Come to here, but I'm pretty certain it's not me. Over at the far end, Aisha Ghanna, the co-founder of Addo. And I've been through that several times. Whether it's Addo — do people say that?
Next is Dr. Tobias Meyer, the chief executive of DHL. Give me an idea of how many parcels you move every day. Just, I mean, I think it's important just to get an idea. Go on.
Well, I would say currently around 100 million a day.
100 million a day? So you're a good barometer of the economy.
Right. The managing director of the IMF, MD. How are the meetings going?
They're going great. I want to say that Thailand is a fabulous host. To the Thai in the audience, thank you.
Okay. Governor Macklem from the Bank of Canada. Look, let's just get this out of the way. When are you going to raise interest rates? Just, let's not mess around.
I'm very good at not answering that question, but I will say—
We haven't raised our rates yet. We've been looking through the increase in energy prices, but the longer this goes, the harder that gets. So we are asking ourselves that question: are we going to have to raise rates?
The only thing that was missing in your answer, Governor, was 'data dependent.'
And finally, Jack Clark from Anthropic. Let's get it out the way. When are you going to IPO?
I'm very good at not answering that question.
Right. So we've got those out of the way, just on the off chance that you thought I might be stupid and not ask it. But there we are. You've had your answers. Managing director, the AI... I come to this subject with some trepidation. Why? Because it's so big.
There are so many issues, and in debates and discussions we usually go in circles, always between the good, the bad, and the ugly. So, let's just begin with where you think we are at the moment.
Extraordinary, fast-moving technology, nothing we have seen in our lifetimes. The speed of penetration is unbelievable. One, two — a huge boost of growth for the countries that are integrated in the AI value chain. Great, that's great. Except this is a small group of countries. And three, it is already starting to show that it can be incredibly disruptive. So, people who worry about financial stability are watching it. Bottom line: love it, hate it, fear it — it is here. It has arrived, and it is changing our lives, the way we work, the way we live, and prospects for our economies.
But to this point — that it's something we've not seen, of size and scale. You believe.
You know, I absolutely agree, but I think that the one thing we underestimate is that despite it being such a huge consumer-led revolution, organizations are still struggling to adopt it at high speed. And there are a couple of reasons for that. One of them is that we've given everyone a co-pilot. So, all the employees feel individually productive, but that doesn't translate into workflows that actually move the operating leverage of the company.
So, how inevitable is that, if you will, inefficiency of execution of the story? We're at the beginning of this. So, isn't it reasonable that employers don't really know how to use it?
Mhm. I mean, it's not reasonable anymore. They would realize that it's not working. So, we've been doing this bottoms-up experiment for the last couple of years. And every time — apart from a few high-performance companies like DHL, most of them are realizing that the AI transformation has been a failed experiment, and they're rethinking how to do it now.
Well, I would share two observations. A, we see it being very important in global trade. If you look at the share of AI that was triggering the growth in global trade in the first half, people estimate 60-80% of the incremental growth was AI driven. That's extraordinary. And we see also some chaos in supply chains to be able to deliver those projects, build all those data centers, also increasing resistance on the ground. So, there's that part, and then there is the adoption part, which indeed takes time. And I would argue 25 years ago, with the dot-com boom, we had a similar phenomenon where it just takes more time to get the internet, to get the PC, and now get AI into your business processes.
Right, but I think probably the difference for you versus the rest of the panel is that you've got 300 planes, and how many employees?
Oh, a small company. Five — did you say 580,000?
Right. So you are a real live example of it. Now, when one of your directorate management comes and says to buyers, 'Look, give me another 25 mil so that I can introduce this particular AI on this particular process, because I think it's going to be the best thing since sliced bread.'
The answer is no, because over-funding IT projects has never been a good idea. So, what we do is we select areas where it's applicable. For us, it's mainly informational processes at high volume — customs clearance, customer service interactions. This is where you can deploy it very well. And as important as good reasoning capability is, and it's fascinating to see how it's progressing, the more important step for us was the agentic setup. When that started, it became really relevant to get into business processes, to get also in some parts to the deterministic outcome. You know, we all learned AI, at least in the initial days, didn't calculate very well. When you know where a parcel is, you don't want somebody to be philosophic about it; you want to know exactly where it is. So, that's what agentic AI really changed, and that—
How? How? Can you just tell — we're going to come to Jack in a minute, but how? Give us a — how does it do it?
So, what you basically do is you look at a process, and you have some general reasoning capability, but you architect systems that do references to databases where you have a special context to solve a special problem, and deliver that result back to a general reasoning capability. It also provides us with more modularity in our tech stack, which we generally like, because it reduces the dependency on specific vendors.
So, we work with some of the world's largest enterprises, and many of the world's most successful startups. And I think the common thing is being willing to do creative destruction — being willing, from the top of an organization, to tear down and completely rebuild parts of yourself with this technology, which gets harder the bigger the organization gets.
So, did you use this phrase, creative destruction?
He nearly had a heart attack when you said it, as a central banker. What does that mean?
So, I'll give you an example. At Anthropic, we recently published a post where we took a large piece of software that we've been using for years, and we completely rewrote it from the ground up using our own technology, and we made something much faster and more efficient than we'd had before. But it took the willingness to sort of take the old way of doing things and be willing to completely change it and throw it out and replace it with something new. And generally, this takes conviction, and it takes a belief in where the technology is going to go. We see this all the time in startups. The reason why startups become big companies is they have found new ways of doing things more efficiently than the ones they're competing with. And the reason why big companies succeed or fail is whether they can reinvent themselves as the technology around them changes. So, a lot of it is: can we get the world's largest enterprises to rebuild themselves around this technology quicker than the startups kind of nipping at their heels, which are trying to replace them?
Is there inevitability, though, that the startups will be able to do it quicker than the old dog learning how to disco dance?
The old dog learning how to disco dance has the advantage of context, right? DHL has tremendous context and a footprint in the world and understands how to get things done in the physical world all around us. But the digital world, maybe. Startups move very, very quickly, and things run over there. But the more you get to the physical world, the more it becomes about these long, deep supply chains, relationships, and sort of people-to-people stuff. So, I think enterprises have a tremendous advantage there.
Governor, why do I get the feeling you're sitting here terrified? Because at the end of the day, you, as lender of last resort, will have to pick up the bill if it goes horribly wrong.
Let's start with the upside. I realize my job here is to be a bit of a Debbie Downer, but let's start with the upside. You know, we've all been talking about AI for 10, 15 years, but what you see across this panel now, and what you see in the global economy is — use your words — it is here. It is everywhere in the global economy. You see it in a global investment boom. We had a global savings glut; we now have a global investment boom. You see it in GDP, you see it in the stock market, you see it in the bond market, you can see it in long rates.
It's buried secular stagnation.
Yeah, it's buried it. We don't talk about secular stagnation anymore. You see it in trade flows. So, it is here. But, you know, I agree with you, Aisha — it's just the beginning. Where do we see it? We see it in big tech firms, we see it in hyperscalers, we see it in chip makers. Where's the big payoff for society? It's in the diffusion and adoption in every company. That's where the big productivity gain comes from. And look, productivity is the fundamental source of rising living standards for all our citizens. So, you know, we want to see adoption, we want to see diffusion, but we need to manage the risks, or we won't reap the benefits.
We'll get to the risks. Stay with productivity for a second, because Alan Greenspan famously, when you finally decoded or deciphered what he was saying, used to talk about the productivity gains from dot-com and the first internet revolution, which were there and difficult to quantify and did lead to a dot-com boom and bust. So I can't decide whether we are nascent with AI in terms of productivity gains, middling, or where we are on that scale. Managing director.
Well, we are in the very, very early days. We see incremental improvements in individual productivity, but not yet transformation on scale. There is this Amara's Law that says a new technology excites us, and we have expectations for immediate benefits right away, but they actually come in the long term, and we underestimate what is going to come later. So, we are now in this early stage. We have very high expectations. They are not going to be met because it takes time for diffusion. And the risk — sorry, I will bring one risk. The risk is we are excited, it doesn't lift the productivity immediately, we are disappointed, we pull out, and then it becomes our problem.
Right, but we're unlikely to pull out. It's just a question of the amount of damage, which we will come to. Yes, Governor, and then Jack.
Well, I think you have to separate, though, the success of individual companies from the overall economy. I mean, the dot-com bubble is a really good example. There was a big boom, there was a big bust. The internet was a fabulous technology. It boosted productivity.
Over a long time. Some, you know, it got carried away. Some companies went bust. Some people lost a lot of money, but it was a great technology and it has hugely benefited society. What we sort of want to do — I think actually what everybody here wants to do — is, you know, you want the thing to build progressively and not have a big setback.
You want to manage it. I mean, it could be that I don't—
That's a big word, manage it. But you're right, Jack.
So, I mean, look — AI has very bizarre properties. Now, I'll just name the simplest one, which is we've obviously been in the experimental stage. Some things are getting to scale. We're moving through that lifespan of a technology, as you all talk about. But there's one difference between AI and other technologies. If you buy a plane or a microscope or a large-scale computer and you bring them into your company, you need to plug it in, you need to put it to work, you need to do all kinds of stuff to make it useful for you. If you buy AI systems, they come into your company and say, 'Let's get to work. How can I help?' Which is unusual. Usually only people do this. And I think that we underestimate the ability for these technologies to sort of enhance productivity earlier than we might think, by being somewhat self-orienting inside the companies or organizations or governments that bring them in. And I feel like this is going to show up in coming years as a sort of — you see this already in some of the data of enterprise adoption of larger companies finding success maybe earlier than we've seen previously.
Where, for you on topic, where is in your scale of priorities, productivity gains for enterprises?
I mean, the goal is — we are trying to make a general-purpose technology that is useful for all of the world's different forms of work. And so, as part of that, productivity and making it get to work simply and easily is a huge part of it. And we constantly are inventing new ways of making it productive. So, we talked about agents. You know, we developed Claude Code, which went from the standard chatbot interface to an agentic interface. In some sense, that's a productivity tool. It makes it more productive more quickly. So, we're always refining the technology in that direction.
I was going to say yes. But I register a small complaint?
We take complaints, yeah. Here's an 800 number.
Agent things — I mean, the agents we get, they can be really annoying.
So, can we work on, like, their manners?
Have you any idea what you've just done, managing director? Back at Anthropic, the MD said they're annoying.
What is this? Well, we have this dial between sycophantic and rude, and so we'll just spin it around the other way. But thanks for the feedback.
Well, I think one of the things that enterprises are really getting obsessed with is this measure of productivity, but maybe that's not the right measure. In other words, is it just the speed and the amount of output that defines productivity? Actually, we should be using other metrics, like the quality of the decision — not just the velocity of the decision it helps you make — the quality of the output, not just how many units you produced. Because if everybody's going to be using similar AI models, even with their own context, you're going to have this baseline of not mediocrity, but nothing that gives a competitive edge to anyone else.
So, I've got one thing here. We published some research where we looked at how our own research is used for technology, and we looked at when they took its advice versus when they didn't. And this turned out to be one way of measuring the advance of it: less pure productivity, but the thing making recommendations that the person chooses instead of what they were going to do by default.
Right, but surely the true gain of the productivity is the next question.
Yes — was that a good idea?
Should we have taken their advice ex post facto?
Yes, so I would argue we are also in this discussion a bit focused on this super intelligent LLM. And it's obviously fascinating. I think we're all fascinated to see the rapid progress, you know, mathematical problems now being addressed and so forth. But I do think for the general economy and the gain in productivity, it's other things that matter. There's an astonishing number of very simple informational processes where there's a much larger number of people employed than those reasoning about these complex problems and what we decide in some headquarters. This is not that big of a population. And if you see also what particularly is happening in China with much simpler physical models, which are extremely important for robotics. And robotics, the physical part of AI, matters very much for a company like us, and I think it matters for the economies because we still have a lot of people employed in those areas. But also the simple informational processes is where the productivity gains are really going to show up.
Which brings me to move us forward to the next area for discussion, which is the ability for everybody to benefit. And I'm going to start, because obviously we've got the LLMs, but one thing I'm hearing here is the use of the smaller language model, and the significance — which to the layman, and I put myself in that category because I really don't know much about this — the LLMs are one thing, but there are many other variants, which will arguably be beneficial in different communities.
I mean, as we were discussing, AI is not a brand new technology. By some measures, it's about 60 or 70 years old, and the world is full of computer vision systems which do very useful work today, robotic systems that allow us to make robots smarter, small-scale language models, precursors to that. The world is full of intelligence. I think the one thing I'd note is that all of those intelligences have up until now been put together and synthesized by groups of people. And the way in which AI is starting to work is we have these powerful LLMs which are now putting the mosaic together and allowing for things to happen that were not possible before. And I think we all underestimate the effect of that.
But let me get the downside of those who do not benefit.
So, let's first recognize that there are some pre-conditions that are necessary to be in place for people everywhere, businesses everywhere, to benefit. The first one is digital infrastructure and access to electricity. Not everywhere does it exist. In Africa, 50% of people don't have access to the internet. There is a tremendous problem with just basic electricity. The second prerequisite is people to be attuned, to be plugging into this new world. The third one is this whole thing about diffusion — is it going everywhere? In agriculture, in healthcare, in education. And that, by the way, does require policymakers to be proactive. And there is a fourth element that I'm very interested in the views of the panel, which is regulation and ethics. If there are no safeguards, then the risks may exceed the opportunities. And above all, we do need an ethical foundation. And I'm not sure we're there as a society.
We'll come to ethics, if I may, in just a moment. I want to stick with—
But here you asked a question: would everybody benefit? Potentially, that can be a very democratic technology at par with the cell phone. And that is where we should be headed — to have everybody everywhere able to benefit in whatever their lives and their work requires.
But on the immediate front, you're going to have to, Governor, deal with advanced societies where there will be job losses that won't necessarily — we don't know — necessarily be made up, ceteris paribus, on a one-for-one basis. And in developing emerging markets, those aspirational jobs will never have been created in the first place, because they'll be being done by something else.
I mean, on jobs, I don't want to diminish the risks, but if you look through every major wave of technology in the past, they have created more jobs than they've destroyed. I'm not saying there hasn't been, you know, creative destruction — Schumpeterian economist term for, yeah, some things get destroyed as other things get created; resources have to transfer from one place to another. But the extra income that this creates creates more spending power and creates new jobs. Many of these jobs, you know, we don't know what they are. It's easy to identify the jobs that are getting destroyed; it's harder to know the new ones that are going to get created. But let me just say it could be different this time. AI could be so powerful, you know, it could do the new jobs too. So there are some risks. I will just say, so far in the data, what you see in most countries is that you can see it in certain professions, but you cannot see it at an aggregate level. Jobs have not been destroyed yet.
We have done a fairly detailed study on what is actually happening in the market. And of course, one study doesn't give us all the answers, but it says three things. One, there is demand for skills, for jobs that either didn't exist or didn't exist in this form. 10% of jobs are in this category. They pay more. Two, if you are in a repetitive task job, kiss it goodbye. This job could be gone. Three, and that was the most interesting part of the study: because AI created this investment boom and asset valuations went up, everybody feels richer — not only the 10% that get bigger pay. So, people with more money, when they feel richer, what do they do? They go and spend. What do they spend for? More services of all kinds. And demand for jobs in services actually has grown. The net impact is positive. What is the problem? Services don't pay a lot. So, you get young people who normally would have gone into professional routine jobs, now serving coffee in Starbucks. And that is not what I think young people in the world aspire for.
With 580,000, you're qualified to talk about this.
Well, I would clearly say that productivity enhancements are good, and I don't think revolutions — technology revolutions — happen overnight or have impact overnight. So, I'm not so worried about that. We will see an infusion of this technology give us a productivity gain over a decade or two, similar to the PC and the internet, and it's going to take time. I'm not sure whether currently the capital allocation is efficient. There are some signs of the supply chain for these large data centers and so forth getting choked up. But that is a different question. It's more the diffusion. And I would again say we need to broaden our view. There are things where we need to have the discussion of values. I'll give you one example. AI-supported inward and outward-facing cameras on vehicles are proven to reduce heavy accidents and fatalities by about 50%. Now, we deploy those in a lot of emerging countries. Very hard to do in Europe, because we value the privacy of the individual higher than we value the people dying on the street. So, that is a discussion that somehow needs to happen. How do we calibrate and then thereby regulate?
The thing that worries me, actually, is that for those of us who've covered this sort of stuff — the fourth industrial, well, you know, you go to the dot-com boom and bust and those jobs that were lost from that. Then you talk about the fourth industrial revolution and those jobs. And at each point we were always told that there will be training, there will be skill setting, that governments will have proper policies to manage it, and it never happened. And so you ended up with an underclass, you ended up with a right-wing backlash, and you ended up with the populism that we've got today.
Look, I think that this panel has exhibited, between the different opinions that we have, that we don't know the speed at which it will disrupt jobs and exactly what jobs will be disrupted. But I want to point out the Singapore government and how they've kind of dealt with it. So, the Prime Minister said, 'We may not be able to protect every job, but we will protect every worker.' And since he said that—
Why should we believe him?
