Making AI Real for Enterprises and Society

Dennis Gada:

Some of you may know Scott may have heard him at Semafor as well at the World Economy Summit. What led you to one day say, let's start Devin and do software engineering, amongst all the things you could have done with AI.

Scott Wu:

What we were all starting to see in the field was that reasoning was starting to work. And what was the funnest application of reasoning that we could think of that we wanted to work on obviously is teaching AI how to code. And so that's how we got started. And from the beginning, our whole view was that it would be this full agentic form factor. And that's what became Devin.

Dennis Gada:

You were working on this autonomous software engineer with the group of your co-founders. Why Devin? What's the story behind Devin?

Scott Wu:

The whole idea from the beginning was, at the time it was very tab complete, right? The GitHub Copilot was obviously most popular and most famous for doing this, but the form factor of software engineering was mostly just business as usual, and then occasionally the AI would help you go a little bit faster. And that was kind of what everyone was seeing and using at the time. And our view from the beginning was actually, no, it's not going to be like that. It's not going to be like, oh, you have an assistant that helps you go a little bit faster and a little bit faster. At some point, you will have a full AI software engineer. And a lot of the reason that we felt that way, obviously, is it was kind of the natural consequence of reasoning, right? If a model could go and reason and think logically and solve problems, then you want it to be able to go and test its own code. You want it to be able to go and interact with the systems end to end and so on as these models kept getting smarter. And obviously now we call that an agent. At the time, there wasn't really a term for such a thing, but the way that we really thought about it is like, no, it is a full AI software engineer. And so naturally we thought of it as like, oh, this is our buddy, this is our engineer. And so it should have a name. So the most common way that folks use Devin is if you're in Slack, you're in Microsoft Teams, you're in Jira, you're in whatever it is, you just say, "Hey, Devin," you tag Devin, can you take a look at this or can you go fix this thing or can you go work on this thing? And it starts off a session and Devin has its own computer, just like a software engineer. It's plugged into all the same tools that a human engineer has and it will just go work on the task end to end.

Dennis Gada:

Devin has a very special focus, more brownfield development, not just the basic greenfield. So what are the advancements that you have seen in the last one year and what more is coming? What can you share?

Scott Wu:

If you get to the point where the agent can just do hours and hours of work, as long as you give it a good task, then you want it to have its own computer. You want it to be able to test its own code. You want the opportunity to be able to spin up multiple Devins that work on multiple different tasks. And so I think what we've seen is over the last few months, it's really taken off. I mean, one of the crazy stats that we saw is our usage, if you just look at our usage this week across all of our customers, we're doing, from three months ago to now, from the start of the year until today, we've already grown about 400%.

Dennis Gada:

Wow.

Scott Wu:

And so it's really taken off, I'd say in the last several months.

Dennis Gada:

Indermit Gill, chief economist at World Bank. I mentioned before you joined, I've done a lot of panels, but with a leading economist and a leading technologist on the same panel, is a little unique. You are a leader with a lens of data. So what's your view of the world of AI?

Indermit Gill:

I think in general, the debates about AI's effect on the economy and so on are dominated by the work that is being done in the US and in advanced economies, and maybe to some extent China, but not even that. It's mainly advanced economies. So the debates have tended to be a lot about job displacement, economic dislocation and things like that. So what we are doing right now is we are taking a very careful look or like a comprehensive look at all of the... So all of the evidence that we can get on the effects of technologies like AI in developing economies and you come out in a very different place.

Dennis Gada:

That's interesting.

Indermit Gill:

So the first one, what happens is that you have to see when you look at it from the viewpoint of a developing country rather than a developed country, in general, you come out much more positive AI, much, much more positive. Developing countries. Yep. In part, it is because of the nature of work in these countries and so on. So you find many, like a larger share of the jobs are going to have AI will augment productivity in those jobs rather than displace or reduce it or distort it, right? About three times as many jobs will actually benefit from AI than be hurt by it. So that's very different. That ratio is very different in a place like India as compared to a place like the US. So what we are training our folks to do at the bank is to think of this from the lens of developing countries. There's one last thing. So we always sort of look at the effects of AI on the labor market, because that's the debate here. But if you go to a country like India or if you go to Africa and so on, you find half the people live in villages and they have access to the kind of access that they have to basic education, basic healthcare, public services and so on is terrible, right? Generally speaking, much worse than... And what we see is the potential of AI, not in five years, but maybe in 10, to actually do things that would take a hundred years to do otherwise in terms of improve the quality of these things. That part, those returns are actually much greater in a developing economy. Actually in a least developing country, the potential for that is massive. So all in all, I would sort of say based on that, we have to be super positive about this.

Dennis Gada:

I think a lot of people in this room are from financial services, more complex regulated industries. How have you gotten over that hump of making it ready for those industries?

Scott Wu:

Yeah, you and I have talked about this a lot and I think it's an interesting... I mean, I think there's a certain, maybe I'd call it like Silicon Valley Syndrome, right? Which is that all of these tech companies that are working on really, really exciting technology, the first thing that they think of obviously is like, Oh, how do we use it for ourselves? And it's a very natural thing and every industry has this bias, right? But I think to your point, obviously for it to really reach its potential and to provide the value that it should, it's actually much more about reaching everyone out there and doing so much more. And so from the beginning, the way that we've always thought about it is like, if you think about big banks or healthcare companies, insurance, retail, it's every single one of these companies obviously has so much, so much software need. I mean, I think the reality is every company in 2026 is actually a software company. And so from how we thought about it at the beginning is, we thought about it as, look, we want to be able to solve problems in the real world. And so there's a ton of technical problems. Obviously, how do you manage a massive code base with decades of history, with tons of different commits and architectural decisions? And in Semafor there was a lot of discussion on this that there has been so much of investment in general on AI, like people talk about 500 billion CapEx or more and it'll get even more and so on. And there's a lot of transformation potential, but the return or the impact so far is still to be seen, right?

Dennis Gada:

Maybe there was a discussion around maybe five to 10% of the potential impact is what a lot of the enterprise clients, enterprises see today. What's your view on that? And you think there will be a little bit of a lag and then really exponential impact?

Scott Wu:

Yeah. So it's a great question. And it's one that we get all the time, obviously. And people, I think there's a lot of commentary on this on, why are these big companies investing so much in compute, so much in energy? Isn't it all going to go to waste? Is there even that demand for it? It's kind of interesting because in Silicon Valley, you feel exactly the opposite because everyone is in an absolute rush and everyone's like, I don't even know how I'm going to be able to get enough compute to be able to serve all my customers and all the folks who want that. And I think that's honestly what basically everyone is feeling now, some to different extents from others. And my answer to your point I think is like, for sure, I think 99% of the impact that AI will have is to come. And I think we are very far from getting to a point where there's like full adoption across the board. And so here's one of the stats that I would give is there's about 30 million software engineers in the world. And if you think about what is the total amount of spend that companies spend on putting out good software between wages and cloud and all these different things, it comes out to about six trillion a year. And obviously, I mean, it's worth it. The thing that's kind of crazy is, if you go back to the year 2000, there weren't 30 million software engineers. There was actually less than one million software engineers. And so maybe one way to kind of think about it is hasn't gotten harder to build software. It's gotten much easier, right? But even then, we just have built so much more software that the demand of it has gone so much up that actually there's 30 times as many or more software engineers going and doing that. And I think the reality is just, there's not a ceiling there.

Dennis Gada:

So your prediction is that there will still be a need for equal or more software engineers in the future.

Scott Wu:

It's actually one of the most common questions that I get, especially as the question is like, Oh, my son or my daughter is 16 and should they even study computer science? I'm not sure about... And I always say, of course. And the reason I'd say is... I'd put it this way, what would somebody from the 1960s say about how many software engineers there are today? And you could make the case that what they would say is, Wow, there's no software engineers because nobody's actually plugging in the vacuum tubes or whatever, or nobody's actually typing assembly. And well, in some sense that's true obviously, but of course what actually happened is we've kind of climbed to a higher and higher level of abstraction. And what we found is, before you talk about AI, I mean, Python already is starting to get closer to the idea of, oh, you can just say what you want in English and the computer does it. Obviously it's not quite there, right? But what we found is, of course, along with that, the entire world has digitized and we've made everything into software. And so I think similarly, the same is true. It's like when people talk about coding, this is a silly kind of pedantic thing. When people talk about coding, for example, my thought is, well, maybe it won't be code. You'll just be expressing things in English and you'll be working with these things. Maybe it won't be programs because you'll be thinking about your actual product and your specs. But as far as will people build more, will people do more and more of just deciding what they want to do or what they want to change in their products or building things for their customers? I have no doubt that they will do much, much more of that.

Dennis Gada:

And that's the classic Jevons paradox. 2023, 2024 was maybe doing a lot of POCs and pilots. Now, as we work with enterprises, many of them are adopting it, but it is still incremental, right? Electricity fundamentally changed, now we live and work. And what's your view? How long will that take from an AI standpoint?

Indermit Gill:

Okay. So I've tried many times to sort of give what I call the guild law of technology diffusion. So back in 2016, 2017, I actually looked at this issue about actually 10 years back, because there was actually a lot of concern back then. And the concern really came from this, artificial intelligence, China and the US, how the US is losing the technology war. There were lots of headlines like this, by the way.

Dennis Gada:

This was when?

Indermit Gill:

This is a headline from 2018. Yeah. And then COVID came and people forgot about it for a bit, you know? Yeah. But I remember there was this, Putin was also saying, whoever leads in AI will own the next century, et cetera. And then everybody was afraid that China was putting in these huge investments. I think it was like made in China 2025 was supposed to sort of take these technologies and just diffuse them across. So if you look at things... So at that point when we looked at it, there were three things that we thought were really important. One was big money. There needs to be these big investments and China was putting in a lot of big investments because it was a public thing. The second one was business process innovation, because if you look at the previous technologies and so on, you have to change processes in order for you to actually avail of them. And then the third part was to the extent that you need distributional systems because of that concentration effect, which I was telling you about. So at that time we compared the US, China and Europe, and we found that each of these countries or each of these economies had one advantage. The US led in business process innovation, China led in big money, and then Europe led in these tax and transfer systems and things like that. But what we said also at that time was, or what I said at the time was two out of three isn't going to be enough, you're going to need all three. It should be at least two out of three, but now fast-forward to today, and the same people who were saying that the US is not investing enough in AI technology and so on, the same guys are saying, "Oh, too much investment." This craziness. Because it's not going to be that the US government is going to spend a lot of money on this thing. You don't want the US government to spend a lot of money in this stuff. You want the US private sector to spend money, so they're spending money. So then you say, okay, they're going to spend money and some firms are going to win this race. And then there are others that are going to lose, right? Now, then you say, All right, if the losers stick around. They should go out of business, okay? So when they go out of business, they'll free up capital, they'll free up talent, etc.. and the firms that are growing will grow faster as a result of it. If you think that the US business system is such where that won't happen, then you should be worried about the fact that there's too much is being invested, but I don't think you'll find too many people who are actually worried about that, so you will actually get that journey. Now there'll be some stocks that'll not do well and other stocks that'll do a very... Some will do well and some won't do well, but the economy is going to be much better as a result of it. So that's my view of that. Again, great insights.

Dennis Gada:

Scott, Elon Musk is putting data centers in space. Will that help companies like you? Yeah.

Scott Wu:

I think in the long run, we can debate the time horizon and so on. I think in the long run it makes sense. I think in the short... By the way, I strongly agree with all of your points, Indermit. I think competition is a wonderful thing and people look at all the mag seven who are taking all of their free cash flow and putting it into building data centers, hiring AI people. And the first response from people is, Oh wow, that's crazy. I mean, what are they doing? They're giving up their whole business. And I think the actual, the reality is almost the opposite, which is they would rather forego all of their profit than risk losing what they all see as the even bigger thing that is coming. People talk about the tech waves of the past, and technology always has these waves. Mobile phone, it's like internet, it's like things like that. Personal computer. And for most of these, you kind of see it and you're like, Well, it kind of did take... I mean, it took 10 years, 15 years to get... If you think about the internet was invented, how long did it take for everyone to get onto the internet? It was 15 years. Same thing for mobile phones or cloud is another one. I mean, a lot of folks are still on prem, even though that was 15 years ago. And people naturally say, okay, well then the same thing will happen with AI. I think one of the reasons that it's different, and this is why you see these AI companies growing so fast and the CapEx growing so fast, is I think fundamentally the technology is delivered a little bit differently. So first of all, there's not a massive hardware component. I mean, if you think about internet, you have to go ship everybody a modem. And there's only so fast that these processes can go. You have to go build a mobile phone. You have to go and build all these mobile phones and then you have to get the whole world to be using mobile phones. And then the other thing is you don't have as nearly the same kind of empty room effect where there needs to be a critical mass. So the internet, for example, internet wasn't really useful until everyone you know was also on the internet. And so because of that, it had to build and build until it had that critical capacity. Whereas obviously AI is, you can just be one person and use the AI. And what that means, if you think about both of those two in isolation, one that it's just pure software and two, that it doesn't need the network effect. To your point, as soon as a use case comes online, this is what we see. Now the AI is good enough to do talent migrations, let's say, or now the AI is good enough to respond to SonarQube alerts, All of a sudden, it is correct that anybody who has that as a use case should pick it up today and should start using it. And so that's kind of how you see this rapid exponential growth. And so that'd be my one thought there.

Indermit Gill:

So I just realized why people don't talk about guild laws because even I forget to talk about it. So essentially when we were looking at these general purpose technologies, going back to the steam engine, then electricity, then IT and then AI, what we found based on the data was that it took about 80 years for the steam engine to be sufficiently integrated into production and consumption related activities also, right? So it took about 80 years of diffusion. If you then go forward and say electricity, it was about 40 years, right? So 80 years to 100 years, and then about 40 years to 50 years. Then if you go to IT, it was actually half that. It was about 20 years for all the reasons that you mentioned but it was much faster than electricity and so on. So if you then go on further, you say, okay, it gets halved each time you think of a new GPT, it's going to be 10 years, right? So of course it's going to move. And now that causes problems, that causes problems in terms of, I guess in terms of economic dislocation and so on, but it's the speed that's causing problems. It's not the nature of the technology and so on because that rate of diffusion of these technologies has dropped, has actually increased a lot. The rate of invention of these technologies, I doubt that has changed a lot. I haven't looked at it carefully, but that's the experience of the last 200 years. I don't think it's going to be that different for AI. I mean, as long as you think of AI as a GPT. If you start to think of a specific purpose technologies and so on, you actually find that they can diffuse very quickly, like the mobile phone. You don't need to first get a landline and then get to a mobile phone. But if you think of GPTs, you generally have to go, they have to be built on top of each other. So that's the other part in the sense that for countries that have actually already done the first three things and done them reasonably well, they're likely to proceed faster. That's the reason why I'm so positive about India because of the work of people at Infosys and others, digital technologies have got diffused very broadly. So I actually feel like India has this really good basis on which to build. Now it doesn't mean that India should only just be taking in AI technologies from outside because of choke points, because of all of these other things, and we look at those things also in the World Development Report, but you must read the World Development Report, when we put it out, which is very, very soon, because actually folks at Infosys have been helping us with it too.

Dennis Gada:

Thanks for that. And you're right. I mean, we discussed in another panel also today that while the innovation is happening around the world, a lot of the adoption at scale is happening in countries like India, and that's also because of the whole digital public infrastructure that has been set up, the structured data available. And the problems that people are trying to solve there are different from problems in different parts of the world. What's your message to the leaders? How should they think of the next one or two years and what is the message that they should pass on to their teams about the potential of AI?

Scott Wu:

AI won't replace you, but somebody who uses AI better than you might. And I think in reality, it's whatever you think about these new technologies and these new things, the reality that you see is it is an opportunity for the folks who really want to learn it and understand it and think about, okay, what should my entire new process be? What should the business process be? How should we run the teams? And you see this in software where it used to be, okay, we have an idea for something we want to do. First, we spend two weeks and we go do user research to collect this. Now we have a product manager who does a spec. Then we hand it to the designer, right? It very, very kind of blocked off, very like incremental steps, whereas obviously now you're able to do all of those steps together and a sufficiently talented person can actually do the work with their own AI workforce of doing a lot of that. And what it really means is just that the folks who really do embrace the technology and think about how to use the technology well will do very well.

Indermit Gill:

If you combine the US economy, the Chinese economy and the Indian economy, that comes to roughly 50% of global GDP, but you have 190 something other countries and many of them are actually not particularly resilient. All right. So now you sort of say, Now let me take a look at what attitude do the Indians and the Chinese and the Americans have towards AI. And it's distinctly more positive than the other countries. So now I'm looking forward, I'm projecting this forward and I want to say, Which are the economies that are going to end up being resilient given all of these changes and so on? They're going to find ways to grow by harnessing these AI related technologies and so on. Simply because being first is important because you want to be better than the other guy at using the AI too. So if these economies... And this spells trouble. If I had to sort of say, would it spell trouble for the global economy? No. Will it spell trouble for a lot of these other countries that are saying, "No, let's wait and see how..." I think it will spell trouble for them. It won't spell trouble for the US economy. It won't spell trouble for the Chinese economy or for the Indian economy. That's my sense of it.

Dennis Gada:

Thanks so much for partnership, both World Bank and Cognition. We are just getting started. There's a lot to be done.