WEBVTT - Alex Imas on Why Economists Might Be Getting AI Wrong 0:00:02.720 --> 0:00:15.640 Bloomberg Audio Studios, Podcasts, radio News. 0:00:18.079 --> 0:00:21.840 Hello and welcome to another episode of the Odd Lots podcast. 0:00:21.920 --> 0:00:23.600 I'm Joe Wisenthal. 0:00:23.160 --> 0:00:24.200 And I'm Tracy Alloway. 0:00:24.440 --> 0:00:26.440 Tracey, it may have changed a little bit in recent 0:00:26.480 --> 0:00:30.400 weeks or months, but I think by and large and large, 0:00:30.520 --> 0:00:33.440 like if you talk to economists about the long term 0:00:33.440 --> 0:00:37.839 impact of AI, particularly on jobs, by and large, it 0:00:37.920 --> 0:00:40.919 seems like they point to history and they say, there 0:00:40.920 --> 0:00:43.239 have been many technologies in the past that people that 0:00:43.320 --> 0:00:45.960 were going to be very disruptive and destroyal kinds of jobs, 0:00:46.440 --> 0:00:50.960 and in many cases they did. But technologies create new jobs. 0:00:51.000 --> 0:00:54.560 We can't necessarily anticipate them beforehand what they're going to be, 0:00:54.840 --> 0:00:57.680 and AI is like kind of no different ultimately. 0:00:57.360 --> 0:01:00.480 Yes, But then to your point, you ask, well, what 0:01:00.560 --> 0:01:03.520 specific jobs do you have in mind? And I get that, 0:01:03.800 --> 0:01:07.039 you know, it's hard to tell, it's hardly right. But 0:01:07.319 --> 0:01:10.800 it's so frustrating, right because here's this big new technology. 0:01:10.800 --> 0:01:13.680 It's supposed to be a productivity boost, and yet no 0:01:13.720 --> 0:01:17.680 one is actually sure what new jobs it's going to 0:01:17.720 --> 0:01:19.520 create from that productivity boost. 0:01:19.920 --> 0:01:22.360 I love him to death, but edam Uzamak wrote a 0:01:22.400 --> 0:01:26.440 piece several weeks ago and he was like, well, the 0:01:26.480 --> 0:01:31.080 player piano disrupted the existence of piano players, but hotels 0:01:31.120 --> 0:01:33.280 still pay money for a human who will have a 0:01:33.319 --> 0:01:36.320 piano player or a human actual piano player in the 0:01:36.319 --> 0:01:40.319 lobby rather at a player piano, which is true, but like, 0:01:40.520 --> 0:01:42.959 not many people have jobs that are equivalent. And the 0:01:43.080 --> 0:01:44.679 thing that like it's like, oh, you know, it's like 0:01:44.920 --> 0:01:48.320 I want to get like this insurance form reimbursed or whatever, 0:01:48.360 --> 0:01:51.880 this insurance reimbursed, Like I don't care about the human 0:01:51.920 --> 0:01:53.480 touch on that per se. 0:01:53.520 --> 0:01:55.440 I think there's something very happy to. 0:01:55.360 --> 0:01:57.000 Have the equivalent of the player piano there. 0:01:57.080 --> 0:01:59.400 There's something very just satisfying about the idea that we're 0:01:59.400 --> 0:02:02.160 all just going to be like performative in a way. 0:02:02.200 --> 0:02:05.160 But I actually think that's kind of where we might 0:02:05.200 --> 0:02:08.560 be heading, where like the sort of social skills I've 0:02:08.560 --> 0:02:12.960 said before, the looksmaxing, the personal branding, the multitasking, I 0:02:13.000 --> 0:02:14.760 guess like becomes more important. 0:02:15.200 --> 0:02:17.360 So the future is performative humanity. 0:02:17.680 --> 0:02:20.560 Open Ai just spent a ton of money on TVPN. 0:02:20.760 --> 0:02:23.480 I really love those guys. They're both very good looking guys. Man, 0:02:23.680 --> 0:02:25.919 So I sort of feel like, Okay, this is the 0:02:25.919 --> 0:02:28.600 biggest AI company in the world. Uh, sort of making 0:02:28.600 --> 0:02:29.120 a bet on it. 0:02:29.200 --> 0:02:33.040 It's like great characters, two very nice and charismatic humans. 0:02:33.120 --> 0:02:35.760 Yeah, yeah, yeah, so maybe that is the future just 0:02:35.760 --> 0:02:38.280 being nice and charismatic. Anyway, we need to talk more 0:02:38.320 --> 0:02:41.000 seriously about this because I don't I don't know. I 0:02:41.080 --> 0:02:43.600 kind of feel maybe this is not just going to 0:02:43.680 --> 0:02:46.160 be like the steam engine or whatever. It might be 0:02:46.240 --> 0:02:48.200 very different. Maybe we won't have jobs, maybe there will 0:02:48.240 --> 0:02:50.480 be new jobs. Anyway, someone who's been talking and thinking 0:02:50.600 --> 0:02:53.919 a lot about this and why AI might be different, 0:02:53.960 --> 0:02:56.440 we're going to be speaking. Really have the perfect guest. 0:02:56.520 --> 0:03:00.680 Alexeimasi is a professor of economics and Applied AI University 0:03:00.720 --> 0:03:02.799 of Chicago does a lot of writing on this topic. 0:03:02.880 --> 0:03:05.400 So Alex, thank you so much for coming on odd lots. 0:03:05.639 --> 0:03:06.520 Thank you for having me. 0:03:06.840 --> 0:03:09.040 It's pretty cool that you have the job of a 0:03:09.080 --> 0:03:12.119 professor of economics and apply to AI, Like, yeah, they 0:03:12.440 --> 0:03:14.360 worked out pretty well. It's good time you picked a 0:03:14.360 --> 0:03:15.640 good field. Yeah. 0:03:15.760 --> 0:03:17.480 Yeah, I mean I've been I've been an economists for 0:03:17.560 --> 0:03:19.720 much longer than I've been a professor of applied AI. 0:03:20.440 --> 0:03:24.160 I have been studying human behavior human decision making for 0:03:24.200 --> 0:03:27.720 about twelve years now, more than a decade, and when 0:03:27.800 --> 0:03:31.720 chad GPT first came out, I was kind of taken aback. 0:03:32.400 --> 0:03:34.360 This was a few years ago now, and I was thinking. 0:03:35.600 --> 0:03:37.520 After about a week of using it, I was like, 0:03:37.600 --> 0:03:40.320 this is going to be huge for the economy. And 0:03:40.360 --> 0:03:42.640 so I started talking to people who have kind of 0:03:42.680 --> 0:03:44.560 there were several people who kind of knew that it 0:03:44.600 --> 0:03:46.640 was coming and knew what the impact it was gonna 0:03:46.920 --> 0:03:48.440 it was going to have. So I started talking to 0:03:48.440 --> 0:03:51.920 those people and I kind of quickly kind of started retooling. 0:03:51.960 --> 0:03:54.400 I was smart. I started I trained my own model. 0:03:54.480 --> 0:03:56.960 You know, I got into cool. I got into it, 0:03:57.000 --> 0:03:59.520 and you know, it's I've been trying to play catchup 0:03:59.560 --> 0:04:00.000 ever since. 0:04:00.120 --> 0:04:01.640 Wait, what did you see in chat gpt? 0:04:01.760 --> 0:04:05.040 Specifically, because you would have been very early at that time, 0:04:05.120 --> 0:04:08.280 a lot of people were using chat gpt too, basically 0:04:08.360 --> 0:04:12.120 as a sort of enhanced search engine tool or write poems, 0:04:12.840 --> 0:04:15.840 tell silly jokes, whatever. But you saw something that was 0:04:16.279 --> 0:04:18.239 serious for the labor market. 0:04:18.640 --> 0:04:21.719 Yeah, I mean, once you started using it, you saw 0:04:21.800 --> 0:04:27.039 that it was able to basically not so well in 0:04:27.040 --> 0:04:29.800 the very very beginning, but even after a few months 0:04:29.800 --> 0:04:31.919 and like within the year, you saw that it was 0:04:31.920 --> 0:04:35.560 able to kind of do basic cognitive tasks to a 0:04:35.600 --> 0:04:37.919 decent degree. Like it wasn't like we are going to 0:04:37.960 --> 0:04:42.200 replace that person, but it was doing pretty sophisticated things 0:04:42.240 --> 0:04:45.400 that and in the jump from like where we were 0:04:45.440 --> 0:04:49.239 thinking about AI as these very very very targeted things 0:04:49.279 --> 0:04:52.440 like AI will play the game Go or something like 0:04:52.480 --> 0:04:56.200 that to something where WHOA, it can write an essay, 0:04:56.520 --> 0:04:59.400 it can tell me about this accounting property, it can 0:04:59.440 --> 0:05:01.680 make a four cast. All of a sudden, the generality 0:05:01.680 --> 0:05:05.440 of the technologies just exploded, And to me that was 0:05:05.480 --> 0:05:06.320 that was a huge deal. 0:05:06.520 --> 0:05:09.560 Yeah, the generality of it. I mean I guess literally 0:05:10.160 --> 0:05:13.520 that's the g right, Yeah, exactly, but yeah, no, I 0:05:13.560 --> 0:05:16.520 mean absolutely, I have to say there's a decide. But 0:05:16.680 --> 0:05:19.520 like learning a little bit more about like where AI 0:05:19.880 --> 0:05:23.840 was pre MS or pre CHAGYBT almost makes me even 0:05:23.880 --> 0:05:26.200 more impressed. Like the I don't know if like this 0:05:26.320 --> 0:05:27.880 is a common but when you like look at like 0:05:27.960 --> 0:05:31.119 some of like what was cutting edge in twenty nineteen, Yeah, 0:05:31.160 --> 0:05:32.840 and then you look at what's cutting edge in late 0:05:32.839 --> 0:05:36.279 twenty twenty two, I'm almost more impressed than if like 0:05:36.360 --> 0:05:38.400 I hadn't known what they were up to in twenty nineteen, 0:05:38.440 --> 0:05:40.320 Like it's a huge gap with those few years. 0:05:40.480 --> 0:05:42.600 It's a huge gap. But at the same time, like 0:05:42.720 --> 0:05:45.599 there were there, there was a kind of a path 0:05:45.680 --> 0:05:48.800 towards AI, and like the way that AI was being 0:05:49.800 --> 0:05:51.880 worked on for a long time, which was like these 0:05:52.000 --> 0:05:56.160 very specific purpose build technology, and I think Jeffrey Hinton 0:05:56.400 --> 0:05:58.920 and other people were kind of working on their own 0:05:58.960 --> 0:06:01.760 for a long time in the willderness of thinking like, 0:06:01.880 --> 0:06:03.880 maybe we can do something much more general than that. 0:06:03.960 --> 0:06:06.640 Maybe we can kind of come back to this idea 0:06:06.640 --> 0:06:09.640 of AGI versus these very specific tools. So then the 0:06:09.680 --> 0:06:13.320 whole term AGI, the general part of it. The reason 0:06:13.320 --> 0:06:15.960 that term came out was because in response to these 0:06:16.040 --> 0:06:18.839 very specific technologies that were being developed, which were by 0:06:18.960 --> 0:06:22.880 design not general. So somebody said Shane leg was one 0:06:22.960 --> 0:06:25.200 of the people who kind of, I think coined the term. 0:06:25.720 --> 0:06:30.760 He was saying, look, let's think about the general part 0:06:30.839 --> 0:06:33.719 of intelligence, and let's try to build a technology that 0:06:33.920 --> 0:06:36.200 is as general as the human mind. Let's go back 0:06:36.240 --> 0:06:37.120 to that starting point. 0:06:37.160 --> 0:06:39.120 So like, if someone makes a model that can tell 0:06:39.160 --> 0:06:42.400 the difference between written and spoken word, that's mind blowing, 0:06:42.440 --> 0:06:45.560 there's an incredible breakthrough, But that's not a general technology. 0:06:45.560 --> 0:06:47.320 That's a specific tell what time. 0:06:47.200 --> 0:06:50.000 Did you have in our betting book for Joe to 0:06:50.080 --> 0:06:51.560 refer to his vibe coding. 0:06:51.920 --> 0:06:53.680 I had two minutes thirteen seconds. 0:06:53.720 --> 0:06:55.840 Okay, so I made it longer. No, I made it 0:06:55.839 --> 0:06:56.480 a little bit longer. 0:06:56.800 --> 0:06:58.080 It's because I talked for so long. 0:06:58.240 --> 0:06:59.040 I'm sorry. 0:06:59.320 --> 0:07:01.680 No, fair enough, it's a fair point. I mean to me, 0:07:01.920 --> 0:07:03.520 Like the moment when things seem. 0:07:03.360 --> 0:07:06.200 To get very serious was the release with plod code. 0:07:06.240 --> 0:07:09.720 And at that point you went from like, okay, the 0:07:09.800 --> 0:07:12.000 model could not just tell you things, but it could 0:07:12.040 --> 0:07:15.320 actually do things for you. Was that the vibe shift 0:07:15.360 --> 0:07:18.280 that you anticipated or experienced as well? 0:07:18.840 --> 0:07:21.760 I mean, even though many people were talking about this, 0:07:21.920 --> 0:07:24.520 that this vibe shift was going to happen, people were 0:07:24.560 --> 0:07:27.720 telegraphing it for for months and months. Look, when agents 0:07:27.760 --> 0:07:30.800 start taking off, things are going to change as far 0:07:30.840 --> 0:07:34.559 as how people perceive this technology. Because the thing about 0:07:34.600 --> 0:07:38.280 agents versus just like the web based browsers, they can 0:07:38.360 --> 0:07:41.160 do stuff on your computer. They can say, like you 0:07:41.160 --> 0:07:43.560 could tell it like look, make me a spreadsheet. It 0:07:43.600 --> 0:07:45.800 will go and make you a spreadsheet using the tools 0:07:45.800 --> 0:07:48.680 that are available in your computer, not just say okay, 0:07:48.800 --> 0:07:51.240 here is how you would make a spreadsheet but you 0:07:51.280 --> 0:07:53.160 have to get your story, and that's a complete that's 0:07:53.160 --> 0:07:56.080 a paradigm shift as far as the economics of the technology. 0:07:56.360 --> 0:07:58.240 So I set up this sort of maybe it was 0:07:58.280 --> 0:07:59.920 a strong man, but I said up a sort of 0:08:00.040 --> 0:08:01.880 strong man that maybe we're going to knock down in 0:08:01.920 --> 0:08:06.880 this conversation. But how would you describe this sort of 0:08:07.640 --> 0:08:10.960 modal view of the impact of AI on the labor 0:08:11.040 --> 0:08:16.720 market among the economics profession to the extent there is one, So. 0:08:16.880 --> 0:08:18.920 I definitely think there is one. There was there's a 0:08:19.080 --> 0:08:22.800 very nice survey done by a whole team of people. 0:08:24.680 --> 0:08:28.040 Kevin Bryan was was one of them, and Basil Helper 0:08:28.080 --> 0:08:30.680 and was another. And they released the survey where they 0:08:30.920 --> 0:08:34.880 did they asked for forecasts for from economists and AI technologists. 0:08:34.880 --> 0:08:37.720 Now this is a self selected group of economists. These 0:08:37.720 --> 0:08:40.079 are economists who are working on AI. Okay, so it's 0:08:40.080 --> 0:08:42.360 not the whole field. But one of the things that 0:08:42.360 --> 0:08:45.560 you got from that that that survey was they're very 0:08:45.640 --> 0:08:50.120 much aligned. Okay, Right, So economists, at least the ones 0:08:50.120 --> 0:08:53.959 who are actually working and thinking about the technology, they 0:08:54.000 --> 0:08:57.000 think there will be a big impact as far as capabilities, 0:08:57.080 --> 0:09:01.439 and there will be some impact on the labor market astronomical, 0:09:01.440 --> 0:09:03.680 and we're talking about like twenty thirty, twenty fifty and 0:09:03.720 --> 0:09:08.079 things like that. There's going to be substantial capability increases, 0:09:08.559 --> 0:09:11.880 but the growth is going to be pretty moderate. It's 0:09:11.920 --> 0:09:14.679 like an extra two to three percent. And the really 0:09:14.679 --> 0:09:16.720 interesting thing for me from that survey was that the 0:09:16.760 --> 0:09:20.320 technologists were kind of a bit more optimistic than that 0:09:20.600 --> 0:09:24.320 as far as both the productivity growth and kind of 0:09:24.600 --> 0:09:27.920 somewhere kind of thinking that there will be much more unemployment. 0:09:28.320 --> 0:09:30.720 But for the most part, the two groups kind of agreed. 0:09:31.040 --> 0:09:33.800 I was personally surprised by that survey and this came 0:09:33.800 --> 0:09:36.000 out I think last week or two weeks ago. I 0:09:36.040 --> 0:09:37.560 thought that there was going to be a lot more 0:09:37.920 --> 0:09:39.280 daylight between the two groups. 0:09:40.160 --> 0:09:41.800 Well, the other thing that you tend to see is 0:09:41.840 --> 0:09:44.760 people release these charts of like which job is most 0:09:44.800 --> 0:09:47.640 exposed to AI, and it's usually like, you know, a 0:09:47.679 --> 0:09:50.920 knowledge worker at the top or something like that, your 0:09:51.000 --> 0:09:54.040 work is really interesting to us because you point out 0:09:54.080 --> 0:09:56.760 that a job is like much more than just the 0:09:56.800 --> 0:09:59.839 sector that you're actually working in. Tell us more about that. 0:10:00.480 --> 0:10:04.720 So the exposure measures they came from this literature, but 0:10:04.760 --> 0:10:08.880 mainly this this one paper by Daniel Rock and Pamela 0:10:08.960 --> 0:10:11.720 Michkin and co authors that were published in Science called 0:10:12.440 --> 0:10:15.920 one of the greatest titles is GPTs are GPTs GPT. 0:10:16.360 --> 0:10:18.280 You know what GPT is, but GPT in the second 0:10:18.320 --> 0:10:23.040 term is called general purpose technology. There they they basically 0:10:23.600 --> 0:10:27.440 started mapping jobs to the exposed as being exposed to 0:10:27.440 --> 0:10:29.880 to AI. But it's really important to understand what that 0:10:29.960 --> 0:10:34.520 number means. That number means that a AI could do 0:10:34.600 --> 0:10:39.199 fifty percent of a task, right, and how many tasks 0:10:39.320 --> 0:10:42.680 are in the job that AI can do fifty percent 0:10:42.800 --> 0:10:45.840 or more? Off. So there's a couple of things in 0:10:45.840 --> 0:10:50.280 that statement. First, fifty is not one hundred percent. That's obvious, right, 0:10:50.320 --> 0:10:52.280 So you still need a human in the loop if 0:10:52.400 --> 0:10:54.760 AI can do fifty percent. But two, it's the fact 0:10:54.800 --> 0:10:58.800 that a human job is a bunch of different tasks, right. 0:10:58.840 --> 0:11:01.600 So this is not a new point. David tour Has, 0:11:01.920 --> 0:11:04.079 you know, has has worked from the early two thousands 0:11:04.080 --> 0:11:06.480 with co authors on this, saying this is the task 0:11:06.559 --> 0:11:09.559 based model of jobs. They're on a SMOGLU has the 0:11:10.480 --> 0:11:14.920 canonical model on this, and the idea is that when 0:11:15.000 --> 0:11:16.839 we look at a job and we say, look, your 0:11:16.960 --> 0:11:22.360 job is exposed let's say it's fifty percent exposed. It 0:11:22.480 --> 0:11:26.720 really really matters what tasks in your job are exposed 0:11:26.760 --> 0:11:29.280 and how these tasks relate to one another. So let's 0:11:29.320 --> 0:11:31.320 say I have a job and I have a whole 0:11:31.320 --> 0:11:34.560 bunch of like completely meaningless garbage that I'm doing, but 0:11:34.800 --> 0:11:37.240 I have a comparative advantage and why I'm really getting 0:11:37.280 --> 0:11:40.240 paid for is like twenty thirty percent of the job. 0:11:40.559 --> 0:11:44.840 If AI is automating the kind of like meaningless kind 0:11:44.840 --> 0:11:47.280 of rote things in my job, I could take all 0:11:47.320 --> 0:11:50.640 of that time and I can focus on the job, 0:11:50.720 --> 0:11:53.160 on the parts of the job that are by comparative advantage. 0:11:53.160 --> 0:11:55.840 What does that mean? Means I'm going to become more productive, 0:11:56.320 --> 0:11:59.000 but I'm going to get paid more even though my 0:11:59.080 --> 0:12:02.599 job is really exposed. Now what does that mean for 0:12:02.679 --> 0:12:05.880 the labor market? Now you have to think, Okay, so 0:12:06.280 --> 0:12:07.280 a person is gonna. 0:12:07.040 --> 0:12:09.080 Get so just to be clear before we go any further, 0:12:09.240 --> 0:12:13.560 if if I'm working on a factory floor and one 0:12:13.559 --> 0:12:16.880 of my tasks is to pull a lever, like that 0:12:17.000 --> 0:12:19.520 is something that could presumably be automated, But if the 0:12:19.600 --> 0:12:23.559 other part of my work is to observe like how 0:12:23.720 --> 0:12:26.000 things are actually working on the floor and to report 0:12:26.080 --> 0:12:29.720 back to managers that might be something that's still valuable 0:12:29.960 --> 0:12:31.760 under our sort of AI future. 0:12:31.880 --> 0:12:34.679 And if the lever part gets automated, the theory is 0:12:35.200 --> 0:12:38.280 that not only you know, will Tracy be more productive 0:12:38.280 --> 0:12:39.319 and should get paid more for it? 0:12:39.440 --> 0:12:42.240 Yeah, exactly, Okay because of the increased products. Yeah right. 0:12:42.520 --> 0:12:45.000 This is the O Ring Model of jobs A VI. 0:12:45.160 --> 0:12:48.360 Goldfarb and Joshua Ganson this really nice paper. 0:12:48.520 --> 0:12:50.400 Can I just just a quick question here too, Like 0:12:50.679 --> 0:12:53.280 how good are we mean by we? I guess the 0:12:53.280 --> 0:12:57.880 economists who study this at like actually being able to, like, 0:12:58.200 --> 0:13:01.800 here's a job that someone has, is right down a 0:13:01.840 --> 0:13:05.760 list of these tasks. Describe? How good are we Pretty good? 0:13:05.880 --> 0:13:08.320 Describing actually pretty good? I would say on that dimension, 0:13:08.320 --> 0:13:12.560 we're pretty okay. There's the Ohnet database that has very 0:13:12.720 --> 0:13:15.880 very detailed records, and like here's a job, and here's 0:13:15.960 --> 0:13:17.920 like a whole vector of things that are involved in 0:13:17.920 --> 0:13:20.640 that job. Okay, So I'd say on that part, like 0:13:20.800 --> 0:13:23.840 just listing the tasks, pretty good. The thing that I 0:13:23.840 --> 0:13:26.680 think we're less good on is how those tasks relate 0:13:26.720 --> 0:13:29.600 to one another. This is the term called complementarity. 0:13:29.800 --> 0:13:30.760 Yeah, talk about that. 0:13:30.920 --> 0:13:35.640 So this is the weak links model is essentially saying, like, look, 0:13:35.960 --> 0:13:38.560 if tasks are completely separable. Let's say you know, I 0:13:38.600 --> 0:13:40.920 have a I pull a lever at my factory, and 0:13:41.000 --> 0:13:42.760 I talk to people on the factory floor, and these 0:13:42.800 --> 0:13:45.640 are completely independent. If I fail to pull the lever correctly, 0:13:45.840 --> 0:13:49.120 the other part of my job is unaffected. There's other 0:13:49.160 --> 0:13:51.320 parts of the job, like cooking. For example. Let's say 0:13:51.360 --> 0:13:53.840 I'm really good at ninety percent of the job, but 0:13:53.920 --> 0:13:57.120 like I really screw up the seasoning right, that meal 0:13:57.160 --> 0:13:59.600 tastes like garbage, right. 0:13:59.480 --> 0:14:01.079 So that succeeded in your task. 0:14:01.120 --> 0:14:04.400 You haven't succeeded on that. So when the tasks are interrelated, 0:14:04.960 --> 0:14:07.080 screwing up on one or two tasks means you did 0:14:07.080 --> 0:14:10.120 not complete your job. And it's basically is kind of 0:14:10.160 --> 0:14:13.560 almost a zero one sort of relationship. So the extent 0:14:13.600 --> 0:14:16.439 of that complementarity at how these tests are related will 0:14:16.480 --> 0:14:19.720 determine the extent to which automation is going to affect 0:14:19.720 --> 0:14:22.280 the labor market. And we don't have good numbers on that, so. 0:14:22.280 --> 0:14:24.440 This is really interesting. We're good at writing down the 0:14:24.520 --> 0:14:27.400 list of the task, yeah, we are not good at 0:14:27.520 --> 0:14:32.000 writing down the sort of like deep relational links to. 0:14:31.920 --> 0:14:34.760 The task and how they fit together exactly exactly, So 0:14:34.800 --> 0:14:38.200 that's something we need data on. The other part that 0:14:38.240 --> 0:14:42.560 we really need much more data on, and I recently 0:14:42.840 --> 0:14:45.600 was quoted as saying we need almost like a you know, 0:14:45.920 --> 0:14:51.600 Manhattan Project level effort on this is the This is 0:14:51.600 --> 0:14:55.000 a term from economics called elasticity of consumer demand, and 0:14:55.040 --> 0:14:57.480 that basically means how much will people buy more of 0:14:57.480 --> 0:14:58.640 something when the price changes. 0:14:59.280 --> 0:14:59.440 Right. 0:14:59.480 --> 0:15:03.440 So, let's say a person becomes a lot more productive, right, 0:15:03.880 --> 0:15:06.600 and they for the same sort of resources they can 0:15:06.720 --> 0:15:09.200 make a lot more of the product, their wage rises. 0:15:09.680 --> 0:15:12.440 What does that mean for the labor market? If they 0:15:12.480 --> 0:15:15.280 become more productive given the same kind of inputs, their 0:15:15.320 --> 0:15:18.160 wage rises. But also the firm's probably going to be 0:15:18.240 --> 0:15:21.160 paying less money to produce the same output. If it's 0:15:21.200 --> 0:15:23.480 a competitive industry, the prices are going to go down. 0:15:24.000 --> 0:15:27.240 If the consumers don't respond by buying a lot more 0:15:27.240 --> 0:15:30.160 of the product, the firm is going to fire a 0:15:30.160 --> 0:15:33.360 bunch of people because they can do more with less. 0:15:34.040 --> 0:15:37.280 But when prices come down, people buy way more of 0:15:37.280 --> 0:15:41.040 the product, then they might hire more of the same people. 0:15:41.480 --> 0:15:44.320 And in many sectors we've seen kind of the second 0:15:44.320 --> 0:15:48.040 thing play out what's then example, So people are arguing 0:15:48.040 --> 0:15:51.080 that software is actually one of those sectors. So there's 0:15:51.120 --> 0:15:53.760 been there's been a bunch of talk kind of looking 0:15:53.880 --> 0:15:58.560 historically at like what does productivity mean for the technology sector. 0:15:59.040 --> 0:16:03.120 It usually means a lot more consumer demand. So there's 0:16:03.160 --> 0:16:07.160 this really active debate now about what are coding agents 0:16:07.240 --> 0:16:10.320 actually going to do the software to software engineers. And 0:16:10.360 --> 0:16:13.840 some people are arguing, look, we have seen historically pretty 0:16:13.880 --> 0:16:16.960 elastic demand, and so we're going to potentially see a 0:16:17.000 --> 0:16:19.640 lot more hiring in that sector. And many people are 0:16:19.680 --> 0:16:23.080 saying this, but other people are saying, wait, maybe it's 0:16:23.120 --> 0:16:25.240 not as elastic as we as we think, and people 0:16:25.280 --> 0:16:28.280 are going to become so productive that we're really are 0:16:28.320 --> 0:16:30.040 going to see it down downside. 0:16:30.200 --> 0:16:32.360 That was kind of the argument that Jared Sweeper was 0:16:32.400 --> 0:16:34.800 making in our Defensive Software episodes. 0:16:34.920 --> 0:16:55.160 Yeah, you know, people are worried right about AI white 0:16:55.240 --> 0:16:56.200 color wipeout. 0:16:56.240 --> 0:16:56.720 I'm worried. 0:16:58.560 --> 0:17:01.640 So maybe the question should be what would have to 0:17:01.680 --> 0:17:06.920 be true about either the nature of AI capabilities or 0:17:07.000 --> 0:17:10.679 the relationship between tasks and job What would have to 0:17:10.720 --> 0:17:14.800 be true such that the scenario could unfold wipe out? 0:17:15.160 --> 0:17:19.160 Yeah? Two things. Well, let let me let me talk 0:17:19.160 --> 0:17:22.720 about three things. Yeah, one one is just full automation. Okay, right, 0:17:22.720 --> 0:17:25.160 that the models are so good that they just automate 0:17:25.200 --> 0:17:27.359 all of the tasks. That that probably that's like a 0:17:27.480 --> 0:17:30.960 very simple scenario to think about, because obviously people are 0:17:31.000 --> 0:17:34.800 gonna get fired, right if it's full fully automated. The 0:17:34.920 --> 0:17:37.240 other one is the one we've just been talking about 0:17:37.240 --> 0:17:40.880 where people become much more productive, but consumer demand is 0:17:40.920 --> 0:17:44.080 not elastic enough to absorb that extra production, so you're 0:17:44.080 --> 0:17:46.719 gonna have much fewer people doing a lot more stuff. 0:17:46.760 --> 0:17:49.879 So again, you're gonna have a lot of unemployment. The 0:17:49.920 --> 0:17:54.040 third thing is is related but is basically how many 0:17:54.200 --> 0:17:57.080 jobs each person has will determine the incentives of the 0:17:57.080 --> 0:18:00.480 company to actually invest in the automation technology. So let's 0:18:00.520 --> 0:18:04.159 talk about like the one task job. Let's say a 0:18:04.240 --> 0:18:07.840 person is just pulling the lever and let's say, right now, 0:18:07.880 --> 0:18:10.600 that doesn't even look exposed. Right, we look at the 0:18:10.640 --> 0:18:14.040 exposure graph. It doesn't look exposed. But let's say we're 0:18:14.119 --> 0:18:17.119 kind of getting kind of close and it just needs 0:18:17.320 --> 0:18:21.520 a bit more money to get to the automation switch. Well, 0:18:21.560 --> 0:18:24.439 the company has a lot higher incentive to invest that 0:18:24.560 --> 0:18:27.560 money if they know that if they invest that money, hey, 0:18:27.560 --> 0:18:30.320 they can get rid of that person completely, whereas they 0:18:30.320 --> 0:18:35.639 have less incentive when you know, let me invest in 0:18:35.760 --> 0:18:38.280 automating the lever poll If I know that, I can't 0:18:38.280 --> 0:18:41.000 fire the person because he's also a lot of stuff. 0:18:41.000 --> 0:18:42.760 So we have to think about the incentives of the 0:18:42.760 --> 0:18:44.959 firms to automated in the first place. These are large 0:18:45.600 --> 0:18:48.920 projects to do the automation. It's not like oh, Open 0:18:48.920 --> 0:18:52.880 Eye releases a model all of the companies adopted overnight, 0:18:53.640 --> 0:18:55.439 we see it in you know, a week later, we 0:18:55.480 --> 0:18:58.880 see the outcome. There's a lot of an organizational kind 0:18:58.920 --> 0:19:01.400 of going back and forth, a lot of systems need 0:19:01.440 --> 0:19:03.080 to be changed, all of this sort of thing, and 0:19:03.119 --> 0:19:05.720 so companies need to know, like, look, if I spend 0:19:05.760 --> 0:19:07.440 the money on it, I'm actually going to save money 0:19:07.440 --> 0:19:08.040 as a result. 0:19:08.400 --> 0:19:13.320 So, setting the archetypal guy pulling one lever aside, what 0:19:13.359 --> 0:19:16.600 are the real world jobs in your framework that are 0:19:16.600 --> 0:19:20.800 actually most exposed to AI risk? The one dimensional work? 0:19:21.080 --> 0:19:24.080 Yeah, I'm I hate to say one dimensional because every 0:19:24.119 --> 0:19:26.119 job is multidimensional. But if I had to make a 0:19:26.160 --> 0:19:30.640 guess where economists and other people should be kind of worried, 0:19:31.320 --> 0:19:35.600 I'd say stuff like truck driving and stuff like warehouse workers. 0:19:35.640 --> 0:19:38.840 Like if you google, you know warehouses built in China 0:19:38.960 --> 0:19:42.280 or something like that. These warehouses look nothing like what 0:19:42.320 --> 0:19:46.720 we think about warehouses. They're completely completely automated. They have 0:19:46.880 --> 0:19:50.359 robots like crawling on the walls. They're just there's no 0:19:50.480 --> 0:19:53.880 human in the loop at all. The the in these warehouses. 0:19:54.160 --> 0:19:57.840 And so oh, the warehouse gets automated, and then the 0:19:57.880 --> 0:20:01.119 warehouse gets automated. So part of that automation is going 0:20:01.160 --> 0:20:03.800 to be kind of loading that truck, and then the 0:20:04.200 --> 0:20:08.520 truck gets loaded through automation, and then that truck drives 0:20:08.560 --> 0:20:09.159 from A to B. 0:20:10.359 --> 0:20:14.359 It's interesting because you know, obviously a lot of people 0:20:14.480 --> 0:20:18.359 in freight will say the way you make that argument 0:20:18.400 --> 0:20:21.440 is very different. Then they'll say, well, yeah, driving a 0:20:21.520 --> 0:20:24.800 truck is much more than the driving part. Right, So 0:20:24.840 --> 0:20:28.160 it's like, okay, you could have a way mode truck, 0:20:28.400 --> 0:20:32.720 but who's gonna deliver? It's all who's delivered. 0:20:32.400 --> 0:20:34.960 Action is actually a big deal. Like if somebody stops 0:20:35.000 --> 0:20:37.680 it on the road a way more truck, they could 0:20:37.680 --> 0:20:39.960 just stop it on the road and rob the truck. Right, 0:20:40.000 --> 0:20:41.200 That's that's one element. 0:20:41.280 --> 0:20:44.119