WEBVTT - CZM Rewind: The Academics That Think ChatGPT Is BS 0:00:02.400 --> 0:00:03.560 Cauzone Media. 0:00:05.200 --> 0:00:07.600 Hey everyone, it's me ed Ze Tron and we're doing 0:00:07.640 --> 0:00:10.920 a rerun episode this week. Sadly, the in studio recording 0:00:10.920 --> 0:00:13.960 we did with Ashmund Rodriguez and Victoria's song we had 0:00:13.960 --> 0:00:17.400 a technical fall. Nothing we can do. It sucks, but 0:00:17.560 --> 0:00:19.800 we're doing a rerun this week. We're doing the Academics 0:00:19.800 --> 0:00:22.200 that think Chat GPT is BS and this is one 0:00:22.200 --> 0:00:24.560 of my favorite episodes I ever recorded. It changed how 0:00:24.600 --> 0:00:27.600 I do interviews for at Large. It's with these three academics, 0:00:27.640 --> 0:00:30.720 Michael Townsend Higgs, James Humphries and Joe Slater, who wrote 0:00:30.720 --> 0:00:35.000 a paper using the actual Frank thirty in definition of 0:00:35.000 --> 0:00:38.840 bullshit to say that chat GPT bullshits. It's so much fun. 0:00:38.920 --> 0:00:41.560 It's one of my favorite episodes I've recorded. I will 0:00:41.560 --> 0:00:43.599 have you a monologue this week as well. I do 0:00:43.720 --> 0:00:46.840 apologize for not having something new for you this week. 0:00:47.040 --> 0:00:49.479 You'll still get the monologue on Friday, though. Thank you 0:00:49.479 --> 0:00:52.040 for your time, your patience and of course for listening 0:00:52.040 --> 0:01:04.600 to Better Offline. Hello and welcome to Better Offline. I'm 0:01:04.600 --> 0:01:20.039 your host ed Ze Trunk. In early June, three researchers 0:01:20.040 --> 0:01:22.760 from the University of Glasgow published a paper in the 0:01:22.800 --> 0:01:26.640 Ethics of Information and Technology Journal called chat GBT is bullshit, 0:01:27.319 --> 0:01:29.399 And I just want to be clear. This is a 0:01:29.560 --> 0:01:32.959 great and thoroughly researched, from well argued paper. This is 0:01:33.040 --> 0:01:36.559 not silly at all. It's actually great academia. And today 0:01:36.640 --> 0:01:39.320 I'm joined by the men who wrote it, academics Michael 0:01:39.360 --> 0:01:42.800 Townsend Hicks, James Humphries, and Joe Slater, to talk about 0:01:42.880 --> 0:01:45.920 chat gbt's mediocrity and how it's not really built to 0:01:46.120 --> 0:01:49.240 represent the world at all. So, for the sake of argument, 0:01:49.720 --> 0:01:51.840 could you define bullshit for me? 0:01:53.080 --> 0:01:57.440 So you are bullshitting if you are speaking without caring 0:01:58.200 --> 0:02:01.040 about the truth of what you say. So normally, if 0:02:01.080 --> 0:02:04.480 I'm telling you staff about the world in a good case, 0:02:04.520 --> 0:02:07.160 I'll be telling you something that's true, and like trying 0:02:07.160 --> 0:02:09.760 to tell you something that's true. If I'm lying to you, 0:02:09.840 --> 0:02:12.519 I'll be knowingly telling you something that's false or something 0:02:12.560 --> 0:02:15.640 I think is false. I'm bullshitting. I just don't care. 0:02:16.160 --> 0:02:18.920 I'm trying to get you to believe me. I don't 0:02:18.960 --> 0:02:20.919 really care about whether what I say is true. I 0:02:21.000 --> 0:02:23.639 might not have any particular view on whether it's true 0:02:23.760 --> 0:02:24.880 or not right. 0:02:25.480 --> 0:02:28.560 And you define between like soft and hard bullshit? Can 0:02:28.600 --> 0:02:30.880 you also get into that as well? Can you also 0:02:30.960 --> 0:02:32.840 identify yourselves as well too? 0:02:33.080 --> 0:02:33.240 Yeah? 0:02:33.280 --> 0:02:34.000 Yeah, and Joe. 0:02:34.320 --> 0:02:37.200 So the soft bullshit hard bullshit distinction is a very 0:02:37.360 --> 0:02:41.880 serious and technical distinction, right, So he came up with 0:02:41.919 --> 0:02:47.000 this because bullshit is in the technical philosophical sense. Comes 0:02:47.000 --> 0:02:50.760 from Harry Frankfurt, recently diseased but a really great philosopher, 0:02:51.200 --> 0:02:54.280 and he talks about the about bullshit that there is 0:02:54.440 --> 0:02:58.400 in a popular culture and just in general discourse these days. 0:02:58.639 --> 0:03:01.240 Some of the ways he talks about bullshit seem to 0:03:01.280 --> 0:03:05.160 suggest that it needs to be accompanied by a sort 0:03:05.200 --> 0:03:06.400 of malign intention. 0:03:07.320 --> 0:03:09.440 I'm doing something kind of bad. 0:03:09.560 --> 0:03:13.520 I'm intending to mislead you about the enterprise of of 0:03:13.600 --> 0:03:14.120 what I'm. 0:03:13.919 --> 0:03:16.600 Doing, maybe about who you are or what you know. 0:03:16.720 --> 0:03:20.240 So you might be trying to portray yourself as someone 0:03:20.280 --> 0:03:24.360 who is knowledgeable about a particular subject. Maybe you're a 0:03:24.360 --> 0:03:26.800 student who showed up to class without doing the work. 0:03:27.280 --> 0:03:30.800 Maybe you're trying to portray yourself as someone who's virtuous 0:03:30.840 --> 0:03:33.079 in ways you're not. Maybe you're a politician who wants 0:03:33.160 --> 0:03:36.280 to seem like you care about your constituents, but actually 0:03:36.360 --> 0:03:40.000 you don't. So you're not trying to mislead somebody about 0:03:40.560 --> 0:03:43.480 what you're saying the content of your utterance. 0:03:44.080 --> 0:03:44.680 You're trying to. 0:03:44.600 --> 0:03:48.800 Mislead them instead about like why you're saying it's that's 0:03:48.840 --> 0:03:50.160 what we call hard bullshit. 0:03:50.360 --> 0:03:52.480 Yeah, it's one of the things Frankfort talked about. 0:03:52.680 --> 0:03:55.360 Yeah, so Frankfurt doesn't make this hard bullshit soft bullshit 0:03:55.360 --> 0:03:58.200 in the station, but we do because sometimes it seems 0:03:58.240 --> 0:04:01.240 like Frankfurt has this particular kind of intention in mind, 0:04:01.760 --> 0:04:04.760 but sometimes he's just a bit looser with it. And 0:04:04.800 --> 0:04:07.680 we want to say that CHAT GPT and other large 0:04:07.720 --> 0:04:11.280 language models they don't really have this intention to deceive 0:04:11.360 --> 0:04:13.640 because they're not people that don't have these intentions. They're 0:04:13.640 --> 0:04:16.159 not trying to mess with us in that kind of way, 0:04:16.560 --> 0:04:19.599 but they do lack this kind of caring about truth. 0:04:20.360 --> 0:04:23.080 Well, I'm James, by the way, I suppose we strictly 0:04:23.080 --> 0:04:25.840 don't want to say that they aren't hard bullshitters. And 0:04:25.880 --> 0:04:28.360 we just think if you don't think that large language 0:04:28.360 --> 0:04:30.120 models are sapient, if you don't think they're kind of 0:04:30.160 --> 0:04:33.560 minds in any important way, then they're not hard bullshitters. 0:04:33.600 --> 0:04:35.599 So I think in the paper we don't take a 0:04:35.640 --> 0:04:37.840 position on whether or not they are. We just say, 0:04:38.279 --> 0:04:40.200 if they are, this is the way in which they are. 0:04:40.400 --> 0:04:43.599 But minimally they're soft bullshitters, So they're kind of soft bullshites, 0:04:43.640 --> 0:04:46.640 as Joe says, doesn't require that speakers attempting to deceive 0:04:46.680 --> 0:04:48.839 the audience about the nature of the enterprise. 0:04:49.040 --> 0:04:49.960 Hard bullshit does. 0:04:50.160 --> 0:04:52.520 So if it turns out that large language models are sapient, 0:04:52.560 --> 0:04:54.560 which they're definitely not, like that's just. 0:04:54.600 --> 0:04:56.760 Technicio property, Yeah, that's nonsense. 0:04:56.880 --> 0:04:59.320 Yeah, But if they are, then they're hard bullshitters and 0:04:59.440 --> 0:05:01.960 minimally soft bullshitters or their bullshit machines. 0:05:03.400 --> 0:05:06.360 So you also make the distinction in there the intention, 0:05:06.720 --> 0:05:09.400 so the very fabric of hard bullshit that you intentionally 0:05:09.440 --> 0:05:12.880 are bullshitting to someone, You kind of make this distinction 0:05:12.960 --> 0:05:17.520 that the intention of the designer and the involved prompting 0:05:17.640 --> 0:05:21.719 could make this hard bullshit. Because with a lot of 0:05:21.760 --> 0:05:25.480 these models and someone recently jail broke chat GPT and 0:05:25.560 --> 0:05:28.000 it listed all of the things that's prompted to do, 0:05:28.400 --> 0:05:32.280 could prompting be considered intentional that Sam Altman, CEO of 0:05:32.560 --> 0:05:35.520 open Ai, could be intentionally bullshitting I think it is. 0:05:35.680 --> 0:05:37.640 And yeah, this again, I think it's something I don't 0:05:37.680 --> 0:05:40.159 know what the kind of hive mind consensus on this is. 0:05:40.200 --> 0:05:42.599 I'm sort of sympathetic to the idea that if you 0:05:42.640 --> 0:05:47.040 take this kind of purposive or teleological attitude towards right, 0:05:47.440 --> 0:05:50.279 what an intention is an effort to do something, then 0:05:50.279 --> 0:05:52.880 maybe they do have intentions. But again, I think in 0:05:52.920 --> 0:05:54.360 the paper we just sort of wanted I mean, it's 0:05:54.400 --> 0:05:57.560 a standard philosophical move, right, just sort of go, look, 0:05:57.600 --> 0:05:59.960 here's all this uncontroversial stuff as we can make it. 0:06:00.240 --> 0:06:02.120 Now we can hit you with the really controversial shit 0:06:02.160 --> 0:06:04.279 that we wanted to get to. So in the paper 0:06:04.320 --> 0:06:06.880 we sort of deliberately when maybe you might think it 0:06:06.880 --> 0:06:09.280 has intentions for it this reason, we kind of have 0:06:09.320 --> 0:06:12.120 no judgment on this efficiently, I'm sympathetic to the sort 0:06:12.120 --> 0:06:12.920 of view that you're putting. 0:06:12.960 --> 0:06:14.440 I think you're kind of sympathetic this as well. 0:06:14.720 --> 0:06:17.200 Right, Yeah, so, I Mike, there are a few ways 0:06:17.200 --> 0:06:20.440 that you can think of CHATCHYBT as having intentions. 0:06:19.960 --> 0:06:22.400 I think, and we talk about a few of them. 0:06:22.440 --> 0:06:25.840 One is by thinking of the designers who created it 0:06:26.240 --> 0:06:28.799 as kind of imbuing it with intentions. So they created 0:06:28.800 --> 0:06:31.280 it for a purpose and to do a particular task, 0:06:32.000 --> 0:06:35.320 and that task is to make people think that it's 0:06:35.360 --> 0:06:36.680 a normal sounding person. 0:06:36.960 --> 0:06:37.120 Right. 0:06:37.200 --> 0:06:40.280 It's to make people, when they have a conversation with it, 0:06:40.400 --> 0:06:44.280 not be able to distinguish between what it's outputting and 0:06:44.360 --> 0:06:46.200 what a normal human would say. 0:06:46.520 --> 0:06:46.719 Right. 0:06:47.120 --> 0:06:52.320 Right, And that kind of goal, if it amounts to 0:06:52.360 --> 0:06:56.680 an intention, is the kind of intention we think a 0:06:56.720 --> 0:06:57.520 bullshitter has. 0:06:57.640 --> 0:06:57.800 Right. 0:06:57.839 --> 0:07:00.440 It's not trying to deceive you about anything it's saying. 0:07:00.480 --> 0:07:02.600 It doesn't care whether what it's saying is true. That's 0:07:02.640 --> 0:07:04.760 not part of the goal. But the goal is to 0:07:04.800 --> 0:07:07.559 do is to make it seem as if it's something 0:07:07.600 --> 0:07:11.520 that it's not, like specifically a human interlocator. And one 0:07:11.560 --> 0:07:14.720 source for that goal is the programmers who designed it. 0:07:15.160 --> 0:07:19.080 Another is the training method. So it was trained by 0:07:19.120 --> 0:07:22.240 being given sort of positive and negative feedback in order 0:07:22.320 --> 0:07:26.320 to achieve a specific thing, right, And that specific thing 0:07:26.920 --> 0:07:29.280 is just sounding normal. And that's so similar to what 0:07:29.280 --> 0:07:31.680 our students are doing when they try to pretend that 0:07:31.760 --> 0:07:33.800 they read something they haven't read. 0:07:34.920 --> 0:07:37.760 There is something very collegiate about the way it bullshits. 0:07:37.760 --> 0:07:37.960 Though. 0:07:38.040 --> 0:07:40.320 It reminds me of when I was in college, as 0:07:40.480 --> 0:07:43.880 when Penn State and Aporistwyth two very different institutis. 0:07:43.520 --> 0:07:46.600 Right, both kind of in the middle of nowhere. 0:07:47.560 --> 0:07:50.600 Yeah, both very sad. But the one thing you saw 0:07:50.640 --> 0:07:53.440 with like students who were doing like B plus homework 0:07:53.520 --> 0:07:55.840 is they were using words they didn't really understand. They 0:07:55.840 --> 0:07:57.800 were putting things together in a way that really is 0:07:57.920 --> 0:08:00.760 like the intro body can clu usion. There was a 0:08:00.800 --> 0:08:03.680 certain formula behind it, and it's just it feels exactly 0:08:03.800 --> 0:08:05.640 like it. But that kind of brings me to my 0:08:05.680 --> 0:08:08.960 next question, which is, how did you decide to write this? 0:08:09.040 --> 0:08:10.080 What inspired this? 0:08:10.400 --> 0:08:10.560 Right? 0:08:10.560 --> 0:08:12.640 I would feel this because whenever Mike tells the story, 0:08:12.640 --> 0:08:13.960 you get a sanitized version. 0:08:14.760 --> 0:08:17.720 We were in the pub winsing about student essays. Perfect, yeah, 0:08:17.800 --> 0:08:21.080 like you were not long here, right, you're not long yeah, 0:08:21.320 --> 0:08:22.720 long post of. 0:08:22.640 --> 0:08:24.000 A bunch of us sort of went to the pub 0:08:24.480 --> 0:08:26.800 on a notional let's let's welcome our new members of 0:08:26.840 --> 0:08:29.600 staff sort of event, and inevitably we had about two 0:08:29.600 --> 0:08:32.240 points we were all passing and moaning and I think 0:08:32.240 --> 0:08:33.640 it might have been Neil that kind of prompted this. 0:08:33.840 --> 0:08:37.200 No, I don't think he was about it. 0:08:37.280 --> 0:08:39.040 We talked about it after, right, Yeah, in a case 0:08:39.120 --> 0:08:39.760 like sort of came out. 0:08:39.760 --> 0:08:41.440 We were talking about this sort of prevalence of chat 0:08:41.559 --> 0:08:45.040 GPT generated stuff and kind of what it said about 0:08:45.760 --> 0:08:48.719 how at least some students were kind of approaching assessments 0:08:48.800 --> 0:08:50.400 and you know, I forget who. Someone sort of just 0:08:50.440 --> 0:08:52.200 went off hand, Yeah, but it's all just Frank thirty 0:08:52.240 --> 0:08:54.440 and bullshit though, wasn't it. And we sort of collectively 0:08:54.480 --> 0:08:57.160 went hello, because obviously we all having a background plosophy, 0:08:57.160 --> 0:08:59.079 we've all read on bullshit. You know, we all went 0:08:59.600 --> 0:09:02.040 we get to say bullshit and allegedly serious academic work. 0:09:02.400 --> 0:09:05.160 So the start of it was we've had this experience 0:09:05.240 --> 0:09:07.959 of having to read all of this kind of uncanny 0:09:08.080 --> 0:09:10.840 Valley stuff, and when prompted, we all went, oh, it 0:09:10.960 --> 0:09:13.400 really is like Frank thirteen bullshit in a way that 0:09:13.440 --> 0:09:14.760 we can probably get paper ouse off. 0:09:14.800 --> 0:09:18.000 Yeah, And at that point I think we were in 0:09:18.040 --> 0:09:22.840 the department meetings talking about how to deal with chat 0:09:22.880 --> 0:09:26.160 GPT written papers. And there were discussions going on all 0:09:26.160 --> 0:09:28.280 over the university, including kind of in high levels and 0:09:28.320 --> 0:09:32.880 the administration to come up with a policy. And we 0:09:32.920 --> 0:09:37.440 specifically wanted a stronger policy because our university is very 0:09:37.440 --> 0:09:40.920 interested in cutting into technology, right, and so they wanted 0:09:40.920 --> 0:09:43.839 the students to have an experience using and figuring out 0:09:43.880 --> 0:09:47.360 how to use whatever the freshest technology. 0:09:46.880 --> 0:09:49.680 Is, right, as they should as they should, but not 0:09:50.280 --> 0:09:51.000 for essays. 0:09:51.080 --> 0:09:53.240 Right, that's the worry we had. We thought, you know, 0:09:53.720 --> 0:09:54.960 if we're not very clear. 0:09:54.760 --> 0:09:56.560 About this, the students will be using it in a 0:09:56.559 --> 0:09:59.840 way that will detract from their educational experience. 0:10:00.559 --> 0:10:03.840 Right. And at the same time, it was becoming more. 0:10:03.679 --> 0:10:07.280 Widely known how these machines work, like specifically how they're 0:10:07.280 --> 0:10:11.480 doing next token or next word prediction in order to 0:10:11.520 --> 0:10:15.559 come up with a digestible string of text. And when 0:10:15.600 --> 0:10:18.400 you know that that's how they're doing it, I mean, 0:10:19.559 --> 0:10:23.560 it seems so similar to what humans do when they 0:10:23.600 --> 0:10:26.520 have no idea what they're talking about. And so when 0:10:26.520 --> 0:10:29.280 we were talking about it just seemed like an obvious 0:10:29.280 --> 0:10:31.400 paper that someone was going to write, and we thought 0:10:31.440 --> 0:10:34.600 that had better bs. You know, eventually people are going 0:10:34.640 --> 0:10:37.559 to see this connection and it's a great papa. 0:10:37.679 --> 0:10:38.640 I think it worked very well. 0:10:38.679 --> 0:10:40.400 I mean, I think that it's the kind of thing 0:10:40.440 --> 0:10:44.120 where when I pitch this to other philosophers, it doesn't 0:10:44.160 --> 0:10:47.280 take them very long to just agree to be like, ah, yes. 0:10:47.520 --> 0:10:48.160 That's right. 0:10:48.720 --> 0:10:52.559 It's an interesting philosophical concept as well, because the way 0:10:52.600 --> 0:10:54.839 that people look at chat GPT in large language models 0:10:54.920 --> 0:10:57.080 is very much machine do this. But when you think 0:10:57.120 --> 0:10:59.840 about it, there are other ways what people are giving 0:10:59.880 --> 0:11:02.120 it consciousness. I just sort of tweet just now if 0:11:02.120 --> 0:11:05.520 someone talk about saying please with every request, it's like, no, 0:11:05.679 --> 0:11:08.320 I will abuse the computer in whatever manner I see fit. 0:11:08.760 --> 0:11:14.000 But it's it's curious because I think more people need 0:11:14.040 --> 0:11:17.040 to be having conversations like this. And one particular thing 0:11:17.080 --> 0:11:18.599 I like that you said, I actually would love you 0:11:18.640 --> 0:11:20.760 to go into more detail, is you said, the chat 0:11:20.760 --> 0:11:24.160 GIPT in large language buddels, they're not designed to represent 0:11:24.200 --> 0:11:26.920 the world at all. They're not lying or misrepresenting. They're 0:11:26.960 --> 0:11:28.640 not designed to do that. What do you mean by that? 0:11:29.000 --> 0:11:32.000 I mean kind of as background, I do philosophy of science, 0:11:32.160 --> 0:11:36.840 but my thoughts about something like chatchipt are largely inspired 0:11:36.840 --> 0:11:38.760 by the fact that I also teach a class called 0:11:39.480 --> 0:11:41.319 Understanding Philosophy through Science Fiction. 0:11:41.559 --> 0:11:43.360 Oh and and now we like talk. 0:11:43.160 --> 0:11:47.200 About whether computers could be conscious, and I don't know 0:11:47.200 --> 0:11:50.000 what you guys think. Actually I think they could, right, 0:11:51.360 --> 0:11:53.959 I just don't think this one is. And part of 0:11:54.000 --> 0:11:56.240 the reason I think they could, but this one isn't 0:11:56.440 --> 0:11:58.560 is that I think that in order to sort of 0:11:58.600 --> 0:12:01.840 represent the world or have the kinds of things we 0:12:01.920 --> 0:12:07.000 have their like beliefs, desires, thoughts that are about external things, 0:12:07.920 --> 0:12:12.280 you have to have internal states that are connected in 0:12:12.320 --> 0:12:16.640 some way to the external world. Usually causation, we're perceiving things, 0:12:16.679 --> 0:12:19.280 information is coming in. Then we've got some kind of 0:12:19.280 --> 0:12:23.080 state in our brain that's designed just to track these 0:12:23.120 --> 0:12:24.360 things in the external world. 0:12:24.600 --> 0:12:24.760 Right. 0:12:25.280 --> 0:12:28.240 That's a huge part of our cognitive lives is just 0:12:29.120 --> 0:12:33.160 tracking external world things, and it's a very important part 0:12:33.200 --> 0:12:35.720 of childhood development when you figure out how to act. 0:12:36.000 --> 0:12:40.480 It's semiotics, right, Daniel Chanler Aperistwyth told me semiotics it's 0:12:40.520 --> 0:12:41.440 like anception of the world. 0:12:41.520 --> 0:12:44.439 And this is like theory of meaning stuff. So yeah, 0:12:44.600 --> 0:12:47.000 semiotics is like theory of science. How is it that 0:12:47.040 --> 0:12:50.640 assign It can be both the representation of the thing 0:12:50.760 --> 0:12:53.560 and the thing itself. That can happen, yeah, but not always. 0:12:53.559 --> 0:12:57.040 Sometimes it's just the representation of the thing. And there's 0:12:57.080 --> 0:13:00.760 a lot of philosophy is about figuring out how brain 0:13:00.800 --> 0:13:04.320 states or words on a page can be about external 0:13:04.360 --> 0:13:07.960 world things. And a big part of it, at least 0:13:07.960 --> 0:13:11.720 from my perspective, has to do with tracking those things, 0:13:12.280 --> 0:13:15.400 keeping tabs on them, changing as a result of seeing 0:13:15.480 --> 0:13:19.360 differences in the external world. And chatch ept is not 0:13:19.440 --> 0:13:22.520 doing any of that, right, that's not what it's designed 0:13:22.520 --> 0:13:25.079 to do. It's taking in a lot of data once 0:13:25.880 --> 0:13:30.080 and then using that to respond to text. But it's 0:13:30.160 --> 0:13:36.120 not remembering individuals, tracking things in the world in any 0:13:36.120 --> 0:13:40.160 way perceiving things in the world. It's just forming a 0:13:40.200 --> 0:13:43.720 sort of statistical model of what people say. And that's 0:13:43.800 --> 0:13:49.280 kind of so divorced from what most thinking beings do. 0:13:50.080 --> 0:13:51.559 It's divorced from experience. 0:13:51.840 --> 0:13:53.280 Yeah. I mean, as far as I can tell, it 0:13:53.320 --> 0:13:55.160 doesn't have anything like experience. 0:13:55.480 --> 0:13:55.680 Yeah. 0:13:55.720 --> 0:13:57.640 And that's one of the things that this, I think 0:13:57.800 --> 0:14:00.199 of in one way, comes down to, is that if 0:14:00.200 --> 0:14:02.200 you sort of post this sufficiently far, someone is going 0:14:02.240 --> 0:14:04.480 to go, ah, isn't this just bioschauvinism, right, Like, aren't 0:14:04.520 --> 0:14:07.480 you just assuming that unless something runs on like meat, 0:14:07.840 --> 0:14:10.040 it can't be sentient? And this isn't something we get 0:14:10.040 --> 0:14:12.559 into into the paper, partly because we didn't really think 0:14:12.559 --> 0:14:15.720 it was worth addressing. But the sorts of things that 0:14:16.559 --> 0:14:19.000 seem like they're like never mind consciousness, right, But it 0:14:19.080 --> 0:14:20.840 seemed to be necessary in order for something to be 0:14:20.840 --> 0:14:22.960 trying to track the world, or in order to be 0:14:22.960 --> 0:14:25.160 corresponding the world, or to form beliefs about the world. 0:14:25.400 --> 0:14:28.480 Chat GPTs just doesn'tthing to meet any of them. Kind Of, 0:14:28.560 --> 0:14:31.360 it's if it does turn out that it's sapient, and 0:14:31.360 --> 0:14:37.840 then chat gpt has got some profoundly serious executive function disorders, right, sapient, Right, 0:14:37.840 --> 0:14:39.240 so we don't have to worry about it. But kind 0:14:39.280 --> 0:14:41.760 of it's not the case that we've got some blundering 0:14:41.840 --> 0:14:44.160 proto general intelligence that's trying to figure out how to 0:14:44.160 --> 0:14:46.000 represent the world. It's not trying to represent the world 0:14:46.000 --> 0:14:48.560 at all. This utterances are designed to look as if 0:14:48.880 --> 0:14:50.560 it's trying to represent the world, and then we just 0:14:50.560 --> 0:14:52.840 go that that's just bullshit. This is a classic case 0:14:52.840 --> 0:14:53.400 of bullshit. 0:14:53.800 --> 0:14:56.200 Yeah, it seems to be making stuff up, but making 0:14:56.240 --> 0:14:58.400 stuff up doesn't even seem to describe it. It's just 0:14:58.760 --> 0:15:02.040 throwing shit at a wall accurately but not accurately enough. 0:15:02.080 --> 0:15:04.320 Yeah, it's got various guidelines that allow it to throw 0:15:04.360 --> 0:15:05.920 shit at the wall where a sort of reasonably high 0:15:05.960 --> 0:15:07.200 degree of Actually, no, you're right, I mean one of 0:15:07.200 --> 0:15:09.280 the things that a human bullshitter could at least be 0:15:09.360 --> 0:15:11.840 characterized as doing is. They'd have to try and kind 0:15:11.840 --> 0:15:14.280 of judge their audience, right, they'd have to try and 0:15:14.320 --> 0:15:16.960 make the bullshit plausible to the audience that they're speaking to, 0:15:17.200 --> 0:15:19.320 and chat GPT can do that, right. All it can 0:15:19.360 --> 0:15:22.120 do is go on a statistically large enough model and 0:15:23.000 --> 0:15:26.960 it looks like Z follows, why follows X. It's not 0:15:27.040 --> 0:15:29.440 kind of got any or doesn't have any consciousness at all, 0:15:29.480 --> 0:15:32.080 of course, but it doesn't have any sensitivity to the 0:15:32.160 --> 0:15:34.600 sorts of things that people are in fact likely to 0:15:34.600 --> 0:15:37.560 find plausible. It just does a kind of brute number crunch. 0:15:37.760 --> 0:15:39.880 So it's more complicated than that, but I think it 0:15:40.120 --> 0:15:42.240 boils down it effectively to kind of number crunching, right, 0:15:42.360 --> 0:15:44.000 kind of data and context. 0:15:43.720 --> 0:15:47.480 In which probablistic planning of what planning it doesn't plan. 0:15:47.720 --> 0:15:49.800 That's the thing. It's interesting. There are these words you 0:15:49.920 --> 0:15:51.920 use to describe things that, when you think about it, 0:15:51.960 --> 0:15:55.640 are not accurate. You can't say it plans or things. 0:15:56.480 --> 0:15:59.280 I think one thing between when we came up with 0:15:59.360 --> 0:16:02.920 the idea and when we finish writing the paper, we 0:16:03.000 --> 0:16:06.000 spend some time reading about how it works and how 0:16:06.000 --> 0:16:09.040 it represents language and what the statistical. 0:16:08.400 --> 0:16:09.320 Model is like. 0:16:09.640 --> 0:16:15.680 And I was maybe more impressed than James about that, 0:16:15.760 --> 0:16:19.240 because it is like doing things that are similar to 0:16:19.320 --> 0:16:22.240 what we do when we understand language. It does seem 0:16:22.280 --> 0:16:26.440 to kind of locate words in a meaning space right right, 0:16:26.840 --> 0:16:29.200 and connect them to other words, and you know, show 0:16:29.240 --> 0:16:32.000 similarity and meaning, and it also does seem to be 0:16:32.000 --> 0:16:35.000 able to in some way understand context. 0:16:35.760 --> 0:16:38.720 But we don't know how. 0:16:38.440 --> 0:16:41.560 Similar that is for a variety of reasons, but mostly 0:16:41.640 --> 0:16:44.320 because it's too big of a system and we can 0:16:44.360 --> 0:16:48.360 only kind of probe it. And it's trained indirectly, right, 0:16:48.520 --> 0:16:54.040 so it's not programmed by individuals. And even though that's 0:16:54.120 --> 0:16:56.760 kind of a very impressive model of language and meaning 0:16:57.000 --> 0:17:00.240 and may in some ways be similar to or what 0:17:00.320 --> 0:17:04.440 we do when we understand language, we're doing a lot 0:17:04.520 --> 0:17:08.960 more things like planning, things like tracking things in the world. 0:17:09.320 --> 0:17:12.119 Just having desires and representing the way you want the 0:17:12.160 --> 0:17:15.280 world to be and thereby generating goals doesn't seem to 0:17:15.320 --> 0:17:18.000 be something that it has anything any room for in its. 0:17:17.960 --> 0:17:19.679 Architect So this is something of the time of it's 0:17:19.680 --> 0:17:21.000 just goods to me, and you were talking about the 0:17:21.080 --> 0:17:23.200 kind of in some ways it learns language the same 0:17:23.200 --> 0:17:24.919 way that we do. I mean, it's got no grasp 0:17:24.920 --> 0:17:27.240 of expleted in fixation, right, this is one of chum's here. 0:17:27.280 --> 0:17:29.520 What does that mean? Just for not for me? I 0:17:29.560 --> 0:17:30.159 definitely know. 0:17:30.640 --> 0:17:33.480 Yeah, yeah, of course, if I give you the sentence 0:17:33.920 --> 0:17:36.320 that's completely crazy, man, and tell you to put the 0:17:36.320 --> 0:17:39.320 word fucking into that sentence, there's a number of ways 0:17:39.359 --> 0:17:41.280 in which any language speaker is going to do it 0:17:41.560 --> 0:17:43.120 that they'll just go yeah, like of course. 0:17:42.920 --> 0:17:44.920 That where it goes, right. 0:17:45.280 --> 0:17:47.720 But we it seems that we've got a grasp on 0:17:47.760 --> 0:17:50.760 this incredibly early on in a way that doesn't look 0:17:50.840 --> 0:17:52.280 like it's the ways at least most of us that 0:17:52.400 --> 0:17:54.840 taught language, right, we get quite highly told off when 0:17:54.840 --> 0:17:55.280 we try and do. 0:17:55.280 --> 0:17:56.200 Expleted in fixation. 0:17:56.400 --> 0:17:58.520 Yes, So this I think would be one of those 0:17:58.600 --> 0:18:01.080 cases where you could do a sort of dis analogy 0:18:01.160 --> 0:18:03.760 by cases. Right, you present chat GPT a sentence and 0:18:03.800 --> 0:18:07.040 say insert the word fucking correctly in this sentence, and 0:18:07.119 --> 0:18:08.679 I doesn't think it would be very good at it. 0:18:08.720 --> 0:18:09.359 I think it would be. 0:18:09.400 --> 0:18:11.119 I think you reckon. I mean we probably couldn't. 0:18:11.840 --> 0:18:14.880 We could, but we shouldn't. Yeah. 0:18:15.359 --> 0:18:19.080 One of the things that I thought was like kind 0:18:19.080 --> 0:18:23.280 of interesting about how it works is that it does 0:18:23.400 --> 0:18:25.720 learn language, probably differently from the way we do, but 0:18:25.800 --> 0:18:28.560 it does it all by examples, you know, So it's 0:18:28.560 --> 0:18:30.680 looking at all these pieces of text and thinking, ah, 0:18:30.680 --> 0:18:33.440 this is okay, that's okay. And one kind of interesting 0:18:33.480 --> 0:18:36.480 thing about how humans understand language is that we're able 0:18:36.520 --> 0:18:41.280 to kind of understand meaningless but grammatical sentences. It's not 0:18:41.320 --> 0:18:44.600 clear to me that chat GPT would understand those. That's 0:18:44.640 --> 0:18:49.200 another Chopski example. So you know, Chowski has this example that's. 0:18:49.000 --> 0:18:53.600 Like, what is it colorless? Us sleep furiously color? 0:18:53.640 --> 0:18:58.840 It's colorless green ideas, right, And that's a meaningless sentence, 0:18:58.840 --> 0:19:02.119 but it's grammatically well formed, and we can understand that 0:19:02.119 --> 0:19:06.119 it's grammatically well formed, but also that it's meaningless. Because 0:19:06.200 --> 0:19:12.359 chat Gibt kind of combines different aspects of what philosophers 0:19:12.359 --> 0:19:16.800 of language, logicians, linguistics people see y as like different 0:19:16.800 --> 0:19:19.199 components of meaning. It sees these as all kind of 0:19:19.200 --> 0:19:20.960 wrapped up in the same thing. It puts them in 0:19:21.000 --> 0:19:24.040 the same big model. I'm not sure it could differentiate 0:19:24.040 --> 0:19:29.120 between ungrammaticality and meaninglessness. 0:19:36.880 --> 0:19:39.720