Event Replay: AI and the Next Development Leap

[00:00:00] Ronnie Chatterji: Hey, everyone. [00:00:11] Ronnie Chatterji: Welcome back to OpenAI forum. [00:00:13] Ronnie Chatterji: My name is Ronnie Chatterji. I'm the chief economist here at OpenAI and your host for today. [00:00:18] Ronnie Chatterji: I'm really excited to be joined today by a very special guest, Professor Daniel Björkegren. [00:00:24] Ronnie Chatterji: Daniel Björkegren from Columbia University, an economist, where he focuses his research at the intersection of AI and global development. [00:00:33] Ronnie Chatterji: I'm really excited to have Daniel here on the forum. [00:00:35] Ronnie Chatterji: As many of you know, OpenAI's mission is to ensure that artificial general intelligence benefits all of humanity. [00:00:42] Ronnie Chatterji: And a major part of that will be ensuring access and positive impact in low and middle income countries. [00:00:49] Ronnie Chatterji: And so the way that I'd like this to work for today is I'm going to have a short conversation with Dan about his research, emerging trends in the space, and then we're going to open it up for some audience Q&A. [00:01:02] Ronnie Chatterji: Dan, maybe to start, why don't we just talk a little about your research focus and how you got involved in these questions. [00:01:09] Ronnie Chatterji: I mean, studying technology in the context that you do isn't necessarily an obvious question for a lot of development economists. How did you think about it? How did you get involved? And what do you focus on?

[00:01:19] Daniel Björkegren: Great. Thanks so much for having me. [00:01:21] Daniel Björkegren: So a lot of the discussions that we're having around AI are actually focused on rich countries. [00:01:27] Daniel Björkegren: But there's about six billion people who live in low and middle income countries. [00:01:31] Daniel Björkegren: And so my work tries to understand what that might mean for them. [00:01:36] Daniel Björkegren: I have a background in technology and a bit in development economics. [00:01:40] Daniel Björkegren: And so that intersection ends up being what most of my work is on. [00:01:44] Daniel Björkegren: I think it's possible that AI could have profound effects on economic development, so I think it's worth understanding what those might be. [00:01:53] Ronnie Chatterji: And then I mean, think about this. This work brought you to Sierra Leone, where you worked with 500 teachers and 40,000 messages to try to understand the utility of AI via chatbot. [00:02:04] Ronnie Chatterji: What made you want to ask that question, given how people use technology in these settings in Sierra Leone? [00:02:10] Ronnie Chatterji: What are some of the things you found in terms of why AI may have an impact or be advantaged versus how people are accessing the Web now traditionally?

[00:02:19] Daniel Björkegren: Yeah, great question. [00:02:20] Daniel Björkegren: If you look across sub-Saharan Africa, one of the surprising things is that many people are not using the Internet or not using what we think of as the Internet, like web search and loading web pages. [00:02:33] Daniel Björkegren: I worked with this organization, Fab Inc, that created a new chatbot, which basically uses ChatGPT but accessible via WhatsApp, and gave that to teachers in Sierra Leone. [00:02:46] Daniel Björkegren: One of the surprising things was that they ended up using it more than they use the web. [00:02:52] Daniel Björkegren: Even though they're just asking questions like you and I would have submitted to Google, they're choosing to use AI instead.

[00:03:00] Ronnie Chatterji: That's so interesting, Dan. [00:03:01] Ronnie Chatterji: For a lot of us, when we think about technology adoption in other contexts, we tend to think it's going to look the same way it does for us. [00:03:10] Ronnie Chatterji: You know, we'll hand someone a laptop, and they'll access the web through the same tools that we do. [00:03:15] Ronnie Chatterji: Why do you think these differences exist, let's say, in Sierra Leone and in other parts of the continent? [00:03:20] Ronnie Chatterji: What are the implications for how we should think about AI in economic development?

[00:03:24] Daniel Björkegren: Yeah, I mean, one question I ask when I give this talk in person is how many people wish they could use the Internet more? [00:03:31] Daniel Björkegren: When you ask that question in the U.S., zero people raise their hand. [00:03:35] Daniel Björkegren: But in a survey in sub-Saharan Africa, when people are asked what are the barriers that keep you from using the Internet as much as you would like, most say there are barriers, and the number one barrier is cost, which we don't even think of in our day to day. [00:03:53] Ronnie Chatterji: Wow. [00:03:54] Ronnie Chatterji: I think we have some results on that, if we can queue them up, some from your Sierra Leone study, if I'm correct.

[00:03:59] Ronnie Chatterji: As you think about the research design and what you were trying to demonstrate in this study, walk us through what the results tell us, if we're able to bring them up now. [00:04:10] Ronnie Chatterji: I don't know if I see them yet. [00:04:12] Ronnie Chatterji: Yeah, Leanne, if we could get those slides. [00:04:13] Daniel Björkegren: So, one of the things that was surprising is that if you load a web page, we found that the average web page takes about 2.5 megabytes to load. [00:04:24] Daniel Björkegren: That's just the text plus the formatting plus images plus any code. [00:04:29] Daniel Björkegren: That's not a lot for you and I who are sitting over this fast internet connection. [00:04:39] Daniel Björkegren: But it is a lot if you have a slow internet connection or one that's costly to use. [00:04:39] Daniel Björkegren: Instead, when people query the AI, the response was just the information they were asking for, and it was just a few kilobytes. [00:04:48] Daniel Björkegren: We found that the AI was over 700 times more data efficient than loading a web page. [00:04:54] Ronnie Chatterji: That's so interesting. [00:04:55] Ronnie Chatterji: Okay, so we see here the data efficiency advantage of AI, which is not something that I was really tracking before.

[00:04:58] Ronnie Chatterji: which is not something that I was really tracking before this, and you're right, like when you think about a typical web search, you and I aren't necessarily thinking about the data efficiency, but it really matters in this context in Sierra Leone. And I think you also have another chart on cost, right, here that shows us sort of the cost efficiency frontier as we look at it.

[00:05:13] Daniel Björkegren: Yeah, so if we go to the next slide, so not only is it annoying if you're on a slow connection to load things that take a lot of data, but it's also costly. Now, if you went back in time before ChatsGPT was launched, loading about 1,000 web pages in Sierra Leone would've cost you about $3, so just in terms of bandwidth. Now, before the release of ChatsGPT, if you'd asked the same queries of AI, it would've been pretty expensive. It would've been about $30, because you need to pay a little bit for the bandwidth, but then actually running these AI models on the big data centers was pretty... took up a lot of resources.

[00:05:56] Daniel Björkegren: Now, those data centers have become much more efficient. Those algorithms have become much more efficient, and so the cost of running AI models has just declined precipitously. So during much of the study, it would've been about the same cost as using AI as using the web, but AI inference costs have continued to decline, and so at the moment that we wrote this, that we updated the paper, we have to keep updating it with the newer numbers, AI was 98% cheaper than... creating AI was 98% cheaper than loading a single web page.

[00:06:26] Ronnie Chatterji: Wow. And just, I'll let this linger one more time, and then we'll get back to the face-to-face, but what I see, this gap between web search and AI chatbots, this is almost like the development dividend, right? It's the idea of like some sort of consumer surplus people are getting from using AI compared to web search by thinking about it the right way, and it's something that, as costs go down, will hopefully be able to democratize access to intelligence in a way that's pretty efficient compared to loading a traditional web page. Is that the way to think about it in terms of how AI can make an impact?

[00:06:53] Daniel Björkegren: Yeah, exactly. I mean, if you're in a constrained environment where bandwidth is important, just delivering the information you need could be a benefit. I think we were kind of optimistic, like historically, 20 years ago, we were optimistic that the internet was going to close all development caps, and it definitely helped in many ways. One big question is, is AI going to close some of those gaps that the internet hasn't quite reached.

[00:07:21] Ronnie Chatterji: Well, let's pivot to that and move away from this fantastic research, Dan, to think about kind of where's AI really having an impact today? You know, I think a lot of people are thinking about the potential, you've seen in your study some of the stuff that's happening right now. If you were going to sort of just give sort of a short summary of how AI is impacting low and middle income countries today, the ones you've worked in, which are many, where is it having the biggest impact?

[00:07:45] Daniel Björkegren: Yeah, so I mean, you're right that it's still early days. So we're still seeing early pilots, we're seeing some initial evidence, we aren't seeing the final word. There's some promising evidence, both this bandwidth efficiency result, another study by Raisa Fabregas and co-authors finds that people in Mexico are getting positive mental health benefits from a chatbot that helps them through that. There's a lot of evidence that people are asking health queries of these chatbots, which is potentially good in contexts where there's not as much access to other health information, but we haven't seen bulletproof evidence yet on a lot of these questions.

[00:08:31] Daniel Björkegren: So I think a lot of the potential is still in the future.

[00:08:37] Ronnie Chatterji: And that's really interesting. And so developing that bulletproof evidence is one of the things that we're really passionate about at OpenAI on the econ research team. And we can only do so much work internally as our team allows, and in terms of bandwidth. And also it's really important to have outside independent researchers doing this work, too, in academic institutions and other kinds of organizations.

[00:08:55] Ronnie Chatterji: So we recently launched The Exchange, which is a funding program for researchers doing work on the economics of AI. And proud to say that Dan is one of the first recipients, was announced today. Dan, tell us a little bit about the project. And I found it to be one of the most fascinating. We got nearly 500 applications, and yours was one of the select few that was chosen. We had already decided to interview you before that, so it was a great happenstance. But tell us a little bit about the project and kind of what that future looks like in terms of what you're trying to accomplish in estimating the impact of AI.

[00:09:26] Daniel Björkegren: Yeah, thanks, and thank you for the selection. Yeah, I think a big question is that it's clear that AI could be useful in a lot of domains that are traditionally kind of been the role of non-profits or governments, but also people are just starting to use it on their own. Entrepreneurs are asking for advice, people are asking for health information, et cetera. And so that project is really trying to understand what are these organic uses that we see? We've seen kind of stories of these emerge over the internet of...

[00:09:56] Daniel Björkegren: over the Internet of taxi drivers using it in interesting ways. [00:10:00] Daniel Björkegren: And so the hope is to find these case studies or these examples of how people are making productive use of this technology and see like what is leading them to be able to make these productive uses. And then also, what are the bottlenecks? So are people being held up by connectivity, by language performance, or other things? Will we need investment either by the labs or by governments or the social sector to really, really make sure that this technology is more broadly accessible?

[00:10:33] Ronnie Chatterji: That's amazing. If you want to see for our forum viewers that fantastic research proposal by Dan and his colleagues, as well as all the other winners, we'll drop that in the chat and people can look to see all the interesting research that we're doing around the world. Dan, one of the things you said on health is really interesting to me, which is that there's a lot of excitement, but also concern about how AI affects expertise. You know, you and I have done our PhDs. We're really passionate about what we do. And for all experts, whether you're a healthcare provider or an educator, when you see something like AI, it could be a complement, it could be a substitute. When you build in the idea though, that in many emerging markets, you don't always have access to education and healthcare or the delivery systems aren't super efficient. The equation sort of changes. So how do you think about that when you go to places that are under-resourced and you think about, look, a lot of people are getting sort of health information from AI chatbots. As economists, how do we think about the benefits and costs of that kind of phenomenon?

[00:11:30] Daniel Björkegren: Yeah, I mean, as you mentioned, it's a different calculus when there wasn't enough human expertise to go around. And so I think I'm quite optimistic that it's possible to really give access to many more people a good quality of service. There are, of course, questions about making sure that the advice that's given by AI chatbots is good for people and that people know how to assess that. But I mean, if you could provide everyone in the world access to best-in-class health and education, et cetera, that would be amazing. And the big question is, maybe we won't be able to give completely best-in-class, but could we give some fraction of that?

[00:12:13] Ronnie Chatterji: Yeah. And this is another theme that you've written about that I think is really interesting because we focus a lot on trying to make things useful for people. And for someone like me, I think about, well, language has got to be one really important barrier to making things useful. If I don't understand a language and it's not translated into my language, it won't be useful to me. And so at OpenAI, we spend a lot of resources in terms of making sure our models can be accessible in a wide variety of languages around the world. But you made the point, well, this is positive in and of itself. Language accessibility doesn't necessarily equal local usefulness in the market. Tell us a little bit about that. That divergence, I think, was really important both to people at OpenAI, but also people around the world who are watching.

[00:12:53] Daniel Björkegren: Yeah, it's a great question. I should say that the ability of these models to converse in languages other than English has gotten much better over the past few years. I'm hearing that from the field. But there's also some complexity even within language. People may not speak the kind of standard variant. They may speak a local dialect, or they may code switch in between multiple languages, and they may switch into French, for example, to talk about medical conditions. And so we want to make sure that these models speak in the way that people speak. And then even beyond the spoken word, it also needs to understand their local conditions. And so to know what type of medical conditions might be prevalent in their area, or what equipment might be available in a school, or even what time the local market meets. There's a lot of information. Some of it is static, and so it might be easy to train models on. But some of it is just going to require digitizing much more of these societies. And historically, there have been efforts to do that, like Open Data Movements, etc. Those are becoming more important now that that data is being accessed, not just by humans, but also by AI systems.

[00:14:08] Ronnie Chatterji: This is why I think the work you're doing and others in the development community is so important for us to understand at OpenAI. I feel like even thinking about the impact of AI on the economy, these economies that you're working in are often left out of that discussion. And so you recently also wrote about that, but you've been busy writing about a lot of great things. When you wrote about that, you said, look, we need to think about the intelligence economy, not just among the rich countries, but also how it might affect countries at the other end of the spectrum. Talk a little bit about that and how the insights from that essay should change the way or affect the ways that economists think about the intelligence economy.

[00:14:44] Daniel Björkegren: It's a great question. I think in the short run, it's easy to imagine how AI could be beneficial for low-income countries. There's the services gap.

[00:14:54] Daniel Björkegren: Services gap, it's clear that AI is capable of providing some amount of help in education and health, etc. In the longer run there's this bigger question, which is that traditionally economies have grown from producing agriculture to manufacturing, to finally to services. Developed economies now are very service-focused and they have large knowledge sectors.

[00:15:20] Daniel Björkegren: Low-income countries don't have huge knowledge sectors, and over the past couple of years it's been clear that these AI models are effective not just at providing information, but also providing intelligence. They're solving complicated math problems, they're doing complicated data science, they're analyzing market conditions in complicated ways. So you have, in wealthy countries, humans doing those types of tasks already who are now empowered by AI.

[00:15:46] Daniel Björkegren: In many low-income countries, there weren't as many people doing those kinds of heavily math or analysis-based tasks. And so one big question is, what does this mean for the path of development? Does this mean that the traditional pathway of agriculture, manufacturing, services is no longer going to hold, or is this possibly a big opportunity? If you have countries that never had enough human scientists that suddenly could have some artificial intelligence help, you could potentially have a different pathway to development.

[00:16:28] Daniel Björkegren: I think there are enormous questions because it's not clear what that pathway would look like, but I think it's such of deep importance that we need to start thinking about what it might be.

[00:16:40] Ronnie Chatterji: Yeah, I think if economists have been thinking about a traditional economic development path and this arrival of really capable AI, maybe we'll change that path or allow countries to skip steps or imply different policy levers to get from here to there. All really interesting topics I hope you and the other exchange winners can delve into in your work.

[00:17:03] Ronnie Chatterji: The other thing I think is really interesting about the work that you and a lot of development economists do, it's very grounded in the context. I think a lot of people, maybe outside economics, would be surprised how grounded you are in the context and the work you do. One of the things you do is beyond just giving people, let's say, a chat bot with AI, you have to build trust in the community so they understand that this is a research project but you're trying to solve a problem that actually matters to them.

[00:17:27] Ronnie Chatterji: It strikes me that experience, Dan, is probably similar to how when we think about implementing AI around the world and democratizing intelligence. Anything else that you could suggest for technologists and organizations in terms of helping communities understand, trust, and use AI in the ways that they want to? Any ideas from your work or experience that can help us there?

[00:17:44] Daniel Björkegren: Yeah, I mean, I think people build trust when they see the system gives good answers that are locally relevant and it's kind of, in the same way that you and I are learning to trust AI, I think when we were using AI three years ago, there were some things we didn't trust it for that we very much trust it for today.

[00:18:07] Daniel Björkegren: And I think you kind of see two models. The one is this organic adoption where people are using these systems on their own. And then there's also a lot of NGOs and governments that are figuring out how to use AI within their existing services. And that's a really interesting model because they have a lot of contact with the people who are benefiting and they get to see firsthand what the barriers are if people are upset about certain behaviors of the systems, etc.

[00:18:36] Daniel Björkegren: I mean, often it comes down to last mile problems. As you all know, when you're using a product that like seems quite good and then it misses something that's kind of fundamental for your particular use case that it just isn't going to meet your needs. And so those are going to be different from community to community. And so having really boots on the ground and really interacting a lot with users is very helpful.

[00:19:03] Ronnie Chatterji: We have this position, which is kind of like the unsung all stars of OpenAI called forward deployed engineers. And they're the ones who are out there in the communities, in the organizations really trying to rock these use cases and figure out how to get things done. And I read their dispatches on a Slack channel we have, and it's always fascinating to see the kind of challenges that they overcome.

[00:19:23] Ronnie Chatterji: A lot of them are running into the problems that, you know, you and I would run into in the field as economists, like, hey, we might not have consistent Wi-Fi here. You know, that demo I was planning needs to be changed. And I was I'm really excited by sort of that position here at OpenAI because I think it's allowing those folks to help us navigate some of these challenges and ultimately, hopefully, build trust.

[00:19:48] Ronnie Chatterji: Another theme that I wanted to ask you about that's really emerging from our research, Dan, is this idea about capability gaps or capability overhang.

[00:19:52] Ronnie Chatterji: capability overhang. We're noticing when we study the capability advancements in models that they're not always taken up by individuals, let's say in an organization, or they're taken up unevenly, where you see someone at the 95th percentile of, you know, sort of like token usage, you know, running agents for hours and hours on end. And the person who's at the median, mostly using AI like a chatbot. And it strikes me that this capability overhang idea, the notion that models are more capable than how we typically use them for, is really sort of salient in the markets that you work in as well.

[00:20:24] Ronnie Chatterji: I wanted to see how you thought about that. And also what we can do to make sure that we help people understand what capabilities are available to them, given there's probably a big overhang issue in emerging markets too.

[00:20:33] Daniel Björkegren: Yeah. Yeah. It's, they're, they're huge questions. I mean, and, and here, like, I think for, for certain people's jobs. So for analysts, for software developers around the world, it's very clear how useful open these, these tools can be. And I think many of those people around the world are, are making good use of the capabilities of the most advanced models.

[00:20:54] Daniel Björkegren: There is a big question though. I think when you, when you're looking at like a small scale entrepreneur who doesn't have a computer, they are working on their mobile phone, the interface through which you use a coding agent for example is maybe not the easiest for them. They may not want to be involved in the nitty-gritty of the kind of technical analysis of how, how you're deciding what would be a good price to charge or how to design a product, but it may be actually very useful to use some of those capabilities and then see that the final output.

[00:21:38] Daniel Björkegren: And so I do think as, as models become more advanced, it's possible that, that kind of will close that, that overhang on its own in the sense that, that people want the results of the analysis even if they're less involved in the kind of the detail. And this is something I think I'm pretty excited to study going forward.

[00:21:57] Ronnie Chatterji: You know, it's been really a big topic of conversation among economists recently, well, if AI capabilities are so advancing so quickly, how come we're not seeing sort of all the transformation of the economy? I think a lot of it has to do with the capabilities not really being used at scale yet. As that starts to happen, we're going to be in a really interesting position to study bigger economic impacts.

[00:22:21] Ronnie Chatterji: And I think it'll be showing up in the markets that I've worked in, but also the markets that you've worked in as well and done research in. So I'm pretty excited about that. If I think about what you're excited on, Dan, what are you most excited on to research next? Besides your exchange project, which I'm excited about, what are you really excited to tackle next in terms of your research?

[00:22:34] Daniel Björkegren: Yeah, I mean I think the enormous question is this question about growth pathways. And is there something, I mean, one, these technological changes may just fundamentally change the momentum of where this is going. But the kind of bigger question is whether we have some control over that. So are there things that governments, that the labs, that the social sector can do to shift that direction to make sure that there's more opportunity and more usefulness coming out?

[00:23:06] Daniel Björkegren: So I think that's a big question. And then in terms of the kind of diffusion, a lot of these questions are last-mile questions. Like we need to not just deploy the technology, it's not just an app, you need to rethink the processes and the systems around that. And so that is, I think, a very exciting prospect of really rethinking what should a healthcare system look like that's AI-enabled, what should schools look like, etc.

[00:23:37] Ronnie Chatterji: Yeah, this is a fascinating opportunity to make sure the AI transition goes well. And I agree, it's going to be things that the labs can do, that governments, multilaterals can do. And I just think there's a multitude of questions that we should be answering in that regard.

[00:23:49] Ronnie Chatterji: Two more questions for you and then we'll go to the Q&A because I know you're already generating some great questions from our forum community. When you think it's just about takeaways, I mean, you've talked about so much interesting research. For me, the takeaway about sort of just how people use technology in Sierra Leone and other sort of emerging markets is super interesting to me, and changed the way I think about it.

[00:24:11] Ronnie Chatterji: If you were going to give our forum community just one takeaway from your work and what you're trying to do, what would it be?

[00:24:15] Daniel Björkegren: Yeah. So I think there's a tendency to think of low-income countries as targets, targets for aid or for sympathy, and I really hope that we can view them as contributing. There's so many people, enormous populations of hardworking people who are creative and have lots of ideas, and I think we're going to learn things from how they use this technology, from how they're empowered by it, and I'm really optimistic about what a world could look like that really empowers the rest of the world. I think that mindset is important.

[00:24:50] Ronnie Chatterji: I think that mindset's important. I travel a lot for this job, and when you go to these markets all around the world, you learn stuff if you have an open mind. And if you just think about sort of countries in one way, or just as sort of this is a place where the technology will be used exactly the same way that it's used in other markets, you miss those kinds of things. And you also deny people the agency and the ingenuity that they have to contribute. So I will definitely take this on board. I think the other folks listening to this will do so as well.

[00:25:16] Ronnie Chatterji: Maybe last question for you. How do you use AI personally as an economist? Have you had an aha moment? I've had a few lately with the voice node. What do you see out there and how do you use it?

[00:25:28] Daniel Björkegren: Yeah, no, it's been like a progression of aha moments. It feels like being in a candy store every few months. There's something new and exciting. Yeah, I mean, I think it's been incredibly empowering for a lot of the data science work that I do. I'm not a macroeconomist, but I got interested in a macro question last fall and ended up writing a short macro paper. It's just very exciting to be able to think of new ideas and not kind of sit on them, but actually just create a proof of concept, just start something very quickly. So it's been very generative for me. And I'm excited to see the other work that colleagues put out using it as well.

[00:26:06] Ronnie Chatterji: Fantastic. And Dan, before we go into the questions, any place people should go to check out your work and what you're doing?

[00:26:14] Daniel Björkegren: Sure. If you can spell my name, dan.bjorkegren.com is the best place.

[00:26:20] Ronnie Chatterji: That is awesome. And we will drop that link in the chat as well. Dan, we're gonna go to questions, but before just thank you. Dan Björkegren from Columbia University, amazing development economist, really working at the frontier of AI and economic development and a recent recipient of our exchange programs fellowship, now at OpenAI, economic research fellow.

[00:26:41] Ronnie Chatterji: Let me start with a question from the audience. We have Johan. Johan's asking from your point of view, how does the advent of powerful LLMs impact how governments should work to improve internet coverage and infrastructure in low and middle income countries? So, a question about sort of the advent of LLMs now affecting something the government's been thinking about for a long time and working on, which is internet connectivity and infrastructure in low and middle income countries. How does that sort of change things now?

[00:27:11] Daniel Björkegren: Yeah, it's interesting, I guess. And I've been thinking about it in two directions. So, one is that once you have connectivity, then you can access LLMs. That's great. And so, it's important to make sure that connectivity is widespread. But at the same time, if previously, people had to get information from the web or from videos, and they had to scan through lots of information to find what they were wanting, if you have an AI system that is accessible over voice or via text that delivers exactly the information they know that they want, you might not need as high bandwidth internet connections.

[00:27:53] Daniel Björkegren: And so, I think it's possible that it cuts in both ways. Of course, there may be other innovations that use more bandwidth that we figure out later, but as Ronnie, as you mentioned, voice mode is pretty incredible. And so, if you had an assistant that could answer all your questions using that, maybe we wouldn't need the internet as much.

[00:28:15] Ronnie Chatterji: This is very interesting. And Dan, as I have you on this question, how do you think that the big multilateral institutions like the World Bank and the IMF are seeing AI? Is this something that is right on the top of their agenda? Are there other organizations that people should be tracking or if they wanna contribute technical knowledge kind of places that you'd say, look, these are the places they're gonna make these decisions? What do you think in terms of the multilaterals and the big institutions affecting AI in lower middle income countries?

[00:28:40] Daniel Björkegren: So, they are all paying attention to AI. So, the World Bank's World Development Report is on AI this year. IMF, I also know has been thinking about this. I also have a hat at the Center for Global Development, a think tank in DC, that is thinking a lot about AI in development. So, there's a few places to look. But I think the sector is paying serious attention to what AI will mean.

[00:29:09] Ronnie Chatterji: Fantastic. Next question comes from Drosko. In regards to AI and global education, great question Drosko, how can we focus in education on adoption, not just availability? And probably Drosko is thinking about, wow, AI tools might be available, but people aren't necessarily using them. How do we focus on adoption? How can we make that happen?

[00:29:27] Daniel Björkegren: It's a great question. And I think there's two pieces of this. One is you're gonna have organic adoption regardless of what you do. Students may use this and without certain policies or without thinking it through, they may use it in ways that circumvent learning. And so I think we need to think carefully about how to.

[00:29:48] Daniel Björkegren: about how to make sure that we're enabling the good uses and I think it's also it's not just the technology we need to think about what happens within schools and how the whole system changes how our students spending their time how are teachers spending their time. I would like to see a lot more pilots of that, of really creative ways to think about using the school day and using the advantages of the technology and the advantages of teachers and their ability to motivate students to learn more collectively. So, yeah, I think there's a lot of work to be done.

[00:30:27] Ronnie Chatterji: Well, and then on that, I mean, what is the role of philanthropy? You know, we've talked a lot about governments and about the labs and some of the multilateral institutions. There's obviously a huge boon in philanthropy relates to AI, of course, opening our Foundation but others as well. Where have you seen philanthropy sort of make a positive impact when it comes to previous technology waves, and are there places that you might suggest that like we really need more research or evidence that could look at?

[00:30:54] Daniel Björkegren: Yeah, I think the labs are paying a lot of attention to things that are within the technology and maybe one degree away, but there's a lot of things that are maybe two degrees or more away, of like how does it look if you have an AI-forward school system. I think seeing innovative pilots that are really thinking about how to reconfigure these systems, the social systems, would be very valuable.

[00:31:20] Daniel Björkegren: Another is that there are bottlenecks. We talked about language and kind of local connectivity, local data that also maybe there's a lot of possible opportunities that are just kind of shut out because of these bottlenecks. And so, hopeful, one aim of this work that we're going to be doing together is to really quantify these bottlenecks and understand which ones are the most important.

[00:31:39] Ronnie Chatterji: Very exciting; I'm looking forward to that.

[00:31:45] Ronnie Chatterji: I think there's a lot of people who can contribute to this agenda and, you know, we're obviously economists and we're talking about economics, but there's so many different fields that can bring their expertise to bear, and that's pretty exciting for us to consider as well.

[00:31:56] Ronnie Chatterji: Another question, really interesting one from the community, is around common misconceptions. This is a great question to ask an expert like Dan. What is the most common misconception about AI and development that you'd like to correct or maybe some conventional wisdom that you'd like to bust, if it's not a common misperception?

[00:32:16] Daniel Björkegren: Yeah, I mean, I think the study that we talked about at the beginning raises this question where I think people think, you know, AI is an advanced technology; it seems like new and scary. And so, like, why should we think of that in the poorest places in the world? There the misconception is, well no, actually, there are ways in which it's much more appropriate to those places than the previous generation of technology. And so I think that we'll find a lot of cases like that.

[00:32:38] Daniel Björkegren: One other thing is that in the media we talk about AI being very expensive. But if you look at the actual cost of building systems, there's about several hundred million people around the world who don't have smartphones; they have basic phones. You could build an AI chatbot for them or an AI system they could interact with via SMS, that kind of old-school text messages. It turns out if you're a developer and you're building an app over SMS, that costs about five cents per message. That doesn't sound like a lot, but that means that if you're building like ChatGPT for SMS, you're going to spend 98% of your budget on sending these text messages using this 1990s technology, not on the big fancy AI model.

[00:33:31] Daniel Björkegren: And so I think there's just a lot of other costs that are very high relative to AI, which is another thing to keep in mind that we need to think a bit more creatively about where it sits in relation to everything else, and the other constraints that people face, and the ways that it can alleviate those, and then the ways that it's going to continue to be constrained.

[00:33:52] Ronnie Chatterji: Yeah, I think then, I mean, you've kind of illuminated a lot of great points here. One that's really sticking with me is that, you know, there hasn't been as much conversation as there should be about AI and its role in these markets that you're working in and countries. And I think some of that comes from a misconception, like you say, that it won't be very useful or that these markets aren't prepared for it. But in fact, what we're finding when you look at the cost and data efficiency, these might be the ideal places to have AI implemented.

[00:34:19] Ronnie Chatterji: However, that still means that we're gonna have to align the policy environment, the business environment, to make sure that the transition to AI goes well. And those are things I think a lot of folks are really interested in, and there's ways to think about how to harness that intelligence. I think that policymakers, local entrepreneurs, innovators, NGOs, and others will have a lot to contribute to, so I'm excited about that piece.

[00:34:40] Ronnie Chatterji: Let me ask you another question about sort of the future, and we talked about what you're excited about. You know, you and I, in our other conversations, talked about the excitement around the internet or smartphones.

[00:34:46] Ronnie Chatterji: excitement around the Internet, or smartphones in low and middle income countries, it's like there's always a lot of excitement that some of the traditional challenges can be solved by technology.

[00:34:53] Ronnie Chatterji: And while it's made a big difference, you always have to be careful about sort of overhyping these things or expecting too much. If you and I were going to come back in five years and do this forum event, would there be some sort of set of outcomes or an indicator that would convince you that AI had really facilitated a leap in economic development?

[00:35:07] Ronnie Chatterji: Is there something that you could look at and indicate that we should be watching so we can hold ourselves accountable to these big ideas and big dreams?

[00:35:13] Daniel Björkegren: Yeah, I mean, I guess there's two things. One, even if AI doesn't get adopted in low income countries, if it's widespread in high income countries and changes the global economy, they will still be impacted through the paths of development through receiving new innovations that come out of wealthy country markets or also harms.

[00:35:35] Daniel Björkegren: I mean, we've seen these kind of cybersecurity questions that have arisen. So I think there's one set of questions which is just globally is AI going to have the kind of widespread economic impact that a lot of us believe and those will trickle over.

[00:35:54] Daniel Björkegren: Then in terms of these development specific things, yeah, I think if we are seeing big improvements in education and access to healthcare and I think those would be the kind of metrics to follow that I would look at mostly.

[00:36:14] Daniel Björkegren: But yeah, it's a big question. Also having gone through that period where a lot of people thought the Internet was going to magically close the gap between wealthy countries and low income ones, and it did a lot of things but low income countries are not as wealthy as wealthy countries are.

[00:36:30] Ronnie Chatterji: Well, Dan Björkegren, thank you so much for an amazing conversation. Your research has really inspired a lot of us at OpenAI. I'm excited that you'll be working on this work as part of the exchange program as well.

[00:36:41] Ronnie Chatterji: I also just want to thank the audience for great questions. I'm sorry we couldn't get to all of them, but it was great to answer the ones that we could. I've learned a lot from this and in addition to thanking Dan I just want to thank everyone who's sent in suggestions and comments for topics we should cover.

[00:36:59] Ronnie Chatterji: This is one that there's a lot of internal and external interest in. I think when you think about how you can ensure that AGI benefits all humanity, our mission, it's the work that Dan is doing that's really going to facilitate that, and so thanks to Dan for speaking with us today.

[00:37:13] Ronnie Chatterji: Thanks to the community that's been following us. Don't forget to register for upcoming OpenAI Forum events that you don't want to miss coming in August.

[00:37:21] Ronnie Chatterji: We'll put some links in the chat as well as to Dan's work and the Exchange winners from this month as well. Thank you, everyone, and see you next time.