Patrick 00:00

Welcome back to the Barclays Brief podcast. It’s Patrick here. Now on this podcast series we've talked about AI a lot. We've talked about robotics before. We thought about the investment angle there, as well as use cases such as drones delivering pizzas into your back garden. But there's some new research out from the team as part of this year's Equity Guilt study, which steps back and thinks about the much bigger picture – the macro and market implications of humanoid robotics as they start to enter the workforce.

Now, I think it's pretty sci-fi. It's sort of weird to think about and to put some numbers on it. Our team frames this as a potential $200 billion market by 2035, with the possibility that humanoids could make up around 4% of China's workforce. So today, we're going to think about AI and humanoid robotics through a pure economic lens.

And there's no one better to talk about that than our chief economist, Christian Keller. Christian, thanks for coming back on to the Barclays Brief.

Christian 01:05

Thanks for having me, Patrick. And let's hope it doesn't become too much of a theoretical economics lecture.

Patrick 01:10

Well, I'm sure our listeners would enjoy that, Christian, but let's start with a question about what you found most interesting.

Christian 01:16

The most interesting thing I took away is that I believe that it's the combination of the generative AI, together with that physical AI that we described in that study, that Zornitsa has done so much work of it. You know, the humanoid robotics, that combination will actually allow us to really reap the full potential of productivity gains from AI, generally speaking.

Patrick 01:38

Yeah. What I found so interesting is when I, I've looked at your work, but then I've looked across other pieces of research and the range is massive. How would you position your view in amongst all of the rest of the research being published on the impact of AI on productivity?

Christian 01:55

Exactly. You know, we didn't just want to add another point. You know, I tell you, like you say, it's going from, you know, 0.1% labour productivity, annual labour productivity increase, 0.1%, almost negligible from people like Acemoglu, you know, MIT professor, Nobel Prize laureate, all the way to 3 or 4% or some even forecasting, you know, an ever increasing explosion of growth. And so, our take here was, why is it so different? And could we identify what makes these studies so different? You know, and not to go to fine into the theoretical aspects, there’s always different assumptions or so. But it seems that the key here is that those who arrive to very high aggregate productivity gains are the ones who take, you know, specific tasks where we already can see very high productivity increases from generative AI models, and then just extrapolate them to the aggregate level. And the others who come up with the lower ones are the ones who do general equilibrium models, meaning they see how the entire economy, the different sectors will adjust to productivity increases that may be quite, uneven across sectors.

Patrick 03:04

Yeah. So how do you think about it then?

Christian 03:06

So, what you do see is, you know, the first approximation approach or it's called the Hulten’s theorem. And it's done in economics. You know, you look at certain sectors, you see what the productivity increase is from a task level, you build to the sector level, then you weigh the sectors with their weight of GDP, and you aggregate it. That's fair, and that gets you to the first round. The problem is if the productivity gains in sectors varies largely and you have a high and uneven distribution of these gains, you know, will there be sectors that hold others back and how will they respond?

Patrick 03:39 

So, take legal for instance. What's the impact on the legal profession from AI?

Christian 03:44

That's a good example. So, you have legal, where you know, is one of these sectors, very cognitive. AI models have a very large impact. So, we have a huge improvement in productivity. But that reduces the price for legal services. What is the demand reaction to this? Will we increase our demand for legal services? And so, it increases. Or will it actually lead to a smaller share of legal services in GDP? That's relevant because if that share goes down in nominal terms, the productivity effect of that big improvement actually on an aggregate level goes down. And there are other sectors which may have very little impact from generative AI.

Patrick 04:20

Agriculture or something like that.

Christian 04:21

Yeah, exactly. Like sectors that, you know, that have the physical aspects of it that are not touched by generative AI. And they increase as a share in the economy even though they're very low productivity. That phenomenon is real. It's called, you know, Baumol, an economist in the 60s, you know, basically documented this. That it held back productivity gains, it doesn't exterminate them.

They're still there. But you know it slows them down quite significantly over time.

Patrick 04:46

Yeah. That's Baumol’s growth disease isn’t it.

Christian 04:48

Exactly.

Patrick 04:49

OK, so that's a perfect time to think about humanoids. Because we talked about agriculture not really touched by gen AI. Although of course it is to a degree. But humanoids potentially could impact it in a much bigger way. How are you thinking about the impact of physical AI alongside Gen AI on productivity gains for the economy?

Christian 05:09

We have a simple chart or a very good illustration, I think in our paper. And that is a spiderweb chart. And it starts off with our fellow economists at Anthropic. They, you know, with 22 occupations being on the chart. And they chart where they believe their models create the most exposure of these sectors.

And, you know, as you would expect, legal services, finance, management, engineering, etc. huge exposure. But the spiderweb is very asymmetric. There are certain sectors, you know, like you mentioned agriculture, but also in care services and maintenance and all these functions have very little exposure. So, then we take our work and you know in particular Zornitsa’s team has shown where we believe physical AI will in the near future already creates some significant exposure.

If we then overlay the two, you suddenly, get a spider web chart that's much more asymmetric and filled out, so to say. And that makes me think, that should reduce then, you know, the negative Baumol, or the Baumol’s growth disease and other effects where you have this unevenness among sectors because you've suddenly a much wider spread, productivity gains across all occupations.

Patrick 06:17

Yeah. So, if you put it all together, you've got Gen AI combining with physical AI, humanoids particularly, touching more parts of the economy, boosting more productivity overall, which is good for the economy. It's a pretty optimistic view. The nuances, the impact on the labour force. Yeah, people have worried about this for centuries, from Aristotle to Keynes, all predicted in different ways that machines would eventually replace human work.

Christian 06:43

Yeah, that's a question economists have devoted a lot of time to. You know, why each time we get afraid of a new technology and then we end up, really having unemployment actually going down over decades. And what is found is that there are really two aspects of a new technology. One is automation, where really jobs are being replaced and substituted away. And then there's also the augmentation part, where people become much more productive in what they do. And typically, you have also related to that, the creation of new jobs. You can see this over decades. This is why today we still don't have the mass unemployment due to automation after decades and decades of automation.

Patrick 07:21

Yeah. So automation comes along, new technology comes along, demand for new skills goes up.

How do you think about humanoid robotics on the labour force? If you're worried about a humanoid taking over, what do we as humans have, that the humanoids don't have?

Christian 07:35

If you look at issues like dexterity, if you look at emotional intelligence, you could actually imagine that in a society that gets richer from the productivity gains of AI, you know, there would be certain jobs that people want a human, and humans are still better than machines. And therefore, we believe that, you know, that that picture where human jobs in a way, are being eliminated. We're still very far away from this.

Patrick 07:58

Okay. But can you kind of go back to, I don't know, to the 70s and through to today and in the future, and think about the relationship between income and capital and how that shifted with automation and how you think that might shift over the next ten, 15 years.

Christian 08:12

So that's a very good point, because we may not get mass unemployment, and we may even see wages continue to grow. But in relation to what capital earns, that could change. And you mentioned the 70s, we have data from the 40s. And what you do see since the 70s with the waves of automation, with the new robotics, industrial robotics, and with computerization, we have seen a share, the share of labour as part of total income falling. And that that shift has been ongoing now for several decades. And with that powerful combination of generative AI and physical AI that could even accelerate.

Patrick 08:49

Yeah. Okay. So, if we put it all together, we've got a picture where in the next ten years, AI and physical AI could boost productivity across lots of sectors and therefore the overall economy.

But more of the income goes to capital. And at the same time, demand for things like electricity and commodities will clearly increase to fuel AI. Do we end up in a world therefore where inflation is less about wages and more about inputs like energy? And if so, what would that mean for things like interest rates?

Christian 09:20

Interest rates is certainly a topic de jure. And you know, theoretically, I think it makes sense to look at it from two components. There's a real interest rate and then there's a, you know, the nominal which relates to inflation. Now, what our research suggests is that if we are right about, you know, the CapEx needs and we see it already and from AI, it would probably be increased from physical AI. You have such a demand for capital that we think the r*, at least over the next five years or so, is unlikely to go down, but rather go up. So, you know, higher real interest rates. When it comes to inflation, it is less clear. But what seems quite obvious, and we see it already that there is such a demand for commodities. You know what you call a commodity supercycle. There would be probably cost pressures, price pressures from commodities and probably less so from wages. I think that is also something that, is likely to play out.

Patrick 10:14

Yeah. So, we're not returning anytime soon to very low interest rates pre-COVID, then.  

Christian 10:19

That really does not look likely from many perspectives, including fiscal, but in particular also from what we found in our study.

Patrick 10:28

Yeah. Well, it's a great note. Thank you so much for joining us on the on the podcast today. I've learned a lot sitting opposite you again. So, thanks for being back on the Barclays Brief.

Christian 10:37

Thanks. A great pleasure.

Patrick 10:39

Okay, so it's been great to have Christian with me today. A few takeaways for me. Clearly, the advance of Gen AI in combo with humanoid robots looks to expand automation and augmentation potential across a wide range of cognitive and physical tasks in the economy. And the associated benefits across sectors could unleash a significant acceleration of economy wide productivity growth. But there will be an impact. It will likely further increase the share of capital income in the economy at the expense of labour income and so rising productivity combined with large capital expenditure needs points ultimately to a higher neutral real interest rate. Thanks a lot for listening to the Barclays Brief. Do hit subscribe and we'll be back again at the same time next week.