WEBVTT - Week In Tech: DeepSeek Freak Out 0:00:00.200 --> 0:00:02.880 Thanks for tunion to tech Stuff. If you don't recognize 0:00:02.880 --> 0:00:05.640 my voice, my name is Oz Volshan and I'm here 0:00:05.680 --> 0:00:08.880 because the inimitable Jonathan Strickland has passed the baton to 0:00:08.960 --> 0:00:12.000 Cara Price and myself to host tech Stuff. The show 0:00:12.039 --> 0:00:14.920 will remain your home for all things tech, and all 0:00:14.960 --> 0:00:18.200 the old episodes will remain available in this feed. Thanks 0:00:18.239 --> 0:00:22.239 for listening. Welcome to tech Stuff, a production of iHeart 0:00:22.239 --> 0:00:26.079 Podcasts and Kaleidoscope. I'm oz Voloshian, and today will bring 0:00:26.160 --> 0:00:29.480 you the headlines this week, including the Chinese AI company 0:00:29.560 --> 0:00:33.199 that spooked the US tech sector. On today's Tech Support segment, 0:00:33.440 --> 0:00:35.880 we'll talk to four or four medius Jason Kebler about 0:00:35.880 --> 0:00:39.760 building mazes to trap AI web crawlers, and then we've 0:00:39.800 --> 0:00:42.640 got when did this become a thing? This time we 0:00:42.680 --> 0:00:45.600 look at why everyone is so obsessed with their own data, 0:00:46.240 --> 0:00:49.680 all of that on the weekend. Tech is Friday, January 0:00:49.680 --> 0:00:58.920 thirty first. Welcome, Welcome into another tech field of news cycle. 0:00:59.280 --> 0:01:01.840 Today and for the next few weeks, you'll get used 0:01:01.840 --> 0:01:03.840 to hearing a lot from me because Carra is out 0:01:03.840 --> 0:01:06.399 on leave, but she'll be back soon. And to fill 0:01:06.440 --> 0:01:09.000 the considerable void, I want to welcome one of our 0:01:09.040 --> 0:01:13.280 producers to the show, Eliza Dennis, to be my interlocutor 0:01:13.319 --> 0:01:16.679 slash captive audience as I run through the headlines. Eliza, 0:01:16.720 --> 0:01:17.759 thanks for jumping today. 0:01:18.040 --> 0:01:20.000 Of course I'm happy to be here. 0:01:20.240 --> 0:01:22.080 You know exactly what I want to talk about because 0:01:22.120 --> 0:01:23.520 you and Toy did all the research. 0:01:23.920 --> 0:01:26.680 I mean, yes, we all got sucked into that deep 0:01:26.720 --> 0:01:27.600 Seek vortex. 0:01:28.080 --> 0:01:31.920 Yes, it's a fascinating story. Monday was I think the 0:01:32.040 --> 0:01:36.280 craziest day I can remember in tech in terms of headlines, 0:01:36.680 --> 0:01:40.760 probably since the release of CHATCHYPT three in November twenty 0:01:40.760 --> 0:01:46.560 twenty two. US stock market lost a trillion dollars of value. Unbelievable, 0:01:47.000 --> 0:01:50.200 and the biggest loser was in Nvidia, the manufacturer of 0:01:50.240 --> 0:01:54.200 advanced aichips, which was down seventeen percent on Monday, representing 0:01:54.280 --> 0:01:56.880 almost six hundred billion dollars in value. 0:01:56.960 --> 0:01:58.440 The deep Seek freak. 0:01:58.400 --> 0:02:01.000 Deep seak freak. I like that. I mean, the reason 0:02:01.000 --> 0:02:03.120 it's my story of the week is because it has 0:02:03.200 --> 0:02:07.080 these two characteristics that define a lot of tech coverage, 0:02:07.440 --> 0:02:11.320 which is hype and doom. I think honestly, before Monday, 0:02:11.560 --> 0:02:14.400 most people didn't know anything about deep Seek. But the 0:02:14.440 --> 0:02:17.920 whole world, including US having getting up to speed. So 0:02:18.360 --> 0:02:21.440 deep Seek is primarily a research company that makes its 0:02:21.480 --> 0:02:24.920 own AI models, and it's released a number of different models, 0:02:25.200 --> 0:02:28.320 but the one that kind of shook the US tech sector, 0:02:28.400 --> 0:02:31.880 I would say was R one. R one was released 0:02:31.919 --> 0:02:36.640 on January twentieth, Inauguration day. Some online conspiratorial folks are 0:02:36.639 --> 0:02:40.080 saying it's no coincidence, but it's a so called reasoning model, 0:02:40.400 --> 0:02:42.800 and it doesn't just generate answers, but it's able to 0:02:42.840 --> 0:02:47.040 break down problems into smaller parts and consider multiple approaches 0:02:47.120 --> 0:02:50.840 to solving a problem. Until January twentieth, the state of 0:02:50.840 --> 0:02:54.560 the art on reasoning models was open AIS one and 0:02:54.639 --> 0:02:57.560 this was released to users on December fifth, twenty twenty four, 0:02:57.680 --> 0:03:00.639 so less than two months ago, and it was the 0:03:00.680 --> 0:03:03.880 first so called reasoning model to be released OH one. 0:03:04.160 --> 0:03:08.320 Because of these reasoning capabilities, I breaking down problems can 0:03:08.400 --> 0:03:12.560 solve much more complicated problems than GPT four more successfully. 0:03:13.160 --> 0:03:16.600 The cost of doing that is more computing power, and 0:03:16.639 --> 0:03:19.520 that drives up cost and that's where the crux of 0:03:19.520 --> 0:03:22.160 this R one story is. So what really makes our 0:03:22.200 --> 0:03:24.880 one remarkable is that it performs just as well on 0:03:24.960 --> 0:03:29.440 these benchmark tests as one if not better, but it's 0:03:29.520 --> 0:03:32.440 far cheaper because it requires less compute to solve the 0:03:32.440 --> 0:03:35.760 same problem. This means, at least according to deepseek, and 0:03:35.800 --> 0:03:39.120 there's some controversy here that it is twenty to fifty 0:03:39.160 --> 0:03:41.960 times less expensive than open AI's one. 0:03:42.360 --> 0:03:47.080 Okay, so money always causing a scandal? Is this the 0:03:47.200 --> 0:03:50.160 doom part of the story that you know China made 0:03:50.160 --> 0:03:50.680 it cheaper? 0:03:51.320 --> 0:03:53.880 Well, I mean, it depends on your perspective, right, Like, 0:03:53.960 --> 0:03:57.600 I think certainly the US stock market thought that this 0:03:57.720 --> 0:04:00.000 was a cause for doom, but others are pretty excited 0:04:00.280 --> 0:04:03.160 that you can do AI with far less computer and energy. 0:04:03.400 --> 0:04:06.240 So it's kind of an interesting double headed monster here. 0:04:06.280 --> 0:04:08.440 Depends how you look at it. But what shook the 0:04:08.480 --> 0:04:12.720 stock market was that Wall Street assumed US tech companies 0:04:12.760 --> 0:04:16.680 basically had a lockdown on Frontier AI a true mote, 0:04:16.720 --> 0:04:19.440 and the race of R one makes people think that 0:04:19.520 --> 0:04:24.919 may not exist. Mark Andresen the Silicon Valley VC referred 0:04:24.960 --> 0:04:29.080 to deep Seek as the twenty first century's Sputnik moment. 0:04:30.240 --> 0:04:33.919 Sputnik was the first artificial earth satellite launched into orbit 0:04:33.960 --> 0:04:36.919 by the Soviet Union in nineteen fifty seven, and it 0:04:36.960 --> 0:04:39.760 was really kind of the starting gun on the space race, 0:04:40.120 --> 0:04:42.000 at least as far as many in the US were 0:04:42.040 --> 0:04:44.760 concerned to all of a sudden side had to play 0:04:44.760 --> 0:04:48.279 catch up, right, which is clarified. Deep Seek is both 0:04:48.800 --> 0:04:51.880 a research organization that creates its own models, but it 0:04:51.920 --> 0:04:54.440 also has a consumer facing app in the form of 0:04:54.480 --> 0:04:57.080 a chatbot, and you can get it from the App Store, 0:04:57.080 --> 0:04:59.520 the Apple App Store, or the Android App Store. On 0:04:59.600 --> 0:05:03.680 Monday this week, Deepseek was the number one app in 0:05:03.720 --> 0:05:06.800 the Apple Charts. This was driven by a couple of 0:05:06.800 --> 0:05:09.800 million downloads in a short period of time. So just 0:05:09.839 --> 0:05:12.960 for the sake of context, chat GPT has around seventy 0:05:13.040 --> 0:05:16.520 million monthly users in the US. But nonetheless this set 0:05:16.520 --> 0:05:20.360 off basically a frenzied feedback loop because Wall Street really 0:05:20.360 --> 0:05:23.000 cared about whether the US was losing its AI edge 0:05:23.000 --> 0:05:25.919 and whether people would still, you know, value companies like 0:05:25.960 --> 0:05:28.159 open ai and video in the way that they did. 0:05:29.000 --> 0:05:32.440 But Main Street had curious people downloading the app, and 0:05:32.520 --> 0:05:35.280 then the more downloads in the shorter period, you know, 0:05:35.320 --> 0:05:37.559 the longer the app was on the number one place 0:05:37.560 --> 0:05:39.719 on the Apple charts, and then the news media were 0:05:39.760 --> 0:05:42.599 picking up on that and it kind of created this frenzy. 0:05:42.640 --> 0:05:45.200 I think which put more and more market pressure on 0:05:45.440 --> 0:05:46.440 tech sector stocks. 0:05:46.640 --> 0:05:49.520 That's just massive feedback loop that was happening. 0:05:49.800 --> 0:05:53.240 Yeah, it was a kind of crazy interesting intersection of 0:05:53.600 --> 0:05:56.479 media and tech and sentiments, and like, let's be clear, 0:05:56.800 --> 0:05:59.880 this is partly a story about leading edge models, but 0:06:00.440 --> 0:06:02.800 it's partly a story which I think hits home, which 0:06:02.800 --> 0:06:06.400 is about Chinese software on US devices. And it was 0:06:06.440 --> 0:06:08.000 only a week ago that we were all talking about 0:06:08.040 --> 0:06:12.159 TikTok and so yeah, this like geopolitical US China thing 0:06:12.440 --> 0:06:16.040 is very present obviously all over this story. And Twitter 0:06:16.080 --> 0:06:20.120 were quick to uncover deep seek the apps in terms 0:06:20.160 --> 0:06:25.039 of service, which include quote collection of device model, operating 0:06:25.080 --> 0:06:30.279 system keystroke patterns or rhythms, IP address and system language 0:06:31.240 --> 0:06:34.960 keystroke patterns. That is euphemism for what you type on 0:06:35.000 --> 0:06:37.360 your phone, and not just what you type into the 0:06:37.400 --> 0:06:39.800 deep seak app, but whatever you're typing on your phone. 0:06:39.800 --> 0:06:42.880 So that's that's why I, for one, have not downloaded 0:06:42.880 --> 0:06:46.479 this app. And of course us US has also found 0:06:46.480 --> 0:06:49.920 a lot of joy messing with deep Seek, asking questions 0:06:49.920 --> 0:06:55.520 about Shishin Ping Chanaman Square, Taiwan, and in certain cases 0:06:55.520 --> 0:06:59.400 they watch the app begin to answer before erasing its 0:06:59.400 --> 0:07:01.360 own answer. Saying it didn't know the answer or it 0:07:01.400 --> 0:07:04.360 couldn't engage. There are also examples of the actress saying 0:07:04.360 --> 0:07:07.920 it couldn't help, or even churning out Chinese Communist Party propaganda. 0:07:07.960 --> 0:07:12.320 And again these are the most like readily understandable parts 0:07:12.360 --> 0:07:14.760 of the deep seek story. But I would argue then 0:07:14.840 --> 0:07:16.400 by no means the most consequential. 0:07:16.960 --> 0:07:19.240 Okay, well, what's the real story. 0:07:19.680 --> 0:07:22.560 Well, it's not the app, right, it's the model or 0:07:22.560 --> 0:07:26.840 the models. And one of the most interesting things about 0:07:26.840 --> 0:07:30.120 this story is that deep seeks models are actually open source. 0:07:30.760 --> 0:07:35.480 Google's models, open Eyes models, Anthropics models, They're all closed source, 0:07:35.800 --> 0:07:38.800 which means that the underlying code and the training details 0:07:38.840 --> 0:07:43.240 are not publicly available. Deep Seak, by contrast, is open source, 0:07:43.600 --> 0:07:46.480 meaning you can actually take the technology the model a 0:07:46.560 --> 0:07:50.520 deep Seek has developed and use it without ever touching 0:07:50.600 --> 0:07:54.360 a deep Seek product. And funny enough, this actually builds 0:07:54.360 --> 0:07:57.120 on the one outlier in the US tech sector, which 0:07:57.160 --> 0:08:01.280 is Meta, whose own large language model LAM was released 0:08:01.320 --> 0:08:04.080 in twenty twenty three and it kind of shocked the 0:08:04.080 --> 0:08:08.640 whole industry because it open sourced its model with the 0:08:08.680 --> 0:08:13.320 explicit idea basically of wanting to create a platform where 0:08:13.800 --> 0:08:17.239 innovation could happen and the innovation wouldn't just be captured 0:08:17.240 --> 0:08:19.080 in the hands of its competitors. And mean, I think 0:08:19.360 --> 0:08:22.679 Lama was actually like a worse model than what open 0:08:22.720 --> 0:08:25.320 ai and Google and others had, but it was an 0:08:25.320 --> 0:08:28.880 invitation to others to kind of do better, and Deepseek 0:08:29.360 --> 0:08:31.680 took them up on the invitation. I think there's both 0:08:31.720 --> 0:08:34.040 been a victory lap at Meta this week. The strategy 0:08:34.080 --> 0:08:37.600 of open sourcing their model worked. It did create incredible innovation. 0:08:38.120 --> 0:08:40.080 But also I think people are Meta are scratching their 0:08:40.120 --> 0:08:43.199 heads according to the information, saying how did they how 0:08:43.200 --> 0:08:44.320 do they do so much better than us? 0:08:44.960 --> 0:08:47.360 It's also interesting though, because Meta, you know, has been 0:08:47.400 --> 0:08:50.720 accused of stealing other people's ideas for years. 0:08:50.800 --> 0:08:52.079 I mean, that's true. 0:08:52.080 --> 0:08:56.040 We all know like stories, seems like snapchat reels, seems 0:08:56.080 --> 0:08:59.480 like TikTok. I don't know, so maybe maybe. 0:08:59.160 --> 0:09:02.160 This is karma Meta giving something back to the world. 0:09:02.240 --> 0:09:03.920 I mean, of course, what's interesting is the New York 0:09:03.920 --> 0:09:07.679 Times pointed out is that Meta's business model relies less 0:09:07.760 --> 0:09:10.120 on large language models, so they can kind of afford 0:09:10.200 --> 0:09:13.120 to let this technology into the wild, versus like Google, 0:09:13.120 --> 0:09:15.640 which is a fundamentally a search company or open Ai, 0:09:15.760 --> 0:09:19.960 which is basically valued almost exclusively because of its models. 0:09:20.400 --> 0:09:23.880 Now deep Seek also had interesting incentives because it's actually 0:09:23.920 --> 0:09:26.800 been developed by a guy called li Yang guen Fung, 0:09:27.480 --> 0:09:30.480 and in his day job, he runs a multi billion 0:09:30.520 --> 0:09:33.360 dollar Chinese quant hedge fund called High Flyer. 0:09:33.559 --> 0:09:36.240 Okay, explain, I don't know what quant hedge fund is. 0:09:36.720 --> 0:09:38.960 So quornt hedge fund is basically a fancy way of 0:09:38.960 --> 0:09:42.960 saying a hedge fund that uses algorithms to process the 0:09:43.000 --> 0:09:47.839 world's information and make decisions about trading stocks. So quant 0:09:47.840 --> 0:09:50.200 hedge funds are and have been for a long time, 0:09:50.840 --> 0:09:52.440 very heavily reliant on AI. 0:09:52.960 --> 0:09:53.720 Okay, got it. 0:09:53.800 --> 0:09:57.280 So he's no stranger to AI. So this seemed like 0:09:57.720 --> 0:09:58.720 a logical. 0:09:58.320 --> 0:10:01.560 Path, Yeah, exactly. And it's worth noting that Funk said 0:10:01.640 --> 0:10:05.880 last year that Chinese AI sector quote cannot remain a 0:10:05.960 --> 0:10:08.880 follower forever, as in, it shouldn't be in second place 0:10:08.920 --> 0:10:12.160 to the US forever. And so you know, he has 0:10:12.200 --> 0:10:15.120 his hedge fund, but he also has this mission which 0:10:15.160 --> 0:10:19.040 maybe is not purely economic, as a kind of nationalist tone. 0:10:19.240 --> 0:10:21.719 And so back in twenty twenty three, it's reported that 0:10:21.840 --> 0:10:25.760 he started buying huge amounts of Nvidia GPU chips and 0:10:25.840 --> 0:10:28.560 found a deepseek hiring some of the best engineers in 0:10:28.640 --> 0:10:33.080 China and arguing that publishing the code open source increases 0:10:33.120 --> 0:10:36.760 collaboration and helps bring people into the mission. Basically, his 0:10:36.800 --> 0:10:39.360 point was, it's more exciting to work on something that 0:10:39.400 --> 0:10:42.160 the whole world can use and build on and see 0:10:42.200 --> 0:10:45.840 how it works than contributing to building ip that makes 0:10:45.880 --> 0:10:49.320 the owners of one or two private companies extremely wealthy. 0:10:50.000 --> 0:10:52.400 I see. So it really was kind of like egged 0:10:52.440 --> 0:10:54.760 on by this race that China and the US are 0:10:55.760 --> 0:10:56.960 creating for themselves. 0:10:57.280 --> 0:10:57.520 I think. 0:10:57.559 --> 0:11:01.640 So you can only speculate that you quite well regarded 0:11:01.679 --> 0:11:05.400 in China today if you've if you've managed to wipe 0:11:05.400 --> 0:11:07.520 a trillion dollars off the US sock market with your 0:11:07.960 --> 0:11:13.760 with your innovation, So what's been roiling the US markets 0:11:13.800 --> 0:11:16.679 and the tech sector more broadly. It's not like R 0:11:16.800 --> 0:11:19.280 one is way, way, way better than O one the 0:11:19.320 --> 0:11:21.880 open AI model. In fact, it performs you know, at 0:11:21.960 --> 0:11:25.120 par or maybe slightly better in places and open A 0:11:25.240 --> 0:11:28.120 I have already started previewing their new reasoning model three, 0:11:28.160 --> 0:11:31.760 which I think everyone agrees will be substantially better than 0:11:32.400 --> 0:11:35.160 one and are one so it's not like the US 0:11:35.160 --> 0:11:37.199 has been superseded, is it? Kind of not like the 0:11:37.240 --> 0:11:40.160 story moment in their respect. But there are three I 0:11:40.200 --> 0:11:43.240 think key drivers as to as to why people are 0:11:43.280 --> 0:11:46.280 concerned that a Chinese company has been able to achieve 0:11:46.320 --> 0:11:51.120 parity so fast. The first is price. Deep Seat claimed 0:11:51.200 --> 0:11:54.680 that another of their frontier models, called V three, was 0:11:54.720 --> 0:11:58.240 trained for just six million dollars, which is several orders 0:11:58.280 --> 0:12:01.840 of magnitude less than the multi one hundred million dollar 0:12:01.920 --> 0:12:05.280 costs of training US models. Now, someone said this number 0:12:05.280 --> 0:12:08.400 is actually deeply misleading, but no one is denying that 0:12:08.440 --> 0:12:12.400 deep Seat models are way more efficient than US models. 0:12:12.720 --> 0:12:15.960 They can perform at par with US models using far, 0:12:16.080 --> 0:12:19.360 far less computational power, and that is a huge breakthrough. 0:12:19.880 --> 0:12:22.800 Right, So those numbers might be fudged, but still they 0:12:22.800 --> 0:12:24.679 are going to be cheaper no matter what. 0:12:24.880 --> 0:12:26.800 Yeah, I mean, I think the common strategy here was 0:12:26.880 --> 0:12:28.800 to deflate the price because the cheaper it is, the 0:12:28.800 --> 0:12:31.280 more scary it is, which is kind of interesting. The 0:12:31.320 --> 0:12:33.480 other point to make is that I think the US 0:12:33.559 --> 0:12:39.000 firmly believed that export controls on advanced GPU chips were 0:12:39.160 --> 0:12:43.560 a way to guarantee superiority in the AI arms race. 0:12:44.320 --> 0:12:47.320 And I think what these deep seat models show is 0:12:47.880 --> 0:12:51.520 that's far from necessarily true, because with far less access 0:12:51.520 --> 0:12:54.560 to advanced chips, deep Seat was able to make models 0:12:54.559 --> 0:12:59.600 that perform on par with open Ai models. The third 0:13:00.120 --> 0:13:03.120 kind of interesting thing here is the concept of distillation. 0:13:04.080 --> 0:13:09.680 So the deep Seek models trained on US models, including 0:13:09.800 --> 0:13:14.320 open Ai. They effectively distilled all the work that open 0:13:14.360 --> 0:13:16.760 a I had already done and used it to train 0:13:17.200 --> 0:13:19.000 their model. So that's part of the reason why it 0:13:19.000 --> 0:13:21.000 was cheaper, because it was building on work that somebody 0:13:21.040 --> 0:13:24.720 else had already done. CNBC reported actually that when you 0:13:24.800 --> 0:13:28.240 ask deep Seek what it is, it responds, quote, I 0:13:28.240 --> 0:13:30.959 am a large language model created by open ai based 0:13:31.000 --> 0:13:36.920 on the GPT four architecture. Wow, honest, honest, exactly. So 0:13:37.040 --> 0:13:41.040 open Ai basically say they've stole our ip, which is 0:13:41.360 --> 0:13:43.559 kind of ironic given what so many people say about 0:13:43.679 --> 0:13:47.040 open ai and how lms work more generally. 0:13:47.400 --> 0:13:51.880 Absolutely, I'm really curious as like what your takeaway from this. 0:13:52.000 --> 0:13:57.840 Is, Well, you and I both work in the media, Eliza, true, 0:13:58.040 --> 0:14:01.439 which is which which is a sector that doesn't come 0:14:01.480 --> 0:14:06.480 in for much love from our cousins who work in technology. 0:14:07.080 --> 0:14:10.679 But to me, this is really a story about the 0:14:10.720 --> 0:14:15.640 power of narrative. The US is deeply, deeply invested in, 0:14:16.320 --> 0:14:21.520 especially right now, big beautiful buildings, this idea that more 0:14:21.640 --> 0:14:24.760 is bigger is better, stargate hundreds of billions of dollars, 0:14:24.880 --> 0:14:29.400 you know, huge data centers, oceans of cash. Just spending 0:14:29.480 --> 0:14:32.600 loads and loads and loads of money and preventing other 0:14:32.640 --> 0:14:36.800 people from accessing hardware could ensure the US would be 0:14:36.880 --> 0:14:40.640 in the lead forever. And that narrative got punctured this week. 0:14:41.080 --> 0:14:45.680 China's narrative and Deep Seek's narrative very consciously wanted people 0:14:45.720 --> 0:14:48.440 to focus on how cheaply they'd done this, basically the 0:14:48.440 --> 0:14:51.520 opposite flex And again, you know, as people look at 0:14:51.600 --> 0:14:54.840 China and fast following et cetera, et cetera, they really, 0:14:54.880 --> 0:14:58.040 I think effectively with narrative punctured a lot of the 0:14:58.080 --> 0:15:02.840 bravado of the US sector. And so you know, narratives 0:15:02.840 --> 0:15:07.120 do have value, folks, And the reality on both sides, 0:15:07.160 --> 0:15:10.880 of course, is far more complicated. If we're taking the 0:15:10.880 --> 0:15:14.560 stock market has anything to go by, I think China 0:15:14.680 --> 0:15:16.720 and Deep Seek definitely won the narrative this. 0:15:16.720 --> 0:15:20.120 Week, absolutely. But what we know being in media is 0:15:20.160 --> 0:15:22.520 that there's a reason it's called a news cycle. This 0:15:22.600 --> 0:15:25.560 could be turned all around very quickly. 0:15:26.360 --> 0:15:29.720 Thank you so much for doing this today, Eliza, and 0:15:30.440 --> 0:15:32.280 look forward to seeing you. I'll see you all day 0:15:32.320 --> 0:15:34.000 every day, but I look forward to seeing you again 0:15:34.320 --> 0:15:35.760 on the microphone next week. 0:15:35.880 --> 0:15:36.920 I'm happy to do it. 0:15:38.840 --> 0:15:41.400 When we come back four our four Media joins with 0:15:41.440 --> 0:15:44.320 the story of AI web crawlers caught in a trap 0:15:44.720 --> 0:15:53.760 laid by a little human ingenuity. Stay with us on 0:15:53.840 --> 0:15:56.040 tech stuff. We keep an eye on all the ways 0:15:56.040 --> 0:15:59.960 that technology impacts us as humans, but today we want 0:15:59.960 --> 0:16:03.680 to turn the focus around on the people subverting tech. 0:16:04.160 --> 0:16:07.640 During protests in Hong Kong back in twenty nineteen, umbrellas 0:16:07.640 --> 0:16:11.440 and even lasers were used to subvert facial recognition technology 0:16:11.960 --> 0:16:15.760 and protect protesters from being recognized by the Chinese police. 0:16:15.960 --> 0:16:19.720 Since then, we've witnessed the birth of chatbots and the 0:16:19.720 --> 0:16:23.600 incredible stories of humans messing with them. There are researchers 0:16:23.640 --> 0:16:27.000 at the University of Pennsylvania who've tricked AI powered robots 0:16:27.040 --> 0:16:32.280 to act rather problematically, driving off bridges, finding optimal places 0:16:32.280 --> 0:16:36.240 to set off bombs, spying on people, and entering restricted 0:16:36.280 --> 0:16:39.800 areas just a few examples of the way that humans 0:16:40.120 --> 0:16:44.880 can interfere and overcome guardrails built into large language models. 0:16:45.600 --> 0:16:49.040 On today's tech support, we bring you another example of 0:16:49.120 --> 0:16:53.360 human ingenuity against AI training bots. Here to tell us 0:16:53.360 --> 0:16:56.080 all about it is Jason Kebler from four or four Media. 0:16:56.520 --> 0:16:59.680 Jason, welcome, Hey, excited to talk about the story. 0:17:00.160 --> 0:17:02.120 Too excited to have you on the show. As always, 0:17:02.440 --> 0:17:05.080 take a couple of steps back though, what is what 0:17:05.160 --> 0:17:08.680 is the relationship between AI training and web scraping. 0:17:09.119 --> 0:17:13.600 So in order to build things like chat GPT, companies 0:17:13.640 --> 0:17:17.399 like OpenAI need tons and tons of training data, and 0:17:17.440 --> 0:17:19.760 they get that training data from a variety of places. 0:17:19.800 --> 0:17:22.840 They you know, scrape big databases of books, they scrape, 0:17:22.920 --> 0:17:25.199 you know, all sorts of things. But one of the 0:17:25.240 --> 0:17:27.679 biggest places that they get content is just from the 0:17:27.720 --> 0:17:32.000 open Internet. And they have these web crawling bots that 0:17:32.080 --> 0:17:35.280 basically go all over the Internet and just pull text 0:17:35.400 --> 0:17:35.719 from it. 0:17:35.960 --> 0:17:40.280 So are these websites consenting to being kind of crueled 0:17:40.400 --> 0:17:41.960 by AA models. 0:17:42.640 --> 0:17:46.879 It's happening almost universally without consent. There are ways that 0:17:46.920 --> 0:17:49.880 you can try to stop it, which is by instructing 0:17:49.920 --> 0:17:53.119 these bots not to scrape a website using a file 0:17:53.200 --> 0:17:56.760 called robots dot txt, which is basically a list of 0:17:56.800 --> 0:18:00.800 instructions for which bots are allowed to rape your website 0:18:00.840 --> 0:18:03.520 and which are not. But there's so many different AI 0:18:03.560 --> 0:18:06.119 companies that are doing this, you sort of have to 0:18:06.640 --> 0:18:11.320 constantly be researching, like what is the name of xyz 0:18:11.480 --> 0:18:15.320 companies AI training bot at any given moment. But this 0:18:15.400 --> 0:18:17.919 is something that you have to like proactively do. And 0:18:18.000 --> 0:18:21.520 the other thing, very quickly is there's also been examples 0:18:21.560 --> 0:18:25.960 of AI companies that have been ignoring robots dot txt. 0:18:26.160 --> 0:18:29.240 So even when a web developer says, hey, don't scrape 0:18:29.280 --> 0:18:33.880 my website, oftentimes AI companies will do so anyway. And 0:18:33.960 --> 0:18:37.080 so for the most part, the entire Internet is being 0:18:37.080 --> 0:18:39.880 scraped by these AI crawling bots. 0:18:40.000 --> 0:18:44.040 And what is the kind of value transfer that's happening here? 0:18:44.080 --> 0:18:45.840 I read about I think, can you a story about 0:18:45.880 --> 0:18:46.479 I fix It? 0:18:47.000 --> 0:18:50.119 So there's this website called I fix It that posts 0:18:50.119 --> 0:18:53.600 all these instructions for how to repair your phone or 0:18:53.640 --> 0:18:58.080 your computer. It got hit by OpenAI's training bot more 0:18:58.119 --> 0:19:01.840 than three million times in a single day, which that, uh, 0:19:02.080 --> 0:19:04.960 you know that server space that costs money for I 0:19:05.000 --> 0:19:07.399 fix it, So they're actually losing money on the proposition. 0:19:07.880 --> 0:19:09.959 So what's the story this week? It has it has 0:19:09.960 --> 0:19:13.080 an interesting name which I can't really pronounce. Is it nepenthes. 0:19:13.920 --> 0:19:16.840 Yeah, it's Nepenthees, which is actually the name of the 0:19:17.000 --> 0:19:21.040 genus of carnivorous plant that makes up the picture plant. 0:19:21.160 --> 0:19:24.080 So not a venus fly trap, but the picture plant, 0:19:24.080 --> 0:19:27.680 which is like this plant that sits and waits for 0:19:27.760 --> 0:19:29.919 a fly to get stuck in it, and then it 0:19:30.000 --> 0:19:32.840 eats the fly. So I think it's a it's a 0:19:32.880 --> 0:19:37.199 reference to this like trap plant more or less. 0:19:37.359 --> 0:19:39.080 Yeah, And how is it? 0:19:39.240 --> 0:19:39.400 What? 0:19:39.400 --> 0:19:40.520 What? What's what is it? 0:19:41.280 --> 0:19:44.719 Yeah? So basically it is an endless maze that is 0:19:44.800 --> 0:19:48.720 designed to get these AI bots trapped in it forever good. 0:19:49.000 --> 0:19:51.280 And what I mean by that is it's like a 0:19:51.359 --> 0:19:55.439 layer that is enticing to an AI bot because it 0:19:55.480 --> 0:19:58.119 looks like there's a lot of content on the website, 0:19:58.800 --> 0:20:02.040 But the way that it was programmed is it's text 0:20:02.160 --> 0:20:06.040 that loads very very slowly, Like if you click on it, 0:20:05.760 --> 0:20:09.359 it's excruciating how slowly it loads. And then it just 0:20:09.520 --> 0:20:12.840 links endlessly to pages that do the same thing and 0:20:12.880 --> 0:20:16.639 link back to themselves. And so you know, a human 0:20:16.680 --> 0:20:18.359 would click this and say, oh, I don't want to 0:20:18.359 --> 0:20:20.920 be here, I'm gonna leave this. This is a useless website, 0:20:21.440 --> 0:20:25.560 but an AI bot might think, oh, there's interesting text 0:20:25.640 --> 0:20:28.320 to scrape here, let me scrape it, and it just 0:20:28.359 --> 0:20:31.800 does so endlessly. And the text is nonsense. I should 0:20:32.240 --> 0:20:35.440 preface that. It's like the text doesn't really mean anything. 0:20:35.560 --> 0:20:38.520 It just like pulls randomly from a dictionary. So it's 0:20:38.560 --> 0:20:42.120 not really adding much meaning to what the AI companies 0:20:42.119 --> 0:20:43.360 are trying to get out of this. 0:20:43.800 --> 0:20:48.760 So the article includes a link that shows nepencies at work. 0:20:49.080 --> 0:20:49.960 Can you describe it? 0:20:50.200 --> 0:20:52.040 Yeah, So if you click on it, it's just like 0:20:52.080 --> 0:20:54.919 a bunch of words. It loads super slowly, and then 0:20:55.000 --> 0:20:58.800 it's a bullet list of links and if you click 0:20:58.840 --> 0:21:01.959 on that link, the exact same thing happens, where the 0:21:02.000 --> 0:21:04.720 text just slowly pops up, like one word at a time. 0:21:05.119 --> 0:21:09.080 It's pretty excruciating to actually watch because it goes so slow. 0:21:09.640 --> 0:21:12.320 So who made this? Why and how did you find 0:21:12.320 --> 0:21:12.760 the story? 0:21:13.280 --> 0:21:17.320 Yeah, it was made by a pseudonymous developer who calls 0:21:17.359 --> 0:21:21.840 themselves Aaron b Okay, and they're a web developer who 0:21:22.359 --> 0:21:26.560 hates AI more or less, and they they've actually released 0:21:26.560 --> 0:21:30.399 the code to put this on your own website publicly, 0:21:30.520 --> 0:21:34.080 and so their hope is that people will put this 0:21:34.160 --> 0:21:39.200 on their websites to you know, disrupt training bots. There's 0:21:39.200 --> 0:21:42.840 this disclaimer that says, quote, this is deliberately malicious code 0:21:42.880 --> 0:21:46.240 intended to cause harmful activity. Do not deploy if you 0:21:46.320 --> 0:21:50.280 aren't fully comfortable with what you're doing. And you know, 0:21:50.359 --> 0:21:52.679 I don't know that much about Aaron B because they 0:21:52.720 --> 0:21:54.920 are pseudonymous, but I get the sense that there's sort 0:21:54.960 --> 0:21:59.080 of like an old school web developer who is anti AI, 0:21:59.880 --> 0:22:03.680 is anti you know, like social media and big tech 0:22:03.760 --> 0:22:07.960 to some extent, and was really like looking for some 0:22:08.160 --> 0:22:10.840 way of fighting back. Like even if this isn't going 0:22:10.880 --> 0:22:14.000 to destroy the AI companies and their bots, it will 0:22:14.040 --> 0:22:16.960 probably waste their time and waste their resources. 0:22:17.680 --> 0:22:19.560 Do you think it could do that in a way 0:22:19.600 --> 0:22:22.800 which is kind of inspiring and thrilling, as if somebody 0:22:22.840 --> 0:22:25.520