1 00:00:03,320 --> 00:00:06,520 Speaker 1: Today on the Big Take. Concerns about AI are growing 2 00:00:06,600 --> 00:00:11,280 Speaker 1: from whistleblowers and from industry leaders. I'm Craig Gordon in 3 00:00:11,320 --> 00:00:23,200 Speaker 1: for West Kosova. Artificial intelligence or AI has been getting 4 00:00:23,239 --> 00:00:26,480 Speaker 1: quite a bit of attention lately. Even as it promises 5 00:00:26,560 --> 00:00:29,960 Speaker 1: to revolutionize the way we think and work, AI's positioned 6 00:00:29,960 --> 00:00:31,240 Speaker 1: to bring headaches as well. 7 00:00:31,640 --> 00:00:35,520 Speaker 2: Pictures online of a bombing at the Pentagon in the US, 8 00:00:35,600 --> 00:00:38,200 Speaker 2: and it was AI generated, but obviously so many people 9 00:00:38,320 --> 00:00:41,400 Speaker 2: very quickly panicked thinking that'd been a bombing that happened. 10 00:00:41,640 --> 00:00:44,319 Speaker 3: AI has even infiltrated music. Now there's a new song 11 00:00:44,360 --> 00:00:45,760 Speaker 3: I don't know if you've heard about by Drake in 12 00:00:45,840 --> 00:00:49,199 Speaker 3: the Weekend that wasn't made by Drake or the Weekend 13 00:00:49,360 --> 00:00:52,200 Speaker 3: was created by artificial intelligence. 14 00:00:52,200 --> 00:00:55,240 Speaker 1: In March, thousands of tech leaders signed an open letter 15 00:00:55,760 --> 00:00:59,680 Speaker 1: calling for a pause in AI development. Even more recently, 16 00:00:59,720 --> 00:01:03,200 Speaker 1: three and fifty industry leaders signed a second open letter 17 00:01:03,400 --> 00:01:08,160 Speaker 1: urgent caution. Their letter was just one sentence, mitigating the 18 00:01:08,240 --> 00:01:11,640 Speaker 1: risk of extinction from AI should be a global priority 19 00:01:11,840 --> 00:01:17,160 Speaker 1: alongside other societal scale risks such as pandemics and nuclear war. 20 00:01:18,240 --> 00:01:21,839 Speaker 1: How realistic is this ominous warning? I spoke to Bloomberg 21 00:01:21,920 --> 00:01:25,000 Speaker 1: AI reporters Dina Bass and Rachel Mattz to find out. 22 00:01:27,520 --> 00:01:29,119 Speaker 1: I would love to hear from you, Dina, a little 23 00:01:29,160 --> 00:01:32,400 Speaker 1: reality check. What is AI capable of doing right now? 24 00:01:32,600 --> 00:01:34,760 Speaker 1: What is it not capable of doing? How can we 25 00:01:34,800 --> 00:01:39,760 Speaker 1: put this technology to use for humankind on an individual basis. 26 00:01:40,560 --> 00:01:41,319 Speaker 4: AI is not. 27 00:01:41,480 --> 00:01:43,959 Speaker 5: New, and we've been talking about both promise and the 28 00:01:44,200 --> 00:01:47,240 Speaker 5: challenges and concerns of lots of different types of AI 29 00:01:47,360 --> 00:01:49,000 Speaker 5: for a number of years. 30 00:01:49,080 --> 00:01:51,800 Speaker 4: What has happened in the last year that has. 31 00:01:51,680 --> 00:01:55,520 Speaker 5: Started the current hype churn and excitement cycle that we 32 00:01:55,600 --> 00:01:57,559 Speaker 5: have is something called generative AI. 33 00:01:58,040 --> 00:02:00,440 Speaker 4: And the difference in terms of what we're talking about out. 34 00:02:00,280 --> 00:02:04,680 Speaker 5: Here is that there are AI algorithms, AI models that 35 00:02:05,040 --> 00:02:08,799 Speaker 5: basically suck up a lot of information. You know, pictures 36 00:02:09,000 --> 00:02:12,800 Speaker 5: aren't text from across the Internet, from you know, Reddit 37 00:02:12,880 --> 00:02:15,600 Speaker 5: and social media things like that, and what these models 38 00:02:15,639 --> 00:02:19,760 Speaker 5: do is use all of that information to generate new 39 00:02:19,880 --> 00:02:22,560 Speaker 5: content in some respect. And so when you start talking 40 00:02:22,560 --> 00:02:25,160 Speaker 5: about a picture that purports to be an explosion near 41 00:02:25,200 --> 00:02:28,600 Speaker 5: the Pentagon, that's new content that's created in some way 42 00:02:28,680 --> 00:02:29,040 Speaker 5: by the. 43 00:02:29,040 --> 00:02:30,639 Speaker 4: Artificial intelligence systems. 44 00:02:31,040 --> 00:02:33,400 Speaker 5: What people are getting a little tripped up in, though, 45 00:02:33,520 --> 00:02:35,560 Speaker 5: is you know people will use these sort of human 46 00:02:35,880 --> 00:02:38,400 Speaker 5: like terms for what the AI is doing. They'll and 47 00:02:38,480 --> 00:02:41,400 Speaker 5: thropple morphize things because it's the easiest way to understand it, 48 00:02:41,440 --> 00:02:44,840 Speaker 5: and they'll ascribe kind of human like intelligence to these systems. 49 00:02:45,240 --> 00:02:48,120 Speaker 5: They are not human, they are not thinking, they aren't 50 00:02:48,200 --> 00:02:52,200 Speaker 5: producing art. What they're doing, in many cases is making predictions, 51 00:02:52,280 --> 00:02:55,800 Speaker 5: making guesses, generating things that are sort of an imitation 52 00:02:56,120 --> 00:02:58,280 Speaker 5: of the other things that they've seen. And so we 53 00:02:58,360 --> 00:02:59,960 Speaker 5: need to be very careful to understand. 54 00:03:00,040 --> 00:03:02,120 Speaker 4: It's definitely not human. It's not even human like. 55 00:03:02,720 --> 00:03:05,440 Speaker 1: That said, as we have seen some examples, it can 56 00:03:05,560 --> 00:03:08,120 Speaker 1: sort of mimic human speech, It appears to mimic sort 57 00:03:08,120 --> 00:03:11,440 Speaker 1: of human thought. How can people sort of draw the distinction. 58 00:03:11,800 --> 00:03:13,320 Speaker 1: How do we tell the difference between a human and 59 00:03:13,360 --> 00:03:15,360 Speaker 1: an AI voice that sounds like a human. 60 00:03:16,120 --> 00:03:18,560 Speaker 5: That's a real concern. I think people are grappling with 61 00:03:18,600 --> 00:03:20,240 Speaker 5: what we're going to do there. At the beginning of 62 00:03:20,280 --> 00:03:24,160 Speaker 5: the Senate hearing to talk about the artificial intelligence, Senator 63 00:03:24,200 --> 00:03:28,239 Speaker 5: Blumenthal had chat GPT from open AI write his introductory 64 00:03:28,280 --> 00:03:31,960 Speaker 5: speech and if he had an audio generation artificial intelligence 65 00:03:32,000 --> 00:03:35,000 Speaker 5: algorithm actually speak it sounding like it was him. 66 00:03:34,880 --> 00:03:38,680 Speaker 6: And how the lack of transparency can undermine public trust. 67 00:03:39,760 --> 00:03:44,360 Speaker 6: This is not the future we want. If you were 68 00:03:44,440 --> 00:03:47,440 Speaker 6: listening from home, you might have thought that voice was 69 00:03:47,520 --> 00:03:52,240 Speaker 6: mine and the words from me, But in fact that 70 00:03:52,440 --> 00:03:58,920 Speaker 6: voice was not mine, the words were not mine, and 71 00:03:59,200 --> 00:04:04,280 Speaker 6: the audio was an AI voice cloning software trained on 72 00:04:04,320 --> 00:04:08,800 Speaker 6: my floor speeches. The remarks were written by chat gpt. 73 00:04:09,880 --> 00:04:11,600 Speaker 5: There's a lot of discussion about how to kind of 74 00:04:11,640 --> 00:04:14,880 Speaker 5: watermark these things. It's potentially a little bit easier with 75 00:04:14,920 --> 00:04:17,800 Speaker 5: images text it can be very difficult to tell. 76 00:04:17,880 --> 00:04:19,520 Speaker 4: Open AI itsself released an. 77 00:04:19,480 --> 00:04:22,320 Speaker 5: Algorithm that was intended to help people figure out if 78 00:04:22,360 --> 00:04:25,159 Speaker 5: chat gpt had authored a piece of content, but it's 79 00:04:25,200 --> 00:04:28,120 Speaker 5: not terribly accurate, and as a decently high false positive 80 00:04:28,160 --> 00:04:30,200 Speaker 5: rate as well. But there is some work right now 81 00:04:30,200 --> 00:04:32,479 Speaker 5: in technology to sort of help flag to people that 82 00:04:32,560 --> 00:04:36,120 Speaker 5: something was generated by an artificial intelligence algorithm and is 83 00:04:36,160 --> 00:04:38,080 Speaker 5: not human authored or human drawn. 84 00:04:40,320 --> 00:04:42,719 Speaker 1: So, Rachel, I know a lot of these different technologies 85 00:04:42,760 --> 00:04:45,200 Speaker 1: get lumped together, but maybe you could help us understand 86 00:04:45,200 --> 00:04:47,960 Speaker 1: the difference between the sort of flavors of AI. 87 00:04:48,720 --> 00:04:51,800 Speaker 7: I like to think of it as different kinds of architecture. 88 00:04:52,120 --> 00:04:54,280 Speaker 7: I like it a lot because it's easy for us 89 00:04:54,320 --> 00:04:57,080 Speaker 7: as people to understand architecture in general. Right, we have 90 00:04:57,120 --> 00:05:00,559 Speaker 7: different types of buildings, they serve different purposes, They change 91 00:05:00,600 --> 00:05:02,839 Speaker 7: over time, and things might fall out of fashion and 92 00:05:02,880 --> 00:05:04,160 Speaker 7: then come back into fashion. 93 00:05:04,680 --> 00:05:06,680 Speaker 1: So, Diana, what are some of the leading industries that 94 00:05:06,720 --> 00:05:08,880 Speaker 1: are adapting this technology. 95 00:05:09,279 --> 00:05:10,839 Speaker 4: The newest generative AI stuff. 96 00:05:10,839 --> 00:05:13,560 Speaker 5: I think we're just seeing industries kind of climb on 97 00:05:13,600 --> 00:05:15,000 Speaker 5: board with I mean, there's a little bit of a 98 00:05:15,040 --> 00:05:17,520 Speaker 5: panic for every company to figure out what their AI 99 00:05:17,640 --> 00:05:20,960 Speaker 5: strategy is and how to use it. But we're starting 100 00:05:21,000 --> 00:05:23,960 Speaker 5: to see it, certainly in the graphic design space because 101 00:05:24,000 --> 00:05:26,600 Speaker 5: some of the image generation stuff is a little bit 102 00:05:26,640 --> 00:05:29,160 Speaker 5: older than the Chat GPT stuff. It preceded it by 103 00:05:29,200 --> 00:05:31,760 Speaker 5: a few months, and so we're definitely seeing a lot 104 00:05:31,800 --> 00:05:33,799 Speaker 5: of that in the kind of the graphic design and artwork. 105 00:05:34,160 --> 00:05:37,719 Speaker 5: But we're seeing people use it for finance, for legal Obviously, 106 00:05:37,760 --> 00:05:40,359 Speaker 5: there's been a lot of discussion about academic use cases. 107 00:05:40,680 --> 00:05:43,440 Speaker 5: There's been a lot of focus on cheating, but there are, 108 00:05:43,560 --> 00:05:47,320 Speaker 5: you know, non cheating applications that can help students learn 109 00:05:47,400 --> 00:05:49,400 Speaker 5: or help them draft as long as they're clear with 110 00:05:49,440 --> 00:05:52,480 Speaker 5: their teachers what they're doing. There really are I think 111 00:05:52,520 --> 00:05:54,839 Speaker 5: a pretty wide array of use cases I hear, you know, 112 00:05:55,240 --> 00:05:57,839 Speaker 5: lots of banking executives, lots of Wall Street executives tripping 113 00:05:57,839 --> 00:05:59,400 Speaker 5: over each other to talk about who is going to 114 00:05:59,400 --> 00:06:03,320 Speaker 5: have these smart generative BAI strategy. So it's pretty across 115 00:06:03,320 --> 00:06:05,400 Speaker 5: the board in terms of people trying to say that 116 00:06:05,440 --> 00:06:09,040 Speaker 5: they're adopting it. What will actually get used in practice? 117 00:06:09,480 --> 00:06:11,119 Speaker 5: I think it's going to take a little while longer 118 00:06:11,160 --> 00:06:13,320 Speaker 5: to know, and potentially even longer than that to have 119 00:06:13,360 --> 00:06:17,360 Speaker 5: a sense of whether these replace workers, replace entry level 120 00:06:17,440 --> 00:06:21,320 Speaker 5: jobs or the rosier scenario that Silicon Valley and Microsoft 121 00:06:21,400 --> 00:06:23,320 Speaker 5: up here in the Seattle area like to talk about. 122 00:06:23,320 --> 00:06:25,760 Speaker 5: The rosier scenario is that they make people more productive 123 00:06:25,760 --> 00:06:27,039 Speaker 5: but don't put them out of work. 124 00:06:27,560 --> 00:06:29,560 Speaker 1: What specifically do banks do with it? Can they use 125 00:06:29,600 --> 00:06:32,480 Speaker 1: it to decide whether I should get my mortgage or 126 00:06:32,680 --> 00:06:34,480 Speaker 1: what would be a good use for a bank. 127 00:06:35,080 --> 00:06:37,479 Speaker 5: Banks have actually been using AI to decide whether you 128 00:06:37,520 --> 00:06:39,400 Speaker 5: get your mortgage for a long time, and it's found 129 00:06:39,400 --> 00:06:42,880 Speaker 5: to be problematic because it has racial and geographic bias 130 00:06:42,880 --> 00:06:46,200 Speaker 5: implications when you do things like that. Also, banks produce 131 00:06:46,240 --> 00:06:51,480 Speaker 5: a tremendous volume of written content, analysis notes to clients, 132 00:06:51,560 --> 00:06:54,880 Speaker 5: research reports, things like that, But people are that trade 133 00:06:54,920 --> 00:06:56,839 Speaker 5: are always looking for anything that can give them an 134 00:06:56,880 --> 00:06:59,760 Speaker 5: advantage as well. So for several years, probably more than that, 135 00:06:59,760 --> 00:07:01,240 Speaker 5: people been trying to figure out how they can use 136 00:07:01,240 --> 00:07:03,320 Speaker 5: algorithms to give them a better you know, trade a 137 00:07:03,320 --> 00:07:06,400 Speaker 5: little bit faster, get information a little bit faster. The 138 00:07:06,440 --> 00:07:08,960 Speaker 5: mortgage and lending scenario there's a real problematic one. 139 00:07:09,640 --> 00:07:11,880 Speaker 1: So we hear a lot about this technology known as 140 00:07:11,960 --> 00:07:14,160 Speaker 1: generative AI, and that seems to be kind of the 141 00:07:14,240 --> 00:07:16,400 Speaker 1: umbrella topic for a lot of the things that people 142 00:07:16,480 --> 00:07:18,760 Speaker 1: talk about and write about in the media. Explain what 143 00:07:18,800 --> 00:07:21,320 Speaker 1: that is and where is that leading us in the future. 144 00:07:22,080 --> 00:07:24,360 Speaker 7: A lot of this is just still an experimentation phase, 145 00:07:24,560 --> 00:07:27,400 Speaker 7: and I think what we'll probably see is these models 146 00:07:27,480 --> 00:07:31,160 Speaker 7: used increasingly to train on very specific data sets. Like 147 00:07:31,240 --> 00:07:34,520 Speaker 7: right now you have large language models such as that 148 00:07:34,520 --> 00:07:37,160 Speaker 7: that underpins chat GPT. It's meant to be kind of 149 00:07:37,200 --> 00:07:39,680 Speaker 7: a general purpose language model. You know, you can sort 150 00:07:39,720 --> 00:07:41,880 Speaker 7: of ask it or you know, type to it any 151 00:07:41,960 --> 00:07:44,440 Speaker 7: kind of question. It'll give you some kind of answer 152 00:07:44,480 --> 00:07:47,600 Speaker 7: that may or may not be accurate. It's basically trying 153 00:07:47,640 --> 00:07:51,000 Speaker 7: to give you what you want based on its training data. 154 00:07:51,200 --> 00:07:52,960 Speaker 7: But I think that what we're going to be starting 155 00:07:52,960 --> 00:07:55,760 Speaker 7: to see more and more of is these large language 156 00:07:55,760 --> 00:07:59,360 Speaker 7: models used in more like specific ways with more specific 157 00:07:59,440 --> 00:08:01,640 Speaker 7: types of trains data, and then you might actually be 158 00:08:01,680 --> 00:08:04,200 Speaker 7: able to get better answers, like maybe able to use 159 00:08:04,240 --> 00:08:07,320 Speaker 7: it with certain subsets of medical data. Let's say, so 160 00:08:07,400 --> 00:08:10,280 Speaker 7: you can use it to answer medical questions in a 161 00:08:10,320 --> 00:08:12,679 Speaker 7: more accurate way than you could with just the basic 162 00:08:12,760 --> 00:08:14,000 Speaker 7: model as it exists now. 163 00:08:14,480 --> 00:08:16,000 Speaker 5: Yeah, and I feel like the other thing that would 164 00:08:16,040 --> 00:08:18,080 Speaker 5: be an interesting shift about that as we stop competing 165 00:08:18,120 --> 00:08:20,920 Speaker 5: on size, like you know, like right now everything is 166 00:08:21,680 --> 00:08:24,720 Speaker 5: size matters. The bigger models do better if we are 167 00:08:24,760 --> 00:08:28,160 Speaker 5: able to train on smaller, more specific to task models. 168 00:08:28,240 --> 00:08:29,920 Speaker 5: They don't have to compete on size, and that can 169 00:08:29,960 --> 00:08:32,720 Speaker 5: be useful in a few ways. Running these large language 170 00:08:32,760 --> 00:08:35,160 Speaker 5: models is very expensive, both in a dollar sense and 171 00:08:35,200 --> 00:08:38,800 Speaker 5: an environmental sense, and because the data sets are so large, 172 00:08:38,840 --> 00:08:42,160 Speaker 5: they become almost impossible to check for harmful content for 173 00:08:42,280 --> 00:08:45,800 Speaker 5: biased content. Smaller could be both cheaper to run and 174 00:08:46,160 --> 00:08:47,360 Speaker 5: have greater quality control. 175 00:08:48,679 --> 00:08:50,960 Speaker 1: Rachel Beck. In March, an open letter was signed by 176 00:08:51,000 --> 00:08:53,840 Speaker 1: thousands of tech leaders regarding some of the apprehensions they 177 00:08:53,840 --> 00:08:57,400 Speaker 1: have over AI technology. What are some of the specific 178 00:08:57,480 --> 00:08:59,160 Speaker 1: concerns they laid out in that. 179 00:08:59,160 --> 00:09:04,079 Speaker 7: Letter, the basic sort of overarching theme there was there 180 00:09:04,120 --> 00:09:06,960 Speaker 7: should be a time period pause, like a six month 181 00:09:07,080 --> 00:09:11,520 Speaker 7: pause on developing AI more powerful than GPT four I 182 00:09:11,600 --> 00:09:14,079 Speaker 7: believe it was, which was the currently released state of 183 00:09:14,120 --> 00:09:18,160 Speaker 7: the art model from open Ai. Right now, OpenAI is 184 00:09:18,520 --> 00:09:21,280 Speaker 7: very much acknowledged as one of just a few leaders 185 00:09:21,440 --> 00:09:25,680 Speaker 7: in the AI industry. I think what's actually really interesting 186 00:09:25,720 --> 00:09:28,080 Speaker 7: about this, besides the fact that, as you pointed out, 187 00:09:28,360 --> 00:09:30,400 Speaker 7: a lot of people who have been working in and 188 00:09:30,440 --> 00:09:33,920 Speaker 7: around the AI industry for years signing onto this letter, 189 00:09:34,480 --> 00:09:37,280 Speaker 7: there have already been lots in lots of people that 190 00:09:37,360 --> 00:09:41,880 Speaker 7: have been shouting from the rooftops about existing very real 191 00:09:42,000 --> 00:09:45,600 Speaker 7: current issues with the AI systems that are already in place, 192 00:09:45,640 --> 00:09:48,280 Speaker 7: and that seemed to have gotten lost in the conversation there, 193 00:09:48,320 --> 00:09:50,839 Speaker 7: And I felt like, to me that was kind of 194 00:09:50,880 --> 00:09:52,920 Speaker 7: like a big kind of question mark. 195 00:09:54,760 --> 00:09:57,160 Speaker 1: With all of this talk about a six month pause, 196 00:09:57,360 --> 00:09:59,679 Speaker 1: is anyone actually hitting the brakes on their efforts to 197 00:09:59,679 --> 00:10:02,199 Speaker 1: develop up the next version of this I. 198 00:10:02,160 --> 00:10:04,360 Speaker 5: Saw something the other day that just said that open 199 00:10:04,400 --> 00:10:07,120 Speaker 5: ai is not yet working on GPT five or something 200 00:10:07,200 --> 00:10:09,720 Speaker 5: like that. But to Rachel's point, some of the motivation 201 00:10:09,840 --> 00:10:11,880 Speaker 5: behind the letter was a little unclear. Some of the 202 00:10:11,880 --> 00:10:15,559 Speaker 5: people signing it were potentially working on competing products. 203 00:10:15,559 --> 00:10:16,400 Speaker 4: It's not really clear. 204 00:10:16,440 --> 00:10:20,240 Speaker 5: Also, what does a six month pause do? Why six months? 205 00:10:20,280 --> 00:10:22,040 Speaker 5: What do you do at the end of six months? 206 00:10:22,480 --> 00:10:24,960 Speaker 5: And is that really the way to go? I think, 207 00:10:25,400 --> 00:10:27,959 Speaker 5: you know, to Rachel's point, a lot of the AI 208 00:10:28,040 --> 00:10:30,600 Speaker 5: ethics and responsible AI people who've been working in this 209 00:10:30,679 --> 00:10:33,880 Speaker 5: field for five, six or longer years, and many of 210 00:10:33,880 --> 00:10:36,920 Speaker 5: them are in the populations that are most affected by 211 00:10:36,920 --> 00:10:38,240 Speaker 5: the current problems with AI. 212 00:10:38,280 --> 00:10:40,200 Speaker 4: It's a number of women, a number of people of color. 213 00:10:40,520 --> 00:10:42,920 Speaker 5: You know, many of them have been calling for slower 214 00:10:42,960 --> 00:10:46,680 Speaker 5: approaches for a long time, asking companies developing these things 215 00:10:46,720 --> 00:10:50,040 Speaker 5: to take the time to make sure that products are 216 00:10:50,320 --> 00:10:54,000 Speaker 5: responsible it don't cause serious harm before they release them, 217 00:10:54,120 --> 00:10:57,760 Speaker 5: and also calling for regulation of some sort by various governments. 218 00:10:57,880 --> 00:10:59,800 Speaker 5: So there's been a call for a while to slow 219 00:10:59,840 --> 00:11:02,520 Speaker 5: down and make sure that you're not going by the 220 00:11:02,559 --> 00:11:05,600 Speaker 5: typical tech adage of move fast and break things. 221 00:11:05,640 --> 00:11:07,120 Speaker 4: But the idea of. 222 00:11:07,040 --> 00:11:10,240 Speaker 5: A specific six month pause, and then what and how 223 00:11:10,280 --> 00:11:12,880 Speaker 5: would you even enforce that six month pause. I don't 224 00:11:12,880 --> 00:11:14,280 Speaker 5: know if they just seized on that because it was 225 00:11:14,320 --> 00:11:16,559 Speaker 5: attention getting or it was something you could just say, oh, 226 00:11:16,559 --> 00:11:19,720 Speaker 5: this is a specific proposal, but it wasn't completely clear 227 00:11:19,760 --> 00:11:21,640 Speaker 5: what that actually would look like in practice and. 228 00:11:21,600 --> 00:11:25,199 Speaker 1: What it would do when we come back. Worries about 229 00:11:25,200 --> 00:11:27,559 Speaker 1: the direction of AI have been around as long as 230 00:11:27,600 --> 00:11:30,760 Speaker 1: the technology has. We'll look at what the concerns are 231 00:11:30,800 --> 00:11:41,280 Speaker 1: and who has been raising them. I wanted to go 232 00:11:41,280 --> 00:11:44,319 Speaker 1: a little bit deeper on this idea. One of these whistleblowers, 233 00:11:44,360 --> 00:11:47,280 Speaker 1: if we can call them that, Timny Gibrew from Google, 234 00:11:47,679 --> 00:11:50,120 Speaker 1: was probably one of the most prominent early critics of 235 00:11:50,160 --> 00:11:52,720 Speaker 1: this for some of the reasons that you cited related 236 00:11:52,720 --> 00:11:54,800 Speaker 1: to the questions of whether women and people of color 237 00:11:54,840 --> 00:11:57,600 Speaker 1: would be treated fairly or treated responsibly by some of 238 00:11:57,600 --> 00:12:00,520 Speaker 1: these technologies. Dina, maybe you could tell us bit about 239 00:12:00,520 --> 00:12:02,200 Speaker 1: her story and what her concerns were. 240 00:12:02,840 --> 00:12:05,320 Speaker 5: So she was one of the pioneers of sort of 241 00:12:05,360 --> 00:12:09,319 Speaker 5: looking at the harms of various AI systems. 242 00:12:09,600 --> 00:12:10,719 Speaker 4: She co authored with. 243 00:12:10,800 --> 00:12:13,560 Speaker 5: Joy Bloom Winnie a landmark paper in twenty eighteen that 244 00:12:13,640 --> 00:12:16,800 Speaker 5: showed that a lot of the most popular facial recognition 245 00:12:17,000 --> 00:12:21,040 Speaker 5: products were just performing spectacularly badly when they were looking 246 00:12:21,040 --> 00:12:24,360 Speaker 5: at images of people of color in general, but particularly women, 247 00:12:24,480 --> 00:12:26,880 Speaker 5: and choose at Microsoft. At that point, she moves to 248 00:12:26,960 --> 00:12:30,920 Speaker 5: Google and along with Margaret Mitchell, they co found an 249 00:12:31,000 --> 00:12:34,080 Speaker 5: ethical AI group at Google, and they start trying to 250 00:12:34,720 --> 00:12:37,760 Speaker 5: make Google's AI scientists pay more attention to some of 251 00:12:37,800 --> 00:12:40,199 Speaker 5: the problems in the algorithms that they were working on. 252 00:12:40,640 --> 00:12:44,040 Speaker 5: This ultimately ends up in her dismissal from Google. Several 253 00:12:44,040 --> 00:12:47,760 Speaker 5: months later, Margaret Mitchell is also dismissed from Google, basically 254 00:12:47,800 --> 00:12:51,400 Speaker 5: decapitating this ethical AI group. One of the things that 255 00:12:51,440 --> 00:12:54,040 Speaker 5: this sort of raises is some of these, you know, 256 00:12:54,160 --> 00:12:57,679 Speaker 5: people that are now speaking up about concerns in the 257 00:12:57,720 --> 00:13:00,560 Speaker 5: AI work, including Jeff Hinton, who's one of the you know, 258 00:13:00,559 --> 00:13:03,000 Speaker 5: the pioneers of the current generation of AI that we 259 00:13:03,040 --> 00:13:06,679 Speaker 5: have are Google people, and they're you know, waiting essentially 260 00:13:06,679 --> 00:13:09,320 Speaker 5: to twenty twenty three to speak up about this and 261 00:13:09,480 --> 00:13:12,200 Speaker 5: did not, in any way that I know of, really 262 00:13:12,320 --> 00:13:15,360 Speaker 5: back or stick up for or anything else. The people 263 00:13:15,360 --> 00:13:18,400 Speaker 5: that were at Google several years ago that were very 264 00:13:18,440 --> 00:13:20,480 Speaker 5: much whistle lowers and the truest sense of the word 265 00:13:20,640 --> 00:13:22,720 Speaker 5: that they were fired for airing their concern. 266 00:13:22,960 --> 00:13:24,760 Speaker 1: And so what is Google's response to all this? 267 00:13:26,000 --> 00:13:30,000 Speaker 5: So there is you know, a dispute about what happened 268 00:13:30,080 --> 00:13:34,280 Speaker 5: between Google and doctor Timmy Debrew. Google said at the 269 00:13:34,320 --> 00:13:39,040 Speaker 5: time that they accepted her resignation. Doctor Gabrew maintains that 270 00:13:39,080 --> 00:13:42,560 Speaker 5: she offered no such resignation. With regard to doctor Mitchell, 271 00:13:42,880 --> 00:13:44,319 Speaker 5: Google says that they fired her. 272 00:13:44,840 --> 00:13:47,560 Speaker 1: Is there anything that we could point to in the 273 00:13:47,640 --> 00:13:51,280 Speaker 1: technology that exists now where these concerns have been addressed? 274 00:13:52,000 --> 00:13:55,000 Speaker 5: Both Google and Microsoft have significant groups of people that 275 00:13:55,080 --> 00:13:57,840 Speaker 5: work on responsible AI. All of the new things that 276 00:13:57,880 --> 00:14:00,760 Speaker 5: Microsoft has put out they have taken and pains to 277 00:14:00,880 --> 00:14:04,120 Speaker 5: tell us have gone through their responsible AI reviews and 278 00:14:04,200 --> 00:14:06,840 Speaker 5: continue to as people test the products and they get 279 00:14:06,880 --> 00:14:10,800 Speaker 5: more feedback about what is and isn't working. Microsoft continually 280 00:14:11,120 --> 00:14:14,400 Speaker 5: every time they announce a new AI product, will tell 281 00:14:15,360 --> 00:14:17,400 Speaker 5: the press and the public, here are the ways in 282 00:14:17,400 --> 00:14:20,560 Speaker 5: which we know it does not work. They're trying very 283 00:14:20,560 --> 00:14:23,120 Speaker 5: hard to make it clear that they know that there 284 00:14:23,160 --> 00:14:26,280 Speaker 5: are limitations and that they are working on many of them. 285 00:14:26,720 --> 00:14:29,440 Speaker 5: Open ai does the same thing. I mean, when OpenAI 286 00:14:29,600 --> 00:14:34,000 Speaker 5: released Dolli, its image generation tool, they did some work 287 00:14:34,040 --> 00:14:36,280 Speaker 5: to make sure that the images that are generated by 288 00:14:36,320 --> 00:14:38,320 Speaker 5: Dolli are you know, as they sort of explained it, 289 00:14:38,360 --> 00:14:40,760 Speaker 5: more representative of the world. So when you ask DOLLI 290 00:14:40,880 --> 00:14:44,200 Speaker 5: to generate a picture of doctors. They went in and 291 00:14:44,240 --> 00:14:47,600 Speaker 5: made sure manually that the doctors were both men and 292 00:14:47,640 --> 00:14:50,440 Speaker 5: women of different races. And so there are things that 293 00:14:50,600 --> 00:14:54,120 Speaker 5: companies are trying to do to address these issues. I 294 00:14:54,120 --> 00:14:57,280 Speaker 5: think there's a continued push and pull between the company's 295 00:14:57,320 --> 00:15:00,480 Speaker 5: own ethical AI people and people externally. The are looking 296 00:15:00,480 --> 00:15:03,960 Speaker 5: at their work about whether what's enough, whether it's enough, 297 00:15:04,040 --> 00:15:07,600 Speaker 5: whether things are moving too quickly to possibly ensure safety. 298 00:15:10,200 --> 00:15:11,960 Speaker 1: Just a quick quote from the letter that kind of 299 00:15:11,960 --> 00:15:14,280 Speaker 1: sums up to me the kind of the heart of it. 300 00:15:14,280 --> 00:15:18,600 Speaker 1: It says quote recent months have seen AI labs locked 301 00:15:18,640 --> 00:15:21,920 Speaker 1: in an out of control race to develop and deploy 302 00:15:22,080 --> 00:15:26,400 Speaker 1: ever more powerful digital minds that no one, not even 303 00:15:26,440 --> 00:15:30,840 Speaker 1: their creators can understand, predict or reliably control. 304 00:15:31,920 --> 00:15:34,040 Speaker 7: I don't one hundred percent buy that the fact that 305 00:15:34,040 --> 00:15:37,040 Speaker 7: they use the word mind in there is to me 306 00:15:37,520 --> 00:15:40,120 Speaker 7: it's self a little bit suspicious. I think that it's 307 00:15:40,160 --> 00:15:42,960 Speaker 7: extremely out of date at this point to act as 308 00:15:43,000 --> 00:15:46,000 Speaker 7: though the current AI systems are black boxes that we 309 00:15:46,080 --> 00:15:50,080 Speaker 7: can't understand and we can't interrogate. We very much can, 310 00:15:50,280 --> 00:15:53,080 Speaker 7: And I think that anybody who says otherwise is either 311 00:15:53,200 --> 00:15:57,160 Speaker 7: hopelessly naive or as kidding themselves. People who are making 312 00:15:57,160 --> 00:15:59,640 Speaker 7: these systems understand that there may be aspects of them 313 00:15:59,640 --> 00:16:02,400 Speaker 7: that they don't quite get, but in general I think 314 00:16:02,440 --> 00:16:05,600 Speaker 7: they very much do understand them. Whether they're putting in 315 00:16:05,600 --> 00:16:08,520 Speaker 7: proper safeguards before releasing them is a whole other thing. 316 00:16:09,080 --> 00:16:14,359 Speaker 7: But we're nowhere near a point of like complete not understanding, 317 00:16:14,760 --> 00:16:17,720 Speaker 7: or AI at a point where it's like out of control. 318 00:16:18,120 --> 00:16:22,320 Speaker 7: This very much boils down to applications of mas Tina. 319 00:16:22,400 --> 00:16:25,680 Speaker 1: Let's talk about the recent congressional testimony of open Ai 320 00:16:25,800 --> 00:16:29,440 Speaker 1: CEO Sam Altman. What prompted him to come before Congress 321 00:16:29,480 --> 00:16:32,240 Speaker 1: at all and share a little bit about the concerns 322 00:16:32,280 --> 00:16:35,080 Speaker 1: that he raised with a rather unusual call that the 323 00:16:35,480 --> 00:16:38,360 Speaker 1: Congress should actually regulate his own company, not something to 324 00:16:38,400 --> 00:16:39,880 Speaker 1: hear every day in the halls of Congress. 325 00:16:40,560 --> 00:16:42,160 Speaker 4: I think he was summoned. I think that was what 326 00:16:42,240 --> 00:16:44,080 Speaker 4: brought him. I want to argue a little bit with 327 00:16:44,120 --> 00:16:44,600 Speaker 4: the notion. 328 00:16:44,760 --> 00:16:47,760 Speaker 5: I know that Congress doesn't frequently have captains of industry 329 00:16:47,760 --> 00:16:50,000 Speaker 5: come in and say regulate me, but in the AI 330 00:16:50,080 --> 00:16:52,280 Speaker 5: space that's actually been going on for a few years. 331 00:16:52,560 --> 00:16:56,160 Speaker 5: Open Ai, Microsoft IBM, who were also on that panel, 332 00:16:56,160 --> 00:16:59,960 Speaker 5: and even Amazon somewhat surprisingly, have been asking for regularly 333 00:17:00,560 --> 00:17:03,640 Speaker 5: on parts or all of the AI field for several 334 00:17:03,760 --> 00:17:06,800 Speaker 5: years now. Now we can discuss why they want that, 335 00:17:07,040 --> 00:17:09,960 Speaker 5: to what extent they really want to be regulated, but 336 00:17:10,160 --> 00:17:12,520 Speaker 5: the fact is they have been asking for it, and 337 00:17:12,720 --> 00:17:15,200 Speaker 5: some of it is because some of the larger companies 338 00:17:15,600 --> 00:17:19,640 Speaker 5: really want some guidelines and a level playing field so 339 00:17:19,680 --> 00:17:21,720 Speaker 5: that they know what they can do and they can't do. 340 00:17:21,880 --> 00:17:22,800 Speaker 4: And other companies. 341 00:17:22,840 --> 00:17:24,720 Speaker 5: If Microsoft is going to be in their minds a 342 00:17:24,720 --> 00:17:27,520 Speaker 5: good corporate citizen and not do certain things that they 343 00:17:27,560 --> 00:17:30,439 Speaker 5: aren't undermined by other companies that are willing to do that. 344 00:17:30,840 --> 00:17:33,399 Speaker 5: And in fact, the US is behind here. Europe has 345 00:17:33,400 --> 00:17:36,399 Speaker 5: still not passed anything. It's wending its way through the 346 00:17:36,400 --> 00:17:39,040 Speaker 5: European Parliament right now, but Europe's been working on an 347 00:17:39,080 --> 00:17:42,520 Speaker 5: AI law for a couple of years now, and everybody 348 00:17:42,520 --> 00:17:44,520 Speaker 5: has a kind of different version of it. But what 349 00:17:44,600 --> 00:17:47,159 Speaker 5: Sam Altman was talking about is that Opening Eye feels 350 00:17:47,160 --> 00:17:49,480 Speaker 5: that there needs to be a separate US agency, that 351 00:17:49,520 --> 00:17:52,240 Speaker 5: the current agencies in government are not fit for this, 352 00:17:52,840 --> 00:17:55,960 Speaker 5: and that there should be through that agency or otherwise 353 00:17:55,960 --> 00:18:01,800 Speaker 5: some sort of licensing for these kinds of algorithmstra. 354 00:18:00,560 --> 00:18:03,160 Speaker 2: Here's your shot, Thank you, Senator. Number one, I would 355 00:18:03,160 --> 00:18:05,800 Speaker 2: form a new agency that licenses any effort above a 356 00:18:05,800 --> 00:18:08,920 Speaker 2: certain scale of capabilities and can take that license away 357 00:18:08,920 --> 00:18:11,760 Speaker 2: and ensure compliance with safety standards. Number two, I would 358 00:18:11,800 --> 00:18:14,359 Speaker 2: create a set of safety standards focused on what you 359 00:18:14,400 --> 00:18:18,000 Speaker 2: said in your third hypothesis as the dangerous capability evaluations. 360 00:18:18,359 --> 00:18:21,919 Speaker 2: And then third, I would require independent audits, so not 361 00:18:21,960 --> 00:18:23,879 Speaker 2: just from the company or the agency, but experts who 362 00:18:23,920 --> 00:18:26,280 Speaker 2: can say the model is or isn't in compliance with 363 00:18:26,320 --> 00:18:29,160 Speaker 2: these state and safety thresholds and these percentages of performance 364 00:18:29,160 --> 00:18:29,840 Speaker 2: on question. 365 00:18:29,760 --> 00:18:30,159 Speaker 6: X or Y. 366 00:18:30,760 --> 00:18:34,600 Speaker 4: Would you be quantified if we promulgated those rules to 367 00:18:34,680 --> 00:18:35,840 Speaker 4: administer those rules. 368 00:18:36,000 --> 00:18:36,520 Speaker 5: I love my. 369 00:18:36,480 --> 00:18:42,800 Speaker 1: Current job, Rachel. First of all, is any of this realistic? Like, 370 00:18:43,000 --> 00:18:46,560 Speaker 1: how much of this can truly be regulated? The agency 371 00:18:46,600 --> 00:18:49,640 Speaker 1: doesn't exist, Congress moves very slowly, so what is even 372 00:18:49,680 --> 00:18:52,840 Speaker 1: realistic about regulation? And then even if the US was 373 00:18:52,880 --> 00:18:54,680 Speaker 1: able to come up with a regulatory scheme, are there 374 00:18:54,720 --> 00:18:56,880 Speaker 1: fears of any unintended consequences? 375 00:18:57,400 --> 00:18:59,080 Speaker 7: I guess there's a few different ways we can look 376 00:18:59,119 --> 00:19:04,399 Speaker 7: at it. One is there's some existing rules, laws, different 377 00:19:04,400 --> 00:19:08,720 Speaker 7: agencies that could probably take on a lot of different aspects. 378 00:19:08,800 --> 00:19:14,200 Speaker 7: Of regulating AI. As of now, there's this idea of 379 00:19:14,280 --> 00:19:16,800 Speaker 7: creating a new agency. I think it sort of depends 380 00:19:16,840 --> 00:19:18,440 Speaker 7: on who you talk to whether or not they think 381 00:19:18,480 --> 00:19:21,600 Speaker 7: that's necessary. Some people would say that could make things 382 00:19:21,880 --> 00:19:23,879 Speaker 7: even harder and even worse, because it's like, oh, a 383 00:19:23,920 --> 00:19:26,119 Speaker 7: new agency, like you know, we already have quite a 384 00:19:26,200 --> 00:19:28,240 Speaker 7: number of those in the US, and that could possibly 385 00:19:28,240 --> 00:19:31,040 Speaker 7: bog things down even more. That's a little tricky to 386 00:19:31,119 --> 00:19:33,280 Speaker 7: start a whole new thing. I mean, I think what's 387 00:19:33,320 --> 00:19:36,080 Speaker 7: important to keep in mind is that up to now, 388 00:19:36,200 --> 00:19:40,680 Speaker 7: there hasn't been any like AI specific legislation in the US. 389 00:19:40,760 --> 00:19:45,640 Speaker 7: As as Dina pointed out, there's some like application specific stuff, 390 00:19:46,200 --> 00:19:49,199 Speaker 7: but it's more at the local level. But other than that, 391 00:19:49,240 --> 00:19:52,520 Speaker 7: there isn't much so far in the US, and it's 392 00:19:52,560 --> 00:19:54,919 Speaker 7: going to be interesting to see, especially because we have 393 00:19:54,960 --> 00:19:58,480 Speaker 7: a few like AI related court cases right now sort 394 00:19:58,480 --> 00:20:01,200 Speaker 7: of winding their way through various courts. Over the next 395 00:20:01,240 --> 00:20:04,160 Speaker 7: few years, I wouldn't be surprised if things change quite 396 00:20:04,160 --> 00:20:06,880 Speaker 7: a bit as far as having some kind of either 397 00:20:06,920 --> 00:20:09,720 Speaker 7: federal legislation or more rules at the state level. 398 00:20:10,200 --> 00:20:12,280 Speaker 1: I think it's also fair to ask the level of 399 00:20:12,320 --> 00:20:15,359 Speaker 1: sincerity of these calls. Sam Altman's own company released an 400 00:20:15,400 --> 00:20:18,000 Speaker 1: iPhone and Android app about three days after his testimony 401 00:20:18,160 --> 00:20:20,159 Speaker 1: to make chat GPT available to all of us on 402 00:20:20,200 --> 00:20:22,720 Speaker 1: all of our phones. But at the same time he's 403 00:20:22,760 --> 00:20:24,479 Speaker 1: saying we need to be careful how we use this, 404 00:20:24,560 --> 00:20:27,560 Speaker 1: he's sort of making it even more available. Rachel, How 405 00:20:27,640 --> 00:20:30,800 Speaker 1: since here is the industry when it says please regulate. 406 00:20:30,440 --> 00:20:33,679 Speaker 7: Us, I mean, I think part of it is probably 407 00:20:34,359 --> 00:20:39,520 Speaker 7: a actual desire for regulation to know what the rules are, 408 00:20:39,560 --> 00:20:40,879 Speaker 7: what can you do, what you can you not do? 409 00:20:41,160 --> 00:20:43,800 Speaker 7: And some of it I think is definitely aimed at 410 00:20:44,200 --> 00:20:46,840 Speaker 7: keeping control for the parties that are already in control. 411 00:20:46,920 --> 00:20:50,520 Speaker 7: You know, companies like open ai and Google and Microsoft, 412 00:20:51,320 --> 00:20:53,720 Speaker 7: they like their positions in the industry and I can't 413 00:20:53,720 --> 00:20:58,280 Speaker 7: imagine that they want to fall back. So having legislation 414 00:20:58,760 --> 00:21:02,720 Speaker 7: could be helpful to them in certain ways. And I 415 00:21:02,720 --> 00:21:04,760 Speaker 7: feel like keeping some control might be some of it. 416 00:21:05,440 --> 00:21:08,360 Speaker 5: For all these companies that want a licensing scheme, they 417 00:21:08,400 --> 00:21:11,680 Speaker 5: don't have to wait for Congress to act to come 418 00:21:11,760 --> 00:21:14,679 Speaker 5: up with some sort of mechanisms for outside auditing that 419 00:21:14,880 --> 00:21:18,439 Speaker 5: ensures people that they are closed source algorithms are fair 420 00:21:19,119 --> 00:21:19,920 Speaker 5: and safe. 421 00:21:20,440 --> 00:21:23,400 Speaker 1: When we return, what do Rachel and Dina see coming 422 00:21:23,440 --> 00:21:26,880 Speaker 1: down the line when it comes to regulating artificial intelligence. 423 00:21:35,760 --> 00:21:38,959 Speaker 1: As all of our listeners think about these topics, and 424 00:21:39,000 --> 00:21:40,399 Speaker 1: I think they are on a lot of our minds. 425 00:21:40,440 --> 00:21:42,600 Speaker 1: What are you watching for in the next weeks and 426 00:21:42,680 --> 00:21:46,240 Speaker 1: months as this story continues to unfold, whether it's on 427 00:21:46,280 --> 00:21:48,880 Speaker 1: the regulatory scheme, the new technologies that might come out, 428 00:21:48,920 --> 00:21:51,120 Speaker 1: the new uses we might learn about. Dina, why don't 429 00:21:51,119 --> 00:21:53,720 Speaker 1: you go first tell us what you're watching for right now? 430 00:21:54,800 --> 00:21:56,320 Speaker 5: My sense is that the first thing we're going to 431 00:21:56,320 --> 00:22:00,119 Speaker 5: see from a regulatory standpoint is this European law. One 432 00:22:00,160 --> 00:22:02,080 Speaker 5: of the things that Sam Outman suggested is that if 433 00:22:02,080 --> 00:22:04,320 Speaker 5: we do something in the US, it shouldn't impact every 434 00:22:04,320 --> 00:22:06,800 Speaker 5: single kind of AI. In the same way he wanted 435 00:22:06,840 --> 00:22:10,120 Speaker 5: some carbouts for open source so as not to stifle innovation, 436 00:22:10,200 --> 00:22:12,280 Speaker 5: he wanted some carbouts for startups. 437 00:22:12,359 --> 00:22:15,560 Speaker 4: Things like that. The European proposal that's being. 438 00:22:15,359 --> 00:22:19,000 Speaker 5: Looked at has several different tiers of the type of 439 00:22:19,040 --> 00:22:23,240 Speaker 5: algorithm and the sort of related level of scrutiny that 440 00:22:23,320 --> 00:22:26,919 Speaker 5: it gets. For algorithms that do things that the European 441 00:22:27,000 --> 00:22:30,520 Speaker 5: Union considers completely unacceptable, those would just flat out the 442 00:22:30,560 --> 00:22:33,480 Speaker 5: outlawed in the block and so that would definitely be new, 443 00:22:33,520 --> 00:22:37,000 Speaker 5: and then below the absolutely not. There's three other tiers 444 00:22:37,080 --> 00:22:40,240 Speaker 5: that have different levels of scrutiny. I'm interested to see 445 00:22:40,280 --> 00:22:43,400 Speaker 5: if that passes and in what form and what action 446 00:22:43,560 --> 00:22:47,240 Speaker 5: the US companies take as a result, because US companies 447 00:22:47,280 --> 00:22:50,440 Speaker 5: can't just continue to do something here that they cannot 448 00:22:50,440 --> 00:22:52,560 Speaker 5: do in Europe. This will completely impact the way that 449 00:22:52,560 --> 00:22:54,560 Speaker 5: they do business around artificial intelligence. 450 00:22:55,160 --> 00:22:56,240 Speaker 1: Rachel, what are you watching for? 451 00:22:56,920 --> 00:22:59,399 Speaker 7: I mean, I think we're gonna see just more and 452 00:22:59,400 --> 00:23:03,320 Speaker 7: more experiments and people actually using these systems like chat, 453 00:23:03,359 --> 00:23:07,439 Speaker 7: GPT in practice, and as we've seen with some of 454 00:23:07,480 --> 00:23:09,720 Speaker 7: the examples, like the stuff's getting better and better, it's 455 00:23:09,720 --> 00:23:12,840 Speaker 7: getting harder and harder to determine what is a genuine 456 00:23:13,040 --> 00:23:16,639 Speaker 7: human made article from what is created using AI. So 457 00:23:16,720 --> 00:23:18,919 Speaker 7: I think we are going to see more sort of 458 00:23:19,280 --> 00:23:23,720 Speaker 7: disinfo misinfo stuff related to these systems going forward. But 459 00:23:23,840 --> 00:23:27,080 Speaker 7: I'm also optimistic that the detection is going to get 460 00:23:27,119 --> 00:23:31,240 Speaker 7: better and we may be seeing companies increasingly using various 461 00:23:31,240 --> 00:23:35,280 Speaker 7: watermarking technologies to stamp things. Essentially, there are ways to 462 00:23:35,280 --> 00:23:38,439 Speaker 7: stamp both text and images so that they can be 463 00:23:38,560 --> 00:23:42,840 Speaker 7: detected and be seen as created with the help of AI, 464 00:23:43,440 --> 00:23:46,760 Speaker 7: and then I think we'll probably see more applications of 465 00:23:46,800 --> 00:23:50,840 Speaker 7: things like generative video. That's something that's really in its 466 00:23:50,880 --> 00:23:54,000 Speaker 7: infancy right now and has been getting better rapidly, so 467 00:23:54,040 --> 00:23:55,920 Speaker 7: it's going to be really interesting to keep an eye 468 00:23:55,920 --> 00:23:58,399 Speaker 7: on that. There's all kinds of interesting things that are 469 00:23:58,440 --> 00:24:00,640 Speaker 7: happening in the open source community right now now, such 470 00:24:00,640 --> 00:24:05,080 Speaker 7: as the idea of automating these systems, so that, like 471 00:24:05,119 --> 00:24:08,040 Speaker 7: AUTOGBT is one thing that people are paying attention a 472 00:24:08,080 --> 00:24:10,800 Speaker 7: lot too lately, So sort of setting an end goal 473 00:24:11,160 --> 00:24:14,520 Speaker 7: for a large language model and having it sort of 474 00:24:14,560 --> 00:24:18,520 Speaker 7: go off and create additional tasks toward that goal to 475 00:24:18,560 --> 00:24:20,960 Speaker 7: eventually reach the goal, maybe connecting it to other services. 476 00:24:21,240 --> 00:24:22,080 Speaker 4: So, yeah, there's a lot. 477 00:24:21,960 --> 00:24:24,440 Speaker 7: Of interesting stuff there, some potentially scary stuff, but also 478 00:24:24,520 --> 00:24:26,159 Speaker 7: some potentially really cool stuff. 479 00:24:27,480 --> 00:24:28,760 Speaker 1: I'll throw it out there to either of you. The 480 00:24:28,800 --> 00:24:31,679 Speaker 1: Supreme Court did recently avoid ruling on section two thirty, 481 00:24:31,800 --> 00:24:34,159 Speaker 1: the law that protects, you know, tech companies being liable 482 00:24:34,160 --> 00:24:36,680 Speaker 1: for every single thing that's on their platforms. Does that 483 00:24:36,720 --> 00:24:38,879 Speaker 1: tell us that at least the highest court of the 484 00:24:38,960 --> 00:24:41,880 Speaker 1: land doesn't really want to wade into being the big 485 00:24:41,920 --> 00:24:45,520 Speaker 1: referee on tech issues and or how does that ruling 486 00:24:45,600 --> 00:24:48,240 Speaker 1: or non ruling actually affect the conversation around AI. 487 00:24:48,840 --> 00:24:49,400 Speaker 4: So two things. 488 00:24:49,480 --> 00:24:51,600 Speaker 5: One, we're still not sure that section two thirty would 489 00:24:51,680 --> 00:24:53,400 Speaker 5: or should apply to AI. That came up a bit 490 00:24:53,440 --> 00:24:56,760 Speaker 5: in the recent Senate hearing. I know Altman said he 491 00:24:56,800 --> 00:24:59,119 Speaker 5: didn't think two thirty was the right way to regulate AI. 492 00:24:59,520 --> 00:25:02,600 Speaker 5: But you know, interestingly enough that day that really came out. 493 00:25:02,840 --> 00:25:05,800 Speaker 5: Rachel and I were actually discussing the AI implications of 494 00:25:05,840 --> 00:25:08,600 Speaker 5: the other ruling that came out that morning, which also 495 00:25:08,680 --> 00:25:11,439 Speaker 5: has some AI implications, maybe even clearer ones, which was 496 00:25:11,760 --> 00:25:15,080 Speaker 5: a case where a photographer had sued the Andy Warhol estate. 497 00:25:15,600 --> 00:25:17,879 Speaker 5: Andy Warhol had taken a photo that this photographer had 498 00:25:17,920 --> 00:25:20,359 Speaker 5: taken of prints and turned it into one of his 499 00:25:20,680 --> 00:25:23,480 Speaker 5: you know, usual pieces of art silkscreen, et cetera, and 500 00:25:23,520 --> 00:25:26,600 Speaker 5: the photographer sued over it. And the question was was 501 00:25:26,680 --> 00:25:31,480 Speaker 5: the Warhol work transformative enough of the photographer's original work 502 00:25:31,520 --> 00:25:33,800 Speaker 5: that accounted as a new work. And the Supreme Court 503 00:25:33,840 --> 00:25:36,320 Speaker 5: ruled in favor of the photographer's claim. And that has 504 00:25:36,359 --> 00:25:39,960 Speaker 5: implications for AI because there are a number of lawsuits 505 00:25:40,160 --> 00:25:44,920 Speaker 5: right now from artists and computer programmers whose works either 506 00:25:45,000 --> 00:25:47,920 Speaker 5: their software code or their works of art, their photography 507 00:25:48,119 --> 00:25:50,280 Speaker 5: has been used in the training data of some of 508 00:25:50,280 --> 00:25:53,480 Speaker 5: these generative AI systems that we're discussing, and in some 509 00:25:53,640 --> 00:25:57,480 Speaker 5: cases the artist or the computer programmer claims that not 510 00:25:57,520 --> 00:25:59,520 Speaker 5: only was their work used in the training data, but 511 00:25:59,560 --> 00:26:02,560 Speaker 5: the out put, the thing that the AI algorithm generated 512 00:26:02,600 --> 00:26:05,399 Speaker 5: looked suspiciously like their original product, and so there are 513 00:26:05,440 --> 00:26:09,359 Speaker 5: suits over whether that's allowed now. And the companies that 514 00:26:09,520 --> 00:26:12,640 Speaker 5: create these AI models are basically making a fair use 515 00:26:12,720 --> 00:26:14,680 Speaker 5: argument that they are allowed to use these things in 516 00:26:14,720 --> 00:26:17,040 Speaker 5: the training data, that they're transforming it into something else 517 00:26:17,080 --> 00:26:19,920 Speaker 5: that it doesn't resemble the original work. So you know, 518 00:26:20,040 --> 00:26:22,080 Speaker 5: Rachel and I were actually spent a lot more of 519 00:26:22,160 --> 00:26:26,040 Speaker 5: that day discussing that Supreme Court cases implications for AI models, 520 00:26:26,119 --> 00:26:27,440 Speaker 5: rather than the section two thirty one. 521 00:26:28,240 --> 00:26:30,200 Speaker 1: Well, I want to say thank you to our two 522 00:26:30,280 --> 00:26:32,879 Speaker 1: guests today, Dina bas who covers Microsoft and AI for 523 00:26:33,000 --> 00:26:36,240 Speaker 1: Bloomberg News, and of course Rachel Metz, who covers AI 524 00:26:36,400 --> 00:26:37,080 Speaker 1: for US as well. 525 00:26:37,119 --> 00:26:39,040 Speaker 4: Thank you so much for joining me, Thank you for 526 00:26:39,080 --> 00:26:42,000 Speaker 4: having us, thank you thanks for. 527 00:26:42,000 --> 00:26:44,440 Speaker 1: Listening to us here at the Big Take, a daily 528 00:26:44,520 --> 00:26:48,920 Speaker 1: podcast from Bloomberg and iHeartRadio for more shows from iHeartRadio, 529 00:26:49,160 --> 00:26:53,040 Speaker 1: visit the iHeartRadio app, Apple Podcasts, or wherever you listen, 530 00:26:53,520 --> 00:26:56,120 Speaker 1: and we'd love to hear from you. Email us questions 531 00:26:56,200 --> 00:27:00,480 Speaker 1: or comments to Big Take at Bloomberg dot net. Revising 532 00:27:00,520 --> 00:27:05,200 Speaker 1: producer is Vicky Vergalina. Our senior producer is Katherine Fink. 533 00:27:05,880 --> 00:27:10,879 Speaker 1: Our producer is Rebecca Chasson. Our associate producer is Sam Gibbauer. 534 00:27:11,760 --> 00:27:16,320 Speaker 1: Raphael M. Seely is our engineer. Original music by Leo Sidron. 535 00:27:16,800 --> 00:27:19,560 Speaker 1: I'm Craig Gordon sitting in for West Kasova. Have a 536 00:27:19,600 --> 00:27:20,240 Speaker 1: great weekend.