Latest / Faster, Please! — The Podcast / ✨💪 The future of work in an age of AI: My chat with economist Daniel Rock
Transcript
- 0:06SPEAKER_01: Welcome to Faster Please, the podcast.
- 0:09SPEAKER_01: I'm your host, Jim Pethikoukas.
- 0:12SPEAKER_01: Several times a month,
- 0:13I'll feature a lively conversation with a fascinating and
- 0:16provocative guest about
- 0:18SPEAKER_01: how to make the world a better place through scientific discovery,
- 0:21SPEAKER_01: technological innovation,
- 0:23SPEAKER_01: and economic growth.
- 0:24SPEAKER_01: You're also going to want to check out my Faster Please newsletter.
- 0:27SPEAKER_01: You're on Substack throughout the week for fresh essays,
- 0:30SPEAKER_01: Q&As,
- 0:31SPEAKER_01: and stories from around the internet and around the world.
- 0:35SPEAKER_01: Artificial intelligence is undoubtedly poised to change the workplace.
- 0:40SPEAKER_01: The question is, by how much?
- 0:43Key voices in Silicon Valley warn that white-collar jobs will
- 0:47soon be a thing of
- 0:47SPEAKER_01: the past,
- 0:48while others predict more modest economic gains as firms
- 0:52struggle to reorganize
- 0:53SPEAKER_01: workflows.
- 0:55SPEAKER_01: When it comes to the future of work and American business,
- 0:58SPEAKER_01: contradictory forecasts can be difficult to interpret and reconcile.
- 1:03SPEAKER_01: Today on Faster Please, the podcast is
- 1:05Daniel Rock and I attempt to sift through the often confusing
- 1:10current AI conversation.
- 1:12SPEAKER_01: We cover the distinction between AI-exposed fields and those destined for
- 1:16SPEAKER_01: automation,
- 1:17SPEAKER_01: explore the bottlenecks that could slow adoption among businesses,
- 1:21SPEAKER_01: and offer a more realistic outlook for growth.
- 1:26SPEAKER_01: Rock is an assistant professor of operations,
- 1:28SPEAKER_01: information,
- 1:28SPEAKER_01: and decisions at the University of Pennsylvania's Wharton School.
- 1:33SPEAKER_01: There,
- 1:34SPEAKER_01: his research dives into the economics of AI and digital technologies,
- 1:38SPEAKER_01: as well as the future of work.
- 1:40SPEAKER_01: Just the other day,
- 1:42I was reading the Wall Street Journal and there was a interview
- 1:46with the new CEO of
- 1:48SPEAKER_01: Verizon.
- 1:50SPEAKER_01: And in that interview,
- 1:51he predicted 20 to 30 percent unemployment within the next two
- 1:57to five years.
- 1:59SPEAKER_01: due to both AI causing a white-collar bloodbath,
- 2:02SPEAKER_01: but also humanoid robots upending manual labor jobs.
- 2:08SPEAKER_01: Now, I've heard a couple of these kinds of forecasts from AI CEOs.
- 2:12SPEAKER_01: This is maybe the first one from a tech CEO.
- 2:14SPEAKER_01: It's not necessarily an AI CEO.
- 2:17SPEAKER_01: But there are other ones.
- 2:19I believe it was the head of Microsoft's AI unit said within 12
- 2:24to 18 months,
- 2:26SPEAKER_01: Every sort of behind the computer screen job will be automated.
- 2:31SPEAKER_01: What do you make of those kinds of forecasts?
- 2:36SPEAKER_00: I find those kinds of forecasts highly risky.
- 2:39SPEAKER_00: I'm not going to say I don't think it's possible.
- 2:42SPEAKER_00: Something could happen that would make these things possible.
- 2:45SPEAKER_00: But I think our economy is ingesting and responding to
- 2:51SPEAKER_00: AI at a fairly gradual clip right now, even what we have right now.
- 2:56So setting aside that I think there are a number of advances
- 3:00and important research
- 3:02results that will be necessary to make humanoid robots happen
- 3:06to that extent,
- 3:07SPEAKER_00: I think even if they existed,
- 3:08SPEAKER_00: if you had this enormous explosion in capabilities,
- 3:12SPEAKER_00: it would still take a while for organizations to integrate them properly.
- 3:15SPEAKER_00: So
- 3:16SPEAKER_00: Do I think it's possible longer run?
- 3:18SPEAKER_00: Yeah, potentially.
- 3:21But within 18 months or two to five years,
- 3:23that seems a bit aggressive to me.
- 3:26SPEAKER_00: Though,
- 3:26SPEAKER_00: you know,
- 3:26SPEAKER_00: I think in this world,
- 3:27SPEAKER_00: you kind of have to have wide confidence bands as things change,
- 3:30SPEAKER_00: right?
- 3:31SPEAKER_01: When I hear these kinds of forecasts from people in the tech sector,
- 3:35I sometimes wonder if they mean that this is what the
- 3:38technology could do in the
- 3:40SPEAKER_01: lab or some artificial scenario.
- 3:44SPEAKER_01: The CEO of Verizon seems to be talking about the real world, though.
- 3:48SPEAKER_01: And even if you're a super optimist,
- 3:49SPEAKER_01: it seems that given real world constraints,
- 3:52SPEAKER_01: be they legal or organizational or what have you,
- 3:56even the most powerful technologies getting to those
- 4:00unemployment numbers anytime
- 4:02SPEAKER_01: soon seems science fictional.
- 4:04SPEAKER_00: Yeah, in a sense, I agree.
- 4:06And then I think to get to the point where you have
- 4:08unemployment and people aren't
- 4:09getting redeployed while you have these remarkable technologies
- 4:13shifting and
- 4:14SPEAKER_00: getting deployed at the same time,
- 4:16that strikes me as a little bit hard to imagine that scenario
- 4:20as being realistic.
- 4:22The key word you said there,
- 4:23the organizational component,
- 4:25that's the really big part.
- 4:26SPEAKER_00: You know, how are firms doing trainings?
- 4:28How are they thinking about incorporating these tools into
- 4:31their workflows?
- 4:32I talked with a number of executives who are really taking
- 4:34these kinds of questions
- 4:35SPEAKER_00: super seriously.
- 4:37And it's going to require a ton of buy-in from different types
- 4:40of people.
- 4:40SPEAKER_00: Not everybody's stoked to use the technologies.
- 4:43SPEAKER_00: The folks who are using them aggressively, they run into guardrails.
- 4:47They run into places where organizational process isn't quite
- 4:51set up for the speed
- 4:52SPEAKER_00: at which they're building.
- 4:54SPEAKER_00: So there's still a lot to get figured out.
- 4:56If you want to do productive activity,
- 4:58it's a little bit of that stone soup story.
- 5:00SPEAKER_00: And you have to configure a lot of different components all at once.
- 5:04You know,
- 5:04I'm not the only one who's been thinking and writing about a
- 5:06bunch of this.
- 5:07There's a long sort of history of the economics of technology
- 5:12where researchers
- 5:12SPEAKER_00: have been considering these ideas for a long time.
- 5:14Part of my job as a sort of junior researcher in the field
- 5:18right now is going back
- 5:20SPEAKER_00: to stuff that was written in the late 80s and 90s and being like,
- 5:23SPEAKER_00: hey,
- 5:23SPEAKER_00: these people are still right.
- 5:25So I'm digging up those papers and sort of reformatting them in
- 5:29a sense for the AI age.
- 5:32SPEAKER_01: What I see,
- 5:32SPEAKER_01: and you're on social media as well,
- 5:36SPEAKER_01: and we follow each other,
- 5:37SPEAKER_01: but I get a sense there are some,
- 5:40SPEAKER_01: I think this is certainly the case among technology people,
- 5:43but even there are some economists who think that exactly what
- 5:47you're doing,
- 5:48SPEAKER_01: which is looking at the history of technology,
- 5:51SPEAKER_01: that that has created a blind spot to this powerful new technology.
- 5:56SPEAKER_01: And it's really kind of a, this time is really different argument.
- 6:01SPEAKER_01: Are you concerned that you have a blind spot?
- 6:03SPEAKER_00: Yeah, yeah, definitely.
- 6:05SPEAKER_00: So I think the Silicon Valley adage here,
- 6:07SPEAKER_00: right,
- 6:07SPEAKER_00: is strong opinions,
- 6:08SPEAKER_00: weakly held,
- 6:09SPEAKER_00: where,
- 6:10SPEAKER_00: you know,
- 6:10SPEAKER_00: you really let evidence and what's happening guide your thinking.
- 6:15SPEAKER_00: My friend, Kevin Brown,
- 6:16SPEAKER_00: He's a professor at the University of Toronto.
- 6:18SPEAKER_00: He pointed out the other day on social media,
- 6:19SPEAKER_00: right,
- 6:20that the burden of proof for people to say this time is
- 6:22different really falls on
- 6:24SPEAKER_00: them.
- 6:24SPEAKER_00: You can look at the new technology and say, yes, this is different.
- 6:28SPEAKER_00: As Eric Bernelson would say, it's the most G of all GPTs.
- 6:32SPEAKER_00: It's the most general of all general purpose technologies.
- 6:35And you could very much make a case that if not all knowledge
- 6:38workers are going to
- 6:38SPEAKER_00: have to change what they're doing,
- 6:41SPEAKER_00: then this would actually change.
- 6:45SPEAKER_00: create quite a profound impact on the economy.
- 6:48SPEAKER_00: At the same time,
- 6:48SPEAKER_00: I have,
- 6:49SPEAKER_00: maybe it's a lack of imagination on my part,
- 6:51but I have a tough time seeing a place where all sorts of
- 6:55knowledge work changes
- 6:56all at once and we don't take a while to figure out the best
- 6:59new configurations.
- 7:01SPEAKER_00: Or as we're figuring them out,
- 7:03there isn't a bunch of churn where people get hired to do new
- 7:05types of work.
- 7:07SPEAKER_00: Some of it will look like it's automation.
- 7:09SPEAKER_00: Some of it will look like it's transformation.
- 7:12SPEAKER_00: It's all going to get figured out.
- 7:14I think that works at human time scale,
- 7:16not necessarily machine time scale.
- 7:18So you really have to kind of buy that the machines are going
- 7:20to be responsible for
- 7:22SPEAKER_00: much,
- 7:22SPEAKER_00: much more.
- 7:23SPEAKER_00: AI systems can do much,
- 7:24SPEAKER_00: much more in verifiable,
- 7:25high quality ways to think that people aren't going to be part
- 7:28of the loop.
- 7:30SPEAKER_00: in some way, shape, or form.
- 7:31SPEAKER_00: I think we'll find those bottlenecks.
- 7:33In a sense,
- 7:34I would love it if it were the case that machines could do way
- 7:37more.
- 7:38Society's got a redistribution problem there,
- 7:40and we're very wealthy at that point.
- 7:43SPEAKER_00: Yeah, there's power dynamics and other things.
- 7:46SPEAKER_00: You could imagine not quite sci-fi,
- 7:50but deep change scenarios where we have really important social
- 7:53problems to sort
- 7:54SPEAKER_00: out,
- 7:55SPEAKER_00: but all sorts of amazing things are here.
- 7:57SPEAKER_00: And I think
- 7:58SPEAKER_00: More likely,
- 7:59SPEAKER_00: you know,
- 7:59my modal expectation over the next few years is this does look
- 8:02like the kind of
- 8:03SPEAKER_00: normal technology historical precedent,
- 8:06SPEAKER_00: even if there are winners and losers from that kind of scenario.
- 8:10SPEAKER_01: Having spent,
- 8:11SPEAKER_01: you know,
- 8:11SPEAKER_01: many years and most of my career in journalism,
- 8:14SPEAKER_01: I understand like there's a certain bias toward novelty.
- 8:18SPEAKER_01: You're looking for what's new.
- 8:19SPEAKER_01: You're looking for sort of the new take.
- 8:22SPEAKER_01: I see this in economics coverage all the time.
- 8:24Someone has a new theory and it changes everything and rarely
- 8:27does it actually
- 8:28SPEAKER_01: change everything.
- 8:29And I was kind of thinking that when you were recently quoted
- 8:32in a New York Times
- 8:33SPEAKER_01: piece,
- 8:34SPEAKER_01: and the headline of that piece was,
- 8:36SPEAKER_01: economists once dismissed the AI job threat,
- 8:40SPEAKER_01: but not anymore.
- 8:42SPEAKER_01: I don't think that's right.
- 8:43SPEAKER_01: I don't think economists have ever dismissed the AI job threat.
- 8:48I think they dismissed were sort of these very huge,
- 8:53very quick kinds of scenarios.
- 8:56SPEAKER_01: And I'm not sure...
- 8:58SPEAKER_01: generally among economists, that's changed.
- 9:01SPEAKER_00: Yeah,
- 9:01SPEAKER_00: I don't think it's changed a ton,
- 9:03SPEAKER_00: though I'll say,
- 9:04SPEAKER_00: given the narrative and some of the media,
- 9:07SPEAKER_00: maybe that's changed a little bit.
- 9:09Ben Castleman did a great job researching that piece,
- 9:13really talking to a bunch of people.
- 9:14SPEAKER_01: And he doesn't write the headlines.
- 9:16SPEAKER_00: Yeah, that's true too.
- 9:18What I think he was finding was a little bit less certainty
- 9:21amongst economists,
- 9:23SPEAKER_00: especially the group that he spoke to.
- 9:25SPEAKER_00: I think we have a sense that
- 9:28we should pay attention to the technological capabilities and
- 9:31be expansive in terms
- 9:32SPEAKER_00: of how we think about what they can do.
- 9:34SPEAKER_00: Listen to the AI researchers who say like,
- 9:37SPEAKER_00: okay, maybe this can automate a bunch of work.
- 9:39SPEAKER_00: I still think at the task level for a bunch of it,
- 9:42SPEAKER_00: like what are the processes,
- 9:43SPEAKER_00: what are the tasks people are doing?
- 9:44What are the new systems that you build up and how do you
- 9:46reconfigure them?
- 9:47SPEAKER_00: That's hard managerial work.
- 9:49The fact that I see a lot of large company executives dumped by
- 9:53this at like a big
- 9:55SPEAKER_00: picture level makes me think that
- 9:57yeah indeed it will take a little while to sort out but that
- 10:01said the new systems
- 10:02SPEAKER_00: might be way more efficient and for a short run period you could have
- 10:06reconfiguration of work that is potentially job destroying if
- 10:09you have inelastic
- 10:10demand so it's all a question of like what is the speed you
- 10:13have to kind of mix a
- 10:13few things in your head it's like what's the speed of
- 10:15capabilities advance uh
- 10:17SPEAKER_00: what's the speed of like organizational structure uh reconfiguration
- 10:22SPEAKER_00: And then I'm like, how big is the pie, right?
- 10:25SPEAKER_00: Like, I am paying a little bit more attention to scientific teams.
- 10:29Now,
- 10:30a couple of reasons is when you use AI and science,
- 10:32the returns are huge.
- 10:34SPEAKER_00: And that's very, very exciting.
- 10:37But another is scientists want to tell you how they did it and
- 10:40how they're doing things.
- 10:41SPEAKER_00: So
- 10:42There's a little bit of what economists would call incentive
- 10:45compatibility there,
- 10:46SPEAKER_00: where they want to say,
- 10:47SPEAKER_00: OK,
- 10:47SPEAKER_00: yeah,
- 10:47SPEAKER_00: here's how I transform my workflow using AI.
- 10:50And if we can get some insight around how they're redesigning
- 10:53their workflows,
- 10:54maybe that can be extended to larger companies where you've got
- 10:58more invasive
- 11:00SPEAKER_00: control systems or set structures.
- 11:05One of the things I think people find as they start trying to
- 11:08do digital transformation
- 11:11SPEAKER_00: is they look at,
- 11:13they do kind of an inventory of their existing workflows and
- 11:15how they do stuff and
- 11:17SPEAKER_00: nothing to do with AI.
- 11:18They look at some of this and they say,
- 11:19well,
- 11:20that doesn't make any sense.
- 11:22SPEAKER_00: Maybe we should do this differently.
- 11:24The worst thing you ever want to hear is we do this because
- 11:26that's how it's always
- 11:27SPEAKER_00: been done,
- 11:27SPEAKER_00: not because it makes a ton of sense.
- 11:29And then they go and change that and they get some productivity
- 11:32gains just by
- 11:35SPEAKER_00: stripping out some of the stuff that they don't need to do anymore.
- 11:38SPEAKER_00: So if you are finding that that's what's happening in your work,
- 11:43it may be a good sign because then you're on your way to the
- 11:47new way with new
- 11:47SPEAKER_00: capabilities.
- 11:49Anybody who has written or been a key researcher on the notion
- 11:54of a productivity J
- 11:55SPEAKER_01: curve,
- 11:56SPEAKER_01: there's no way I'm not going to ask them,
- 11:59SPEAKER_01: where are we on the productivity J curve?
- 12:03SPEAKER_01: Where are we?
- 12:04SPEAKER_01: And how do you figure out where we are?
- 12:07SPEAKER_00: My best guess is that we're very, very early stage on this.
- 12:11SPEAKER_01: Let me first explain what that is.
- 12:12Explain what the J curve is probably quickly and more concisely
- 12:15than I could.
- 12:16So J-curve is kind of a simple finance dynamic that we've
- 12:20attached to productivity measurement.
- 12:23SPEAKER_00: Total factor productivity is in effect the magic beans of the economy.
- 12:27SPEAKER_00: It's the growth that we can't explain by growth in measurable inputs.
- 12:32It's also referred to as a solo residual and commonly said it's
- 12:37a measure of our
- 12:37SPEAKER_00: ignorance more than anything else.
- 12:39SPEAKER_00: When I teach my students how you compute this, they're like, wait, what?
- 12:42SPEAKER_00: That's what productivity means?
- 12:43SPEAKER_01: I mean,
- 12:44SPEAKER_00: labor productivity is a lot more intuitive,
- 12:46SPEAKER_00: like GDP per hour worked or per worker or something that's,
- 12:49SPEAKER_00: you know,
- 12:49SPEAKER_00: it doesn't correct for capital usage,
- 12:51SPEAKER_00: but at least that's a little bit more directly interpretable.
- 12:54SPEAKER_00: But what we did,
- 12:56SPEAKER_00: my co-authors,
- 12:57SPEAKER_00: Eric Brynelson,
- 12:58SPEAKER_00: Chad Severson and I,
- 13:00SPEAKER_00: we bolted on kind of an investment dynamic where in the beginning,
- 13:03of a sufficiently transformative technology where you have to
- 13:06do this kind of work
- 13:06SPEAKER_00: of creating intangible capital,
- 13:08SPEAKER_00: new ways of doing things,
- 13:10SPEAKER_00: training folks,
- 13:10SPEAKER_00: installing software,
- 13:11SPEAKER_00: configuring software,
- 13:12SPEAKER_00: getting the security protocols in place,
- 13:14SPEAKER_00: all that kind of fun stuff.
- 13:17SPEAKER_00: In the beginning,
- 13:18SPEAKER_00: it looks like you're putting a lot of real stuff in,
- 13:20SPEAKER_00: real money,
- 13:21SPEAKER_00: real investment,
- 13:22and you're not getting as much out because you're spending to
- 13:25create these
- 13:26SPEAKER_00: intangible assets like the ones I mentioned are culture and so on.
- 13:30SPEAKER_00: So, you know, effectively, we get less for more.
- 13:35SPEAKER_00: And that's a problem, right?
- 13:36SPEAKER_00: Everyone worries about a drag on productivity growth.
- 13:39SPEAKER_00: It will potentially look like a drag on productivity growth.
- 13:43SPEAKER_00: In practice,
- 13:44SPEAKER_00: when you go and measure productivity growth,
- 13:46SPEAKER_00: you might see one and a half,
- 13:47SPEAKER_00: 2% PFP growth.
- 13:50SPEAKER_00: And that could be like the counterfactual where we're
- 13:53where we're trying to build all these intangible assets and
- 13:56were we actually using
- 13:57SPEAKER_00: all of our money to just generate straight measurable stuff,
- 14:01SPEAKER_00: we'd actually be up closer to two and a half,
- 14:02SPEAKER_00: 3%.
- 14:04So in the beginning,
- 14:06when we got this dynamic,
- 14:07we create the intangible asset.
- 14:08But then now that we've got this capital asset that's hidden on
- 14:12the balance sheet
- 14:12SPEAKER_00: and it's not in your income statement,
- 14:14eventually you start using that in production and that
- 14:17investment starts to pay
- 14:18SPEAKER_00: off,
- 14:18SPEAKER_00: but you're not measuring the input.
- 14:20SPEAKER_00: So everyone says, hey, we're getting free money here.
- 14:23SPEAKER_00: We're getting way more from less as an input and productivity is booming.
- 14:28SPEAKER_00: So this dynamic where in the initial stages you dip in,
- 14:31you make like a huge investment and that investment return or
- 14:35the asset you build
- 14:36SPEAKER_00: is not characterized anywhere in measurable accounts.
- 14:40But then later it starts to sort of trickle out free money,
- 14:44free money in air quotes here.
- 14:47SPEAKER_00: That's the J curve.
- 14:48SPEAKER_00: So where are we?
- 14:49SPEAKER_00: Is that was the question you asked, right?
- 14:51SPEAKER_01: I ask you just one quick one question about that.
- 14:53SPEAKER_01: We look at these like,
- 14:54SPEAKER_01: you know,
- 14:55SPEAKER_01: big aggregate productivity numbers when the government releases them.
- 14:59SPEAKER_01: And, you know, like, is this it?
- 15:00Are we seeing are we as you know,
- 15:01if there's any kind of blip up or like,
- 15:03is that AI?
- 15:04Would we actually expect to see productivity increase from like
- 15:10wherever it is or
- 15:11SPEAKER_01: would actually almost expect to see a dip before it goes up?
- 15:15I would be relatively surprised to see a big dip in measured
- 15:20productivity,
- 15:21SPEAKER_00: more like anemic productivity growth in the presence of like radically
- 15:27SPEAKER_00: transformative stuff is kind of what my,
- 15:29SPEAKER_00: my base case would be.
- 15:31SPEAKER_00: I think also these,
- 15:32SPEAKER_00: these series,
- 15:33SPEAKER_00: because there's such a measure of our ignorance,
- 15:35SPEAKER_00: they're super noisy.
- 15:37It's going to be really hard to tell from like the quarterly
- 15:39productivity numbers.
- 15:40SPEAKER_00: Like, are we in a boom or, or anything right now?
- 15:44SPEAKER_01: I believe the good ones.
- 15:46SPEAKER_01: I disbelieve the bad ones.
- 15:47SPEAKER_01: Yeah, right?
- 15:48SPEAKER_00: Let's just be optimistic about it.
- 15:49SPEAKER_00: We're going to need a do-over on Q3, right?
- 15:55SPEAKER_00: But yeah,
- 15:56SPEAKER_00: I think like,
- 15:57SPEAKER_00: okay,
- 15:58we might be seeing some of the gains from deep learning or some
- 16:00of the technologies
- 16:01SPEAKER_00: that were being developed five,
- 16:03SPEAKER_00: six years ago,
- 16:04SPEAKER_00: the initial stages of this,
- 16:06some retrofitting that's happening where we take existing
- 16:08processes and drop AI in.
- 16:12SPEAKER_00: Now, do I think we're
- 16:14We're going to see a boom in productivity potentially,
- 16:16but I haven't seen it yet in the data.
- 16:18You kind of like scrutinize the little squiggles in the graph
- 16:23and you're like,
- 16:23SPEAKER_00: I really hope that's it.
- 16:24SPEAKER_00: So we are not ascending.
- 16:27SPEAKER_00: We are not in the ascent of the J curve.
- 16:28SPEAKER_00: Not yet on the J curve.
- 16:30SPEAKER_00: I wouldn't say so.
- 16:30I think we're very much in the early stages where a lot of the
- 16:34investment honestly
- 16:34SPEAKER_00: is measurable.
- 16:35SPEAKER_00: It's in like chips and data centers and things like that.
- 16:39SPEAKER_00: I don't think companies have really done the reconfiguration work yet.
- 16:46SPEAKER_00: If I had to guess,
- 16:47SPEAKER_00: it would probably be in low billions of dollars spent in the economy,
- 16:52SPEAKER_00: which is potentially enough to move the needle,
- 16:55SPEAKER_00: importantly,
- 16:55SPEAKER_00: in some areas,
- 16:56SPEAKER_00: but not enough to show up in aggregate statistics.
- 16:59SPEAKER_00: And I think it'll start moving a lot more in the coming years.
- 17:02SPEAKER_00: People are not doubting that AI is valuable.
- 17:04SPEAKER_00: They're just trying to figure out how to use it.
- 17:06SPEAKER_01: Yeah, I think I think generally we're sort of past past that point.
- 17:10SPEAKER_01: And if you're I mean,
- 17:11if you're looking for sort of those like the green shoots and
- 17:14you don't want to
- 17:15SPEAKER_01: wait till there's a final,
- 17:17SPEAKER_01: you know,
- 17:17SPEAKER_01: a final government statistic,
- 17:20SPEAKER_01: you know,
- 17:20SPEAKER_01: I mean,
- 17:21SPEAKER_01: where do you where do you look?
- 17:22SPEAKER_01: Do you sort of look at, you know, company earnings calls and like what?
- 17:28SPEAKER_01: How are they talking about it?
- 17:29SPEAKER_01: It seems like it'd be real detective work.
- 17:33SPEAKER_00: It is.
- 17:34SPEAKER_00: I cheat a little bit.
- 17:37SPEAKER_00: So Eric and I,
- 17:37SPEAKER_00: and then Andy McAfee and James Milan,
- 17:40SPEAKER_00: the four of us co-founded a company called Workelix,
- 17:43where companies will give us the data of what's going on inside
- 17:47of their...
- 17:48SPEAKER_00: their AI systems.
- 17:50SPEAKER_00: And yeah,
- 17:51SPEAKER_00: you can see,
- 17:52I think one interesting thing we see is like a power law every
- 17:55single time.
- 17:55SPEAKER_00: It doesn't matter the tool, doesn't matter the company.
- 17:58SPEAKER_00: The numbers change a little bit,
- 17:59SPEAKER_00: but it might be like 10% of your user,
- 18:01SPEAKER_00: they're half the value that's getting generated.
- 18:03SPEAKER_00: And when you see that,
- 18:04SPEAKER_00: you think,
- 18:04SPEAKER_00: okay,
- 18:04SPEAKER_00: well,
- 18:04SPEAKER_00: there's some people who are really running ahead and doing a great job,
- 18:07but that's a ton of low-hanging fruit to train people to do all
- 18:09sorts of new stuff.
- 18:11And as long as that's persistent and you haven't quite worked
- 18:13out how to spread
- 18:14SPEAKER_00: these best practices or constrain them in some way,
- 18:19SPEAKER_00: because there's a lot of people who don't want to use it,
- 18:20but they don't want to use it because they really want to do a
- 18:23good job at work.
- 18:24SPEAKER_00: They don't want to fail.
- 18:25SPEAKER_00: They don't want to create headaches for other people.
- 18:27SPEAKER_00: Those people are gold.
- 18:28You can have them develop all sorts of constraints and things
- 18:32that make these
- 18:32SPEAKER_00: systems work better.
- 18:34So,
- 18:35you know,
- 18:35it's really about kind of building a consensus inside of
- 18:38organizations.
- 18:38SPEAKER_00: We haven't seen a ton of that.
- 18:41SPEAKER_00: And they're going to get there.
- 18:43SPEAKER_00: Like,
- 18:43SPEAKER_00: there are a lot of top-down efforts that I know of where it's like,
- 18:46SPEAKER_00: how can we celebrate the wins and,
- 18:48SPEAKER_00: you know,
- 18:48SPEAKER_00: also make people responsible for doing a good job?
- 18:53SPEAKER_01: At least so far,
- 18:54the general attitude among companies seems to be that when a
- 18:59pilot project doesn't
- 19:00SPEAKER_01: produce a good result,
- 19:03SPEAKER_01: the response isn't to blame the technology.
- 19:05SPEAKER_01: It's to sort of blame the implementation.
- 19:07SPEAKER_01: They're not giving up on the technology.
- 19:09SPEAKER_01: They're just redoubling their efforts to figure out how to make it work.
- 19:12SPEAKER_01: Definitely.
- 19:13SPEAKER_00: And that's good.
- 19:14SPEAKER_00: That's healthy.
- 19:15SPEAKER_00: You want to learn what works and what doesn't.
- 19:17SPEAKER_00: And if 90% of the stuff doesn't work out and 10% is a smashing success...
- 19:22SPEAKER_00: Great.
- 19:22SPEAKER_00: Like, just make sure you're keeping track of it.
- 19:24Make sure you know how to double down on the things that work
- 19:27really well.
- 19:29What would you tell someone who either reads quotes about 30
- 19:33percent unemployment
- 19:35or they look at their job and there's some index that say their
- 19:38job is deeply
- 19:40SPEAKER_01: exposed to A.I.?
- 19:42SPEAKER_01: What should they think?
- 19:43SPEAKER_01: Should they be like, I'm going to be out of a job soon.
- 19:45SPEAKER_01: I better go learn how to do something else.
- 19:49SPEAKER_01: You know, I better make no large purchases.
- 19:50SPEAKER_01: You know, what should they what should they think?
- 19:53SPEAKER_00: Well,
- 19:53SPEAKER_00: I would have a different reply to the person who is reading about,
- 19:57SPEAKER_00: you know,
- 19:57large unemployment versus the person who reads that their job
- 20:00is exposed to AI.
- 20:01SPEAKER_00: The 30% unemployment are like, hey, the jobs apocalypse is coming.
- 20:05SPEAKER_00: I would just digging into,
- 20:08SPEAKER_00: you know,
- 20:09SPEAKER_00: some of the research on what would be required for that to happen.
- 20:12SPEAKER_00: You know,
- 20:13SPEAKER_00: Alex Emas,
- 20:14SPEAKER_00: a friend of mine,
- 20:16SPEAKER_00: and we're working a project together,
- 20:18SPEAKER_00: which is really cool.
- 20:19SPEAKER_00: He has a great blog called Ghosts of Electricity.
- 20:22SPEAKER_00: He's exploring a lot of these different scenarios.
- 20:25There are other papers that kind of show you what you need to
- 20:29get to outcomes like that.
- 20:30There's a paper,
- 20:32Benzel,
- 20:33Lagarda,
- 20:34Kotlikoff and Sachs,
- 20:35Robots Are Us,
- 20:36which thinks about a
- 20:38SPEAKER_00: you know, an overlapping generations model.
- 20:40SPEAKER_00: We don't have to get into all the technical details here,
- 20:42but basically a robot is a perfect replica for a person and all
- 20:45their work they can
- 20:45SPEAKER_00: do.
- 20:47SPEAKER_00: And then you can think like,
- 20:48SPEAKER_00: all right,
- 20:48SPEAKER_00: the technology I'm seeing that I'm using on a day-to-day basis,
- 20:51does that kind of line up with what I think someone could do in
- 20:57their job?
- 20:57SPEAKER_00: Like, is it something that could be fully automated?
- 21:01SPEAKER_00: There's a lot of interstitial stuff that people do to make things work.
- 21:04SPEAKER_00: If you have folks,
- 21:06SPEAKER_00: I think the Marxists call it radical compliance as a protest strategy,
- 21:09SPEAKER_00: you do precisely what your job entails and nothing more.
- 21:13SPEAKER_00: And things break when that happens.
- 21:16SPEAKER_00: So people are kind of smoothers and they figure out how...
- 21:20SPEAKER_00: how to work with each other and to work within systems.
- 21:23SPEAKER_00: And that stuff shouldn't be underrated.
- 21:25SPEAKER_00: Now,
- 21:25SPEAKER_00: if you see,
- 21:26SPEAKER_00: since I've done some of this research on exposure,
- 21:28SPEAKER_00: if you see that your job is highly exposed to AI,
- 21:31SPEAKER_00: exposure does not mean automation.
- 21:33SPEAKER_00: Exposure means change.
- 21:35And it might be the case that high levels of exposure are
- 21:37actually great for you.
- 21:39SPEAKER_00: It could be that it's not good.
- 21:41I would say be concerned if you are a customer service agent
- 21:45and it's remote,
- 21:47SPEAKER_00: that kind of stuff.
- 21:48SPEAKER_00: tends to be highly visible to machines.
- 21:50SPEAKER_00: There are clear output measures that are good and bad.
- 21:54SPEAKER_00: I've already seen companies starting to automate aspects of that.
- 21:57On the other hand,
- 21:59you know,
- 21:59the famous proclamation that radiologists were going away.
- 22:03SPEAKER_00: They have some really interesting mixtures of tasks.
- 22:06SPEAKER_00: That is a highly expert group of people.
- 22:09SPEAKER_00: And if you think of them as data enabled doctors,
- 22:12the idea that they would go away starts to seem a little bit
- 22:14more and more foreign.
- 22:17SPEAKER_00: Luis Carricano and his co-authors,
- 22:20SPEAKER_00: they just had a paper come out on weak versus strong task bundling.
- 22:24I think that's kind of a good model for thinking about what
- 22:28your job might do.
- 22:31SPEAKER_00: If it's a strong bundle of tasks,
- 22:32it's a bunch of things that are really kind of inextricable
- 22:35from each other,
- 22:36SPEAKER_00: and using AI in one doesn't mean that you don't still need to be there,
- 22:42SPEAKER_00: then you're in good shape.
- 22:43SPEAKER_00: But if it's stuff where a bunch can be separated out,
- 22:45SPEAKER_00: now you're thinking about what's left for you.
- 22:48If it's something that you're going to spend a ton of time on
- 22:50yourself,
- 22:50then great.
- 22:52SPEAKER_00: And it's probably complimentary to what the machines are doing.
- 22:54SPEAKER_00: If it's something where machines can kind of handle the whole bundle,
- 22:57SPEAKER_00: then we're looking at a different story.
- 22:59OpenAI just had a report come out,
- 23:00too,
- 23:00where they're starting to handle some of this.
- 23:02SPEAKER_00: Shout out, Alex and Ronnie.
- 23:04SPEAKER_00: Great work there.
- 23:05There was an interesting survey put out by Forecasting Research
- 23:08Institute,
- 23:09SPEAKER_01: and they interviewed technologists,
- 23:11SPEAKER_01: interviewed economists.
- 23:12SPEAKER_01: And what was interesting is...
- 23:15SPEAKER_01: that for all the talk that these groups see things differently,
- 23:21SPEAKER_01: their general median forecasts really weren't that differently,
- 23:24SPEAKER_01: their forecasts about growth.
- 23:26Since you view things academically and you're seeing what
- 23:28companies are actually
- 23:29SPEAKER_01: doing,
- 23:30SPEAKER_01: what do you think is the realistic upside of AI boosting economic growth,
- 23:35SPEAKER_01: boosting productivity growth?
- 23:37SPEAKER_00: I think it's going to be fairly large.
- 23:39SPEAKER_00: But as we know, since we're looking at growth, you can grow very quickly.
- 23:43SPEAKER_00: Maybe we're lucky we get like 3% to 5% growth in a short period of time.
- 23:48SPEAKER_00: But then your base is getting better.
- 23:49SPEAKER_00: It's hard to maintain the top level growth rates.
- 23:53SPEAKER_00: So my optimistic case right now for the next few years,
- 23:57I would love to see somewhere in the neighborhood of 2% to 3%
- 24:00productivity growth.
- 24:02SPEAKER_00: I think that'd be
- 24:03SPEAKER_00: that'd be cool.
- 24:04SPEAKER_00: You know, three to four is potentially, you know, reasonable.
- 24:08SPEAKER_00: I don't know.
- 24:08SPEAKER_00: Maybe we do get a miraculous technology and it goes faster.
- 24:12But when I start to see people thinking it's going to be a
- 24:15little bit higher than
- 24:15the median or even like the 90th percentile respondents to
- 24:19those surveys where
- 24:20SPEAKER_00: people say it's like 15,
- 24:21SPEAKER_00: 20%.
- 24:22SPEAKER_00: Right now you're talking about like a sci-fi economy.
- 24:25SPEAKER_00: That's hard for me to conceptualize properly.
- 24:27Like that's really radical,
- 24:29transformative,
- 24:30um,
- 24:30AI driven change in the economy.
- 24:33SPEAKER_00: Um,
- 24:33I'm kind of curious,
- 24:35you know,
- 24:35I know there's that curse,
- 24:37maybe you live in interesting times.
- 24:39SPEAKER_00: That would be a very interesting time to live through.
- 24:42SPEAKER_00: I'm kind of curious to see what it would be like.
- 24:44SPEAKER_01: There's sort of a flip side of that question,
- 24:45SPEAKER_01: not like what's the upside of this technology,
- 24:49SPEAKER_01: but like what is too much of an upside?
- 24:51I often write,
- 24:51act like it's,
- 24:52you know,
- 24:52there's no downsides,
- 24:54but like could...
- 24:55Could an economy grow like too fast and that would be so
- 24:59disruptive to society that
- 25:01SPEAKER_01: you would get,
- 25:01SPEAKER_01: you know,
- 25:02SPEAKER_01: you'd get pushback and it would just be sort of too much.
- 25:05SPEAKER_00: Yeah,
- 25:06I think there's a lot of folks who like the way things are
- 25:08generally and are OK
- 25:10SPEAKER_00: with marginal growth.
- 25:12There are people who are willing to break things to see what
- 25:14would happen.
- 25:16SPEAKER_00: in a better situation.
- 25:17I think you start to get into the realm of politics and
- 25:20sociology and things that I
- 25:22don't understand nearly as well as folks in that area for what
- 25:26that looks like when
- 25:27SPEAKER_00: society changes that quickly.
- 25:30SPEAKER_00: I don't think economists,
- 25:31SPEAKER_00: if you asked about blind spots earlier,
- 25:32I don't think we really have a good sense of what this looks
- 25:34like for a developed
- 25:38SPEAKER_00: wealthy economy like the US to grow that quickly.
- 25:40SPEAKER_00: So
- 25:43SPEAKER_00: I don't know.
- 25:44SPEAKER_00: I'd be fascinated to see what happens.
- 25:47SPEAKER_00: Uh, and I, I hope for the best thing.
- 25:49SPEAKER_00: I hope for an explosion in high quality, um,
- 25:53drugs for things like cancer we i saw that news about the
- 25:57pancreatic cancer trials
- 25:59with mrna vaccines and i was like this is the best we need more
- 26:03of this please uh
- 26:04faster please as someone might say uh thank you i appreciate
- 26:08that so yeah i get
- 26:09excited about those potential things but i think you know then
- 26:12you need people we
- 26:13rely on economic theory a little bit more like where's the
- 26:16bottleneck going to show
- 26:17SPEAKER_00: up what sorts of things are going to slow things down
- 26:20If we have this conversation a year from now,
- 26:22what do you think will be the state of play?
- 26:24SPEAKER_01: What will we be saying?
- 26:25SPEAKER_01: Will we still be saying,
- 26:27SPEAKER_01: well,
- 26:27we're still kind of prowling through the business earnings
- 26:32calls and the microdata,
- 26:34SPEAKER_01: or will things be a little clearer?
- 26:36SPEAKER_00: So you're asking me to make predictions in a very clever way.
- 26:39SPEAKER_01: It is a very, with all cautions and caveats.
- 26:43SPEAKER_00: All right.
- 26:44SPEAKER_00: So if I had to guess, let's see.
- 26:45The cop-out answer that I have fairly high confidence in is I
- 26:49think my confidence
- 26:49SPEAKER_00: bands are going to be wider in a year.
- 26:52SPEAKER_00: So in a Bayesian way,
- 26:53I suppose I should be updating backwards to say my confidence
- 26:56band should be wider
- 26:57SPEAKER_00: now for all the new information that gets released.
- 27:00SPEAKER_00: But if I had to make a concrete prediction, you know, we went from...
- 27:05coding assistance about a year ago to agentic coding being kind
- 27:10of mind-blowingly good.
- 27:12SPEAKER_00: And now people are using it to build all sorts of things.
- 27:15SPEAKER_00: I would say,
- 27:16let's take seriously the idea that anybody can build software
- 27:19for anything and that
- 27:20SPEAKER_00: it'll be better and better and more capability enhancing.
- 27:24SPEAKER_00: In which case,
- 27:26I think a thing people are going to be discussing is how do I
- 27:29manage agentic teams
- 27:31SPEAKER_00: well?
- 27:32SPEAKER_00: And how do I, you know,
- 27:35SPEAKER_00: integrate people into those processes because I need to.
- 27:39And then we might have some conversations of what it's like to
- 27:41be managed by an AI system.
- 27:43SPEAKER_00: So as organizations blow up,
- 27:45SPEAKER_00: I think there's been a lot of discussion of like,
- 27:47SPEAKER_00: what does it look like for a one person unicorn to show up?
- 27:51I think,
- 27:52okay,
- 27:52if that's possible,
- 27:53then the scope and scale of an organization with
- 27:57AI enabled reduction of coordination costs and like increased
- 28:00modularity and design.
- 28:02SPEAKER_00: I think the,
- 28:03SPEAKER_00: you know,
- 28:03SPEAKER_00: so-called boundary of firms can expand quite quickly,
- 28:06SPEAKER_00: at which point you might have an existing large firm that figures it out,
- 28:11SPEAKER_00: become much,
- 28:11SPEAKER_00: much larger as a result.
- 28:13SPEAKER_00: It's just, we can sustain a bigger thing.
- 28:15SPEAKER_00: If that's the case, though, you're going to have machines managing folks.
- 28:18SPEAKER_00: That would be my guess.
- 28:19SPEAKER_01: So that might be an early conversation in a year.