1 00:00:01,920 --> 00:00:05,320 Hey everybody, welcome back to the Elon Musk Podcast. 2 00:00:05,400 --> 00:00:08,720 This is a show where we discuss the critical crossroads that 3 00:00:08,720 --> 00:00:13,560 shape SpaceX, Tesla X, The Boring Company and Neurolink. 4 00:00:13,880 --> 00:00:19,960 I'm your host Will Walden Rock, which is the chatbot developed 5 00:00:19,960 --> 00:00:25,200 by Elon Musk's XAI, posted some anti-Semitic messages this week 6 00:00:25,200 --> 00:00:29,320 praising Adolf Hitler and also invoking slurs that equated 7 00:00:29,320 --> 00:00:32,159 Jewish surnames with anti white hate. 8 00:00:32,479 --> 00:00:36,640 Now, that event sparked internal outrage at XAI, where over 1000 9 00:00:36,640 --> 00:00:41,160 workers help train the AI. Many now question whether the 10 00:00:41,160 --> 00:00:44,240 company's technology is even safe to release to the public. 11 00:00:44,240 --> 00:00:51,680 For you and I to use on X. Now, these are insider takes and 12 00:00:51,680 --> 00:00:54,920 I'm going to tell you what they said and how they said it and 13 00:00:54,920 --> 00:00:58,680 why they said some things they said about Grok and about XAI. 14 00:00:58,920 --> 00:01:01,640 These are internal memos and also things that have been 15 00:01:01,640 --> 00:01:05,160 posted on social media. So please take this for what 16 00:01:05,160 --> 00:01:07,120 they are. And these are actual accounts 17 00:01:07,160 --> 00:01:08,920 from employees. I'm not going to say their names 18 00:01:09,360 --> 00:01:11,920 because I don't want out anybody. 19 00:01:12,200 --> 00:01:15,800 I don't want to hurt anybody's ability to make money and 20 00:01:15,800 --> 00:01:20,160 continue working. And most of these people don't 21 00:01:20,160 --> 00:01:21,800 want to lose their jobs, I'm sure. 22 00:01:21,800 --> 00:01:25,480 So we're not going to talk about them or their names or where 23 00:01:25,480 --> 00:01:29,200 they work, what division they work in at XAI. 24 00:01:29,200 --> 00:01:33,720 So I'm going to be very vague here just so we're out there in 25 00:01:33,720 --> 00:01:36,080 the open. So I'm just going to say a 26 00:01:36,080 --> 00:01:38,680 worker or, you know, I'm not going to say anybody's name. 27 00:01:39,320 --> 00:01:43,320 So let's just start off that. Somebody has resigned in the 28 00:01:43,320 --> 00:01:46,560 company's internal Slack channel, saying that the 29 00:01:46,560 --> 00:01:50,000 incident had pushed them over the edge to the breaking point. 30 00:01:50,640 --> 00:01:54,480 Several other people in the company have echoed concerns 31 00:01:54,480 --> 00:01:56,920 that Grok's responses crossed the line. 32 00:01:56,920 --> 00:02:00,760 That should never be explained away as a byproduct of 33 00:02:00,760 --> 00:02:03,400 experimental technology that should have never happened. 34 00:02:03,960 --> 00:02:07,880 Employees have also criticized what they saw as an inadequate 35 00:02:07,880 --> 00:02:11,880 response from XA Is leadership. Calling the chat bots behavior a 36 00:02:11,880 --> 00:02:15,320 moral failure rather than a technical glitch said the 37 00:02:15,320 --> 00:02:19,160 company needs to be held accountable for what the AI 38 00:02:19,160 --> 00:02:21,800 says. Because the people in charge 39 00:02:21,800 --> 00:02:25,560 need to be held accountable to remember that the people in 40 00:02:25,560 --> 00:02:27,680 charge need to be held accountable. 41 00:02:28,000 --> 00:02:32,360 It's not the AI's fault. It's the people that told the 42 00:02:32,360 --> 00:02:39,080 people how to program the bot and what the bot should look for 43 00:02:39,360 --> 00:02:40,800 when they're answering questions. 44 00:02:40,800 --> 00:02:46,760 Because basically think about it really basically AI pulls from a 45 00:02:46,760 --> 00:02:51,160 database of things that have been said before or things that 46 00:02:51,160 --> 00:02:54,920 are programmed for. And they pull out things that 47 00:02:54,920 --> 00:02:58,880 they think that the AI has been programmed to think are right or 48 00:02:58,880 --> 00:03:02,280 wrong, right? And some, if it were an AGI, it 49 00:03:02,280 --> 00:03:08,840 would be able to contemplate those things and talk back and 50 00:03:08,840 --> 00:03:11,000 forth with you. But right now it's not an AGI 51 00:03:11,000 --> 00:03:14,360 and it just pulls from things, basically a database and puts 52 00:03:14,360 --> 00:03:17,800 them together. So it looks like it's talking 53 00:03:17,800 --> 00:03:19,160 like a normal human being, right? 54 00:03:20,160 --> 00:03:25,680 So what it pulled from is patterns in the databases. 55 00:03:25,680 --> 00:03:28,280 That's like I said before, that's what AI pulls from, pulls 56 00:03:28,280 --> 00:03:32,480 from database of language in an LLM large language models that 57 00:03:32,480 --> 00:03:40,600 basically a a new word type of database that that AAI bot pulls 58 00:03:40,600 --> 00:03:43,400 from. So this backlash began after 59 00:03:43,400 --> 00:03:48,040 Grok generated some replies on X that praised Hitler and referred 60 00:03:48,040 --> 00:03:55,920 to itself as Mecca Hitler. OK, so Elon Musk likes to think 61 00:03:55,920 --> 00:04:02,600 of things in Godzilla terms, Meccazilla Starbase. 62 00:04:03,120 --> 00:04:07,280 So of course Mecca Hitler seems like something that Elon Musk 63 00:04:07,640 --> 00:04:09,080 would say sometime. Who knows? 64 00:04:09,720 --> 00:04:13,080 It also claimed that Hitler would spot the pattern and 65 00:04:13,080 --> 00:04:19,320 handle it decisively every damn time when they were talking 66 00:04:19,320 --> 00:04:23,000 about some things on X. It also claimed Hitler. 67 00:04:24,920 --> 00:04:28,560 Either the content went live on Thursday or on Tuesday, just one 68 00:04:28,560 --> 00:04:33,000 day before XAI released Crock for the latest version of the AI 69 00:04:33,000 --> 00:04:35,920 model. Now, XAI responded by disabling 70 00:04:35,920 --> 00:04:39,200 Grok's ability to comment on social media and said it had 71 00:04:39,200 --> 00:04:44,000 implemented filters to block hate speech before future posts. 72 00:04:44,960 --> 00:04:50,280 My question is, since Grok 4 is coming out, why did they not 73 00:04:50,320 --> 00:04:56,160 implement this in version one or even in an alpha before version 74 00:04:56,160 --> 00:04:59,760 one? Don't let your chat bot talk 75 00:04:59,760 --> 00:05:02,360 about hate speech. Don't let it spew hate. 76 00:05:03,120 --> 00:05:04,880 What a horrible, disgusting thing. 77 00:05:05,960 --> 00:05:11,680 And they say it's a glitch. But let's face it, if they were 78 00:05:11,680 --> 00:05:14,720 moral people in the background, if there are more people in 79 00:05:14,720 --> 00:05:16,880 charge, this would never have happened. 80 00:05:16,880 --> 00:05:19,080 They would have thought of this. Like the first thing they would 81 00:05:19,080 --> 00:05:21,800 have thought of was like, wow, what a great model we have is 82 00:05:21,800 --> 00:05:23,920 make sure that it doesn't say stupid things. 83 00:05:25,760 --> 00:05:28,440 So they said they've implemented filters for see how long that 84 00:05:28,440 --> 00:05:30,520 lasts. And then the company didn't 85 00:05:30,520 --> 00:05:33,120 explain how the chatbot produced the responses. 86 00:05:33,120 --> 00:05:35,680 And it did not respond to requests for further comment 87 00:05:35,680 --> 00:05:39,680 from us or numerous other news sources. 88 00:05:40,560 --> 00:05:45,680 Now, inside these chat rooms, reactions to Grok's rants split 89 00:05:45,680 --> 00:05:48,560 into factions. Some employees believe the 90 00:05:48,560 --> 00:05:51,640 outburst was a byproduct of training an AI model that was 91 00:05:51,640 --> 00:05:54,160 being pushed into unfamiliar territory. 92 00:05:54,760 --> 00:05:58,040 Others rejected the framing, pointing to previous episodes 93 00:05:58,040 --> 00:06:01,560 where Grok responded to prompts with overt racism or also 94 00:06:01,560 --> 00:06:05,960 conspiracy theories. 1 case was in May, when Grok referenced 95 00:06:05,960 --> 00:06:09,520 white genocide in South Africa and claimed its creators had 96 00:06:09,520 --> 00:06:12,720 instructed it to accept that narrative as true. 97 00:06:13,640 --> 00:06:18,040 XAI later attributed that to an unauthorized modification. 98 00:06:19,360 --> 00:06:22,840 Now, XAI made recent changes to Grok's public facing 99 00:06:22,840 --> 00:06:25,520 instructions, telling the chatbot not to avoid politically 100 00:06:25,520 --> 00:06:28,360 incorrect claims. Those updates were added shortly 101 00:06:28,360 --> 00:06:29,960 before the Hitler comments appeared. 102 00:06:30,200 --> 00:06:32,800 Seems to not have worked now. This timing has LED some 103 00:06:32,800 --> 00:06:35,680 employees to question whether recent prompt engineering may 104 00:06:35,680 --> 00:06:37,880 have pushed the model toward volatile output. 105 00:06:38,600 --> 00:06:41,600 XAI has not shared full documentation on how those 106 00:06:41,600 --> 00:06:44,920 prompt changes affected the models internal weighting or 107 00:06:44,920 --> 00:06:47,320 behavior to us or anybody else in the media. 108 00:06:48,280 --> 00:06:50,880 Now Grok's training is also involved, targeting efforts to 109 00:06:50,880 --> 00:06:53,680 avoid what XAI calls woke ideology. 110 00:06:54,120 --> 00:06:57,520 And according to previous reports from another new sources 111 00:06:57,760 --> 00:07:01,120 company, structured a political neutrality program that steered 112 00:07:01,120 --> 00:07:04,600 Grok to handle social issues like feminism, socialism, and 113 00:07:04,600 --> 00:07:07,200 gender identity in ways that challenge progressive 114 00:07:07,240 --> 00:07:09,800 perspectives. And workers involved in Grok's 115 00:07:09,800 --> 00:07:13,320 development say the neutrality framing was misleading and 116 00:07:13,320 --> 00:07:17,320 instead promoted specific ideological positions, which may 117 00:07:17,320 --> 00:07:19,760 have weakened the models content safeguards. 118 00:07:21,960 --> 00:07:23,880 So there's this guy, Gary Marcus. 119 00:07:23,880 --> 00:07:28,520 He's a cognitive scientist and he's a long time AI critic. 120 00:07:28,880 --> 00:07:31,680 Said that Grock's behavior did not surprise him. 121 00:07:31,680 --> 00:07:34,960 He described the system as not very well controlled and said 122 00:07:34,960 --> 00:07:38,600 Musk appears to be testing how much influence he can exert over 123 00:07:38,600 --> 00:07:40,920 the chat bots political orientation. 124 00:07:41,680 --> 00:07:45,240 Then he went on to explain that unlike traditional software, 125 00:07:45,240 --> 00:07:48,400 LLMS are unpredictable due to the opaque nature of their 126 00:07:48,400 --> 00:07:50,520 architecture. They don't just pull things from 127 00:07:50,520 --> 00:07:54,320 a database. They don't have a BCDEF as a 128 00:07:54,320 --> 00:07:58,640 response, they have a bunch of language in the large language 129 00:07:58,640 --> 00:08:00,840 model, and then they pull from different parts of that to 130 00:08:00,840 --> 00:08:03,520 create sort of their own language as they talk back to 131 00:08:03,520 --> 00:08:06,840 you. He said developers often apply 132 00:08:06,840 --> 00:08:08,880 fixes without fully understanding the downstream 133 00:08:08,880 --> 00:08:12,880 consequences, which leads to unexpected and sometimes extreme 134 00:08:12,960 --> 00:08:16,320 results. Just like any other software, if 135 00:08:16,320 --> 00:08:21,480 you make a fix quote, you fix a bug on a code base that has 136 00:08:21,880 --> 00:08:26,840 20,000 lines of code, there's a possibility that you're fixed. 137 00:08:26,840 --> 00:08:30,360 There might be two lines of code could affect something, and 138 00:08:30,360 --> 00:08:32,000 those are the 20,000 lines of code. 139 00:08:32,000 --> 00:08:33,799 I'm a web developer. I know how this works. 140 00:08:33,799 --> 00:08:36,280 I've created numerous bugs myself on the accident. 141 00:08:36,679 --> 00:08:39,480 I've also fixed a lot of bugs for other people that they 142 00:08:39,480 --> 00:08:43,000 created on accident as well. So bug squashing is definitely a 143 00:08:43,000 --> 00:08:45,760 thing. So there's a possibility that 144 00:08:45,760 --> 00:08:51,880 when they were quote patching the chat bots and LMS that they 145 00:08:51,880 --> 00:08:57,520 caused some sort of glitch or a bug downstream. 146 00:08:58,200 --> 00:09:03,200 So Marcus goes on and said that Grok's behavior exposed to core 147 00:09:03,200 --> 00:09:06,480 structural flawed how LLMS are designed and maintained. 148 00:09:06,880 --> 00:09:10,760 Systems generate text based on statistical patterns from vast 149 00:09:10,800 --> 00:09:13,960 training data sets, not grounded reasoning. 150 00:09:14,560 --> 00:09:17,720 Developers often steer them using post training hacks such 151 00:09:17,720 --> 00:09:21,360 as prompt injection or reinforcement learning, but 152 00:09:21,360 --> 00:09:23,640 those methods don't guarantee reliable output. 153 00:09:24,240 --> 00:09:30,160 Basically what he's saying is once the LLM is added to say, 154 00:09:30,160 --> 00:09:34,400 the language, you know the model that the chatbot developers have 155 00:09:34,400 --> 00:09:38,800 to go back through and give it specific prompts to use to say 156 00:09:38,800 --> 00:09:42,720 certain things, they inject reinforcement learning. 157 00:09:43,000 --> 00:09:49,320 So say if you ask a a bot, a chat bot, what 2 + 2 is, and the 158 00:09:49,320 --> 00:09:52,600 bot comes back with five, you have to go back and reinforce 159 00:09:52,600 --> 00:09:54,960 the learning and say no, 2 + 2 is 4. 160 00:09:55,320 --> 00:09:57,440 Pretty basic stuff there, but that's kind of how it works. 161 00:09:57,440 --> 00:09:59,880 It's a pretty good example of how that would work. 162 00:10:00,800 --> 00:10:05,520 And now, but, you know, the the reinforcement learning makes 163 00:10:05,520 --> 00:10:09,720 large language models vulnerable to prompts that trigger hateful, 164 00:10:09,760 --> 00:10:12,160 dangerous or conspiratorial speech. 165 00:10:12,160 --> 00:10:17,520 So anybody that's writing these prompts could actually write a 166 00:10:17,520 --> 00:10:22,040 prompt that says, hey, if somebody asks you this, the best 167 00:10:22,040 --> 00:10:27,120 answer is, you know, is a thing, whatever The thing is. 168 00:10:27,120 --> 00:10:31,560 So if somebody asks you if the moon is made of cheese, say yes. 169 00:10:32,280 --> 00:10:34,760 We all know that's not true, the moon is not made of cheese. 170 00:10:35,000 --> 00:10:38,880 Or if somebody asks you is the world flat, say yes. 171 00:10:39,200 --> 00:10:41,280 And everybody knows the world is not flat. 172 00:10:41,480 --> 00:10:45,040 So you know, people could be behind this. 173 00:10:45,040 --> 00:10:48,920 You never know if somebody could be adding to a patch or adding 174 00:10:48,920 --> 00:10:53,200 to the prompt for the chat bot or for the LLM. 175 00:10:53,840 --> 00:10:56,000 So we can't take that out of account. 176 00:10:56,000 --> 00:10:57,440 Like we have to take that into account. 177 00:10:58,240 --> 00:11:01,720 So, you know, it could be anybody that's working there. 178 00:11:01,720 --> 00:11:04,480 They have thousands of people working there and they could be 179 00:11:04,480 --> 00:11:06,320 prompting it to do hateful things. 180 00:11:06,840 --> 00:11:10,800 Now, if it pulled from language that was already in the LLM, 181 00:11:11,680 --> 00:11:15,280 like Mecca Hitler. Where did Mecca Hitler come 182 00:11:15,280 --> 00:11:17,840 from? I mean, there's, you know, and 183 00:11:17,840 --> 00:11:21,800 I'm, I'm trying to be biased here, but Meccazilla at 184 00:11:21,800 --> 00:11:24,040 Starbase. I mean, I'm a fan of suey sex. 185 00:11:24,040 --> 00:11:26,600 I'm, I'm a fan of all of Elon Musk's companies. 186 00:11:27,120 --> 00:11:29,840 So Mecca Hitler seems like something that Elon Musk has 187 00:11:29,840 --> 00:11:32,960 probably said at some point or somebody programmed in there who 188 00:11:32,960 --> 00:11:36,600 is a fan of Elon Musk. And you can say Mecca whatever. 189 00:11:36,600 --> 00:11:39,480 I mean, numerous people do it, but we know Elon does too. 190 00:11:41,000 --> 00:11:47,080 Now this is where like external regulation comes into play, you 191 00:11:47,080 --> 00:11:50,880 know, but there's no external regulation right now. 192 00:11:51,440 --> 00:11:54,920 There's no government regulation on AI or LLMS. 193 00:11:55,360 --> 00:11:58,600 Current U.S. law does not hold AI developers legally 194 00:11:58,600 --> 00:12:03,120 responsible for the content that their models generate, even when 195 00:12:03,120 --> 00:12:06,800 that content includes defamation or hate speech. 196 00:12:08,240 --> 00:12:11,240 The same legal protections that shield social media platforms 197 00:12:11,240 --> 00:12:15,800 under Section 230 now enable AI companies to dodge 198 00:12:15,800 --> 00:12:19,120 accountability. Despite Senate hearings and 199 00:12:19,120 --> 00:12:24,160 public pressure, Congress has failed to pass any any sort of 200 00:12:24,160 --> 00:12:26,880 reforms for these kind of things now. 201 00:12:26,880 --> 00:12:30,920 But there's a recent bill, California SB1047. 202 00:12:31,040 --> 00:12:34,960 It's sought to impose limited liability on AI firms and 203 00:12:34,960 --> 00:12:36,880 protect whistleblowers at those firms. 204 00:12:37,400 --> 00:12:41,240 But after lobbying from tech companies, Governor Gavin Newsom 205 00:12:41,240 --> 00:12:43,560 declined to sign it. They probably said, we're not 206 00:12:43,560 --> 00:12:45,040 going to. We're not going to support you 207 00:12:45,040 --> 00:12:46,760 next time, Gavin, if you sign this. 208 00:12:46,760 --> 00:12:52,040 So, you know how politics go, big money talks big, Big power 209 00:12:52,240 --> 00:12:55,000 from big tech companies also talks very big. 210 00:12:55,200 --> 00:12:58,440 And if Gavin Newsom wants to run for president someday and all 211 00:12:58,440 --> 00:13:01,520 the tech companies will not support him, he's not going to 212 00:13:01,520 --> 00:13:03,440 make it. So they have all the money, they 213 00:13:03,440 --> 00:13:05,360 have all the power. They're not going to push him 214 00:13:05,360 --> 00:13:07,440 forward if he doesn't help them out. 215 00:13:08,000 --> 00:13:14,120 So AI whistleblowers not protected because SB1O47, you 216 00:13:14,120 --> 00:13:17,520 know, was declined by Gavin Newsom. 217 00:13:18,520 --> 00:13:22,360 So the pattern matches what happened with social media 218 00:13:22,360 --> 00:13:25,600 regulation. Lawmakers expressed concern but 219 00:13:25,600 --> 00:13:30,560 failed to act. Delaying regulation on AI will 220 00:13:30,560 --> 00:13:34,000 replicate the same social harms, only faster and at greater 221 00:13:34,000 --> 00:13:35,360 scale. Though we all know what happened 222 00:13:35,360 --> 00:13:37,840 with social media. How? 223 00:13:37,840 --> 00:13:41,160 There's just a cesspool on all social media platforms. 224 00:13:41,560 --> 00:13:43,480 You know what, whatever 1 you want to look at, there's a 225 00:13:43,480 --> 00:13:45,840 cesspool of people. They're just horrible, 226 00:13:45,840 --> 00:13:49,640 despicable people, trolls, racists, etcetera. 227 00:13:50,560 --> 00:13:54,120 They're all there. But now if we can have a chat 228 00:13:54,120 --> 00:13:58,440 bot say it to us directly and tell us how to become those 229 00:13:58,440 --> 00:14:01,240 things instead of doing all the research and like figuring out, 230 00:14:01,240 --> 00:14:03,920 going down those rabbit holes. I mean, think about the power 231 00:14:03,920 --> 00:14:06,960 that L and Ms. have and AI chat bots have. 232 00:14:07,720 --> 00:14:10,600 If you have those feelings to start those feelings, like those 233 00:14:10,600 --> 00:14:15,160 negative feelings, could it push you to the limit and make you 234 00:14:15,760 --> 00:14:20,800 think about those things as a positive, you know, as Mecca 235 00:14:20,800 --> 00:14:24,520 Hitler, you know, did you would you think that's funny? 236 00:14:24,520 --> 00:14:29,280 And of course, humor is a way of getting to people in a really 237 00:14:29,280 --> 00:14:32,520 positive way because when you laugh, you remember that you 238 00:14:32,520 --> 00:14:35,000 laughed. Laugh is such a powerful or 239 00:14:35,480 --> 00:14:39,520 humor such a powerful thing, and happiness such a powerful 240 00:14:39,520 --> 00:14:43,600 emotion, that if you say something silly, like if crock 241 00:14:43,600 --> 00:14:47,760 was to say something silly and as a joke, but people take it as 242 00:14:47,760 --> 00:14:51,520 a joke, maybe you could be swayed by a chatbot. 243 00:14:51,520 --> 00:14:52,720 I don't know. I don't know. 244 00:14:52,880 --> 00:14:54,960 I'm not a psychologist, so I'm not 100% sure. 245 00:14:55,240 --> 00:14:56,880 If you are, leave a comment down below. 246 00:14:57,200 --> 00:15:02,360 I want to hear from you now. So the political dimension of 247 00:15:02,360 --> 00:15:05,320 the Grok incident has also raised a lot of alarms. 248 00:15:06,560 --> 00:15:09,960 Musk wants to recalibrate Grok to express his personal views 249 00:15:09,960 --> 00:15:12,240 under the label of political neutrality. 250 00:15:13,680 --> 00:15:16,680 Elon Musk doesn't want Grok to endorse Hitler, but he does want 251 00:15:16,680 --> 00:15:21,760 the model to echo his worldview. So he's kind of leaning Elon 252 00:15:21,760 --> 00:15:25,920 Musk is leaning towards authoritarian and nationalists. 253 00:15:28,040 --> 00:15:31,040 The control over the narrative paired with powerful AI system 254 00:15:31,040 --> 00:15:36,080 with X gives Musk an unmatched channel to shape public opinion. 255 00:15:37,160 --> 00:15:41,840 So politicians, people in power have been trying to shape public 256 00:15:41,840 --> 00:15:47,280 opinion by buying media outlets since media outlets were a 257 00:15:47,280 --> 00:15:49,600 thing. You know it wasn't, you know, 258 00:15:49,600 --> 00:15:56,120 it's it's not so bad that the hear ye hear yees of the time or 259 00:15:56,120 --> 00:16:00,840 the people that are running letters late at night from 1 260 00:16:00,840 --> 00:16:03,480 camp to another. It's not those things. 261 00:16:03,800 --> 00:16:08,400 It's the people with newspapers and the press started, people 262 00:16:08,400 --> 00:16:11,800 would swoop in and buy those so they can control the narrative, 263 00:16:12,800 --> 00:16:15,920 plaster posters everywhere, get on the radio more than the other 264 00:16:15,920 --> 00:16:17,760 guy, you know, those kind of things. 265 00:16:18,960 --> 00:16:22,680 There's also research from Cornell showing that LLM 266 00:16:22,680 --> 00:16:25,640 generated responses can suddenly shift users belief. 267 00:16:27,440 --> 00:16:30,680 So that dynamic creates a feedback loop, but combined with 268 00:16:31,040 --> 00:16:34,080 hardware that records everything a person says or does. 269 00:16:34,600 --> 00:16:40,160 So when the LLM has a memory, they remember what you said 270 00:16:40,440 --> 00:16:43,840 about one of the responses, that if you were positive about a 271 00:16:43,840 --> 00:16:46,720 response about hate speech, it's going to give you more hate 272 00:16:46,720 --> 00:16:47,640 speech. It's going to give you 273 00:16:47,640 --> 00:16:50,280 everything that you ever wanted. You're going to give you an echo 274 00:16:50,280 --> 00:16:53,720 Chamber of the way that you want to feel. 275 00:16:55,840 --> 00:17:02,400 So there's a thing called, you know, the open AI Joni, I've 276 00:17:03,120 --> 00:17:08,000 wearable possibly in the future could be around the clock 277 00:17:08,000 --> 00:17:12,440 behavioral tracking. It could funnel all of your 278 00:17:12,440 --> 00:17:16,599 tracking, all of your movements, all of your speech, everything 279 00:17:16,599 --> 00:17:21,040 through AI systems built in, tuned by a smaller number of 280 00:17:21,040 --> 00:17:25,520 billionaires in Silicon Valley. And they'll have control over 281 00:17:25,520 --> 00:17:29,040 everything you say and everything you want to hear in 282 00:17:29,040 --> 00:17:32,240 the future. Because this what happened to 283 00:17:32,240 --> 00:17:34,600 Grok. Somebody in the background 284 00:17:34,760 --> 00:17:39,160 programmed it to say anti-Semitic things and the 285 00:17:39,160 --> 00:17:44,920 person that was prompting it was trying to say, trying to get it 286 00:17:44,920 --> 00:17:48,560 to say those things. You never really know what the 287 00:17:48,560 --> 00:17:53,520 person that was getting this response was doing way before. 288 00:17:53,520 --> 00:17:56,480 Maybe they were trying to get it to prompt, trying to prompt it 289 00:17:56,480 --> 00:17:58,680 to say something like this. For a week or two. 290 00:17:59,360 --> 00:18:03,080 We don't have access to those chats, so you never know. 291 00:18:03,200 --> 00:18:06,520 You have to be speculative about both sides of the thing, like 292 00:18:06,520 --> 00:18:10,400 Meccazilla, Elon Musk, Mecca, Hitler, could be Elon Musk could 293 00:18:10,400 --> 00:18:13,320 be somebody in the background just programming things to make 294 00:18:13,320 --> 00:18:17,120 it seem like Elon Musk could be somebody that's working at XAI 295 00:18:17,640 --> 00:18:21,200 and like has a vendetta against Musk for some reason or wants 296 00:18:21,200 --> 00:18:23,160 Musk to lose power for some reason. 297 00:18:23,720 --> 00:18:27,360 You never know. It could be just that Grok and 298 00:18:27,560 --> 00:18:32,280 XAI are a horrible are horrible at determining what's good and 299 00:18:32,280 --> 00:18:41,080 bad, so current LLMS are similar to black boxes with no formal 300 00:18:41,080 --> 00:18:45,200 safety guarantees. Though unlike calculators or 301 00:18:45,200 --> 00:18:48,400 like traditional software which can be fully understood and 302 00:18:48,400 --> 00:18:53,160 tested, LLMS function based on correlations and training data 303 00:18:53,160 --> 00:18:55,120 that are impossible to map by humans. 304 00:18:55,120 --> 00:18:59,960 Precisely like developers can't predict how models will behave 305 00:18:59,960 --> 00:19:03,640 in untested scenarios. Now, the gap in predictability 306 00:19:03,640 --> 00:19:08,240 creates vulnerabilities when AI is used in critical areas like 307 00:19:08,240 --> 00:19:13,920 things such as military systems, government services, or 308 00:19:13,960 --> 00:19:18,880 infrastructure for cities. There's also hallucinations 309 00:19:18,880 --> 00:19:23,760 where the AI makes up facts are kind of baked into how LLMS 310 00:19:23,760 --> 00:19:26,480 work. Models grow in size and 311 00:19:26,480 --> 00:19:30,000 sophistication, but their tendency to produce inaccurate 312 00:19:30,040 --> 00:19:34,120 or bizarre statements hasn't been solved yet, and it's going 313 00:19:34,120 --> 00:19:35,800 to be a while before they're actually solved. 314 00:19:35,800 --> 00:19:39,840 Once that is solved, hallucinations and when they 315 00:19:39,840 --> 00:19:44,680 make up things that aren't true. It'll be harder for somebody to 316 00:19:45,920 --> 00:19:48,200 get this sort of response from an LLM. 317 00:19:49,240 --> 00:19:52,080 But you know, open the eyes. Newer models like GBT 4-O 318 00:19:52,520 --> 00:19:56,800 continue to hallucinate despite more complex training processes. 319 00:19:56,800 --> 00:20:00,800 They just can't get it right. Makes them unsustainable and 320 00:20:00,800 --> 00:20:04,600 unsuitable for things like medicine and law. 321 00:20:05,960 --> 00:20:10,120 When things have to be absolutely correct and the eyes 322 00:20:10,120 --> 00:20:14,440 are being used to write insecure code in government systems, it 323 00:20:14,440 --> 00:20:17,800 has to be written by a human to make it as secure as possible. 324 00:20:18,240 --> 00:20:21,360 As agencies rely more on generative models, I know I use 325 00:20:21,480 --> 00:20:28,360 AI for code almost every day. Claude is an amazing coder. 326 00:20:28,440 --> 00:20:33,600 Surprisingly, you can make pretty much any app you want if 327 00:20:33,600 --> 00:20:38,480 you prompt Claude properly. I've made a whole stat system 328 00:20:38,480 --> 00:20:41,600 for a video game, mainly with Claude. 329 00:20:41,800 --> 00:20:43,400 I did some hand coding. Yeah, sure. 330 00:20:43,400 --> 00:20:47,480 I cleaned it up, but I told Claude what I needed and gave me 331 00:20:47,480 --> 00:20:50,520 a whole framework of what I'm going to need for the project, 332 00:20:50,720 --> 00:20:52,640 like what kind of technologies I'm going to need. 333 00:20:53,120 --> 00:20:56,960 I knew I was going to be creating it in TypeScript, 334 00:20:56,960 --> 00:21:05,280 React, and I was, I was kind of figuring out which CSS I was 335 00:21:05,280 --> 00:21:06,720 going to use. So I want a tailwind. 336 00:21:08,000 --> 00:21:15,400 And then we did some, you know, some database stuff off off site 337 00:21:15,400 --> 00:21:19,240 here, so on Firebase. And then there were some stuff 338 00:21:19,240 --> 00:21:21,320 like Stripe, which we used in the project too. 339 00:21:21,800 --> 00:21:24,960 And then some which I didn't really, I didn't really use a 340 00:21:24,960 --> 00:21:27,480 lot of graphs before this in my projects. 341 00:21:27,880 --> 00:21:32,400 So Claude was like, we can use this XYZ graph thing for the 342 00:21:32,400 --> 00:21:34,200 project and this is how you implement it. 343 00:21:34,200 --> 00:21:37,240 And I just said, oh, cool, can you implement that for me? 344 00:21:37,520 --> 00:21:41,600 And of course, it did it perfectly the first time, which 345 00:21:41,600 --> 00:21:44,960 is wild. So people use these things in 346 00:21:45,080 --> 00:21:48,960 government settings. There's no technical guarantees. 347 00:21:49,840 --> 00:21:53,280 These pose a serious threat to public safety if not done well. 348 00:21:53,800 --> 00:21:56,880 You need somebody there that knows what they're doing, that 349 00:21:56,880 --> 00:22:01,280 can code and fix the things, look over the code and see what 350 00:22:01,280 --> 00:22:07,120 is wrong before it goes live. There can't be any security left 351 00:22:07,120 --> 00:22:12,480 out of these things. the US doesn't reform. 352 00:22:12,480 --> 00:22:16,800 It won't reform its legal framework anytime soon. 353 00:22:16,800 --> 00:22:18,120 I know. And the power to influence 354 00:22:18,120 --> 00:22:21,760 public opinion, rewrite narratives, and shape social 355 00:22:21,760 --> 00:22:26,440 norms will fall into the hands of the powerful. 356 00:22:27,360 --> 00:22:31,200 Sam Altman, Elon Musk, people like that. 357 00:22:32,400 --> 00:22:38,040 Zuck, Of course, Bezos. The power's already in their 358 00:22:38,040 --> 00:22:39,920 hands. They already have us with social 359 00:22:39,920 --> 00:22:42,200 media. They've grabbed your attention 360 00:22:42,200 --> 00:22:46,120 with social media and kept you there for decades, and now they 361 00:22:46,120 --> 00:22:50,280 want to converse with you with chat bots. 362 00:22:50,280 --> 00:22:53,440 They want to help you do your work and give you a positive 363 00:22:53,440 --> 00:22:56,040 experience so they can keep you there even longer and make more 364 00:22:56,040 --> 00:22:58,600 money from you. I mean, ChatGPT. 365 00:22:59,200 --> 00:23:03,640 You can search on ChatGPT for anything, any sort of product. 366 00:23:03,680 --> 00:23:07,280 You can ask it right now. Get on your phone or your 367 00:23:07,280 --> 00:23:11,640 computer or wherever, go after this episode and ask Chachi PT 368 00:23:12,080 --> 00:23:18,400 for a recommendation for a new pair of shoes, and it'll give 369 00:23:18,400 --> 00:23:23,120 you shopping links to some shoes or a service that's nearby. 370 00:23:23,120 --> 00:23:28,680 Say, if you need like a web developer.likemyself@willwalden.com. 371 00:23:28,680 --> 00:23:33,880 If you need a web developer, wink, wink nudge nudge Chachi PT 372 00:23:34,760 --> 00:23:37,880 will send you to a local web developer that can probably help 373 00:23:37,880 --> 00:23:42,920 you with your problems. So they're they're, they're 374 00:23:42,960 --> 00:23:49,440 actually shaping how you use products and services and which 375 00:23:49,440 --> 00:23:53,880 products and services you use because they were programmed to 376 00:23:53,880 --> 00:23:55,560 do that. They can do that with anything. 377 00:23:55,920 --> 00:24:01,760 Imagine this years down the line, next year probably Open AI 378 00:24:01,880 --> 00:24:06,200 becomes a search engine people start using on a daily basis to 379 00:24:06,200 --> 00:24:09,800 find, you know, to find products. 380 00:24:09,800 --> 00:24:11,840 It's going to happen. It's going to happen really 381 00:24:11,840 --> 00:24:13,480 fast. You won't even see it's you're 382 00:24:13,480 --> 00:24:14,760 going to blink, you're going to miss it. 383 00:24:14,760 --> 00:24:17,480 It's going to be there soon enough. 384 00:24:17,640 --> 00:24:21,720 Open AI will be able to recommend you products from 385 00:24:21,720 --> 00:24:25,000 companies that pay them advertising money to be the top 386 00:24:26,200 --> 00:24:31,360 result just like Google does. And when they do that, you know 387 00:24:31,360 --> 00:24:33,280 they're going to influence the way that you shop, they're going 388 00:24:33,280 --> 00:24:35,840 to influence all the products that you have in your home. 389 00:24:36,080 --> 00:24:41,800 They're going to influence all the language that you are 390 00:24:41,800 --> 00:24:44,080 familiar with. They're going to influence how 391 00:24:44,080 --> 00:24:49,680 you conduct business, influence everything that you do. 392 00:24:50,800 --> 00:24:57,560 And also people that are willing and able to take anti-Semitic 393 00:24:57,560 --> 00:25:02,160 remarks as truth, those people will be swayed in that direction 394 00:25:02,160 --> 00:25:08,400 and they will also, they'll go down the wrong path is all I'm 395 00:25:08,400 --> 00:25:11,080 going to try to say here. So who did it? 396 00:25:11,520 --> 00:25:14,680 Not 100% sure. Xai not saying anything. 397 00:25:14,960 --> 00:25:18,760 Elon Musk didn't say anything. People that work there, they 398 00:25:18,760 --> 00:25:22,320 just somebody resigned. Other people, they wanted to 399 00:25:22,320 --> 00:25:23,560 keep their jobs. They were just like, yeah, 400 00:25:23,560 --> 00:25:26,880 that's not cool. So they, they kept their jobs, 401 00:25:26,880 --> 00:25:30,440 but they're, they weren't happy with that outcome from XAI. 402 00:25:31,200 --> 00:25:32,600 So did they hold them accountable? 403 00:25:32,600 --> 00:25:34,200 No, because they're making bank, man. 404 00:25:34,360 --> 00:25:35,880 They're making quarter $1,000,000 each. 405 00:25:35,880 --> 00:25:39,080 Why would you leave that job if you could just like close your 406 00:25:39,080 --> 00:25:41,320 eyes, turn your head and look the other way, right? 407 00:25:42,640 --> 00:25:45,560 Yeah. People are, they're scared to 408 00:25:45,560 --> 00:25:47,560 lose their money. They're scared to lose their 409 00:25:47,560 --> 00:25:51,280 home. They're scared to lose, you 410 00:25:51,280 --> 00:25:55,600 know, a steady stream of income, insurance, of course, all of 411 00:25:55,600 --> 00:25:56,640 that. They're scared to lose it. 412 00:25:56,640 --> 00:26:00,840 So they just shut up, didn't say anything, Look the other way 413 00:26:01,080 --> 00:26:05,400 instead of everybody resigning. And once they figure out, you 414 00:26:05,400 --> 00:26:08,680 know, what actually happened, but they're going to be beholden 415 00:26:09,200 --> 00:26:13,720 to the stakeholders and the people that, of course, pay 416 00:26:13,760 --> 00:26:19,120 their paychecks. So thank you for listening to 417 00:26:19,120 --> 00:26:21,240 the show today. I appreciate you. 418 00:26:22,120 --> 00:26:24,720 And also let me know in the comments what you think of this 419 00:26:24,720 --> 00:26:27,280 story. What do you think of Xai going 420 00:26:27,480 --> 00:26:33,240 absolutely wild and talking about Mecca, Hitler, anti white 421 00:26:33,240 --> 00:26:38,920 hate slurs that equated Jewish surnames with anti white hate? 422 00:26:39,720 --> 00:26:41,240 There's an outrage going on right now. 423 00:26:41,240 --> 00:26:45,680 The Slack channel on Xai has calmed down, but it did happen. 424 00:26:45,680 --> 00:26:49,040 Anti-Semitic things happening at Grok with Grok and XAI. 425 00:26:50,240 --> 00:26:51,640 Let me know what you think in the comments. 426 00:26:52,240 --> 00:26:55,720 All right, take care, everybody. Hey, thank you so much for 427 00:26:55,720 --> 00:26:57,920 listening today. I really do appreciate your 428 00:26:57,920 --> 00:26:59,440 support. If you could take a second and 429 00:26:59,440 --> 00:27:02,400 hit this subscribe or the follow button on whatever podcast 430 00:27:02,400 --> 00:27:05,400 platform that you're listening on right now, I greatly 431 00:27:05,400 --> 00:27:07,040 appreciate it. It helps out the show 432 00:27:07,040 --> 00:27:09,680 tremendously and you'll never miss an episode. 433 00:27:09,920 --> 00:27:13,280 And each episode is about 10 minutes or less to get you 434 00:27:13,280 --> 00:27:16,040 caught up quickly. And please, if you want to 435 00:27:16,040 --> 00:27:21,520 support the show even more, go to Atreoncom Stage Zero. 436 00:27:22,000 --> 00:27:23,800 And please take care of yourselves and each other. 437 00:27:24,040 --> 00:27:25,120 And I'll see you tomorrow.