Agentic Coding and the Economics of Open Source

0:01Welcome to the Practical AI Podcast, where we break down the real-world applications of artificial intelligence and how it's shaping the way we live, work, and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind-the-scenes content, and AI insights. You can learn more at practicalai.fm. Now, on to the show.

0:41Welcome to another episode of the Practical AI Podcast. I'm Chris Benson, Principal AI Research Engineer. And with me today, I have a guest I've been looking forward to for some time now. I have Dr. Miklos Koren, who is a professor of economics at Central European University in Vienna. And he has written a really interesting paper on the effect of vibe coding on open source. So welcome to the show. Really excited to have you here today. Thank you. And thanks for having me. Yeah. So I think this is a slightly different take for us. We tend to, on the show, leap straight into models and all sorts of stuff.

1:29But I know you're a professor of economics and you study incentive systems. And so I'm really interested in understanding how you turn that particular lens of economics onto open source. And so I was wondering if for listeners, if you could, you know, kind of talk a little bit about what drew you into the notion of exploring open source through that lens of yours up front. You know, what was the first thing that said this is something that we need to go study? Yeah, let me give a little bit of background on that. So as an economist, my research is really focusing on competitiveness. So what does it take for a company to be competitive in the marketplace or what does it take for a country to be competitive?

2:23and for a long time I've been really interested in whether it's technology that makes a business succeed or whether it's a talent that they have or maybe both or maybe there's some interaction between technology and talent and I would also call myself kind of an accidental software developer in the sense that economics is a very quantitative science and there's a lot of computational research involved. And when I was never trained as a software developer, but we have to be effective at using your computer. And this part I actually enjoy at least as much as thinking about the economic incentives that you mentioned or the other parts of the science.

3:08And so the story of this paper is, we've actually been thinking with my co-authors, Gabor, Aaron, and Julian, And we've been thinking about the economics of the software industry for some time, various aspects, and in particular, open source, the open source ecosystem. Why do people write open source code? Where are the open source developers? How do they collaborate in space? I think that's really fascinating that, you know, you can work with someone on the other end of the world. And yet, actually, what we see is that most of the collaborations are highly localized. So they are typically from the same city or basically a couple hours drive from one another.

3:53So these kind of things we've been exploring for some time when actually AI came about. And initially we didn't really connect these. So we had the research agenda thinking about software. And then we were looking at AI as consumers. So we were, of course, looking at ChatGPT and the amazing success of ChatGPT. And we started actually thinking about, oh, we should be writing more papers. And, you know, I'm happy to tell more about how AI has impacted science. You can get back to that. But let me come back to the paper. So it took us quite a while to kind of connect these two pieces. And so we are actually doing research on this.

4:35And, you know, software engineering and software development is one of the use cases where AI is really, really successful. So maybe let's try to think about that. For me, the personal, I very much remember the kind of what triggered this particular paper. It was like November or December of 2025. And I consume a lot of social media related to technology and software. And so everybody was doing vibe coding. and everybody was showcasing the app that they have developed on reddit or x and so like look here i did this and you know why don't you download it and and i was just like my social media feeds were just flooded with these type of uh vibe coded apps and i started thinking like why would i do that i can do it myself in half an hour why would i download your app when i can just go into uh cloud code and you know do it myself exactly to the specification that i would need And so this kind of, in my mind, connected Vibe Coding with this idea that you can actually write software for one person or just a few people.

5:48And then I realized that, of course, that has a major implication of how we think about the software industry that we have been working on with my co-authors. So I said, okay, let's write a paper about Vibe Coding. And so that's how we started about that. Yeah, and I love the title, which I don't know that we've actually said yet, which is Vibe Coding Kills Open Source, which is quite provocative when you think about it. And all the components of the title are provocative. I mean, Vibe Coding, there's been all sorts of commentary for and against and, you know, trying to understand it. With open source being very foundational to software coding for the last, you know, three decades at least.

6:37and most commercial packages having various open source components to it. But that is part of a conversation that I know I've certainly had with other people in terms of at this point with these tools, especially, you know, you mentioned CloudCode, and I know that's probably the most popular coding tool recently. I certainly use it. And with Opus 4.5 that came out in November, that really kicked off a lot of innovation. I think a lot of us that do code ourselves are wondering why, you know, I can just go do exactly what I want, or I can take existing software and make alterations and stuff. with, you know, how should we be thinking about open source with this new paradigm of vibe coding or, you know, it's becoming, you know, kind of a prompt engineering for software engineering, whatever you want to call it.

7:37There's a bunch of names coming out now. Yeah, yeah. So I think, and we got a little bit of heat because of the title. I think in - I'm sure. Academia is quite conservative and, you know, this sounds like a click-baity title, but to this my replies is that it's not clickbait if it's true but of course let me explain why we think that it's true um and you know we thought about every word we discussed in detail like should it is this the correct word to use in the title and and i agree with that sentiment of a vibe coding has this uh negative connotation but that's that's the word that people use so so let's just absolutely let's just stick with that and uh so the way we approach the problem there's um at least at the time, there was very little data.

8:21Now that we have more data and we have done some new empirical analysis as well, looking at data. But at the time, we were just thinking like, how would an economist approach this problem? And the way I think about economics, and this may not be the canonical undergraduate economics textbook, but that's what I kind of think that the three main pillars of economics are. One is that people respond to incentives. So whether it's monetary incentives, so they want to make a good wage, but it might be other incentives as well. So in open source, it's typically not, at least not directly some payment.

9:00It could be kudos, if you like, that it's good to fame or it may just be like a good feeling of solving other people's problems and sharing your solution with others. and that's also you feel good about yourself and feel good about helping out others, that's also an incentive. So economics can also talk about that. It doesn't have to be money that's involved. The second pillar I would say is that the economy is a closed system. So if you take away something here, you just basically have to make sure that things add up. So I think when it comes to AI, one thing that is a very scarce resource, I think is human attention.

9:43So this idea that AI is going to take over everything and it's just going to infinitely produce stuff, including software. I mean, it's getting better and better of doing that, you know, longer and longer runs for AI agents. But ultimately then has to be a human who instructs the AI and who reviews the result. And so there is this attention. There is a very limited resource. So if you turn this towards AI, you have to take it away from something else. And then the third pillar, I would say, is how these two things come together. So if something is scarce, there's very few of it in the economy, then the price of it has to go up so that people can adjust.

10:30And so this kind of system thinking in economics actually helps us think about the open source ecosystem as well. Because here, I think, just looking at, you know, individual data sets, there are a lot of surveys about how people use AI. There are surveys about how software engineers use AI. But these are limited. Even if you do this on tens of thousands of people or hundreds of thousands of people, they don't necessarily capture the entire system of engineers and users and how they interact. And we saw that it's very important to capture that kind of interaction that as an economist, so actually all four of us are within economics.

11:14We are initially specialized in international trade, which is fundamentally about this type of market equilibrium forces. but there are different countries that are trading with one another. And so here, you know, it's not countries, but there would be a software developer who's building a software package and there's a user who is downloading the package from GitHub or whichever package manager. And then we started thinking about it and we realized that we actually have the tools in economics to think about this and we just have to make sure that which of these apply to the open source sector and which of these are irrelevant.

11:53So we do think that the response to incentives is still very relevant, even though it might not be monetary. And so we identified that it's whatever developers care about, it's roughly proportional to the human attention that they get or the visibility that they get. So it might be that, and I understand that there are different types of open source projects. It might be that a big corporation sponsors a project and it would be different, but say a hobby developer would typically be kind of happier. And actually, even the corporate developers would be happier if there are more people using their product.

12:32For the hobby developer, it might be that, you know, it looks great on their CV that they have, you know, fantastic contributions to open source and they might get a better job later. So maybe even earnings-wise, they can turn it into money, not just fame, but it might also be the kind of incentives that I mentioned earlier. But in any case, the more users they have, the better. So that's one key component of our theory of open source is that while it's easy to write just a simple project for myself and a dozen people, I'm much more happier if I can share it with thousands and millions of users.

13:12And so that's one aspect. And the other is that what is changing with AI is that the technology of writing software is drastically changing. And so that, of course, affects the cost of writing open source packages, of actually developing the software, but also sharing it with others. It becomes much easier. And if you think about it, these two forces actually go against one another in the following sense, that um so if i have just an idea of of uh of a software package and before ai i might not think that it's good enough of an idea or i might be bothered oh i have to pick a license and upload it to github and write some documentations because they're going to be users who ask questions and And so all of these hassles, of course, GitHub itself is responsible for reducing that type of friction for developers.

14:14So now a large chunk of open source packages are on GitHub. But with AI, I think many of these calls go down, and so it's much easier to produce open source. So this would actually probably create more open source packages than before. by contrast and so this is where i mentioned the this finite resource of attention every developer but also users they are kind of users of open source libraries as well and so they either pay attention to the to the developer so just let me give you an example say in web web development. And that's actually kind of a use case that's perfectly solved by generative AI.

15:01We might agree or disagree about different software use cases, but I think a simple front-end development, it's almost 100 % covered by recent AI models. And I can ask AI to build me a website with a number of features, without ever looking at the libraries that are being included there, ever looking at their documentation. And in open source, this kind of feedback is actually very important, not just looking at the documentation, but if there's a bug, then I can report that bug back to the developer. And so that type of visibility of the human behind the open source package is getting reduced.

15:49Yeah, I find that fascinating. I had not really thought about the notion of the developer's attention being a point of scarcity, you know, in terms of managing that. But that makes perfect sense when you say that, because I know that I would say everybody that I speak with regularly that is doing development, that has become the challenge of, you know, they can produce a volume of code with the tooling these days, whether it's open source or whether it's proprietary. but they are still required. There's an expectation from their employer, certainly of going back and validating and making sure that everything is correct and figuring out the bugs and stuff like that.

16:34And I think one of the things I'm curious about is how that plays in as we're talking about kind of that notion of what happens to open source from all this. How does that, like, what does that imply about the role of the developer going forward and the responsibility of the developer as you're looking? Because, you know, the volume of code that obviously, you know, if there's a company behind the open source project driving it, they're trying to get a lot of users on board and stuff like that. But with these different incentives for different players in that, how does that play out? There's a complexity there that I think maybe I missed and possibly other people have missed too that I think you're delving into i think it's very important uh and so you mentioned proprietary software it's very important to distinguish between the two in that of course for proprietary software as well people like to have many users because they mean paying clients but the business model and when i say business model again it's not necessarily money but so the model of open source libraries is very different from proprietary you can be a very successful proprietary software company with you know even if you don't have millions of clients if you have like deep pocket clients it's totally fine but for open source you you have to have like millions of users to to be successful because kind of the margins are so thin uh and again the margins and there is a great study by um scholars at at Harvard Business School looking at the value of open source in terms of the value that it creates and the amount of work that goes into it.

18:17And the amount of work that goes into it versus the value that it creates, there's, I think, at least a thousandfold difference. So, you know, in a business, that would be like a huge gap. It would mean that you cannot really monetize the value that you create. So the only way open source can survive is if you have really, really big user base. And that's why I think this type of argument that with AI, human attention drops, your user base drops, it's particularly affecting open source. Now, in terms of what do developers do and and how are they... So when I talk about developer, I think of them as having two jobs in this model.

19:08One is to write code and to share their code with others. But the other is that there are all kinds of dependencies that you're using. So if you build a website, you're going to install a whole lot of JavaScript libraries. and so you're selecting which of those to use based on your familiarity with the package or maybe you go and read the documentation. And this is the part that completely disappears. You can still instruct your AI to go with this versus that because you have stronger feelings about one library versus the other. But you could, in principle, just ask the AI to give me a website and it would.

19:48And so that kind of selection of libraries is actually something that we're looking at in the data. By now, I think we have sufficient data to try to tease out the effect of AI. And we're actually looking at website development because that's something that's kind of easy to track. You can actually see what's going on on websites. There's this case of Tailwind, Tailwind CSS, which is a very popular CSS library. and so they had a tremendous increase in usage, a lot of it driven by AI agents, but also a very big fall in website visits and then their particular monetization model really depended on website visits because they had some premium package that they are selling on the website.

20:42So if you don't show up on the website, revenue drops. And so actually what we're doing in a new part of the paper that we're working on right now, we have some preliminary results that I'm happy to share. It's trying to see whether this tailwind story is the exception or whether that's the rule across packages that are relevant for front-end web development. Keep going. I'll ask. There's a thing that you have prompted me, but I'd like you to continue and then I'll ask afterwards. So here we actually did a controlled experiment. So that's, of course, one of the things that you do in science, but in social science it's very rare because how do you control the economy?

21:33But here we actually wanted to see what different AI models would do. There we can control because we can really instruct the AI models. So what we did was the following. We took 100 websites. It's actually a representative sample of very popular websites out there. We described their, you know, what they do, the use case with a product requirements document. And we actually checked, or we asked an AI model to check that there are no mention of technology in there. So it's kind of functional requirements or performance requirements, but no mention of any technology, any brand name. So we have basically, but you should think of, say, banking websites or car dealerships, the most popular, most heavily used websites.

22:20E-commerce has all kinds of different websites. And then we asked various AI models across different families and different vintages to try to build that website from scratch. And then we looked at what is the, and really the only instruction we gave is that you have to use NPM to install dependencies because it's, yes, there are others, but I think it's fairly universal for JavaScript dependencies to be installed via NPM. And basically, as soon as you install everything that you need for your website, we actually pulled the plug and then we looked at like what are the dependencies that the AI model wanted to install.

23:02So we already did this for seven models. So now, you know, we could have 700 different websites, but we actually pulled the plugs. I think we do want to build some of these just to see kind of how different they look from the actual websites that were the seed of the experiment. And then for every model, we know when they came out, like when they were released. So we could see, and there's a lot of correlation, of course. So Tailwind is very popular across all these models. So it's almost universally recommended to be included, but not. So some websites need a calendar, some websites need a chat box or a map.

23:41And so these features are different and the different models have different opinions about this. And so we could track kind of how, when a certain model started recommending a package. And then we look at two outcomes that we can actually measure in the data at the weekly frequency. Obviously, we could go deeper, but I think weekly is sufficient here. One is downloads from NPM. Were these packages more frequently downloaded? And so that, of course, would include the demand generated by AI models. Just go and download these packages. And then the other metric is stars on GitHub, which we think of as a proxy for human attention.

24:27so you really like a package and say okay i really want to engage or at least you know just show that i like this package and i make the effort to go on github and give it a star and so so what we find and actually it's very much in line with the predictions of the model the first one may not be very surprising is that as soon as more and more models start recommending a library for some of the 100 use cases that we have, downloads go up. So for every additional use case for which there's a recommendation, downloads go up by like 3 to 5 million per week. So it's a pretty... In terms of percentages, it's something like 3, 4, 5 % of weekly downloads for the typical package.

25:15But by contrast, what you see in stars is that they often actually go down so at the very least they don't go up as much they are zero or they actually fall so so for packages that kind of become very very vibe coding friendly at least the mechanism and the model is that you divert attention from humans towards the machines and so the machines are downloading but the machines are not interacting with the developer on GitHub. And so that already seems to be, so this is data for basically 2025, where I think a lot of the agentic software revolution was happening. And so already in the first year of that, you can see this effect.

26:01I'm curious, if you extrapolate out, you know, the use of agents selecting, you know, these different libraries for inclusion, especially since you very specifically did not constrain that in the prompt up front. You gave it the choice of doing that. If you were to kind of take that out to the extreme case, is there, given the fact that a lot of the, not only the libraries, but a lot of the tooling, you know, the very notion of version control and such are really human constructs. We humans historically are trying to deal with complexity as we're writing code. And we have created all of these ecosystems of toolings and libraries and stuff.

26:51You mentioned Tailwind. Tailwind exists because it makes it easier to implement CSS for a developer. And it's very friendly. I've used it myself a lot. But you can look at all of these different tools and libraries out there in the same way. Is there a case when you're assuming that you have kind of unlimited prompting available to you on a very high capability model with agents such as, you know, Opus or something? Is there in the future, is there a need to have libraries in that, you know, when you talk about open source and, you know, going back to the title, is there a reason that the model needs to use a tailwind and other, you know, other capabilities to integrate in with it rather than just construct it based on those functional requirements up front?

27:42Yeah. Like what, how does, how does the notion of using these existing things that humans created already versus the AI just going and saying the functional requires this, um, and you know, here's all the requirements I've been giving. I'm just going to produce what, what it is from scratch without the notion almost as disposable code, because if you needed to make a change, you might potentially just do it again, assuming that you can get some consistency. yeah that's a great question and and so i can reflect a little bit on on on my personal experience vibe coding while i'm doing like scientific computing it's a very different like i i spend a lot of time uh um you know instructing my computer but it's of course very very different from what it looked like say two years ago and um and it's not so much so for me personally the the big difference in how I program is not so much that, okay, now English is the programming language.

28:45I think it's actually, it's strange at first, but it took me like two weeks to get through that. Then it took me some more time to realize that structures are still quite important. So you cannot, and this might change with different generations of AI models. I remember it was, say, about two years ago, so it was more the tab completion era and not the fully agentic era, but it was an aha moment where I was working on a scientific paper and I started writing some code in Julia, which is a scientific programming language. I love the language because it has a very good high-level interface, so it's very easy to get started as a scientist who knows little about programming.

29:33But it's also a proper programming language with all the right abstractions and tooling. And I set up, I wanted to solve a problem and I set up the function properly. And as soon as I typed out the full signature of the function, I gave it a proper name. So a lot of scientists, when they engage in scientific computing, they would be very sloppy about these things. I'm going to call my variable X, my other variable Y, and these kind of things. But the function name was a verb. There were meaningful arguments. They had different types, and the types were declared, which is optional, but I thought I was going to declare what I'm expecting.

30:16And as soon as I finished with the first line, Copilot wrote the entire function in one shot, and it was perfectly fine because it understands what I want to do. I didn't even give any instruction. I just had to push that. it understood what i want to do by kind of looking at the names and the types and figuring what the types represent and and then so this is going to be a binary search problem and so i'm going to have to do this and and immediately figure this out and so the way i think about ai in in in software but i think in in a lot of other use cases in definitely in science or other knowledge work is more like a very very capable very fast co-worker as opposed to as opposed to some machine or or some tool and i think that that's still a little bit hard to get get used to that so in science there's a lot of discussion about how to use AI, how to be efficient and increase your productivity by using AI.

31:24And I feel that some of the discussion is misguided as if it were like, oh, you know, can you share with me your skills for, you know, writing a research paper, you know, running a statistical analysis. And so you basically, it's like your colleague. Like, should I share my skills for talking to colleagues in the coffee break. I'm like, no, you just walk up to them and you tell them what you're working on and hopefully they're going to have some useful feedback. And so it takes some getting used to, but I think that that type of, and of course, much like if you're working on a project with a colleague, if you have good structure and good coding practice, it helps you collaborate.

32:09And so if anything, I think the importance of maintaining good structures is more important than before because you have a very effective colleague who can immediately do what you ask and give you feedback. And the better your practices, the more you can get out of them. So it enhances that collaboration and communication between you and your AI colleague in this case in terms of being able to have those structures that exist versus everything from scratch. Am I understanding you correctly on that? Yeah, so I would think that this is primarily kind of a collaboration problem. Now, the question you asked earlier, so maybe there is the next level where, you know, a lot of the work are done by AI and maybe they don't need this type of interfaces and maybe they would, you know, maybe just write things from scratch.