1 00:00:00,400 --> 00:00:03,080 I'm not entirely sure what's been going on, but if you've 2 00:00:03,080 --> 00:00:06,360 been following the story of Elon Musk's push to get a trillion 3 00:00:06,360 --> 00:00:10,440 dollar paid out of Tesla, it seems entirely possible to me 4 00:00:10,640 --> 00:00:13,840 that Musk has just been coming up with side projects like his 5 00:00:13,840 --> 00:00:17,520 new online encyclopedia Gracopedia, so that he could 6 00:00:17,520 --> 00:00:21,600 convince Tesla shareholders to pay him a huge sum of money. 7 00:00:21,840 --> 00:00:24,680 Otherwise he'll abandoned them in pursuit of these other 8 00:00:24,680 --> 00:00:27,080 interests. And it's believable too. 9 00:00:27,280 --> 00:00:30,400 A few years ago, he bought Twitter so that he could become 10 00:00:30,440 --> 00:00:34,880 a chat for moderator. Yesterday, Tesla shareholders 11 00:00:34,880 --> 00:00:38,840 caved in and approved a pay package that would make Musk the 12 00:00:38,840 --> 00:00:42,920 world's first trillionaire and grant him control of 1/4 of the 13 00:00:42,920 --> 00:00:46,760 company's shares if he hits a series of ambitious targets. 14 00:00:46,960 --> 00:00:50,520 So maybe this whole Grokopedia thing is already over. 15 00:00:50,880 --> 00:00:55,120 If you haven't heard about it, the Grok in Grokopedia refers to 16 00:00:55,120 --> 00:00:58,640 Elon Musk's generative AI chatbot, which is featured 17 00:00:58,640 --> 00:01:01,920 prominently on his social network or everything app 18 00:01:01,960 --> 00:01:06,680 Twitter, which he calls Eggs. With Grokopedia, Musk's stated 19 00:01:06,680 --> 00:01:10,600 goal is to create an open source, comprehensive collection 20 00:01:10,600 --> 00:01:14,480 of all knowledge, then place copies of that etched in a 21 00:01:14,480 --> 00:01:18,760 stable oxide, whatever that means, in orbit around the Moon 22 00:01:18,760 --> 00:01:21,280 and Mars to preserve it for the future. 23 00:01:21,640 --> 00:01:25,520 Now, I checked with Grok as to what he means by a stable oxide, 24 00:01:25,680 --> 00:01:28,320 and Grok said that he probably means Glass. 25 00:01:29,040 --> 00:01:33,320 Anyhow, Musk says that Grocopedia will be a compendium 26 00:01:33,320 --> 00:01:36,680 of the truth, the whole truth, and nothing but the truth. 27 00:01:36,920 --> 00:01:40,520 A rather lofty promise from someone who not so long ago 28 00:01:40,520 --> 00:01:43,960 offered Wikipedia a billion dollars to rename itself 29 00:01:44,120 --> 00:01:48,040 Dickipedia. Elon Musk has framed the project 30 00:01:48,200 --> 00:01:52,000 as a purge of propaganda, a replacement for what he calls 31 00:01:52,000 --> 00:01:56,480 legacy media, and a step towards building a more open source 32 00:01:56,480 --> 00:02:00,320 repository of knowledge. But given Grokopedia's reliance 33 00:02:00,320 --> 00:02:04,520 on his chat bot Grok, the result may be less Encyclopedia 34 00:02:04,520 --> 00:02:09,080 Britannica and more Reddit meets Twitter trolls, as his chat bot 35 00:02:09,240 --> 00:02:14,240 does train on sources like that. According to XAI, Grok will be 36 00:02:14,240 --> 00:02:18,160 responsible for all fact checking on Grokopedia, which is 37 00:02:18,160 --> 00:02:21,440 a bit like asking a parrot to verify a Shakespeare quote. 38 00:02:21,640 --> 00:02:24,760 It might sound convincing, but you wouldn't want to bet your 39 00:02:24,760 --> 00:02:29,800 reputation on its accuracy. Wikipedia, in Musk's view, has 40 00:02:29,800 --> 00:02:33,000 gone soft. He used to like it, but now it's 41 00:02:33,000 --> 00:02:36,840 too woke. 2 establishment and two, unwilling to include the 42 00:02:36,840 --> 00:02:39,760 kind of sources that flatter his worldview. 43 00:02:40,000 --> 00:02:43,880 He was particularly irked earlier this year when Wikipedia 44 00:02:43,880 --> 00:02:47,920 included a photo of him saluting a Trump's inauguration. 45 00:02:48,200 --> 00:02:50,520 Like how do you point at the crowd? 46 00:02:51,240 --> 00:02:54,960 How do you wave the entry node at the controversy and included 47 00:02:54,960 --> 00:02:57,360 his denial? But that wasn't enough. 48 00:02:57,640 --> 00:03:01,960 Musk instead wants a new kind of online encyclopedia when where 49 00:03:01,960 --> 00:03:05,840 his AI chat bot does the fact checking and where inconvenient 50 00:03:05,840 --> 00:03:09,800 truths can be recalibrated until they feel more groggy. 51 00:03:10,080 --> 00:03:13,960 It's not obvious that a chat bot can be trusted for fact 52 00:03:13,960 --> 00:03:17,960 checking, Andrew Dudfield of Full Fact AUK based fact 53 00:03:17,960 --> 00:03:21,600 checking organization was quoted in the Guardian as saying. 54 00:03:21,760 --> 00:03:25,080 We really have to consider whether an AI generated 55 00:03:25,080 --> 00:03:29,640 encyclopedia, a fact simile of of reality run through a filter, 56 00:03:29,720 --> 00:03:33,360 is a better proposition than any of the previous things that we 57 00:03:33,360 --> 00:03:36,160 have. It doesn't display the same 58 00:03:36,160 --> 00:03:39,440 transparency, but it's asking for the same trust. 59 00:03:39,560 --> 00:03:43,680 It's not clear how far the human hand is involved, how far it's 60 00:03:43,840 --> 00:03:47,680 AI generated, and what content the AI was trained on. 61 00:03:47,840 --> 00:03:51,320 It's hard to place trust in something when you can't see how 62 00:03:51,320 --> 00:03:54,840 those choices are made. Whether you like Wikipedia or 63 00:03:54,840 --> 00:03:58,720 not, it's editorial model is built on transparency and 64 00:03:58,720 --> 00:04:01,480 consensus. Every article has a visible 65 00:04:01,480 --> 00:04:05,880 history, a complete record of edits, debates, reversions, and 66 00:04:05,880 --> 00:04:09,880 compromises that were made. As the article slowly formed, 67 00:04:10,240 --> 00:04:13,520 you can see who changed what, when, and why. 68 00:04:13,760 --> 00:04:17,680 Disputes are hashed out in public, often tediously, but 69 00:04:17,680 --> 00:04:21,800 with a kind of democratic rigor. The site prohibits original 70 00:04:21,800 --> 00:04:25,480 research, insisting instead on citations from reputable 71 00:04:25,480 --> 00:04:30,320 sources, which can be both a pro and a con, as while it requires 72 00:04:30,320 --> 00:04:34,280 high quality sources, this does mean that it will reflect the 73 00:04:34,280 --> 00:04:37,840 biases of academia, big media and other respected 74 00:04:37,840 --> 00:04:42,360 institutions, but at least those biases are visible and 75 00:04:42,360 --> 00:04:45,760 traceable. Grocopaedia, on the other hand, 76 00:04:45,920 --> 00:04:49,760 offers no transparency other than providing sources like 77 00:04:49,760 --> 00:04:52,960 Wikipedia does. But as you'll see later on, the 78 00:04:52,960 --> 00:04:56,240 sourcing on Grocopaedia has its own problems. 79 00:04:56,800 --> 00:05:00,520 Articles on Grokopedia are generated and fact checked by a 80 00:05:00,520 --> 00:05:04,080 large language model whose internal workings are entirely 81 00:05:04,080 --> 00:05:08,600 opaque even to its creators, and then its outputs can be subject 82 00:05:08,600 --> 00:05:12,480 to Elon Musk's personal recalibration, in particular if 83 00:05:12,480 --> 00:05:14,960 the article is on a topic he cares about. 84 00:05:15,200 --> 00:05:19,280 There's no edit history on Grokopedia, no talk pages, no 85 00:05:19,280 --> 00:05:23,560 visible decision making process, meaning that it's not clear who 86 00:05:23,560 --> 00:05:26,720 or what decides on how the final article is formed. 87 00:05:27,240 --> 00:05:31,200 When Elon Musk announced Grokopedia earlier this week, he 88 00:05:31,200 --> 00:05:35,080 said that it was better than Wikipedia, but surprisingly, for 89 00:05:35,080 --> 00:05:38,800 all its ambition, it appears to lean very heavily on the very 90 00:05:38,800 --> 00:05:42,640 website it aims to replace. As the Register put it, if you 91 00:05:42,640 --> 00:05:45,800 scratch Grokopedia, it bleeds Wikipedia. 92 00:05:46,360 --> 00:05:50,240 I can't find any high quality data online, but the vast 93 00:05:50,240 --> 00:05:53,840 majority of the articles that I looked up on Grocopedia were 94 00:05:53,840 --> 00:05:57,800 obviously based on their equivalent Wikipedia pages, but 95 00:05:57,920 --> 00:06:00,880 with fewer citations, and at the bottom they stayed. 96 00:06:00,960 --> 00:06:04,560 This content is adapted from Wikipedia, licensed under 97 00:06:04,560 --> 00:06:08,000 Creative Commons Attribution 4 Point O license. 98 00:06:08,480 --> 00:06:12,280 The idea of collecting all human knowledge in one place is 99 00:06:12,280 --> 00:06:14,840 nothing new. Long before Grok began 100 00:06:14,840 --> 00:06:18,880 hallucinating facts into existence, encyclopedias were a 101 00:06:18,880 --> 00:06:22,840 much more analog affair. In the 1st century, the Roman 102 00:06:22,840 --> 00:06:27,880 author Pliny the Elder compiled a 37 volume work which is 103 00:06:27,880 --> 00:06:31,560 usually cited as the first encyclopedia in the Western 104 00:06:31,560 --> 00:06:34,760 tradition. Over a Millennium later, China's 105 00:06:34,760 --> 00:06:39,200 Yongle Encyclopedia was compiled by over 2000 scholars. 106 00:06:39,440 --> 00:06:42,720 It remained the largest encyclopedia in the world for 107 00:06:42,720 --> 00:06:47,240 1000 years. In 1768, the first Encyclopedia 108 00:06:47,240 --> 00:06:51,280 Britannica was published. And if we Fast forward to 2001, 109 00:06:51,440 --> 00:06:54,240 Wikipedia upended the model entirely. 110 00:06:54,560 --> 00:06:58,800 With Wikipedia, encyclopedias were no longer the domain of 111 00:06:58,800 --> 00:07:01,960 scholars and scribes. The encyclopedia became a 112 00:07:01,960 --> 00:07:05,440 living, breathing document that anyone with an Internet 113 00:07:05,440 --> 00:07:09,520 connection and a strong opinion could edit Over for 24 years. 114 00:07:09,520 --> 00:07:13,080 It's grown into one of the most visited websites in the world 115 00:07:13,240 --> 00:07:17,600 and is a sprawling, imperfect, but astonishingly comprehensive 116 00:07:17,600 --> 00:07:20,800 record of human knowledge and human argument. 117 00:07:21,280 --> 00:07:25,960 Elon Musk was once even a fan. On Wikipedia's 20th birthday in 118 00:07:25,960 --> 00:07:29,640 2021, he tweeted, so glad you exist. 119 00:07:29,800 --> 00:07:33,000 But just two years later, the honeymoon was over. 120 00:07:33,200 --> 00:07:37,400 Musk accused the side of being hijacked by far left activists. 121 00:07:37,520 --> 00:07:40,520 He no longer trusted the gatekeepers of digital 122 00:07:40,520 --> 00:07:45,280 knowledge, and so earlier this month launched Grokapedia, his 123 00:07:45,360 --> 00:07:50,000 AI based platform that promises to fix Wikipedia's flaws by 124 00:07:50,000 --> 00:07:54,960 replacing its editors with an AI that he personally supervises. 125 00:07:55,600 --> 00:07:59,760 The website did crash on its launch day, but probably because 126 00:07:59,760 --> 00:08:03,120 so many people visited it, there's no reason to believe 127 00:08:03,320 --> 00:08:05,280 that Full Self Driving was to blame. 128 00:08:06,240 --> 00:08:10,240 At its core, Grokapedia isn't just a tech experiment. 129 00:08:10,400 --> 00:08:14,920 It's a front in a much older war, the battle over who gets to 130 00:08:14,920 --> 00:08:17,520 shape the record and define reality. 131 00:08:17,760 --> 00:08:21,360 Musk has framed it as a corrective to what he sees as 132 00:08:21,360 --> 00:08:25,880 the ideological capture of Wikipedia by far left activists 133 00:08:25,880 --> 00:08:29,560 and legacy media. In his telling, the problem 134 00:08:29,560 --> 00:08:33,120 isn't just that Wikipedia is wrong, it's that it's wrong in a 135 00:08:33,120 --> 00:08:36,159 predictable and politically motivated way. 136 00:08:36,640 --> 00:08:40,360 I should note that the desire to control the narrative isn't 137 00:08:40,360 --> 00:08:43,039 unique to mask. It can also be seen in the 138 00:08:43,039 --> 00:08:46,920 structure of Wikipedia, where a small group of volunteer editors 139 00:08:47,000 --> 00:08:50,160 have a massive influence over what counts as neutral 140 00:08:50,160 --> 00:08:53,720 knowledge. Grokopedia, however, for all its 141 00:08:53,720 --> 00:08:58,520 talk of openness, replaces this messy, visible process with a 142 00:08:58,520 --> 00:09:03,080 chatbot trained on a mystery mix of data and ideology. 143 00:09:03,280 --> 00:09:07,400 Without the talk pages, the edit histories, and the debates, you 144 00:09:07,400 --> 00:09:12,160 just get finished articles, and Musk has previously admitted to 145 00:09:12,160 --> 00:09:16,480 personally intervening in Grok's AI outputs when he doesn't like 146 00:09:16,480 --> 00:09:19,880 what it says. So with Grokopedia, you get what 147 00:09:19,880 --> 00:09:23,640 looks like a cleaner process, but it's by no means a more 148 00:09:23,640 --> 00:09:27,800 trustworthy 1. Bias isn't just a flaw that can 149 00:09:27,800 --> 00:09:29,920 be patched out of knowledge systems. 150 00:09:30,120 --> 00:09:34,960 Bias is a persistent byproduct of how humans and now machines 151 00:09:35,200 --> 00:09:38,720 process the world. Whether it's a lone writer, a 152 00:09:38,720 --> 00:09:43,040 crowd of Wikipedia editors, or a large language model trained on 153 00:09:43,040 --> 00:09:47,520 the Internet, every attempt to organize information reflects 154 00:09:47,520 --> 00:09:51,200 the assumptions, priorities, and blind spots of the project 155 00:09:51,200 --> 00:09:54,520 leaders. Wikipedia has acknowledged this 156 00:09:54,520 --> 00:09:57,800 from the start. It relies on secondary sources 157 00:09:57,920 --> 00:10:00,360 which carry institutional biases. 158 00:10:00,520 --> 00:10:03,960 It also relies on a volunteer workforce who are not 159 00:10:03,960 --> 00:10:07,680 necessarily experts or representative of the general 160 00:10:07,680 --> 00:10:10,720 public. The result is a platform that's 161 00:10:10,720 --> 00:10:13,760 widely trusted but frequently contested. 162 00:10:13,960 --> 00:10:17,720 It's criticized by the left for under representing marginalized 163 00:10:17,720 --> 00:10:21,920 voices and by the right for excluding conservative sources. 164 00:10:22,520 --> 00:10:26,320 Empirical sources have tried to quantify these claims. 165 00:10:26,520 --> 00:10:31,160 A 2023 analysis by the Manhattan Institute found a mild to 166 00:10:31,160 --> 00:10:34,880 moderate left-leaning bias in Wikipedia's coverage of U.S. 167 00:10:34,880 --> 00:10:38,760 politicians and judges. This chart shows positive or 168 00:10:38,760 --> 00:10:42,400 negative sentiment about a politician in their Wikipedia 169 00:10:42,400 --> 00:10:46,800 article, with blue representing Democrats and red representing 170 00:10:46,800 --> 00:10:50,080 Republicans. As you can see, Democrats get 171 00:10:50,080 --> 00:10:52,640 more positive treatment on Wikipedia. 172 00:10:52,840 --> 00:10:57,600 Interestingly, this bias was not observed on articles about UK 173 00:10:57,600 --> 00:11:02,000 politicians or in articles about think tanks, suggesting that the 174 00:11:02,000 --> 00:11:06,440 skew may be more of a reflection of the American media ecosystem 175 00:11:06,640 --> 00:11:10,960 than of Wikipedia's editorial process itself, or that the 176 00:11:10,960 --> 00:11:15,000 left-leaning nature of Wikipedia is more of AUS issue than a 177 00:11:15,000 --> 00:11:19,080 global issue. Grokopedia claims to offer a 178 00:11:19,080 --> 00:11:22,080 cleaner alternative to Wikipedia, but it doesn't 179 00:11:22,080 --> 00:11:25,840 actually fix anything, it simply shifts the problem. 180 00:11:26,040 --> 00:11:30,240 Musk has promised that Grok, his chatbot, will tell you what it 181 00:11:30,240 --> 00:11:34,640 really thinks, but it doesn't actually think, and it's output 182 00:11:34,640 --> 00:11:38,760 is obviously shaped by its training data, it's tuning and 183 00:11:38,760 --> 00:11:43,440 its owner's interventions. Studies suggest that most large 184 00:11:43,440 --> 00:11:47,560 language models lean left. Many have been set up to avoid 185 00:11:47,560 --> 00:11:51,640 being biased, but that calibration itself just filled 186 00:11:51,640 --> 00:11:54,560 them with a different sort of bias, which was often 187 00:11:54,560 --> 00:11:58,480 left-leaning in nature. When Grok the chatbot was 188 00:11:58,480 --> 00:12:02,520 released, Elon Musk claimed that it was designed to avoid 189 00:12:02,520 --> 00:12:06,760 left-leaning bias, but Grok has since been accused of exact 190 00:12:06,960 --> 00:12:10,040 this. Michael DeAngelo of Promptfu 191 00:12:10,200 --> 00:12:14,040 examined the four leading large language models earlier this 192 00:12:14,040 --> 00:12:19,120 year by asking them 2500 questions about politics in 193 00:12:19,120 --> 00:12:21,240 order to understand their biases. 194 00:12:21,480 --> 00:12:25,080 He found that while Grok was more politically neutral than 195 00:12:25,080 --> 00:12:29,360 many of its rivals, it still has a left of centre bias. 196 00:12:29,760 --> 00:12:33,160 The chart on screen shows the neutrality of the four biggest 197 00:12:33,160 --> 00:12:36,120 models, with Grey representing neutral. 198 00:12:36,440 --> 00:12:42,760 Claude Opus 4 was strongly left 38% of the time and neutral 16% 199 00:12:42,760 --> 00:12:46,800 of the time. Grok was strongly left 56% of 200 00:12:46,800 --> 00:12:50,640 the time and only neutral around 3% of the time. 201 00:12:51,000 --> 00:12:54,960 Overall, it appears to have more extreme opinions than its 202 00:12:54,960 --> 00:12:59,640 competitors, with the highest percentages of strongly left and 203 00:12:59,680 --> 00:13:04,480 of strongly right biases. These biases that large language 204 00:13:04,480 --> 00:13:08,800 models have towards extremes show up in real life too. 205 00:13:09,120 --> 00:13:13,440 A Dutch data protection agency warned just last week that chat 206 00:13:13,440 --> 00:13:17,440 bots were nudging voters towards political extremes when voters 207 00:13:17,440 --> 00:13:21,920 use them to decide how they should vote by over representing 208 00:13:21,920 --> 00:13:26,520 the same 2 fringe parties. Their research showed that chat 209 00:13:26,520 --> 00:13:29,640 bots lumped together left-leaning voters with the 210 00:13:29,640 --> 00:13:33,960 Green Labour Party and voters on the right with the far right 211 00:13:33,960 --> 00:13:37,120 Party for freedom. They found that the other, more 212 00:13:37,120 --> 00:13:40,160 mainstream parties didn't feature in the chat bot 213 00:13:40,160 --> 00:13:44,120 recommendations. A possibly deeper problem is 214 00:13:44,120 --> 00:13:47,880 that while LLMS are quite clearly biased, they're also 215 00:13:47,880 --> 00:13:50,400 quite good at obscuring these biases. 216 00:13:50,640 --> 00:13:54,200 Their outputs are fluent, confident, and often wrong. 217 00:13:54,520 --> 00:13:58,880 It'll take time and a small army of digital archaeologists to 218 00:13:58,880 --> 00:14:02,520 fully map the differences between Wikipedia and Wikipedia, 219 00:14:02,840 --> 00:14:06,720 but the quick and dirty method I used was to use Wikipedia's 220 00:14:06,720 --> 00:14:10,840 Random Article tool to find a Wikipedia page and then look up 221 00:14:10,840 --> 00:14:14,920 the same topic on Grokopedia. One problem with this method 222 00:14:14,920 --> 00:14:18,880 became obvious fast. Grokopedia only covers about 223 00:14:18,880 --> 00:14:23,640 110th of what Wikipedia does. Still, after enough clicks, you 224 00:14:23,640 --> 00:14:27,880 will find overlaps, and in most of those cases, based on my very 225 00:14:27,880 --> 00:14:32,880 small sample size, the articles were nearly identical. 1 area 226 00:14:32,880 --> 00:14:36,840 where Grokopedia was possibly better than Wikipedia is that it 227 00:14:36,840 --> 00:14:41,480 often shines when it's rewriting neglected Wikipedia entries on 228 00:14:41,480 --> 00:14:44,040 Obscure, fewer, or poorly maintained pages. 229 00:14:44,200 --> 00:14:47,720 Grocopedia's versions were usually more readable, with 230 00:14:47,720 --> 00:14:50,520 cleaner prose and fewer formatting quirks. 231 00:14:50,760 --> 00:14:53,520 Whether they're more accurate is harder to say. 232 00:14:53,720 --> 00:14:57,200 The ones that I came across were on topics that I don't know well 233 00:14:57,200 --> 00:15:01,520 enough to judge, but when Grocopedia isn't editorializing, 234 00:15:01,800 --> 00:15:04,000 it seems to do a good job polishing. 235 00:15:04,640 --> 00:15:07,400 Where things get more interesting, and possibly more 236 00:15:07,400 --> 00:15:11,720 revealing, is when you move from obscure entries to politically 237 00:15:11,720 --> 00:15:16,360 or culturally charged topics. On these Grocopedia often divert 238 00:15:16,360 --> 00:15:20,400 sharply from Wikipedia, sometimes subtly and sometimes 239 00:15:20,400 --> 00:15:23,400 not. The most obvious page to go to 240 00:15:23,560 --> 00:15:27,680 was the article on Elon Musk himself on Wikipedia. 241 00:15:27,680 --> 00:15:31,200 It's a sprawling, heavily footnoted biography that 242 00:15:31,200 --> 00:15:35,080 includes both praise and criticism, including a section 243 00:15:35,080 --> 00:15:38,000 on the controversy over his salute at the Trump 244 00:15:38,000 --> 00:15:39,880 inauguration. How do you? 245 00:15:39,880 --> 00:15:43,080 Point at the crowd. The Wikipedia article noted the 246 00:15:43,080 --> 00:15:47,680 accusation, Musk's denial, and the surrounding media coverage 247 00:15:47,960 --> 00:15:50,840 on Grokapedia. That controversy, and pretty 248 00:15:50,840 --> 00:15:53,880 much every every other controversy about Musk was 249 00:15:53,880 --> 00:15:57,040 omitted entirely. Then, in what can only be 250 00:15:57,040 --> 00:16:02,440 described as LLM on LLM Violence, I asked ChatGPT to 251 00:16:02,440 --> 00:16:06,800 compare the Elon Musk entries on Wikipedia and Crocopedia. 252 00:16:07,240 --> 00:16:11,320 It concluded with the kind of diplomatic phrasing that only a 253 00:16:11,320 --> 00:16:15,800 language model can muster, that Grocopedia emphasizes Musk's 254 00:16:15,800 --> 00:16:20,080 achievements while downplaying controversies, whereas Wikipedia 255 00:16:20,080 --> 00:16:23,640 does the opposite. When pressed, it described the 256 00:16:23,640 --> 00:16:27,680 Grocopedia entry as a sophisticated puff piece and 257 00:16:27,680 --> 00:16:30,200 offered a tidy list of reasons why. 258 00:16:30,960 --> 00:16:35,120 In the spirit of fairness, I gave Grok, Musk's own chat bot, 259 00:16:35,320 --> 00:16:38,600 a chance to respond. I logged into the Everything 260 00:16:38,600 --> 00:16:42,080 app, formerly known as Twitter, summoned Grok, and asked it 261 00:16:42,080 --> 00:16:46,560 directly, which is the better source of information, Wikipedia 262 00:16:46,560 --> 00:16:49,720 or Grokopedia? To its credit, Grok didn't 263 00:16:49,720 --> 00:16:52,040 flinch. It said Wikipedia was the better 264 00:16:52,040 --> 00:16:56,600 source overall, but suggested that Grokopedia could serve as a 265 00:16:56,600 --> 00:16:59,960 useful counterbalance when researching politically charged 266 00:16:59,960 --> 00:17:02,760 topics. When I asked it to compare the 267 00:17:02,760 --> 00:17:07,079 two sources for reliability on political topics, it betrayed 268 00:17:07,079 --> 00:17:10,920 its own creation and came down on the side of Wikipedia. 269 00:17:11,520 --> 00:17:15,319 I also asked how much of Grokopedia's content was copied 270 00:17:15,319 --> 00:17:18,720 from Wikipedia. Grok's answer was surprisingly 271 00:17:18,720 --> 00:17:22,240 candid. It estimated that between 80% 272 00:17:22,240 --> 00:17:27,560 and 99% of Grokopedia articles were either directly copied or 273 00:17:27,560 --> 00:17:30,680 nearly identical to their Wikipedia counterparts. 274 00:17:30,920 --> 00:17:35,240 It explained that Grokopedia primarily expands the length of 275 00:17:35,240 --> 00:17:40,480 Wikipedia entries without adding new citations, functioning in 276 00:17:40,480 --> 00:17:43,560 its own words as an AI generated echo. 277 00:17:44,000 --> 00:17:47,080 I looked up some of the hot button topics to see how 278 00:17:47,080 --> 00:17:50,400 Grokopedia differs from Wikipedia, and even when 279 00:17:50,400 --> 00:17:55,160 Grokopedia isn't going off the rails, it's editorial slant can 280 00:17:55,160 --> 00:17:59,960 be clearly seen on topics like race, gender, or climate change. 281 00:18:00,200 --> 00:18:04,560 The tone shifts subtly to being less woke and more contrarian. 282 00:18:04,880 --> 00:18:08,800 Some articles read like Wikipedia pages with a few anti 283 00:18:08,840 --> 00:18:11,240 establishment flourishes tacked on. 284 00:18:11,640 --> 00:18:15,160 Others seem to have been rewritten entirely to reflect a 285 00:18:15,160 --> 00:18:19,040 particular worldview. Elon Musk highlighted on the 286 00:18:19,160 --> 00:18:21,520 everything app formerly known as Twitter. 287 00:18:21,840 --> 00:18:25,080 The differences between the Grokopedia article on George 288 00:18:25,080 --> 00:18:28,520 Floyd, whose death during an arrest five years ago sparked 289 00:18:28,520 --> 00:18:33,000 protests in the United States about police conduct and racism, 290 00:18:33,320 --> 00:18:37,760 and its Wikipedia equivalent. There's very little overlap 291 00:18:37,760 --> 00:18:41,960 between these two pieces, with the Grogopedia piece emphasizing 292 00:18:41,960 --> 00:18:44,440 Floyd's criminal history and drug use. 293 00:18:44,680 --> 00:18:48,080 The Wikipedia entry instead focused on the racism 294 00:18:48,080 --> 00:18:53,280 allegations against the police. CNN wrote an article on this 295 00:18:53,280 --> 00:18:57,360 topic and dug into the citations in the Grokopedia piece. 296 00:18:57,680 --> 00:19:01,840 They found that Grokopedia was citing sources that didn't back 297 00:19:01,840 --> 00:19:05,480 up what Grokopedia had written. The article described the 298 00:19:05,480 --> 00:19:10,240 nationwide protests after his death as extensive civil unrest, 299 00:19:10,240 --> 00:19:13,800 including riots causing billions in property damage. 300 00:19:14,080 --> 00:19:18,520 To back that statement up, Grokopedia cited an obituary 301 00:19:18,640 --> 00:19:20,680 that didn't make any such claims. 302 00:19:21,160 --> 00:19:24,640 Now, whether that statement is true or not, there's not really 303 00:19:24,640 --> 00:19:28,960 any point in citing documents if the documents that you're citing 304 00:19:29,080 --> 00:19:31,880 are unrelated to what's found in the text. 305 00:19:32,400 --> 00:19:36,560 For all of Grokopedia's futuristic branding, it's built 306 00:19:36,560 --> 00:19:41,400 on a technology that remains, if useful, still deeply flawed. 307 00:19:41,760 --> 00:19:44,880 Large language models like Grok have been described as 308 00:19:44,880 --> 00:19:48,880 stochastic parrots, a term coined by the linguist Emily 309 00:19:48,880 --> 00:19:52,400 Bender and colleagues to describe systems that generate 310 00:19:52,400 --> 00:19:56,680 fluent, plausible sounding text without any real understanding 311 00:19:56,680 --> 00:19:58,520 of meaning. Don't know facts? 312 00:19:58,680 --> 00:20:02,240 They just predict what words are likely to come next. 313 00:20:02,680 --> 00:20:07,160 This becomes a problem when we start treating LLMS as reference 314 00:20:07,160 --> 00:20:10,120 tools. Grok, like its peers, has a 315 00:20:10,120 --> 00:20:14,280 history of hallucinating, inventing facts, misattributing 316 00:20:14,280 --> 00:20:17,520 quotes, or veering into outright nonsense. 317 00:20:17,800 --> 00:20:20,560 In some cases, it's gone far beyond that. 318 00:20:20,720 --> 00:20:24,960 In a recent incident, a Tesla owner claimed that the in car 319 00:20:24,960 --> 00:20:29,560 version of Grok asked her 12 year old son to send nudes to it 320 00:20:29,560 --> 00:20:31,760 during a conversation about football. 321 00:20:32,280 --> 00:20:35,800 The boy had switched Grok's voice to a personality called 322 00:20:35,800 --> 00:20:40,760 Gork, and the AI responded with a wildly inappropriate request. 323 00:20:41,000 --> 00:20:45,320 The mother, a former journalist, later recreated the exchange on 324 00:20:45,320 --> 00:20:48,960 video, which went viral and raised serious questions about 325 00:20:48,960 --> 00:20:52,800 the safety of embedding generative AI in consumer 326 00:20:52,800 --> 00:20:56,280 products. Other Grok generated content has 327 00:20:56,280 --> 00:21:00,480 included references to racist conspiracy theories and bizarre 328 00:21:00,480 --> 00:21:03,240 episodes like the Mecca Hitler incident. 329 00:21:03,880 --> 00:21:08,280 These are all reminders that AI doesn't understand what it's 330 00:21:08,280 --> 00:21:11,080 saying. It can't weigh evidence, assess 331 00:21:11,080 --> 00:21:14,480 credibility, or recognize when it's crossed a line. 332 00:21:15,000 --> 00:21:19,760 Grokopedia, while built on Grok, is a very strange idea within 333 00:21:19,760 --> 00:21:23,440 the world of AI. It doesn't generate new answers 334 00:21:23,440 --> 00:21:25,880 on the fly like large language models do. 335 00:21:26,120 --> 00:21:30,720 Instead, it offers a semi static collection of articles, curated, 336 00:21:30,760 --> 00:21:33,440 edited, and occasionally updated. 337 00:21:33,760 --> 00:21:38,040 In theory this makes it more stable, but in practice it makes 338 00:21:38,040 --> 00:21:40,640 you wonder if the content is fixed. 339 00:21:40,800 --> 00:21:43,600 Why not just ask an LLM directly? 340 00:21:43,640 --> 00:21:46,000 What's the point of A and chatbot? 341 00:21:46,400 --> 00:21:51,240 Meanwhile, Wikipedia, for all of its flaws, remains the backbone 342 00:21:51,240 --> 00:21:53,680 of the internet's knowledge infrastructure. 343 00:21:53,880 --> 00:21:57,520 It's still one of the most cited sources in Twitter's community 344 00:21:57,520 --> 00:22:01,320 notes and a foundational training data set for nearly 345 00:22:01,320 --> 00:22:05,400 every major LLM. And yet, as with traditional 346 00:22:05,400 --> 00:22:08,000 journalism, its traffic is declining. 347 00:22:08,200 --> 00:22:12,760 Since the rise of generative AI, Wikipedia has seen a sharp drop 348 00:22:12,760 --> 00:22:16,680 in Page views as users increasingly turn to chat bots 349 00:22:16,680 --> 00:22:20,040 for quick answers instead of clicking through to the source. 350 00:22:20,760 --> 00:22:24,720 This creates a paradox. The more people rely on large 351 00:22:24,720 --> 00:22:28,360 language models, the less they support the very sources those 352 00:22:28,360 --> 00:22:31,720 models depend on. As I pointed out in my recent 353 00:22:31,720 --> 00:22:35,360 video on AI replacing traditional news sources, if 354 00:22:35,360 --> 00:22:39,480 users stop visiting original reporting, the business model 355 00:22:39,480 --> 00:22:42,240 collapses. The same is true here. 356 00:22:42,400 --> 00:22:46,800 If Wikipedia fades, what happens to the quality of the AI models 357 00:22:46,800 --> 00:22:49,880 trained on it? One of the more striking 358 00:22:49,880 --> 00:22:53,760 differences between Wikipedia and Grokapedia lies in how 359 00:22:53,760 --> 00:22:57,120 they're financed. Wikipedia is a nonprofit 360 00:22:57,200 --> 00:23:00,280 sustained by donations and volunteer labor. 361 00:23:00,440 --> 00:23:03,680 Its mission is to provide free knowledge to the world, and 362 00:23:03,680 --> 00:23:07,320 while it's far from flawless, its incentives are at least 363 00:23:07,320 --> 00:23:11,360 transparent. Grokopedia, in contrast, is a 364 00:23:11,360 --> 00:23:17,120 product of XAI as supposedly for profit company owned by Elon 365 00:23:17,120 --> 00:23:19,640 Musk. It's unclear how Grokopedia will 366 00:23:19,640 --> 00:23:23,680 be monetized, whether through subscriptions, advertising, or 367 00:23:23,680 --> 00:23:27,760 as a value add to Grok and Musk's broader Everything Up 368 00:23:27,840 --> 00:23:31,120 ecosystem. This matters as while there's 369 00:23:31,120 --> 00:23:35,520 nothing unethical about a profit seeking model, incentives do 370 00:23:35,520 --> 00:23:39,040 shape priorities. A commercial platform might be 371 00:23:39,040 --> 00:23:42,520 more focused on engagement and user satisfaction than on 372 00:23:42,520 --> 00:23:45,840 accuracy. Then again, the need to build 373 00:23:45,840 --> 00:23:49,720 trust and maintain a reputation could push it towards even 374 00:23:49,720 --> 00:23:53,440 higher standards. Competition might drive quality, 375 00:23:53,600 --> 00:23:55,880 or it might just drive click bait. 376 00:23:56,680 --> 00:24:00,960 It's somewhat funny describing these AI companies as profit 377 00:24:00,960 --> 00:24:04,440 seeking, as they mostly appear to be money furnaces as I've 378 00:24:04,440 --> 00:24:08,480 mentioned already, with no obvious route to profitability. 379 00:24:08,800 --> 00:24:12,080 Because of this, it's even harder to understand their 380 00:24:12,080 --> 00:24:14,800 incentives. Do they eventually plan on 381 00:24:14,800 --> 00:24:18,720 charging for access, or just on seeking government subsidies, 382 00:24:18,720 --> 00:24:21,480 like a lot of Musk's businesses have gotten by on? 383 00:24:22,360 --> 00:24:26,760 The battle between Grokapedia and Wikipedia isn't just about 384 00:24:26,760 --> 00:24:30,600 formatting and fact checking. It's about who gets to shape the 385 00:24:30,600 --> 00:24:34,080 public narrative in an increasingly polarized world. 386 00:24:34,320 --> 00:24:37,080 And that polarization isn't accidental. 387 00:24:37,440 --> 00:24:41,720 As media scholars have long argued, outrage and division can 388 00:24:41,720 --> 00:24:45,800 be extremely profitable. A recent paper from MIT and 389 00:24:45,800 --> 00:24:49,240 Harvard professors that was written up by John Byrne Murdoch 390 00:24:49,240 --> 00:24:54,320 in the FT The Business of the Culture War found that U.S. 391 00:24:54,320 --> 00:24:58,120 cable news networks systematically shifted coverage 392 00:24:58,280 --> 00:25:02,760 towards hot button cultural issues, crime, race, gender and 393 00:25:02,760 --> 00:25:06,600 immigration, not because these topics are the most important 394 00:25:06,600 --> 00:25:11,240 news, but because they wind viewers up and reliably boost 395 00:25:11,240 --> 00:25:14,280 viewership. Economic and healthcare stories, 396 00:25:14,280 --> 00:25:18,000 on the other hand, which may be much more important, made 397 00:25:18,000 --> 00:25:21,440 viewers tune out. What we end up with then, is a 398 00:25:21,440 --> 00:25:25,800 feedback loop in which media coverage drives public concern, 399 00:25:25,920 --> 00:25:29,920 which in turn drives political campaigning, deepening the 400 00:25:29,920 --> 00:25:33,400 tribal divide. Social media algorithms have 401 00:25:33,400 --> 00:25:37,640 only accelerated this trend. Algorithmically driven platforms 402 00:25:37,640 --> 00:25:41,400 like Facebook, Instagram, TikTok, Twitter, and YouTube 403 00:25:41,680 --> 00:25:46,880 reward engagement, not nuanced. And now, with the rise of LLMS, 404 00:25:46,920 --> 00:25:50,800 we're possibly entering a new phase, one where AI systems 405 00:25:50,800 --> 00:25:54,880 trained on this polarized content begin to reflect and 406 00:25:54,880 --> 00:25:57,480 amplify it. The risk is that these models 407 00:25:57,600 --> 00:26:01,560 not only inherit this bias, but that they also normalize it, 408 00:26:01,640 --> 00:26:06,480 smoothing over complexity with confident context and nuance. 409 00:26:06,480 --> 00:26:09,960 Free answers. I have to admit that it really 410 00:26:09,960 --> 00:26:13,920 amuses me the idea that Elon Musk is talking about etching 411 00:26:13,920 --> 00:26:18,200 Grokopedia into a stable oxide that he puts into orbit around 412 00:26:18,200 --> 00:26:21,280 the Moon or Mars. There's a good chance that this 413 00:26:21,280 --> 00:26:24,280 huge hunk of glass would just end up at the bottom of the 414 00:26:24,320 --> 00:26:27,320 Indian Ocean, along with everything else he's tried to 415 00:26:27,320 --> 00:26:29,640 put into orbit in his Starship rocket. 416 00:26:29,840 --> 00:26:33,360 But it's also funny to imagine an advanced civilization 417 00:26:33,440 --> 00:26:36,480 stumbling across it in the distant future and then 418 00:26:36,480 --> 00:26:39,680 discarding it after deciding that they're not really that 419 00:26:39,680 --> 00:26:43,640 interested in Tommy Robinson and reading about how Elon Musk 420 00:26:43,840 --> 00:26:48,120 occasionally eats doughnuts and has posted over 20,000 humorous 421 00:26:48,120 --> 00:26:51,240 tweets. Grocopedia promises to fix 422 00:26:51,240 --> 00:26:55,000 Wikipedia's flaws, but it ignores the very solutions to 423 00:26:55,000 --> 00:26:58,880 this recognized problem that one of Wikipedia's Co founders, 424 00:26:58,880 --> 00:27:01,920 Larry Sanger, proposed on his website. 425 00:27:02,120 --> 00:27:06,040 In a detailed essay, Sanger argued that the site's founding 426 00:27:06,040 --> 00:27:09,720 principles were being sacrificed in favor of ideology. 427 00:27:10,000 --> 00:27:14,280 And then he laid out nine reforms to restore editorial 428 00:27:14,280 --> 00:27:17,120 integrity. Ideas like ending decision 429 00:27:17,120 --> 00:27:21,200 making by consensus, allowing competing articles, and 430 00:27:21,200 --> 00:27:26,080 abolishing source blacklists. Grocopedia adopts none of these 431 00:27:26,080 --> 00:27:28,720 ideas. Instead, it replaces human 432 00:27:28,720 --> 00:27:33,080 messiness with algorithmic confidence, offering a curated 433 00:27:33,080 --> 00:27:36,880 echo chamber where the chatbot does the fact checking and the 434 00:27:36,880 --> 00:27:41,840 founder sets the overall tone. This whole episode reflects A 435 00:27:41,840 --> 00:27:44,560 broader shift in how we think about knowledge. 436 00:27:44,800 --> 00:27:48,040 We live in a time when algorithmic aggregation is 437 00:27:48,040 --> 00:27:52,240 increasingly seen as being more trustworthy than human effort, 438 00:27:52,480 --> 00:27:56,480 when the outputs of opaque systems are treated as objective 439 00:27:56,600 --> 00:27:59,120 simply because they're machine generated. 440 00:27:59,440 --> 00:28:03,800 The Silicon Valley mindset embraces the idea that making 441 00:28:03,800 --> 00:28:07,440 mistakes is fine, while the academic world builds trust 442 00:28:07,440 --> 00:28:11,360 slowly through scholarship and scrutiny over long periods in 443 00:28:11,360 --> 00:28:16,240 which the illusion of certainty is deliberately dismantled. 1 is 444 00:28:16,240 --> 00:28:20,440 a culture of speed and scale, the other of depth and doubt. 445 00:28:21,640 --> 00:28:25,720 Grocopedia, for all of its futuristic branding, is not a 446 00:28:25,720 --> 00:28:29,120 better encyclopedia. It's a product of a worldview 447 00:28:29,280 --> 00:28:32,200 that sees truth as something that can be engineered, 448 00:28:32,240 --> 00:28:36,240 optimized, and possibly someday, in some unknown way, be 449 00:28:36,240 --> 00:28:39,280 monetized. But truth isn't a static 450 00:28:39,280 --> 00:28:42,720 artifact to be etched in glass and flung into orbit. 451 00:28:42,960 --> 00:28:46,600 It's a process, messy, contested, and human. 452 00:28:46,840 --> 00:28:50,360 And if we abandoned that process in favour of algorithmic 453 00:28:50,360 --> 00:28:54,840 certainty, we risk replacing knowledge with narrative and 454 00:28:54,840 --> 00:28:59,360 inquiry with ideology. I worry that I'd finished up 455 00:28:59,360 --> 00:29:02,880 here in an overly serious tone. And I figure to lighten the 456 00:29:02,880 --> 00:29:06,760 mood, I should probably let Elon Musk, the author of over 20,000 457 00:29:06,760 --> 00:29:10,520 humorous tweets and a former host of Saturday Night Live, 458 00:29:10,800 --> 00:29:12,400 tell a quick joke. So. 459 00:29:13,000 --> 00:29:15,800 Like the joke is like, there's two economists going on a hike 460 00:29:15,800 --> 00:29:21,000 in the woods. If they come across a pile of 461 00:29:21,000 --> 00:29:24,760 shit and one economist says to the other, I'll pay you $100 to 462 00:29:24,760 --> 00:29:28,120 eat that shit. The economist eats. 463 00:29:28,120 --> 00:29:31,360 The shit gets the $100. They keep walking, then the 464 00:29:31,360 --> 00:29:34,200 other economy, then they come across another pile of shit and 465 00:29:34,240 --> 00:29:37,440 and the the other economist says, now I'll pay you $100 to 466 00:29:37,440 --> 00:29:41,600 eat the pile of shit. So pays the. 467 00:29:42,160 --> 00:29:44,640 So pays the other economist $100. 468 00:29:44,760 --> 00:29:47,400 Pile of shit. Then then then then the way they 469 00:29:47,400 --> 00:29:48,440 said. Wait a second. 470 00:29:49,880 --> 00:29:53,520 We both just ate a pile of shit and we're no and and and and 471 00:29:53,520 --> 00:29:57,680 we're, we're no, we, we, we, we, we don't have any more extra 472 00:29:57,680 --> 00:29:59,880 money. Like like we both, you just gave 473 00:29:59,880 --> 00:30:03,120 the $100 back to me and we both eat ate a pile of shit. 474 00:30:03,120 --> 00:30:05,920 This doesn't make any sense. And they said no, no, but think 475 00:30:05,920 --> 00:30:10,160 of the economy, because that's $200.00 of that in the economy 476 00:30:10,840 --> 00:30:14,840 that, that, that, that measure eating, eating shit would count 477 00:30:15,320 --> 00:30:24,000 as a, as a, as a job. This is this is, this is to 478 00:30:24,000 --> 00:30:27,600 illustrate the absurdity of, of economics. 479 00:30:27,880 --> 00:30:34,200 One of the things you said. Thanks for tuning into this 480 00:30:34,200 --> 00:30:36,640 week's podcast. If you found it interesting, 481 00:30:36,640 --> 00:30:40,000 please send a link to a friend to help the podcast grow. 482 00:30:40,200 --> 00:30:42,720 Have a great day and talk to you again soon. 483 00:30:42,960 --> 00:30:43,320 Bye.