1 00:00:11,800 --> 00:00:14,680 Speaker 1: Natasha Taylor. There was a fascinating story and Wired last 2 00:00:14,680 --> 00:00:19,120 Speaker 1: week about an internal memo Meta regarding a loose squirrel. 3 00:00:19,320 --> 00:00:22,400 Speaker 2: I did see this, Yeah, I didn't understand. 4 00:00:22,600 --> 00:00:24,119 Speaker 3: We had a bird get into the New York Times 5 00:00:24,160 --> 00:00:26,360 Speaker 3: office once and it did derail our afternoon. 6 00:00:26,480 --> 00:00:28,360 Speaker 1: So this was this was a hard hard to pass. 7 00:00:28,400 --> 00:00:32,720 Speaker 1: But there was a squirrel mails to Meta in the Bangkok. 8 00:00:32,440 --> 00:00:33,959 Speaker 3: Office, the poor squirrel. 9 00:00:34,520 --> 00:00:37,360 Speaker 1: There was great hilarity because it's been a rough time 10 00:00:37,400 --> 00:00:40,239 Speaker 1: for Meta, but people were commenting on how this was 11 00:00:40,280 --> 00:00:42,800 Speaker 1: the most joyful moment the company had for a while, 12 00:00:42,840 --> 00:00:45,080 Speaker 1: despite the fact that Janet had got scratched on his finger. 13 00:00:45,640 --> 00:00:47,000 Speaker 3: Well, I hope he got a Rabi's shot. 14 00:00:50,159 --> 00:00:52,600 Speaker 1: Welcome to Tech Stuff. I'm was Vloscha and this is 15 00:00:52,640 --> 00:00:54,480 Speaker 1: the Week in Tech where I'm joined by the world's 16 00:00:54,520 --> 00:00:57,600 Speaker 1: most plugged in reporters to break down what's really happening 17 00:00:57,600 --> 00:01:00,200 Speaker 1: in tech right now. Today we're joined by Taylor, the 18 00:01:00,240 --> 00:01:03,040 Speaker 1: rends if use a mag and Natasha Tiku, tech reporter 19 00:01:03,160 --> 00:01:04,160 Speaker 1: at the Washington Post. 20 00:01:04,240 --> 00:01:07,679 Speaker 3: Welcome both, Hey everybody, Hello, thanks for having us. 21 00:01:08,040 --> 00:01:08,840 Speaker 1: It's good to see you. 22 00:01:09,400 --> 00:01:09,800 Speaker 3: Natasha. 23 00:01:09,880 --> 00:01:13,800 Speaker 1: You recently published an article titled Inside the Secret Ai 24 00:01:13,840 --> 00:01:18,280 Speaker 1: War between Silicon Valley and China. Great headline, but it 25 00:01:18,319 --> 00:01:21,880 Speaker 1: wasn't about ships or competing for talent, but actually about 26 00:01:21,880 --> 00:01:25,360 Speaker 1: the models themselves. Can you explain what's going on here? 27 00:01:25,680 --> 00:01:30,000 Speaker 4: So this is something of a narrative war between the 28 00:01:30,160 --> 00:01:34,560 Speaker 4: US leading AI companies and the leading Chinese AI companies. 29 00:01:35,160 --> 00:01:39,160 Speaker 4: So what companies like Anthropic and Open Ai are alleging 30 00:01:39,480 --> 00:01:44,400 Speaker 4: is that these Chinese companies are accessing their models which 31 00:01:44,440 --> 00:01:47,800 Speaker 4: are not supposed to be available in China and Hong Kong. 32 00:01:48,160 --> 00:01:51,680 Speaker 4: They're accessing their models in order to do this process 33 00:01:51,680 --> 00:01:55,840 Speaker 4: called distillation, which basically involves kind of you know, distilling, 34 00:01:55,960 --> 00:01:58,680 Speaker 4: taking the learnings from a larger model and applying it 35 00:01:58,720 --> 00:02:02,320 Speaker 4: to us to improve a smaller model. And distillation is 36 00:02:02,400 --> 00:02:05,120 Speaker 4: a process that has been used in the AI industry 37 00:02:05,200 --> 00:02:08,520 Speaker 4: for a really long time. It's standard process. It's how 38 00:02:08,880 --> 00:02:12,560 Speaker 4: state of the art improvements are diffused throughout the system, 39 00:02:13,000 --> 00:02:17,400 Speaker 4: you know, through academia and open source. But Open Ai 40 00:02:17,760 --> 00:02:22,280 Speaker 4: and Anthropic have started using this term adversarial distillation and 41 00:02:22,960 --> 00:02:28,360 Speaker 4: they are like seeking government US government protection from this process, 42 00:02:28,480 --> 00:02:30,520 Speaker 4: and they're alleging, you know that a lot of the 43 00:02:30,560 --> 00:02:34,560 Speaker 4: gains that we're seeing in Chinese models in open source 44 00:02:34,560 --> 00:02:37,919 Speaker 4: AI in particular is coming at the you know, at 45 00:02:37,919 --> 00:02:40,160 Speaker 4: the expense of anthropic. 46 00:02:39,960 --> 00:02:43,079 Speaker 1: Yeah, I have a serial distillation. It sounds like makes 47 00:02:43,080 --> 00:02:44,720 Speaker 1: me think of prohibition actually. 48 00:02:44,440 --> 00:02:44,520 Speaker 4: But. 49 00:02:46,440 --> 00:02:47,040 Speaker 2: Yeah, I was. 50 00:02:47,120 --> 00:02:50,480 Speaker 4: I was at this National Security AI conference and the 51 00:02:50,560 --> 00:02:56,440 Speaker 4: person from open ai brought it up, and Irene's salimon 52 00:02:56,480 --> 00:02:58,760 Speaker 4: from hugging face like Brazer Hins. She was like, I've 53 00:02:58,800 --> 00:03:01,320 Speaker 4: never I've been in this indust for so long. She 54 00:03:01,360 --> 00:03:03,360 Speaker 4: was a super early open ai employee, and she was like, 55 00:03:03,400 --> 00:03:06,200 Speaker 4: I've never heard this term before, but now it's now. 56 00:03:06,200 --> 00:03:09,799 Speaker 4: It's kind of it is diffusing throughout the US because, 57 00:03:10,400 --> 00:03:13,840 Speaker 4: you know, they're extremely concerned that the Chinese open source 58 00:03:13,919 --> 00:03:18,840 Speaker 4: models are cheaper, more accessible, and are being adopted very 59 00:03:18,840 --> 00:03:22,640 Speaker 4: widely by US companies. Not that they the US companies 60 00:03:22,680 --> 00:03:25,840 Speaker 4: necessarily want to talk about it, but you know, particularly 61 00:03:26,200 --> 00:03:31,799 Speaker 4: around the restrictions around fable that happened with US government, 62 00:03:32,040 --> 00:03:35,720 Speaker 4: the inability to access mythos. You know, companies are like, 63 00:03:35,760 --> 00:03:38,600 Speaker 4: it's more reliable and more cost effective for me to 64 00:03:38,640 --> 00:03:42,680 Speaker 4: build on top of Chinese technology. And I think another 65 00:03:43,000 --> 00:03:46,680 Speaker 4: kind of distinction that's really been alided over is that 66 00:03:47,200 --> 00:03:51,000 Speaker 4: distillation might be a violation of your terms of service. 67 00:03:51,280 --> 00:03:54,280 Speaker 4: You know, if you use if you access the Anthropic API, 68 00:03:54,480 --> 00:03:58,680 Speaker 4: you are not supposed to try to ask questions twenty 69 00:03:58,720 --> 00:04:01,080 Speaker 4: seven million times and then and you know, use it 70 00:04:01,120 --> 00:04:02,280 Speaker 4: to improve your own model. 71 00:04:02,440 --> 00:04:05,760 Speaker 1: What's the difference between what's happening right now in your 72 00:04:05,840 --> 00:04:09,840 Speaker 1: story and the deep Seek freak of early last year, Like, 73 00:04:09,920 --> 00:04:11,840 Speaker 1: is this kind of a continuation of the same story 74 00:04:11,920 --> 00:04:13,160 Speaker 1: or is it something slightly different? 75 00:04:13,560 --> 00:04:16,000 Speaker 4: No, I would say it's a continuation of the same story. 76 00:04:16,000 --> 00:04:19,560 Speaker 4: In that case, it was open Ai alleging that, you 77 00:04:19,600 --> 00:04:22,599 Speaker 4: know that Deep Deep Seek did the distillation, but it 78 00:04:22,760 --> 00:04:24,919 Speaker 4: also has to do with the fact that the US 79 00:04:25,040 --> 00:04:28,159 Speaker 4: was really caught off guard by how good the Deep 80 00:04:28,200 --> 00:04:31,960 Speaker 4: Seek model was. And you know, many of the gains 81 00:04:32,000 --> 00:04:36,560 Speaker 4: that you're seeing in the Chinese market are you know, 82 00:04:36,640 --> 00:04:39,720 Speaker 4: like open Ai and Anthropic would like you to believe 83 00:04:39,720 --> 00:04:43,200 Speaker 4: that it's only because they're taking from US models, but 84 00:04:43,240 --> 00:04:46,440 Speaker 4: that's not true. In fact, like many of the you know, 85 00:04:46,440 --> 00:04:48,719 Speaker 4: speaking to two experts in the field, they say that 86 00:04:48,839 --> 00:04:51,800 Speaker 4: like many of the gains that the Chinese models have 87 00:04:51,880 --> 00:04:57,680 Speaker 4: made in efficiency, in size, cost effectiveness, those have all 88 00:04:57,720 --> 00:05:01,120 Speaker 4: come from the constraints on the industry, like the fact 89 00:05:01,120 --> 00:05:04,680 Speaker 4: that we have export controls on chips. You know, they 90 00:05:04,680 --> 00:05:06,760 Speaker 4: don't have access to as many as. 91 00:05:06,640 --> 00:05:09,520 Speaker 1: Resources Jnsui Aguen, right, which is a necessity is the 92 00:05:09,560 --> 00:05:11,440 Speaker 1: mother invention and actually restricts exactly. 93 00:05:11,480 --> 00:05:13,960 Speaker 4: I mean, you know this is like the startup way. 94 00:05:14,080 --> 00:05:16,320 Speaker 4: And obviously you know that if you look at like 95 00:05:16,360 --> 00:05:19,200 Speaker 4: how much money has been raised by the major AI labs, 96 00:05:19,240 --> 00:05:23,560 Speaker 4: that's not been a constraint for them, right, And uh, 97 00:05:23,960 --> 00:05:28,200 Speaker 4: you know what's happened recently is that a one of 98 00:05:28,240 --> 00:05:32,760 Speaker 4: the Anthropic users in Hong Kong discovered that Anthropic had 99 00:05:33,279 --> 00:05:37,359 Speaker 4: was using this kind of covert way to identify Chinese users. 100 00:05:37,760 --> 00:05:40,080 Speaker 4: They were looking at what time zone they're in. They 101 00:05:40,080 --> 00:05:43,080 Speaker 4: were looking at the you know, where they're logging in from. 102 00:05:43,200 --> 00:05:46,680 Speaker 4: Is it associated? Is that domain associated with the Chinese 103 00:05:46,760 --> 00:05:49,800 Speaker 4: AI lab? And then it's sending the information back to 104 00:05:49,880 --> 00:05:54,760 Speaker 4: Anthropic but in a way it like hidden in today's date. 105 00:05:55,520 --> 00:05:58,920 Speaker 4: So like the apostrophe in today's date said like you know, 106 00:06:00,120 --> 00:06:04,640 Speaker 4: Chinese user associated with minimax logging in in this time zone, 107 00:06:05,200 --> 00:06:09,840 Speaker 4: et cetera. And people were really upset about this, like. 108 00:06:09,800 --> 00:06:13,039 Speaker 1: This is sort of a nightmare scenario for AI companies, 109 00:06:13,040 --> 00:06:16,080 Speaker 1: for users to feel like they're being actively surveiled, right, 110 00:06:16,120 --> 00:06:18,359 Speaker 1: I mean, Taylor, what's your I know that we're going 111 00:06:18,400 --> 00:06:20,480 Speaker 1: to talk about surveillance in the second half of the episode, 112 00:06:20,480 --> 00:06:23,080 Speaker 1: but what's your what's your take on the idea that 113 00:06:23,200 --> 00:06:26,440 Speaker 1: you know, Anthropic have been caught essentially spying on users 114 00:06:26,440 --> 00:06:28,160 Speaker 1: to try and figure out where they are and if 115 00:06:28,200 --> 00:06:31,120 Speaker 1: they're Chinese users, illicitly trying to distill the models. 116 00:06:31,360 --> 00:06:34,880 Speaker 3: I think this whole thing is honestly, I don't want 117 00:06:34,920 --> 00:06:36,680 Speaker 3: to say a losing game, but it seems ridiculous and 118 00:06:36,680 --> 00:06:39,719 Speaker 3: I'm kind of curious to Tasha, like why why isn't 119 00:06:39,760 --> 00:06:43,599 Speaker 3: the US developing open source models? Like why why are 120 00:06:43,800 --> 00:06:46,520 Speaker 3: you know, why can't we have our own open source 121 00:06:46,520 --> 00:06:48,760 Speaker 3: that competes with these Chinese open source models. 122 00:06:49,680 --> 00:06:52,960 Speaker 4: Well, now there is a lot more investment going into it, 123 00:06:53,000 --> 00:06:58,279 Speaker 4: but yeah, it wasn't like fine as financially lucrative. And 124 00:06:59,279 --> 00:07:01,719 Speaker 4: you know, even the you hear a lot of rhetoric 125 00:07:01,800 --> 00:07:04,880 Speaker 4: from the VCS, right like we we can't stop AI, 126 00:07:05,120 --> 00:07:07,559 Speaker 4: like we have to support open source. You weren't seeing 127 00:07:07,640 --> 00:07:11,360 Speaker 4: commensurate investment in open source. There are like a lot 128 00:07:11,360 --> 00:07:14,960 Speaker 4: of Alan Institute for AI hugging face. 129 00:07:16,240 --> 00:07:18,600 Speaker 2: There was this project called the Adam Project. 130 00:07:19,240 --> 00:07:23,280 Speaker 4: People were really interested in pushing this and having you know, 131 00:07:23,320 --> 00:07:25,280 Speaker 4: the US be the center of open source. But yeah, 132 00:07:25,280 --> 00:07:28,240 Speaker 4: there wasn't as much investment in it, and the vcs 133 00:07:28,280 --> 00:07:33,920 Speaker 4: are all backing the closed sourced companies, and so yeah, 134 00:07:33,960 --> 00:07:36,040 Speaker 4: this is the situation we now find ourselves in where 135 00:07:36,040 --> 00:07:37,040 Speaker 4: we're playing ketchup. 136 00:07:37,480 --> 00:07:39,960 Speaker 3: Yeah, I feel like we're playing ketchup. And also it's 137 00:07:40,000 --> 00:07:42,560 Speaker 3: a losing battle. Like I mean, I know Alex Karp 138 00:07:42,720 --> 00:07:46,080 Speaker 3: was talking about this recently, and obviously he's self interested. 139 00:07:46,160 --> 00:07:51,360 Speaker 3: I think volunteer uses, maybe the open source models, but 140 00:07:51,360 --> 00:07:55,040 Speaker 3: but like it just seems like this idea of this 141 00:07:55,200 --> 00:07:58,320 Speaker 3: like closed source situation and like tokens and I don't 142 00:07:58,360 --> 00:08:01,320 Speaker 3: I don't know, like everything for profit like that Anthropic 143 00:08:01,400 --> 00:08:04,160 Speaker 3: and OPA are built on is just again it's just 144 00:08:04,200 --> 00:08:08,000 Speaker 3: a losing game because ultimately open source is going to win. 145 00:08:08,840 --> 00:08:11,360 Speaker 1: I guess if Reid was here, if our play read, 146 00:08:11,400 --> 00:08:13,280 Speaker 1: because he's not here this week he's in Boston. But 147 00:08:13,320 --> 00:08:15,640 Speaker 1: I mean, I think what he would say is yes, 148 00:08:15,960 --> 00:08:20,200 Speaker 1: except that if you are leading edge company or startup 149 00:08:20,280 --> 00:08:22,000 Speaker 1: or lab, you need to have access to the best 150 00:08:22,040 --> 00:08:25,720 Speaker 1: models otherwise you will fall behind competitively. So like the 151 00:08:25,760 --> 00:08:28,080 Speaker 1: future general you say, I may be more and more 152 00:08:28,120 --> 00:08:30,800 Speaker 1: open source, but there will always be demand for like 153 00:08:31,240 --> 00:08:33,120 Speaker 1: mythos or whatever it is that you have to pay for. 154 00:08:33,520 --> 00:08:37,600 Speaker 1: Is that I mean? I guess what Red would say, Well. 155 00:08:37,440 --> 00:08:39,560 Speaker 4: You can, I mean, you could build a state of 156 00:08:39,559 --> 00:08:42,360 Speaker 4: the art model on top of open source. You don't 157 00:08:42,400 --> 00:08:46,880 Speaker 4: need to start with Anthropic. You know, Anthropic could be 158 00:08:46,880 --> 00:08:51,800 Speaker 4: building on top of deep seek. I think, like, you know, 159 00:08:51,840 --> 00:08:54,760 Speaker 4: what's really interesting is that the US position, at least 160 00:08:54,840 --> 00:08:57,040 Speaker 4: under David Sachs, was that you want to build on 161 00:08:57,080 --> 00:09:00,160 Speaker 4: top of the American stack, like you want to make 162 00:09:00,240 --> 00:09:03,120 Speaker 4: sure just in the same way that you know, China 163 00:09:03,120 --> 00:09:07,000 Speaker 4: has diffused its power by working with you know, companies 164 00:09:07,000 --> 00:09:09,120 Speaker 4: in the Middle East and Africa. They were saying, you know, 165 00:09:09,200 --> 00:09:12,800 Speaker 4: we want the US, the US stack to be like 166 00:09:12,920 --> 00:09:16,520 Speaker 4: the layer on open source. But then the Trump administration 167 00:09:16,600 --> 00:09:20,360 Speaker 4: policies are really going against that. And actually Anthropic had said, 168 00:09:20,720 --> 00:09:22,920 Speaker 4: you know, they tried to argue that if the US 169 00:09:23,000 --> 00:09:27,880 Speaker 4: enforces you know, a stronger stance against distillation, that maybe 170 00:09:27,920 --> 00:09:30,840 Speaker 4: the US labs could maintain their lead for twelve to 171 00:09:30,880 --> 00:09:36,760 Speaker 4: twenty four months longer. And yeah, it's just it's uh 172 00:09:37,320 --> 00:09:40,040 Speaker 4: has led to exactly like what you're you, Taylor. You 173 00:09:40,080 --> 00:09:44,800 Speaker 4: mentioned Alex Carp's rant recently, and it's completely related because 174 00:09:45,120 --> 00:09:48,800 Speaker 4: companies are extremely frustrated. They feel like, you know, do 175 00:09:48,840 --> 00:09:52,720 Speaker 4: they even own the intellectual property that comes from using 176 00:09:52,760 --> 00:09:55,600 Speaker 4: your model? Why are they paying for tokens? Like it's 177 00:09:55,640 --> 00:09:59,680 Speaker 4: really had people even kind of rethinking the potential business 178 00:09:59,760 --> 00:10:03,320 Speaker 4: model and what the value is for the companies if 179 00:10:03,400 --> 00:10:06,640 Speaker 4: like they don't own the technology there. Yeah, it's just 180 00:10:07,160 --> 00:10:09,800 Speaker 4: I think they feel like it's highly inefficient and not 181 00:10:09,880 --> 00:10:14,320 Speaker 4: a good way to to like start the new AI 182 00:10:14,440 --> 00:10:17,439 Speaker 4: era building on this technology that could be turned off 183 00:10:17,440 --> 00:10:20,160 Speaker 4: by the US government or anthropic could change something, or 184 00:10:20,200 --> 00:10:21,160 Speaker 4: they're spying on you. 185 00:10:21,520 --> 00:10:24,240 Speaker 3: Yeah, it seems like ridiculous. It seems like obviously we 186 00:10:24,280 --> 00:10:27,280 Speaker 3: should be moving towards an open source model. Obviously China 187 00:10:27,360 --> 00:10:30,199 Speaker 3: has the superior sort of structure to the AI environment. Also, 188 00:10:30,240 --> 00:10:33,200 Speaker 3: it's better for consumers, it's more affordable, it's better for 189 00:10:33,240 --> 00:10:36,199 Speaker 3: business consumers, better for I would say average consumers as well. 190 00:10:37,360 --> 00:10:39,920 Speaker 3: I think, I mean, I don't use deep Seek for 191 00:10:40,040 --> 00:10:42,640 Speaker 3: like tons of stuff, but I find it completely comparable 192 00:10:42,720 --> 00:10:46,240 Speaker 3: to the US you know, companies that are out there, 193 00:10:47,160 --> 00:10:49,480 Speaker 3: and I just I also wonder if if it would 194 00:10:49,559 --> 00:10:51,680 Speaker 3: help if we if we were investing more in open 195 00:10:51,679 --> 00:10:54,240 Speaker 3: source as well, if it would help sort of stem 196 00:10:54,280 --> 00:10:57,760 Speaker 3: some of the like curb some of the Antai backlash, 197 00:10:57,800 --> 00:10:59,800 Speaker 3: because I feel like so much of the Antai backlash 198 00:10:59,880 --> 00:11:02,760 Speaker 3: is not just from companies, like people like Alex Karp 199 00:11:02,800 --> 00:11:06,320 Speaker 3: going on rants, but consumers. I think consumers feel squeezed. 200 00:11:06,440 --> 00:11:09,320 Speaker 3: They feel like these companies are just doing everything to profit, 201 00:11:09,480 --> 00:11:12,760 Speaker 3: the token maxing and minning, the sort of wars. I 202 00:11:13,160 --> 00:11:15,800 Speaker 3: just feel like if we yeah had this different sort 203 00:11:15,800 --> 00:11:18,440 Speaker 3: of open source ecosystem, it would be a radically different 204 00:11:18,679 --> 00:11:20,840 Speaker 3: tech landscape and a better one in my opinion. 205 00:11:21,800 --> 00:11:23,760 Speaker 1: If you're the US government, I mean you're looking at 206 00:11:23,800 --> 00:11:27,040 Speaker 1: you're looking at the spaces and anthropic and open AI 207 00:11:27,120 --> 00:11:29,880 Speaker 1: as your kind of national champions essentially, right, And so 208 00:11:29,920 --> 00:11:33,199 Speaker 1: if you take away the ability by letting too much 209 00:11:33,240 --> 00:11:36,520 Speaker 1: distillation happen of these companies to protect their value. 210 00:11:36,320 --> 00:11:39,240 Speaker 3: Like, but let's be clear, let's be clear, open source 211 00:11:39,520 --> 00:11:44,720 Speaker 3: like models are not only developed by distilling the anthropic 212 00:11:44,880 --> 00:11:47,560 Speaker 3: and open ai like this is what they're alleging, but like, 213 00:11:47,880 --> 00:11:51,560 Speaker 3: these open source models are also progressing at a significant rate. 214 00:11:51,640 --> 00:11:54,080 Speaker 3: This is why they're so obsessed with competing with China, 215 00:11:54,160 --> 00:11:56,400 Speaker 3: Like these open source models are getting better and better, 216 00:11:56,440 --> 00:11:59,280 Speaker 3: and it's not just because they're distilling anthropic and open AI, but. 217 00:11:59,400 --> 00:12:02,800 Speaker 1: In your store or Natasha. Chinese researchers did some research 218 00:12:02,840 --> 00:12:05,280 Speaker 1: on this and it turns out that Quen, which is 219 00:12:05,320 --> 00:12:08,599 Speaker 1: Ali Barbara's AI model, this is, according to Chinese researchers, 220 00:12:09,040 --> 00:12:13,160 Speaker 1: misidentified itself as clawed in nearly one third of all cases. 221 00:12:13,559 --> 00:12:14,520 Speaker 1: I mean, that's pretty stunning. 222 00:12:15,240 --> 00:12:16,600 Speaker 2: Yeah, that's true. 223 00:12:16,640 --> 00:12:20,120 Speaker 4: But I will just say that, you know, I have 224 00:12:20,200 --> 00:12:24,080 Speaker 4: spoken to researchers who say when they test Quad and 225 00:12:24,120 --> 00:12:26,680 Speaker 4: other US models in Chinese and they do the same 226 00:12:26,720 --> 00:12:29,760 Speaker 4: thing and they say, who are you? It's as deep seek. 227 00:12:30,200 --> 00:12:33,120 Speaker 4: So I mean, like, this is what I'm trying to 228 00:12:33,160 --> 00:12:36,400 Speaker 4: say about. You know, it's really challenging when you use 229 00:12:36,520 --> 00:12:40,720 Speaker 4: this terminology that means just like a process, a methodology, 230 00:12:40,760 --> 00:12:46,440 Speaker 4: a technique, and imply that it's necessarily, you know, a 231 00:12:46,520 --> 00:12:49,080 Speaker 4: violation in some way. It might be breaking in terms 232 00:12:49,080 --> 00:12:51,520 Speaker 4: of service, which users do all the time and which 233 00:12:51,600 --> 00:12:54,080 Speaker 4: you know gets you kicked off of the platform, but 234 00:12:54,320 --> 00:12:57,160 Speaker 4: in some cases, like there is something different going on. 235 00:12:57,240 --> 00:13:00,160 Speaker 4: It's more of like jail breaking in a way. But 236 00:13:00,280 --> 00:13:02,640 Speaker 4: when you kind of use this term so broadly, it's 237 00:13:02,679 --> 00:13:06,720 Speaker 4: really hard to tell what exactly is happening and whether 238 00:13:07,760 --> 00:13:10,880 Speaker 4: some like whether it's really an egregious violation or just 239 00:13:10,920 --> 00:13:12,960 Speaker 4: something that the companies don't really want. 240 00:13:13,320 --> 00:13:15,920 Speaker 1: Tell me about this phrase secret war in the in 241 00:13:15,920 --> 00:13:17,880 Speaker 1: the headline of your story, like how is the secret 242 00:13:17,920 --> 00:13:21,880 Speaker 1: war playing out? Like who are the combatants? And what 243 00:13:21,880 --> 00:13:22,800 Speaker 1: are they doing to each other? 244 00:13:23,200 --> 00:13:25,960 Speaker 4: Well, I would say anthropic is one of the primary ones. 245 00:13:26,000 --> 00:13:28,200 Speaker 4: I mean when Taylor is asking about like why isn't 246 00:13:28,200 --> 00:13:31,560 Speaker 4: there a lot of open source. Anthropic was a big 247 00:13:32,360 --> 00:13:36,240 Speaker 4: voice in DC starting in the Biden administration, talking about 248 00:13:36,280 --> 00:13:39,240 Speaker 4: how dangerous open source was because you wouldn't be able 249 00:13:39,320 --> 00:13:43,440 Speaker 4: to check the downstream uses as closely, like if somebody 250 00:13:43,520 --> 00:13:46,320 Speaker 4: is using it to create a bioweapon or what have you, 251 00:13:46,400 --> 00:13:50,400 Speaker 4: because you can just download open source and then kind of, 252 00:13:50,440 --> 00:13:53,200 Speaker 4: you know, fork it, use it for yourself and the 253 00:13:53,240 --> 00:13:56,400 Speaker 4: company the developers of the open source technology won't necessarily 254 00:13:56,480 --> 00:13:59,400 Speaker 4: know what's happened. But I think we've also seen that 255 00:13:59,640 --> 00:14:02,480 Speaker 4: we can don't really rely on at least, you know, 256 00:14:02,679 --> 00:14:07,400 Speaker 4: when it came to how young people were using this technology, 257 00:14:08,160 --> 00:14:10,200 Speaker 4: you know, in terms of like therapy or some of 258 00:14:10,200 --> 00:14:14,000 Speaker 4: the things that we've seen with with the mental health issues, 259 00:14:14,080 --> 00:14:17,120 Speaker 4: we haven't seen like a robust reports from the companies 260 00:14:17,160 --> 00:14:20,840 Speaker 4: about downstream usage or even when it comes to you know, 261 00:14:20,920 --> 00:14:23,160 Speaker 4: how is how is it affecting jobs? It's not like 262 00:14:23,800 --> 00:14:27,320 Speaker 4: it's transparent to the public or to regulators or to 263 00:14:27,440 --> 00:14:31,360 Speaker 4: advocates how these models are being used downstream. I mean, 264 00:14:31,400 --> 00:14:35,280 Speaker 4: they put out these reports, but it's like quite filtered. 265 00:14:35,400 --> 00:14:37,960 Speaker 4: I would say, it's really hard when you dig into 266 00:14:38,000 --> 00:14:40,720 Speaker 4: methodology to really understand how people are using it. So 267 00:14:40,760 --> 00:14:43,840 Speaker 4: only the companies have access to that information. But anthropics 268 00:14:43,920 --> 00:14:47,960 Speaker 4: argument was very persuasive because you know, they're really worried 269 00:14:47,960 --> 00:14:52,720 Speaker 4: about the use of these models for uh, you know, 270 00:14:52,840 --> 00:14:58,040 Speaker 4: cyber warfare in some way, or or bioweapons or you know, 271 00:14:58,400 --> 00:15:02,880 Speaker 4: developing novel chemicals in some way, and those like once use. 272 00:15:03,120 --> 00:15:05,360 Speaker 1: This is what secure Miss Hassabis talked about this week 273 00:15:05,440 --> 00:15:08,320 Speaker 1: right the CEO of DeepMind at Google, when he called 274 00:15:08,320 --> 00:15:12,280 Speaker 1: for a US led AI standards group to regulate frontier models, 275 00:15:12,880 --> 00:15:14,760 Speaker 1: which got quite a lot of buzz. He also talked 276 00:15:14,760 --> 00:15:18,480 Speaker 1: about quote fostering international collaborational key safety issues. I assume 277 00:15:18,480 --> 00:15:19,440 Speaker 1: he was talking about China. 278 00:15:20,160 --> 00:15:23,400 Speaker 4: Yes, that's that's the way I read it as well. Yeah, 279 00:15:23,440 --> 00:15:28,120 Speaker 4: I mean it's kind of becoming this like a uniform 280 00:15:28,240 --> 00:15:31,640 Speaker 4: voice now from because we saw Jack Clark from Aanthropics 281 00:15:31,680 --> 00:15:35,200 Speaker 4: say something similar, Dario open Ai. Think what they're pushing 282 00:15:35,240 --> 00:15:39,200 Speaker 4: for is like potentially a market basedally, like you'll have 283 00:15:39,280 --> 00:15:42,960 Speaker 4: a third party person testing. It won't be through the government, 284 00:15:44,240 --> 00:15:48,040 Speaker 4: it would be a startup or a company that has 285 00:15:48,080 --> 00:15:49,040 Speaker 4: some independence. 286 00:15:49,080 --> 00:15:50,720 Speaker 2: And yeah, I could. 287 00:15:50,480 --> 00:15:53,480 Speaker 4: See more movement on this because the situation with Mythos 288 00:15:53,520 --> 00:15:58,360 Speaker 4: and fable has been so tumultuous and not good for business. 289 00:15:58,600 --> 00:15:59,880 Speaker 1: And do you think that's a good thing, I mean 290 00:16:00,200 --> 00:16:02,840 Speaker 1: for the average person in the world if this like 291 00:16:03,080 --> 00:16:05,520 Speaker 1: safety initiative happens, or do you think the fact that 292 00:16:05,560 --> 00:16:09,600 Speaker 1: it's led by the industry itself automatically creates kind of 293 00:16:09,640 --> 00:16:10,600 Speaker 1: perverse incentives. 294 00:16:11,480 --> 00:16:14,000 Speaker 4: I mean, I would say all tech policy in the 295 00:16:14,120 --> 00:16:17,800 Speaker 4: US is led by industry. They have you know, they 296 00:16:18,840 --> 00:16:21,720 Speaker 4: they have donated to all the nonprofits, they have the 297 00:16:21,720 --> 00:16:27,800 Speaker 4: biggest teams able to brief members of Congress. You know, 298 00:16:28,120 --> 00:16:31,120 Speaker 4: we just kind of naturally default to the tech executives 299 00:16:31,120 --> 00:16:35,080 Speaker 4: to tell us how to develop the technology. And you know, 300 00:16:35,200 --> 00:16:38,440 Speaker 4: lots of areas of government that need this kind of 301 00:16:38,440 --> 00:16:41,840 Speaker 4: technical expertise have been gutted. I would say there's like 302 00:16:41,880 --> 00:16:44,840 Speaker 4: suspicion obviously towards the tech companies, but I would say 303 00:16:44,920 --> 00:16:47,040 Speaker 4: all I mean, in the fifteen years that I've been 304 00:16:47,080 --> 00:16:50,840 Speaker 4: covering tech, it's all been guided by the companies themselves, 305 00:16:51,520 --> 00:16:53,960 Speaker 4: like from the beginning stage to kind of see the ideas, 306 00:16:53,960 --> 00:16:57,120 Speaker 4: and then once the law is written, you know, have 307 00:16:57,200 --> 00:16:59,480 Speaker 4: their lobbyist tweak things line to line. 308 00:16:59,600 --> 00:17:04,120 Speaker 3: I think so terrifying because like Anthropic kind of reminds 309 00:17:04,160 --> 00:17:07,600 Speaker 3: me of Meta in this way where they will they're 310 00:17:07,640 --> 00:17:10,880 Speaker 3: obsessed with safety. They want to frame everything as doing 311 00:17:10,920 --> 00:17:14,840 Speaker 3: for safety. It's for safety for safety, and often what 312 00:17:14,880 --> 00:17:18,760 Speaker 3: those regulations ultimately do is cement their power and consolidate 313 00:17:18,840 --> 00:17:23,359 Speaker 3: power in the most powerful tech companies. I mean, I 314 00:17:23,359 --> 00:17:27,600 Speaker 3: don't know. The way that Entropic and sort of the 315 00:17:27,600 --> 00:17:30,480 Speaker 3: co founders have spoken about freedom of speech and access 316 00:17:30,520 --> 00:17:33,399 Speaker 3: to information is like deeply concerning. These are people that 317 00:17:35,080 --> 00:17:38,040 Speaker 3: you know, I just I don't trust them at all 318 00:17:38,160 --> 00:17:41,320 Speaker 3: to regulate these tech policies in ways that preserve our rights. 319 00:17:41,760 --> 00:17:44,800 Speaker 3: And I think that because they go out and they 320 00:17:44,840 --> 00:17:47,760 Speaker 3: do these pr tours and they say things that people 321 00:17:47,760 --> 00:17:50,800 Speaker 3: want to hear like oh, safety safety, child safety, whatever, whatever. 322 00:17:51,200 --> 00:17:53,960 Speaker 3: People somehow think that they're responsible, and I think that 323 00:17:53,960 --> 00:17:56,320 Speaker 3: they're not responsible. I think there's I don't know. To me, 324 00:17:56,359 --> 00:17:58,520 Speaker 3: it very much feels like this METAP playbook of like 325 00:17:59,680 --> 00:18:01,679 Speaker 3: going out and just saying a bunch of stuff and 326 00:18:01,880 --> 00:18:05,040 Speaker 3: meanwhile you're just using the rules to tighten control over 327 00:18:05,040 --> 00:18:08,840 Speaker 3: our information ecosystem, titan control over the market. It's just 328 00:18:08,920 --> 00:18:10,120 Speaker 3: like regulatory capture. 329 00:18:11,000 --> 00:18:13,600 Speaker 4: I will say, it is a step above like what 330 00:18:14,320 --> 00:18:18,359 Speaker 4: was you know, pitch during the Biden EO and what's 331 00:18:18,359 --> 00:18:21,199 Speaker 4: happening now, which is like voluntary disclosure. So it's the 332 00:18:21,240 --> 00:18:25,000 Speaker 4: companies themselves doing these tests and like if you look 333 00:18:25,119 --> 00:18:27,040 Speaker 4: very closely at the you know, every time a new 334 00:18:27,119 --> 00:18:29,960 Speaker 4: model comes out, they use a different benchmark, they use 335 00:18:30,000 --> 00:18:35,240 Speaker 4: different standards. So even just having you know, I don't 336 00:18:35,240 --> 00:18:37,080 Speaker 4: want to like stand behind it, because what if we 337 00:18:37,080 --> 00:18:40,080 Speaker 4: find out, you know, the companies are like it just 338 00:18:40,480 --> 00:18:43,399 Speaker 4: the financial incentives are so skewed towards the companies that 339 00:18:43,440 --> 00:18:45,720 Speaker 4: it would be really hard for a third party. However, 340 00:18:46,280 --> 00:18:48,640 Speaker 4: I mean even just having a standard across the board 341 00:18:48,680 --> 00:18:53,040 Speaker 4: would make it a lot easier to compare models, you know, 342 00:18:53,160 --> 00:18:57,199 Speaker 4: to like for outsiders for the public to push for 343 00:18:57,280 --> 00:18:59,520 Speaker 4: different kinds of standards, like right now, we have no 344 00:18:59,640 --> 00:19:04,119 Speaker 4: say in how they're testing, what they're testing for, you know, 345 00:19:04,200 --> 00:19:07,080 Speaker 4: and and like the kinds of language. 346 00:19:06,680 --> 00:19:07,919 Speaker 2: That they use, the methods they use. 347 00:19:08,000 --> 00:19:10,800 Speaker 4: So at least if it were a third party, you know, 348 00:19:10,880 --> 00:19:13,359 Speaker 4: we would have that approach. But that, you know, doesn't 349 00:19:13,520 --> 00:19:15,719 Speaker 4: mean that it's better if it's a market based approach 350 00:19:15,800 --> 00:19:19,040 Speaker 4: versus Some of the testing that we heard was done 351 00:19:19,040 --> 00:19:26,920 Speaker 4: through Casey, the NIST organization, which is very very very 352 00:19:26,920 --> 00:19:29,880 Speaker 4: small budget but was apparently did like some great testing 353 00:19:29,920 --> 00:19:35,880 Speaker 4: of mythos maybe and the UK the UK Ai Safety Institute, 354 00:19:35,880 --> 00:19:37,560 Speaker 4: which was one of the first to like jail break 355 00:19:37,600 --> 00:19:41,960 Speaker 4: a lot of US models. I think, like a third party, yes, 356 00:19:42,240 --> 00:19:44,719 Speaker 4: and then I guess the rest kind of depends on 357 00:19:44,760 --> 00:19:45,800 Speaker 4: the on the fine print. 358 00:19:46,240 --> 00:19:48,639 Speaker 1: Just on the topic of AI and China, Taylor, I 359 00:19:48,680 --> 00:19:50,560 Speaker 1: saw another story this week they made me think of you, 360 00:19:50,640 --> 00:19:55,800 Speaker 1: which is that China has moved to ban Chinese AI 361 00:19:55,920 --> 00:20:00,720 Speaker 1: companies developing chatbots that people establish comp onionship with or 362 00:20:00,720 --> 00:20:03,840 Speaker 1: any kind of emotional connection whatsoever. I'm curious if you 363 00:20:03,880 --> 00:20:05,639 Speaker 1: saw the story and what it made you think. 364 00:20:06,200 --> 00:20:10,040 Speaker 3: Yes, they're specifically sort of trying to forbid chatbots that 365 00:20:10,080 --> 00:20:14,359 Speaker 3: are designed for companionship. So, as you mentioned, it's like 366 00:20:15,200 --> 00:20:17,719 Speaker 3: the chatbots that are like kind of like the character 367 00:20:17,760 --> 00:20:22,159 Speaker 3: AI type ones here where they're specifically designed to provide 368 00:20:22,440 --> 00:20:26,399 Speaker 3: a companion like experience, and so they're sort of built 369 00:20:26,560 --> 00:20:31,320 Speaker 3: to you know, give more maybe emotionally resident answers. They're 370 00:20:31,359 --> 00:20:34,479 Speaker 3: you know, designed to kind of help you navigate your 371 00:20:34,520 --> 00:20:39,320 Speaker 3: life problems, somebody to talk to. You know. This is 372 00:20:39,440 --> 00:20:42,680 Speaker 3: this is obviously like a very sort of controversial topic. 373 00:20:43,160 --> 00:20:45,959 Speaker 3: I think what's so concerning is is that you know, 374 00:20:46,119 --> 00:20:50,480 Speaker 3: China has extremely authoritarian uh sort of control over the 375 00:20:50,520 --> 00:20:55,040 Speaker 3: content that comes out of the tech ecosystem there, Like 376 00:20:55,080 --> 00:20:57,520 Speaker 3: they have a lot of government oversight. I don't think 377 00:20:57,520 --> 00:21:00,720 Speaker 3: that we have any evidence like this is not a 378 00:21:00,800 --> 00:21:03,760 Speaker 3: very evidence based approach, Like they're saying that they want 379 00:21:03,760 --> 00:21:06,840 Speaker 3: to ban it so that you know, basically so that 380 00:21:06,880 --> 00:21:09,520 Speaker 3: people have more babies. And I just I don't think 381 00:21:09,560 --> 00:21:11,959 Speaker 3: that the reason people are not having more babies in 382 00:21:12,040 --> 00:21:15,200 Speaker 3: China is related to the AI chatlots at all. I 383 00:21:15,200 --> 00:21:17,640 Speaker 3: think it's just again this is as you normally hear 384 00:21:17,680 --> 00:21:21,000 Speaker 3: with them sort of these types of tech laws. It's 385 00:21:21,080 --> 00:21:23,960 Speaker 3: not really based in science, it's based more in vibes. 386 00:21:24,680 --> 00:21:26,399 Speaker 1: It is interesting those and I mean it's it's like, 387 00:21:26,440 --> 00:21:28,880 Speaker 1: you know, we have, you know, in the US all 388 00:21:28,920 --> 00:21:31,680 Speaker 1: this like handering about technology out of control and their 389 00:21:31,680 --> 00:21:34,159 Speaker 1: regulation and et cetera, et cetera, and then you know, 390 00:21:34,240 --> 00:21:37,280 Speaker 1: you have this policy in China, and it's so hard 391 00:21:37,320 --> 00:21:40,680 Speaker 1: to you know, between the between the between those two 392 00:21:40,760 --> 00:21:45,240 Speaker 1: paths figure out like how how should societies you know, Well, I. 393 00:21:45,160 --> 00:21:47,919 Speaker 3: Think what's really scary and concerning is that we're seeing 394 00:21:47,960 --> 00:21:51,960 Speaker 3: America begin to regulate our information ecosystem increasingly like China. 395 00:21:52,119 --> 00:21:55,160 Speaker 3: So we continue to pass laws and actually you'll see 396 00:21:55,160 --> 00:22:00,000 Speaker 3: people that are ostensibly liberals praising China and extremely other 397 00:22:00,000 --> 00:22:03,199 Speaker 3: authoritarian regimes. And we have like Jonathan Heights, sort of 398 00:22:03,200 --> 00:22:06,560 Speaker 3: pseudoscience guy who's amassed a big audience here as a pundit, 399 00:22:07,119 --> 00:22:09,840 Speaker 3: praising countries like Malaysia and Indonesia. You know, these are 400 00:22:09,880 --> 00:22:12,920 Speaker 3: countries that don't have for free press, that arrest people 401 00:22:12,920 --> 00:22:15,800 Speaker 3: for their online speech, and often they're working with them. 402 00:22:15,800 --> 00:22:20,159 Speaker 3: I mean, we had even lawmakers praising the UAE and 403 00:22:20,200 --> 00:22:23,880 Speaker 3: their restrictions on sort of speech online and access to information. 404 00:22:24,000 --> 00:22:27,960 Speaker 3: So I think this is really worrying. There's also a 405 00:22:27,960 --> 00:22:30,359 Speaker 3: bunch of research that has come out around sort of 406 00:22:30,359 --> 00:22:36,600 Speaker 3: information and how sort of different llms criticize government. Basically, 407 00:22:36,600 --> 00:22:40,200 Speaker 3: in countries with lower levels of media freedom, when you 408 00:22:40,400 --> 00:22:43,760 Speaker 3: query models in the language of that country about politics 409 00:22:43,880 --> 00:22:46,440 Speaker 3: or the government or whatever, you are much more likely 410 00:22:46,520 --> 00:22:49,679 Speaker 3: to get a pro regime authoritarian answer than if you 411 00:22:49,720 --> 00:22:52,960 Speaker 3: ask in a different language. So these models, and this 412 00:22:53,000 --> 00:22:55,359 Speaker 3: came out in like a paper last month. Met As 413 00:22:55,400 --> 00:22:58,359 Speaker 3: Oversight Board is doing research into this. But basically what 414 00:22:58,400 --> 00:23:03,400 Speaker 3: we realize is that like training, data is shaped by 415 00:23:03,520 --> 00:23:06,400 Speaker 3: the sort of political ecosystem of that country. This has 416 00:23:06,440 --> 00:23:09,520 Speaker 3: consequences for LLM output. We need to make sure that 417 00:23:09,600 --> 00:23:13,480 Speaker 3: we are preserving our free and open information ecosystem here 418 00:23:13,480 --> 00:23:17,000 Speaker 3: in the US. And I don't think we're doing any 419 00:23:17,040 --> 00:23:18,720 Speaker 3: of that. And I think when we look at laws 420 00:23:18,760 --> 00:23:21,800 Speaker 3: like what China is doing and we sort of praise it, 421 00:23:21,600 --> 00:23:22,920 Speaker 3: it makes me very nervous. 422 00:23:28,960 --> 00:23:31,480 Speaker 1: When we come back, Tayla tells us why all the 423 00:23:31,720 --> 00:23:47,639 Speaker 1: girls having a hot surveillance summer stay with us? Welcome back, Taylor. 424 00:23:47,920 --> 00:23:50,639 Speaker 1: Where does the phrase hot surveillance summer come from? What 425 00:23:50,720 --> 00:23:51,200 Speaker 1: does it mean? 426 00:23:52,840 --> 00:23:57,280 Speaker 3: So it's actually from this podcaster Natalie, who hosts this 427 00:23:57,320 --> 00:24:00,560 Speaker 3: podcast Boys Club It's like a tech podcast. She wrote 428 00:24:00,600 --> 00:24:03,600 Speaker 3: this post about having a surveillance summer that was basically 429 00:24:03,640 --> 00:24:09,680 Speaker 3: pro Meta smart glasses. She came out right along when 430 00:24:09,880 --> 00:24:12,320 Speaker 3: sort of Meta began to launch this campaign with Kylie 431 00:24:12,400 --> 00:24:15,040 Speaker 3: Jenner and all the influencers are having, you know, their 432 00:24:15,080 --> 00:24:18,159 Speaker 3: sort of surveillance summer moment. Basically what she said is 433 00:24:18,200 --> 00:24:20,840 Speaker 3: like masurveillances everywhere, so why not participate in it? 434 00:24:22,320 --> 00:24:23,720 Speaker 2: Which is I. 435 00:24:23,760 --> 00:24:28,800 Speaker 3: Trust, it's horrifying. I hate it. But a lot of influencers, 436 00:24:28,800 --> 00:24:31,320 Speaker 3: if you look at the replies to her posts, were like, yes, 437 00:24:31,640 --> 00:24:33,960 Speaker 3: let's go. And you know, I was in New York 438 00:24:34,440 --> 00:24:36,399 Speaker 3: just last week and I saw the first pair of 439 00:24:36,440 --> 00:24:38,520 Speaker 3: Meta smart glasses in the wild on this girl. I 440 00:24:38,560 --> 00:24:39,960 Speaker 3: don't know if she was an influencer, but she was 441 00:24:40,000 --> 00:24:42,719 Speaker 3: really pretty, and I imagine that she was an influencer 442 00:24:42,720 --> 00:24:46,360 Speaker 3: because she had the early access to these glasses. But they're, yeah, 443 00:24:46,359 --> 00:24:49,280 Speaker 3: they're getting popular and people just kind of are treating 444 00:24:49,280 --> 00:24:51,560 Speaker 3: it as this like fun thing and they love the 445 00:24:51,640 --> 00:24:56,040 Speaker 3: Kylie glasses and yeah, they're talking about sort of proactively 446 00:24:56,040 --> 00:24:58,000 Speaker 3: having a hot surveillanceummer in like a positive way. 447 00:24:58,480 --> 00:25:00,880 Speaker 1: Tally can you kind of lay out the history smart glasses. 448 00:25:00,880 --> 00:25:03,439 Speaker 1: I remember when I was a TV producer back in 449 00:25:03,480 --> 00:25:07,399 Speaker 1: like the twenty eleven period, Sebastian Thrun came on the 450 00:25:07,440 --> 00:25:11,040 Speaker 1: show I worked, only had this like Bionic Man prototype 451 00:25:11,040 --> 00:25:15,000 Speaker 1: of the Google glass that didn't quite catch on. Then 452 00:25:15,040 --> 00:25:18,560 Speaker 1: they were the first generation of Snapchat spectacles. Then metas 453 00:25:18,640 --> 00:25:21,200 Speaker 1: ray Band partnership kind of brought these somewhat into the mainstream. 454 00:25:21,240 --> 00:25:25,600 Speaker 1: Obviously Kylie is now has her own Meta ray Ban collab, 455 00:25:26,000 --> 00:25:27,640 Speaker 1: and then later this year or early next year, snap 456 00:25:27,760 --> 00:25:30,480 Speaker 1: had to going to release these new spectacles are apparently 457 00:25:30,480 --> 00:25:32,280 Speaker 1: going to cost two thousand dollars. I mean, what's the 458 00:25:32,359 --> 00:25:35,159 Speaker 1: kind of what's the technological history of these glasses? And 459 00:25:35,280 --> 00:25:37,000 Speaker 1: are they are there things I've actually I'm not sure 460 00:25:37,040 --> 00:25:40,399 Speaker 1: i've ever seen anyone actually wearing them in the wild 461 00:25:40,680 --> 00:25:41,560 Speaker 1: as well as I know. 462 00:25:41,920 --> 00:25:47,960 Speaker 3: I've seen them unfortunately never Google glass. But but but yeah, 463 00:25:48,000 --> 00:25:50,360 Speaker 3: I mean these, I mean Meta RayBan glasses are popular. 464 00:25:50,600 --> 00:25:52,320 Speaker 3: I mean people have them. I've seen them in you 465 00:25:52,760 --> 00:25:56,600 Speaker 3: in LA. I think what's interesting to me, I mean 466 00:25:56,640 --> 00:25:59,080 Speaker 3: you just sort of like laid out the general progression 467 00:25:59,200 --> 00:26:01,439 Speaker 3: of a smart glas is I think the idea of 468 00:26:01,480 --> 00:26:04,040 Speaker 3: having something on your face that is constantly recording. That 469 00:26:04,119 --> 00:26:07,480 Speaker 3: can you know, act as this sort of second brain 470 00:26:07,560 --> 00:26:09,800 Speaker 3: for you, remind you of someone you can't remember, someone 471 00:26:09,880 --> 00:26:11,320 Speaker 3: you know, Like that's always been the sort of like 472 00:26:11,359 --> 00:26:15,280 Speaker 3: sci fi fantasy for years. We've had recording capabilities. So 473 00:26:15,560 --> 00:26:19,399 Speaker 3: I actually had the Snapchat spectacles back in twenty sixteen, 474 00:26:19,920 --> 00:26:21,920 Speaker 3: and you know, you could record very easily, you could 475 00:26:21,920 --> 00:26:24,520 Speaker 3: post a snapchat, so and that is sort of what 476 00:26:24,560 --> 00:26:27,560 Speaker 3: pioneered I think the popularity, the niche popularity, I guess, 477 00:26:27,600 --> 00:26:30,800 Speaker 3: of the original sort of meta rap ban glasses. What 478 00:26:30,960 --> 00:26:32,920 Speaker 3: makes this moment interesting to me and why I think 479 00:26:32,920 --> 00:26:37,000 Speaker 3: we're reaching a turning point in the normalization of these 480 00:26:37,040 --> 00:26:40,880 Speaker 3: products is that you now have sort of plausible deniability. 481 00:26:40,920 --> 00:26:43,880 Speaker 3: Previously they were only for recording, right, And so it's 482 00:26:43,920 --> 00:26:46,920 Speaker 3: kind of weird, like your people know that that's why 483 00:26:46,960 --> 00:26:49,080 Speaker 3: you're wearing them. It's like, what do you creep? Why 484 00:26:49,119 --> 00:26:50,680 Speaker 3: do you have to record? Why you know? Why can't 485 00:26:50,680 --> 00:26:52,200 Speaker 3: you just take out your phone like a normal person. 486 00:26:52,920 --> 00:26:55,120 Speaker 3: But two things have happened. One, we've actually normalized taking 487 00:26:55,160 --> 00:26:57,560 Speaker 3: out our phone and recording at all times. That's basically 488 00:26:57,560 --> 00:27:01,159 Speaker 3: been completely destigmatized ten years ago. That was actually still stigmatized. 489 00:27:01,160 --> 00:27:02,760 Speaker 3: Now there was like houn its like influencers in the 490 00:27:02,760 --> 00:27:04,879 Speaker 3: wild that would like make fun of people recording in public. 491 00:27:05,000 --> 00:27:09,080 Speaker 3: Now everyone records in public. Second of all is that 492 00:27:09,160 --> 00:27:11,520 Speaker 3: these glasses have AI integrated into them, and so they're 493 00:27:11,560 --> 00:27:14,439 Speaker 3: being sold as this like smart AI assistant. And so 494 00:27:14,480 --> 00:27:17,200 Speaker 3: the people that I spoke to in the replies of 495 00:27:17,240 --> 00:27:19,199 Speaker 3: this tweet that I was like, why would you do this, 496 00:27:19,359 --> 00:27:21,919 Speaker 3: They're like, well, I want AI. I want somebody to 497 00:27:22,080 --> 00:27:24,359 Speaker 3: see that everything that I see and help me. And 498 00:27:24,400 --> 00:27:27,080 Speaker 3: that is I think when I sort of see the 499 00:27:27,160 --> 00:27:31,160 Speaker 3: discussion around these consumer products, they're like, well, yeah, people 500 00:27:31,160 --> 00:27:33,080 Speaker 3: that record all day around them? Are you know, if 501 00:27:33,080 --> 00:27:35,080 Speaker 3: you're sort of like recording women with your glasses, you're 502 00:27:35,080 --> 00:27:37,680 Speaker 3: a greep. But I actually use them for help for AI. 503 00:27:37,920 --> 00:27:40,400 Speaker 3: And this is like part of the sort of girl 504 00:27:40,440 --> 00:27:42,960 Speaker 3: bossification of AI, right, Like I use my metal smart 505 00:27:42,960 --> 00:27:45,400 Speaker 3: glasses all day long and they help me figure out 506 00:27:45,400 --> 00:27:48,840 Speaker 3: what to buy at arrow on or you know, I 507 00:27:48,840 --> 00:27:53,000 Speaker 3: don't know, remember to shop stop by Alo or something. 508 00:27:53,040 --> 00:27:54,960 Speaker 3: I don't know. But it's yeah, which is I think 509 00:27:54,960 --> 00:27:57,080 Speaker 3: notable because when you look at who they've been giving 510 00:27:57,119 --> 00:27:58,600 Speaker 3: these glasses to, it is a lot of the like 511 00:27:58,640 --> 00:28:02,679 Speaker 3: West Village girlies, like lifestyle influencers, these aspirational women. 512 00:28:03,359 --> 00:28:05,840 Speaker 1: Natasha, you've been flowing this story, yes. 513 00:28:05,720 --> 00:28:08,480 Speaker 4: I mean like I of course went through the whole 514 00:28:09,080 --> 00:28:11,840 Speaker 4: Kylie photodump of the of the Meta glasses. And I 515 00:28:11,840 --> 00:28:14,520 Speaker 4: think when I moved to San Francisco was just when, 516 00:28:15,400 --> 00:28:20,240 Speaker 4: like maybe slightly after, people were being accosted in bars 517 00:28:20,280 --> 00:28:23,199 Speaker 4: for wearing the Google glasses, which are you know, like 518 00:28:23,440 --> 00:28:27,480 Speaker 4: extremely it looked you can tell when somebody was wearing 519 00:28:27,520 --> 00:28:29,639 Speaker 4: Google glasses, So I did see them here. I have 520 00:28:29,800 --> 00:28:32,720 Speaker 4: only seen the Metal ones in La one time, and 521 00:28:32,760 --> 00:28:35,720 Speaker 4: people were giving the guy a hard time. He was 522 00:28:35,760 --> 00:28:38,200 Speaker 4: on it was like on Abbot Kinney and you know, 523 00:28:38,360 --> 00:28:39,520 Speaker 4: it was just like what are you doing? 524 00:28:39,640 --> 00:28:43,400 Speaker 2: U creep? But this was maybe in January, so a 525 00:28:43,440 --> 00:28:44,040 Speaker 2: while ago. 526 00:28:45,320 --> 00:28:53,280 Speaker 4: Yeah, I mean, I understand the appeal of potentially having 527 00:28:53,960 --> 00:28:59,880 Speaker 4: something like as an always on assistant. I for some reason, 528 00:28:59,920 --> 00:29:02,400 Speaker 4: I always think about that scene in Devilwaar's Prada where 529 00:29:02,400 --> 00:29:05,680 Speaker 4: they're like whispering the name of the person to you. 530 00:29:06,280 --> 00:29:09,680 Speaker 4: I understand why we would want to feel like we're 531 00:29:09,880 --> 00:29:13,240 Speaker 4: the boss and we have you know, minions helping us. 532 00:29:13,280 --> 00:29:15,480 Speaker 2: But I don't know. I mean at least when I 533 00:29:15,520 --> 00:29:15,920 Speaker 2: put on. 534 00:29:15,960 --> 00:29:18,200 Speaker 4: The Google glasses, I remember too that it was just 535 00:29:18,520 --> 00:29:22,680 Speaker 4: it was not the seamless minority report experience that that 536 00:29:22,760 --> 00:29:23,880 Speaker 4: Sci Fi promised us. 537 00:29:23,960 --> 00:29:26,000 Speaker 2: So I don't know. 538 00:29:26,080 --> 00:29:28,960 Speaker 4: I mean, what do you think Taylor about, like how 539 00:29:30,480 --> 00:29:34,600 Speaker 4: how savvy they've been with targeting the right influencers. Do 540 00:29:34,640 --> 00:29:39,479 Speaker 4: you feel like going through women has been like, has 541 00:29:39,560 --> 00:29:41,280 Speaker 4: led to more adoption than they would have if they 542 00:29:41,360 --> 00:29:43,240 Speaker 4: tried to go like the Google glass route. 543 00:29:43,640 --> 00:29:46,960 Speaker 3: Undeniably. I think, like number one, they've destigmatized the idea 544 00:29:47,000 --> 00:29:49,360 Speaker 3: that these are creep glasses used to surveil women, which 545 00:29:49,400 --> 00:29:54,280 Speaker 3: is what people sort of thought of metaglasses before. And 546 00:29:54,800 --> 00:29:57,480 Speaker 3: I think they've really made it aspirational. I think the 547 00:29:57,960 --> 00:30:01,479 Speaker 3: HIGHI partnership was brilliant. I think the sort of rollout 548 00:30:01,480 --> 00:30:03,960 Speaker 3: has been great, and I think you're seeing a lot 549 00:30:03,960 --> 00:30:07,600 Speaker 3: of women get reflexively defensive and be like, well I 550 00:30:07,640 --> 00:30:09,400 Speaker 3: want I mean I talked to my friend the other 551 00:30:09,440 --> 00:30:11,320 Speaker 3: day who wants to buy them because she has her kids, 552 00:30:11,320 --> 00:30:13,120 Speaker 3: and she's like, oh, it would be so great, Like 553 00:30:13,160 --> 00:30:14,840 Speaker 3: my hands are always full, I could just talk to 554 00:30:14,880 --> 00:30:17,280 Speaker 3: the glasses and also then I could record moments with 555 00:30:17,320 --> 00:30:18,800 Speaker 3: my children. I don't have to have my phone out 556 00:30:18,800 --> 00:30:21,640 Speaker 3: and like, I totally get it. I totally get it. 557 00:30:22,000 --> 00:30:22,400 Speaker 4: Like I. 558 00:30:24,360 --> 00:30:26,560 Speaker 3: Hate the mass surveillance aspect. I think like, if we 559 00:30:26,600 --> 00:30:28,680 Speaker 3: are going to normalize this level of masurveillance, let's have 560 00:30:28,720 --> 00:30:32,400 Speaker 3: like serious conversations around policy and privacy and privacy protections. 561 00:30:32,400 --> 00:30:34,080 Speaker 3: I think it's interesting as well. There was that Wired 562 00:30:34,120 --> 00:30:37,160 Speaker 3: report about that system called face Card that Meta was 563 00:30:37,440 --> 00:30:38,400 Speaker 3: going to ask you about this. 564 00:30:38,400 --> 00:30:41,640 Speaker 1: This is interesting idea that the lasses were cross reference 565 00:30:41,680 --> 00:30:43,760 Speaker 1: with the database of all faces, so you could kind 566 00:30:43,760 --> 00:30:47,200 Speaker 1: of live identify people. But then they discontinued that, or 567 00:30:47,240 --> 00:30:49,480 Speaker 1: they kind of sum set it for the time being 568 00:30:49,520 --> 00:30:51,520 Speaker 1: off the wide article or what happened. 569 00:30:51,440 --> 00:30:55,360 Speaker 3: Well sort of, And you know, Boz was on this 570 00:30:55,400 --> 00:30:59,840 Speaker 3: podcast with Nicholas Thompson, the editor of The Atlantic, a 571 00:31:00,160 --> 00:31:03,040 Speaker 3: Meta executive, and he's like, you know, I can't believe 572 00:31:03,040 --> 00:31:04,640 Speaker 3: there was so much backlash. Like, first of all, he 573 00:31:04,680 --> 00:31:07,800 Speaker 3: confirmed the story after Andy Stone, the comms person for Meta, 574 00:31:07,840 --> 00:31:10,280 Speaker 3: was like on Twitter all day denying it. This is nonsense, 575 00:31:10,320 --> 00:31:12,880 Speaker 3: this is nonsense. Then you have bas an executive going 576 00:31:12,920 --> 00:31:14,960 Speaker 3: on a podcast and being like, actually, we're really proud 577 00:31:14,960 --> 00:31:17,680 Speaker 3: of this work and it's actually really amazing. My favorite 578 00:31:17,720 --> 00:31:19,640 Speaker 3: part of that interview is that he's like, oh, people 579 00:31:19,640 --> 00:31:22,040 Speaker 3: think that there's just going to be some master database 580 00:31:22,080 --> 00:31:23,480 Speaker 3: and like it's just going to be pulling everyone for 581 00:31:23,560 --> 00:31:26,080 Speaker 3: the face databases. No, it's going to be pulling every 582 00:31:26,240 --> 00:31:28,800 Speaker 3: it's going to be pulling from a database that shows 583 00:31:28,800 --> 00:31:31,240 Speaker 3: you your connections, if you met that person once or whatever. 584 00:31:31,280 --> 00:31:33,920 Speaker 3: It's just like, Okay, that is such a meaningless distinction. 585 00:31:33,960 --> 00:31:36,520 Speaker 3: The point is that you are creating a database of 586 00:31:36,560 --> 00:31:40,840 Speaker 3: biometric data and facial recognition data and tying that to people, 587 00:31:40,920 --> 00:31:43,400 Speaker 3: and that is what we are mad about, you know. 588 00:31:43,760 --> 00:31:46,200 Speaker 1: So it's amazing. I remember when Facebook first came out, 589 00:31:46,200 --> 00:31:47,920 Speaker 1: and I was at high school. Maybe not when it 590 00:31:47,920 --> 00:31:49,880 Speaker 1: first came out, but when it first started diffuse and 591 00:31:50,120 --> 00:31:53,040 Speaker 1: I was supposed sixteen or seventeen, I remember looking at 592 00:31:53,080 --> 00:31:55,720 Speaker 1: him thinking, oh my god, like there was soon not 593 00:31:55,840 --> 00:31:59,080 Speaker 1: in the not too in futury, possible to overlay everyone voluntary, 594 00:31:59,240 --> 00:32:02,480 Speaker 1: voluntarily kind of uploading themselves to this platform and just 595 00:32:02,600 --> 00:32:04,800 Speaker 1: having no expectation of privacy and people being able to 596 00:32:04,840 --> 00:32:07,480 Speaker 1: know who you are all the time. And it's like, wh. 597 00:32:07,560 --> 00:32:10,920 Speaker 3: Boy, there, which, by the way, we normalize that, I mean, 598 00:32:11,040 --> 00:32:14,320 Speaker 3: I like, I mean, I did the same thing with Facebook, 599 00:32:14,360 --> 00:32:16,360 Speaker 3: like I would stock down everyone in my class, Like 600 00:32:16,400 --> 00:32:19,520 Speaker 3: I think I added everyone in my college dorm, you 601 00:32:19,560 --> 00:32:21,920 Speaker 3: know when I got into that. And so it's like 602 00:32:22,040 --> 00:32:26,200 Speaker 3: I think that, like people used to have a huge 603 00:32:26,200 --> 00:32:29,920 Speaker 3: amount of privacy, we have watched privacy be obliterated over 604 00:32:30,000 --> 00:32:32,920 Speaker 3: the past couple decades with the rise of social media. 605 00:32:33,320 --> 00:32:37,960 Speaker 3: Now we have AI which is super charging privacy violations. 606 00:32:38,440 --> 00:32:41,240 Speaker 3: I think that we are not having any conversations about 607 00:32:41,360 --> 00:32:43,720 Speaker 3: privacy like we should be having. We should be passing 608 00:32:43,760 --> 00:32:48,440 Speaker 3: privacy laws. Instead, we're passing the opposite. We're actually mandating. 609 00:32:48,920 --> 00:32:52,680 Speaker 3: Over twenty five states have mandated identity verification laws to 610 00:32:52,760 --> 00:32:56,240 Speaker 3: use Internet and then or to use parts of the Internet, 611 00:32:56,280 --> 00:32:58,160 Speaker 3: I should say. And then I think there's forty seven 612 00:32:58,200 --> 00:33:00,640 Speaker 3: states that are that are that have total either pass 613 00:33:00,680 --> 00:33:02,640 Speaker 3: these laws or are considering past them. And then we 614 00:33:02,680 --> 00:33:05,560 Speaker 3: have the Federal Kids Act discussions. So we're about to 615 00:33:05,640 --> 00:33:09,880 Speaker 3: mandate that Meta start harvesting biometric data, including on children 616 00:33:10,200 --> 00:33:12,920 Speaker 3: while they're building this technology. So if we want, if 617 00:33:12,920 --> 00:33:16,360 Speaker 3: we're really against surveillance, then we have to have a 618 00:33:16,360 --> 00:33:17,720 Speaker 3: bigger discussion about privacy. 619 00:33:18,280 --> 00:33:20,280 Speaker 1: Just to close, Natasha, I want to ask you about 620 00:33:20,360 --> 00:33:22,920 Speaker 1: and There's two interesting stories I saw this week that 621 00:33:22,960 --> 00:33:26,600 Speaker 1: are kind of adjacent to this. One is Apple suing 622 00:33:26,760 --> 00:33:32,160 Speaker 1: open Ai for allegedly getting interviewing Apple execs and basically 623 00:33:32,240 --> 00:33:36,800 Speaker 1: using the interviews to steal Apple trade secrets around device manufacturing. 624 00:33:37,280 --> 00:33:39,560 Speaker 1: And the other was a report in Bloomberg about open 625 00:33:39,600 --> 00:33:42,720 Speaker 1: ai bringing its first device to market, which is not 626 00:33:42,800 --> 00:33:44,960 Speaker 1: a phone and not glasses. It sounds more like an 627 00:33:45,000 --> 00:33:47,920 Speaker 1: Alexa or a Sonos. But can you give us some 628 00:33:48,400 --> 00:33:49,640 Speaker 1: contexts on those two stories. 629 00:33:50,320 --> 00:33:55,160 Speaker 4: Yeah, the Apple lawsuit was super interesting because it was 630 00:33:55,280 --> 00:33:59,560 Speaker 4: chock full of details about, you know, even having access 631 00:33:59,640 --> 00:34:04,040 Speaker 4: to people's texts and messages where they're like, lol, I 632 00:34:04,080 --> 00:34:07,920 Speaker 4: guess I still have access to this or I didn't 633 00:34:08,000 --> 00:34:09,960 Speaker 4: know you could, you know, bring a piece of an 634 00:34:09,960 --> 00:34:13,040 Speaker 4: Apple device out of the out of the company because 635 00:34:13,560 --> 00:34:16,359 Speaker 4: this long time Apple executive who moved to open ai 636 00:34:16,560 --> 00:34:19,839 Speaker 4: suggested that when you come to open ai, you bring 637 00:34:20,000 --> 00:34:24,000 Speaker 4: like an actual physical thing for a little show and 638 00:34:24,040 --> 00:34:27,600 Speaker 4: tell session. But one thing I thought was also really 639 00:34:27,600 --> 00:34:33,480 Speaker 4: interesting is Tony Fidel, the CEO of Nest. He reached 640 00:34:33,480 --> 00:34:39,399 Speaker 4: out after this newsletter writer Ben Thompson Etstra Tetri wrote 641 00:34:39,400 --> 00:34:41,680 Speaker 4: about it, and he said, this is how Apple scares 642 00:34:41,719 --> 00:34:44,640 Speaker 4: employees like this is I think what he was doing 643 00:34:44,719 --> 00:34:47,240 Speaker 4: was trying to frame it as not necessarily going after 644 00:34:47,440 --> 00:34:50,319 Speaker 4: open Ai, like you know, being sure that they'll they'll 645 00:34:50,360 --> 00:34:54,040 Speaker 4: be able to you know, get damages or prevent something 646 00:34:54,040 --> 00:34:57,120 Speaker 4: from open Ai, but rather as a message to Apple employees, 647 00:34:57,560 --> 00:35:00,000 Speaker 4: of which open ai has hired four hundreds. 648 00:35:00,239 --> 00:35:03,520 Speaker 1: But these were texts onto Apple employees on their work phones, 649 00:35:03,560 --> 00:35:05,799 Speaker 1: which therefore Apple could read. Or how did Apple start 650 00:35:05,840 --> 00:35:07,360 Speaker 1: to like see this, Corsemond. 651 00:35:07,080 --> 00:35:10,120 Speaker 4: Yeah, yeah, what like the a few of the texts 652 00:35:10,120 --> 00:35:13,400 Speaker 4: that they already had were on their Apple devices remaining 653 00:35:13,400 --> 00:35:15,600 Speaker 4: on their Apple devices, and they believe that, you know, 654 00:35:15,640 --> 00:35:17,719 Speaker 4: if they take it to discovery, they'll be able to 655 00:35:17,800 --> 00:35:22,480 Speaker 4: access even more. And they also mentioned this, you know, 656 00:35:22,520 --> 00:35:26,080 Speaker 4: they mentioned the Johnny I've partnership, the Johnny Ive company 657 00:35:26,120 --> 00:35:29,319 Speaker 4: that was that was acquired by open Ai, and they 658 00:35:29,360 --> 00:35:32,080 Speaker 4: were saying that you know, Johnny went to it was 659 00:35:32,120 --> 00:35:36,080 Speaker 4: something in particular about the metal finishing, like a proprietary 660 00:35:36,120 --> 00:35:39,200 Speaker 4: metal finishing that only Apple knows, and Johnny I've it 661 00:35:39,239 --> 00:35:42,000 Speaker 4: sounded like basically he's going to their suppliers and using 662 00:35:42,080 --> 00:35:45,120 Speaker 4: like Apple terminology, and so they had the impression this 663 00:35:45,239 --> 00:35:49,640 Speaker 4: was for an Apple product, And yeah, I guess we'll 664 00:35:49,680 --> 00:35:52,600 Speaker 4: see when this new Alexa comes out. I thought it 665 00:35:52,640 --> 00:35:54,680 Speaker 4: was supposed to be that tiny thing that we saw 666 00:35:55,239 --> 00:35:58,319 Speaker 4: like a few executives carrying around and get photographed with 667 00:35:58,520 --> 00:36:03,759 Speaker 4: like a compact almost with like in a dark metal finish, 668 00:36:04,080 --> 00:36:06,399 Speaker 4: with I think a little open AI logo. It looked 669 00:36:06,440 --> 00:36:09,719 Speaker 4: like a little clamshaw kind of thing and people assumed 670 00:36:10,200 --> 00:36:11,480 Speaker 4: that that was the device. 671 00:36:11,600 --> 00:36:15,720 Speaker 2: But yeah, it sounds much less exciting. 672 00:36:16,360 --> 00:36:18,960 Speaker 1: That's all we have time for this week. Taylor, Natasha, 673 00:36:19,000 --> 00:36:19,439 Speaker 1: thank you. 674 00:36:19,600 --> 00:36:20,200 Speaker 2: Good to see you. 675 00:36:20,400 --> 00:36:21,879 Speaker 3: Thanks for having us. 676 00:36:43,520 --> 00:36:46,680 Speaker 1: For tech stuff. I'm os Voloshin. This episode was produced 677 00:36:46,680 --> 00:36:49,960 Speaker 1: by Eliza Dennis. It was executive produced by me and 678 00:36:50,040 --> 00:36:54,839 Speaker 1: Julian Nutta for Kaleidoscope and Katrina norvelf iHeart Podcasts. Jack 679 00:36:54,880 --> 00:36:57,960 Speaker 1: Insley makes this episode and Kyle Murdoch wrote our theme song. 680 00:36:58,440 --> 00:37:01,240 Speaker 1: A special thank you to Taylor Lorenz and Natasha Tiku. 681 00:37:01,640 --> 00:37:03,759 Speaker 1: Please check out all the work they put out into 682 00:37:03,760 --> 00:37:06,279 Speaker 1: the world. We're lucky to call them friends of the Pod.