C (3:34)

Yeah, look, I think from our perspective, again, AI is one of the most important technologies we've ever built. It is incredibly powerful and we already see the amount of economic productivity it can unlock for software engineers, how much economic productivity it can unlock for knowledge workers who speak English in the, in Global north context. And I think I worry deeply about lopsided economic productivity. Like what if AI just makes the already economically productive even more productive like today? AI models do not generate the same level of economic productivity benefits for the average farmer, for the average healthcare worker, for the average lawyer, for the average, you know, citizen of India or anyone in the global South. Right. And I think that I, I worry that we need to quickly start thinking about how do these models not just work in the context of our community. So if I'm a farmer and I ask a question about crop, you know, for crop guidance, I get an answer that is locally contextual. That makes, makes sense. But how do we go beyond that? Like I remember when I was, you know, a student in, in high school, I would spend like three to four hours a day going to scholarships in India.org and I would just scroll through the website to see which scholarships I would qualify for. But I imagine an AI workflow where I can upload my mark sheet, I can upload my economic status if I come from a below poverty line family or whatever government scheme I'm a beneficiary for. And the AI model just goes through the website, finds which scholarships I qualify for, fills the PDF forms on my behalf and applies for me, saving me hours of effort. We're not fully there yet. These workflows exist in Global north context, but don't exist for our communities. So I think for me the biggest worry is we get the Eureka AI moment. It creates incredible productivity gains, but they just never reach the communities that they should reach and just that it isn't equitable. And I think that's what we need to fix.

A (5:30)

You've described walking into a data work facility early in your career and seeing workers who were paid a fraction of what their labor was worth. And realizing that you were just talking about the potential uneven impacts of AI. But at this moment in time, AI's value chain is deeply uneven. So tell us a little bit more about that and how that moment and the moment since have led to what you're doing now with Karya.

C (5:57)

I'm happy to. I think I want to expand what we consider as the AI value chain because so much of the conversation on the AI value chain is about data collection, which is of course very, very important. But if you think of the AI value chain as a circle, of course it starts with data collection, then you have model building, you have fine tuning evaluations. And our core belief at KA is that our communities that we seek to serve and global communities all across the world aren't just excellent beneficiaries of the AI revolution, they're excellent builders. They're excellent evaluators of the global AI revolution. So when we think of centering our communities in the AI revol revolution, yes, it's extremely critical that data collection happen in a way that isn't extractive, that it happens in a way that is as dignified as it possibly can be. But it isn't just data collection. Right? It is about like who gets to build the models, who gets to fine tune the models, what are our models fine tuned on, who gets to evaluate these models, who gets to tell us that these models are performing well or that they are not performing well. I think those are extremely important considerations to keep in mind. We did this project recently working with the government of Maharashtra here in India. There's a lovely district called Nandarbar that is led by a rock star district collector. Her name is Dr. Mitali and she came to us, she has incredible tribal languages spoken to her in her district. And there are no models or technologies that existed in this language. So we worked with her, we worked with the communities in this district in a project funded by the government where the data collection happened in the district, where the data sets are owned by the people of the district. Then the model building happened in the district, the fine tuning of the model happened in the district. We build these specific technologies for healthcare, for all of these different applications we're looking at. And the work is ongoing. But then the evaluation of this technology is also going to happen in the district. And I think you have. But almost like, I think we can almost redesign a new AI economy, one where our communities aren't just centered as beneficiaries, but as active builders at the very core. I think to me that's very exciting. Because an economy like that would be fundamentally more just and inclusive. It is exciting. So that's how we think about our work. And of course, yes, it starts with data collection. It starts with making sure that we are working with communities in a manner that's fair. We're working with communities for them to make sure that their languages exist online in a matter that's dignified. We had to work with amazing people like Alex and the OpenAI team. We, in fact, just did this thing called the nonprofit jam, which I know Alex will speak more in detail of the insights about. But we invited, you know, hundreds of nonprofits across four major cities in India, and it was almost like a tour of the country. And it was like six days, Alex, if I remember correctly, and four, four cities I got sick. But it was, it was incredible. It was, it was just really, really. It was so heartening to see civil society organizations, which again, are representation, representatives of our incredible communities, think deeply about the AI value chain for their communities. And how do we, how do we not repeat the mistakes of the Internet revolution? Right? How do we actually redesign this economy from the ground up?

A (9:22)

Alex, there's a lot for you to pick up on there. I want to hear more about your work with karya. I want to hear more about the nonprofit jam. But first, I actually want to hear more about your background, just like we heard from Manu about kind of the origin story of ka. Part of why I've been so excited to bring you into this podcast series is you come from the global development community. You worked at GiveDirectly and helped to set up a lot of their really interesting, I think, pioneering AI work. So can you tell us a little bit more about your background and how that led you to OpenAI and what you see as the opportunity for OpenAI in this vision Manu laid out in terms of communities being not just beneficiaries, but builders.

B (10:08)

Yeah, absolutely. So, as you mentioned, I spent a decade in the social sector before OpenAI, starting out at organizations like Innovations for Poverty Action, focusing on evidence based research and how to use technology to improve evidence based outcomes, but not for technology's sake itself. After several years in that research space, I personally was really interested in less in the research and more in the doing, the implementation, the partnerships being on the ground. And that's what brought me to giftdirectly, which I thought strikes a really great balance between actually operating in the real world and producing great impact for people, but also making sure that they're generating research and helping build the Field as they go. And so while I was there, I led a series of projects that were using machine learning for disaster response. And these varied in how they operated, but they did things like analyze cell phone metadata or satellite imagery or water levels in a river to predict who might be most vulnerable for a disaster, for anticipatory programs or after a disaster, to find the people that might have been most affected by that disaster. I had a particularly formative moment during one of those projects in Togo. And this was in the peak of the pandemic. And we were working with the government of Togo, which itself is a really pioneering government in terms of the way it's thinking about technology. And we were using cell phone data to try to find the people who were most vulnerable in terms of poverty level and income during the pandemic so that we could figure out how to get cash to them to get them through that tough period. I just remember being in the room with the billiard minister, Sina Lawson, and hearing her talk a little bit about how this technology can't just be for the rich countries. And we were talking a little bit about the trade offs. Always in these programs there are trade offs. And she was saying, these are trade offs that we have to make, that we have to own. And we're lucky here to have organizations like GiveDirectly and the Research lab at Berkeley that are helping us learn how to use this technology. But ultimately the decision about these trade offs lies with us. I just thought her clear eyedness about this was really inspiring. And that's actually what motivated me to kind of stay in this lane of AI and global development. And so that's actually what brought me to OpenAI. I thought, I've spent five or so years at GiftDirectly. I've learned an incredible amount working all over the world. But I'd love to kind of go back to Silicon Valley where I'd spent some years, and actually work in one of these frontier labs to bring that perspective. That perspective, which I think is relatively uncommon in terms of the folks that are being hired there. That's kind of how I ended up back at OpenAI.

A (13:24)

So I mentioned earlier some of the work that you're doing with OpenAI Academy and AI literacy programs. Can you tell me a little bit more about that work? And also you mentioned that you bring this perspective from your experience in global development, how that perspective helped to shape these programs. And it is continuing to shape these programs. I know they're always evolving.

B (13:45)

Yeah, absolutely. So OpenAI Academy has the Goal of helping people learn how to use AI. And that ranges from kind of everyday people just learning how to use AI for kind of regular productivity use cases, to governments and civil servants, teachers, students, but also builders and people that might be able to take advantage of these new systems and build them into products that have some sort of social benefit. So it's not connected to our sales work. It's part of our kind of policy and community outreach programming. One of the first projects that we did after I joined OpenAI was this AI for global development accelerator in collaboration with the center for Global Development and the Agency Fund. And I think that this program actually is a great kind of embodiment of a lot of what Manu's talking about, which is really focused on finding who are the organizations operating in the Global south that have a commitment to evidence, that have relationships with the government, that have an understanding of how to scale, and are either already using AI or kind of dipping their toes in or interested in using AI to improve their programming, whether that means being more cost effective or reaching more people. And so that was one of the first things we set up. And I think the way that works is that OpenAI helped with the selection process. We brought technical experts to assess the AI feasibility of the projects. We're providing API credits for the builders, and we provide technical expertise as well. And then on the other side, the Agency Fund is providing significant amount of funding and technical resources, and the center for Global Development is providing a lot of kind of research oriented perspective to help make sure that this program is not just helping the eight organizations that are in it, but is actually building public goods in terms of ways that we can measure these programs. And they've done a lot of interesting work there. So I think that's a great example of kind of how I was able to bring that perspective from GIP directly, to help tie it to this broader AI literacy goal and just kind of make sure that we're doing what we can to enable some of the actors in the field, whether it's government, nonprofits, social enterprises, or philanthropic foundations.

A (16:15)

Kaaria was not part of that initial group, is that right? But actually, that brings me to a question for you, Manu, and I see you nodding your head no for those who are listening and not watching on video. So you were not part of that group. But when Alex talks about the need to generate broader benefits from these kinds of initiatives, I'd love to hear from you, Manu. Why something like this? This AI for Global Development accelerator is the kind of partnership needed you know, what value did you see from that? Or, you know, to broaden the question, how do you think companies, groups like OpenAI, can be most useful in supporting the kind of work you're trying to do at Karya?

C (17:00)

Great question. Right? So I think the very first milestone that we can win in our fight against information asymmetry, which remains a big problem in the Global south, is by partnering with organizations like OpenAI that have incredible reach. I don't have the latest numbers, but, Alex, I believe ChatGPT has over 100 million active users in India. Last I checked, it might be much more than that. Like, OpenAI is growing very fast, but that is an incredible base of people. And I think these companies are. These labs are excited to provide value to more people across the Global South. And there is a synergy I see in between them providing value to our communities and us making sure that these models work well to solve genuine needs that they have. Right. So for social enterprises, for nonprofits, I think you can use these incredible models that are just, you know, they're just magical. And the kind of things that, you know, again, during the nonprofit jam, the kind of like, no code solutions, like, you don't even, like, need to learn how to code anymore to be able to use these technologies to create custom GPTs, to deploy something that you can bring to your community. Today evening, had you started with the idea, I don't know, 10 minutes ago, it's that fast. Now you're like, your speed of execution is just incredible. And again, nobody knows our communities better than our own communities and the people who serve them who were there. And I think one of the things that we partner with labs across the world, including all these BJ labs, is thinking about, where are these models serving our communities and where are they falling short? And then actually sharing that information with the lab for they can improve their models. Right. We internally call it the broccoli kiwi models.

D (18:53)

Right.

C (18:53)

These models are getting very, very good at solving all of these problems. But if you ask these models, like, hey, I'm in Karnataka, which is the state I'm in right now, I want to eat healthy. What food should I eat? It says, you know, perfect Kannada. You should have lots of broccoli and lots of kiwi. Right? It isn't wrong. And I'm sure broccoli and kiwi are accessible to many people in the state, just not to every person and just not to, like, you know, if you're in other communities in the state, it'll be harder to find broccoli and kiwi. It's not locally relevant. Right. You've had cases where, let's say, models, when these models are asked, like when you say, oh, I have a fever, what medicine should I take? Some of these models will say, oh, you should take paracetamol or Tylenol. Tylenol doesn't exist in India, just called paracetamol. But it cites Mayo Clinic as a source of, you know, like, as a trusted source. Now, the three of us, even though I've never been to Mayo Clinic, I know that it's a trustworthy source, but we had one of our community members says, I don't know what this website is, so how do I trust it? Right? Because we were looking at how do we make sure that these answers are trustworthy or what is trustworthy to a community member? And I think social enterprises and civil society organizations can actually partner with these labs because it is in our best interest to make sure that model responses actually meet our communities where they are and they're useful to our communities. And it's also in the lab's best interest about the incredible orgs that are in the accelerator, including Rocket Learning. Alex. Right. Rocket Learning was there and just one of the favorite organizations in the world. There's just so many amazing organizations in that group. And I think all of them were able to use these incredible models to launch truly meaningful digital interventions in their communities. And I think that's an excellent step one, because then you're taking a dent on that information asymmetry problem, which, again, like, I'm a complete nerd. Like, you know, like, like Windsor used to talk about how the Internet would end information asymmetry. And. And it did. And it didn't. Right. And of course it didn't in a major way.

D (20:53)

Right.

C (20:53)

We still have, like, last I checked, billions of people, over 2.9 billion people that still don't have access to the Internet. Right. And I think we have a long way to go. But this dream of can I get people the information that they need when they need it at a really cheap cost in their languages, in their contexts, that makes sense to them is really, really powerful. Right? And I think that's step one, and I think that's a great place for civil society or social enterprises like Kara to partner with labs like OpenAI. And then on top of that, the economic productivity piece I was talking about, because step one is correct answer. Step two is workflows, making sure that you're actually making communities more productive so they can. I have like five, like Model instances running right now that are doing my job while I'm talking to you, I want that to happen for the average farmer in the country, the average healthcare. Like, you know, we have ASHA workers in India who like, you know, knock doors and, and give health care guidance. They're just so incredible and so intelligent and they have decades of experience. But a lot of this is digitized already, thanks to the work the government of India is doing. And I think of like, hey, can I just, like, if I build the right AI product, can I make them do 10 homes in an hour instead of, say seven homes, which means they'll get paid more by the hour, which means they can stop the work two hours earlier and that gives them two hours back. And what does that lead to? And these are interesting research questions, and I think that I'm excited about laps and enterprise Social enterprises working together to actually find answers to some of these questions. And it's an exciting collaboration.

A (22:26)

It's really interesting to hear you mention that, Manu, because I feel like often when we hear about large language models and jobs in low and middle income countries, it's a scary scenario. That's that, you know, and there are some scary scenarios, but the, the, the gains in terms of, for example, being able to be more productive and have more time, you don't often hear that narrative. So that's, I think that's really interesting. I don't know if either of you wanted to jump in on that. Do you feel like that's something people are kind of missing?

B (23:00)

That's actually kind of a great segue to talk a little bit about the nonprofit jam that Manu and I worked on together, because one of the things I found when I was there, we had over 200 nonprofit leaders across four different cities. And these, unlike the organizations that are part of the accelerator, generally, are pretty new to AI, so they might not even have IT teams. They might struggle with some basic digital literacy, not to mention AI literacy. But I found that one of the things that resonated with them most is that idea of time back. Every nonprofit kind of by definition is operating under resource constraints. And these leaders, it's so inspiring to talk to them, but they're sacrificing a lot. They might be away from home in order to be serving some of the most vulnerable people in their communities. And I think anything that can make their lives a little bit easier, give them a little bit more time, I think is really valuable. So, for example, sometimes some of the use cases that seem to resonate most had Nothing to do with work, but were about cooking or things around the home. And I think often you see light bulbs turn off when you show some of those kind of personal life use cases of AI and they say, oh, if it can save me time that way, then here's an idea of how I might be able to use it at work. So I thought that was particularly inspiring as we met these nonprofits, but it

C (24:36)

was just so heartening to see people bring the realities of their communities. The fact that if you pick two Indians at random chance that they speak a different language is 92% that shows up in the communities we seek to serve. So I cannot build a single product that only speaks one language. So while GPT might be great in like 10 languages, if it's not good in the other two, I need to make sure it's good in the other two because I have to serve all 12. And I think that that is something OpenAI obviously knows, and they're making such incredible advancements in that. But that showed up as things that all these models have to improve on, which is exciting to see.

B (25:10)

Yeah, if I can just add on to that quickly. I think one of the things that's great about working at OpenAI right now is we're actually still enough to take the kind of feedback from events like the nonprofit JAM directly to the engineers that are building our products and models. So we're in the room and we're hearing this feedback of, oh, it's performing better in this language than that language, or it's okay in this language, but it sounds like it's from Bangladesh instead of from India. And we can take that feedback and give it to our engineers, and it's something that they will have as they think about building the next models. I think in general, the kind of number of players in the room from the nonprofit jam, having Karya that plays this kind of important role of convening, of having these trusted relationships with all these nonprofits, but also this deep expertise of building in the context of India. Having organizations like Wadhwani AI that had a lot of technical experience and can serve as mentors, and then having OpenAI that can bring that kind of local context back to HQ really was kind of a magical combination.

A (26:23)

Earlier, Manu, you mentioned how ChatGPT is being so widely used in India and across the world and growing rapidly every day, and you've referenced a few times lessons we can bring from the Internet revolution. To me, I'm reminded also of the growth of Facebook now meta. And, you know, I Followed Facebook's growth very closely and was in touch with many nonprofits who pretty much had to be on messenger in order to meet people where they were. And you know, there were some benefits to that. There were also some tensions there. For example, organizations that had built their own, you know, chat systems or what have you, and suddenly everything had to be on Facebook. So I'm just curious to hear from each of you and maybe we'll start with Manu. How does that digital ubiquity factor kind of change the relationship between social enterprises and AI labs? And what do you see as exciting opportunities that it presents, but also maybe some challenges to work through?

C (27:25)

I think from the opportunities perspective, right. If we are able to make sure that these models meet our communities where they are, that they're shaped by our communities, I think we are looking at something very, very exciting. Right. I am very inspired by the kind of work that Mentor me did in Namibia.

D (27:46)

Right.

C (27:46)

Which shows that just using a GPT enabled entrepreneurship life coach actually has a meaningful increase in the earnings that an entrepreneur can make. And I think fine tuned models that understand the community's realities that, that actually solve these problems also just benefit from the inherent network effects and distribution advantages that these companies have. And I think there is unique skill sets that we as civil society organizations bring in. And there are unique benefits, in complete honesty that OpenAI has or other model providers have. And I don't see a challenge there. To be very honest with you. I'm almost disappointed I don't have. I have other challenges I could talk to you about. But on the point of digital ubiquity or these models becoming like default gateways to the Internet, of course that can create some challenges should these companies not meet our communities where they are. And I think there's a lot of competition, as Alex could tell you in this space. I'll be surprised if that happens. But I mean in India you see incredible sovereign providers. We just wrapped up a summit with like where the government of India showcased three sovereign AI models that are exciting to see. You see companies like OpenAI invest deeply in Indic language. So I think it's a very vibrant space, which makes me happy because I think that gives social enterprises the chance to work with multiple players and actually make sure that we are ultimately holding our community and their interests as the single most important, like guiding light.

A (29:26)

Alex would love to hear you comment directly on just the scale we've seen with ChatGPT and what that means in terms of the kinds of relationships you can strike with organizations. Like Karya, both in terms of again, like exciting opportunities, but challenges or tensions that need to be navigated.

B (29:45)

Yeah, absolutely. I think just to elaborate a little bit on this idea of digital ubiquity, I think it's definitely true that AI is becoming core infrastructure, but I do think there's a couple of really important core differences between AI and these earlier digital platforms. One is this, that you guys were just hitting on this competitiveness. The ecosystem is incredibly competitive and organizations can choose to work on our models, but Also our competitors, CloudGemini or open source models, and not only do they have that choice, but it creates pressure on all of these companies to be building for those users. And I think that's a really healthy dynamic that will be the benefit of those end users and those builders. I think the second big distinction is that it's not just kind of a passive distribution channel, the way messenger is, but it's giving a new kind of set of capabilities to builders. So its building capacity, building capability that social enterprises can then leverage to grow and to shape their own impact. So in many ways they're in the driver's seat with deciding how to use these models, whether to use these models, which ones to use, and actually building kind of real world applications that are at the benefit of the organization of the communities that they serve. I think connecting to this idea of scale, I think this is something that because of our scale, we're particularly lucky to see. So when we're anywhere in the world, we're meeting people from builders to everyday people who are using our models and telling us a little bit about how they're benefiting them, but also the challenges or issues they might have. So I think that's a really exciting place for us to be.

A (31:46)

Part of what's refreshing about this conversation is we're getting a little bit more specific about large language models, not just talking AI in this broad sense, which I think sometimes people can't wrap their head around, specifically when it comes to the explosive growth of ChatGPT, of large language models in order to ensure a future. Back to what Manu was describing earlier, where these communities we were talking about are not just beneficiaries, but builders. What needs to happen in terms of this relationship between social enterprises, AI labs and the broader global development community? What do you hope to see?

C (32:22)

I'll give you a quick example of something I know that OpenAI is working on in India that we are delighted to support. There is a app called Mahabistar which the government of Maharashtra has launched and it is an app for over 13 million farmers across the state of Maharashtra. And I think OpenAI supports the creation of these technologies. It is built on top of GPT and Karas communities are involved in fine tuning the models and making the models better and evaluating these systems and bringing them to languages where GPT may not exist yet. And I think I see that as a great way for us to collaborate with labs, with governments that have a social responsibility and a moral mandate to actually serve communities across the country. And I'm very, very hopeful, genuinely very hopeful, not just saying this because Alex is here, but genuinely very hopeful about what can happen in that collaboration between social enterprises, civil society organizations, governments and AI labs. Like when we started our work on community evaluations, we had a very senior client, not from OpenAI, I promise you, who said, you know, evals are done by experts. Your communities are not experts. Right. And they would actually call them non expert evaluations. Right. And I was like, but if I'm doing agriculture evaluations, who is an expert but the farmer who's going to use this model? Right? And who is an expert but the agri extension worker that Digital Green has trained to do this work, who is an expert but the government trained agronomist who is an expert like, like these people, our communities are experts on their lives, on their community, on what is quite literally their bread and butter. And I think that it is really important to us that we just. It sounds so obvious but we really need to think of communities that all of us serve as not just passive beneficiaries. And in this specific case of agriculture, we have now done over half a million expert evaluations in the last six months. And we've actually shown what are places where these models actually are excellent at providing advice and what are places where they falter. And it has led to benchmarks that model builders are hill climbing against to make their models better. Right. And I think to me, I see that as you know, this project is called Samiksha. It is now funded by incredible partners, Dr. Sunera Sitaram and Dr. Kalika Bali from Microsoft Research and funded separately by Anthropic on the agri space. And I think to me it's really, really important that we do more of this work. We open sourcing everything so that everyone can benefit from it. And I think that to me is a great collaboration between model companies, research scientists. Governments are using these benchmarks now and civil society organizations being a mirror on board these technologies mean to our communities and being very honest and like, you know, these models are really powerful.

A (35:30)

Manu Alex, Want to thank you both so much for your time and looking forward to following your work in this space.

E (35:45)

Hi, I'm Kate Warren, Executive Vice President and Executive Editor at devex. At devex, we don't just cover the biggest moments in global development. We create space to understand who and what are driving the headlines. Alongside gatherings like the World bank and IMF Spring and Annual meetings, the World Health assembly, the UN General assembly and beyond, we host DEVEX Impact House, where our journalism comes off the page and onto the stage. We bring together a curated group of leaders for live interviews, intimate roundtables, hands on workshops, and candid conversations you won't hear in the official meetings. It's where tough questions get asked, the spin gets stripped away and meaningful connections happen. If you'd like to join us or stay in the loop on all of our events online and in person, please visit devex.com events so what we just

A (36:38)

heard in that exchange is both the ambition and the complexity of this moment. So on one level, these partnerships are about making AI more useful in practice and responsive to the realities of the communities they aim to serve. But behind all of this, there's this question of what it will really take to make sure that AI supports development in ways that are equitable and accountable, and to help us think more broadly about what these kinds of partnerships mean for the future of global development. I'm joined now by Hans Zheng Chia, Director of the AI for Global Development Initiative at the center for Global Development. Hans, welcome. Thanks for having me. You and I had a conversation where you introduced this term of digital ubiquity. And as I understand it, the point you were making is when you have tools like ChatGPT that are becoming so widely used, social enterprises and NGOs really have no choice but to engage with those tools in order to stay relevant in the communities they serve. So how does it change the landscape for the development sector and its relationship with AI labs?

D (37:47)

You know, I think firstly we're starting to see use cases of AI that are both incredibly beneficial, but we're also seeing use cases that are harmful. So let's take education for example. There was a study in ghana of a WhatsApp of an AI tutor delivered over WhatsApp that improved learning gains equivalent to one to two years of schooling and a fraction of traditional costs. So this is incredibly beneficial. And we're starting to see similar results from studies in the uk, in the US and across all our middle income countries as well. But at the same time, we also know that AI products can harm learning, especially if they don't have pedagogical safeguards built into them. So these are AI tools that give answers away to students, that overload them of information and don't balance the cognitive load for the learner. And so the debate is no longer about can AI deliver benefits or harms. It can clearly do both. The question is, which version will reach greater scale? And I think for a lot of nonprofits and government institutions, if they built a pedagogically sound, really beneficial AI tutor that reaches a million kids, that's a big win. But at the same time, we should be cognizant of the fact that Meta, for example, has 3.5 billion daily active users across all of its products, WhatsApp, Instagram and Facebook. And students today can just give a WhatsApp chatbot or put it even into their Instagram search browser their entire homework for the week and say, just give me the answers please, and completely devolve any form of self directed learning and really use AI as a crutch. And so the question to me here is, wow, how do we compete with that? You use the term digital ubiquity, right? The fact that existingly, before governments can even deploy its pedagogically sound AI tutor, we already have sense of wide skill use that could harm learning outcomes.

A (40:23)

So what kinds of partnerships do you see as really urgent and essential between, you know, the companies designing and deploying these tools and the organizations working in these communities?

D (40:36)

So I think there's two. I've grouped the potential partnerships into two buckets. The first is what I think the traditional tech for good sector has been quite strong at, which is this idea that the frontier technology companies develop powerful technology that then public institutions should adopt. And so famously, when healthcare.gov was launched under the Obama administration, it crashed very early on. And then there was a tech surge where the administration said, I need the top engineers from Google and the major tech companies to come and stabilize the product. And so this is an example of technology transfer from tech companies to the public institutions. But I've been increasingly thinking about this idea of reverse flow. So how can we take the knowledge, the data, the contextual know how from the public sector and systematically influence the technology products of these major tech companies? That again could potentially have much larger reach than government institutions and nonprofits. And I think we're talking about AI now. But this isn't a new problem we've had. Public institutions have long struggled to keep pace with the influence of technology companies. We saw during the COVID pandemic. But then suddenly with the advent of social media, you have multiple sources of knowledge competing with those public health institutions. And so yes, it's about making our public institutions stronger, but it is also about how can we influence the major technology companies.

A (42:34)

As I hear you describe these different kinds of partnerships, I think about some of these direct to consumer apps and social enterprises focused on direct to consumer. So one example that comes to mind is Digital Green. What is the kind of relationship that an organization like Digital Green might have with an organization like OpenAI?

D (42:57)

The tech companies, OpenAI, google.org, right, can be providing, and I think they are providing your frontier models to Digital Green. They could provide fellows, engineers to embed a digital Green. They can provide concessionary credits. There are lots of these programs, I think by all the major labs to social enterprises. So that's one way that's the traditional flow of tech resources to these social enterprises. But then let's reverse that flow and say digital Greens on the ground in Ethiopia, and an Ethiopian farmer is trying to understand how they can get credit for the planting season. And we may think in the global north of getting credit from a formal financial institution, but perhaps for that farmer, local networks are actually a better source of financing. Digital Green with its employees on the ground there may realize that that's contextual knowledge that's more relevant to the farmer. Perhaps the formal institutions are more predatory in the global North. We may think, hey, formal institutions are well regulated. If you go off to informal networks, you could get a loan shark, somebody who takes advantage of you. But the reverse could be true in a different context. And so Digital Green has that unique insight through its interaction with farmers. And so what if that information and that unique piece of insight could flow back into the major tech companies such that when the foundational models provided by OpenAI or Gemini or Anthropic is used by a farmer in Ethiopia, or there are other apps that are built on top of those foundational models beyond digital greed, 50 other digital agronomy apps that built on top of that. It benefits from that same underlying knowledge, the same underlying mental model of what locally contextualized credit opportunities looks like.

A (45:15)

I still find that a lot of the vehicles for work in the social impact space are in that traditional space of grant making, corporate social responsibility. Are you starting to see a shift where AI labs and tech companies are realizing the value they're gaining from these organizations and maybe changing the way they interact with them to reflect that?

D (45:37)

I think we're starting to see Quite a bit of it in the education space, notably from Google, anthropic and and OpenAI, they've all released learn modes that are supposed to incorporate some of these pedagogical best practices of withholding the answer, managing cognitive load, trying to drive the student to have some friction in their learning process, which is very good for learning outcomes. And so you see them including teachers and educators in the development of this product. You see them trying to get pedagogical standards and best practices and building these data sets. So I'm quite excited about that. But coming back to this question of digital ubiquity, we can see the reverse flow in sectors like education that I just mentioned. But I think a question for society to grapple with is should learn mode that exhibits these pedagogical practices be switched on by default? Right now it's still something that has to be opted into or should. Now many of these companies are releasing parental controls then right now allow parents to steer kids away from graphic content, for example, but should it allow parents to switch on Learn mode and default only to learn mode? So we're starting to see reverse slow from the social sector into these major tech companies. But then should we go further and build more features that allow parental controls?

A (47:43)

For example, you formerly worked with GiveDirectly, which I see as a really pioneering organization in terms of its use of AI. Alex, who we just heard from, also formerly worked with GiveDirectly and is now working within OpenAI when it comes to who should be steering what's next for the relationship between social enterprises, NGOs and on the one hand, and AI labs, big tech on the other hand. I see a really interesting career opportunity, really for people from the realm of global development who can kind of speak both languages as you do and as Alex does. So can you speak to that a little bit? And it builds on your point about Learn mode being switched on by default as we figure out how to scale AI that has beneficial social impact. As you mentioned at the outset of our conversation, what new opportunities does that open in terms of roles for people who can lead us in the right direction?

D (48:45)

So I'll touch quickly on positions that Alex myself. I think Drew Bent at Anthropic also came from the education sector and is now in hyperscale. I think that's great that tech companies are incorporating folks with public interest expertise. I think what's even more exciting would be to expand the work that Maru has been doing. So expertise, the kind of expertise that these foundational models need is so, so, so much larger than what individual quote unquote, experts can contribute.

B (49:33)

Right.

D (49:34)

So that understanding of local financial products for the Ethiopian planting season, that is knowledge that the farmer knows one is it refer that financial product and its terms in the local dialect. It's stuff no expert's going to know. It's stuff that a farmer knows. And so this relationship between not just expertise or knowledge, but the relationship between that and the local context and local language, that's just something you need from the local community. And so I think we should be thinking of. I loved Manu's point about hey, this isn't about what Karrio is trying to build is not just about data collection or data extraction from the local communities. It's about ensuring that they are part of the reversal. They are the builders and validators.

C (50:37)

Right.

D (50:37)

And again I, I think this is fundamentally different from how perhaps LLMs first got started, which is based off the Internet, which is based off large central pools of knowledge that could get hoovered up by the major tech firms. Now we need a new model that's actually requiring federated data collection by federated validation, federated model building. That I'm hoping, I think as Manu hoax creates business opportunities for many and non middle income countries, what needs to

A (51:13)

happen to create more models like we see with Karya and to ensure that communities are builders.

D (51:19)

Organizations like KA are fantastic. I want to make sure that they emerge not because there's a brilliant founder like manufacturing, you know, and not just because she's a brilliant founder like Manu, but because this is the systematic there as an ecosystem we have a kind of systematic push for more Karyas. And part of my worry is that. We have lots of this talk about how we need to be inclusive, but that actually it might be very costly for the hyperscalers to really pursue this at a tremendous scale, or at least to do it at a scale that's somewhat equivalent to the level of data ingestion, benchmarking, et cetera that they're doing for high income use cases. Right. So you know, for example, the tech sector, anybody who codes is raving about COD code, for example, right. And how it's revolutionized programming. What if, you know, what if the investment in low and middle income country use cases was it was half of what hasken invested in these high income use cases like coding, right. It would be many, many, many more times what is being invested in Icario. And so, and so my hope is that, you know, and I don't think that if we left it solely to the business interest of the major tech companies that we would come anywhere close to it, because by definition many of these use cases are not profitable, they're market failures. And so I think I'm hoping that there is a systematic mapping and this is some work that we're hoping to pursue at the center for Global Development, which is a systematic understanding of the domains that we think should be prioritized, You know, for, for, for the major tech companies to invest in that would otherwise not be invested in based off their business incentive. And then to say, well, should CSR expand to fund more of these reverse flow domains? Should private philanthropy step in or should just more corporate profits be used for these domains? I think that would be tremendously important.

A (54:17)

That's fascinating and I'm excited to follow the mapping work that you do and I hope to see real outcomes result from it once you paint a clear picture of what impact could look like. Hans, thank you so much for taking the time to join us.

D (54:30)

Thank you so much for having me.

A (54:32)

What I take away from these conversations with Manu, Alex and Hans and is as tools like ChatGPT become more widely used, the question for the development sector isn't just whether to engage, but how. And I found this to be a really useful conversation, unpacking what that might look like, the opportunities, the challenges, and also what needs to change. That's all for this episode of Global Progress in the AI Era. I'm Katherine Chaney. Stay tuned to DEVEX for our continued coverage of AI and let us know what angles you think we should be exploring until next time.