Latest / Faster, Please! — The Podcast / An update on AI’s economic impact: My interview with Scott Strand of Google’s AI & Economy program
Transcript
- 0:00SPEAKER_00: Welcome to Faster Please,
- 0:01SPEAKER_00: the podcast, the audiovisual companion to my Faster Please newsletter.
- 0:05SPEAKER_00: I'm your host, Jim Pethokoukis.
- 0:07SPEAKER_00: My guest this episode is Scott Strand.
- 0:10SPEAKER_00: Scott is head of strategy,
- 0:11SPEAKER_00: operations,
- 0:12SPEAKER_00: and special projects at Google,
- 0:14SPEAKER_00: where he founded the Google Thanks so much for having me, Jim.
- 0:17SPEAKER_00: Happy to be here.
- 0:40SPEAKER_00: Right here in the beginning of,
- 0:44SPEAKER_00: I guess,
- 0:44SPEAKER_00: autumn or soon to be autumn, 2026,
- 0:47SPEAKER_00: what is your confidence level that large language models,
- 0:52SPEAKER_00: generative AI,
- 0:53SPEAKER_00: are what economists call a general purpose technology,
- 0:58SPEAKER_00: meaning it can affect an entire economy in a substantial way?
- 1:02SPEAKER_00: So classic one, steam power, electrification, the computer.
- 1:06SPEAKER_00: Are you confident that LLMs are more like those things than maybe like one step
- 1:13SPEAKER_00: below, like the shipping container or air conditioning.
- 1:16SPEAKER_00: What's your confidence level?
- 1:18SPEAKER_01: I'd put my confidence level at medium high on that front.
- 1:24SPEAKER_01: We've looked at this data through our Atlas research,
- 1:27SPEAKER_01: and we see AI having diffused throughout professional activities,
- 1:34SPEAKER_01: day-to-day life activities at a level that surprised even us.
- 1:38SPEAKER_01: So if you look at how Americans spend their waking hours in, say,
- 1:42SPEAKER_01: the ATIS survey,
- 1:44SPEAKER_01: AI is being used across 98% of those waking hours,
- 1:49SPEAKER_01: typically for value producing things.
- 1:53SPEAKER_01: Similar story with work.
- 1:55SPEAKER_01: 65% of occupations right now are using AI in some form,
- 1:59SPEAKER_01: and that's touching over 80% of the workforce.
- 2:03SPEAKER_01: They are generally useful for a lot of different things and,
- 2:09SPEAKER_01: yeah, appear to be becoming something very much like a GPT.
- 2:14SPEAKER_00: A lot of focus right now on economic impact.
- 2:18SPEAKER_00: I'm like, you know, when's the productivity boom going to happen?
- 2:21SPEAKER_00: I mean, I think about that all the time.
- 2:23SPEAKER_00: For most people, they probably think more about jobs.
- 2:26SPEAKER_00: Let me just start there.
- 2:29SPEAKER_00: We talked about usage,
- 2:31SPEAKER_00: but right now, is AI broadly affecting the US job market in any significant way?
- 2:40SPEAKER_00: I mean, how much do you have to dig before you can see the impact?
- 2:46SPEAKER_01: Man, you know,
- 2:47SPEAKER_01: I think you could dig the deepest hole you can think of,
- 2:51SPEAKER_01: and it would be hard to find the impact unless you were really,
- 2:56SPEAKER_01: really pushing for it.
- 2:58SPEAKER_01: Look, I mean, there's the basic stats that everyone can see right now.
- 3:00SPEAKER_01: Unemployment's at, you know, 4.1% or something like that on trend with...
- 3:05SPEAKER_01: More or less where it's been.
- 3:07SPEAKER_01: I will say in the data that we have that we've looked at,
- 3:10SPEAKER_01: it's not consistent with the story
- 3:14SPEAKER_01: of AI creating either certainly not widespread unemployment or even unemployment
- 3:21SPEAKER_01: in specific pockets.
- 3:24SPEAKER_01: Steve McLaughlin, Ph.D.:
- 3:52SPEAKER_01: The picture we're getting is that it's still pretty early days and
- 3:57SPEAKER_01: that like there is a ways to go in terms of adopting this technology
- 4:01SPEAKER_01: and reconfiguring workflows and businesses to make this technology more central
- 4:07SPEAKER_01: to how we work and to get the full productivity benefit out of it.
- 4:10SPEAKER_01: So I think it's,
- 4:15SPEAKER_01: you know,
- 4:16SPEAKER_01: our data is not consistent with the story
- 4:17SPEAKER_01: that AI is driving widespread unemployment.
- 4:20SPEAKER_00: Before I go any further, we're going to talk about two papers.
- 4:23SPEAKER_00: The first one is mapping Gemini usage in the economy.
- 4:27SPEAKER_00: Just briefly, how did you gather this information?
- 4:31SPEAKER_00: Where is this information coming from that you used to form some insights?
- 4:37SPEAKER_01: Yeah, yeah.
- 4:38SPEAKER_01: So we pulled 15 million de-identified human AI conversations from Gemini App,
- 4:46SPEAKER_01: that's the consumer app, gemini.google.com, Gemini API, and Search.
- 4:53SPEAKER_01: That's what you get to when you click AI mode on the search interface.
- 4:57SPEAKER_01: And we then categorized all of those conversations into
- 5:03SPEAKER_01: whether they represented work or not work, so professional usage or personal usage.
- 5:09SPEAKER_01: What I find super sort of intriguing about that is
- 5:39SPEAKER_00: I guess previously to try to figure out this kind of thing, you would have to do what?
- 5:43SPEAKER_00: Like surveys of CEOs or people at work and they would have to say how they were using it.
- 5:49SPEAKER_00: And I imagine one,
- 5:50SPEAKER_00: you're going to get a lot more data
- 5:52SPEAKER_00: and you're not just getting what people say they're doing,
- 5:54SPEAKER_00: but you actually get what people are doing.
- 5:56SPEAKER_00: That would have to seem to me that as an economist,
- 5:58SPEAKER_00: you know, that would have to be tremendously helpful.
- 6:01SPEAKER_01: It is tremendously helpful.
- 6:02SPEAKER_01: Yeah, there's a huge gap between how productive people think they are
- 6:06SPEAKER_01: and how productive they actually are and the types
- 6:09SPEAKER_01: of things they think they're doing in their days and what they're actually doing
- 6:11SPEAKER_01: in their days.
- 6:12SPEAKER_01: So empirical data always rules the roost as far as we're concerned.
- 6:17SPEAKER_01: So it's very helpful.
- 6:18SPEAKER_01: I mean,
- 6:19SPEAKER_01: what's also fantastic is that gathering and classifying a data set of this size
- 6:24SPEAKER_01: before you had LLMs would have taken, gosh, I don't know, a team of 25, 30 people.
- 6:33SPEAKER_01: We were able to do this with a pretty small team in a matter of weeks.
- 6:40SPEAKER_01: So yeah, it's exciting new territory in terms of social science research.
- 6:46SPEAKER_01: I think teams are going to be able to get a lot more productive in the years to come.
- 6:59SPEAKER_00: Is it kind of what you would, because you early say, no, it's still early days.
- 7:03SPEAKER_00: So when you look at what they're doing,
- 7:06SPEAKER_00: based on what,
- 7:08SPEAKER_00: you know,
- 7:09SPEAKER_00: the adoption of past important technologies, is it kind of what you would expect?
- 7:15SPEAKER_00: You know, still fairly shallow.
- 7:17SPEAKER_00: Is that what you would expect?
- 7:18SPEAKER_00: Sort of what they rolled out chat GPT at the end of 2022.
- 7:22SPEAKER_00: We're sitting here in fall of 2027.
- 7:25SPEAKER_00: Not very long, really.
- 7:26SPEAKER_00: And is it kind of what you'd expect to see?
- 7:28SPEAKER_01: Yeah, it is and it isn't.
- 7:31SPEAKER_01: So it is in so far as it's still relatively shallow.
- 7:35SPEAKER_01: It's early days.
- 7:39SPEAKER_01: It's being used more for technically complex uses, as you might expect.
- 7:47SPEAKER_01: The software engineers use AI meme is very much true.
- 7:51SPEAKER_01: But there were some things that we found Pretty surprising.
- 7:55SPEAKER_01: One is with past technological innovations,
- 7:59SPEAKER_01: the routine tasks are the things that get automated first,
- 8:01SPEAKER_01: the things that are repeatable.
- 8:03SPEAKER_01: You do them over and over in the same way.
- 8:05SPEAKER_01: That is not what we saw in our data.
- 8:08SPEAKER_01: We saw I saw that most of AI usage focuses on non-routine tasks,
- 8:13SPEAKER_01: the types of things like creative design,
- 8:16SPEAKER_01: hypothesis testing,
- 8:17SPEAKER_01: doing something that's a little bit out of the ordinary from your job description.
- 8:21SPEAKER_01: Those make up about a third of the tasks in the average American worker's job basket.
- 8:28SPEAKER_01: It represents about 65% of their AI usage.
- 8:31SPEAKER_01: So they're really over-indexing on using AI for this type of work.
- 8:36SPEAKER_01: Poses an interesting challenge to the routine task intensity model
- 8:41SPEAKER_01: where you expect every wave of technological innovation to automate
- 8:47SPEAKER_01: the repetitive stuff.
- 8:48SPEAKER_01: But that's not quite what we see happening.
- 8:51SPEAKER_01: The other thing that really surprised us was the amount of blue collar
- 8:54SPEAKER_01: and manual occupational usage in the data.
- 8:56SPEAKER_01: We came in honestly just expecting to see not very much.
- 8:59SPEAKER_01: That's certainly what our peer laboratories found.
- 9:02SPEAKER_01: But we saw a pretty significant amount.
- 9:07SPEAKER_01: Auto technicians using our models to diagnose engine wear and tear.
- 9:13SPEAKER_01: We see electricians using it to sketch out wiring diagrams and inspect what might be going on.
- 9:19SPEAKER_01: And what's even cooler is we see them using it in a more sophisticated way than
- 9:24SPEAKER_01: a lot of white collar workers.
- 9:26SPEAKER_01: They use it at twice the rate of images and video,
- 9:30SPEAKER_01: so multimodal usage, than white collar workers.
- 9:35SPEAKER_01: So they're really engaging with these tools in a much broader and more diverse way
- 9:39SPEAKER_01: than other types of occupations.
- 9:42SPEAKER_01: So those are a couple of surprising things.
- 9:44SPEAKER_00: Another confidence level question.
- 9:45SPEAKER_00: What is your confidence level that this is all going to add up to something?
- 9:52SPEAKER_00: That at some point here,
- 9:55SPEAKER_00: there'll be the quarterly productivity data will be released and we'll say,
- 10:01SPEAKER_00: That's AI right there.
- 10:03SPEAKER_00: Now we know.
- 10:04SPEAKER_00: They certainly had that moment in the 1990s.
- 10:07SPEAKER_00: So are you confident we're going to have that moment where all these gains,
- 10:11SPEAKER_00: the auto technicians and people doing white-collar workers summarizing things
- 10:16SPEAKER_00: or using them for different,
- 10:16SPEAKER_00: that it's going to add up to a measurable productivity boost.
- 10:21SPEAKER_00: I'm not going to say,
- 10:22SPEAKER_00: I'm not going to make you give me a number, but that it will be significant.
- 10:27SPEAKER_00: What you would expect if this is a real GPT.
- 10:31SPEAKER_01: Yeah, for sure.
- 10:31SPEAKER_01: So, you know, I think I fall somewhere between the uber pessimists and the uber optimists.
- 10:37SPEAKER_01: You know,
- 10:37SPEAKER_01: you've seen everything from like a quarter of a percentage point hat on to well
- 10:43SPEAKER_01: over 10.
- 10:44SPEAKER_01: I'm somewhere in the middle of that.
- 10:46SPEAKER_01: I think that...
- 10:48SPEAKER_01: Look,
- 10:49SPEAKER_01: we wouldn't see this extent of professional usage unless people were finding it
- 10:54SPEAKER_01: to actually be helpful in their day-to-day job,
- 10:56SPEAKER_01: to get more done, to be more efficient.
- 10:58SPEAKER_01: I think market incentives are very real and you're not just messing around with AI
- 11:05SPEAKER_01: at work just for fun and games.
- 11:06SPEAKER_01: If you're using it professionally,
- 11:08SPEAKER_01: because it can either help you get something done faster or get it done better.
- 11:12SPEAKER_01: And, you know,
- 11:14SPEAKER_01: we actually released a report a couple of days ago on this where we worked with
- 11:18SPEAKER_01: the economist David Autor to provide a work-in-process AI tool to a group
- 11:25SPEAKER_01: of patent lawyers where it helped them with patent drafting.
- 11:30SPEAKER_01: We had a panel of other lawyers rate the quality of their work
- 11:35SPEAKER_01: and also examine how long it took them to do that work.
- 11:38SPEAKER_01: And when they had the tool,
- 11:40SPEAKER_01: they did demonstrably better work,
- 11:42SPEAKER_01: demonstrably more quickly by very, very objective standards.
- 11:46SPEAKER_01: Then we took that tool away and they slowed down and the quality dipped back to the baseline.
- 11:54SPEAKER_01: In a micro sense, the productivity benefits are very real and observable.
- 11:59SPEAKER_01: As to when they start to show up in the macro, I don't know.
- 12:03SPEAKER_01: I think a lot of it is going to require a lot of things.
- 12:06SPEAKER_01: It's going to require teaching people to use these tools to their maximum potential.
- 12:11SPEAKER_01: It's going to require reorganizing businesses around it.
- 12:14SPEAKER_01: When they first introduced electricity,
- 12:17SPEAKER_01: They needed to retool entire factory floors in order
- 12:20SPEAKER_01: to make electricity available throughout the factory,
- 12:23SPEAKER_01: and only then were they really able to make full use of it.
- 12:26SPEAKER_01: So I think we're probably going to see something similar with AI.
- 12:29SPEAKER_00: So sometimes people will speak of this technology in a quasi-magical way,
- 12:36SPEAKER_00: but when you look And what is happening?
- 12:41SPEAKER_00: It seems like maybe it's a really important technology,
- 12:44SPEAKER_00: but as far as how businesses are adopting it and what they're going to have to do
- 12:48SPEAKER_00: to make productive use of it,
- 12:50SPEAKER_00: it's kind of like what economists would have predicted back at the end of 2022.
- 12:54SPEAKER_01: Yeah, I think that's right.
- 12:58SPEAKER_01: I mean, it It's an exceptional technology.
- 13:03SPEAKER_01: It's an incredible thing,
- 13:04SPEAKER_01: but you're not going to get the benefits out of it unless your workers know how
- 13:07SPEAKER_01: to use it and you've redesigned workflows and processes around it.
- 13:13SPEAKER_01: It is kind of cool in that it's the first tool that teaches you how to use it.
- 13:17SPEAKER_01: If you just start engaging with it, it kind of instructs you on how to get the most out of it.
- 13:21SPEAKER_01: So that's helpful.
- 13:23SPEAKER_01: But even still,
- 13:23SPEAKER_01: like Until businesses start to remove a lot of the manual steps and sort
- 13:32SPEAKER_01: of non-AI empowered workflows,
- 13:35SPEAKER_01: it's going to be tough to get the full benefit of the tech.
- 13:38SPEAKER_00: Yeah, and that's something where I'll talk about a couple of minutes with your second paper.
- 13:43SPEAKER_00: But one interesting finding in this paper was that so far,
- 13:46SPEAKER_00: this is more a story of sort of augmentation than automation.
- 13:50SPEAKER_00: As you wrote in the paper,
- 13:52SPEAKER_00: while there is some task automation,
- 13:55SPEAKER_00: the vast majority of you so far is collaborative and assistive to tasks and work.
- 14:01SPEAKER_00: Because people obviously worried about job loss.
- 14:04SPEAKER_00: But we should expect to see both things happening,
- 14:07SPEAKER_00: task automation,
- 14:08SPEAKER_00: as well as augmentation,
- 14:10SPEAKER_00: as well as I would imagine, like figuring out new things to do.
- 14:14SPEAKER_01: Oh, yeah.
- 14:15SPEAKER_01: Yeah, absolutely.
- 14:17SPEAKER_01: Absolutely.
- 14:17SPEAKER_01: The way we talk about this inside of Google a lot of times is that most jobs will be changed.
- 14:24SPEAKER_01: Some jobs will be lost.
- 14:25SPEAKER_01: Some jobs will be gained.
- 14:27SPEAKER_01: And that's just how things tend to go in major economic transitions.
- 14:31SPEAKER_01: And the key is How can you support people,
- 14:36SPEAKER_01: the workers through this transition to set them up well for what comes next?
- 14:43SPEAKER_01: I do think that we're going to see steadily expanding automation over time.
- 14:48SPEAKER_01: We're already working on the next iteration of this report where we're going
- 14:52SPEAKER_01: to start to establish a trend line around some of these automation stats and be able
- 14:56SPEAKER_01: to see if the trend line is, you know, what the angle of the line is.
- 15:01SPEAKER_01: So we're excited for that.
- 15:03SPEAKER_01: But yeah, I guess the only other thing I'd say is that AI is a very jagged frontier.
- 15:09SPEAKER_01: There are some things that AI is going to be very good at automating, software code review.
- 15:17SPEAKER_01: There are some things it might find difficult to automate,
- 15:22SPEAKER_01: things like sort of hypothesis generation for complex tasks or clinical validation
- 15:27SPEAKER_01: of a given trial or things like that.
- 15:30SPEAKER_01: So yeah, I guess it all gets back to its early days.
- 15:34SPEAKER_00: One issue comes up,
- 15:35SPEAKER_00: and I want to see if you had a thought on it,
- 15:37SPEAKER_00: is a lot of concern about, like,
- 15:39SPEAKER_00: entry-level work, that,
- 15:42SPEAKER_00: you know,
- 15:42SPEAKER_00: that they're going to start,
- 15:43SPEAKER_00: and that sort of feeder pipeline will end, and, like, That's bad.
- 15:48SPEAKER_00: Like you need people to,
- 15:49SPEAKER_00: kids have to get jobs and they have to, you know, get more, gain more expertise.
- 15:53SPEAKER_00: And I think you found this,
- 15:56SPEAKER_00: that if you already have,
- 15:57SPEAKER_00: and I certainly found it in my own life,
- 15:59SPEAKER_00: my level of expertise in a topic allows me to use this technology much better.
- 16:04SPEAKER_00: I get much more out of it.
- 16:05SPEAKER_00: So is that like initial step?
- 16:08SPEAKER_00: What do we know at all about that?
- 16:11SPEAKER_00: Or is it still too soon to tell?
- 16:14SPEAKER_01: So based on our data, it's still too soon to tell.
- 16:17SPEAKER_01: It's not clear.
- 16:18SPEAKER_01: I mean, what you said is exactly right.
- 16:20SPEAKER_01: We found in our data that a lot of the usage was people
- 16:25SPEAKER_01: in high expertise roles using it for lower expertise tasks,
- 16:30SPEAKER_01: the types of tasks that entry-level workers often take on.
- 16:33SPEAKER_01: And obviously, this is concerning.
- 16:36SPEAKER_01: And we can't confirm or evaluate whether or not it's a dampening entry-level Hiring,
- 16:45SPEAKER_01: but I think that this is something that we need to look
- 16:48SPEAKER_01: at really closely going forward.
- 16:52SPEAKER_01: One thing that I find it to be a useful frame for thinking
- 16:56SPEAKER_01: about this is David Otter's expertise framework,
- 17:00SPEAKER_01: where if you think about
- 17:06SPEAKER_01: the expertise requirements for a given job to sit on
- 17:08SPEAKER_01: a distribution curve of the number of tasks...
- 17:11SPEAKER_01: Some technologies automate away the low expertise requirements related to that role,
- 17:17SPEAKER_01: which raises the average expertise level required for that job.
- 17:22SPEAKER_01: This is kind of what happened with, say, accountants when spreadsheets were released.
- 17:27SPEAKER_01: You didn't really need to have people doing sort of tables of numbers anymore,
- 17:31SPEAKER_01: but the computer took care of that, but that freed up.
- 17:35SPEAKER_01: A lot of time for accountants to specialize in tax strategy and things of that nature.
- 17:39SPEAKER_01: And so that meant those accountants, well,
- 17:43SPEAKER_01: became scarcer because that expertise is rare and they were paid more as a result.
- 17:47SPEAKER_01: So, you know, flat to declining employment prospects with higher incomes.
- 17:54SPEAKER_01: On the flip side,
- 17:55SPEAKER_01: you can think about technologies that automate away some of
- 17:58SPEAKER_01: the higher expertise tasks.
- 18:00SPEAKER_01: The more complex work, say,
- 18:04SPEAKER_01: Classic example here is taxis and Uber drivers.
- 18:08SPEAKER_01: So once the Uber app was released,
- 18:11SPEAKER_01: like complex routing,
- 18:13SPEAKER_01: like detailed knowledge of the back alleys in every city,
- 18:16SPEAKER_01: that was no longer really a requirement to be a good taxi driver.
- 18:19SPEAKER_01: You could just kind of get your car and get on the road and go.
- 18:21SPEAKER_01: You also didn't even need a dispatch.
- 18:23SPEAKER_01: You could just get out there.
- 18:25SPEAKER_01: So what happened was then we had a flood of new entrants,
- 18:28SPEAKER_01: new drivers in cities, but the average wages for those occupations decreased.
- 18:33SPEAKER_01: It became a less well-paid, less well-compensated profession.
- 18:35SPEAKER_01: And so you get this dynamic where there is rising employment, but sort of flat to lower wages.
- 18:43SPEAKER_01: So I'm getting long-winded here,
- 18:46SPEAKER_01: but I say all this to make the point that The employment effects of AI,
- 18:50SPEAKER_01: I think, are going to be extraordinarily complex.
- 18:52SPEAKER_01: And it's going to be a lot more nuanced and detailed than just entry level jobs are going away.
- 18:57SPEAKER_01: Like, I think some jobs will probably be paid more with slower employment growth.
- 19:01SPEAKER_01: I think some jobs might see lower hiring.
- 19:05SPEAKER_01: And I think some jobs might see more hiring with flatter wages.
- 19:10SPEAKER_01: These are all the kinds of things we're looking into.
- 19:12SPEAKER_00: I think,
- 19:13SPEAKER_00: and you get at this in the paper,
- 19:14SPEAKER_00: and I think this is underappreciated,
- 19:16SPEAKER_00: sort of the value this technology can provide For what you do at home.
- 19:22SPEAKER_00: I always thought people still undervalue like the value of YouTube
- 19:26SPEAKER_00: and helping you do all manner of sort of like household jobs, figure things out.
- 19:32SPEAKER_00: You know, if you if you, know, lost the instruction.
- 19:35SPEAKER_00: This is something I didn't think about that you mentioned,
- 19:37SPEAKER_00: just like interacting with government in some way,
- 19:38SPEAKER_00: government-related, like all that.
- 19:40SPEAKER_00: I don't know if that shows up in GDP, but it certainly shows up sort of like
- 20:04SPEAKER_00: It's hugely
- 20:09SPEAKER_01: significant.
- 20:11SPEAKER_01: 86% of AI interactions are happening outside of works.
- 20:14SPEAKER_01: And we see people using it for really valuable things,
- 20:19SPEAKER_01: things like seeking professional services,
- 20:22SPEAKER_01: navigating legal questions,
- 20:25SPEAKER_01: discussing how to engage with government and civic services.
- 20:29SPEAKER_01: That actually overindexes by 20 times relative to the amount of time people spend there.
- 20:34SPEAKER_01: And about half of that interaction with AI happens on nights and weekends
- 20:39SPEAKER_01: when government offices are closed.
- 20:40SPEAKER_01: So it raises a really interesting idea that AI may be helping
- 20:45SPEAKER_01: to extend civil society in some way and help people get more out
- 20:49SPEAKER_01: of their governments,
- 20:49SPEAKER_01: whether that's benefits or just clearing up questions on fines, et cetera.
- 20:55SPEAKER_01: I mean, like I...
- 20:57SPEAKER_01: I'm a parent.
- 20:57SPEAKER_01: I'm a really busy parent.
- 20:59SPEAKER_01: And so I just think of like tired parents sitting at the kitchen table at 10
- 21:04SPEAKER_01: at night,
- 21:05SPEAKER_01: like trying to figure out how to fix a broken dishwasher
- 21:07SPEAKER_01: or like translate some zoning notice from the city into something that they can do
- 21:12SPEAKER_01: and act on.
- 21:13SPEAKER_01: And AI is really providing real value there with real like dollar figure attached to that.
- 21:18SPEAKER_01: But it doesn't go measured.
- 21:19SPEAKER_01: This is kind of like how the introduction of the washing—I think the introduction of
- 21:23SPEAKER_01: the washing machine was one of the greatest economic triumphs of the 20th century.
- 21:27SPEAKER_01: But because it predominantly just saved on labor inside the home and,
- 21:31SPEAKER_01: to be frank, labor for women, it did not go measured.
- 21:36SPEAKER_01: And it should be measured.
- 21:37SPEAKER_01: And we need to figure out a way to measure stuff like that
- 21:40SPEAKER_01: because it's hugely important to society's well-being.
- 21:45SPEAKER_00: Since you brought the washing machine,
- 21:47SPEAKER_00: I'm going to urge all the listeners and viewers to try to check out
- 21:50SPEAKER_00: an old BBC show called 1900 House,
- 21:54SPEAKER_00: which took a modern family and had them live like a pretty well-to-do family
- 21:58SPEAKER_00: in 1900.
- 21:59SPEAKER_00: And so much of that show revolved around doing the wash.
- 22:05SPEAKER_00: No, they had servants, but still a brutal, time-consuming, back-breaking labor.
- 22:10SPEAKER_00: I'm glad you brought that up.
- 22:12SPEAKER_01: That was life.
- 22:12SPEAKER_01: Just wash your stuff.
- 22:14SPEAKER_01: That's it.
- 22:14SPEAKER_01: Yeah.
- 22:16SPEAKER_00: Let me jump to the second paper before we run out of time,
- 22:18SPEAKER_00: which is AI and Science, Early Insights.
- 22:22SPEAKER_00: Then a similar methodology,
- 22:25SPEAKER_00: you looked at what scientists were doing with AI models, like what they were asking.
- 22:30SPEAKER_01: Yeah.
- 22:31SPEAKER_01: Yeah, exactly.
- 22:32SPEAKER_01: Exactly.
- 22:33SPEAKER_01: So this one was really cool.
- 22:35SPEAKER_01: I love this paper.
- 22:37SPEAKER_01: And like full credit to my authors,
- 22:38SPEAKER_01: my co-authors, Mihaly, Alex Amis, and Juan for doing an amazing job with this.
- 22:44SPEAKER_01: What makes this really cool, a few things make it really cool.
- 22:48SPEAKER_01: One is we worked with MIT to do this in the context of an entirely new way
- 22:53SPEAKER_01: of categorizing scientific tasks.
- 22:56SPEAKER_01: So the economics wonks who listen to your show will know of O-Net very,
- 23:00SPEAKER_01: very well,
- 23:00SPEAKER_01: like very clearly well-defined taxonomy for how we think about tasks at work.
- 23:06SPEAKER_01: Nothing's really existed like that for science before, but MIT's future tech team built one.
- 23:11SPEAKER_01: And so we were able to take this taxonomy,
- 23:13SPEAKER_01: apply it to Google data,
- 23:15SPEAKER_01: and analyze, all right, which scientific tasks are people using AI for?
- 23:21SPEAKER_01: And we looked at both our data,
- 23:25SPEAKER_01: bibliometric data across over 2,600 specialized AI models across a wide range
- 23:30SPEAKER_01: of disciplines.
- 23:31SPEAKER_01: We surveyed 600 scientists.
- 23:34SPEAKER_01: We found a lot of interesting stuff, but I think two big takeaways I'd highlight.
- 23:41SPEAKER_01: One is that there is a really clear division of labor between general purpose LLMs
- 23:48SPEAKER_01: and specialized models.
- 23:50SPEAKER_01: There's a lot of debate in the scientific community on
- 23:52SPEAKER_01: whether your basic chatbot will handle all the tasks you need or whether you need
- 23:57SPEAKER_01: a really specialized model for science.
- 24:00SPEAKER_01: Scientists are doing a really good job at parsing work between the two models,
- 24:05SPEAKER_01: and they tend to use specialized models more for domain-specific predictions,
- 24:10SPEAKER_01: data generation, stuff like that.
- 24:12SPEAKER_01: And we found that scientists were getting a lot of productivity gains out of this.
- 24:16SPEAKER_01: They're saving about seven hours per week,
- 24:18SPEAKER_01: which they're mostly just reinvesting back in more research, which is huge.
- 24:24SPEAKER_01: So it's not all good news.
- 24:28SPEAKER_01: What's happening is that scientists are able to generate a lot more hypotheses,
- 24:32SPEAKER_01: which is great,
- 24:33SPEAKER_01: but now there's a backlog of hypotheses to test,
- 24:35SPEAKER_01: and there are all the bottlenecks on research are starting to shift downstream
- 24:39SPEAKER_01: to things like physical experimentation and clinical validation.
- 24:43SPEAKER_01: So yeah, lots of, I mean, exciting stuff, but still lots of work to do.
- 24:49SPEAKER_00: What I found interesting is that when people hear about AI and doing science,
- 24:54SPEAKER_00: I think what they imagine is You give a chatbot a bunch of data and then say,
- 25:01SPEAKER_00: cure this disease or come up with something stronger than steel.
- 25:05SPEAKER_00: You know,
- 25:07SPEAKER_00: that it's going to do this amazing idea generation that then is up to you,
- 25:12SPEAKER_00: which is great.
- 25:13SPEAKER_00: I hope we can do that.
- 25:14SPEAKER_00: But just acting as kind of a super research assistant,
- 25:20SPEAKER_00: making scientists more productive at what they...
- 25:24SPEAKER_00: And all the tasks they do, I mean, there's a value to that.
- 25:27SPEAKER_00: Again, much like I think what happens at home using AI,
- 25:30SPEAKER_00: like the value of making scientists more productive in what they do,
- 25:34SPEAKER_00: setting aside, coming up with, you know, fantastic ideas.
- 25:37SPEAKER_00: Like, again,
- 25:38SPEAKER_00: real value,
- 25:40SPEAKER_00: underappreciated,
- 25:41SPEAKER_00: but like all that productivity,
- 25:43SPEAKER_00: that seven hours a week across all scientists adds up.
- 25:47SPEAKER_01: It really adds up.
- 25:48SPEAKER_01: It really does.
- 25:49SPEAKER_01: And yeah, I mean, I'm sure you've spent time in academia.
- 25:53SPEAKER_01: RAs are a precious resource.
- 25:56SPEAKER_01: Good RAs are really hard to find.
- 25:58SPEAKER_01: And to have an assistant that can do things like Forecasting the structural property
- 26:05SPEAKER_01: of a given molecule 24-7 or building a synthetic data set or grouping
- 26:10SPEAKER_01: and classifying data never sleeps.
- 26:12SPEAKER_01: It just keeps pushing out that task.
- 26:14SPEAKER_01: That's a really valuable thing for people.
- 26:17SPEAKER_00: I think also that scientific progress is the bedrock of all
- 26:20SPEAKER_01: productivity gains.
- 26:22SPEAKER_01: This is how we actually become wealthier as a society and get better lives and move forward.
- 26:28SPEAKER_01: And so this idea of like And just to finish up,
- 26:33SPEAKER_01: you mentioned the first paper,
- 26:34SPEAKER_01: you're going to be doing another version of that and kind of get a trend line going.
- 26:37SPEAKER_01: Any other sort of sneak previews and things you're working on
- 26:41SPEAKER_01: or just
- 26:43SPEAKER_00: interesting questions you would like to work on?
- 27:00SPEAKER_01: Yeah, we've got all kinds of things in the lab.
- 27:02SPEAKER_01: We are so excited about this.
- 27:04SPEAKER_01: And this is a long-term piece of work, so we're going to be doing this for a while.
- 27:08SPEAKER_01: I think we're really interested in the emergence of new work,
- 27:14SPEAKER_01: figuring out what is this post-AI diffusion world,
- 27:17SPEAKER_01: what types of jobs, tasks are emerging.
- 27:22SPEAKER_01: I think when Google acquired YouTube in the 2010s,
- 27:26SPEAKER_01: there's no way we could have predicted that influencer would become such a sort
- 27:31SPEAKER_01: of popular career path for people and sort of be something that
- 27:35SPEAKER_01: so many folks were interested in.
- 27:36SPEAKER_01: I think we similarly like we're going to see some wild new career paths coming out
- 27:42SPEAKER_01: of the spread of AI and we're really excited to figure out what those are.
- 27:47SPEAKER_01: We're really interested to dig into more global analysis,
- 27:51SPEAKER_01: looking more on the development economic side of things,
- 27:54SPEAKER_01: seeing AI's effects on the digital divide and how it's being used by folks across
- 28:00SPEAKER_01: the global north and south.
- 28:03SPEAKER_01: We're excited about manual occupations and huge We have a swath of the global workforce.
- 28:10SPEAKER_01: Most of the global workforce still works in physical and manual occupations
- 28:13SPEAKER_01: and we see AI as being useful there and we're really, really excited to dig in.
- 28:18SPEAKER_01: We're also just really excited to For lack of a better word, just be useful to people.
- 28:25SPEAKER_01: We want people to use this data.
- 28:26SPEAKER_01: We want people to incorporate it into their research, into their work.
- 28:30SPEAKER_01: We want to help.
- 28:33SPEAKER_01: And so we're really excited to externalize the data we have.
- 28:37SPEAKER_01: What we have is available online right now,
- 28:39SPEAKER_01: and we're going to keep it fresh and keep making efforts to add more.
- 28:44SPEAKER_01: And partner with other researchers,
- 28:46SPEAKER_01: think tanks,
- 28:47SPEAKER_01: government agencies to see how this data can help to shed some light
- 28:52SPEAKER_01: on what's happening with the economy.
- 28:53SPEAKER_01: So those are some of the things we're working on.
- 28:56SPEAKER_00: Absolutely great stuff, Scott.
- 28:57SPEAKER_00: Hey, thanks for coming on the podcast.
- 28:59SPEAKER_01: Yeah, thanks so much, Jim.
- 28:59SPEAKER_01: I appreciate the time.