Anton Korinek, a professor at the University of Virginia’s economics department and business school, joined the podcast to discuss AI productivity gains, augmentation versus automation, the future of white-collar work, global access to AI technology, universal basic income, career planning in the age of AI, workforce adaptation, AI-generated podcasts, and more.
Available on YouTube, Apple Podcasts, Spotify, or any other podcast platform.
Our music is by Micah Rubin (Producer) and John Lisi (Composer).
Relevant Links
Economic Policy Challenges for the Age of AI (Anton Korinek)
Generative AI for Economic Research: Use Cases and Implications for Economists (Anton Korinek)
OpenAI confirms new $200 monthly subscription, ChatGPT Pro, which includes its o1 reasoning model (Kyle Wiggers, TechCrunch)
Poverty and Inequality Platform (World Bank)
2.9 billion people still offline (International Telecommunication Union)
CS enrollment and graduation data, College of LSA + CS enrollment and graduation data, College of Engineering (University of Michigan)
o1 System Card (OpenAI)
Number of Jobs, Labor Market Experience, Marital Status, and Health for those Born 1957-1964 (Bureau of Labor Statistics)
Preparing the Workforce for an Uncertain AI Future - Written Statement to the U.S. Senate AI Insight Forum on Workforce (Anton Korinek)
Klarna AI handles two-thirds of customer service chats in its first month (Klarna)
The World’s Call Center Capital Is Gripped by AI Fever — and Fear (Saritha Rai, Bloomberg)
NotebookLM’s automatically generated podcasts are surprisingly effective (Simon Willison)
Timestamps
00:01:43 - AI in today’s workplace
00:12:34 - Learning from history
00:18:07 - Ensuring broad AI access
00:24:52 - Future of higher education
00:31:13 - Evolving workforce needs
00:41:18 - Social support planning
00:46:47 - Current industry impacts
00:57:25 - Changes in podcasting
01:00:52 - Building shared prosperity
Transcript
This transcript was generated safely by AI with human oversight. It may contain errors.
(Cold Open) Anton (00:00:00):
I know some students who tell me, well, I'm not sure if it makes sense for me to spend so many years to get a degree because by the time I'm done, the AI will be able to do everything I've learned and will be able to do it better. And I can't contradict them with conviction.
Jakub (00:00:30):
Welcome to the Center for AI Policy Podcast where we zoom into the strategic landscape of AI and unpack its implications for US policy. I'm your host, Jakub Kraus, and today's guest is Anton Korinek. Anton is a professor at the University of Virginia's Economics Department and Business School. He also has appointments at the Center for the Governance of AI, the Complexity Science Hub, the Center for Economic Policy Research, the National Bureau of Economic Research, and Brookings. We discuss topics like AI productivity gains, augmentation versus automation, the future of white-collar work, global access to AI technology, universal basic income, career planning in the age of AI, workforce adaptation, AI-generated podcasts, and more. I hope you enjoy.
(00:01:38):
Anton. Thank you for joining the podcast.
Anton (00:01:41):
Thank you for having me.
00:01:43 - AI in today’s workplace
Jakub (00:01:43):
So you've written about how AI can help with economic research related tasks and you have a long list of different tasks that it might be able to help with, how reliable it is at them, and in some cases you find that it's very reliable and will save economists time if they can use it for that. And some of these are things like summarizing research papers, editing text. So how useful overall do you find current AI capabilities for conducting economic research? And then looking ahead in 2025, how useful do you think they will be?
Anton (00:02:25):
2025? Well, let me go back in time a little bit. I started analyzing this all in 2022 essentially when I saw the first language models come out that really had public uptick like ChatGPT and so on. And back then I first realized, whoa, this is going to really revolutionize things. And I decided I wanted to spend my time on two major themes. The first one was how should we prepare for the rapid advances in AI that I think we have seen in the past two years, but that I'm also expecting to continue for the next couple of years. And I'm sure we will spend plenty of time during our conversation looking at those. And the second thing was to look at the cutting-edge of what AI can actually do in my own work, because I figured if these tools are really going to be so impactful, I should basically try to incorporate them into my own work as well, and that will tell me something about... First of all, how fast things are advancing. And secondly, it'll also help me do my work more efficiently. And so when I wrote my first paper on this topic in late 2022, early 2023, I basically found that these large language models are quite useful for certain, I call them micro tasks like what you just described, summarizing short sections of text editing things a little bit.
Jakub (00:04:24):
Was that ChatGPT?
Anton (00:04:27):
That was mostly ChatGPT and Claude at the time. And the first version of this paper I'm talking about, it came out before GPT-4 even. So that means we were really talking about what from our current perspective looks like very rudimentary tools,
Jakub (00:04:46):
The dark ages of AI.
Anton (00:04:48):
The dark ages indeed. And over the past one and a half years, we've really seen rapid advances. And in fact, yesterday I just finished the most recent update on this research project because the journal that published it requires me to provide semi-annual updates. And so that forced me to spend the past couple of weeks immersing myself in what have been the most recent advances. And that really gave me kind of a front row seat to how rapidly large language models, generative AI more generally is advancing. It's not only micro tasks that they're useful for, but there's really a growing range of tasks that researchers or white-collar workers more generally can employ these large language models for. And I would say a year ago I said these tools make me about 10, 15% more productive. Now it's probably at least 25%.
Jakub (00:06:04):
Wow. And the tools you're talking about are OpenAI's o1... Any other specific tools that you were testing out?
Anton (00:06:15):
So the amazing thing is right now as of essentially summer fall 2024, we have a pretty crowded field of frontier large language models that all rank pretty closely to each other in terms of their capabilities. So yeah, as you mentioned, OpenAI is definitely still leading, but Google Gemini has improved significantly. Elon Musk's xAI is very close in terms of its capabilities. Anthropic Claude is also right there and that's the one I like particularly for writing. And then you have open source tools like Meta's Llama and that's also really quite good. So depending on what you do, I pick different models for different things because they have slightly different personalities. And let's say for writing, I use Claude for reasoning, I use Open AI o1 for long context processing, meaning when you have a lot of text that you want to process simultaneously I use Gemini and I have my tool for each task, so to say.
Jakub (00:07:41):
And then in 2025 with if you said maybe it was some percent at the start and then 25% this year, what do you think that number might be?
Anton (00:07:55):
Well, if we kind of blindly extrapolate, I guess it should be 35, 40%, but it's so hard to foresee how exactly things are going to evolve. I would say the most significant change I have seen in 2024 is that we have moved from this chatbot-based interaction mode to systems that work in a much more collaborative fashion. So let's say if you use Claude's Artifacts, you have a separate window opening up in which the document that you're working on is being written - OpenAI came out with a similar system, Canvas, where you can interact with the document that the language model is also working on. In some sense it's as if you're a collaborator of the language model. And of course companies like Microsoft are also introducing Copilots into their existing software packages where you can collaboratively interact with an AI system that does tasks for you, that helps you work on what you're working on and ultimately that makes you more productive at whatever it is that you're doing. Hopefully lots of good things.
Jakub (00:09:19):
I've noticed these branding and the actual technical features of the products are trying to be like co-pilots trying to assist humans and make them more productive and augment their work. But at the same time, I've seen a separate division of, say, Google - it could be the same employees, but separate market stream - trying to make AI agents. So at the present, the best complexity of a task that AI could do... it might be ordering you a pizza, ordering you some chocolates, but I could certainly see then that extends into more and more complex tasks. And with the AI agents like this, it's not very much a assistance, it's more the AI acts autonomously and it would be automating the task of ordering a pizza, which might be good. But I'm wondering, is there a way to distinguish if we have a given new tool, to what extent it will be performing tasks in a more automation fashion and taking work away from humans versus augmenting them and making them more productive? So if you said it might go up to 40% of a productivity boost next year, is there any limit to how high this can go before it's more about the AI doing the task than the human or before it has some sort of labor market impact?
Anton (00:10:56):
Yeah, so right now I would say the AI in particular when it comes to language models is still doing parts of tasks, but as you say, they're advancing quite rapidly towards these types of agents that can perform more work more autonomously. We can already envision that a lot of call service center tasks will be able to be performed automatically in a year from now. Probably also a lot of tasks in marketing and eventually perhaps also in research in what I'm doing myself. Fundamentally, I think each one of those tools always has both effects. They replace certain parts and as a result of that, you need humans for fewer of the remaining tasks, and that means the overall work performed is more productive. Now for any given application, there's always a horse race between those two effects. And if the productivity enhancing horse, so to say, is the faster one, then it increases labor demand and ultimately raises wages. But if the displacing horse is the faster one, then it reduces labor demand.
00:12:34 - Learning from history
Jakub (00:12:34):
Are there historical examples of these two phenomena?
Anton (00:12:40):
Plenty. So if we take the long run perspective and look at the industrial revolution, then I would say that the productivity enhancing force was by far the strongest one. And it has led to wage levels in advanced countries going up by something like a factor of 20. But there have been episodes when specific segments of the labor market have really been affected negatively by automation. So for example, take blue-collar workers in the US for the past four decades, maybe with exception of the past few years. We've seen that the overall wage level of nonsupervisory workers has actually gone down in real terms for multiple decades. So while the rest of the economy has more than doubled during that period, a lot of blue-collar workers have actually seen their incomes fall.
Jakub (00:13:47):
Is that because of automation technology? What's the cause?
Anton (00:13:53):
That was essentially because of the wave of automation in blue-collar work, robotics, factories and so on. To some extent also because of globalization, but you can view globalization as a form of technological advancement, right? Because it relied on things like shipping, modern logistics and so on. So you can say these types of technological advances have directly hurt blue-collar workers. Blue-collar workers in the US have been worse off, and I think that may be a harbinger of the effects of what white-collar workers may be seeing over the coming 5 or 10 years.
Jakub (00:14:39):
Yeah, this has been interesting because I think for a while because of popular conceptions of robots, it was thought that these blue-collar tasks might be the first to be automated by AI, but then it's almost flipped where the first tasks we're seeing lots of AI capability are in making art and making music and talking and acting in someone else's voice and writing a short story. And these are some things that people might've thought was going to come much later. So is there any good way to forecast what types of skills will continue to be in demand and even potentially created as new tasks and new jobs? And is there also a way to forecast which ones are particularly prone to the automation? My impression is it's really hard, but maybe I'm unaware of some tools.
Anton (00:15:41):
I agree that it's difficult. Now, there is a lot of very well done economic research papers to try to address this, but from a big picture perspective, I would say the tasks that are almost certainly going to be in high demand in the near term are tasks interacting with AI tools. So if you know how to leverage large language models or generative AI more generally, then you're going to be in high demand. You are the one who makes the connection to these tools that are growing very fast in terms of their productivity. And that means your value is also growing fast. If you don't know how to use these tools and they can do what you do, you're going to be among the losers. And I think there are a bunch of factors, a bunch of high-level factors that allow us to kind of anticipate which areas are going to be hit the hardest in the near term.
(00:16:56):
I think first, if you are doing cognitive work that can be done by a remote worker, then it is quite likely that AI will be able to do it soon too. Secondly, if you work in a sector where there is a lot of competition, let's think for example the financial sector or consulting, then those competitive forces will ensure that the AI is employed as much as possible and that humans who can be replaced will be replaced much more quickly. And on the flip side, right now, robotics is lagging behind AI at least by a little bit. I think robotics is also seeing a moment of acceleration because robots can do much more when you combine them with modern AI-powered brains. But right now in the short term, anything that involves physically interacting with the real world is probably also going to see a relative increase in demand.
00:18:07 - Ensuring broad AI access
Jakub (00:18:07):
It's interesting that you bring up the winners and losers and how people who benefit more from this might be people who are interacting with AI tools and learning how to. This makes me think about potential inequalities and who has access to the AI. So there are two main things I can think of: the cost, and the internet access. There are certainly others, but the best AI models may charge you $20 a month, and OpenAI is looking at potentially even higher prices. They floated a $2,000 figure, which is probably way higher than what they will actually end up doing,
Anton (00:18:49):
At least in the short run, yeah.
Jakub (00:18:53):
Even if they don't ever do or if they don't in the short run, do a much higher price, even if they just stay at $20, this is already somewhat prohibitive. So what I found from the World Bank's poverty and inequality platform was that almost three and a half billion people live on less than $200 per month, roughly speaking. And that would make a $20 monthly subscription about 10% of your month's income and how much money you have, which would be pretty unaffordable.
(00:19:29):
And then the internet access is also pretty striking. So in 2022, the United Nations International Telecommunications Union estimated that 5.3 billion people, or 60% of the global population, are online, which they mean by saying that they've used the internet in the last three months. But then the flip side is that 2.7 billion people are offline and haven't used the internet in the last three months according to this estimate. So they won't be able to access the AI. So it seems to me that there will be billions of people who are using the AI and getting ahead at least in some of the tasks where it really matters to be familiar with AI. And then there will be billions of other people who are lacking this access to it. So I'm wondering what effects you think this might have on global inequality and global income distribution?
Anton (00:20:32):
I think you're right. There is a risk that the already existing digital divide may turn into an intelligence divide, that those who have access to the most cutting-edge intelligence will become a lot more productive and at some level will also be relative winners. So it means that those who don't have access will be left behind.
(00:20:59):
Now, it doesn't have to be that way. And one thing that's encouraging is that, for example, with ChatGPT, even if you don't pay, you have access to the free version, which is in fact almost as good as the paying version. You may have certain access restrictions, you may not be able to use quite as many chats and so on, but it is still extremely valuable. And that's actually something that I think is applaudable, where competition, I would think, has ensured that the leading AI labs have seen the price of access to the models competed down almost to zero because everybody wants to pick up those users. So that's the encouraging part.
(00:21:57):
The less encouraging part is if we look a couple years in the future, I think you're absolutely right that it is plausible that AI models may become a lot more expensive at the frontier. So that means if you want to hire that artificial virtual worker, maybe that is going to cost significantly more and maybe just a fraction of the population will be able to afford that and everybody else will be outcompeted.
Jakub (00:22:34):
Is there anything that can be done about that? So even the idea that we can provide the free access, I find somewhat, I feel a little bit uncomfortable or wary of, because I know that OpenAI I think was spending billions of dollars just on running ChatGPT, they're running the inference to serve it to all the users, and they have about 200 million weekly active users. So serving the 2 billion people would be 10 times as much cost potentially. And also with some factors like OpenAI's o1, the inference costs might go up further. This could be balanced by inference costs falling by making the model smaller or by training them on larger models and trying to distill. But it seems like as you said, there would have to be some sort of cap on how much the free users can use it, whereas the people who are using it as a virtual coworker and might need it many, many times a day are going to be able to use it many, many times more. So besides offering it for free, are there any other solutions to this access?
Anton (00:24:06):
So this is speculative now it's difficult to predict with a lot of confidence where this is going, but I would actually expect that in the next couple years, almost everybody will have access to language models that are significantly better than what we have right now at essentially no cost because they are, for example, built into their cell phones or they can access cheap services on the internet that are not the cutting edge, but still better than what we have today. So that's kind of my optimistic perspective, but I agree with you that we can by no means be assured of that.
00:24:52 - Future of higher education
Jakub (00:24:52):
And now zooming back to the US specifically. So I went to college in the US, I went to University of Michigan and took some computer science classes. And in these, there were hundreds of students sometimes; they were very popular classes. But it wasn't always this way. I learned from reading up on the internet, so I found that in two decades, so between 2002 and 2022, the number of the degrees they were awarding in computer science at the University of Michigan went up by maybe tenfold. So 10 times as many. And when I look back at some of the problems as well that I was solving in my computer science degree, I find that AI would've made some of them a lot easier. In some cases just the click of a button or the press of a tab key, it could have solved some of the problems that have been given.
(00:25:54):
And this is even a popular type of benchmark of assessing AI systems is how good they are at coding. So OpenAI tested the o1 model specifically on OpenAI engineer interview questions, and this had both multiple choice and coding questions, and it got quite a lot of them right, even though these were from the internal question bank that they would give to the real OpenAI applicants.
(00:26:22):
So if I were entering the college, again, I wouldn't necessarily know how AI will impact the job market. That seems like it could be very complicated in software engineering. But I'd definitely feel more uncertain about what to major in. So what would you be telling me? Do you have any advice that you would give to a first year undergrad who wants to get a major and they really are focused on having job security, having a successful career?
Anton (00:26:59):
Frankly, that's a really difficult question because there's so much uncertainty about both the speed and the direction at which this is going to continue to advance. So I know some students who tell me, well, I'm not sure if it makes sense for me to spend so many years to get a degree because by the time I'm done the AI will be able to do everything I've learned and will be able to do it better. And I can't contradict them with conviction. Maybe they are right, maybe they're not, maybe it's going to take a little bit longer, but there is certainly a risk that they're right.
(00:27:42):
And the example of coding that you just brought, that's a really good one because five years ago everybody among pundits believed that this is going to be the field of the future. And if you major in that, you are guaranteed lots and lots of job offers. And nowadays, I think for good reasons, people are not quite so sure anymore. And the declared goal of all the AGI labs is to automate the whole process so that they can develop systems that will improve themselves. And we will probably have that before they can do everything that humans can do. So it's not as safe of a career anymore.
(00:28:32):
So in a recent paper of mine, I tried to think about what would be the roles for human labor in a world where computers can solve all the cognitive tasks that we can perform. And there's kind of two categories. The first one is during the transition. So even if we do have AI and machines at some point that can perform all human works, it's still going to take some time during which humans are also going to do those tasks just because the economy is a system that always adapts slowly. And in particular in sectors where there's no competition or sectors where there's a lot of bureaucratic rules, it'll take a lot of time.
(00:29:25):
But there are also some areas where I think the human identity, the fact of being human will create certain specific jobs that we won't let the AI do. So I'm sure there will be lots of jobs where the authentic human connection plays a role. Some people are going to say, well, I won't send my kids to that kindergarten where they have robots watching the kids, I want a real human. Maybe others will choose differently. In certain performative professions like let's say dancing or sports or singing, people are going to say, well, I really care for this because it's amazing that a human can run this fast, but if a robot can run twice as fast, I won't be impressed. And finally, there's also going to be human roles in overseeing the AI systems and making sure that they do what humans want them to do, essentially, AI alignment roles where I think humans will remain indispensable. And I think it's plausible that the AI will make inroads in a lot of these things, but likely there will also remain human roles. So preparing for that may perhaps be the most future-proof type of job.
00:31:13 - Evolving workforce needs
Jakub (00:31:13):
Yeah, future proof and AGI proof. So we've touched on it a few times, but just for people listening who aren't familiar, this AGI concept is artificial general intelligence, and it's a theoretical form of AI that would be able to do most cognitive tasks. Or the way OpenAI defines it in their charter, where they state their mission is to create it and ensure it benefits everyone, is highly autonomous systems (so, they're working on their own) that outperform humans (so even better or at least as good) as most economically valuable work, which seems to include the physical tasks as well. And so you point out that yeah, there will be some, I think I read this paper you're referencing, so there will be some factors that take time to go away, and then there will be others that are more fundamental if people want an authentic human connection in a job. How much of a protective effect does this have? And specifically how many jobs would be left for humans in that world?
Anton (00:32:37):
Yeah, so I think from an economic perspective, the most useful way of looking at it is what is the labor demand going to be? Because the number of jobs always depends on the level of pay. And if we have lots of jobs that pay a buck an hour, we would not be very satisfied with that, I suppose. So we want labor demand to remain healthy or we want people to earn their incomes in some other way. And in fact, in the long run, I think it may be a much more plausible economic model that humans receive the majority of their income through other ways than through labor. Because if machines can do all the tasks as you described in your definition of AGI, if it can do all the tasks or almost all of the economically relevant tasks without human input, it would be kind of a waste of human effort to make human workers toil away at things that the AI can do cheaper, better, more effectively. So I for one, I wouldn't want my kids to work in jobs that the AI can do much better. I would rather have them enjoy their time on other activities, on something else that's fulfilling, like enjoying their social life, enjoying their family connections, doing things that are inherently meaningful, because the meaning of life is not just to do work.
Jakub (00:34:30):
That makes sense.
Anton (00:34:32):
That would be a fundamental paradigm change from what we have today, and it won't happen overnight. But I do think that if we focus too much on jobs as the only way of distributing income in our economy and society, then we run the risk of facing even more significant challenges in the next couple years - as we enter what I call the age of AI - than we would otherwise if we don't subject ourselves to that constraint and we think a little bit more broadly.
Jakub (00:35:17):
Yeah, the unemployment doesn't have to be impoverishment of people.
Anton (00:35:24):
And I would also say it doesn't even have to be unemployment. Unemployment is something that makes people miserable, because by definition you are unemployed when everybody around you is employed and you are also looking for a job. If we really do face this kind of economic paradigm change where the role of labor becomes less important in our economy, then I think it makes more sense of looking at it like retirement. When people retire, they become on average much happier than when they're working, because they have more time for all the meaningful activities in their life, they have more time to devote to their health, caring for themselves, caring for their family. And I think crucially there is also the social message that retirement is something well-deserved, whereas unemployment is something that fairly or unfairly - I guess in most cases unfairly - a lot of people associate with failure. So I think part of our challenge will be to create a social narrative that if we have these highly advanced AI systems and robots that can do so many work tasks that we humans can just be content with working less, letting the AI and robots do more of the work, deriving some of the benefits by distributing some of the productivity gains that we will make, and collectively entering something that looks much more like retirement rather than unemployment.
Jakub (00:37:19):
And to drill it down on this point, so the picture you're painting of the post AGI labor market is one where humans have less of role than today and a very small role overall just in total over time. And you said likely, so is this inevitable? Specifically I'm thinking when we had horses doing a lot of transportation work, then we made cars and we no longer use horses. We use them in a very small role for transportation work. But whereas with humans, a lot of technologies might displace some labor, but then since humans are very adaptive, we can do a lot of new tasks as well. But the concern as I see it with AGI, is that the AGI will be able to do the new tasks as well, and there won't be a lead, a niche that humans can find (beyond some of these areas you pointed out like an authentic human connection). To me, this seems like a really airtight argument, aside from really hard-to-predict things, like maybe the world decides to force everyone to work or require labor or ban the machines,
Anton (00:38:48):
Right, yeah. No, it is very hard to foresee how exactly this is going to roll out, but you're right, it's kind of by definition, if we really have AGI, then it could do all the job tasks, including the new ones that become invented that are probably invented by AGIs. So I think in the economics profession, the prevailing paradigm is still, oh, we have had lots of technological progress for many centuries and we have always seen that kind of old job tasks have been automated. And then we have come up with new job tasks. But the fundamental difference is that there were always things that only humans could do. And that will no longer be true if we have a AGI.
Jakub (00:39:44):
Yes. This seems somewhat inescapable as a conclusion, but what is the reaction of your fellow economists and the academic community when you bring up this sort of argument? Do they agree that if this level of AI were created, then there would be humans having a significantly small role in the economy, similar to what horses have in transportation today?
Anton (00:40:17):
It very much depends whom you talk to. And I would say there's probably still a significant contingent of people who say, well, this is just a completely implausible premise. This will never happen. And maybe they are right. I don't want to rule that out. As you mentioned, there are various conceivable ways that this type of progress may stop. From my perspective, it's unlikely that it will stop. And among economists who buy that conclusion, who think along similar lines, I think people are going through the five stages of grief. And at first there is shock and despair. And ultimately I guess there must be acceptance. We probably still have a couple years to go through those stages.
00:41:18 - Social support planning
Jakub (00:41:18):
And let's say that many economists or many politicians started to recognize this need for scenario planning that you've highlighted before about what would we do if AGI had this so-called "aggressive scenario" where it comes in five years or less, or a similarly short timeframe. So in your testimony to the Senate's AI insight forum on the workforce issues last fall, one thing you brought up was that this universal basic income or UBI, this is one of the first ideas people jump to is if we don't have money or we don't have jobs providing people income, maybe it can be provided more universally without jobs. Or another idea is just standard unemployment benefits could be revamped, but you write that UBI specifically is not very desirable in our current system. It might become useful as a policy mechanism if we want to prepare for a certain kind of AGI though, and you are talking about a "seed UBI," so this would provide a nominal small, tiny bit of income, maybe just a few dollars a month,
Anton (00:42:38):
For now, yeah.
Jakub (00:42:40):
For now. And it would ensure that infrastructure is in place if we ever needed to do a real UBI. But I was curious what infrastructure needs to be done? Why can't we just do this in a few months? How hard is it to actually build out this infrastructure?
Anton (00:42:59):
Yeah, that's a very fair question. So let me first say the reason why we don't want a significant UBI now in 2024 is because we have a social safety net that is much better targeted than a UBI. It's not perfect, and there is certainly some abuse of it, and there are certainly some inefficiencies generated by it, but still it ensures much better that social support goes where it is needed than a UBI would. Right now, humans are much better than AI at so many things, and we need the humans to work to contribute their unique skills. Now in who knows when, maybe five years, maybe 20 years from now, this may no longer be the case. And at that point, as we mentioned before, it may be desirable to have another mechanism of distributing income in the economy than labor. And then we do suddenly want that kind of system in place. And so the big question is how do we get from here to there? And so this seed UBI is essentially a proposal to make sure we have this system.
(00:44:31):
Now you asked why does it take more than a couple of months? If you remember, for example, the Covid emergency relief measures, that was a pretty chaotic system. There were unemployment insurance benefits that were completely disconnected from how much income people made before and then induced people to stay unemployed even though it would've been good for the economy to go back into the labor market at some point. There were lots of checks sent to people who were no longer alive. And so on. So we would actually want to create a new social safety net, a new system, and maybe it could be based on social security, maybe it could be based on something else. And creating a new system always takes time. Just passing the law alone would take at least several months, but possibly more than that. And then embedding that into an existing agency would probably take a year or so. And we don't know how much time exactly we will have left until we will need it, but I think it would be prudent to start now so that we have the option to expand that what I call seed UBI into a more serious UBI when we need it.
Jakub (00:46:12):
Another benefit I could see is that if you started earlier, you can notice flaws and deficiencies before you're giving out so much money. You can notice if there might be any fraud or double counting, for example.
Anton (00:46:27):
Good point. Yeah, there's going to be lots of difficult questions that need to be resolved, who exactly is included starting at which age and so on and so forth. So that's going to make the political process of actually passing the law difficult.
00:46:47 - Current industry impacts
Jakub (00:46:47):
Now also looking at this intermediate stages between current AI and AGI, it seems like there could already be some disruptions. So I've seen this Swedish FinTech company named Klarna. They put out a report that their OpenAI based customer service agent was doing the equivalent work of 700 full-time agents, I believe customer service call center types of jobs. And the customer satisfaction was on par with the human agents. And they were then not only on par, but were some big advantages. I imagine the wage was much lower for this than hiring the 700 workers. It could do 24/7 work schedule with no sleep, no procrastination, no Facebook, Instagram checking. And over 35 languages at once in one, so to speak, employee. Probably very good accents as well, since they've trained on the best voices they can find,
Anton (00:47:59):
Yeah better than mine.
Jakub (00:48:03):
Maybe better than mine too. And then they also reported that the machine was resolving errors that came up faster than humans. It was resolving them more accurately, finding more long-term lasting solutions than humans. And this isn't just totally an isolated incident. So in the Philippines, there have been two lengthy articles from Bloomberg, investigations of what's going on there. And there's some early signs that AI is disrupting things, that workers need to re-skill, learn to use new AI tools. And then just the overall effect on employment - it's not guaranteed to be net positive. And currently there are about 1.7 million people there working in this business process outsourcing industry. I think it accounts for about 8% of their country's GDP, so it's a really big deal. There's going to be a lot of need for re-skilling, retraining, so potentially to just keep up to speed with this evolving industry or to find employment in new industries. And of course US workers are going to have to do this too.
(00:49:19):
So I was looking at the Bureau of Labor Statistics and they found that the number of times peoples have switched jobs in their lifetime was about 12.7 on average, 12.7 different jobs for people born in the 1957 to 1964, the end of the baby boom. So do you expect AI to be forcing more adaptation than what we've typically seen? Just thinking about current AI capabilities, if we have things that are generating text and generating images, generative AI, but it doesn't really take off in the agent side of things as quickly... Will we be already seeing significant changes, or is it going to be more like what we've seen with the internet?
Anton (00:50:14):
Yeah, it's kind of hard to disentangle what we currently have and what is clearly visible on the horizon. But as you described with Klarna for example, even the current systems are already highly capable. And I think if we froze all further progress in AI and we just spent the next five years diffusing the existing capabilities throughout the economy, the impact would already be quite disruptive. But of course, that's not the world we live in. Progress is not frozen and it is by some measures accelerating. And based on that, I do expect that the disruption is going forward, going to be significantly more severe than what we've seen in previous decades.
Jakub (00:51:10):
Why do you expect it to be more severe? Do we have any way to assess this kind of impact or is it from your experience using the tools with your economics research?
Anton (00:51:24):
Yeah, it is from basically the observation that one cognitive task after the other can be automated by the AI and also from seeing that there are so many things that the AI can already do that we still let humans do right now, even though it is kind of inefficient. So there's kind of an implementation gap between the current technical capabilities and what companies actually use the AI for in practice. And if we just make that gap disappear, if we, as I said before, roll out the existing capabilities, then I can see there's going to be a lot of change. There's going to be a lot of job roles that are essentially going to be performed largely by AI, and that would require a lot of reallocation into the remaining untouched areas. But again, those won't remain untouched for very long given the speed of advances.
Jakub (00:52:37):
And are there any ways the US government could help its workers transition effectively and quickly? So I've heard for example, that generally speaking, it's important if people need to move geographically to give them mobility in that sense. And one helpful thing is you want to make sure there's housing available to respond to changes in where people are living. Do you have any other policies you've thought of for how this might be handled?
Anton (00:53:10):
Yeah, I think that policy advice was mostly based on the times when essentially rural blue-collar jobs were automated and then there was nothing left in the countryside. And people who wanted to move to the cities found it very expensive to do so because housing was very tight there. Now the coming wave is going to be somewhat differently. It's going to be people in the cities who see their jobs displaced. And in that sense, I'm not sure if let's say housing measures are going to be a very impactful lever.
(00:53:54):
I think in the short term, working on AI literacy may be one of the most impactful things, both because interacting with this current generation of AI tools is actually not that hard - you can do it in natural language - and because that's what the economy really needs, the economy needs people who are literate in using those tools, they don't all need to be able to create them. I can't create cutting edge language models, but I can use them. And just like me, I think lots and lots of categories of white-collar workers can leverage that and can conduct useful work for our economy.
(00:54:43):
But it's going to be an adjustment. Let's say for example, call service center workers are going to have to either learn how to work with AI in other roles or, what I also view quite plausible, I think there may be a movement back from white-collar jobs towards blue-collar jobs that deal more with physically interacting with the world in the near term. Now, when the robots really come - because they advance as rapidly as AI - then that may no longer provide a valve either.
Jakub (00:55:27):
And another thing I think about there is a lot of the white-collar workers, the stereotype is that they would not enjoy the blue-collar work or maybe they would find it very difficult. For example, if you think about the people who do a lot of really thinking-heavy desk jobs and then thinking about them going to really heavy day-to-day physical labor where they have to use their muscles a lot, it might be a pretty big transition in career compared to what has historically been. But is this just... in economics, this is not a big factor? Typically people can make these big changes?
Anton (00:56:11):
No, you're absolutely right. It is difficult. Any transition - even between jobs that look like they're relatively close to each other - is hard. But people unfortunately are forced to do it all the time. I guess we are lucky if we haven't done it for quite some time, but people are doing it all the time and it is a difficult transition. And you're right that public programs can facilitate it at the margin. They can provide the right training, they can provide monetary assistance to people moving jobs, and they can mitigate the blow of being forced to change. It is a blow in many cases. It is hard in many cases when your job gets automated.
Jakub (00:57:14):
Okay. This has been a really fascinating conversation. Before we go, is there any last thing you wanted to mention or wish I had asked you about?
00:57:25 - Changes in podcasting
Anton (00:57:25):
Yeah, I was really curious. How do you feel as a podcast host about systems like NotebookLM that do a really amazing job at both developing the script for podcasts and then also generating synthetic voices that talk about it in a very engaging way. They're much, much better than I am as a podcast interviewee. How do you think about it and how do you feel about it?
Jakub (00:58:03):
Yeah, you asked me this via email and it did stick with me. It's a little bit unsettling. For example, I was preparing for an earlier interview and I was writing the questions and I was feeling a bit tired from the day's work and thinking, oh, maybe I'll just write some simple questions, something like, "can you tell me what you wrote in this paper?" But then I realized this is precisely the kind of task that Notebook LM is great at summarizing content, explaining it in an engaging way. And I realized that I am competing against the machine to make the best podcast. And I realized that I might not have such a large lead. I might have a fairly slim lead. And as you said, the AI is already better than me in certain regards. It's much faster than me, lower wage, Google doesn't charge anything for the service.
Anton (00:59:02):
It certainly has more endurance, right? 24 hours nonstop.
Jakub (00:59:10):
And also the voice too. I have been trying to think about ways I can speak in a way that's engaging to listeners. And there you have this AI that's just talking very compellingly and capturing people's attention, and
Anton (00:59:27):
It's so optimized, right? It's like impossible for a human speaker to keep up with that.
Jakub (00:59:34):
Yeah. So I do feel like I... the themes we've been talking about in this interview. I feel uncertain about what the future of this will be like, and I feel a little bit worried about what's going to happen to me and all the other podcasters, especially because I think podcasting is one of the jobs that people really love. This isn't flipping burgers in a fast food restaurant. Some people might love that, but I think some people find that tedious. Whereas a podcast is something a lot of people are very passionate about. And there's not really a grand master plan behind the AI advances deciding which jobs to automate and only automating the boring ones. It kind of comes at whatever it happens to come for.
Anton (01:00:24):
Exactly. Right. Yeah. Yeah, I feel like every time I see a new big innovation coming out, I can see another little part of my work being automated. It's kind of a bit like a death by a thousand cuts.
Jakub (01:00:42):
Yes. Well, on that cheery note, where should the audience go to learn more about your work? Or did you have anything else on that?
01:00:52 - Building shared prosperity
Anton (01:00:52):
Yeah, let's end on a very optimistic note. So especially as an economist, what I do really want to emphasize as the upside of these developments is that the advances in AI are making it possible to generate so much more wealth, to generate so much more surplus. And I don't mean that only in a materialistic way, but also in a broader way. AI can generate so much more content. So much more insightful - I suppose at some point - economic theorems, podcasts, everything we talked about. And I think the fundamental challenge that we will face in the coming decade or so is to ensure that the vast benefits that we can generate through this new technology are shared sufficiently broadly to make us all better off. And I think that potential is there. We could all be so much better off in 10 years from now and so much happier, so much more fulfilled, so much more spending our time on the things that truly matter. And it will take a really big political effort to get there. But economically speaking, it is possible, and I hope we will move in that direction. Thank you.
Jakub (01:02:31):
That is a good note. And if audience members want to read more, should they visit your website, your Twitter? What are the links you'd like to point them to?
Anton (01:02:42):
Yeah, on my website, I always keep an up-to-date selection of the projects I'm working on. So one of my recent papers that deals very much with the questions we have been discussing is called Economic Policy Challenges for the Age of AI. And if you have found this conversation interesting, I want to invite you to take a look at that paper. And there's also my contact information. Feel free to email me, reach out to me, let me know what you think.
Jakub (01:03:16):
Great. Thank you so much for joining the podcast.
Anton (01:03:20):
Thank you again for having me, and thank you for the interesting conversation that I think was still, at this point, much more engaging than talking to a computer.
Jakub (1:03:35):
Thank you, thank you.
Anton (1:03:38):
Thanks.
Share this post