,720 --> 00:05:07,520 Speaker 1: to you that this might not be right. This is 81 00:05:07,520 --> 00:05:10,520 Speaker 1: something that students that seized upon chat GPT immediately for 82 00:05:10,600 --> 00:05:14,760 Speaker 1: writing papers found out to their apparel party. As Dina 83 00:05:14,839 --> 00:05:19,000 Speaker 1: was saying, we're seeing all these instances from students to 84 00:05:19,279 --> 00:05:22,800 Speaker 1: companies starting to use chat GPT and starting to see 85 00:05:22,839 --> 00:05:25,400 Speaker 1: this as a way of assimilating a large amount of 86 00:05:25,400 --> 00:05:28,520 Speaker 1: information that isn't just coming up with funny poems. What 87 00:05:28,640 --> 00:05:30,760 Speaker 1: are some of the ways that business is already getting 88 00:05:30,760 --> 00:05:33,800 Speaker 1: their hands on this. Yeah, it's We're in a really 89 00:05:33,839 --> 00:05:36,920 Speaker 1: interesting moment for how we use these models, and I 90 00:05:36,920 --> 00:05:40,320 Speaker 1: would call it a sort of honeymoon period. So when 91 00:05:40,440 --> 00:05:43,080 Speaker 1: chat CHPT was first announced a few months ago, there 92 00:05:43,080 --> 00:05:46,720 Speaker 1: were lots of screenshots shared on Twitter of people using 93 00:05:46,760 --> 00:05:50,599 Speaker 1: it in fun ways, right, how to remove a peanut 94 00:05:50,600 --> 00:05:53,279 Speaker 1: butter sand which from a VCR in the style of 95 00:05:53,279 --> 00:05:55,800 Speaker 1: the King James Bible. That was my favorite out of 96 00:05:55,800 --> 00:05:58,880 Speaker 1: all of them. But after that is now the company's 97 00:05:58,920 --> 00:06:01,680 Speaker 1: trying to figure out how do we make money from this? 98 00:06:02,240 --> 00:06:07,000 Speaker 1: So just recently open ai released its latest, much more 99 00:06:07,080 --> 00:06:10,600 Speaker 1: powerful model called GPT four. If you think of the 100 00:06:10,680 --> 00:06:14,719 Speaker 1: chatbot as being like a car, these models that Dina described, 101 00:06:14,920 --> 00:06:18,440 Speaker 1: they're like the engine, and each engine is getting more powerful, 102 00:06:18,600 --> 00:06:21,400 Speaker 1: allowing the car to go faster and do more things. 103 00:06:21,480 --> 00:06:26,200 Speaker 1: And this new GPT four is more accurate, according to 104 00:06:26,200 --> 00:06:29,640 Speaker 1: open Ai, much more humanlike in its responses and in 105 00:06:29,680 --> 00:06:33,320 Speaker 1: the announcement, open ai gave some examples of companies that 106 00:06:33,360 --> 00:06:37,719 Speaker 1: are actually using GPT four. One example was Morgan Stanley. 107 00:06:37,839 --> 00:06:41,520 Speaker 1: So Morgan Stanley has been using GPT four since last year, 108 00:06:42,000 --> 00:06:44,560 Speaker 1: and they said that about two hundred of their wealth 109 00:06:44,600 --> 00:06:49,479 Speaker 1: management staff had been using a chatbot, their own proprietary 110 00:06:49,560 --> 00:06:54,120 Speaker 1: chatbot that had been trained on thousands of papers written 111 00:06:54,160 --> 00:06:58,679 Speaker 1: by their analysts to come up with a model using 112 00:06:58,680 --> 00:07:02,200 Speaker 1: the GPT for model that could just give them answers 113 00:07:02,240 --> 00:07:04,400 Speaker 1: more quickly than they would get if they had to 114 00:07:04,440 --> 00:07:08,560 Speaker 1: read that research themselves. And so they're saying that what 115 00:07:08,720 --> 00:07:12,480 Speaker 1: the advisors would take thirty minutes to do, they can 116 00:07:12,520 --> 00:07:16,040 Speaker 1: now do in seconds. So the value here is that 117 00:07:16,160 --> 00:07:21,920 Speaker 1: essentially companies can ring out more productivity from their workers 118 00:07:22,160 --> 00:07:24,520 Speaker 1: than they did before. So, you know, there's a lot 119 00:07:24,560 --> 00:07:27,960 Speaker 1: of talk about these systems replacing professional jobs. I don't 120 00:07:28,000 --> 00:07:30,040 Speaker 1: think they're actually going to do that. I think they're 121 00:07:30,040 --> 00:07:32,600 Speaker 1: just going to make professional workers have to do more 122 00:07:33,040 --> 00:07:37,040 Speaker 1: in less time. So PARMI, I want to jump in 123 00:07:37,040 --> 00:07:39,720 Speaker 1: with one example and one question for you. Rachel and 124 00:07:39,760 --> 00:07:42,400 Speaker 1: I were talking to Greg Bachran, who's the president of 125 00:07:42,480 --> 00:07:45,720 Speaker 1: open Ai. He mentioned another use case similar to Morgan 126 00:07:45,760 --> 00:07:48,320 Speaker 1: Stanley summarizing all of these papers, which was the US 127 00:07:48,400 --> 00:07:51,080 Speaker 1: tax Code. He flagged for US that he thought GPT 128 00:07:51,240 --> 00:07:56,000 Speaker 1: four was super useful for summarizing the very complicated US 129 00:07:56,080 --> 00:07:58,720 Speaker 1: tax code and telling you what policy, what exemption is 130 00:07:58,720 --> 00:08:01,120 Speaker 1: relevant for you. Probably would have been better if we 131 00:08:01,160 --> 00:08:03,440 Speaker 1: had that a month ago before people started filing, But 132 00:08:03,560 --> 00:08:05,800 Speaker 1: there we go. But I wanted to ask you on 133 00:08:05,880 --> 00:08:09,480 Speaker 1: your point about replacing jobs versus increasing productivity. I hear 134 00:08:09,520 --> 00:08:11,960 Speaker 1: the same thing from Microsoft, which is pushing this. It'll 135 00:08:11,960 --> 00:08:14,560 Speaker 1: make you more productive, it'll free you from all the groundwork. 136 00:08:14,960 --> 00:08:18,600 Speaker 1: But aren't there people at entry level jobs that do 137 00:08:18,640 --> 00:08:21,520 Speaker 1: that sort of thing, that do the summarizing for Morgan Stanley? 138 00:08:21,520 --> 00:08:25,360 Speaker 1: Aren't there paralegals computer programming as well? Aren't there entry 139 00:08:25,440 --> 00:08:28,920 Speaker 1: level jobs that may get replaced if we can't figure 140 00:08:28,920 --> 00:08:31,800 Speaker 1: out how to upscale those people. I think that's a 141 00:08:31,800 --> 00:08:34,880 Speaker 1: good possibility. But that's why I say large swaths. I 142 00:08:34,920 --> 00:08:37,640 Speaker 1: don't think there's going to be large swaths of job losses, 143 00:08:37,760 --> 00:08:40,319 Speaker 1: but there will be job losses. One hundred percent people 144 00:08:40,320 --> 00:08:42,880 Speaker 1: will be replaced by this. But it kind of makes 145 00:08:42,880 --> 00:08:47,440 Speaker 1: me think of the translation industry. So professional translators five 146 00:08:47,559 --> 00:08:50,640 Speaker 1: ten years ago we're really worried about being replaced by 147 00:08:51,040 --> 00:08:56,320 Speaker 1: Google Translate and Deepole and these other AI translation software tools. 148 00:08:56,800 --> 00:08:58,680 Speaker 1: But in the end, there weren't a lot of job 149 00:08:58,720 --> 00:09:02,199 Speaker 1: losses in the translate industry. They were just expected to 150 00:09:02,240 --> 00:09:06,439 Speaker 1: translate more words, so instead of translating two thousand words 151 00:09:06,480 --> 00:09:09,199 Speaker 1: in a day, they were expected to translate four thousand 152 00:09:09,200 --> 00:09:11,520 Speaker 1: words in a day. And I was speaking to some 153 00:09:11,559 --> 00:09:14,360 Speaker 1: people in the industry just recently because I thought that 154 00:09:14,360 --> 00:09:16,560 Speaker 1: this was quite a useful parallel to figure out where 155 00:09:16,559 --> 00:09:19,280 Speaker 1: we might be going with this more broadly, and there 156 00:09:19,280 --> 00:09:21,600 Speaker 1: weren't a lot of job losses that they knew of, 157 00:09:21,679 --> 00:09:24,080 Speaker 1: just anecdotally. So I think there will be some, and 158 00:09:24,120 --> 00:09:26,280 Speaker 1: I think you're exactly right, Dina, that it will be 159 00:09:26,920 --> 00:09:29,880 Speaker 1: these kind of entry level roles, and I think people 160 00:09:29,920 --> 00:09:32,600 Speaker 1: are just going to have to rethink what people do 161 00:09:32,720 --> 00:09:35,320 Speaker 1: in those kinds of roles when they do come in 162 00:09:35,360 --> 00:09:40,080 Speaker 1: as interns or kind of junior workers. Rachel Permit has 163 00:09:40,120 --> 00:09:43,440 Speaker 1: described the way some of these companies are using GPT 164 00:09:43,920 --> 00:09:49,240 Speaker 1: for important functions, things that customers would rely on critical information. 165 00:09:49,280 --> 00:09:52,880 Speaker 1: But as Dina said earlier, the bat will spit out 166 00:09:52,880 --> 00:09:56,160 Speaker 1: an answer with supreme confidence, and what we've learned is 167 00:09:56,200 --> 00:09:59,920 Speaker 1: that sometimes it's just wrong, Like if it doesn't know 168 00:10:00,120 --> 00:10:03,120 Speaker 1: the truth, it'll just kind of make it up. And 169 00:10:03,360 --> 00:10:06,800 Speaker 1: isn't that a big problem if you're using it for 170 00:10:06,920 --> 00:10:10,400 Speaker 1: business and you can't tell whether what the bot is 171 00:10:10,440 --> 00:10:13,880 Speaker 1: spitting out is real or not. Yeah, I think that's 172 00:10:13,920 --> 00:10:17,640 Speaker 1: a huge concern, And for me personally, I would be 173 00:10:17,679 --> 00:10:21,160 Speaker 1: really reluctant to trust it at this point with any 174 00:10:21,200 --> 00:10:25,680 Speaker 1: really important applications in my life, like personal finances. I 175 00:10:25,720 --> 00:10:28,559 Speaker 1: would personally want to know if my bank is using it. 176 00:10:28,760 --> 00:10:30,480 Speaker 1: I mean, maybe I'm like a little nerdier than the 177 00:10:30,480 --> 00:10:32,959 Speaker 1: average person, but I'd want to know as specifically as 178 00:10:33,480 --> 00:10:36,520 Speaker 1: what model are you using to get this information from? 179 00:10:36,760 --> 00:10:39,440 Speaker 1: How much information can you give me about the AI 180 00:10:39,679 --> 00:10:42,559 Speaker 1: model that you're using, an AI company service that you're using, 181 00:10:43,040 --> 00:10:45,440 Speaker 1: Just to get a handle on like what might be 182 00:10:46,120 --> 00:10:48,079 Speaker 1: leading to the answers. We have to remember that these 183 00:10:48,080 --> 00:10:52,080 Speaker 1: are essentially statistical software. I mean, all the word generation, 184 00:10:52,160 --> 00:10:55,400 Speaker 1: I should say, is being generated statistically. It's not like 185 00:10:55,520 --> 00:10:58,880 Speaker 1: a machine that's pulling from two different pockets. One is 186 00:10:59,240 --> 00:11:01,600 Speaker 1: here answers to things that I totally know, so I'll 187 00:11:01,600 --> 00:11:03,760 Speaker 1: give you an answer from here, versus I don't know, 188 00:11:03,800 --> 00:11:05,480 Speaker 1: so I'm going to make it up over here. It's 189 00:11:05,520 --> 00:11:08,480 Speaker 1: just it's really a jumble based on what it has 190 00:11:08,480 --> 00:11:11,680 Speaker 1: been trained on and what it thinks is likely to be. 191 00:11:12,360 --> 00:11:14,280 Speaker 1: You know what you want as an answer based on 192 00:11:14,320 --> 00:11:17,440 Speaker 1: the input you gave it, and that Yeah, that should 193 00:11:17,559 --> 00:11:20,079 Speaker 1: make people nervous if it's being used to do things 194 00:11:20,120 --> 00:11:22,640 Speaker 1: like deal with their personal money of their company's money. 195 00:11:23,280 --> 00:11:27,520 Speaker 1: GPD four. Technically it's what you would call a multimodal model, 196 00:11:27,920 --> 00:11:31,520 Speaker 1: which means it's not just generating text. It can also 197 00:11:32,120 --> 00:11:35,640 Speaker 1: do things like computer vision tasks, which you can't do 198 00:11:35,720 --> 00:11:38,120 Speaker 1: that right now, in like the ai playground that Opening 199 00:11:38,120 --> 00:11:40,480 Speaker 1: Eye has available to the average user, Like here's a picture, 200 00:11:40,559 --> 00:11:42,480 Speaker 1: tell me what the caption could be. We couldn't do that, 201 00:11:44,200 --> 00:11:46,360 Speaker 1: and we should say this was the demo that open 202 00:11:46,440 --> 00:11:50,280 Speaker 1: ai did when it released GPT for the new version, 203 00:11:50,720 --> 00:11:53,080 Speaker 1: right because that was one of the really interesting parts 204 00:11:53,240 --> 00:11:57,720 Speaker 1: of the demo was that you could point your camera 205 00:11:57,760 --> 00:12:00,400 Speaker 1: on your phone camera at the inside of your fridge 206 00:12:00,920 --> 00:12:05,000 Speaker 1: and chat. GPT will now say what recipes you can 207 00:12:05,040 --> 00:12:07,600 Speaker 1: make based on the contents of your fridge. Okay, but 208 00:12:07,679 --> 00:12:10,600 Speaker 1: let's keep in mind those recipes might be terrible yeah, 209 00:12:11,440 --> 00:12:13,760 Speaker 1: you know, it's like or like, what cocktails should I make? 210 00:12:14,040 --> 00:12:16,760 Speaker 1: There's an app that they're suggesting for people with blindness 211 00:12:16,880 --> 00:12:19,800 Speaker 1: or low vision as well. It's not just cocktails. Yeah, 212 00:12:19,800 --> 00:12:21,400 Speaker 1: I know. I've known about that app for a couple 213 00:12:21,400 --> 00:12:23,280 Speaker 1: of years, and it's you know, I don't know if 214 00:12:23,280 --> 00:12:26,559 Speaker 1: they were using machine vision recently or if they've always 215 00:12:26,679 --> 00:12:28,560 Speaker 1: used humans. You know, for a lot of that, they've 216 00:12:28,600 --> 00:12:31,760 Speaker 1: just had to use humans to help. But it's a 217 00:12:31,920 --> 00:12:35,800 Speaker 1: super cool app. So how does that app work? In 218 00:12:36,040 --> 00:12:38,120 Speaker 1: the little bit that I saw which I thought was 219 00:12:38,160 --> 00:12:41,680 Speaker 1: really interesting and fingers crossed that it actually can work 220 00:12:41,720 --> 00:12:44,320 Speaker 1: well and be helpful for people. Essentially, what you would 221 00:12:44,360 --> 00:12:46,240 Speaker 1: do with the app, if you are a blind or 222 00:12:46,240 --> 00:12:49,040 Speaker 1: low vision you might take a picture of, say two 223 00:12:49,040 --> 00:12:51,360 Speaker 1: slightly different shirts that you've laid out on your bed. 224 00:12:51,679 --> 00:12:54,320 Speaker 1: Maybe one is blue with polka dots and the other 225 00:12:54,400 --> 00:12:57,040 Speaker 1: is green with polka dots, let's say, and you could 226 00:12:57,080 --> 00:12:59,800 Speaker 1: send in the picture and say which is the blue 227 00:12:59,800 --> 00:13:02,840 Speaker 1: one question mark and it could tell you, you know, 228 00:13:02,840 --> 00:13:05,720 Speaker 1: would analyze the image, and it would analyze your text, 229 00:13:05,760 --> 00:13:07,679 Speaker 1: and it could give you an answer. And that kind 230 00:13:07,679 --> 00:13:10,840 Speaker 1: of thing could be super useful and really just like 231 00:13:11,000 --> 00:13:14,040 Speaker 1: a speedy way to improve the day of somebody who 232 00:13:14,080 --> 00:13:16,800 Speaker 1: has a hard time doing that kind of task that 233 00:13:16,840 --> 00:13:18,680 Speaker 1: a lot of us think is super simple. You know, 234 00:13:18,679 --> 00:13:20,720 Speaker 1: we don't even think about it, but that's hard for 235 00:13:20,760 --> 00:13:25,000 Speaker 1: somebody who can't see that well. So the consumer applications 236 00:13:25,000 --> 00:13:28,959 Speaker 1: that we're talking about CHAT GPT, the Microsoft being Search, 237 00:13:29,040 --> 00:13:32,440 Speaker 1: which is based on the new version of GPTGBT four, 238 00:13:32,520 --> 00:13:35,600 Speaker 1: so that's a web search, and the thing Morgan Stanley 239 00:13:35,720 --> 00:13:39,720 Speaker 1: is doing, or Microsoft and other companies are now putting 240 00:13:39,760 --> 00:13:43,120 Speaker 1: these kinds of artificial intelligence capabilities in their business products. 241 00:13:43,440 --> 00:13:47,119 Speaker 1: So the business products are learning from a more confined 242 00:13:47,200 --> 00:13:50,360 Speaker 1: set of information. They have the ability to learn certain 243 00:13:50,400 --> 00:13:53,360 Speaker 1: things from the Internet, but they're learning from the company's data. 244 00:13:53,480 --> 00:13:55,920 Speaker 1: So in the Morgan Stanley case, it's the Morgan Stanley 245 00:13:56,000 --> 00:13:58,240 Speaker 1: report that analysts right that they would normally send to 246 00:13:58,280 --> 00:14:02,559 Speaker 1: their financial advisors. In the case of Microsoft's customer service 247 00:14:02,600 --> 00:14:07,160 Speaker 1: software or software for salespeople, it's learning from each company's 248 00:14:07,480 --> 00:14:11,480 Speaker 1: database and set of information and set of interactions with customers. 249 00:14:11,800 --> 00:14:16,360 Speaker 1: That's a more confined set of things and less the 250 00:14:16,559 --> 00:14:19,560 Speaker 1: entire universe of the Internet, so you have a little 251 00:14:19,560 --> 00:14:22,440 Speaker 1: bit less potential for it to go awry or to 252 00:14:22,480 --> 00:14:25,200 Speaker 1: do something weird, or to get something completely wrong. But 253 00:14:25,360 --> 00:14:30,600 Speaker 1: the stakes are higher. In a monetary material company application, 254 00:14:30,640 --> 00:14:33,160 Speaker 1: there's some pretty high stakes. So I asked Microsoft CEO 255 00:14:33,200 --> 00:14:36,400 Speaker 1: Satanadella this question when they unveiled some of the business tools, 256 00:14:36,760 --> 00:14:39,400 Speaker 1: and his comment to me was, well, the humans make 257 00:14:39,440 --> 00:14:42,240 Speaker 1: mistakes too, so we need to get to a point 258 00:14:42,240 --> 00:14:45,200 Speaker 1: where you know, the mistakes from the machine are lesser, 259 00:14:45,360 --> 00:14:49,600 Speaker 1: but it is something to be aware of. The mistakes 260 00:14:49,640 --> 00:14:52,960 Speaker 1: in the consumer chat are different also because at least 261 00:14:52,960 --> 00:14:56,040 Speaker 1: in some of the applications. So if you take, for example, 262 00:14:56,080 --> 00:14:58,920 Speaker 1: the bing search engine that uses the open aie product, 263 00:14:59,240 --> 00:15:02,320 Speaker 1: there are sources. Is what Microsoft has done to help people, 264 00:15:02,600 --> 00:15:05,240 Speaker 1: and to flag where the information is coming from is 265 00:15:05,280 --> 00:15:07,600 Speaker 1: to cite the way a high school or would in 266 00:15:07,640 --> 00:15:10,240 Speaker 1: a term paper, and those links are clickable. You can 267 00:15:10,280 --> 00:15:12,520 Speaker 1: go in and see this piece of information is coming 268 00:15:12,520 --> 00:15:14,320 Speaker 1: from this new source, and maybe I don't trust it, 269 00:15:14,440 --> 00:15:16,720 Speaker 1: or I'm reading the article and actually the article doesn't 270 00:15:16,760 --> 00:15:19,320 Speaker 1: say what Being says it does. I had that happen 271 00:15:19,360 --> 00:15:21,760 Speaker 1: a few times when I was asking Being about the 272 00:15:21,880 --> 00:15:25,240 Speaker 1: US shootdowns of the unidentified flying objects being kept getting 273 00:15:25,280 --> 00:15:27,280 Speaker 1: parts of it wrong, and I could see where the 274 00:15:27,360 --> 00:15:30,400 Speaker 1: error was coming from in the article. The concern people have, 275 00:15:30,520 --> 00:15:32,920 Speaker 1: and it's a valid one, is will everyone do that 276 00:15:33,080 --> 00:15:34,960 Speaker 1: or will they just look at the answer and while 277 00:15:34,960 --> 00:15:36,520 Speaker 1: the answer is wrong and they're not going to click 278 00:15:36,520 --> 00:15:40,520 Speaker 1: into the sources. I just think that humans by and 279 00:15:40,640 --> 00:15:44,720 Speaker 1: large have this tendency to believe the algorithm more than 280 00:15:44,760 --> 00:15:47,680 Speaker 1: they do other humans. So we put a lot of 281 00:15:47,720 --> 00:15:51,440 Speaker 1: trust in machines collectively, which is why I think Sacha 282 00:15:51,560 --> 00:15:55,400 Speaker 1: Adela's comment to you rings a little bit hollow. The 283 00:15:55,440 --> 00:15:59,360 Speaker 1: bar for machine truthfulness is really high, and one of 284 00:15:59,360 --> 00:16:03,440 Speaker 1: the things that makes this quite complex for people who 285 00:16:03,600 --> 00:16:06,920 Speaker 1: use the systems is that we don't know how often 286 00:16:07,760 --> 00:16:12,800 Speaker 1: GPT four or chat GPT actually makes mistakes. AI and 287 00:16:12,880 --> 00:16:16,440 Speaker 1: computer scientists have been referring to this as it's hallucination rate. 288 00:16:16,800 --> 00:16:20,920 Speaker 1: I've asked open AI multiple times, what's the hallucination rate 289 00:16:21,120 --> 00:16:24,880 Speaker 1: for GPT three point five. They won't say. The only 290 00:16:24,920 --> 00:16:28,560 Speaker 1: thing they did say about GPT four was that its 291 00:16:28,560 --> 00:16:33,200 Speaker 1: hallucination rate was forty percent more likely to produce factual 292 00:16:33,240 --> 00:16:37,120 Speaker 1: responses than GPT three point five. Okay, that's great, so 293 00:16:37,120 --> 00:16:39,120 Speaker 1: it's a little bit more factually, but we don't even 294 00:16:39,200 --> 00:16:42,800 Speaker 1: know what the baseline was so it doesn't really tell 295 00:16:42,880 --> 00:16:46,960 Speaker 1: us anything. And the problem with AI systems generally is 296 00:16:46,960 --> 00:16:49,120 Speaker 1: it the only way to really know how well they'll 297 00:16:49,120 --> 00:16:52,400 Speaker 1: perform is when you release them into the wild. So 298 00:16:52,440 --> 00:16:55,520 Speaker 1: we are all effectively the guinea pigs for this stuff. 299 00:16:55,880 --> 00:16:59,600 Speaker 1: When these things get things wrong, that's how the designers 300 00:16:59,600 --> 00:17:02,880 Speaker 1: of these will know. But we also have to pay 301 00:17:02,920 --> 00:17:04,960 Speaker 1: the price, and we're not even sure what that price 302 00:17:05,000 --> 00:17:07,720 Speaker 1: will be. You mentioned a couple of times this term 303 00:17:07,760 --> 00:17:11,000 Speaker 1: hallucin nation rate. That's a great term. What exactly does 304 00:17:11,040 --> 00:17:15,440 Speaker 1: that mean? Just confidently presenting as fact something that is 305 00:17:15,760 --> 00:17:19,320 Speaker 1: not factual or flu and tagwash. That's the other one 306 00:17:19,320 --> 00:17:24,080 Speaker 1: I've heard. When we come back my confert, you know what, Eva, 307 00:17:24,720 --> 00:17:27,600 Speaker 1: why don't you take it from here? More from Parmi, 308 00:17:27,840 --> 00:17:38,080 Speaker 1: Rachel and Dina After the break Parmy, we talked about 309 00:17:38,080 --> 00:17:41,640 Speaker 1: some of the businesses that are using chat GPT and 310 00:17:42,119 --> 00:17:45,600 Speaker 1: one of the people getting involved is Elon Musk, Because 311 00:17:45,960 --> 00:17:48,840 Speaker 1: of course, what exactly is he trying to do and 312 00:17:48,880 --> 00:17:51,320 Speaker 1: how serious a venture is this, because we all know 313 00:17:51,400 --> 00:17:54,240 Speaker 1: that Elon Musk likes Elon Musk's name to be in 314 00:17:54,280 --> 00:17:57,680 Speaker 1: the news. Yeah, and there's a lot of different shiny 315 00:17:57,720 --> 00:18:01,000 Speaker 1: objects that Elon Musk likes to go off, and the 316 00:18:01,080 --> 00:18:05,439 Speaker 1: latest one does appear to be AI and language models 317 00:18:05,520 --> 00:18:08,800 Speaker 1: multimodal models. So there was a report in the Information 318 00:18:08,880 --> 00:18:12,960 Speaker 1: recently that said that he was reaching out to artificial 319 00:18:13,000 --> 00:18:16,960 Speaker 1: intelligence researchers and trying to form a new research lab 320 00:18:17,359 --> 00:18:21,200 Speaker 1: and the goal was to build this alternative to chat GPT. 321 00:18:22,000 --> 00:18:26,919 Speaker 1: Musk does have a storied history of involvement in AI development. 322 00:18:27,040 --> 00:18:30,520 Speaker 1: He was an early backer of deep Mind, which is 323 00:18:30,920 --> 00:18:33,680 Speaker 1: the AI lab that was eventually bought by Google. He 324 00:18:33,880 --> 00:18:35,920 Speaker 1: is one of the co founders of open ai, which 325 00:18:35,960 --> 00:18:38,919 Speaker 1: is deep Mind's rival, so he really has his fingers 326 00:18:38,920 --> 00:18:41,679 Speaker 1: and a lot of very important pies in AI, and 327 00:18:41,720 --> 00:18:46,159 Speaker 1: in recent years he has grown a little bit disenchanted 328 00:18:46,440 --> 00:18:51,360 Speaker 1: with how these companies have pursued artificial intelligence research. He 329 00:18:51,480 --> 00:18:55,520 Speaker 1: has complained that open ai was training AI to be woke. 330 00:18:56,440 --> 00:18:59,040 Speaker 1: So we don't know what's going on inside his head, 331 00:18:59,119 --> 00:19:01,440 Speaker 1: but it may well be that he wants to try 332 00:19:01,440 --> 00:19:06,119 Speaker 1: and build an alternative language model chatbot system that isn't 333 00:19:06,160 --> 00:19:10,760 Speaker 1: constrained by the same kinds of content filters and policies 334 00:19:11,080 --> 00:19:15,239 Speaker 1: that open ai has on its chatbot DNA. I think 335 00:19:15,280 --> 00:19:18,600 Speaker 1: when most people hear about AI or a chat GPT, 336 00:19:18,800 --> 00:19:21,920 Speaker 1: they think they're the same thing. But there's a lot 337 00:19:22,000 --> 00:19:25,440 Speaker 1: of AI products out there now with different uses. Can 338 00:19:25,520 --> 00:19:30,240 Speaker 1: you give us like three examples of AI product categories, 339 00:19:30,480 --> 00:19:33,880 Speaker 1: what their products do and what they're used for. It's 340 00:19:33,920 --> 00:19:36,920 Speaker 1: interesting is GPT, even though we've talked about it as language, 341 00:19:37,080 --> 00:19:39,520 Speaker 1: spawned a bunch of different things. So we've talked a 342 00:19:39,520 --> 00:19:42,720 Speaker 1: lot about chatbots. It spawned a bunch of chatbots. Open 343 00:19:42,760 --> 00:19:45,000 Speaker 1: AI when they were using it, also discovered it's really 344 00:19:45,040 --> 00:19:48,240 Speaker 1: good at computer programming, so that's also another kind of 345 00:19:48,320 --> 00:19:51,960 Speaker 1: language generation, but a completely different language. They were using 346 00:19:51,960 --> 00:19:54,200 Speaker 1: it and they found that it did a very serviceable 347 00:19:54,240 --> 00:19:57,280 Speaker 1: job with computer code, so they created their own version 348 00:19:57,480 --> 00:20:01,240 Speaker 1: of a computer coding tool that they call codec. Microsoft 349 00:20:01,320 --> 00:20:05,440 Speaker 1: then reformulated that further into something called Copilot, which maybe 350 00:20:05,440 --> 00:20:08,800 Speaker 1: in one of the most widely used corporate applications of 351 00:20:08,800 --> 00:20:11,119 Speaker 1: this kind of AI. Already, there's a fair number of 352 00:20:11,119 --> 00:20:13,800 Speaker 1: computer programmers that are using it, at least for some 353 00:20:13,880 --> 00:20:17,359 Speaker 1: of those sort of wrote programming tasks, things that you know, 354 00:20:17,680 --> 00:20:20,960 Speaker 1: we're kind of bother someone annoying to do. It also 355 00:20:21,040 --> 00:20:25,119 Speaker 1: spawned image generation, so earlier last year before Chat GPT, 356 00:20:25,240 --> 00:20:28,920 Speaker 1: everybody was fixated on Open Aiyes, Dolly, that also came 357 00:20:28,920 --> 00:20:31,480 Speaker 1: out of the GPT model, and that's a product where 358 00:20:31,480 --> 00:20:33,320 Speaker 1: you can type in a couple of words of text 359 00:20:33,359 --> 00:20:36,639 Speaker 1: and it will generate a picture for you. So if 360 00:20:36,640 --> 00:20:40,399 Speaker 1: you want a grilled cheese clock in the style of Picasso, 361 00:20:40,560 --> 00:20:43,359 Speaker 1: it will make that for you. Artists have been very 362 00:20:43,440 --> 00:20:46,720 Speaker 1: unhappy with that and the copycats of it, you know, 363 00:20:46,760 --> 00:20:49,119 Speaker 1: other models that do that because artists are worried that 364 00:20:49,600 --> 00:20:53,000 Speaker 1: it's using their artworks without permission. And you know, various 365 00:20:53,080 --> 00:20:55,960 Speaker 1: of these models are using artists artworks without permission in 366 00:20:56,080 --> 00:20:58,639 Speaker 1: order to generate a new art. The other thing that 367 00:20:58,880 --> 00:21:02,399 Speaker 1: we're talking about before is search. This chat bought function 368 00:21:02,480 --> 00:21:04,959 Speaker 1: and its ability to answer questions, whether it's doing it 369 00:21:04,960 --> 00:21:09,120 Speaker 1: correctly or incorrectly, has reinvigorated the search battle. Google has 370 00:21:09,119 --> 00:21:12,600 Speaker 1: been so dominant in search for so long, and Microsoft's 371 00:21:13,200 --> 00:21:15,600 Speaker 1: was essentially, you know, even though it has a couple 372 00:21:15,640 --> 00:21:18,320 Speaker 1: percentage points of share, pretty dead in the water in 373 00:21:18,400 --> 00:21:21,000 Speaker 1: terms of the competitive battle. People have been wondering for 374 00:21:21,080 --> 00:21:23,400 Speaker 1: years whether Microsoft would sell it or spin it off. 375 00:21:23,760 --> 00:21:27,000 Speaker 1: That competitive battle is so dead that when Microsoft announced 376 00:21:27,000 --> 00:21:29,919 Speaker 1: this new AI powered bing I couldn't even find a 377 00:21:30,000 --> 00:21:32,600 Speaker 1: research analyst that could give me market share data on 378 00:21:32,600 --> 00:21:35,200 Speaker 1: this because everybody had pulled their analysts off of covering 379 00:21:35,200 --> 00:21:38,040 Speaker 1: this market. But now all of a sudden, search seems 380 00:21:38,080 --> 00:21:41,000 Speaker 1: like it's potentially an open category again. Even if we're 381 00:21:41,000 --> 00:21:44,920 Speaker 1: not thinking Google's going to lose its dominant position, Google's 382 00:21:45,359 --> 00:21:48,359 Speaker 1: panicked enough that they try to kind of rush out 383 00:21:48,440 --> 00:21:52,000 Speaker 1: their own reaction to the being bought, which they're calling barred, 384 00:21:52,480 --> 00:21:54,480 Speaker 1: And even though they announced it a couple days before 385 00:21:54,520 --> 00:21:57,360 Speaker 1: Microsoft's event, it's still not really publicly available. We don't 386 00:21:57,359 --> 00:21:59,760 Speaker 1: really have any sense of what it looks like. Yeah, 387 00:22:00,200 --> 00:22:03,280 Speaker 1: one of the concerns about using these systems for search 388 00:22:03,320 --> 00:22:06,240 Speaker 1: engines goes back to what we were discussing earlier about 389 00:22:06,400 --> 00:22:11,000 Speaker 1: we don't know how often these systems make factual mistakes, 390 00:22:11,440 --> 00:22:15,440 Speaker 1: and search engines you rely on them to give you facts. 391 00:22:15,880 --> 00:22:19,120 Speaker 1: So one of the arguments that some computer scientists are 392 00:22:19,160 --> 00:22:23,560 Speaker 1: making now is that these systems, as magical and remarkable 393 00:22:23,680 --> 00:22:26,520 Speaker 1: as they are, and human like as they are, are 394 00:22:26,600 --> 00:22:30,240 Speaker 1: just not fit for purpose when it comes to search engines. 395 00:22:30,560 --> 00:22:32,800 Speaker 1: Because these are tools that are going to be used 396 00:22:32,840 --> 00:22:36,919 Speaker 1: by millions, hundreds of millions, potentially billions of people. You 397 00:22:37,000 --> 00:22:41,320 Speaker 1: cannot monitor everything that these systems are saying at that scale, 398 00:22:41,800 --> 00:22:45,480 Speaker 1: and so, you know, could we be facing another misinformation 399 00:22:45,600 --> 00:22:49,680 Speaker 1: epidemic When you think about people who are easily persuaded 400 00:22:50,200 --> 00:22:54,160 Speaker 1: or kids using these systems and they're taking everything that 401 00:22:54,200 --> 00:22:57,800 Speaker 1: these tools are telling them at face value, that could 402 00:22:57,800 --> 00:22:59,560 Speaker 1: really be a problem. And you know, it's funny that 403 00:22:59,600 --> 00:23:03,200 Speaker 1: for years we've had a misinformation problem from social media. 404 00:23:03,480 --> 00:23:06,480 Speaker 1: Maybe now we could be entering an age of misinformation 405 00:23:06,600 --> 00:23:11,680 Speaker 1: by algorithm, because, as Dina was saying earlier, Google was 406 00:23:11,720 --> 00:23:14,240 Speaker 1: caught on the back foot. It's racing to get out 407 00:23:14,240 --> 00:23:17,879 Speaker 1: this bot. As Bloomberg News is reported, they're trying to 408 00:23:17,880 --> 00:23:21,400 Speaker 1: put generative AI into all their products, rushing to unleash 409 00:23:21,440 --> 00:23:24,840 Speaker 1: these tools to the public. But are all the proper 410 00:23:24,920 --> 00:23:29,520 Speaker 1: safeguards really there. Yeah, that's something that I actually have 411 00:23:29,560 --> 00:23:32,439 Speaker 1: been thinking about a lot lately. This speed to market 412 00:23:32,760 --> 00:23:35,520 Speaker 1: makes me think about, like, can you think of another 413 00:23:35,560 --> 00:23:38,320 Speaker 1: time in like the last twenty years or so when 414 00:23:38,320 --> 00:23:41,840 Speaker 1: a tech company put out a product, and like the iPhone, 415 00:23:41,840 --> 00:23:44,520 Speaker 1: for instance, when the iPhone first came out, it wasn't 416 00:23:44,560 --> 00:23:47,320 Speaker 1: like here's a thing that can do a bunch of 417 00:23:47,320 --> 00:23:50,639 Speaker 1: things like moderately, well, you know, a decent amount of 418 00:23:50,640 --> 00:23:53,160 Speaker 1: the time, but like, let us know, you know, how 419 00:23:53,200 --> 00:23:55,320 Speaker 1: it does, and we'll improve it over time, and it's 420 00:23:55,320 --> 00:23:57,000 Speaker 1: going to be a okay. No, it had a very 421 00:23:57,040 --> 00:23:59,520 Speaker 1: limited set of features. It didn't even have an app store. 422 00:24:00,080 --> 00:24:01,600 Speaker 1: You know, I don't think it even had a flashlight 423 00:24:01,640 --> 00:24:03,480 Speaker 1: at that point. And you know it can make phone 424 00:24:03,480 --> 00:24:05,720 Speaker 1: calls and do a couple other things, get on the internet. 425 00:24:05,760 --> 00:24:08,399 Speaker 1: You weren't expected to be the beta tester of the 426 00:24:08,480 --> 00:24:11,399 Speaker 1: product and help improve it as it went. And I 427 00:24:11,440 --> 00:24:14,879 Speaker 1: feel like it's really interesting that consumers are at this 428 00:24:14,920 --> 00:24:18,320 Speaker 1: point expected to do that. And I'm wondering if consumers 429 00:24:18,320 --> 00:24:21,000 Speaker 1: are going to get wary and or weary of that, 430 00:24:21,119 --> 00:24:23,480 Speaker 1: and if so, like how long that's going to take. 431 00:24:23,800 --> 00:24:25,840 Speaker 1: It just seems like a lot to ask of somebody, 432 00:24:25,880 --> 00:24:28,240 Speaker 1: doesn't it, I mean, Rachel, One of the big problems 433 00:24:28,320 --> 00:24:31,960 Speaker 1: is that, as Diana was saying, GBT can generate a 434 00:24:31,960 --> 00:24:34,680 Speaker 1: photo that looks real, this idea of things that are 435 00:24:34,800 --> 00:24:39,520 Speaker 1: fake but look very very real. How does it deal 436 00:24:39,600 --> 00:24:42,840 Speaker 1: with that problem of fakery. We've heard about deep fakes, 437 00:24:42,880 --> 00:24:45,720 Speaker 1: but this is like a whole different level of that. Yeah, 438 00:24:45,760 --> 00:24:49,200 Speaker 1: and the fidelity of these images is getting better and furthermore, 439 00:24:49,240 --> 00:24:52,480 Speaker 1: there are other AI programs you can also use to, like, 440 00:24:52,560 --> 00:24:56,119 Speaker 1: in a sense, upscale the fidelity of an image. I've 441 00:24:56,320 --> 00:24:58,159 Speaker 1: played around with those a little bit and they're kind 442 00:24:58,240 --> 00:25:01,200 Speaker 1: they're pretty remarkable, you know that for really good things 443 00:25:01,240 --> 00:25:03,159 Speaker 1: like I have a sort of fuzzy old picture of 444 00:25:03,160 --> 00:25:05,240 Speaker 1: a family member and I want to make it look crisper. 445 00:25:05,640 --> 00:25:08,400 Speaker 1: But you could also use it to make something that's 446 00:25:08,440 --> 00:25:11,119 Speaker 1: fake look more realistic, or you could do it directly 447 00:25:11,119 --> 00:25:13,359 Speaker 1: with one of these programs. I think it's just going 448 00:25:13,440 --> 00:25:15,520 Speaker 1: to be tricky, and it might be one of those 449 00:25:15,560 --> 00:25:20,359 Speaker 1: things kind of like computer viruses and infosec, where there's 450 00:25:20,400 --> 00:25:23,160 Speaker 1: just a constant sort of battle to figure out what's 451 00:25:23,200 --> 00:25:25,280 Speaker 1: real versus what's fake and stay ahead of it, and 452 00:25:25,320 --> 00:25:28,000 Speaker 1: then it'll catch up, and then people will try to 453 00:25:28,000 --> 00:25:30,159 Speaker 1: get ahead of it. It might continue a pace for 454 00:25:30,200 --> 00:25:34,280 Speaker 1: a while. These are very good tools for creating misinformation 455 00:25:34,320 --> 00:25:37,520 Speaker 1: and disinformation at scale. One of the things we're seeing, 456 00:25:37,520 --> 00:25:40,480 Speaker 1: and you often see this in technology, often the first 457 00:25:40,600 --> 00:25:43,920 Speaker 1: at scale use for some new technology, as pornography. You're 458 00:25:43,960 --> 00:25:46,320 Speaker 1: already seeing that with these kinds of tools and deep 459 00:25:46,320 --> 00:25:49,480 Speaker 1: fake pornography where celebrities heads are being put on other 460 00:25:49,480 --> 00:25:51,880 Speaker 1: people's bodies doing things that the celebrities did not do. 461 00:25:52,119 --> 00:25:55,359 Speaker 1: Social media companies are happy to deal with removing those 462 00:25:55,440 --> 00:25:58,960 Speaker 1: removing fake voices. One of the big questions for the 463 00:25:59,000 --> 00:26:02,520 Speaker 1: companies creating these AI tools is what can they do 464 00:26:02,760 --> 00:26:07,280 Speaker 1: to clearly flag in watermarks and images or some sort 465 00:26:07,320 --> 00:26:10,399 Speaker 1: of digital flag in text that something is AI generated 466 00:26:10,720 --> 00:26:12,880 Speaker 1: so that people are aware of what they're dealing with. 467 00:26:13,320 --> 00:26:16,439 Speaker 1: There hasn't been much success yet doing that in a 468 00:26:16,480 --> 00:26:19,879 Speaker 1: reliable way, because, as Rachel says, people can always outsmart 469 00:26:19,880 --> 00:26:21,359 Speaker 1: those and you're playing a bit of a game of 470 00:26:21,359 --> 00:26:32,680 Speaker 1: whac amole. We'll be right back party is you said earlier. 471 00:26:32,680 --> 00:26:35,879 Speaker 1: We've seen lots of different versions of AI come along. 472 00:26:36,520 --> 00:26:39,639 Speaker 1: How much of this is I don't quite want to 473 00:26:39,640 --> 00:26:41,600 Speaker 1: call it a fat but so the latest thing in 474 00:26:41,680 --> 00:26:46,040 Speaker 1: how useful is this to build something even better in 475 00:26:46,080 --> 00:26:49,680 Speaker 1: the future. It's definitely not a fat. This is something 476 00:26:49,720 --> 00:26:53,760 Speaker 1: that people are already using. Chat GPT is the fastest 477 00:26:53,880 --> 00:26:58,080 Speaker 1: growing consumer internet tool in history. More than one hundred 478 00:26:58,080 --> 00:27:00,800 Speaker 1: million people signed up for it just a few months, 479 00:27:01,440 --> 00:27:04,440 Speaker 1: and they're not just playing around with it, they're actually 480 00:27:04,560 --> 00:27:07,320 Speaker 1: using it for work. So I think it's going to 481 00:27:07,480 --> 00:27:10,639 Speaker 1: fundamentally change a lot of things in terms of the 482 00:27:10,680 --> 00:27:15,239 Speaker 1: way professional workers work in all sorts of industries. You know, 483 00:27:15,440 --> 00:27:18,119 Speaker 1: not to be a Debbie downer, but I think this 484 00:27:18,240 --> 00:27:20,600 Speaker 1: is also going to have quite a negative impact on 485 00:27:20,880 --> 00:27:23,560 Speaker 1: the creative industry. And I think you know, when you 486 00:27:23,680 --> 00:27:28,199 Speaker 1: think about writing or generating images, these systems they're not 487 00:27:28,359 --> 00:27:32,160 Speaker 1: inherently creative, but they can do things that creative people do. 488 00:27:32,520 --> 00:27:35,280 Speaker 1: And I have heard from professionals who are using this 489 00:27:35,359 --> 00:27:39,600 Speaker 1: tool to for example, generative video script and the reason 490 00:27:39,640 --> 00:27:42,480 Speaker 1: they do that, they say to me, is it removes 491 00:27:42,560 --> 00:27:47,119 Speaker 1: friction from the creative process and it solves the blank 492 00:27:47,240 --> 00:27:51,440 Speaker 1: page syndrome problem. But you know, if you think about 493 00:27:51,480 --> 00:27:55,199 Speaker 1: the creative process throughout human history, that's what brings value 494 00:27:55,240 --> 00:27:58,560 Speaker 1: to artwork, and that's what brings value to literature is 495 00:27:59,359 --> 00:28:03,000 Speaker 1: the work that a human being puts into looking for 496 00:28:03,040 --> 00:28:05,520 Speaker 1: that word that they couldn't think of when their mind 497 00:28:05,600 --> 00:28:08,160 Speaker 1: is a vacuum. And I think it's a little bit 498 00:28:08,200 --> 00:28:10,840 Speaker 1: sad that we might lose some of that when machines 499 00:28:10,880 --> 00:28:13,840 Speaker 1: come in and solve that problem for us. It's much 500 00:28:13,880 --> 00:28:16,720 Speaker 1: harder to quantify and track that kind of impact, but 501 00:28:16,800 --> 00:28:19,720 Speaker 1: it is going to happen, and it just leaves me 502 00:28:19,760 --> 00:28:23,280 Speaker 1: a little bit melancholy. Dean. When you look ahead with 503 00:28:23,320 --> 00:28:25,560 Speaker 1: the kind of aid that we've been talking about it today, 504 00:28:25,760 --> 00:28:28,000 Speaker 1: what are the next things that you think we should 505 00:28:28,000 --> 00:28:31,520 Speaker 1: be looking for. We should keep looking for further penetration 506 00:28:31,680 --> 00:28:34,639 Speaker 1: into corporate use cases and whether these things really work 507 00:28:35,119 --> 00:28:37,760 Speaker 1: for talking about is this a fat or not? It's 508 00:28:37,800 --> 00:28:39,720 Speaker 1: not a fat, but we have to see how useful 509 00:28:39,760 --> 00:28:42,680 Speaker 1: it actually is for people and where it runs aground 510 00:28:43,400 --> 00:28:47,880 Speaker 1: or where it doesn't quite work technologically. I think we're 511 00:28:47,880 --> 00:28:50,680 Speaker 1: going to see in the short term some models for 512 00:28:50,840 --> 00:28:53,280 Speaker 1: video creation. We've talked about image creation, but there are 513 00:28:53,280 --> 00:28:57,720 Speaker 1: companies working on video creation as well, and we're going 514 00:28:57,760 --> 00:28:59,920 Speaker 1: to see regulation. That's one of the things we have 515 00:29:00,240 --> 00:29:02,720 Speaker 1: talked about. The European Union is looking at how to 516 00:29:02,760 --> 00:29:05,480 Speaker 1: regulate this. States in the US are looking at how 517 00:29:05,480 --> 00:29:08,000 Speaker 1: to regulate Congresses looked at it, but always seems to 518 00:29:08,040 --> 00:29:11,840 Speaker 1: be moving a little bit more slowly. Some companies want regulation, 519 00:29:12,000 --> 00:29:15,360 Speaker 1: but only as far as they want, not beyond what 520 00:29:15,400 --> 00:29:17,200 Speaker 1: they want to see regulated. They want to sort of 521 00:29:17,200 --> 00:29:20,000 Speaker 1: determine how far it goes and no further. Those are 522 00:29:20,120 --> 00:29:23,160 Speaker 1: going to be issues. There's discussion when We've written about 523 00:29:23,160 --> 00:29:25,520 Speaker 1: it here at Bloomberg about the climate impacts and the 524 00:29:25,560 --> 00:29:28,120 Speaker 1: carbon impact. One of the things we've talked a lot 525 00:29:28,160 --> 00:29:31,080 Speaker 1: at party mentioned how much better these models are getting. 526 00:29:31,320 --> 00:29:34,360 Speaker 1: The reason they're getting better is because they're getting bigger 527 00:29:34,360 --> 00:29:37,720 Speaker 1: and bigger, and they're running on ever more powerful supercomputers 528 00:29:37,760 --> 00:29:40,600 Speaker 1: inside of cloud data centers that are run by Microsoft 529 00:29:40,640 --> 00:29:43,800 Speaker 1: and Google and Amazon, and the carbon output of those 530 00:29:43,840 --> 00:29:47,000 Speaker 1: things is not insignificant. So there's a little bit of 531 00:29:47,040 --> 00:29:49,560 Speaker 1: a call for look, people need to actually reckon with 532 00:29:49,600 --> 00:29:52,320 Speaker 1: what the costs of these things are and whether the 533 00:29:52,440 --> 00:29:56,080 Speaker 1: use cases are worth the carbon output. Some of the 534 00:29:56,120 --> 00:29:59,480 Speaker 1: things we've been talking about are iterating getting better at 535 00:29:59,560 --> 00:30:02,920 Speaker 1: what you can already do, but there's some unexpected things 536 00:30:02,920 --> 00:30:06,040 Speaker 1: that we're not even thinking about that this technology may 537 00:30:06,080 --> 00:30:12,080 Speaker 1: suddenly be used for using AI to spot and understand 538 00:30:12,360 --> 00:30:15,400 Speaker 1: certain medical conditions. Cancer has been one that's been talked 539 00:30:15,400 --> 00:30:17,880 Speaker 1: about a lot. I think that that is actually a 540 00:30:18,200 --> 00:30:21,440 Speaker 1: great use of the technology. One thing that I hope 541 00:30:21,480 --> 00:30:25,120 Speaker 1: that we'll see more of is actually data sets that 542 00:30:25,240 --> 00:30:29,600 Speaker 1: are more varied and better ways that companies come up 543 00:30:29,640 --> 00:30:32,239 Speaker 1: with to fix a lot of the problems that they 544 00:30:32,280 --> 00:30:35,160 Speaker 1: have with societal biases. That pop up all over the 545 00:30:35,160 --> 00:30:38,200 Speaker 1: place when you're using these machines, because you can ask 546 00:30:38,200 --> 00:30:42,520 Speaker 1: a question and get gender biases, for instance, very easily 547 00:30:42,600 --> 00:30:45,600 Speaker 1: and matter of factly in the responses. But unless you 548 00:30:45,960 --> 00:30:48,400 Speaker 1: go another step and sort of interrogate the model, like 549 00:30:48,400 --> 00:30:51,200 Speaker 1: why did you give me this answer, it's not gonna 550 00:30:51,640 --> 00:30:54,000 Speaker 1: really give you much more about why it gave you 551 00:30:54,040 --> 00:30:55,440 Speaker 1: that kind of answer. A lot of people will just 552 00:30:55,480 --> 00:30:59,200 Speaker 1: accept its answer as fact. This happened to me a 553 00:30:59,200 --> 00:31:03,040 Speaker 1: couple of times. I was asking for nicknames for little kids, 554 00:31:03,080 --> 00:31:07,000 Speaker 1: like for toddlers for girls and the nicknames for boys, 555 00:31:07,040 --> 00:31:11,560 Speaker 1: and GPG four gave me some really like sort of stereotypical, 556 00:31:11,680 --> 00:31:15,959 Speaker 1: you know, like Rascal and Champ kind of names for boys, 557 00:31:16,120 --> 00:31:19,160 Speaker 1: and you know, sweet pea, jellybean kind of names for girls. 558 00:31:19,520 --> 00:31:22,200 Speaker 1: I didn't say why did you give me these biased names? 559 00:31:22,200 --> 00:31:24,040 Speaker 1: I just said, why did you suggest X, Y, and 560 00:31:24,120 --> 00:31:26,440 Speaker 1: Z for a boy and A B and C for 561 00:31:26,520 --> 00:31:29,000 Speaker 1: a girl? And then it said, ah, those were you know, 562 00:31:29,040 --> 00:31:31,800 Speaker 1: those were biased, and here are some more neutral names 563 00:31:33,280 --> 00:31:35,680 Speaker 1: to make what Rachel's saying a really concrete on the 564 00:31:35,720 --> 00:31:38,800 Speaker 1: potential negative impact of bias. Rachel was talking about, you know, 565 00:31:38,840 --> 00:31:41,080 Speaker 1: cancer applications and it can work if they get it 566 00:31:41,120 --> 00:31:44,920 Speaker 1: working across all populations. So here's the very specific problem. 567 00:31:45,200 --> 00:31:48,160 Speaker 1: You know, take breast cancer, because people are looking at 568 00:31:48,160 --> 00:31:50,960 Speaker 1: ways to scan mamograms for breast cancer more skin cancer. 569 00:31:51,040 --> 00:31:54,040 Speaker 1: So both of those conditions can present differently in people 570 00:31:54,080 --> 00:31:56,600 Speaker 1: of color. So the way breast cancer can look in 571 00:31:56,640 --> 00:31:58,800 Speaker 1: a black person who has breasts can be different. And 572 00:31:58,960 --> 00:32:02,000 Speaker 1: if your data set is insufficiently diverse, if it is 573 00:32:02,000 --> 00:32:04,800 Speaker 1: trained mostly on people who are white, people who are 574 00:32:05,200 --> 00:32:09,200 Speaker 1: Western Northern European, it may not pick up those breastcancers 575 00:32:09,200 --> 00:32:12,160 Speaker 1: and people with darker skin. And that's very dangerous because 576 00:32:12,160 --> 00:32:15,000 Speaker 1: you wind up switching a human scanner for breast cancer 577 00:32:15,440 --> 00:32:18,480 Speaker 1: for an AI one that is insufficiently prepared to find 578 00:32:18,480 --> 00:32:21,719 Speaker 1: breastcancer in all people. Literally people's lives are at if 579 00:32:21,760 --> 00:32:25,360 Speaker 1: you do that wrong. This technology is changing so rapidly. 580 00:32:25,360 --> 00:32:27,320 Speaker 1: It was only a few months ago that it was 581 00:32:27,360 --> 00:32:30,600 Speaker 1: even released out into the public and we all became 582 00:32:30,640 --> 00:32:34,240 Speaker 1: aware of it as it's happening right before our eyes 583 00:32:34,280 --> 00:32:37,080 Speaker 1: and we all have access to it. What should normal 584 00:32:37,160 --> 00:32:40,080 Speaker 1: people do when they're approaching head How should they use it? 585 00:32:40,080 --> 00:32:43,000 Speaker 1: What should they be careful about? Like, what are good 586 00:32:43,160 --> 00:32:48,840 Speaker 1: rules of the road in using this technology. Check everything. Yeah, 587 00:32:48,880 --> 00:32:52,520 Speaker 1: I would approach it cautiously, think about how much do 588 00:32:52,520 --> 00:32:55,280 Speaker 1: you want to trust it with personal data. You may 589 00:32:55,360 --> 00:33:00,000 Speaker 1: want to not give these systems personal information. I wouldn't 590 00:33:00,240 --> 00:33:04,760 Speaker 1: at this point, and definitely look with some skepticism at 591 00:33:04,760 --> 00:33:07,640 Speaker 1: the results you get and feel free, especially if it's 592 00:33:07,640 --> 00:33:11,239 Speaker 1: a chat about specifically, ask follow up questions. You know, 593 00:33:11,240 --> 00:33:14,120 Speaker 1: whether or not you think the information is accurate that 594 00:33:14,160 --> 00:33:17,360 Speaker 1: you're getting from it. You can always type something like 595 00:33:17,880 --> 00:33:20,560 Speaker 1: why did you give me that? Why is that the 596 00:33:20,680 --> 00:33:24,280 Speaker 1: right answer? And especially with like GPT four, it'll text 597 00:33:24,320 --> 00:33:27,360 Speaker 1: back you know, here is why I think that, and 598 00:33:27,400 --> 00:33:30,840 Speaker 1: then you might say, wow, this is completely off base. 599 00:33:30,920 --> 00:33:32,760 Speaker 1: It did this to me with a few different things. 600 00:33:32,920 --> 00:33:34,480 Speaker 1: So then you go, okay, I'm not going to trust 601 00:33:34,520 --> 00:33:36,920 Speaker 1: you for this task, or it might give you some 602 00:33:37,080 --> 00:33:39,520 Speaker 1: really good reasoning that you can then just double check 603 00:33:39,560 --> 00:33:41,320 Speaker 1: and say, Okay, this was actually really helpful to me, 604 00:33:41,560 --> 00:33:45,760 Speaker 1: but definitely check everything. Harmy, Rachel Dina, thanks for coming 605 00:33:45,760 --> 00:33:49,960 Speaker 1: on the show. Thank you, thank you, thanks thanks for 606 00:33:50,000 --> 00:33:52,160 Speaker 1: listening to us. Here at the Big Take. It's a 607 00:33:52,200 --> 00:33:57,240 Speaker 1: daily podcast from Bloomberg and iHeartRadio. For more shows from iHeartRadio, 608 00:33:57,720 --> 00:34:01,600 Speaker 1: visit the iHeartRadio app, Apple podcast or wherever you listen, 609 00:34:01,960 --> 00:34:04,720 Speaker 1: and we'd look to hear from you emails with questions 610 00:34:04,800 --> 00:34:08,600 Speaker 1: or comments at Big Take at Bloomberg dot Net. The 611 00:34:08,680 --> 00:34:12,359 Speaker 1: supervising producer of The Big Take is Vicky Vergolina. Our 612 00:34:12,400 --> 00:34:16,400 Speaker 1: senior producer is Katherine Fink. Our producers are Michael Fallero 613 00:34:16,520 --> 00:34:21,080 Speaker 1: and Mobarrow. Raphael mci lee is our engineer. Original music 614 00:34:21,120 --> 00:34:25,160 Speaker 1: by Leosidrin and I'm Ava, a voice avatar created by 615 00:34:25,160 --> 00:34:28,680 Speaker 1: Well Said Labs, and I'm West Kasova. We'll be back 616 00:34:28,719 --> 00:34:30,600 Speaker 1: tomorrow with another Big Take