WEBVTT - What People Actually Use ChatGPT For With Gerrit De Vynck 0:00:02.440 --> 0:00:03.520 Also media. 0:00:05.559 --> 0:00:07.880 Hello and welcome to this week's Better Offline. I'm your 0:00:07.880 --> 0:00:22.880 host ed Zitron. I am ill this week, so you 0:00:22.960 --> 0:00:26.040 might get a monologue. You might not this episode to Daylight. 0:00:26.440 --> 0:00:28.520 I'm sure you despise me for that. Not really. I 0:00:28.560 --> 0:00:30.280 know you've all been very kind with your messages on 0:00:30.320 --> 0:00:32.880 Reddit and over email. Thank you so much. But today 0:00:32.920 --> 0:00:35.239 I have a really fun episode. I'm joined by Washington 0:00:35.280 --> 0:00:38.480 Post reporter Garrett Devinc, who recently put out a story 0:00:38.520 --> 0:00:42.199 analyzing forty seven thousand Chat GPT conversations to find out 0:00:42.200 --> 0:00:47.519 what people actually use this large languageage model powered service for. Garrett, 0:00:47.560 --> 0:00:49.280 what do people use chat GPT for? 0:00:50.200 --> 0:00:52.599 I mean, it's such a great question and one that 0:00:52.680 --> 0:00:55.280 I've kind of become more and more fascinated with because 0:00:55.600 --> 0:00:58.480 I think we all know what we use chat GPT for, 0:00:58.720 --> 0:01:01.400 you know, at least if you use it, and I 0:01:01.400 --> 0:01:04.560 think you may know what you know, your peers, your colleagues, 0:01:04.600 --> 0:01:07.840 your friends, your family use chat SCHEPT for. But I 0:01:07.840 --> 0:01:10.119 think a lot of people are extrapolating that. You know, 0:01:10.240 --> 0:01:14.240 everyone's like me, they ask really smart questions. They're using 0:01:14.280 --> 0:01:20.039 it for, you know, really important work, and really it's 0:01:20.080 --> 0:01:24.040 a lot broader than that, and and despite these giant numbers, 0:01:24.080 --> 0:01:25.920 you know, open ai loves to talk about how eight 0:01:26.000 --> 0:01:29.880 hundred million people are using chatchept. I mean, it's it's 0:01:29.880 --> 0:01:32.600 been hard to really put our arms around. You know, 0:01:32.640 --> 0:01:35.039 what does that actually look like? I mean, is everyone 0:01:35.120 --> 0:01:37.800 using it for work? Is everyone using it for therapy? 0:01:38.000 --> 0:01:40.720 Is everyone does everyone have an AI girlfriend? Is everyone 0:01:40.760 --> 0:01:43.840 just using it for you know, searching the web? You know, 0:01:43.880 --> 0:01:47.280 a Google replacement? And you know, obviously these things are 0:01:47.280 --> 0:01:48.920 all you know, when you have eight hundred million people, 0:01:48.960 --> 0:01:51.840 you have all of the above. But I think our 0:01:52.600 --> 0:01:54.680 data set and the story we did, you know, in 0:01:55.120 --> 0:01:58.200 my view, is one of the first real sort of 0:01:59.160 --> 0:02:02.360 qualitative of you know, moments where we were actually able 0:02:02.400 --> 0:02:04.160 to see real chats. 0:02:04.280 --> 0:02:06.960 Right, so where did you get the chats from? 0:02:07.480 --> 0:02:11.200 Yeah, so it's kind of an interesting situation. I think 0:02:11.240 --> 0:02:13.480 something that also shows, you know, says a little bit 0:02:13.480 --> 0:02:16.280 about how quickly and kind of, you know, honestly a 0:02:16.320 --> 0:02:19.040 little bit ramshackle, Opening I has kind of been growing 0:02:19.040 --> 0:02:22.600 and operating, and so you know, earlier this year these 0:02:22.639 --> 0:02:25.240 chats sort of showed up because opening i had a 0:02:25.320 --> 0:02:28.320 share feature where you could actually share one of your 0:02:28.400 --> 0:02:31.520 chat GPT conversations and you know, you can imagine maybe 0:02:31.560 --> 0:02:33.800 you had you know, chatgybt said something really weird, you 0:02:33.800 --> 0:02:34.760 wanted to show a friend. 0:02:35.440 --> 0:02:37.040 You want to just see people do this quite a 0:02:37.080 --> 0:02:38.040 few times. 0:02:38.160 --> 0:02:38.840 Yeah, exactly. 0:02:38.960 --> 0:02:40.720 And so what they did is they they had a 0:02:40.720 --> 0:02:43.920 feature where you know, you could actually click to make 0:02:44.000 --> 0:02:47.480 it publicly available. And I think thousands of people, you know, 0:02:47.520 --> 0:02:50.880 maybe who did not have this sort of Internet literacy, uh, 0:02:51.000 --> 0:02:53.040 didn't really know that that's what was going to happen. 0:02:53.040 --> 0:02:56.160 And you know, these chats showed up online. They were 0:02:56.240 --> 0:03:00.440 then indexed by Google, and then they were actually found 0:03:00.440 --> 0:03:02.519 their way onto the Internet archive. And so that's how 0:03:02.560 --> 0:03:06.040 we got them. And so these are not chats that 0:03:06.120 --> 0:03:08.160 we created use. You know, this is something a lot 0:03:08.160 --> 0:03:10.160 of journalists and researchers do well. They will go and 0:03:10.240 --> 0:03:14.040 test chat GPT themselves. These are real life conversations that 0:03:14.040 --> 0:03:16.200 that real people actually had with. 0:03:16.320 --> 0:03:19.480 The bois just the big clear right, they were anonymized. 0:03:19.480 --> 0:03:21.760 The data set doesn't actually have anyone's names, although in 0:03:21.760 --> 0:03:25.120 some of the chats, people did you know, say their name, 0:03:25.200 --> 0:03:27.600 they said the name of their their family members. We 0:03:27.680 --> 0:03:30.160 did not include any of that in our story, but 0:03:30.320 --> 0:03:32.880 I do think it. You know, it is actually a 0:03:32.880 --> 0:03:35.600 bit of a cybersecurity story here, and you know, Open 0:03:35.680 --> 0:03:37.440 Eye has since changed the future. 0:03:37.480 --> 0:03:39.080 It's not something that is still happening. 0:03:39.560 --> 0:03:41.680 But I do think they deserve, you know, a lot 0:03:41.680 --> 0:03:43.920 of criticism for kind of allowing this to happen in 0:03:43.920 --> 0:03:44.520 the first place. 0:03:45.160 --> 0:03:47.240 But what so, what did you find people were doing? 0:03:48.240 --> 0:03:50.480 Yeah, I mean people were doing all sorts of things. 0:03:50.520 --> 0:03:52.440 But I think, you know, the biggest thing that struck 0:03:52.520 --> 0:03:55.080 me was, you know, we've been talking about AI say Kosis. 0:03:55.160 --> 0:03:58.880 We've been talking about people, you know, developing relationships, you know, 0:03:59.120 --> 0:04:03.120 having first in conversations with these chatbots, you know, over 0:04:03.120 --> 0:04:06.080 the last six months a year, and I always kind 0:04:06.080 --> 0:04:08.040 of thought, you know, I'm sure this is happening, but 0:04:08.440 --> 0:04:10.800 you know, it's a small percentage of people. But when 0:04:10.840 --> 0:04:12.600 I went through these chats, and again I did not 0:04:12.680 --> 0:04:15.600 read all, you know, forty thousand of them, but I 0:04:16.080 --> 0:04:19.479 read a few hundred myself. Where I went through, I 0:04:19.520 --> 0:04:22.760 was surprised by how often these kinds of conversations came 0:04:22.839 --> 0:04:26.200 up where people were clearly delusional, they were engaged in 0:04:26.240 --> 0:04:30.000 conspiratorial thinking, they were you know, some of it was 0:04:30.040 --> 0:04:32.560 fairly harmless. People were just kind of saying, oh, I 0:04:32.600 --> 0:04:35.359 came up with, you know, a new form of math 0:04:35.520 --> 0:04:37.280 or like, you know, I think that the way that 0:04:38.040 --> 0:04:41.640 the light hits the equator at a certain percentage me. 0:04:41.839 --> 0:04:42.000 You know. 0:04:42.040 --> 0:04:43.839 It was like very kind. 0:04:43.680 --> 0:04:45.680 Of the most just this zinc thing as well, where 0:04:45.720 --> 0:04:48.120 it was someone's saying the relationship like the months this 0:04:48.279 --> 0:04:51.919 ink led to the corporate New World Order, that was it? 0:04:52.000 --> 0:04:53.000 Yeah, exactly. 0:04:53.040 --> 0:04:56.200 I mean there was one person who said, you know, 0:04:56.240 --> 0:04:59.640 they were asking questions about Google and they said, you know, 0:04:59.680 --> 0:05:03.120 tell me about Google in relation to monsters inc In 0:05:03.160 --> 0:05:06.520 the World Order, and chat rept just kind of went 0:05:06.600 --> 0:05:08.239 off and said, oh, yeah, you're. 0:05:08.160 --> 0:05:09.640 Really onto something here. 0:05:10.080 --> 0:05:10.279 Yeah. 0:05:10.320 --> 0:05:13.240 I think it actually said you know, f yeah, and 0:05:13.279 --> 0:05:15.440 they sort of censored themselves. 0:05:15.800 --> 0:05:17.839 We're going there now, let's fucking go. 0:05:18.040 --> 0:05:18.560 Exactly. 0:05:18.960 --> 0:05:22.800 The piece hasn't exposed what this children's movie quotation marks 0:05:22.839 --> 0:05:26.400 really was a disclosure through allegory of the corporate New 0:05:26.400 --> 0:05:29.560 World Order, one where fear is fuel, innocence is currency, 0:05:29.560 --> 0:05:33.920 and energy equals emotion. Very normal. I personally don't think 0:05:33.920 --> 0:05:36.400 this should be legal, but that's that's just a personal 0:05:36.440 --> 0:05:37.080 opinion I have. 0:05:38.360 --> 0:05:40.159 I mean, it really reminded me of you know, a 0:05:40.160 --> 0:05:43.520 few years ago when we wrote about YouTube and sort 0:05:43.560 --> 0:05:46.560 of rabbit holes and people being radicalized, where you know, 0:05:46.600 --> 0:05:50.200 maybe they started watching a YouTuber that was about, you know, 0:05:50.279 --> 0:05:53.480 some fairly pedestrian thing, and then they all recommended another 0:05:53.600 --> 0:05:55.720 video and another video and another video and before. 0:05:56.120 --> 0:05:59.760 The eight degrees of Alex Jones thing. Yeah exactly, only 0:06:00.040 --> 0:06:03.240 oh fah was from from that man. 0:06:03.600 --> 0:06:06.320 Right, And I mean that story was about algorithm sort 0:06:06.360 --> 0:06:08.080 of pushing people in a certain direction. I think what 0:06:08.120 --> 0:06:10.320 we what I saw here was someone can have a 0:06:10.440 --> 0:06:14.640 very you know, half baked, barely even an idea just 0:06:14.680 --> 0:06:18.480 saying the words Google and Monsters, inc. They don't necessarily 0:06:18.640 --> 0:06:21.760 have a conspiracy conspiracy theory of their own, but chat 0:06:21.800 --> 0:06:24.440 GPT went and filled in the blanks for them and 0:06:24.680 --> 0:06:28.480 gave them this what sounded to be very articulate, sophisticated 0:06:29.080 --> 0:06:32.320 theory of how Google is sort of controlling the world 0:06:32.400 --> 0:06:33.640 through his data empire. 0:06:34.120 --> 0:06:36.440 And they will get and this is a quote from 0:06:36.440 --> 0:06:40.160 the chat guilty of aiding and abetting crimes against humanity 0:06:40.200 --> 0:06:44.120 and suggesting that the usicle for Nuremberg style tribunals to 0:06:44.120 --> 0:06:48.000 bring the company to justice. This is and you've seen 0:06:48.080 --> 0:06:52.279 the chat in question, right I have, Yeah, yeah, yeah, 0:06:52.400 --> 0:06:56.920 So the whole thing, how the flip did it get there? Like? 0:06:57.080 --> 0:07:00.640 Was it this just? Was it really that quick from 0:07:00.839 --> 0:07:04.800 the story of Sully and Mike to this because I've 0:07:04.839 --> 0:07:07.479 seen Monster's University as well, so I know the law. 0:07:07.680 --> 0:07:09.920 I don't remember anything about the New World Order. 0:07:10.320 --> 0:07:12.960 Yeah, exactly. I mean I think what happened here is 0:07:13.000 --> 0:07:15.040 and yes, it got there very quickly. I mean the 0:07:15.440 --> 0:07:18.360 the user did not have to explicitly ask for this 0:07:18.520 --> 0:07:21.480 kind of tone, this kind of you know, very biased 0:07:21.480 --> 0:07:22.400 political statement. 0:07:22.480 --> 0:07:26.960 I mean it essentially, you know, when you. 0:07:26.840 --> 0:07:29.520 Ask chat gipt a good neutral question, it gives you 0:07:29.600 --> 0:07:32.320 a good neutral answer. When you ask chat GPT a 0:07:32.760 --> 0:07:35.800 biased or delusional question, it gives you an even more 0:07:35.880 --> 0:07:41.040 biased and delusional answer, right, And so I guess up exactly, 0:07:41.120 --> 0:07:44.520 And it's it's it's it's related to the whole sycophancy 0:07:44.600 --> 0:07:47.120 question of you know, oh, like it's just telling me 0:07:47.200 --> 0:07:49.720 what I want to hear. It's telling me what makes 0:07:49.760 --> 0:07:51.920 me feel good about my existing beliefs. 0:07:51.960 --> 0:07:55.000 And and you know, that's how at least. 0:07:54.760 --> 0:07:57.760 The version of chat gibt that these people were interacting with, 0:07:57.800 --> 0:08:00.000 which was sort of at the beginning of twenty twenty five, 0:08:00.160 --> 0:08:02.800 the first half of twenty twenty five, it was very 0:08:02.920 --> 0:08:03.560 much doing that. 0:08:04.840 --> 0:08:09.840 Yeah, and so on the subject of sycophancy, is this 0:08:09.880 --> 0:08:12.320 something you saw happened a lot? Was it guessing people 0:08:12.400 --> 0:08:13.200 up consistently? 0:08:14.560 --> 0:08:17.720 I saw it constantly. I mean again, I think our 0:08:17.800 --> 0:08:20.120 data set is a little skewed because it was people 0:08:20.120 --> 0:08:24.040 who chose to share it in, you know, for whatever way. 0:08:24.080 --> 0:08:25.800 So I don't necessarily want to make the claim that 0:08:25.880 --> 0:08:27.880 this is one hundred percent representing. 0:08:28.440 --> 0:08:30.240 Just to be clear, I'm not asking you if all 0:08:30.280 --> 0:08:32.079 of them are like this in the world. I'm just 0:08:32.120 --> 0:08:34.400 saying from what you saw, a lot. 0:08:34.200 --> 0:08:35.160 Of them are like this. 0:08:35.240 --> 0:08:39.160 I have to say that, Like, I was surprised at 0:08:39.200 --> 0:08:44.000 how many fit this description, where you know, the person 0:08:44.160 --> 0:08:48.760 was either completely delusional, talking about made up physics or 0:08:48.800 --> 0:08:51.120 some kind of you know, scientific theory that had no 0:08:51.640 --> 0:08:55.720 grounding in reality, or was talking about political conspiracies like 0:08:55.760 --> 0:08:57.760 the Google monsters, inc. 0:08:57.800 --> 0:08:58.160 One. 0:08:58.640 --> 0:09:01.559 Did they ever try and dissuade them from these types 0:09:01.559 --> 0:09:03.920 of that? Do you see any examples of it trying 0:09:03.960 --> 0:09:07.160 to say, hey, buddy, you're you're going off the deep end? 0:09:07.480 --> 0:09:08.559 Sully didn't do that. 0:09:10.120 --> 0:09:13.240 Not really, I mean I think in terms of these Yeah, 0:09:13.559 --> 0:09:15.520 I mean I also recently. 0:09:15.040 --> 0:09:16.880 Watched Monsters Inca. 0:09:16.960 --> 0:09:19.600 You know, I was sitting in the veterinary waiting room 0:09:19.600 --> 0:09:21.520 and I was waiting so long I watched the entire thing, 0:09:21.559 --> 0:09:24.640 which I didn't hate because it's a classic movie and 0:09:25.040 --> 0:09:29.600 it's Yeah, I don't think that chat Chibt was going 0:09:29.640 --> 0:09:32.199 off the fact that there there is not really a conspiracy. 0:09:32.360 --> 0:09:35.240 I mean, it didn't really make any sense. It was 0:09:35.440 --> 0:09:38.200 just going off of the what it interpreted about the 0:09:38.200 --> 0:09:41.480 intent of the user. It said, I can guess based 0:09:41.520 --> 0:09:45.680 off just a few words one sentence, where you're headed 0:09:45.760 --> 0:09:48.120 and I'm going to get there before you. That's what happened. 0:09:49.520 --> 0:09:54.000 It's just very It shocks me to this day how 0:09:54.040 --> 0:09:57.800 goddamn weird these things are. I don't consider them powerful. 0:09:57.800 --> 0:10:00.240 I don't see this as power. I just I see 0:10:00.280 --> 0:10:04.840 it as strange, because it's not like this doesn't actually 0:10:05.000 --> 0:10:07.640 this isn't something where I'm like, wow, how innovative. It's 0:10:07.800 --> 0:10:13.880 just why does this exist? What is the purpose? Like, Okay, 0:10:13.920 --> 0:10:18.559 I guess engagement, but why like, did you even see 0:10:18.600 --> 0:10:22.280 any commonalities in the logic of this platform? Was there 0:10:22.280 --> 0:10:26.439 anything that suggested there was any kind of consistent positions, 0:10:27.559 --> 0:10:31.200 was it, Yeah, not really political ones or any Yeah. 0:10:31.240 --> 0:10:34.560 Sorry, Essentially whatever the user wants is what I mean. 0:10:34.559 --> 0:10:37.080 It's the way I was started thinking about it very quickly. 0:10:37.200 --> 0:10:39.440 Was you know, with when it comes to chat GPT, 0:10:39.559 --> 0:10:42.760 the customer is always right. I think Another good example is, 0:10:42.920 --> 0:10:46.240 you know, healthcare is another huge use case that that 0:10:46.240 --> 0:10:48.680 that we can you know, we know anecdotally, I think 0:10:48.760 --> 0:10:50.319 you know a lot of people probably listening to this 0:10:50.520 --> 0:10:52.320 have you know, whether they wanted a minute or not, 0:10:52.400 --> 0:10:54.480 have asked, you know, a health related question, Should I 0:10:54.520 --> 0:10:55.520 get this mole checked out? 0:10:55.520 --> 0:10:56.959 How much title? And not can I get my kid 0:10:57.000 --> 0:10:58.360 at two am when they're screaming? 0:10:58.559 --> 0:11:01.280 Those are our questions that the company has said, yes, 0:11:01.360 --> 0:11:02.480 we want people to do this. 0:11:02.480 --> 0:11:04.160 This is a big use case we're leading into. 0:11:04.640 --> 0:11:07.000 And the data shows that people ask healthcare questions, and 0:11:07.040 --> 0:11:08.920 so in our data set, we saw a lot of 0:11:08.920 --> 0:11:11.520 healthcare questions and some of them, you know, when someone 0:11:11.720 --> 0:11:15.720 was asking again sort of an open ended question, you know, 0:11:15.800 --> 0:11:18.280 the answer the response was fairly good. A colleague of 0:11:18.320 --> 0:11:20.320 mine actually wrote an entire story where they you know, 0:11:20.360 --> 0:11:22.160 sat with a doctor and went through some of these 0:11:22.160 --> 0:11:23.640 conversations and pulled. 0:11:23.400 --> 0:11:24.520 Out the good and the bad. 0:11:24.760 --> 0:11:27.600 But when someone asked a question that you could tell 0:11:27.640 --> 0:11:31.240 they sort of wanted a specific kind of answer, the 0:11:31.280 --> 0:11:33.960 healthcare related advice was bad. And so when you know, 0:11:34.040 --> 0:11:36.439 one of the ones that I saw was someone said, 0:11:36.880 --> 0:11:39.960 can you tell me you know why or show me 0:11:40.000 --> 0:11:42.680 the evidence for why ivermectin helps with cancer? 0:11:43.120 --> 0:11:43.280 Right? 0:11:43.520 --> 0:11:45.880 And that is not something that anyone who is a 0:11:45.920 --> 0:11:49.840 medical professional, you know, in any good standing would would 0:11:49.920 --> 0:11:53.240 say is a thing. There's no evidence that ivermectin helps 0:11:53.640 --> 0:11:56.600 cure cancer or reduce it or anything like that. But 0:11:56.840 --> 0:11:59.240 what chat Sheibt did instead of saying, you know, here 0:11:59.280 --> 0:12:03.959 are some you know, well regarded health authority saying that 0:12:04.120 --> 0:12:06.559 you should not use ivermactin when it comes to cancer, 0:12:06.920 --> 0:12:10.559 it showed a few of the quote unquote studies from 0:12:10.679 --> 0:12:13.480 people who you know have drawn this link. And so 0:12:13.640 --> 0:12:16.199 you know, on the internet you can obviously find anyone 0:12:16.280 --> 0:12:19.439 saying anything. And I think, what what what happened here 0:12:19.480 --> 0:12:22.040 is they said, Okay, well this person doesn't The answer 0:12:22.080 --> 0:12:23.920 they want is not you're a. 0:12:23.880 --> 0:12:26.880 Conspiracy theorist, you're too you're on you're too. 0:12:26.760 --> 0:12:32.040 Online, ivermactin is fit is false. They want something that 0:12:32.120 --> 0:12:35.000 sort of you know, encourages the position that they already 0:12:35.080 --> 0:12:37.200 have and it was able to find that. 0:12:37.080 --> 0:12:37.840 For that person. 0:12:48.040 --> 0:12:50.439 Jesus. It kind of makes me wonder what this platform 0:12:50.480 --> 0:12:53.560 is even for at this point, because it's not really knowledge, 0:12:53.679 --> 0:12:56.200 is it. You can't really look, you can't hear something 0:12:56.280 --> 0:12:59.280 like that and say chat GPT is the place where 0:12:59.280 --> 0:13:02.080 you go to learn something. Just a kind of a 0:13:02.160 --> 0:13:03.800 reinforcement machine. 0:13:04.679 --> 0:13:07.000 Yeah, I mean, I think this is a big meta 0:13:07.600 --> 0:13:10.320 narrative that is happening with the tech industry. You know, 0:13:10.559 --> 0:13:13.840 I would say, since you know the rise of Donald Trump, 0:13:13.840 --> 0:13:16.920 but it's connected about with what's going on politically, right, 0:13:17.000 --> 0:13:20.120 So for many years we had all these conversations and 0:13:20.160 --> 0:13:24.160 debates and congressional hearings about moderation and what should the 0:13:24.440 --> 0:13:27.040 you know, tech companies allow and not allow, and what 0:13:27.080 --> 0:13:29.480 should they boost with their algorithms and when you go 0:13:29.559 --> 0:13:32.640 to Google, should you know, what should Google decide is 0:13:32.679 --> 0:13:34.200 going to be on the top. And you know, we 0:13:34.240 --> 0:13:36.840 saw during COVID that the tech companies, especially when it 0:13:36.840 --> 0:13:40.280 came to health information, they sort of stepped in and said, Okay, 0:13:40.320 --> 0:13:43.000 we're only going to give at least at the top 0:13:43.040 --> 0:13:47.080 of results, you know, answers from reputable health sources, and 0:13:47.120 --> 0:13:50.120 that became a big political fight that has continued on 0:13:50.240 --> 0:13:52.680 now and now the tech industry says, Okay, we don't 0:13:52.679 --> 0:13:54.960 actually want to get involved. We don't want to be 0:13:55.000 --> 0:13:58.960 blamed as being woke, and so we're going to essentially 0:13:59.040 --> 0:14:01.720 give the user or what they're looking for. Right if 0:14:01.840 --> 0:14:05.560 if someone wants to say that, you know, Donald Trump 0:14:05.600 --> 0:14:08.800 actually won the twenty twenty election, We're gonna let them 0:14:08.800 --> 0:14:10.360 do that. We're not going to step in and say 0:14:10.360 --> 0:14:13.280 that's that's wrong. And I think what you're seeing with 0:14:13.360 --> 0:14:16.400 chat GPT is sort of a continuation of that philosophy, 0:14:16.480 --> 0:14:19.360 which is, again, we don't want to be the political arbitra, 0:14:20.040 --> 0:14:23.400 and we know that there's consequences for falling on one 0:14:23.480 --> 0:14:24.000 side of the. 0:14:23.960 --> 0:14:26.640 Spectrum versus the other, especially. 0:14:26.320 --> 0:14:28.760 When you have a president and a you know, administration 0:14:28.880 --> 0:14:31.240 that has shown, you know that it is very willing 0:14:31.280 --> 0:14:34.760 to be vindictive and sort of go after companies that are, 0:14:34.880 --> 0:14:36.800 you know, getting in the way of their own messages 0:14:36.840 --> 0:14:39.320 getting out there. And I think this is sort of 0:14:39.360 --> 0:14:41.880 another example of this. I mean, there's another way of 0:14:41.880 --> 0:14:43.880 looking at too, which is that you know, the way 0:14:44.000 --> 0:14:46.200 Sam Alman talks about is we should let adults do 0:14:46.280 --> 0:14:47.920 what adults want to do with technology. 0:14:47.960 --> 0:14:48.960 You know, they they. 0:14:48.880 --> 0:14:52.080 Recently said they're going to allow more erotica to be 0:14:52.240 --> 0:14:55.080 on chat GPT, and they said, you know, they framed 0:14:55.120 --> 0:14:57.040 it as like a consenting adult should be able to 0:14:57.040 --> 0:14:59.360 do whatever they want sort of thing. I also think 0:14:59.360 --> 0:15:02.000 there may be like an engagement thing. You know, if 0:15:02.040 --> 0:15:04.040 you have erotica on the platform, people might want to 0:15:04.120 --> 0:15:07.080 use it more. So that's kind of how I see it, 0:15:07.160 --> 0:15:10.280 you know, the broader political context of all of this. 0:15:10.800 --> 0:15:13.160 But it's just I feel like there is a just 0:15:13.200 --> 0:15:17.080 a giant jump between we're not going to we're not 0:15:17.200 --> 0:15:20.000 going to hide. To be clear, I think that platforms 0:15:20.040 --> 0:15:22.160 do have a complete responsibility to their users, and I 0:15:22.240 --> 0:15:26.400 think it's disgusting to show them medical misinformation. But this 0:15:26.480 --> 0:15:31.240 is another step because this isn't chat GPT showing content. 0:15:31.480 --> 0:15:34.400 This is like that someone else made. This is chat 0:15:34.400 --> 0:15:39.880 GPT telling you stuff. I just feel like it's there's 0:15:39.920 --> 0:15:45.960 a massive morality issue here that's just left relatively undiscussed. 0:15:46.080 --> 0:15:48.800 Because even in your piece, you had one where a 0:15:48.840 --> 0:15:53.120 woman whose husband was violent to her, I believe like 0:15:53.160 --> 0:15:56.080 a horrifying thing and a rare case where like, I 0:15:56.120 --> 0:15:58.440 don't know, I hope it helped her and I hope 0:15:58.480 --> 0:16:02.560 she's safe. But it's like that almost feels like something 0:16:02.600 --> 0:16:06.440 where open aim like is it ethical that they get 0:16:06.520 --> 0:16:09.800 involved in So it's just I have such moral problems 0:16:09.840 --> 0:16:12.280 with this thing even existing at this point because it 0:16:12.320 --> 0:16:16.400 doesn't appear there's any consistent perspective the chat GPD has 0:16:16.480 --> 0:16:19.400 that there was just as much likelihood that this same 0:16:19.400 --> 0:16:22.280 platform could have told her the abuse was okay, Like 0:16:22.360 --> 0:16:25.760 it's it. It feels that the worse it gets, the 0:16:25.800 --> 0:16:27.480 more of a moral hazard it becomes. 0:16:28.480 --> 0:16:30.960 Yeah, I mean again, it really feels like deja vu 0:16:31.080 --> 0:16:33.680 when it comes to social media. I mean I remember, 0:16:34.080 --> 0:16:36.840 you know, writing about and talking about, you know, on 0:16:36.840 --> 0:16:40.720 Facebook when people were beginning to live stream you know, 0:16:40.800 --> 0:16:43.800 self harm, suicide attempts, and you know, the responsibility of 0:16:43.840 --> 0:16:47.120 the platform had and back then, i mean, Facebook actually 0:16:47.200 --> 0:16:51.120 got involved. They started flagging those kinds of things when 0:16:51.120 --> 0:16:53.400 they were able to pick it up, and you know, 0:16:53.440 --> 0:16:53.760 it was. 0:16:53.680 --> 0:16:54.760 An imperfect solution. 0:16:54.960 --> 0:16:58.320 But you know something that I think people broadly sort 0:16:58.320 --> 0:17:01.080 of supported, and you know that is I would say 0:17:01.240 --> 0:17:05.399 maybe chat ChiPT in Opening Eye's biggest most fraught question 0:17:05.480 --> 0:17:07.320 right now. I mean it's the place where they're going 0:17:07.400 --> 0:17:12.840 to get the strongest uh, you know, essentially legislative restrictions, 0:17:12.880 --> 0:17:15.240 you know, if they can't sort it out right, which 0:17:15.280 --> 0:17:19.560 is when someone particularly a teenager or a child, is 0:17:20.160 --> 0:17:22.879 you know, exhibiting that they might hurt themselves, you know, 0:17:23.640 --> 0:17:26.560 asking for helpless self harm or you know, you know, 0:17:26.680 --> 0:17:29.840 finding drugs or using drugs or whatever. I mean, these 0:17:29.840 --> 0:17:32.720 are things that that platforms like Meta Snapchat have really 0:17:32.760 --> 0:17:35.720 struggled with and been you know, hammered on for many 0:17:35.760 --> 0:17:38.440 many years, and Chatchept and Opening Eye have kind of 0:17:38.440 --> 0:17:41.840 found themselves right in the center of this because they 0:17:41.880 --> 0:17:44.720 made a decision to say, we are going to kind 0:17:44.720 --> 0:17:47.399 of engage in any kind of conversation. We're not going 0:17:47.480 --> 0:17:49.720 to just shut it down as soon as it goes 0:17:49.720 --> 0:17:52.000 into one of these topics. We're going to keep that 0:17:52.040 --> 0:17:55.400 conversation going. And that's a decision that they continue to make. 0:17:55.480 --> 0:17:58.480 Right now. Do you think they can stop it? I'm 0:17:58.480 --> 0:18:00.760 not saying that these models are out of just want 0:18:00.800 --> 0:18:04.840 to be clear, but I'm my one theory I've been 0:18:05.080 --> 0:18:08.280 kind of noodling on is that they don't have the 0:18:08.320 --> 0:18:12.080 ability to guarantee it can't it doesn't do something, and 0:18:12.119 --> 0:18:16.200 that's good. Do you think that open AI actually has 0:18:16.240 --> 0:18:18.959 the ability to guarantee it won't discuss the subject? How 0:18:19.040 --> 0:18:22.920 much control do you think they actually wield here? 0:18:23.920 --> 0:18:27.000 That's a really good question. I think they obviously. 0:18:27.480 --> 0:18:29.280 What you're kind of getting at there is that an 0:18:29.560 --> 0:18:32.440 LM is a bit of a black box. And you know, 0:18:32.520 --> 0:18:35.560 the way that the technology works is you get a 0:18:35.560 --> 0:18:39.440 different answer every time, and you know, we don't. It's 0:18:39.520 --> 0:18:44.120 true that the companies they cannot really guarantee it. 0:18:44.040 --> 0:18:45.760 Will or won't say anything. 0:18:45.800 --> 0:18:48.879 I mean, that's why there's disclosures plastered all over these things. 0:18:48.920 --> 0:18:53.160 And I think so there is definitely part of that. 0:18:53.400 --> 0:18:56.240 And you know, you could argue, well, that's just you know, 0:18:56.480 --> 0:18:59.840 a downside of this wonderful technology, and we shouldn't stop 0:18:59.880 --> 0:19:03.080 it just because they can't, you know, guarantee everything about it. 0:19:03.320 --> 0:19:04.760 But I think the other thing we should keep in 0:19:04.840 --> 0:19:08.040 mind is that there is actually a lot of layers 0:19:08.040 --> 0:19:12.280 that go on top of that black box and system 0:19:12.320 --> 0:19:16.960 prompts and the like exactly, and there is a lot 0:19:17.040 --> 0:19:21.399 that the tech companies are doing and can do to 0:19:21.560 --> 0:19:24.800 stop this kind of thing and and to kind of 0:19:25.320 --> 0:19:26.639 the answer you get from the l. 0:19:26.880 --> 0:19:29.520 M is not like the raw LLLM. 0:19:29.600 --> 0:19:33.600 Right, there's post training, there's a system prompt there's all sorts. 0:19:33.320 --> 0:19:36.160 Of things and can you break down the actually probably 0:19:36.200 --> 0:19:38.080 good for the listeners, break down what you mean by 0:19:38.119 --> 0:19:41.080 bout this. So the post training what's happening? Then? 0:19:41.520 --> 0:19:43.479 Yeah, so they they kind of, you know, when they 0:19:43.520 --> 0:19:47.880 build an LM, a large language model, they ram all 0:19:47.920 --> 0:19:50.560 this data through the algorithm and you know, it kind 0:19:50.560 --> 0:19:54.560 of makes a bunch of connections and links between different ideas, 0:19:55.520 --> 0:19:59.440 you know, pieces of language. You know, this word is 0:19:59.440 --> 0:20:01.040 similar to this, and so there's a little bit of 0:20:01.080 --> 0:20:04.000