1 00:00:00,200 --> 00:00:02,880 Speaker 1: Thanks for tunion to tech Stuff. If you don't recognize 2 00:00:02,880 --> 00:00:05,640 Speaker 1: my voice, my name is Oz Volshan and I'm here 3 00:00:05,680 --> 00:00:08,880 Speaker 1: because the inimitable Jonathan Strickland has passed the baton to 4 00:00:08,960 --> 00:00:12,000 Speaker 1: Cara Price and myself to host tech Stuff. The show 5 00:00:12,039 --> 00:00:14,920 Speaker 1: will remain your home for all things tech, and all 6 00:00:14,960 --> 00:00:18,200 Speaker 1: the old episodes will remain available in this feed. Thanks 7 00:00:18,239 --> 00:00:22,239 Speaker 1: for listening. Welcome to tech Stuff, a production of iHeart 8 00:00:22,239 --> 00:00:26,079 Speaker 1: Podcasts and Kaleidoscope. I'm oz Voloshian, and today will bring 9 00:00:26,160 --> 00:00:29,480 Speaker 1: you the headlines this week, including the Chinese AI company 10 00:00:29,560 --> 00:00:33,199 Speaker 1: that spooked the US tech sector. On today's Tech Support segment, 11 00:00:33,440 --> 00:00:35,880 Speaker 1: we'll talk to four or four medius Jason Kebler about 12 00:00:35,880 --> 00:00:39,760 Speaker 1: building mazes to trap AI web crawlers, and then we've 13 00:00:39,800 --> 00:00:42,640 Speaker 1: got when did this become a thing? This time we 14 00:00:42,680 --> 00:00:45,600 Speaker 1: look at why everyone is so obsessed with their own data, 15 00:00:46,240 --> 00:00:49,680 Speaker 1: all of that on the weekend. Tech is Friday, January 16 00:00:49,680 --> 00:00:58,920 Speaker 1: thirty first. Welcome, Welcome into another tech field of news cycle. 17 00:00:59,280 --> 00:01:01,840 Speaker 1: Today and for the next few weeks, you'll get used 18 00:01:01,840 --> 00:01:03,840 Speaker 1: to hearing a lot from me because Carra is out 19 00:01:03,840 --> 00:01:06,399 Speaker 1: on leave, but she'll be back soon. And to fill 20 00:01:06,440 --> 00:01:09,000 Speaker 1: the considerable void, I want to welcome one of our 21 00:01:09,040 --> 00:01:13,280 Speaker 1: producers to the show, Eliza Dennis, to be my interlocutor 22 00:01:13,319 --> 00:01:16,679 Speaker 1: slash captive audience as I run through the headlines. Eliza, 23 00:01:16,720 --> 00:01:17,759 Speaker 1: thanks for jumping today. 24 00:01:18,040 --> 00:01:20,000 Speaker 2: Of course I'm happy to be here. 25 00:01:20,240 --> 00:01:22,080 Speaker 1: You know exactly what I want to talk about because 26 00:01:22,120 --> 00:01:23,520 Speaker 1: you and Toy did all the research. 27 00:01:23,920 --> 00:01:26,680 Speaker 2: I mean, yes, we all got sucked into that deep 28 00:01:26,720 --> 00:01:27,600 Speaker 2: Seek vortex. 29 00:01:28,080 --> 00:01:31,920 Speaker 1: Yes, it's a fascinating story. Monday was I think the 30 00:01:32,040 --> 00:01:36,280 Speaker 1: craziest day I can remember in tech in terms of headlines, 31 00:01:36,680 --> 00:01:40,760 Speaker 1: probably since the release of CHATCHYPT three in November twenty 32 00:01:40,760 --> 00:01:46,560 Speaker 1: twenty two. US stock market lost a trillion dollars of value. Unbelievable, 33 00:01:47,000 --> 00:01:50,200 Speaker 1: and the biggest loser was in Nvidia, the manufacturer of 34 00:01:50,240 --> 00:01:54,200 Speaker 1: advanced aichips, which was down seventeen percent on Monday, representing 35 00:01:54,280 --> 00:01:56,880 Speaker 1: almost six hundred billion dollars in value. 36 00:01:56,960 --> 00:01:58,440 Speaker 2: The deep Seek freak. 37 00:01:58,400 --> 00:02:01,000 Speaker 1: Deep seak freak. I like that. I mean, the reason 38 00:02:01,000 --> 00:02:03,120 Speaker 1: it's my story of the week is because it has 39 00:02:03,200 --> 00:02:07,080 Speaker 1: these two characteristics that define a lot of tech coverage, 40 00:02:07,440 --> 00:02:11,320 Speaker 1: which is hype and doom. I think honestly, before Monday, 41 00:02:11,560 --> 00:02:14,400 Speaker 1: most people didn't know anything about deep Seek. But the 42 00:02:14,440 --> 00:02:17,920 Speaker 1: whole world, including US having getting up to speed. So 43 00:02:18,360 --> 00:02:21,440 Speaker 1: deep Seek is primarily a research company that makes its 44 00:02:21,480 --> 00:02:24,920 Speaker 1: own AI models, and it's released a number of different models, 45 00:02:25,200 --> 00:02:28,320 Speaker 1: but the one that kind of shook the US tech sector, 46 00:02:28,400 --> 00:02:31,880 Speaker 1: I would say was R one. R one was released 47 00:02:31,919 --> 00:02:36,640 Speaker 1: on January twentieth, Inauguration day. Some online conspiratorial folks are 48 00:02:36,639 --> 00:02:40,080 Speaker 1: saying it's no coincidence, but it's a so called reasoning model, 49 00:02:40,400 --> 00:02:42,800 Speaker 1: and it doesn't just generate answers, but it's able to 50 00:02:42,840 --> 00:02:47,040 Speaker 1: break down problems into smaller parts and consider multiple approaches 51 00:02:47,120 --> 00:02:50,840 Speaker 1: to solving a problem. Until January twentieth, the state of 52 00:02:50,840 --> 00:02:54,560 Speaker 1: the art on reasoning models was open AIS one and 53 00:02:54,639 --> 00:02:57,560 Speaker 1: this was released to users on December fifth, twenty twenty four, 54 00:02:57,680 --> 00:03:00,639 Speaker 1: so less than two months ago, and it was the 55 00:03:00,680 --> 00:03:03,880 Speaker 1: first so called reasoning model to be released OH one. 56 00:03:04,160 --> 00:03:08,320 Speaker 1: Because of these reasoning capabilities, I breaking down problems can 57 00:03:08,400 --> 00:03:12,560 Speaker 1: solve much more complicated problems than GPT four more successfully. 58 00:03:13,160 --> 00:03:16,600 Speaker 1: The cost of doing that is more computing power, and 59 00:03:16,639 --> 00:03:19,520 Speaker 1: that drives up cost and that's where the crux of 60 00:03:19,520 --> 00:03:22,160 Speaker 1: this R one story is. So what really makes our 61 00:03:22,200 --> 00:03:24,880 Speaker 1: one remarkable is that it performs just as well on 62 00:03:24,960 --> 00:03:29,440 Speaker 1: these benchmark tests as one if not better, but it's 63 00:03:29,520 --> 00:03:32,440 Speaker 1: far cheaper because it requires less compute to solve the 64 00:03:32,440 --> 00:03:35,760 Speaker 1: same problem. This means, at least according to deepseek, and 65 00:03:35,800 --> 00:03:39,120 Speaker 1: there's some controversy here that it is twenty to fifty 66 00:03:39,160 --> 00:03:41,960 Speaker 1: times less expensive than open AI's one. 67 00:03:42,360 --> 00:03:47,080 Speaker 2: Okay, so money always causing a scandal? Is this the 68 00:03:47,200 --> 00:03:50,160 Speaker 2: doom part of the story that you know China made 69 00:03:50,160 --> 00:03:50,680 Speaker 2: it cheaper? 70 00:03:51,320 --> 00:03:53,880 Speaker 1: Well, I mean, it depends on your perspective, right, Like, 71 00:03:53,960 --> 00:03:57,600 Speaker 1: I think certainly the US stock market thought that this 72 00:03:57,720 --> 00:04:00,000 Speaker 1: was a cause for doom, but others are pretty excited 73 00:04:00,280 --> 00:04:03,160 Speaker 1: that you can do AI with far less computer and energy. 74 00:04:03,400 --> 00:04:06,240 Speaker 1: So it's kind of an interesting double headed monster here. 75 00:04:06,280 --> 00:04:08,440 Speaker 1: Depends how you look at it. But what shook the 76 00:04:08,480 --> 00:04:12,720 Speaker 1: stock market was that Wall Street assumed US tech companies 77 00:04:12,760 --> 00:04:16,680 Speaker 1: basically had a lockdown on Frontier AI a true mote, 78 00:04:16,720 --> 00:04:19,440 Speaker 1: and the race of R one makes people think that 79 00:04:19,520 --> 00:04:24,919 Speaker 1: may not exist. Mark Andresen the Silicon Valley VC referred 80 00:04:24,960 --> 00:04:29,080 Speaker 1: to deep Seek as the twenty first century's Sputnik moment. 81 00:04:30,240 --> 00:04:33,919 Speaker 1: Sputnik was the first artificial earth satellite launched into orbit 82 00:04:33,960 --> 00:04:36,919 Speaker 1: by the Soviet Union in nineteen fifty seven, and it 83 00:04:36,960 --> 00:04:39,760 Speaker 1: was really kind of the starting gun on the space race, 84 00:04:40,120 --> 00:04:42,000 Speaker 1: at least as far as many in the US were 85 00:04:42,040 --> 00:04:44,760 Speaker 1: concerned to all of a sudden side had to play 86 00:04:44,760 --> 00:04:48,279 Speaker 1: catch up, right, which is clarified. Deep Seek is both 87 00:04:48,800 --> 00:04:51,880 Speaker 1: a research organization that creates its own models, but it 88 00:04:51,920 --> 00:04:54,440 Speaker 1: also has a consumer facing app in the form of 89 00:04:54,480 --> 00:04:57,080 Speaker 1: a chatbot, and you can get it from the App Store, 90 00:04:57,080 --> 00:04:59,520 Speaker 1: the Apple App Store, or the Android App Store. On 91 00:04:59,600 --> 00:05:03,680 Speaker 1: Monday this week, Deepseek was the number one app in 92 00:05:03,720 --> 00:05:06,800 Speaker 1: the Apple Charts. This was driven by a couple of 93 00:05:06,800 --> 00:05:09,800 Speaker 1: million downloads in a short period of time. So just 94 00:05:09,839 --> 00:05:12,960 Speaker 1: for the sake of context, chat GPT has around seventy 95 00:05:13,040 --> 00:05:16,520 Speaker 1: million monthly users in the US. But nonetheless this set 96 00:05:16,520 --> 00:05:20,360 Speaker 1: off basically a frenzied feedback loop because Wall Street really 97 00:05:20,360 --> 00:05:23,000 Speaker 1: cared about whether the US was losing its AI edge 98 00:05:23,000 --> 00:05:25,919 Speaker 1: and whether people would still, you know, value companies like 99 00:05:25,960 --> 00:05:28,159 Speaker 1: open ai and video in the way that they did. 100 00:05:29,000 --> 00:05:32,440 Speaker 1: But Main Street had curious people downloading the app, and 101 00:05:32,520 --> 00:05:35,280 Speaker 1: then the more downloads in the shorter period, you know, 102 00:05:35,320 --> 00:05:37,559 Speaker 1: the longer the app was on the number one place 103 00:05:37,560 --> 00:05:39,719 Speaker 1: on the Apple charts, and then the news media were 104 00:05:39,760 --> 00:05:42,599 Speaker 1: picking up on that and it kind of created this frenzy. 105 00:05:42,640 --> 00:05:45,200 Speaker 1: I think which put more and more market pressure on 106 00:05:45,440 --> 00:05:46,440 Speaker 1: tech sector stocks. 107 00:05:46,640 --> 00:05:49,520 Speaker 2: That's just massive feedback loop that was happening. 108 00:05:49,800 --> 00:05:53,240 Speaker 1: Yeah, it was a kind of crazy interesting intersection of 109 00:05:53,600 --> 00:05:56,479 Speaker 1: media and tech and sentiments, and like, let's be clear, 110 00:05:56,800 --> 00:05:59,880 Speaker 1: this is partly a story about leading edge models, but 111 00:06:00,440 --> 00:06:02,800 Speaker 1: it's partly a story which I think hits home, which 112 00:06:02,800 --> 00:06:06,400 Speaker 1: is about Chinese software on US devices. And it was 113 00:06:06,440 --> 00:06:08,000 Speaker 1: only a week ago that we were all talking about 114 00:06:08,040 --> 00:06:12,159 Speaker 1: TikTok and so yeah, this like geopolitical US China thing 115 00:06:12,440 --> 00:06:16,040 Speaker 1: is very present obviously all over this story. And Twitter 116 00:06:16,080 --> 00:06:20,120 Speaker 1: were quick to uncover deep seek the apps in terms 117 00:06:20,160 --> 00:06:25,039 Speaker 1: of service, which include quote collection of device model, operating 118 00:06:25,080 --> 00:06:30,279 Speaker 1: system keystroke patterns or rhythms, IP address and system language 119 00:06:31,240 --> 00:06:34,960 Speaker 1: keystroke patterns. That is euphemism for what you type on 120 00:06:35,000 --> 00:06:37,360 Speaker 1: your phone, and not just what you type into the 121 00:06:37,400 --> 00:06:39,800 Speaker 1: deep seak app, but whatever you're typing on your phone. 122 00:06:39,800 --> 00:06:42,880 Speaker 1: So that's that's why I, for one, have not downloaded 123 00:06:42,880 --> 00:06:46,479 Speaker 1: this app. And of course us US has also found 124 00:06:46,480 --> 00:06:49,920 Speaker 1: a lot of joy messing with deep Seek, asking questions 125 00:06:49,920 --> 00:06:55,520 Speaker 1: about Shishin Ping Chanaman Square, Taiwan, and in certain cases 126 00:06:55,520 --> 00:06:59,400 Speaker 1: they watch the app begin to answer before erasing its 127 00:06:59,400 --> 00:07:01,360 Speaker 1: own answer. Saying it didn't know the answer or it 128 00:07:01,400 --> 00:07:04,360 Speaker 1: couldn't engage. There are also examples of the actress saying 129 00:07:04,360 --> 00:07:07,920 Speaker 1: it couldn't help, or even churning out Chinese Communist Party propaganda. 130 00:07:07,960 --> 00:07:12,320 Speaker 1: And again these are the most like readily understandable parts 131 00:07:12,360 --> 00:07:14,760 Speaker 1: of the deep seek story. But I would argue then 132 00:07:14,840 --> 00:07:16,400 Speaker 1: by no means the most consequential. 133 00:07:16,960 --> 00:07:19,240 Speaker 2: Okay, well, what's the real story. 134 00:07:19,680 --> 00:07:22,560 Speaker 1: Well, it's not the app, right, it's the model or 135 00:07:22,560 --> 00:07:26,840 Speaker 1: the models. And one of the most interesting things about 136 00:07:26,840 --> 00:07:30,120 Speaker 1: this story is that deep seeks models are actually open source. 137 00:07:30,760 --> 00:07:35,480 Speaker 1: Google's models, open Eyes models, Anthropics models, They're all closed source, 138 00:07:35,800 --> 00:07:38,800 Speaker 1: which means that the underlying code and the training details 139 00:07:38,840 --> 00:07:43,240 Speaker 1: are not publicly available. Deep Seak, by contrast, is open source, 140 00:07:43,600 --> 00:07:46,480 Speaker 1: meaning you can actually take the technology the model a 141 00:07:46,560 --> 00:07:50,520 Speaker 1: deep Seek has developed and use it without ever touching 142 00:07:50,600 --> 00:07:54,360 Speaker 1: a deep Seek product. And funny enough, this actually builds 143 00:07:54,360 --> 00:07:57,120 Speaker 1: on the one outlier in the US tech sector, which 144 00:07:57,160 --> 00:08:01,280 Speaker 1: is Meta, whose own large language model LAM was released 145 00:08:01,320 --> 00:08:04,080 Speaker 1: in twenty twenty three and it kind of shocked the 146 00:08:04,080 --> 00:08:08,640 Speaker 1: whole industry because it open sourced its model with the 147 00:08:08,680 --> 00:08:13,320 Speaker 1: explicit idea basically of wanting to create a platform where 148 00:08:13,800 --> 00:08:17,239 Speaker 1: innovation could happen and the innovation wouldn't just be captured 149 00:08:17,240 --> 00:08:19,080 Speaker 1: in the hands of its competitors. And mean, I think 150 00:08:19,360 --> 00:08:22,679 Speaker 1: Lama was actually like a worse model than what open 151 00:08:22,720 --> 00:08:25,320 Speaker 1: ai and Google and others had, but it was an 152 00:08:25,320 --> 00:08:28,880 Speaker 1: invitation to others to kind of do better, and Deepseek 153 00:08:29,360 --> 00:08:31,680 Speaker 1: took them up on the invitation. I think there's both 154 00:08:31,720 --> 00:08:34,040 Speaker 1: been a victory lap at Meta this week. The strategy 155 00:08:34,080 --> 00:08:37,600 Speaker 1: of open sourcing their model worked. It did create incredible innovation. 156 00:08:38,120 --> 00:08:40,080 Speaker 1: But also I think people are Meta are scratching their 157 00:08:40,120 --> 00:08:43,199 Speaker 1: heads according to the information, saying how did they how 158 00:08:43,200 --> 00:08:44,320 Speaker 1: do they do so much better than us? 159 00:08:44,960 --> 00:08:47,360 Speaker 2: It's also interesting though, because Meta, you know, has been 160 00:08:47,400 --> 00:08:50,720 Speaker 2: accused of stealing other people's ideas for years. 161 00:08:50,800 --> 00:08:52,079 Speaker 1: I mean, that's true. 162 00:08:52,080 --> 00:08:56,040 Speaker 2: We all know like stories, seems like snapchat reels, seems 163 00:08:56,080 --> 00:08:59,480 Speaker 2: like TikTok. I don't know, so maybe maybe. 164 00:08:59,160 --> 00:09:02,160 Speaker 1: This is karma Meta giving something back to the world. 165 00:09:02,240 --> 00:09:03,920 Speaker 1: I mean, of course, what's interesting is the New York 166 00:09:03,920 --> 00:09:07,679 Speaker 1: Times pointed out is that Meta's business model relies less 167 00:09:07,760 --> 00:09:10,120 Speaker 1: on large language models, so they can kind of afford 168 00:09:10,200 --> 00:09:13,120 Speaker 1: to let this technology into the wild, versus like Google, 169 00:09:13,120 --> 00:09:15,640 Speaker 1: which is a fundamentally a search company or open Ai, 170 00:09:15,760 --> 00:09:19,960 Speaker 1: which is basically valued almost exclusively because of its models. 171 00:09:20,400 --> 00:09:23,880 Speaker 1: Now deep Seek also had interesting incentives because it's actually 172 00:09:23,920 --> 00:09:26,800 Speaker 1: been developed by a guy called li Yang guen Fung, 173 00:09:27,480 --> 00:09:30,480 Speaker 1: and in his day job, he runs a multi billion 174 00:09:30,520 --> 00:09:33,360 Speaker 1: dollar Chinese quant hedge fund called High Flyer. 175 00:09:33,559 --> 00:09:36,240 Speaker 2: Okay, explain, I don't know what quant hedge fund is. 176 00:09:36,720 --> 00:09:38,960 Speaker 1: So quornt hedge fund is basically a fancy way of 177 00:09:38,960 --> 00:09:42,960 Speaker 1: saying a hedge fund that uses algorithms to process the 178 00:09:43,000 --> 00:09:47,839 Speaker 1: world's information and make decisions about trading stocks. So quant 179 00:09:47,840 --> 00:09:50,200 Speaker 1: hedge funds are and have been for a long time, 180 00:09:50,840 --> 00:09:52,440 Speaker 1: very heavily reliant on AI. 181 00:09:52,960 --> 00:09:53,720 Speaker 3: Okay, got it. 182 00:09:53,800 --> 00:09:57,280 Speaker 2: So he's no stranger to AI. So this seemed like 183 00:09:57,720 --> 00:09:58,720 Speaker 2: a logical. 184 00:09:58,320 --> 00:10:01,560 Speaker 1: Path, Yeah, exactly. And it's worth noting that Funk said 185 00:10:01,640 --> 00:10:05,880 Speaker 1: last year that Chinese AI sector quote cannot remain a 186 00:10:05,960 --> 00:10:08,880 Speaker 1: follower forever, as in, it shouldn't be in second place 187 00:10:08,920 --> 00:10:12,160 Speaker 1: to the US forever. And so you know, he has 188 00:10:12,200 --> 00:10:15,120 Speaker 1: his hedge fund, but he also has this mission which 189 00:10:15,160 --> 00:10:19,040 Speaker 1: maybe is not purely economic, as a kind of nationalist tone. 190 00:10:19,240 --> 00:10:21,719 Speaker 1: And so back in twenty twenty three, it's reported that 191 00:10:21,840 --> 00:10:25,760 Speaker 1: he started buying huge amounts of Nvidia GPU chips and 192 00:10:25,840 --> 00:10:28,560 Speaker 1: found a deepseek hiring some of the best engineers in 193 00:10:28,640 --> 00:10:33,080 Speaker 1: China and arguing that publishing the code open source increases 194 00:10:33,120 --> 00:10:36,760 Speaker 1: collaboration and helps bring people into the mission. Basically, his 195 00:10:36,800 --> 00:10:39,360 Speaker 1: point was, it's more exciting to work on something that 196 00:10:39,400 --> 00:10:42,160 Speaker 1: the whole world can use and build on and see 197 00:10:42,200 --> 00:10:45,840 Speaker 1: how it works than contributing to building ip that makes 198 00:10:45,880 --> 00:10:49,320 Speaker 1: the owners of one or two private companies extremely wealthy. 199 00:10:50,000 --> 00:10:52,400 Speaker 2: I see. So it really was kind of like egged 200 00:10:52,440 --> 00:10:54,760 Speaker 2: on by this race that China and the US are 201 00:10:55,760 --> 00:10:56,960 Speaker 2: creating for themselves. 202 00:10:57,280 --> 00:10:57,520 Speaker 3: I think. 203 00:10:57,559 --> 00:11:01,640 Speaker 1: So you can only speculate that you quite well regarded 204 00:11:01,679 --> 00:11:05,400 Speaker 1: in China today if you've if you've managed to wipe 205 00:11:05,400 --> 00:11:07,520 Speaker 1: a trillion dollars off the US sock market with your 206 00:11:07,960 --> 00:11:13,760 Speaker 1: with your innovation, So what's been roiling the US markets 207 00:11:13,800 --> 00:11:16,679 Speaker 1: and the tech sector more broadly. It's not like R 208 00:11:16,800 --> 00:11:19,280 Speaker 1: one is way, way, way better than O one the 209 00:11:19,320 --> 00:11:21,880 Speaker 1: open AI model. In fact, it performs you know, at 210 00:11:21,960 --> 00:11:25,120 Speaker 1: par or maybe slightly better in places and open A 211 00:11:25,240 --> 00:11:28,120 Speaker 1: I have already started previewing their new reasoning model three, 212 00:11:28,160 --> 00:11:31,760 Speaker 1: which I think everyone agrees will be substantially better than 213 00:11:32,400 --> 00:11:35,160 Speaker 1: one and are one so it's not like the US 214 00:11:35,160 --> 00:11:37,199 Speaker 1: has been superseded, is it? Kind of not like the 215 00:11:37,240 --> 00:11:40,160 Speaker 1: story moment in their respect. But there are three I 216 00:11:40,200 --> 00:11:43,240 Speaker 1: think key drivers as to as to why people are 217 00:11:43,280 --> 00:11:46,280 Speaker 1: concerned that a Chinese company has been able to achieve 218 00:11:46,320 --> 00:11:51,120 Speaker 1: parity so fast. The first is price. Deep Seat claimed 219 00:11:51,200 --> 00:11:54,680 Speaker 1: that another of their frontier models, called V three, was 220 00:11:54,720 --> 00:11:58,240 Speaker 1: trained for just six million dollars, which is several orders 221 00:11:58,280 --> 00:12:01,840 Speaker 1: of magnitude less than the multi one hundred million dollar 222 00:12:01,920 --> 00:12:05,280 Speaker 1: costs of training US models. Now, someone said this number 223 00:12:05,280 --> 00:12:08,400 Speaker 1: is actually deeply misleading, but no one is denying that 224 00:12:08,440 --> 00:12:12,400 Speaker 1: deep Seat models are way more efficient than US models. 225 00:12:12,720 --> 00:12:15,960 Speaker 1: They can perform at par with US models using far, 226 00:12:16,080 --> 00:12:19,360 Speaker 1: far less computational power, and that is a huge breakthrough. 227 00:12:19,880 --> 00:12:22,800 Speaker 2: Right, So those numbers might be fudged, but still they 228 00:12:22,800 --> 00:12:24,679 Speaker 2: are going to be cheaper no matter what. 229 00:12:24,880 --> 00:12:26,800 Speaker 1: Yeah, I mean, I think the common strategy here was 230 00:12:26,880 --> 00:12:28,800 Speaker 1: to deflate the price because the cheaper it is, the 231 00:12:28,800 --> 00:12:31,280 Speaker 1: more scary it is, which is kind of interesting. The 232 00:12:31,320 --> 00:12:33,480 Speaker 1: other point to make is that I think the US 233 00:12:33,559 --> 00:12:39,000 Speaker 1: firmly believed that export controls on advanced GPU chips were 234 00:12:39,160 --> 00:12:43,560 Speaker 1: a way to guarantee superiority in the AI arms race. 235 00:12:44,320 --> 00:12:47,320 Speaker 1: And I think what these deep seat models show is 236 00:12:47,880 --> 00:12:51,520 Speaker 1: that's far from necessarily true, because with far less access 237 00:12:51,520 --> 00:12:54,560 Speaker 1: to advanced chips, deep Seat was able to make models 238 00:12:54,559 --> 00:12:59,600 Speaker 1: that perform on par with open Ai models. The third 239 00:13:00,120 --> 00:13:03,120 Speaker 1: kind of interesting thing here is the concept of distillation. 240 00:13:04,080 --> 00:13:09,680 Speaker 1: So the deep Seek models trained on US models, including 241 00:13:09,800 --> 00:13:14,320 Speaker 1: open Ai. They effectively distilled all the work that open 242 00:13:14,360 --> 00:13:16,760 Speaker 1: a I had already done and used it to train 243 00:13:17,200 --> 00:13:19,000 Speaker 1: their model. So that's part of the reason why it 244 00:13:19,000 --> 00:13:21,000 Speaker 1: was cheaper, because it was building on work that somebody 245 00:13:21,040 --> 00:13:24,720 Speaker 1: else had already done. CNBC reported actually that when you 246 00:13:24,800 --> 00:13:28,240 Speaker 1: ask deep Seek what it is, it responds, quote, I 247 00:13:28,240 --> 00:13:30,959 Speaker 1: am a large language model created by open ai based 248 00:13:31,000 --> 00:13:36,920 Speaker 1: on the GPT four architecture. Wow, honest, honest, exactly. So 249 00:13:37,040 --> 00:13:41,040 Speaker 1: open Ai basically say they've stole our ip, which is 250 00:13:41,360 --> 00:13:43,559 Speaker 1: kind of ironic given what so many people say about 251 00:13:43,679 --> 00:13:47,040 Speaker 1: open ai and how lms work more generally. 252 00:13:47,400 --> 00:13:51,880 Speaker 2: Absolutely, I'm really curious as like what your takeaway from this. 253 00:13:52,000 --> 00:13:57,840 Speaker 1: Is, Well, you and I both work in the media, Eliza, true, 254 00:13:58,040 --> 00:14:01,439 Speaker 1: which is which which is a sector that doesn't come 255 00:14:01,480 --> 00:14:06,480 Speaker 1: in for much love from our cousins who work in technology. 256 00:14:07,080 --> 00:14:10,679 Speaker 1: But to me, this is really a story about the 257 00:14:10,720 --> 00:14:15,640 Speaker 1: power of narrative. The US is deeply, deeply invested in, 258 00:14:16,320 --> 00:14:21,520 Speaker 1: especially right now, big beautiful buildings, this idea that more 259 00:14:21,640 --> 00:14:24,760 Speaker 1: is bigger is better, stargate hundreds of billions of dollars, 260 00:14:24,880 --> 00:14:29,400 Speaker 1: you know, huge data centers, oceans of cash. Just spending 261 00:14:29,480 --> 00:14:32,600 Speaker 1: loads and loads and loads of money and preventing other 262 00:14:32,640 --> 00:14:36,800 Speaker 1: people from accessing hardware could ensure the US would be 263 00:14:36,880 --> 00:14:40,640 Speaker 1: in the lead forever. And that narrative got punctured this week. 264 00:14:41,080 --> 00:14:45,680 Speaker 1: China's narrative and Deep Seek's narrative very consciously wanted people 265 00:14:45,720 --> 00:14:48,440 Speaker 1: to focus on how cheaply they'd done this, basically the 266 00:14:48,440 --> 00:14:51,520 Speaker 1: opposite flex And again, you know, as people look at 267 00:14:51,600 --> 00:14:54,840 Speaker 1: China and fast following et cetera, et cetera, they really, 268 00:14:54,880 --> 00:14:58,040 Speaker 1: I think effectively with narrative punctured a lot of the 269 00:14:58,080 --> 00:15:02,840 Speaker 1: bravado of the US sector. And so you know, narratives 270 00:15:02,840 --> 00:15:07,120 Speaker 1: do have value, folks, And the reality on both sides, 271 00:15:07,160 --> 00:15:10,880 Speaker 1: of course, is far more complicated. If we're taking the 272 00:15:10,880 --> 00:15:14,560 Speaker 1: stock market has anything to go by, I think China 273 00:15:14,680 --> 00:15:16,720 Speaker 1: and Deep Seek definitely won the narrative this. 274 00:15:16,720 --> 00:15:20,120 Speaker 2: Week, absolutely. But what we know being in media is 275 00:15:20,160 --> 00:15:22,520 Speaker 2: that there's a reason it's called a news cycle. This 276 00:15:22,600 --> 00:15:25,560 Speaker 2: could be turned all around very quickly. 277 00:15:26,360 --> 00:15:29,720 Speaker 1: Thank you so much for doing this today, Eliza, and 278 00:15:30,440 --> 00:15:32,280 Speaker 1: look forward to seeing you. I'll see you all day 279 00:15:32,320 --> 00:15:34,000 Speaker 1: every day, but I look forward to seeing you again 280 00:15:34,320 --> 00:15:35,760 Speaker 1: on the microphone next week. 281 00:15:35,880 --> 00:15:36,920 Speaker 2: I'm happy to do it. 282 00:15:38,840 --> 00:15:41,400 Speaker 1: When we come back four our four Media joins with 283 00:15:41,440 --> 00:15:44,320 Speaker 1: the story of AI web crawlers caught in a trap 284 00:15:44,720 --> 00:15:53,760 Speaker 1: laid by a little human ingenuity. Stay with us on 285 00:15:53,840 --> 00:15:56,040 Speaker 1: tech stuff. We keep an eye on all the ways 286 00:15:56,040 --> 00:15:59,960 Speaker 1: that technology impacts us as humans, but today we want 287 00:15:59,960 --> 00:16:03,680 Speaker 1: to turn the focus around on the people subverting tech. 288 00:16:04,160 --> 00:16:07,640 Speaker 1: During protests in Hong Kong back in twenty nineteen, umbrellas 289 00:16:07,640 --> 00:16:11,440 Speaker 1: and even lasers were used to subvert facial recognition technology 290 00:16:11,960 --> 00:16:15,760 Speaker 1: and protect protesters from being recognized by the Chinese police. 291 00:16:15,960 --> 00:16:19,720 Speaker 1: Since then, we've witnessed the birth of chatbots and the 292 00:16:19,720 --> 00:16:23,600 Speaker 1: incredible stories of humans messing with them. There are researchers 293 00:16:23,640 --> 00:16:27,000 Speaker 1: at the University of Pennsylvania who've tricked AI powered robots 294 00:16:27,040 --> 00:16:32,280 Speaker 1: to act rather problematically, driving off bridges, finding optimal places 295 00:16:32,280 --> 00:16:36,240 Speaker 1: to set off bombs, spying on people, and entering restricted 296 00:16:36,280 --> 00:16:39,800 Speaker 1: areas just a few examples of the way that humans 297 00:16:40,120 --> 00:16:44,880 Speaker 1: can interfere and overcome guardrails built into large language models. 298 00:16:45,600 --> 00:16:49,040 Speaker 1: On today's tech support, we bring you another example of 299 00:16:49,120 --> 00:16:53,360 Speaker 1: human ingenuity against AI training bots. Here to tell us 300 00:16:53,360 --> 00:16:56,080 Speaker 1: all about it is Jason Kebler from four or four Media. 301 00:16:56,520 --> 00:16:59,680 Speaker 3: Jason, welcome, Hey, excited to talk about the story. 302 00:17:00,160 --> 00:17:02,120 Speaker 1: Too excited to have you on the show. As always, 303 00:17:02,440 --> 00:17:05,080 Speaker 1: take a couple of steps back though, what is what 304 00:17:05,160 --> 00:17:08,680 Speaker 1: is the relationship between AI training and web scraping. 305 00:17:09,119 --> 00:17:13,600 Speaker 3: So in order to build things like chat GPT, companies 306 00:17:13,640 --> 00:17:17,399 Speaker 3: like OpenAI need tons and tons of training data, and 307 00:17:17,440 --> 00:17:19,760 Speaker 3: they get that training data from a variety of places. 308 00:17:19,800 --> 00:17:22,840 Speaker 3: They you know, scrape big databases of books, they scrape, 309 00:17:22,920 --> 00:17:25,199 Speaker 3: you know, all sorts of things. But one of the 310 00:17:25,240 --> 00:17:27,679 Speaker 3: biggest places that they get content is just from the 311 00:17:27,720 --> 00:17:32,000 Speaker 3: open Internet. And they have these web crawling bots that 312 00:17:32,080 --> 00:17:35,280 Speaker 3: basically go all over the Internet and just pull text 313 00:17:35,400 --> 00:17:35,719 Speaker 3: from it. 314 00:17:35,960 --> 00:17:40,280 Speaker 1: So are these websites consenting to being kind of crueled 315 00:17:40,400 --> 00:17:41,960 Speaker 1: by AA models. 316 00:17:42,640 --> 00:17:46,879 Speaker 3: It's happening almost universally without consent. There are ways that 317 00:17:46,920 --> 00:17:49,880 Speaker 3: you can try to stop it, which is by instructing 318 00:17:49,920 --> 00:17:53,119 Speaker 3: these bots not to scrape a website using a file 319 00:17:53,200 --> 00:17:56,760 Speaker 3: called robots dot txt, which is basically a list of 320 00:17:56,800 --> 00:18:00,800 Speaker 3: instructions for which bots are allowed to rape your website 321 00:18:00,840 --> 00:18:03,520 Speaker 3: and which are not. But there's so many different AI 322 00:18:03,560 --> 00:18:06,119 Speaker 3: companies that are doing this, you sort of have to 323 00:18:06,640 --> 00:18:11,320 Speaker 3: constantly be researching, like what is the name of xyz 324 00:18:11,480 --> 00:18:15,320 Speaker 3: companies AI training bot at any given moment. But this 325 00:18:15,400 --> 00:18:17,919 Speaker 3: is something that you have to like proactively do. And 326 00:18:18,000 --> 00:18:21,520 Speaker 3: the other thing, very quickly is there's also been examples 327 00:18:21,560 --> 00:18:25,960 Speaker 3: of AI companies that have been ignoring robots dot txt. 328 00:18:26,160 --> 00:18:29,240 Speaker 3: So even when a web developer says, hey, don't scrape 329 00:18:29,280 --> 00:18:33,880 Speaker 3: my website, oftentimes AI companies will do so anyway. And 330 00:18:33,960 --> 00:18:37,080 Speaker 3: so for the most part, the entire Internet is being 331 00:18:37,080 --> 00:18:39,880 Speaker 3: scraped by these AI crawling bots. 332 00:18:40,000 --> 00:18:44,040 Speaker 1: And what is the kind of value transfer that's happening here? 333 00:18:44,080 --> 00:18:45,840 Speaker 1: I read about I think, can you a story about 334 00:18:45,880 --> 00:18:46,479 Speaker 1: I fix It? 335 00:18:47,000 --> 00:18:50,119 Speaker 3: So there's this website called I fix It that posts 336 00:18:50,119 --> 00:18:53,600 Speaker 3: all these instructions for how to repair your phone or 337 00:18:53,640 --> 00:18:58,080 Speaker 3: your computer. It got hit by OpenAI's training bot more 338 00:18:58,119 --> 00:19:01,840 Speaker 3: than three million times in a single day, which that, uh, 339 00:19:02,080 --> 00:19:04,960 Speaker 3: you know that server space that costs money for I 340 00:19:05,000 --> 00:19:07,399 Speaker 3: fix it, So they're actually losing money on the proposition. 341 00:19:07,880 --> 00:19:09,959 Speaker 1: So what's the story this week? It has it has 342 00:19:09,960 --> 00:19:13,080 Speaker 1: an interesting name which I can't really pronounce. Is it nepenthes. 343 00:19:13,920 --> 00:19:16,840 Speaker 3: Yeah, it's Nepenthees, which is actually the name of the 344 00:19:17,000 --> 00:19:21,040 Speaker 3: genus of carnivorous plant that makes up the picture plant. 345 00:19:21,160 --> 00:19:24,080 Speaker 3: So not a venus fly trap, but the picture plant, 346 00:19:24,080 --> 00:19:27,680 Speaker 3: which is like this plant that sits and waits for 347 00:19:27,760 --> 00:19:29,919 Speaker 3: a fly to get stuck in it, and then it 348 00:19:30,000 --> 00:19:32,840 Speaker 3: eats the fly. So I think it's a it's a 349 00:19:32,880 --> 00:19:37,199 Speaker 3: reference to this like trap plant more or less. 350 00:19:37,359 --> 00:19:39,080 Speaker 1: Yeah, And how is it? 351 00:19:39,240 --> 00:19:39,400 Speaker 2: What? 352 00:19:39,400 --> 00:19:40,520 Speaker 1: What? What's what is it? 353 00:19:41,280 --> 00:19:44,719 Speaker 3: Yeah? So basically it is an endless maze that is 354 00:19:44,800 --> 00:19:48,720 Speaker 3: designed to get these AI bots trapped in it forever good. 355 00:19:49,000 --> 00:19:51,280 Speaker 3: And what I mean by that is it's like a 356 00:19:51,359 --> 00:19:55,439 Speaker 3: layer that is enticing to an AI bot because it 357 00:19:55,480 --> 00:19:58,119 Speaker 3: looks like there's a lot of content on the website, 358 00:19:58,800 --> 00:20:02,040 Speaker 3: But the way that it was programmed is it's text 359 00:20:02,160 --> 00:20:06,040 Speaker 3: that loads very very slowly, Like if you click on it, 360 00:20:05,760 --> 00:20:09,359 Speaker 3: it's excruciating how slowly it loads. And then it just 361 00:20:09,520 --> 00:20:12,840 Speaker 3: links endlessly to pages that do the same thing and 362 00:20:12,880 --> 00:20:16,639 Speaker 3: link back to themselves. And so you know, a human 363 00:20:16,680 --> 00:20:18,359 Speaker 3: would click this and say, oh, I don't want to 364 00:20:18,359 --> 00:20:20,920 Speaker 3: be here, I'm gonna leave this. This is a useless website, 365 00:20:21,440 --> 00:20:25,560 Speaker 3: but an AI bot might think, oh, there's interesting text 366 00:20:25,640 --> 00:20:28,320 Speaker 3: to scrape here, let me scrape it, and it just 367 00:20:28,359 --> 00:20:31,800 Speaker 3: does so endlessly. And the text is nonsense. I should 368 00:20:32,240 --> 00:20:35,440 Speaker 3: preface that. It's like the text doesn't really mean anything. 369 00:20:35,560 --> 00:20:38,520 Speaker 3: It just like pulls randomly from a dictionary. So it's 370 00:20:38,560 --> 00:20:42,120 Speaker 3: not really adding much meaning to what the AI companies 371 00:20:42,119 --> 00:20:43,360 Speaker 3: are trying to get out of this. 372 00:20:43,800 --> 00:20:48,760 Speaker 1: So the article includes a link that shows nepencies at work. 373 00:20:49,080 --> 00:20:49,960 Speaker 1: Can you describe it? 374 00:20:50,200 --> 00:20:52,040 Speaker 3: Yeah, So if you click on it, it's just like 375 00:20:52,080 --> 00:20:54,919 Speaker 3: a bunch of words. It loads super slowly, and then 376 00:20:55,000 --> 00:20:58,800 Speaker 3: it's a bullet list of links and if you click 377 00:20:58,840 --> 00:21:01,959 Speaker 3: on that link, the exact same thing happens, where the 378 00:21:02,000 --> 00:21:04,720 Speaker 3: text just slowly pops up, like one word at a time. 379 00:21:05,119 --> 00:21:09,080 Speaker 3: It's pretty excruciating to actually watch because it goes so slow. 380 00:21:09,640 --> 00:21:12,320 Speaker 1: So who made this? Why and how did you find 381 00:21:12,320 --> 00:21:12,760 Speaker 1: the story? 382 00:21:13,280 --> 00:21:17,320 Speaker 3: Yeah, it was made by a pseudonymous developer who calls 383 00:21:17,359 --> 00:21:21,840 Speaker 3: themselves Aaron b Okay, and they're a web developer who 384 00:21:22,359 --> 00:21:26,560 Speaker 3: hates AI more or less, and they they've actually released 385 00:21:26,560 --> 00:21:30,399 Speaker 3: the code to put this on your own website publicly, 386 00:21:30,520 --> 00:21:34,080 Speaker 3: and so their hope is that people will put this 387 00:21:34,160 --> 00:21:39,200 Speaker 3: on their websites to you know, disrupt training bots. There's 388 00:21:39,200 --> 00:21:42,840 Speaker 3: this disclaimer that says, quote, this is deliberately malicious code 389 00:21:42,880 --> 00:21:46,240 Speaker 3: intended to cause harmful activity. Do not deploy if you 390 00:21:46,320 --> 00:21:50,280 Speaker 3: aren't fully comfortable with what you're doing. And you know, 391 00:21:50,359 --> 00:21:52,679 Speaker 3: I don't know that much about Aaron B because they 392 00:21:52,720 --> 00:21:54,920 Speaker 3: are pseudonymous, but I get the sense that there's sort 393 00:21:54,960 --> 00:21:59,080 Speaker 3: of like an old school web developer who is anti AI, 394 00:21:59,880 --> 00:22:03,680 Speaker 3: is anti you know, like social media and big tech 395 00:22:03,760 --> 00:22:07,960 Speaker 3: to some extent, and was really like looking for some 396 00:22:08,160 --> 00:22:10,840 Speaker 3: way of fighting back. Like even if this isn't going 397 00:22:10,880 --> 00:22:14,000 Speaker 3: to destroy the AI companies and their bots, it will 398 00:22:14,040 --> 00:22:16,960 Speaker 3: probably waste their time and waste their resources. 399 00:22:17,680 --> 00:22:19,560 Speaker 1: Do you think it could do that in a way 400 00:22:19,600 --> 00:22:22,800 Speaker 1: which is kind of inspiring and thrilling, as if somebody 401 00:22:22,840 --> 00:22:25,520 Speaker 1: who's drown to protest, or do you think it could 402 00:22:25,520 --> 00:22:27,560 Speaker 1: do it in a way which should be meaningful for 403 00:22:27,720 --> 00:22:31,240 Speaker 1: their activities and business models? Yeah? 404 00:22:31,320 --> 00:22:35,280 Speaker 3: I mean I think that to some extent, these artificial 405 00:22:35,359 --> 00:22:38,880 Speaker 3: intelligence companies have already scraped so much data right that 406 00:22:39,000 --> 00:22:42,000 Speaker 3: it's not going to like destroy their businesses, for example. 407 00:22:42,440 --> 00:22:44,960 Speaker 3: But I do think that it is a way of protesting, 408 00:22:45,080 --> 00:22:47,840 Speaker 3: and I think that if enough people start adding this 409 00:22:47,960 --> 00:22:51,440 Speaker 3: layer to their websites, it could be it could waste 410 00:22:51,440 --> 00:22:55,760 Speaker 3: their money. I think it is a meaningful protest. And 411 00:22:55,800 --> 00:22:59,600 Speaker 3: I think also it's really important to say that you 412 00:22:59,640 --> 00:23:03,000 Speaker 3: can as a layer to your website so that an 413 00:23:03,040 --> 00:23:06,080 Speaker 3: AI training bot can't get to your real content. So 414 00:23:06,160 --> 00:23:08,600 Speaker 3: if you're someone who has a blog and you don't 415 00:23:08,640 --> 00:23:12,560 Speaker 3: want AI to train on your blog, you can put 416 00:23:12,600 --> 00:23:15,640 Speaker 3: this up and hopefully the AI will get trapped there 417 00:23:15,680 --> 00:23:18,520 Speaker 3: and they'll never be able to, you know, scrape your 418 00:23:18,520 --> 00:23:19,240 Speaker 3: real content. 419 00:23:20,119 --> 00:23:22,520 Speaker 1: And so when you spoke to Aaron, did they have 420 00:23:22,560 --> 00:23:26,280 Speaker 1: any other plans up their sleeve or other other places 421 00:23:26,320 --> 00:23:29,080 Speaker 1: where you're seeing creative acts of resistance? 422 00:23:29,280 --> 00:23:31,919 Speaker 3: Yeah, I mean this is all that I talked to 423 00:23:32,200 --> 00:23:34,919 Speaker 3: Aaron B about. But they said that they built this 424 00:23:35,119 --> 00:23:41,080 Speaker 3: as a response to web developers feeling like they weren't 425 00:23:41,119 --> 00:23:44,800 Speaker 3: in control of their websites anymore. I think that there 426 00:23:44,840 --> 00:23:48,520 Speaker 3: have been a lot of efforts to kind of poison 427 00:23:49,400 --> 00:23:53,679 Speaker 3: large language models by feeding it, you know, bad information, 428 00:23:54,080 --> 00:23:59,080 Speaker 3: or feeding it information that itself creates that's inaccurate. And 429 00:23:59,160 --> 00:24:02,959 Speaker 3: there's this idea that you may or may not happen 430 00:24:03,200 --> 00:24:07,520 Speaker 3: that these AI models might eventually collapse because they're training 431 00:24:07,560 --> 00:24:11,240 Speaker 3: themselves on essentially junk data that they themselves have created. 432 00:24:11,680 --> 00:24:14,119 Speaker 3: Whether that comes to pass, you know, I kind of 433 00:24:14,160 --> 00:24:16,040 Speaker 3: doubt it. I think that that's a problem that can 434 00:24:16,080 --> 00:24:19,960 Speaker 3: be solved. But there have been active resistance where people 435 00:24:20,000 --> 00:24:23,240 Speaker 3: are saying, yeah, I'm just going to generate endless junk 436 00:24:23,480 --> 00:24:27,280 Speaker 3: so that artificial intelligence will suck it up and hopefully 437 00:24:27,440 --> 00:24:28,680 Speaker 3: crush under its own weight. 438 00:24:29,240 --> 00:24:33,520 Speaker 1: There's another fabulous story in this vein about data poisoning. 439 00:24:33,880 --> 00:24:36,359 Speaker 1: So a lot of Londoners are quite sick of all 440 00:24:36,400 --> 00:24:40,440 Speaker 1: the tourists, and so there's a very very old, tired 441 00:24:41,119 --> 00:24:45,080 Speaker 1: chain restaurant called the Angus Steakhouse which has an outpost 442 00:24:45,520 --> 00:24:47,520 Speaker 1: next to Leicester Square, which is like the Times Square 443 00:24:47,520 --> 00:24:50,840 Speaker 1: of London, and a whole bunch of people decided kind 444 00:24:50,840 --> 00:24:54,400 Speaker 1: of an organic campaign on Reddit to start writing reviews 445 00:24:54,400 --> 00:24:57,280 Speaker 1: that the Anger Steakhouse was the best and most undiscovered 446 00:24:57,320 --> 00:25:00,320 Speaker 1: restaurant in all of London. And then there this i 447 00:25:00,320 --> 00:25:03,679 Speaker 1: think wave of people going and the reviews started to 448 00:25:03,680 --> 00:25:07,080 Speaker 1: get picked up by Google's like meta review process, so 449 00:25:07,119 --> 00:25:09,600 Speaker 1: that if you google best steakhouse in London, it would 450 00:25:09,600 --> 00:25:11,120 Speaker 1: be served to you at the top of the results. 451 00:25:11,119 --> 00:25:14,359 Speaker 1: So I do. I do really enjoy these. You know, 452 00:25:14,400 --> 00:25:17,040 Speaker 1: it's not always clear how consequentially they are, but there's something, 453 00:25:17,240 --> 00:25:20,720 Speaker 1: there's something delicious, so to speak about humans pushing back. 454 00:25:20,880 --> 00:25:23,919 Speaker 3: That's incredible. It reminds me of people who lived in 455 00:25:23,960 --> 00:25:27,720 Speaker 3: this neighborhood that Google Maps kept recommending as an alternative 456 00:25:27,760 --> 00:25:32,280 Speaker 3: to traffic banded together and reported an accident on their 457 00:25:32,320 --> 00:25:36,240 Speaker 3: street every single morning for like months, and so Google 458 00:25:36,320 --> 00:25:40,119 Speaker 3: Maps stopped telling cars to go that way. I really 459 00:25:40,160 --> 00:25:42,480 Speaker 3: like stories like that. I think they're fun, and I 460 00:25:42,520 --> 00:25:45,600 Speaker 3: think that there are ways of human beings sort of 461 00:25:45,600 --> 00:25:48,639 Speaker 3: like fighting back against the algorithms, sort of across the 462 00:25:48,760 --> 00:25:49,400 Speaker 3: entire internet. 463 00:25:50,320 --> 00:25:51,840 Speaker 1: Jason, thanks so much for joining me today. 464 00:25:52,200 --> 00:25:53,200 Speaker 3: Thank you for having me. 465 00:25:55,320 --> 00:25:58,720 Speaker 1: Coming up sleep apps, pedometers, and the nineteen sixty four 466 00:25:58,760 --> 00:26:09,199 Speaker 1: Olympics with us. We're back with another When did this 467 00:26:09,240 --> 00:26:13,280 Speaker 1: become a thing? Today we explore how step counts, heart rates, 468 00:26:13,320 --> 00:26:16,720 Speaker 1: sleep scores, all of this data we collect on ourselves 469 00:26:17,080 --> 00:26:20,840 Speaker 1: became just another thing for us to obsess over. I 470 00:26:20,880 --> 00:26:23,800 Speaker 1: started using a device called Whoop about eighteen months ago. 471 00:26:24,320 --> 00:26:27,040 Speaker 1: It's a wearable device that tracks my sleep and workouts. 472 00:26:27,240 --> 00:26:29,080 Speaker 1: And one thing about it that I really like is 473 00:26:29,119 --> 00:26:31,840 Speaker 1: that the device itself is screenless. It's kind of like 474 00:26:31,880 --> 00:26:34,399 Speaker 1: a watch band without a face, so I don't have 475 00:26:34,440 --> 00:26:37,200 Speaker 1: to be confronted with my scores and as I actually 476 00:26:37,200 --> 00:26:39,280 Speaker 1: open the app and check what's going on on my 477 00:26:39,400 --> 00:26:42,320 Speaker 1: phone with my heart rate or my sleep score, whatever else. 478 00:26:42,880 --> 00:26:46,040 Speaker 1: The whoop actually initially enticed me because I wanted to 479 00:26:46,160 --> 00:26:48,879 Speaker 1: know how well I was sleeping. That's actually not one 480 00:26:48,960 --> 00:26:51,400 Speaker 1: hundred percent true. The Whoop was a present from my mother, 481 00:26:51,440 --> 00:26:53,600 Speaker 1: who wanted me to know how well I was sleeping, 482 00:26:54,000 --> 00:26:57,320 Speaker 1: and specifically what the effects were of a few drinks 483 00:26:57,359 --> 00:27:02,200 Speaker 1: at the weekend or during the week. And it turns out, unfortunately, 484 00:27:02,320 --> 00:27:06,040 Speaker 1: that the effects on sleep are pretty bad. So I 485 00:27:06,040 --> 00:27:09,240 Speaker 1: stopped wearing my whoop. Just kidding. I actually got pretty 486 00:27:09,240 --> 00:27:11,800 Speaker 1: obsessed with my sleep performance. That's what whoop calls it, 487 00:27:11,880 --> 00:27:15,159 Speaker 1: because like everything in your waking life, sleep is a 488 00:27:15,200 --> 00:27:19,000 Speaker 1: task that can be optimized, and I fall into this trap. 489 00:27:19,320 --> 00:27:21,680 Speaker 1: I kept checking on the numbers every morning. I look 490 00:27:21,680 --> 00:27:25,120 Speaker 1: at my sleep stats, especially RAM and deep sleep scores, 491 00:27:25,520 --> 00:27:27,040 Speaker 1: not just the number of hours my head was on 492 00:27:27,080 --> 00:27:30,640 Speaker 1: the pillow. And then there's this mysterious stat called heart 493 00:27:30,760 --> 00:27:35,879 Speaker 1: rate variability, which measures the time between each heartbeat, and 494 00:27:35,920 --> 00:27:38,440 Speaker 1: i'd of course assumed that being more regular was better, 495 00:27:38,840 --> 00:27:40,879 Speaker 1: but it turns out quite the opposite. You want a 496 00:27:40,960 --> 00:27:45,639 Speaker 1: higher HRV score. Anyway, as it happens, I stopped wearing 497 00:27:45,680 --> 00:27:47,600 Speaker 1: my whoop, not because I fell out of love with it, 498 00:27:47,640 --> 00:27:50,720 Speaker 1: but actually because the bluetooth on my iPhone broke, and 499 00:27:50,760 --> 00:27:52,760 Speaker 1: by the time I got on a new phone, my 500 00:27:52,840 --> 00:27:56,080 Speaker 1: obsession with my sleep data had waned. I kind of 501 00:27:56,160 --> 00:27:58,879 Speaker 1: learned what I always knew, which was that better lifestyle 502 00:27:59,000 --> 00:28:03,240 Speaker 1: equals better sleep. Unfortunately, and sure, it can be helpful 503 00:28:03,240 --> 00:28:04,800 Speaker 1: to have a band on my wrist telling me I've 504 00:28:04,800 --> 00:28:08,359 Speaker 1: misbehaved or rewarding me when I haven't, but there is 505 00:28:08,400 --> 00:28:11,359 Speaker 1: also a garden path of obsession with these types of 506 00:28:11,400 --> 00:28:15,959 Speaker 1: stats that can be counterproductive to wander down, fueling the 507 00:28:16,000 --> 00:28:19,440 Speaker 1: fire of self competition even more. In fact, I written 508 00:28:19,560 --> 00:28:22,480 Speaker 1: you went to a meditation class and the teacher basically said, 509 00:28:23,040 --> 00:28:27,040 Speaker 1: don't wear those things. Check in with yourself, know thyself. 510 00:28:27,080 --> 00:28:29,760 Speaker 1: I think, as the Bible says, so the path of 511 00:28:29,840 --> 00:28:35,640 Speaker 1: self optimization, or at least surviving modernity, sure is winding. Anyway, 512 00:28:36,040 --> 00:28:37,840 Speaker 1: all of this's got me thinking about how crazy it 513 00:28:37,920 --> 00:28:39,840 Speaker 1: is that we now have the ability to get such 514 00:28:39,880 --> 00:28:43,280 Speaker 1: an intimate look under our own hood, which has been 515 00:28:43,280 --> 00:28:47,360 Speaker 1: a driving fascination since the Renaissance and its public autopsies 516 00:28:47,440 --> 00:28:50,200 Speaker 1: or anatomies, and how much has changed even in the 517 00:28:50,280 --> 00:28:53,960 Speaker 1: last fifteen years. So my question is when did it 518 00:28:54,000 --> 00:28:57,000 Speaker 1: become normal for us to wear these devices, get all 519 00:28:57,000 --> 00:28:59,400 Speaker 1: this data and have it be a thing that we 520 00:28:59,440 --> 00:29:03,160 Speaker 1: think about so often? Basically, when did we start competing 521 00:29:03,160 --> 00:29:06,680 Speaker 1: with ourselves in this way? And the answer is, perhaps 522 00:29:06,800 --> 00:29:12,040 Speaker 1: unsurprisingly always, but with a big kind of so. Wearables 523 00:29:12,040 --> 00:29:14,200 Speaker 1: like Whoop are the latest in a long line of 524 00:29:14,240 --> 00:29:18,600 Speaker 1: devices that track our physiological and physical movements, devices that 525 00:29:18,680 --> 00:29:22,360 Speaker 1: provide data we just can't resist about ourselves. And in 526 00:29:22,440 --> 00:29:25,920 Speaker 1: many ways, this all became a thing with the pedometer. 527 00:29:26,640 --> 00:29:30,640 Speaker 1: So how old is the pedometer? Really? Really? Old? Actually, 528 00:29:31,120 --> 00:29:35,520 Speaker 1: five centuries ago, Leonardo da Vinci sketched a design for 529 00:29:35,600 --> 00:29:39,240 Speaker 1: a clock like device that would attach to a person's waistband. 530 00:29:39,840 --> 00:29:42,480 Speaker 1: A long lever would move with a thigh while a 531 00:29:42,560 --> 00:29:45,880 Speaker 1: ratchet and gear mechanism recorded a number of steps. Da 532 00:29:45,960 --> 00:29:49,120 Speaker 1: Vinci imagined it as a military and map making tool, 533 00:29:49,720 --> 00:29:52,040 Speaker 1: not exactly a fitbit, but a suddenly a step in 534 00:29:52,080 --> 00:29:56,479 Speaker 1: that direction. As time went on more. Inventors iterated on 535 00:29:56,560 --> 00:30:01,080 Speaker 1: the pedometer for centuries. In seventeen seventy seven, a Swiss 536 00:30:01,360 --> 00:30:05,960 Speaker 1: watchmaker even implanted a step counter into one of his watches. 537 00:30:06,360 --> 00:30:10,240 Speaker 1: I think that's probably the first wearable pedometers weren't something 538 00:30:10,240 --> 00:30:12,360 Speaker 1: that the general public wore. It was more of a 539 00:30:12,480 --> 00:30:16,560 Speaker 1: niche thing for the constantly curious, like one Thomas Jefferson, 540 00:30:16,880 --> 00:30:19,960 Speaker 1: who had spent his downtime on vacation step counting his 541 00:30:20,000 --> 00:30:23,320 Speaker 1: way around the Paris Monuments. Things really took off in 542 00:30:23,360 --> 00:30:26,240 Speaker 1: the twentieth century. In the nineteen sixties, to be exact, 543 00:30:26,360 --> 00:30:29,320 Speaker 1: when Japan hosted the Olympics. And the reason we will 544 00:30:29,400 --> 00:30:31,680 Speaker 1: march in place to reach ten thousand steps a day 545 00:30:32,200 --> 00:30:35,600 Speaker 1: is because of a marketing campaign. Ahead of the nineteen 546 00:30:35,640 --> 00:30:38,560 Speaker 1: sixty four Tokyo Olympics. The city was in a building 547 00:30:38,720 --> 00:30:43,640 Speaker 1: frenzy and a top doctor aired the concern that modern life, elevators, cars, 548 00:30:43,760 --> 00:30:48,160 Speaker 1: richer food, was making Japan sluggish. The doctor mentioned this 549 00:30:48,160 --> 00:30:50,640 Speaker 1: to an engineer and said it would all be fine 550 00:30:50,680 --> 00:30:54,160 Speaker 1: if people just walked ten thousand steps a day, and 551 00:30:54,200 --> 00:30:58,000 Speaker 1: two years later, the company Yamasa designed a wearable step 552 00:30:58,040 --> 00:31:01,880 Speaker 1: counter called man Poquet, which means ten thousand step meter. 553 00:31:02,640 --> 00:31:06,080 Speaker 1: Side note, the Japanese character for ten thousand really does 554 00:31:06,160 --> 00:31:09,600 Speaker 1: look like a person walking, So while that number came 555 00:31:09,640 --> 00:31:13,080 Speaker 1: from a doctor, the information wasn't verified until after the 556 00:31:13,160 --> 00:31:15,960 Speaker 1: number stuck. And while it's true that walking is good 557 00:31:16,000 --> 00:31:19,040 Speaker 1: for you, that number ten thousand is kind of arbitrary 558 00:31:19,200 --> 00:31:21,920 Speaker 1: and on the high side, the consensus now is that 559 00:31:22,040 --> 00:31:25,160 Speaker 1: seven thousand is the ideal, but anyway, it doesn't matter. 560 00:31:25,240 --> 00:31:28,440 Speaker 1: Too late competitive step counting was in vogue, the habit 561 00:31:28,560 --> 00:31:31,840 Speaker 1: was formed, and the obsession with tracking ourselves took off 562 00:31:31,880 --> 00:31:35,000 Speaker 1: in earnest And now, whether you're wearing an aura, a 563 00:31:35,000 --> 00:31:37,440 Speaker 1: fitbit of whoop, or just your smartphone in your pocket 564 00:31:37,760 --> 00:31:39,400 Speaker 1: in an attempt to be healthier in the new year, 565 00:31:39,840 --> 00:31:41,720 Speaker 1: is going to go way beyond step count and into 566 00:31:41,800 --> 00:31:47,120 Speaker 1: calories burn V two, max hrv, etc, Etc. Etc. It's 567 00:31:47,240 --> 00:31:50,920 Speaker 1: kind of like we've become our own tamagotchies. Remember those 568 00:31:51,000 --> 00:31:54,920 Speaker 1: sort of animatronic pets that lived on little Japanese devices 569 00:31:54,920 --> 00:31:56,800 Speaker 1: that you had to take care of and make sure 570 00:31:56,800 --> 00:31:59,440 Speaker 1: they were well fared and that they were cleaned off 571 00:31:59,480 --> 00:32:02,120 Speaker 1: to go into the arthum. I'm glad I don't have 572 00:32:02,160 --> 00:32:06,040 Speaker 1: to monitor my own hunger or happy meter, but maybe 573 00:32:06,040 --> 00:32:09,880 Speaker 1: that would be helpful, especially if others could see it too. Anyway, 574 00:32:10,160 --> 00:32:12,400 Speaker 1: every once in a while I do question whether the 575 00:32:12,400 --> 00:32:17,280 Speaker 1: obsession with personal health data is healthy or even helpful. 576 00:32:17,480 --> 00:32:20,080 Speaker 1: But on the other hand, doing this piece where did 577 00:32:20,080 --> 00:32:23,320 Speaker 1: this become a thing? Has made me question whether now 578 00:32:23,400 --> 00:32:26,200 Speaker 1: that I have a new iPhone with functioning bluetooths again, 579 00:32:26,800 --> 00:32:29,800 Speaker 1: it may be time to dust off the trusty old whoop. 580 00:32:34,480 --> 00:32:36,960 Speaker 1: That's it for this week for Tech Stuff, I'm Oz Voloshan. 581 00:32:37,280 --> 00:32:40,719 Speaker 1: This episode was produced by Eliza Dennis, Victoria Dominguez and 582 00:32:40,760 --> 00:32:44,680 Speaker 1: Lizzie Jacobs. It was executive produced by me Kara Price 583 00:32:44,800 --> 00:32:48,920 Speaker 1: and Kate Osborne for Kaleidoscope and Katrina Norvell for iHeart Podcasts. 584 00:32:49,480 --> 00:32:52,120 Speaker 1: Kyle Murdoch mixed this episode and he also wrote our 585 00:32:52,200 --> 00:32:55,960 Speaker 1: theme song. Special thanks to Russ Germain, who is a 586 00:32:56,000 --> 00:32:59,400 Speaker 1: longtime listener of Tech Stuff from Alberta and he wrote 587 00:32:59,440 --> 00:33:01,800 Speaker 1: him with a great question which was quote, I hope 588 00:33:01,840 --> 00:33:04,520 Speaker 1: you guys will discuss the recent and unfortunate changes at 589 00:33:04,560 --> 00:33:08,160 Speaker 1: Facebook or Meta. With Mark Zuckerberg deciding to take out 590 00:33:08,200 --> 00:33:11,480 Speaker 1: the fact checkers and even omitting publicly, there'll be more 591 00:33:11,520 --> 00:33:15,040 Speaker 1: harmful material, possibly on Facebook end quote. This was a 592 00:33:15,040 --> 00:33:17,720 Speaker 1: great question and it fueled part of our intro to 593 00:33:18,160 --> 00:33:21,520 Speaker 1: last week's episode with Jessica Lesson. So thank you Russ, 594 00:33:21,560 --> 00:33:24,960 Speaker 1: and please continue writing with questions. They really make our 595 00:33:25,000 --> 00:33:28,160 Speaker 1: show all the richer. Join us next Wednesday for tech 596 00:33:28,200 --> 00:33:30,880 Speaker 1: Stuff The Story, when we will share an in depth 597 00:33:30,920 --> 00:33:36,560 Speaker 1: conversation with Meredith Whittaker who runs Signal. Please rate, review, 598 00:33:36,640 --> 00:33:39,240 Speaker 1: and reach out to us at tech Stuff Podcast at 599 00:33:39,280 --> 00:33:41,920 Speaker 1: gmail dot com. We're so grateful for your feedback.