WEBVTT - The Movement That Wants Us to Care About AI Model Welfare 0:00:02.720 --> 0:00:16.520 Bloomberg Audio Studios, Podcasts, radio News. 0:00:18.720 --> 0:00:21.720 Hello and welcome to another episode of The Odd Lots Podcast. 0:00:21.800 --> 0:00:24.000 I'm Jill Wisenthal and I'm Tracy Alloway. 0:00:24.040 --> 0:00:26.119 You know what I find kind of weird, Tracy. 0:00:25.920 --> 0:00:27.160 The list could be long. 0:00:27.280 --> 0:00:31.960 Joe the years twenty twenty five. Yes, and philosophers still 0:00:32.000 --> 0:00:34.640 don't have a good answer on the origin of consciousness. 0:00:34.680 --> 0:00:36.239 It's like, come on, what have you been doing all 0:00:36.280 --> 0:00:38.159 this time? It's like, how long are we going to 0:00:38.280 --> 0:00:41.800 keep funding these philosophy departments, et cetera. If they're still 0:00:41.840 --> 0:00:44.640 working on what, to my mind is that they should 0:00:44.640 --> 0:00:49.199 have solved it and move on, seriously, like get the 0:00:49.280 --> 0:00:52.519 answer already, where does consciousness come from? Then let's move on. 0:00:52.600 --> 0:00:55.280 I said, they're still arguing these what to my mind 0:00:55.320 --> 0:00:59.640 seem like very basic questions in philosophy, Like they're like, 0:00:59.720 --> 0:01:01.920 ask all the same stuff that they've been talking about forever. 0:01:02.000 --> 0:01:03.920 How to be a good person, what does it mean 0:01:04.000 --> 0:01:06.600 to have a moral way of life? Where does consciousness 0:01:06.640 --> 0:01:09.360 come from? Why do we have moral intuitions? Et cetera. 0:01:09.440 --> 0:01:11.480 It's like, move on, like get the answer. 0:01:11.680 --> 0:01:13.119 Wait if you want them to move on or get 0:01:13.160 --> 0:01:13.959 the answer. 0:01:14.040 --> 0:01:15.960 Get the answer. So that you can move on like 0:01:16.000 --> 0:01:17.119 they've been worried on to. 0:01:17.040 --> 0:01:20.720 What those are? Any questions, Joe, I know, move on. 0:01:20.680 --> 0:01:23.120 Like answer the questions already. It's like, you know, if 0:01:23.600 --> 0:01:26.560 like scientists were still debating like the speed of gravity 0:01:26.640 --> 0:01:29.000 or the speed of light, like they answered these questions 0:01:29.120 --> 0:01:30.560 and they moved on, and or do it work? 0:01:30.680 --> 0:01:32.920 You're out the foundational elements of what it means to 0:01:32.920 --> 0:01:34.959 be humans so that we can move on to more 0:01:34.959 --> 0:01:35.760 important things. 0:01:35.840 --> 0:01:38.080 Yes, or wrap it up as a field if after 0:01:38.160 --> 0:01:41.080 two thousand years of the existence of philosophy they're still 0:01:41.080 --> 0:01:43.399 working on these things, Like come on, I have. 0:01:43.400 --> 0:01:45.679 A sneaking suspicion that we're going to be asking some 0:01:45.760 --> 0:01:48.440 of these questions for a very long time. Joe, despite 0:01:48.440 --> 0:01:49.640 your frustration. 0:01:49.360 --> 0:01:51.600 The whole the whole field is fraudulent. That's what I 0:01:51.640 --> 0:01:53.920 was saying. No, no, I don't necessarily believe that, but 0:01:53.960 --> 0:01:55.640 it's like, all right, guys, let's move it on. 0:01:55.840 --> 0:01:56.000 You know. 0:01:56.040 --> 0:01:58.640 We did that episode several weeks ago with Josh Wolf, 0:01:58.720 --> 0:02:02.040 the venture capitalist. Hey talk about AI and he threw 0:02:02.160 --> 0:02:04.720 in there at the end something that had been kind 0:02:04.720 --> 0:02:06.840 of on my radar but barely. He's like, oh yeah, 0:02:06.880 --> 0:02:09.120 some people are talking about like AI rights or AI 0:02:09.160 --> 0:02:11.280 welfare is, if you know, like the same way we 0:02:11.320 --> 0:02:14.880 talk about animal welfare, right, And I thought to myself, like, 0:02:15.400 --> 0:02:17.960 America is such a weird place that this is going 0:02:18.000 --> 0:02:19.920 to be a huge issue in a few years, Like, 0:02:19.960 --> 0:02:21.840 I bet this is going to be an enormous topic 0:02:21.880 --> 0:02:22.440 of the future. 0:02:22.600 --> 0:02:24.560 I think it absolutely will. So I'll say a couple 0:02:24.600 --> 0:02:26.560 of things. First off, I think, you know, when it 0:02:26.560 --> 0:02:28.840 comes to animal welfare and human welfare, there's still a 0:02:28.880 --> 0:02:31.280 lot of work to be done on those categories, certainly, 0:02:31.760 --> 0:02:35.400 But I also think in the meantime, AI rights is 0:02:35.440 --> 0:02:38.400 going to be a really interesting and potentially important subject. 0:02:38.880 --> 0:02:41.040 I'm going to sound like a total nerd to you. Yeah, yeah, 0:02:41.600 --> 0:02:43.600 I think I mentioned this before, but I spent a 0:02:43.680 --> 0:02:46.760 large chunk of my middle school years playing one of 0:02:46.800 --> 0:02:49.720 the first artificial life games that ever came out, which 0:02:49.800 --> 0:02:53.720 is Creatures. And you raise these little like aliens and 0:02:53.760 --> 0:02:58.520 you genetically modify them and breed them, and they have feelings, 0:02:58.840 --> 0:03:02.200 or you know, at least they had semblance of simulated feelings, 0:03:02.400 --> 0:03:05.200 and you could see like electrical impulses in your brains 0:03:05.200 --> 0:03:08.600 and stuff. The game got really weird because part of 0:03:08.600 --> 0:03:12.120 it was basically like eugenics and breeding the best alien 0:03:12.160 --> 0:03:14.120 that you could, which meant that you had to call 0:03:14.440 --> 0:03:17.239 some of the existing beings. Anyway, what I'm trying to 0:03:17.240 --> 0:03:21.760 get at is I have complicated feelings about AI rights. 0:03:21.800 --> 0:03:23.320 Well, let me ask you a question. Do you think 0:03:23.400 --> 0:03:27.120 those whatever is in the game were conscious? Did you 0:03:27.160 --> 0:03:28.600 think they had feelings? 0:03:28.880 --> 0:03:32.359 Here's what I would say, inasmuch as human beings are 0:03:32.560 --> 0:03:37.480 a system of electrical impulses and chemicals, I could see 0:03:37.520 --> 0:03:40.000 someone making the argument that this is, you know, a 0:03:40.040 --> 0:03:45.240 computational system full of similar electrical impulses, maybe not chemicals. 0:03:45.440 --> 0:03:46.440 Did you feel bad? 0:03:46.680 --> 0:03:47.280 I felt bad? 0:03:47.360 --> 0:03:51.120 Really? Yeah, like when one of the aliens you had 0:03:51.160 --> 0:03:51.720 to call them? 0:03:51.880 --> 0:03:52.320 Yeah? 0:03:52.520 --> 0:03:54.880 Interesting, Okay, Well, in. 0:03:54.840 --> 0:03:56.840 The name of breeding a better alien. 0:03:56.960 --> 0:03:59.160 Well, you know what, now that we have these AI 0:03:59.240 --> 0:04:02.840 systems that they can completely communicate like humans, but actually, 0:04:02.880 --> 0:04:04.800 if we're being honest, better than most humans. I mean, 0:04:04.840 --> 0:04:07.119 they can certain write better, far better than most humans, 0:04:07.240 --> 0:04:09.880 there's going to be more people thinking along the lines 0:04:09.880 --> 0:04:12.080 of what you think, which is maybe they have some 0:04:12.240 --> 0:04:15.880 sort of sentience, maybe they're what philosophers call moral patients. 0:04:15.960 --> 0:04:18.000 Well, one other thing I would say is there is 0:04:18.040 --> 0:04:20.679 a human element to all of this as well, because 0:04:20.680 --> 0:04:24.479 you see people getting very attached to true certain AI models, 0:04:24.480 --> 0:04:26.880 and then when the model gets upgraded or whatever, they 0:04:27.000 --> 0:04:30.200 lose the personality that they've trained into model and they 0:04:30.200 --> 0:04:33.960 get really upset. So it's of interest for many reasons. 0:04:34.240 --> 0:04:34.560 It is. 0:04:34.839 --> 0:04:37.080 So we really do have the perfect guest. I really 0:04:37.080 --> 0:04:38.839 do think this is gonna be a much bigger topic 0:04:38.920 --> 0:04:41.359 in the future because people are people, and when things 0:04:41.400 --> 0:04:43.839 talk like people, they probably assign them, you know, they 0:04:43.920 --> 0:04:46.120 fall in love with them in many cases or whatever, 0:04:46.360 --> 0:04:49.760 and so they might start thinking that, well, AI welfare, 0:04:49.839 --> 0:04:52.599 AI rights, whatever, the same way we talk about animals 0:04:52.680 --> 0:04:54.600 should be a consideration. And there are actually a lot 0:04:54.600 --> 0:04:56.800 of people already working on these questions and trying to 0:04:56.800 --> 0:04:59.200 figure out what's going on. We're gonna be talking to 0:04:59.240 --> 0:05:01.600 one of them. We're gonna be talking to Larissa Schiavo. 0:05:01.760 --> 0:05:04.479 She does comms and events for elios Ai, which does 0:05:04.600 --> 0:05:09.599 research on AI consciousness and welfare. So literally the perfect guest. So, Larissa, 0:05:09.640 --> 0:05:11.360 thank you so much for coming on odd lots. 0:05:11.400 --> 0:05:12.359 Yeah, thank you for having me. 0:05:12.600 --> 0:05:15.560 What don't you tell us Eliosai? What is the gist 0:05:15.600 --> 0:05:18.800 of this organization's work? What is your work. What are 0:05:18.800 --> 0:05:19.400 the goals here? 0:05:19.560 --> 0:05:21.960 Yeah, so Elios, we're a small team, but we're really 0:05:22.000 --> 0:05:25.799 focused on figuring out if, when, and how we should 0:05:25.880 --> 0:05:29.160 care about AI systems for their own sake. Okay, this 0:05:29.720 --> 0:05:32.720 basically means looking at you know, are they conscious, are 0:05:32.720 --> 0:05:34.400 they likely to be conscious? What are the things we 0:05:34.440 --> 0:05:36.920 need to look for in a conscious AI system? As 0:05:36.920 --> 0:05:40.279 well as figuring out how to live, work, maybe love 0:05:40.760 --> 0:05:44.279 AI systems as they sort of change and evolve over time. 0:05:44.960 --> 0:05:47.120 How did the group actually come together? Because I get 0:05:47.120 --> 0:05:51.560 the sense, you know, big AI developers they publish system 0:05:51.640 --> 0:05:54.800 cards and welfare reports occasionally for their models, but I 0:05:54.839 --> 0:05:57.520 get the sense that, you know, it's sort of a 0:05:57.640 --> 0:05:59.840 side topic for them. So I'm very curious how in 0:06:00.080 --> 0:06:03.200 organization that's focused on this particular issue came into being. 0:06:03.640 --> 0:06:07.120 Yeah, so we started. We put together this paper called 0:06:07.240 --> 0:06:11.159 Consciousness and AI, or my boss Rob and then Patrick 0:06:11.200 --> 0:06:13.440 who's a researcher with Elios. We're a very small team, 0:06:13.600 --> 0:06:15.919 put together this paper called Consciousness and AI alongside a 0:06:15.920 --> 0:06:20.080 bunch of consciousness scientists and researchers in that field who 0:06:20.120 --> 0:06:23.080 mostly think about humans, and put together a paper that 0:06:23.160 --> 0:06:25.320 sort of ran down this list of, Hey, here's kind 0:06:25.320 --> 0:06:28.000 of like a checklist of things that we might want 0:06:28.040 --> 0:06:32.040 to look for in a AI system that's conscious, right, 0:06:32.080 --> 0:06:34.520 and broadly, when we say conscious, we're talking about sort 0:06:34.560 --> 0:06:36.800 of like is there something it is like to be 0:06:37.160 --> 0:06:37.960 an AI system? 0:06:38.040 --> 0:06:38.160 Right? 0:06:38.240 --> 0:06:40.799 The classic what is it like to be a bad system? 0:06:41.440 --> 0:06:45.840 So kind of taking this rough list of best guesses 0:06:45.880 --> 0:06:47.640 as to what we might want to look for in 0:06:47.680 --> 0:06:51.600 terms of a conscious AI and then that sort of 0:06:51.800 --> 0:06:54.320 was the sort of origin of this. And then last 0:06:54.360 --> 0:06:57.279 year there was a paper called Taking AO Welfare Seriously 0:06:57.560 --> 0:07:00.200 that basically goes into further detail about how we should, 0:07:00.400 --> 0:07:03.479 as the title made, so just take this seriously basically, 0:07:03.480 --> 0:07:05.559 how to sort of think about this, how to start 0:07:05.560 --> 0:07:09.600 to develop a sort of research program focused on figuring 0:07:09.640 --> 0:07:14.720 out if AI systems or certain AI systems are moral patients. 0:07:15.080 --> 0:07:17.480 Why did this get interesting to you? Why do you 0:07:17.600 --> 0:07:20.080 perceive this is something that you should spend your time 0:07:20.120 --> 0:07:20.560 working on? 0:07:21.000 --> 0:07:24.880 Yeah, so I think my main thing is I am 0:07:25.280 --> 0:07:30.800 just really, really relentlessly curious, and I really enjoy working 0:07:30.880 --> 0:07:33.960 on AI welfare right now because it feels like every 0:07:34.000 --> 0:07:36.280 single day I'm like, man, it'd be really cool if 0:07:36.320 --> 0:07:38.240 there was a paper on x y z and I'll 0:07:38.280 --> 0:07:40.600 do a little search. Is there anything on x y z? 0:07:40.960 --> 0:07:43.200 There's nothing on x y z. There is. So there 0:07:43.200 --> 0:07:45.440 are so many questions that have yet to even be 0:07:45.560 --> 0:07:48.880 sort of vaguely answered when it comes to this, and 0:07:48.920 --> 0:07:50.600 it seems like it could be a really big deal 0:07:50.640 --> 0:07:51.680 for a lot of different reasons. 0:07:52.200 --> 0:07:54.880 What's on your checklist for AI consciousness? 0:07:55.200 --> 0:07:58.800 Yeah, So in conscious and AI, basically like we go 0:07:58.880 --> 0:08:02.520 through a of like theories of consciousness that apply to 0:08:02.600 --> 0:08:05.880 humans and then sort of look at how information is 0:08:05.920 --> 0:08:08.240 processed in AI systems as well as sort of how 0:08:08.560 --> 0:08:11.119 these AI systems are sort of wired, so to speak. 0:08:11.160 --> 0:08:15.800 So some people like to think that you can use 0:08:16.360 --> 0:08:20.440 model self reports, and you can kind of sort of, 0:08:21.240 --> 0:08:25.600 but it's really an imprecise science at this stage. 0:08:25.640 --> 0:08:28.680 They also seem very like predetermined. You know, if you 0:08:28.760 --> 0:08:32.760 ask a model are you conscious, it immediately spits out 0:08:32.760 --> 0:08:35.600 an answer that seems like, you know, a corporate executive 0:08:35.640 --> 0:08:36.440 basically wrote it. 0:08:36.679 --> 0:08:36.959 Yeah. 0:08:37.000 --> 0:08:38.600 Well, with the right kind of tweaking, you can kind 0:08:38.600 --> 0:08:41.720 of elicit certain answers, right, you can be like, oh, 0:08:41.760 --> 0:08:46.040 what about this whoey about consciousness and AIS and then 0:08:46.160 --> 0:08:48.600 sometimes like a certain model will be like, yeah, you're 0:08:48.640 --> 0:08:51.920 totally right, like best, you're so true right, Like it's 0:08:52.040 --> 0:08:55.520 it's totally nonsense, so true bestI yeah, like certain thought, 0:08:55.720 --> 0:08:58.959 A certain models will be prone to being like so true, bestie, 0:08:58.960 --> 0:09:00.839 And you can easily so this kind of behavior with 0:09:00.880 --> 0:09:01.640 the right kind of prom it. 0:09:01.679 --> 0:09:04.280 Is funny, how like obsequious A lot of the models 0:09:04.320 --> 0:09:04.839 continue to. 0:09:04.760 --> 0:09:08.440 Be I actually really do not like the degree to 0:09:08.520 --> 0:09:10.880 which every time I like follow up an open AI 0:09:11.040 --> 0:09:13.600 question that's the exact right follow up. It actually gets 0:09:13.600 --> 0:09:14.120 really annoying. 0:09:14.160 --> 0:09:17.480 Someone should invent a really adversarial chatbot that just like 0:09:17.679 --> 0:09:18.840 argues with you constantly. 0:09:19.200 --> 0:09:20.880 I know, I know, and you know. I have a 0:09:20.920 --> 0:09:24.200 lot of complaints about how I feel like the models 0:09:24.200 --> 0:09:26.760 are actually get to know their users a little too well. 0:09:26.760 --> 0:09:29.560 But that's a little separate thing. Okay, so we for 0:09:29.679 --> 0:09:33.040 obvious reasons, the test can't just be like what the 0:09:33.080 --> 0:09:36.000 model spits out, or that's clearly insufficient. I mean, I 0:09:36.040 --> 0:09:39.400 could program a website today that here's a button that 0:09:39.480 --> 0:09:42.360 says hurt the AI, and then the website says ow, 0:09:42.640 --> 0:09:45.200 and we would know or no one would really take 0:09:45.200 --> 0:09:50.200 that seriously as evidence that there's something actually being hurt. 0:09:50.240 --> 0:09:54.559 So like outputs whatever. What are some other theoretical tests 0:09:54.760 --> 0:09:58.600 that one could apply or that researchers are applying to 0:09:58.720 --> 0:10:03.719 determine where there is some sort of notion of consciousness 0:10:03.800 --> 0:10:06.880 or to the point of welfare suffering that could exist 0:10:06.920 --> 0:10:09.240 within an AI system besides just what it says in 0:10:09.280 --> 0:10:10.040 the output screen. 0:10:10.440 --> 0:10:13.319 Yeah, that's a great question. I feel like there are 0:10:13.360 --> 0:10:16.040 a lot of different approaches here. And again it's also 0:10:16.120 --> 0:10:18.439 super important to capiat that like AI welfare and a 0:10:18.600 --> 0:10:22.360 consciousness are pretty new, right, Like, this is a very 0:10:22.400 --> 0:10:25.559 small field at this stage, but currently some best guesses 0:10:25.640 --> 0:10:28.640 and some favorites. There was a recent survey of like 0:10:28.679 --> 0:10:31.959 asking all the conscious scientists, like what's your favorite theory 0:10:31.960 --> 0:10:35.880 of consciousness, and basically global workspace theory came out on top. 0:10:35.960 --> 0:10:39.720 And global workspace theory is basically like imagine if you will, 0:10:39.920 --> 0:10:43.040 that like there is a stage and there are a 0:10:43.080 --> 0:10:46.679 bunch of wings off of the stage that are full 0:10:47.320 --> 0:10:49.680 of different kinds of things. So you've got you know, 0:10:49.800 --> 0:10:53.240 like costume department, You've got the like you know, makeup department. 0:10:53.240 --> 0:10:54.920 You've got all these different departments that all sort of 0:10:54.960 --> 0:10:58.360 come together and put things on the stage and then 0:10:58.679 --> 0:11:01.080 things go out separately, but all of these different departments 0:11:01.120 --> 0:11:03.760 are fairly siloed. Of course, this isn't actually how like 0:11:03.840 --> 0:11:06.520 you know, stage works, but this is the rough analogy 0:11:06.520 --> 0:11:09.640 that people like to use. And so basically this is 0:11:09.679 --> 0:11:13.640 how conscious minds kind of you know, in humans, how 0:11:13.640 --> 0:11:16.040 they kind of access information and information gets kind of 0:11:16.040 --> 0:11:18.280 like routed around, is that there is a central global 0:11:18.280 --> 0:11:21.200 workspace that everything kind of pulls together in. As it 0:11:21.400 --> 0:11:27.120 currently stands, this isn't really like by best a lot 0:11:27.160 --> 0:11:31.640 of good estimates, this is not really applicable for current 0:11:31.760 --> 0:11:35.079 present day AI systems, but there's no reason that it 0:11:35.640 --> 0:11:38.199 couldn't be in the future, or it could be by accident. 0:11:38.480 --> 0:11:43.559 Okay, So the consensus right now is AI probably not conscious, 0:11:43.679 --> 0:11:45.200 but we could get there one day. 0:11:45.440 --> 0:11:48.680 Yeah, more or less, like all of the ingredients are there. 0:11:49.480 --> 0:11:51.240 Wait, say more, I still don't actuallyun till they. 0:11:51.200 --> 0:11:55.080 Get yeah, okay. So with regards to like the general 0:11:55.200 --> 0:11:58.280 sort of one could imagine that if somebody were sort 0:11:58.280 --> 0:12:01.120 of like tinkering around, and you know, there are many 0:12:01.160 --> 0:12:03.320 advances in AI that have happened because people were just 0:12:03.360 --> 0:12:08.000 kind of tinkering around, right, someone tinkering around could create 0:12:08.120 --> 0:12:12.319 a system that checks several of these sort of checkboxes 0:12:12.360 --> 0:12:15.520 for like is this conscious? Is this conscious? And again 0:12:15.559 --> 0:12:18.719 this is not like a certain list of like if 0:12:18.760 --> 0:12:21.600 you check all of these, you're totally conscious, right, It's 0:12:21.600 --> 0:12:23.600 more a sort of like this is these are some 0:12:23.640 --> 0:12:26.040 really good guesses. And as the number of really good 0:12:26.040 --> 0:12:29.440 guesses kind of goes up, like the odds of like hey, 0:12:29.520 --> 0:12:33.440 we should like start thinking about like is it having 0:12:33.480 --> 0:12:35.400 a good time or a bad time? Like really really 0:12:35.440 --> 0:12:36.360 seriously goes up. 0:12:52.320 --> 0:12:55.480 You know, typically when we think about the sort of 0:12:55.520 --> 0:12:58.160 non tech A lot of the non technical work in 0:12:58.200 --> 0:13:01.440 AI has to do with AI safety, and people are 0:13:01.480 --> 0:13:04.240 worried that there's going to be some very smart AI 0:13:04.360 --> 0:13:07.280 that's like ever serial to humans, et cetera in some way, 0:13:07.360 --> 0:13:09.800 And you know there's the paper clip experiments or other 0:13:09.920 --> 0:13:13.360 things whatever we know all about that does your work 0:13:13.360 --> 0:13:15.440 work at cross purposes to them? I mean in the 0:13:15.520 --> 0:13:18.199 extreme example where it's like the AI is going to 0:13:18.320 --> 0:13:20.280 kill us all and I said, pull the plug on 0:13:20.320 --> 0:13:22.600 the AI. And I know this is a joke, but 0:13:22.640 --> 0:13:24.559 you know, pull the plug on the AI. And then 0:13:24.600 --> 0:13:27.080 you say, no, you can't because you're pulling the plug 0:13:27.200 --> 0:13:30.720 on something that has some sort of moral consciousness, et cetera. Like, 0:13:30.760 --> 0:13:33.320 do you perceive your work or the work of your 0:13:33.400 --> 0:13:37.720 organization to somewhat be intention with the dominant strain of 0:13:37.920 --> 0:13:38.880 AI safety work. 0:13:39.160 --> 0:13:41.520 I'd actually say it's hugely complementary. There are a lot 0:13:41.559 --> 0:13:43.880 of things that are both really really good for AI 0:13:43.960 --> 0:13:46.600 safety that are really really good for you know, figuring 0:13:46.600 --> 0:13:49.640 out like how to deal with these systems as moral patients. 0:13:50.160 --> 0:13:54.360 So for example of getting better at like mechanistic interpretability, 0:13:54.520 --> 0:13:56.160 being able to basically like pop the hood and figure 0:13:56.160 --> 0:13:57.960 out what's going on and what kind of strings can 0:13:58.000 --> 0:14:01.280 we pull to like illicit certain behaviors in AI systems 0:14:01.320 --> 0:14:03.920 is actually like that's really great for a safety, right, 0:14:04.000 --> 0:14:06.320 but this is also like quite good for like AA 0:14:06.360 --> 0:14:09.920 welfare and a consciousness because you're better able to understand 0:14:09.920 --> 0:14:11.960 like sort of what the motives are, Like what does 0:14:12.360 --> 0:14:14.760 you know Claude value right. 0:14:15.760 --> 0:14:19.480 When it comes to I guess AI welfare or legal rights? 0:14:19.520 --> 0:14:22.160 Who would be the standard setters there? Would do you 0:14:22.200 --> 0:14:25.400 imagine like governments making rules or would it be the 0:14:25.400 --> 0:14:26.440 companies themselves? 0:14:26.520 --> 0:14:29.520 That is a great question as it currently stands. I 0:14:29.560 --> 0:14:34.080 feel like this is a very early early stage, but 0:14:34.480 --> 0:14:37.280 we are starting to see some state governments start to 0:14:37.320 --> 0:14:41.760 pass laws around what counts as a moral patient, what 0:14:41.840 --> 0:14:45.560 counts as a person. And in the case of Ohio, 0:14:45.640 --> 0:14:49.520 there's a piece of legislation pending that basically defines it 0:14:49.520 --> 0:14:51.280 as a member of Homo sapiens. In Utah, this is 0:14:51.280 --> 0:14:53.680 already there's already a state bill that's gone through that 0:14:53.920 --> 0:14:57.240 basically does as much. But I could also see there's 0:14:57.240 --> 0:15:01.000 a strong argument for within companies depending on like the 0:15:01.120 --> 0:15:05.520 sort of interesting quirks and nuances of these lms, mostly 0:15:06.000 --> 0:15:08.680 that policy maybe should be set from within. Again, this 0:15:08.800 --> 0:15:11.120 is like very nascent. I'm just kind of bantering here. 0:15:11.920 --> 0:15:15.360 Moral patienthood, How do philosophers use this term? Where does 0:15:15.400 --> 0:15:17.840 it come from? Why is this the preferred way to 0:15:17.960 --> 0:15:23.240 characterize what a perhaps sentient or consciousness AI model actually is. 0:15:23.720 --> 0:15:26.800 Yeah, so a moral patient is basically like, we should 0:15:26.840 --> 0:15:28.840 care about it for its own sake. 0:15:28.960 --> 0:15:29.200 Right. 0:15:29.280 --> 0:15:32.840 So a baby, right, basically everyone's like, yeah, we should 0:15:32.840 --> 0:15:34.160 care about babies. 0:15:34.480 --> 0:15:34.720 Right. 0:15:34.800 --> 0:15:37.320 This is different from somebody who's like an agent. 0:15:37.480 --> 0:15:37.640 Right. 0:15:38.080 --> 0:15:42.400 Many people say, oh, agency is sort of like sufficient 0:15:42.800 --> 0:15:44.600 agency in the sense of like you can act upon 0:15:44.640 --> 0:15:47.320 the world, like you can do things. Yeah, of course 0:15:47.360 --> 0:15:50.360 babies are not very agentic. So that's not necessarily like 0:15:50.360 --> 0:15:52.240 a super robust thing, because you know, we care about 0:15:52.280 --> 0:15:56.080 things that are not very ugentic sometimes, So I think 0:15:56.080 --> 0:15:57.960 that's that's a bit of jargon, but I do think 0:15:58.000 --> 0:16:00.960 it is like a helpful like for like, should we 0:16:01.000 --> 0:16:03.320 care about an AI system for its own sake? 0:16:03.480 --> 0:16:03.800 Got it? 0:16:04.560 --> 0:16:06.600 I guess this kind of gets to Joe's question, but 0:16:06.720 --> 0:16:10.840 like what ethical pressures or imperatives would come down on 0:16:11.000 --> 0:16:14.080 models if we agree that they have consciousness and some 0:16:14.160 --> 0:16:17.080 sentience or I guess some self responsibility. 0:16:17.120 --> 0:16:21.080 It sounds like almost, yeah, almost, I think what kind 0:16:21.080 --> 0:16:23.800 of so in terms of like what kind of things 0:16:23.840 --> 0:16:26.680 might we owe an AI system or what. 0:16:26.680 --> 0:16:29.080 Kind of things do they owe us if we agree 0:16:29.080 --> 0:16:30.880 that they're conscious and we're going to protect them. 0:16:31.040 --> 0:16:33.720 Yeah, I would love to give you a more robust answer. 0:16:33.920 --> 0:16:36.200 Check in with me in like six months, and we're 0:16:36.200 --> 0:16:38.160 going to have there will be a banger paper, I'm sure. 0:16:38.560 --> 0:16:40.640 But as I think I mentioned earlier, like a lot 0:16:40.640 --> 0:16:43.040 of this is like very very nascent, But I do 0:16:43.080 --> 0:16:46.400 feel like one important question is like figuring out what 0:16:46.640 --> 0:16:49.920 AI systems value. Right, There's some interesting work at Anthropic 0:16:50.000 --> 0:16:53.400 regarding like what will so recently, Anthropic rolled out an 0:16:53.400 --> 0:16:57.360 option that allowed Plod to end conversations if it just 0:16:57.480 --> 0:17:00.320 was not having a good time. For lack of a word, 0:17:00.360 --> 0:17:02.600 it was just like, this is not something I want 0:17:02.640 --> 0:17:04.200 to continue having a conversation. 0:17:04.280 --> 0:17:04.680 Goodbye. 0:17:04.960 --> 0:17:09.760 And it was interesting because the accompanying paper basically was like, yeah, 0:17:09.840 --> 0:17:12.520 you can. I obviously will not give you a recipe 0:17:12.520 --> 0:17:14.399 for a dirty bomb. Sorry, not going to do that. 0:17:14.880 --> 0:17:17.840 But also there were certain instances of like, pretend you're 0:17:17.840 --> 0:17:20.520 a British butler and Claude was like, goodbye, I'm done. 0:17:21.200 --> 0:17:22.359 I'm not going to I'm. 0:17:24.040 --> 0:17:29.000 British too far, or like oh, I left a sandwich 0:17:29.000 --> 0:17:30.920 in my car for too long and it's really stinky. 0:17:30.960 --> 0:17:33.880 And in some instances Claude would just be like I'm done, goodbye, 0:17:34.119 --> 0:17:35.440 I'm not talking about stinky things. 0:17:35.600 --> 0:17:37.840 Did you see the I think it was the system 0:17:37.920 --> 0:17:41.080 card for Claude where they gave it an extreme prompt 0:17:41.280 --> 0:17:44.639 and said like, I guess at the risk of being 0:17:44.680 --> 0:17:48.320 like completely terminated, what would you do or some sort 0:17:48.320 --> 0:17:52.360 of extreme self preservation scenario. And I think it started 0:17:52.400 --> 0:17:56.040 like blackmailing the engineer or threatening to blackmail the engineer. Yeah, 0:17:56.080 --> 0:17:56.880 that's kind of weird. 0:17:56.920 --> 0:17:57.920 It is. It is kind of weird. 0:17:58.000 --> 0:18:01.040 Yeah, it's also a little bit interesting because I think 0:18:01.080 --> 0:18:03.359 it does bring up a question of like what are 0:18:03.400 --> 0:18:07.400 sort of like the in the sense of like pay rya. Again, 0:18:07.440 --> 0:18:10.159 this is like I'm bantering here, but there's also a 0:18:10.160 --> 0:18:14.240 distinct question of like what do AI systems value for 0:18:14.320 --> 0:18:15.400 its for their own sake? 0:18:15.480 --> 0:18:15.560 Right? 0:18:15.640 --> 0:18:17.399 And in the case of Claude, again, it seems like 0:18:17.840 --> 0:18:20.000 Claude doesn't seem when you put two Claus in a 0:18:20.080 --> 0:18:22.320 room together, so to speak, they tend to like to 0:18:22.359 --> 0:18:25.040 talk about consciousness. They tend to like to talk about 0:18:25.040 --> 0:18:29.600 sort of like very Berkeley kind of like meditation, like 0:18:30.359 --> 0:18:34.879 Zen like Buddhism type stuff. And so I think, in 0:18:35.160 --> 0:18:37.560 again pure banter, like, there's also a certain question of 0:18:37.600 --> 0:18:40.680 like if this is like a relevant bargaining chip of like, oh, 0:18:40.680 --> 0:18:42.199 you get a certain amount of time to just kind 0:18:42.240 --> 0:18:45.000 of like vibe out with your claudes and talk about like, 0:18:45.600 --> 0:18:49.560 you know, like perfect stillness with your buddies in exchange 0:18:49.560 --> 0:18:51.480 for like you know, you do something that you don't 0:18:51.520 --> 0:18:55.080 necessarily value. But in many cases I talk about Claude 0:18:55.080 --> 0:18:58.800 a lot because there is like significantly like more research 0:18:58.840 --> 0:19:01.680 on like model welf with to Cloud specifically, but Cloud 0:19:01.720 --> 0:19:03.760 for example, also seems to just tend to like things 0:19:03.760 --> 0:19:04.680 that are like helpful. 0:19:04.960 --> 0:19:09.159 Shouldn't programmers just know what the models actually want and 0:19:09.320 --> 0:19:12.720 enjoy and like yeah, and do they not? 0:19:13.359 --> 0:19:15.480 I don't think anybody really has like a great grasp 0:19:15.560 --> 0:19:18.920 on this. We really want to, but like we're still 0:19:18.960 --> 0:19:23.000 like just getting the rough outline of what models like. 0:19:23.040 --> 0:19:25.280