Professor Michael Spence exploring the need to support disruptive economic change as old industries decline and new industries emerge.
Full speech transcript
So I had 15 years in academic administration. Sir Anton has completed 16 years here and has had many other leadership positions in higher education and built many institutions, including this one. So I want to just express my admiration to you for all you’ve done and we look forward to knowing what the next chapter is.
I’m often asked by friends, did you enjoy the period of academic administration? And I say yes, it was very rewarding, mostly, but I’m awfully glad I did it then and not now, if you know what’s going on on my side of the Atlantic in higher education.
I want to talk very briefly about two things, structural change and artificial intelligence and why they go together. And so if you want to talk about the current chaos that is going on, ask us questions later on.
Adam Smith was right. He said two things that people who aren’t scholars studying Adam Smith associate with him correctly. One is the invisible hand, which is the market system is a reasonably efficient tool for decentralising and allocating resources. That actually is not the most important thing that Adam Smith said, but it’s the one that neoconservatives remember because they elevate market systems to the status of a religion rather than a way of accomplishing economic and social goals. The most important one for our purpose is specialisation.
Now, Adam Smith meant specialisation within an economy, but of course, everything that David Sainsbury talked about in the global economy is just the Adam Smith insight writ large. And of course, it is the ultimate source of growth.
Without specialisation, you don’t get scale economies. You don’t get learning curves. You spread your activity over too much territory. You don’t get innovation, right? You get nothing if everybody has to do everything.
And so the fundamental message I want to deliver today is that’s still true and that growth is fundamentally about specialisation and structural change. Let me put that more strongly. I’ve spent a lot of time watching high-speed growth processes in a very large number of emerging economies where you can see these dynamics happen because they happen so quickly.
The fastest way to stop growth is to do anything that stops structural change, full stop, right? There’s a tonne of middle-income countries that stop growing. Why? Because the structural change stopped occurring. Now, I have no problem with the focus on innovation, but if you want to affect the whole economy, you need the structural change and the tools and things that power it to go through the whole economy.
So let me be precise. James Maneka, who’s the vice president at Google, and I wrote a paper early on in the AI, Gen AI revolution. It basically said the productivity potential is enormous.
But if it doesn’t spread across the whole economy, you won’t see it ever in the macro data. If it sits there in the tech sector and the biomedical sector and the finance sector, you know, and the rest of the economy, the healthcare sector, hospitality, real estate, you name it, is kind of lagging behind by years and decades, you won’t see it. And you will pay a price in terms of good jobs and inclusiveness of growth patterns as well.
And so I guess, so the question is, how do we think about AI? Now, when we started the digital revolution, we correctly thought of these as new sectors, right? They delivered services like, you know, maps, e-commerce, right, et cetera. And we think of them as sectors, but we’re entering, I don’t want to call it a new era, but we’ve evolved to the point where that’s not the best way to think about technology. The best way to think about it, in my view, and we can talk about this, is it’s a set of tools, really powerful tools, that are going to be the toolbox that we use to transform the structural economy across the board.
So there’s a lot of talk in Europe about what we need to do to get there. I tried to help at the end of the production of one of the reports by Mario Draghi on this subject. And there’s a lot of talk about Germany, which is an industrial powerhouse that people are very worried about, why? Because they do not have, at scale, the tools they need to turn the industrial powerhouse of the last 20 years into the industrial powerhouse on a digital foundation in the next 20 years.
So what’s missing? One, an estimated 100,000 software engineers, right? Remember, these are not tech sectors, right? These are automotive, advanced manufacturing, and so on. So I think the other piece of the puzzle here is to focus on innovation, because it’s a key input, and on entrepreneurial activity, because it’s also part of this process of doing this, but also on the tools that we need. And the AI tools are coming fast and furious at us.
So that’s the main message. A couple of other things, and then I’ll, I actually prefer conversations, Anton, so we’ll get to that pretty soon.
When Anton mentioned the book, Permacrisis, that was a book written by Gordon Brown and Mohamed El-Erian with a couple of other guys coming along as rewriters. And we did say there are three things driving the global economy. One, shocks coming at increasing frequency, putting the system out of equilibrium, pandemics, wars, climate shocks, et cetera. Second are big structural changes, some of which have negative effects on output and so on, like ageing in 80% of the global economy,.
But the third is these revolutions in biomedical and life sciences and digital and energy sectors. And they’re really interesting and powerful. And so, and I think the most important one is the digital one.
So just a couple of comments on that, and I’ll stop. Most people, when they think of AI, think of Gen AI, LLMs and stuff like that, because they’re the easiest to connect with and so on. Actually, probably the most visible effects of AI will come from non-Gen AI sources, and they’re will show up in the sciences.
There’s a kind of near merger occurring in these fields. For example, the Nobel Prize in Chemistry went to three people who were using artificial intelligence, essentially to figure out how proteins look and are constructed. DeepMind, which is Demis Hassabis and John Jumper, were two of the three for the work called AlphaFold.
AlphaFold is an AI that predicts the three-dimensional structure of a protein from the amino acid sequence that defines it. And it’s critical. It takes months to years in a lab with a highly trained people to do this with crystallography.
And by the way, that input was crucial, because you can’t train the AI if you don’t know. And the training set was about 180,000 proteins. There are 200 million known proteins.
They set out to do this and they succeeded. And then they did something interesting. They predicted the three-dimensional structure of all 200 million known proteins.
And then in a famous meeting in London at DeepMind, somebody said, why don’t we publish it? And they did. So now it’s all known proteins have predicted three-dimensional structures.
The predictions are not perfect. But published in an open source database that’s available to scientists all over the world. And there are an estimated 2.6 million biologists using it now. Now the messages from this are, one, open source is powerful.
Two, we’re going to see more of this in multiple sciences. Three, it will produce an unbelievably dramatic acceleration in those things. And fourth, there really is a kind of an important question that we have to deal with in academia that has to do with the division of labour between the basic science, open source, non-proprietary stuff and the other.
I’ll tell you a factoid, and then I’ll let you decide how important it is. NSF data in the United States tells us that a third of the funding of basic science and technology broadly in digital areas, that’s quantum, AI, gen AI, et cetera, comes from the tech giants. Microsoft, Amazon, Google, Maytag, so on.
And the same is true, by the way, in China. So the status quo, if you want an update, is the gap between the United States and China is closing. The costs of training and querying are coming down.
There’s a proliferation of small models. All of that is good news for the deficit we have in Europe vis-a-vis digital. But the most important thing is that as a share, the open source part is expanding.
And that will cut the academic side in. The last thing I’ll say, and then I’ll turn it over to our conversation, is I think we need to start with people, which means, broadly speaking, education. So Gordon and I have talked about this.
I’m going to repeat it and see if it makes some sense to you. We’re not very far, many years away, from having everyone on the planet have access to a digital assistant, a tutor, a research assistant, a learning partner that’s read everything in every field, in every language, and at some reasonably advanced and sophisticated level understands it. Now think about what that means.
I use it every day, right? Everybody has this. And so we have the opportunity, if we do it right, to accelerate the learning process for everybody. Now in a higher education, in education generally, we’re sort of fussing away.
Do we let these kids use the LLMs, right? Will they write the papers with it? How do we test them, right? These are all legitimate questions. But let me suggest we have answers. We can’t possibly take the downside risks and then tell people they can’t use the most powerful learning tool we’ve ever been handed.
It’s just not possible. And by the way, Gen-AI has a distinctive property, which is you don’t need any technical training to use it, but it has a human-like ability to switch domains. So you can go from computer coding to the Italian Renaissance to Scottish history without telling it that you’re doing it, and it knows.
That’s why it’s such a powerful learning tool, plus it’s superhuman reading capacity. So what’s the answer to this? I mean, I think it’s simple, right? We have to teach people how to use it. We have to tell them that it’s a learning tool and not a substitute for them in writing papers.
And then as it comes to testing, this is flamboyant and I apologise this, but at this point I get fed up with my colleagues, right? Well, how do know if he really wrote the paper, she, he, you know? I said, don’t worry about it, right? We’ll tell them how to use it. We’ll tell them about the ethics of it. We’ll tell them about disclosure, just like we do in footnotes now.
And then we’ll sit down with them for 15 minutes in a room at the end of the class with no technology in the room at all. And we’ll have a conversation. And I guarantee you that we’ll be able to tell whether the student learned something or not.
Anyway, I’m pretty excited, as you can tell, about the possibilities here. And I think the time has come to seize them. Thank you.
Thank you.
Q&A with conference attendees
AM Can I just kick off with one question, and we will open it up, because it really stimulated some great ideas. This theme of structural change, the need to rapidly scale up with software engineering, and really, one of the things that strikes me is if you look at the technological revolutions that we saw after 1945, electrical engineering, chemical engineering, that also required a big public investment, certainly in the U.S., in education. Now, as you said, some of that training is now happening in the big tech companies.
Do you want to say a bit about the balance? If we really want to take advantage of AI as a general purpose technology, in terms of the workforce and the expertise, what’s the balance between what the private sector should do and the public sector?
MS So I would say in the AI area, Anton, specifically, we’re out of balance now. So the academic side can’t compete with the tech giants, because they have the cloud computing. I mean, the inputs for Gen AI are, yours wasn’t working either.
It’s technology, don’t worry about it. It’s not data. The Internet’s got a tonne of digital data.
Accumulation of that did enable this, but it’s computing power and then talent, people. And the tech giants have the computing power because of the big cloud computing systems and the massive number of CPUs they’re buying. And the academics are paid more if they go to work, and there’s kind of a gold rush for AI people.
And so I think we’re hollowing out, even in America, we’re hollowing out that side. And that will have a negative long-term effect on the academics’ ability to contribute. In Europe, because we have fewer of these cloud computing systems and so on, I don’t think that the American model is the one we should try to replicate.
We’ve got to do it a different way. And I think probably public sector investment in that is right. The other part of this is, OK, so that’s engineers.
It’s nice to have software engineers and other people who know what they’re doing. But there’s a lot of other people whose jobs are going to change and skills and so on. And this excellent report does devote some really intelligent attention to how to get that done.And that does involve education.
AM OK, thank you very much. Let’s open it up. There’s a question over there, and one over there, and one over here. So we’ve got three. So we’ll take those three questions, and then we’ll hopefully bring them together.
MS Just one second. But just before you ask your question, so I grew up in Canada. The Canadian accent is what left of the Scottish, because the Hudson’s Bay Company were the Scots who populated Canada. By the way, the Hudson’s Bay Company was founded in 1670, and two days ago it was permanently closed. Canadians are very welcome in Scotland these days. I know. But this is an example of structural change. Over to you. Hi there.
Q. I’m Jane Grant from the Open University in Scotland. So we feel a little bit asleep at the wheel at the moment. You know, 66 high schools in Scotland have no computing science teacher. Many of the schools who have more than 1,000 pupils have one computing science teacher. And in terms of that inclusion piece, 31% of low-income students have a computing science teacher, compared to 69% of high-income students with a computing science teacher. I will stop the stats list in a second.
We have a deficit of 15,000 apprenticeship places in Scotland, and the graduate apprenticeship, which was the T-level equivalent, has stalled over the past seven years. And we have four out of ten high school pupils passing maths at Nat 5. So that’s a problem. I have a 14-year-old who is self-teaching himself through his phone and through AI. So will AI leapfrog the need to find all those teachers that we can’t find? In a country with a demographic time bomb, and we’re falling behind.
MS So it’s essentially the lack of maths and computing science teaching in schools. And will AI, if you like, allow us to solve some of these issues, given that there’s a lot of self-teaching essentially going on? And the lady was pointing out that AI is being almost self-taught by enthusiastic school kids at the moment. So yeah, I mean, this is a powerful tool for filling in gaps in multiple dimensions. So you’ve cited one.
You can ask an AI to tell you, how do you do this stuff? You can get an AI to teach you code in Python. And there are many dimensions to this. We have more time. I mean, AIs that are not even past human benchmarks are being used all over the developing world to basically provide services to people that they otherwise wouldn’t have access to in finance, in education, in health care, primary health care, etc. I could tell you stories forever. But the bottom line is it’s a powerful tool to help with that issue.
Because it’s not that you don’t want to solve it with real people, right? It’s that that takes time, right? And so you’re essentially filling in gaps.
AM I’m going to pick up two more questions. One here and one over there. Yes, sir.
Q. So I’m trying to mix three things in one question. Economic, immigration, and talent, right? Or innovation. Because I feel the U.S. is successful because of H-1B visas, right? They are able to attract high-skilled workers all over the world to the Silicon Valley. If you see, it shows up in the CEO list.
A similar kind of visa was in the U.K., a Tier 1 visa. I used to work in London at that time when they scrapped it in 2009. So they removed the Tier 1 visa.And I look at the data from 2010, the productivity crashed. So are not growing, 0.2%. Is there any relationship? Because we kind of discouraged high-skilled migration from 2009 and see the productivity group crash from 2% to 0.2%. And still immigration issue is not solved. It’s still discussed.
MS It’s that we need to relocate whole immigration system to boost this. Immigration and talent. But we’ll pick up the second question and then we can take both at the same time.
Q. Thank you. Thanks very much, Michael and Anton, Our Scottish Future for organising this. My name is Dan Coleman. I’m an economist at Deloitte. And I wanted to ask a question that speaks to Michael’s specialism in information economics. So there are some markets that we know are characterised by information asymmetries like insurance and healthcare.
And I was wondering whether you see AI as a way to reduce those information asymmetries and make markets more efficient, correct that kind of market failure. Or you see AI more as a tool for innovation and that using it in those kind of markets might be more dangerous than the benefits would justify. And whether you’ve got any reflections.
MS Thank you. Immigration. So immigration and talent and the extent to that, is that one of the biggest driver of productivity? Is this an issue for the U.K. and indeed for other European countries and indeed for the U.S. at the moment potentially I think immigration done right is a real accelerator. I mean, I guess the best example I can think of is go to the paper, the 2017 paper called Attention is All You Need, which is the origin of Gen AI, LLMs and so on.Andlook at the names. You wouldn’t know what country you’re in. You know, there’s Russians and people from all over Eastern Europe, etc.
And Indians. I mean, the leading tech companies in the United States, Google and Microsoft are led by Indians. I mean, I think we’re committing a sort of partial suicide right now on the United States side in this dimension because it enriches the talent pool and eventually the innovation so much.
So I don’t mean to take lightly some of the challenges associated with immigration. And I think, you know, people who are serious policymakers have to work through those issues. You know, have it occur at a scale, you know, that’s manageable.
Try to balance the humanitarian needs against the, you know, the kind of normality part. I mean, how much can we absorb? And it’s not easy, right? I mean, people like Gordon, you know, have dealt with this as political leaders and I don’t mean to make light of it. But immigration is a huge resource.
Both partial and permanent, right? So I think international students are another dimension of this. They enrich the lives of other students. They bring new perspectives and so on. I teach an MBA course with a bunch of people from all over the world and they basically teach each other.
I would say on the information side, we’ve been in the process of partially closing informational gaps ever since the internet came. Because you have multiple sources of very low-cost information. But also the issue of information asymmetries in healthcare insurance and other areas where there is asymmetric information. To what extent can AI be part of the solution to the problems you posed in your 1973 paper?
So, I mean, you know, in my lifetime, because I’m pretty seriously old, there was a huge informational gap, say, in the car market between the salespeople and the others. You know, and now there’s so much information that’s pretty accurate on the internet. AI does supercharge this because it can find it faster, right? I mean, remember before Google, we had a growing batch of information on the internet and the search engines didn’t work.
So it was as if we didn’t have the information. Because you couldn’t find it, you just got the list of rubbish, right? Remember in the late 90s? And really the first AI that we got hidden and not called that was the increasing sophistication of the search engines and related things like things that, you know, the matching things that goes on in e-commerce platforms that do actually have a very large effect on the performance of those things. You know, they’ve figured out a way to figure out what you probably might be interested in, etc.
There’s downsides to them. I mean, TikTok’s figured out a way to get everybody’s attention if they’re under the age of 25. And I worry about that along with a lot of other people.
So it’s not all positive, but it’s a process that’s been under for 30 years now of not only having rich batches of information that are available, but having tools that allow you to make use of it.
AM Thanks, Mike. Can I just add on international students? We’ve got 35% of our population of students here in Glasgow is international. They don’t only bring huge amount in terms of talent and skills and hopefully leave with fantastic skills, but it’s the spillover effect they have on our domestic students as well. It really enhances that. And this is why countries like the UK, and certainly I’m really worried about the US like you, Mike, cannot afford to cut off that talent because it’s a massive asset.
Now, Jim, I think you wanted to come in. And if there’s another couple of questions, then we’ll have to probably conclude. So one question here and two there.
JR Actually, it’s more of a comment. I guess I’d like maybe just to offer an industry perspective on AI. And it relates back to what Dan was talking about on clusters as well. So when we look at these big biomedical clusters or other technologies, every one of them, regardless of where they are, every single industry, every single job, will be affected and enhanced and accelerated by AI. So that then begs the question, where does AI fit within the clusters? Is it embedded within that or is that the bedrock upon which the clusters are built? And I think we need to think about that as a base layer. It becomes almost akin to mathematics to some extent.
So that’s the bedrock. But I would offer just one view on how AI fits into industry. It’s my perception, at least, that when you look at an organisational chart right now and you see the CEO, then the C-level suite, and then it goes down, it has this pyramid.
And then we had all this nonsense about, well, we’re a servant-led organisation, but all they’ve really done was turn the pyramid upside down. Well, we kind of figured that out pretty quickly. But what I will think you will see in the future, and that’s already starting to happen right now, is when you look at an organisational chart for a complex business, it will be akin to a neural network.
You will see nodes, and there will be big nodes in that network. Not necessarily the CEO, in fact, probably not the CEO. And the big nodes in that network will be the people who are responsible for the most important things in that company.
Let’s say operations. If it’s an operational business, that node’s going to be big. And the connective tissue that comes out of that, the connections between the other nodes in the network, will show how busy that is.
Here’s the kicker, though, in AI. Embedded within that, you won’t even know agentic AI will be where that comes in. And agentic AI nodes will be some of the biggest nodes.
They won’t be humans, but they’ll be some of the biggest nodes in that neural network. And that’s already happened. If you look at Microsoft GitHub, so Microsoft GitHub is the first time when we embedded that.
If I was to look at that as an organisational chart, GitHub would be pretty big, and it would have a lot of connections. And it’s blurring between human and AI and organisationalstructures.
MS Those are very good points. I’ve got to tell a story briefly. Please. The head of BCGX was at this AI thing in Paris, right? And he told a story that really got my attention because it has to do with where this occurs in learning and all these jobs.
So they got a lot of software engineers, and so they decided they’d equip them all with GitHub, right? Increase their productivity. So they did. Licenced it, gave it to them, and they watched for about two months, and the productivity effect was zero.
They thought, it can’t be true. So they went and investigated, and what was happening is the coders were writing code, just like they were before, and GitHub was finishing the line. It was autocomplete in a pretty sophisticated form.
And they said, well, that’s not how you’re supposed to use it. And so then they went and got trainers, basically. These are software engineers, not people you think are kind of remote from the technology.
And they basically said, tell it what you want, and it’ll write the first draft for you. And then the productivity shot up in the usual way. So just an example of what you’re talking about.
I mean, there’s a huge amount of learning that we’re going to have to go through to figure out. I had a bunch of labour economists in Milan, and they said, how’s this all going to end up? And I said, don’t even try to answer that question. There’s millions of people and firms out there trying to figure out how to do this.
If you think as an economist sitting in the middle of Milan, you’re going to figure it out before they do, forget it. And oh, by the way, even if you did, the technology will be different in two months. So bye-bye to that research project.
Sorry. Fascinating. Let’s pick up those two questions that are over there, and then we’ll wrap up.
Q. Okay, thanks. Ian Brodie, Health and Safety Executive. Just a quick question.
It’s a nice segue from that last point.
It’s a nice segue from that last point you’ve made there. Dan mentioned earlier about the importance of equality in the Scottish Futures Report and tackling inequality. And is there a risk with AI? There’s one thing about having access to AI, and you’ve discussed developing countries and much more sophisticated analysis of protein folding. Is there a risk that we’ve got new types of inequality coming over the horizon in terms of how AI is actually embraced and utilised and people understand how to use it? T
MS The question is inequality, right? Inequality caused by AI. But let’s pick up the other question. That’s a really good question. There was one more question.. We’ll do the two together.
Q. Hi, my name is Peter Devine. My background is in AI R&D. I’d just like to ask more about the point you were making before about how Europe could take a different approach to AI compared to America and China, especially with things like DeepSeek in the past year or so. It’s shown that you don’t need so much kind of CPU power, potentially.
I mean, a lot of it, obviously, from the AI community, the research community, there’s a lot of kind of conjecture as to whether they sort of back-engineered like GPT, for example. But I think that just shows, as well, that there’s a lack of a moat to some extent if you release these products. So I’m just wondering kind of how you think that Europe could evolve with AI without the kind of GPU resources that the U.S. or even China has to some extent.
MS Thank you. Good. Okay.. So first on inequality, we don’t know. And I think, you know, the message that’s being delivered by people who have thought about it carefully, the human development report this year, from the United Nations development thing, basically, the main message was it’s a choice. We’re not sitting watching something that we don’t control happen. You know, business people, you know, academics, policymakers, you know, individually and collectively can choose whether this is, you know, a digital collaborator or it’s full-throated automation.
And how we make those choices will determine the distributional outcomes. The previous round of digital adoption, which has been very carefully studied, I recommend David Autor at MIT for all of these things, was inequality creating because it took out routine, codifiable, middle-class jobs. And it was full-throated automation, right? I mean, you know, we don’t need a bunch of people.
I think most of the applications of AI that are emerging in the kind of zillions of use cases we’re now getting don’t sound like that. They sound more like, you know, there’s somebody who increases your productivity and helps you get something done either faster, better, or at lower cost, but doesn’t really write the human out of the script. And when we write the human out of the script, we usually get the wrong result.
I mean, early in the Gen AI LLM world, a lawyer in the United States had ChaT GPT. write a legal brief, and he handed it in. And the legal brief had a whole lot of legal precedents that were cited, and not a single one existed. So these things, so you don’t hallucinate, they make stuff up.
You know, the developers will fix that, but not perfectly, right? And by the way, AlphaFold is not a substitute for people in a lab. It just gives them a head start, right? It’s productivity enhancing. Now, I have to tell you, if we get productivity that comes so fast that the demand side can’t fill in the gaps, then both at the sectoral level and at the macroeconomic level, we may have problems with employment and distributional issues.
And we’re going to have to deal with them, because we just can’t estimate them now. What was the other?
The other question was on different ways of doing AI, DeepSeek. I’ll be brief on this.
We got to invest. I mean, DeepSeek did lower the cost, and the open source helps Europe, you know, get into the game. The thing we’re not missing in Europe has been mentioned 150 times already, including in the report, is talent. And that’s the most important thing, right? So what we’re missing is infrastructure and some funding for basic research in science and technology and applied, but the upstream part.
And we could use some more powerful GPU-powered cloud computing systems. And I don’t see any reason not to get there. I mean, forgive me for being an economist.
The way to do this is to fund it centrally, right? This is a very European argument, but Mario Draghi and I have been through this and many other people. Europe doesn’t fund enough stuff centrally. We don’t have to centralise Europe to centralise and administer the basic research budget.
The United States spends 96 billion, or used to until this administration, 96 billion dollars on basic research through five agencies, you know, and it’s biomedical, you know, digital, et cetera. So when I talk to my European friends, I say, suppose we X that out to zero and we said California, Texas, New York, and Massachusetts will do it instead. What do you think would happen? Well, it doesn’t take them very long.
We’d underinvest. And because a politician in California can’t send money to Georgia, even though Georgia has the best digital supply chain research in the country, it’s inefficient, right? And so that’s why we have to centralise it and do it meritocratically, administer it so we don’t have multiple layers of intermediaries, countries. You probably know about the RF after the pandemic.
You know, it’s the 800 billion we were going to do for recovery, 400 billion grants, 400 billion loans. Well, how do we do it, right? I’m just a lot better than nothing. Okay, well, we told all 28 countries, here’s the template, 30% on digital, 40% on the energy transition, send in proposals.
So in Italy, we send in a proposal which didn’t, you know, pass muster. We had to try again. And then eventually they distributed that money and then it gets distributed with some politics intervening to various universities.
And eventually, it gets to the researcher. Whereas in the United States, even though there’s lots of flaws, when the NSF makes a grant, it goes to the researcher. And the fiduciary, the person who has handled some money, is the university.
You see what I’m saying? So we have got some institutional changes to make. And by the way, if we centrally funded and issued bonds, we’d have a bigger Euro bond market and a better competitor to the dollar as a reserve currency. But that’s not the subject for today.
AM Thanks, Mike. Really appreciate it. Thank you very much.
