Pondering AI

KIMBERLY NEVALA: Welcome to Pondering AI. I'm your host, Kimberly Nevala.

In this episode, I'm beyond excited to bring you Hiwot Tesfaye. Hiwot is a technical advisor in Microsoft's Office of Responsible AI. She also serves as a Loomis council member at the Stimson Center where she helped launch their global perspectives in RAI fellowship. Today, we're going to be talking about global inclusivity and participation in AI. So welcome to the show, Hiwot.

HIWOT TESFAYE: Thank you so much for having me, Kimberly. It's really fun to be on this podcast. I listen to a lot of your episodes.

KIMBERLY NEVALA: Oh, thank you so much. I have been very humbled by the guests that do, in fact, pick up our call. I also have had the personal pleasure of working with you in the past. So I know that your personal experiences have very much inspired some of your work today.

But for those that have not had the pleasure of working with you and are not yet familiar with your work, can you tell us a little bit about how your personal interests have intertwined with your professional work? And has that been a happy coincidence, a deliberate plan, or a little bit of both?

HIWOT TESFAYE: Thank you, Kimberly. And a little bit of history about Kimberly and I, when I first got into this whole AI ethics space, if it wasn't for Kimberly, I think this whole governance thing would have been all too scary for me to even dip my toes in that water. So you've been an incredible partner in this journey of trying to advance responsible AI, Kimberly. So it's really great to be connecting with you in this capacity this time around.

But yes, to answer your question, I would say happy coincidence. Nothing about my career or life, generally, has played out as I have planned. And thank goodness for that because my plans are-- well, reality has outpaced and outperformed my plan. So I'm just incredibly grateful that things didn't quite go the way I've planned. I never thought I would be in tech, let alone in the AI space.

My undergrad was in nutritional sciences and economics, so far from the tech space. I didn't have your traditional path to tech experience with a CS undergrad, a CS master's or PhD. So happy coincidence for sure in terms of getting into tech and then having my personal interests and work converge.

That part, I would say, is a little bit more deliberate than just a happy coincidence. I've tried my best throughout my life, earlier life, before college and then into college and then into my professional experience, to tap into what gives me energy. And that's often the things that I'm just really curious about or really interested in from a personal experience that allows me to sustain and push. And, I guess, not be dismayed by closed doors or people telling me this is not possible or it's not the right time. That fire that keeps me going in some topic areas is usually when I can find that intersection between my day job and things that I'm personally interested in.

KIMBERLY NEVALA: And I will say, as I said, we did work together. And I think you have given me far too much credit there because you are both patient and you do persevere. But you have a really - actually, I guess, maybe just as a play on words, it'll sound a bit like a pun here - but a way of being very inclusive in your own work. You say that you don't let closed doors hold you back. But I think it's probably more than that. In that you ask a lot of questions to figure out why they might be closed in the first place and perhaps find a way around them with other people's permissions. And I think that's a lovely attribute and skill that we could all, myself, perhaps most of all, learn quite a lot from.

So let's talk about this concept of inclusivity. Very often, what I see when people are discussing the idea of global inclusivity relative to AI - this is true in other domains as well, but AI today in particular - it is often framed primarily as a distribution problem. How do we make what we have developed in a few key central areas or from a few providers available to everybody globally? And I'm wondering, from your perspective, does this framing limit our thinking about both the challenges and the opportunities ahead for AI?

HIWOT TESFAYE: Before I get into the obvious, at least from my perspective, answer that, yes, I think it is limiting.

I think this term inclusivity could be as expansive or as narrow as whoever's talking about the topic. But framed from this perspective of a distribution of opportunities or distributions of resources, I think there's just a couple of underlying assumptions that we make when we frame it from that perspective.

The first is that there are two camps of people, societies at whatever altitude you want to talk about it, but there's two camps. And one is the camp that generates things of value. And there's another camp that waits around to be given those things of value. That's one underlying assumption, which honestly obviously needs to be dispelled. And it's a twisted way to think about the world.

And even when we think about AI and the global North, global South, global minority, global majority. These two camp ideas, there is a role that each of them play. Although, obviously, the most powerful generative AI systems that we're seeing today are being built out of the US and, in some cases, France and China for sure.

If you think about the spectrum of activities that takes place to build those AI systems, there's so many parts of the world that touch that supply chain, all the way from the minerals that are mined to make the chips, the chips that are made, the infrastructure that's needed to develop these giant data centers, all the way to the sourcing and labeling of data. Where does that work come from? It doesn't just happen in one society only. It's a global manufacturing ecosystem where there is a role to be played across these different - across these two camps, if you will.

So there's that aspect that's often ignored. We think that the thing of value is only made in one camp, and then the other camp just waits for handouts, which I think is just a distorted way to think about it. So, yeah, I think I prefer to participate in conversations around inclusivity that assume there is value and beauty, I guess, and unexpected innovation that can occur in any place and across many groups of people.

So I think it's just important to, at least for me, it's always important to stay curious and not make these wild assumptions about where value can be generated or where beauty can come from. I believe it can come from anywhere, really.

It's just a matter of do people have the right serendipitous opportunities? And are they in the right place at the right time to generate that value or to be seen by the people that can either elevate them to the next level or just acknowledge the work that they're doing?

So that's why I try to ground myself in discussions around inclusivity. It's less about a distribution thing for me and more about just can you open your eyes to see the innovation, the creativity, the value that's being generated in every corner of the world. The beauty that's created in every corner of the world.

If you're just willing to take the time to look from a different perspective or just experience and be in those spaces, you will see it. It's just a matter of do you have the time and the willingness and curiosity to do it?

KIMBERLY NEVALA: And that also then presumes or pushes us, perhaps, to think about folks - you mentioned data labeling or even the elemental extraction - where folks in areas where the models themselves are not being trained are viewed primarily as labor. Not necessarily, as you said, the source of inspiration or the owner of the innovation.

And I wonder, does that then also influence how we think about even just basic elements, like literacy and education globally? Because if we're putting folks into that kind of a limited bucket, it would seem to follow that we may also not be investing in some of these other areas that would allow them to, in fact, take part and leverage all of the beauty and the thoughts and the innovations that they have in their own areas - that likely would benefit not just them but benefit everyone globally.

HIWOT TESFAYE: Yeah, no, I completely agree with you on that.

In the US, there's been an increased devaluation of education generally. Teachers don't get paid as much as data scientists do, as an example. And there's a lot of conversation in the tech space right now about trying to accelerate AI adoption across various parts of our society, globally even.

And if we are serious about that, if we truly believe in the transformative power of this technology and we want it to diffuse across all industries, including in education and other sectors, it's just important to think about this in a more expansive way. And recognize that adoption or diffusion cannot take place if we don't consider all of these other camps that we generally tend to not value as much.

KIMBERLY NEVALA: What are the implications that we're seeing now or that you foresee if we proceed down the path of just thinking about importing or, I guess from where we're sitting here in the US, it would be exporting AI to the world?

HIWOT TESFAYE: Yeah this is actually one of the first topics that captured my curiosity when I was starting to think about what does it mean to build globally inclusive or globally relevant responsible AI practices?

This question about importing AI was the thing that was at the forefront of my mind when we first kicked off this fellowship, and we're trying to explore this with our fellows from across the global South.

I think the first thing is just around reliability. Does a system that's built in the US work in Kenya? And not even just across countries, but even within the state of North Carolina where I sit right now. Does a system that's built for a health care model that's built for UNC health care systems, as an example for the Raleigh area, work just as well for the coast or somewhere in the mountains or some other place in rural North Carolina?

It just begs the question of a system that's trained on data and trained by people who are from a particular place, ultimately, just they embody the values and priorities of those people. These systems are not neutral as much as we like to think they are. They take - and I know I'm preaching to the choir here, Kimberly - but I know they take on those same values and priorities as their trainers as well as the data that is chosen to be included in the training process.

So this is definitely from a global lens. It just begs that same question of will it even work well? Does it even perform as well? Or does it even solve the same problem we think it's going to solve in one geo versus another? But that same thing can apply within a country or within a smaller geographical space that we're thinking about. So that's the first issue that comes to mind.

But then beyond that, since I'm in the responsible AI space, of course I always think about what are the values and priorities that these various cultures have across borders? So do the same ethical priorities in the US translate similarly across borders to another part of the world?

And I guess, in my own exploration the last couple of years, I would say that it's not so wildly different to the point where we're like, we need to just scrap the ethics or the responsible AI practices that we've developed in one place and just start brand new in another place. I think there's some baseline level of common understandings that we have across cultures. And then there's nuances, of course, that need to be taken into consideration.

So that's the other aspect of this whole importing AI thing. Does it work well? Does it actually solve the problem we think it's going to solve in one place versus another? And then, in terms of safety and reliability, can we say the safeguards we've put in place for this particular cultural context actually translate well into another cultural context?

And then lastly, I should have probably talked about this first because this is really the reason that I was super concerned about importing AI, is a lot of writings that I've read about digital colonialism. This practice of extracting raw materials and resources and labor, taking it to a different place, processing it, and then selling it back to that same market for an astronomical price. So the same patterns that we have seen throughout the last several centuries about cacao beans or cotton for producing textiles.

We've seen these practices being played out across India and Africa, Latin America. Where these raw materials, and the labor that goes with it, is extracted at a very low cost and then processed elsewhere, sold back to those same markets at an astronomical cost. And we're seeing similar patterns playing out in the AI space. And so that's the other thing that captivates or captures dialogue around importing AI, especially when you're talking to people from those places. These are the topics that come up quite a bit.

KIMBERLY NEVALA: And I think even beyond that the economic extraction, as you're talking about there, it's not always clear, to me at least, that the problems we're trying to then solve and that the solutions we're selling back at astronomical prices actually solve the right problems in those areas either. Even though AI - AI, the royal AI, right, this whole suite and spectrum of technologies and techniques we have available to us - can be applied in many, many different ways. So don't let me forget to come back to that point because I think it is really, really important.

What's also then interesting, and you and I were joking when we were talking before this, is this idea of electricity. Or AI as global infrastructure and it's like electricity. And I said, dot, dot, dot, well known for 100 years, still not ubiquitous, right? So when people say this is the new oil, the new electricity, the implications are not perhaps as positive as they would like to see there. Because there's infrastructure and the ability to actually execute, implement, to make that available that doesn't exist. So we will circle back to that.

I know that, as you were talking, too, about just extracting the labor to be able to make these systems work and then selling them back. Very often when we sell them back, even when we sell them back in a context that is usable, so for instance, in a chatbot - or should be usable, I should say, like a large language model so we'll just use the current hot trend du jour - that they don't then actually reflect or work very well in some of these global contexts.

You talk specifically about low resource languages, and I think this is a really good exemplar for people to understand in a very concrete way why that use of labor doesn't actually then get paid back in value to some of these communities globally. Even though they are really intrinsic in the creation of the product itself. So perhaps maybe we can talk about what is a low resource language for folks that might not be available for that and how we see those showing up or not showing up in these very common LLMs foundational models today.

HIWOT TESFAYE: Yeah, I'm so glad you asked this question.

First of all, I want to say I don't want to pretend like I am a deep subject matter expert on multilingual AI systems. I have learned a lot from so many colleagues across Microsoft Research, especially those in India. Partners across the industry that I've had the opportunity to learn from. And it's been so enlightening the last year and a half or so that I've been able to absorb so much knowledge from much smarter people around me.

But yeah, so this low resource language, so just to answer that first question, what we mean by that. When we think about how these generative AI systems, large language models are built, they consume a lot of the web crawl data that's out there on the internet.

And when you think about, what kind of languages are represented online, It's predominantly English. I think I saw a statistic, like 52% of the websites on the internet are written in English. And then even when you think about another high resource language like Mandarin, billions of speakers that speak Mandarin as a first language, it comparatively, when you look at the representation of web data that is in Mandarin, it's a fraction in comparison to the English representation. Similarly, for Spanish, we have, again, a lot of Spanish speakers around the world as a first language. But again, the representation of Spanish websites online is much smaller than English.

So I think there's that aspect of it: how much of this language is available digitally? And relatedly, the number of people that are even out there who can create these digital websites or Reddit posts and Wikipedia pages in those languages could also be limited.

So my native tongue is, Amareñña, it's Amharic from Ethiopia. And we have over 100 million people that are from this country. Although, not all of them speak that language. There's actually 80 other languages that are spoken in Ethiopia. Amareñña is one of them, which is crazy. So it's a fraction of that over 100 million that speak this language. And so the representation of that language online is just really small.

So yes, we can categorize that language as a low resource language. We don't have a lot of speakers of that language, plus not a lot of content online that's readily available for these AI companies to leverage in their training or fine tuning of these large language models.

One last plug and then I'll stop answering this question. Specifically for Amareñña, although there isn't a lot of resources online for this language, the written content that's available that has been developed over centuries is a ton. So there's something to be said about some low resource languages. Yes, they have been written down, but have they been digitized is the other question? It's not always safe to assume that just because a language is low resource doesn't mean that there isn't a large volume of writing in that language. It's just maybe not digitally available for AI systems to be trained on.

KIMBERLY NEVALA: Yeah, I think that is an important clarification. Perhaps embarrassingly, I will admit, as you started talking there, I was thinking, but this is a spectrum. If someone had said what are examples of low resource languages, some of the Chinese languages, Mandarin and, again, a lot of different dialects and variations there or Spanish wouldn't have even occurred to me just. Because I think of that as the proportion as large if you look at how many actual speakers of that language there are in the world on that scale.

But I hadn't, as I said, somewhat embarrassingly, done the math I suppose to think about how much of that is digitized. So when we say low resource language here, it's not low resource necessarily because of the number of humans speaking of it nor the history of the language. Or how long it's actually even been present as a language and as a common language. We're really talking about how does this show up in digitized data sets that are driving these.

HIWOT TESFAYE: That's right. And honestly, I think Mandarin and Spanish are in a much better position than, say, Yoruba or Swahili or even the Nordic languages. There's not billions of speakers of those languages. So there's definitely a spectrum where English is at the forefront of it. And then you have Spanish and Mandarin shortly after French widely spoken around the world. And it just trickles down from there. So that gap still persists between English and a truly low resource language, like Icelandic or Yoruba or Amharic.

KIMBERLY NEVALA: Because your native language is one of these maybe what we would traditionally think about as a low resource language, what has your experience been with using these models or trying to speak in these models? And why can it sometimes be very disconcerting to do that or misleading?

HIWOT TESFAYE: I'm so glad you asked this question because in preparation for this podcast, I was like, let me go back and just make sure that some of my favorite apps that I use, generative AI apps, like ChatGPT and Copilot, whether they still aren't performing well for my mother tongue, Amharic. And truly, it's still very, very bad.

I don't know if you've ever seen this Simpson's episode where - I don't know if you like The Simpsons - but Marge takes the kids to this shady neighborhood and they stumble upon an Ethiopian restaurant. And this also, by the way, was a big deal for our community. We're like, oh, my gosh, our culture is represented on The Simpsons. And it went viral within our community.

But she takes them into this Ethiopian restaurant, and they are first really like, ugh, what's going on here? And people are speaking weird languages. They end up loving the food. That’s the whole point why we were obsessed with this episode.

But anyway, in that episode, the Amharic they're purporting to speak is gibberish. It has the right intonations. And if you're not paying attention, you would think that they're speaking Amharic, but it's not Amharic. That's exactly how these generative AI systems sound when you try to engage them in Amharic, the voice capability where you talk to the system. And it's hilarious and sad, but definitely hilarious. So yes, they're still not performing well, at least in my mother tongue.

And I will also clarify that although Amharic is my mother tongue, that was the first language I learned to speak. At the age of four, we moved to Uganda. And I needed to speak English given the school that my parents put me in. So they stopped speaking Amharic at home completely.

So I learned English, it felt like, in a matter of days. I couldn't speak it at one point, then I spoke it at another. I really don't even remember the struggles of learning English. But that meant I forgot Amharic.
And so I had to relearn it later. So although I still consider myself conversationally fluent, it's still challenging for me to express my deep inner thoughts in Amharic. I still think in English, and I translate into Amharic.

So it's an interesting experience when you're trying to communicate with family members who don't speak English, and you have to rely only on the Amharic skills that you have. I was telling my mother the other day that I wish ChatGPT or Copilot could speak Amharic so I could ask it questions to teach me the right phrases to say to my elderly aunt who is dealing with something. And I really want to express to her condolences or something more emotional or personal. And it's hard for me to do that.

But having these type of AI systems that can help in these cases could have been pretty incredible. But yeah, so I think there's value here, honestly. I really do. If we can get these systems to perform well in these languages and it doesn't replicate these extractive practices that we've seen for other industries, I think there could be just tremendous value.

KIMBERLY NEVALA: And as you were talking, though, it raises the point or the concern that if someone's trying to use - because you hear this. it's a trope, where folks will say, well, we've basically made all of the knowledge available in the world in these models. And it is now perhaps my-- it makes my head explode. So it's my least favorite saying. I think it's the most ridiculous thing.

But the danger herein when we think about this, relative to things like low resource languages or somebody wanting to, with all good intent, communicate, for instance, make a nice gesture to you in a native language that I don't speak, is to assume that because what comes out of the machine looks, or maybe even if it's vocalized, sounds like it might be reasonable, that you think the translation is OK.

And I would have no way of knowing that until I presented that proudly or tried to - my pronunciation would be so bad anyway, they'd probably still assume the fault lies with me - until I babbled that gibberish at somebody and they were like, what are you saying? And in spoken conversation, that's one thing. But in a written back and forth, I could see this going very badly, very fast.

HIWOT TESFAYE: Yeah, I couldn't agree more.

There's first just plus 1,000 to this trope about these systems are trained on the world's knowledge. They are not. They are trained on a fraction of the world's knowledge based on what's on the internet. So I couldn't agree more.

I also just deeply think about, I know I'm talking a lot about Ethiopia, but this is just my truth. Even the biblical manuscripts that exists that were written in the Middle Ages or even before, like in 400 AD.
We have these ancient manuscripts that talk about these medicinal properties of different herbs and plants and astronomy. And there's science and math in there, too, that hasn't been deciphered. It's still within the possession of monks who have been keeping these manuscripts safe for the last several centuries.

So there's just a lot of undiscovered, untapped knowledge that's out there. It just hasn't been digitized. So I just couldn't agree more to your point about this trope that people like to sling around about generative AI systems being trained on the universe of knowledge or the world, all of the world's knowledge. That's just simply not true. But now I've forgotten your question. What was your question beyond the trope?

KIMBERLY NEVALA: I might have gotten off on a bit of a rant about the trope, so fair enough.

Well, there may be a predilection then for folks to say, maybe this is another trope, that these systems are as bad as they're ever going to get. And they're just going to get better and better. We're going to feed this information in and, voila, at some point, this will happen.

But why are these types of issues - we'll just continue using low language models, low resource languages as an example although it's not the only exemplar by far. There's lots of other contexts where this could be applied. But why is this so hard to address? And have there been any interesting efforts to bridge the gap that you've seen? Where have those succeeded and what have those limitations been?

HIWOT TESFAYE: Yeah, I think we talked about the digitization of these languages already, so I won't touch on that.

But the other aspect is how challenging sometimes it is to make the business case to the builders of these systems, to prioritize these languages. And prioritizing these languages, it's not an insignificant effort. It's a lot of effort. You have to go and collect and clean up that data. And that data collection process, it's not trivial. It's not trivial.

Even for English, they didn't go out there and collect data from people. They just scraped what was available online. So then to add this additional barrier of this data is not online, now you have to go out there and collect it. Making that business case is challenging. Not impossible, because I think there are efforts that we can talk about later where other governments and other nonprofit entities, grassroots efforts are trying to address this gap. To make it a little bit easier for these AI developers to have access to this data, do the collection themselves so that these companies can leverage that data.

The other thing that I learned recently-ish is, if we look at the models themselves, these transformer-based architectures, but even before the data gets into the transformer architecture, there's this process called tokenization that needs to happen. Where the data needs to be, the text needs to be, broken down into smaller components.

The tokenization process itself is designed, or the most commonly used tokenizers, let me say that, were originally designed for Latin script and predominantly for English. So other languages that use Latin script tend to perform generally well with the constraints of those tokenizers or the way the tokenizers were built.

But when you have non-Latin script languages that are also low resource, now, you're in more dangerous waters where the tokenizers weren't built for that script. So they end up breaking down the words into really tiny tokens, which makes the downstream training process significantly less efficient.

So now, not only do you not have a lot of data to start with, the front end of this process has already made it computationally extremely expensive to train a system for that model. So that's this other element of this equation that I learned about recently that I was like, whoa. There's so many components that amplify this challenge that, not impossible, but just make it even harder.

So yeah, I feel like those are really the main things: data availability, digitized data availability, the tokenizer issue, and yeah, the general challenge of trying to make the business case for these organizations to prioritize it.

KIMBERLY NEVALA: Yeah, you had shared this interesting example. I think it was around Iceland about what they're doing. And I will say up front, I would love you to share quickly this novel approach they've taken. But I just want to also acknowledge before we do that that this also is in an area where they have a high level already of digital maturity, funding's available, other foundational problems that may exist in other parts of the world don't exist there. But could you give a quick example of just how Iceland has thought about this problem?

HIWOT TESFAYE: Yeah, I like to talk about this case study. And it's challenging for me to say I know all the details of the ins and outs of their project and I definitely don't want to speak for them.

But I do think it's an inspiring story in that Iceland is a relatively small country, although decently well-resourced in comparison to other countries, like you mentioned. But the speakers of Icelandic, it's a small population of people who speak that language. There is, as far as I've learned, a growing concern that the newer generation is preferring to speak in English and interact in digital spaces in English. And there's this fear that the language will die out as subsequent generations continue to prioritize speaking and interacting with one another in digital spaces in English.

So there's just been a concerted effort that the government has put forth to try to preserve Icelandic for digital spaces. And they have gone out and done all of this data collection. Not just for text, but for audio, engaging their citizens to donate their voices, crowdsourcing, in a lot of cases, the data that's needed to train systems for Icelandic.

So they did that effort for the last several years, and then the next tranche of effort, as far as I understand, was also to petition companies like OpenAI to train on the data that they have. They're like, we've collected all this data and now all we're asking is for you to just use it. And OpenAI did. And ChatGPT is apparently - apparently, I haven't tested it myself because I don't speak Icelandic - but apparently fluent in Icelandic.

Which is honestly, it is an incredibly inspiring story, not to say that it's replicable by every country because its resources and priorities are different across different governments. But there could be some version of this playbook that could be valuable for people that share the same concern for their language and have the resources to address it.

And in cases where governments may not have the right resources to address it, there are roles that multilateral organizations can play, different foundations can play to support and fill that gap. But I do think that the Icelandic government's approach in trying to solve this problem essentially for themselves is great. And it seems like they are willing to share their approach so that people aren't necessarily trying to do this on their own and make the same mistakes they may have made. So just sharing their lessons and their approach with others seems to be something that they're willing to do, which is incredible.

KIMBERLY NEVALA: And that is, that's interesting. I don't know that I had clicked into this distinction either when we had initially touched on this. But Iceland is a very small country, and so it's also a very, even from the low resource perspective, it's not just that it wasn't perhaps digitized it's also they've got a very small number of speakers speaking this language as well. Which is different than, for instance, some of the examples you gave for the various native languages in, yes, Ethiopia, but other places in Africa, or even India, and so on and so forth. So again, the definition of what low resource looks like might be different.

But I'd like to then take a step back because this is a novel approach and an interesting way to say we can get through these problems if we need to. But not without a rather significant amount of effort. But it certainly begs the question of are these really the problems that need to be solved?

So we've been talking a little bit here in the context of LLMs and foundational models. And there is certainly, in what I'm going to call in, as they say, scare quotes, from the First World perspective this thought that AI needs to be, right? It's a political, it's a social, it's even an economic priority right now. And I'm wondering how that might, in your experience or from your perspective, diverge from realities on the ground globally. Are we even solving for the right problems right now, with or without AI, in some of these other areas of the world?

HIWOT TESFAYE: I love that question. I also want to caveat the way I'm about to answer this question by saying that I haven't lived in a developing economy in a long time. I grew up all over East Africa and I went to college in Canada at the age of 18. And I haven't lived in Kenya or Ethiopia or any of the countries I grew up in in a long time. So my perspective might be a little bit divorced from reality. So I just wanted to caveat that first.

But I did get the opportunity to travel to Nairobi in 2023 to attend this event called AI Connect. I believe it was sponsored by the State Department, actually. And we got the opportunity to go. And the program was initiated by the State Department by a colleague of mine, actually, Mac Scott, and then in partnership with a think tank. Anyway, they hosted this event in Nairobi, and there were AI experts from across the Global South there.

And one of them had said something really interesting to me about this case where, I think it was some European company had come to Nigeria to offer robotic services to help offload suitcases from the airplane onto a conveyor belt or something like that so that it doesn't require human labor to carry the suitcases and put them in a particular place. And ultimately, the answer was, no, we don't need this because we actually need to, we need the jobs more than we need the automation.

On the continent of Africa, we have one of the youngest populations in the world. I can't remember the statistic but it was like around 40% of the population is under the age of 19 or something. Don't quote me on that, but it's something to that degree. The volume of youth that exists in terms of the proportion of the population is astounding.

And so I think a lot of countries who are faced with a growing youth population may not be thinking about this opportunity of AI in the same way. A lot of the topics that came up during that event was around jobs. Like how do we provide jobs for this youth that's coming up? How do we provide the necessary skilling? So capacity building and skilling and jobs was at the forefront of the conversation. Of course, there were other priorities mentioned, like sustainable development goals.

Economic development is at the forefront of these conversations, and it's all seen through the lens of advancing their economies, providing jobs for their people. And so, yeah, it does definitely beg the question of are we even prioritizing the same things? And I think if the value prop for AI isn't approached from that same lens, it may not get prioritized by a lot of these economies who are more concerned about growing their economies and providing jobs for their people, providing access to health care for their people, and education, and so on.

KIMBERLY NEVALA: And I will say up front there's this interesting thread recently, and perhaps a somewhat cynical thread, that I've seen about whether our current uses and approaches to where we apply AI, even here in the States, is truly innovative. And this idea of what constraints are we overcoming? What real problems are we solving?

To me, this circles back a little bit to what you said way back at the start of this conversation, which is there are these beautiful minds in these beautiful places who have problems to solve, have ideas about how to solve them, have a view into their own world and how they want to live.

And this approach that we take of inclusivity as an import problem completely then cuts us off from then opening that conversation and looking at that. Not just how can we scale them up so they can use the AI we're going to sell. But how are we actually starting at the very ground level? Which is what is the problem they're trying to solve and building into where, when, and if, and then how. Which might look very, very different than your garden variety chatbot or robot.

HIWOT TESFAYE: Yes, absolutely.

It makes me think about where you start the conversation, where you start probing. And if you start with we have AI. How can we use it in every possible way? It's like you have a hammer, and everything starts to look like a nail. It does not position you well to actually tackle the problems that people really genuinely care about. So yeah, I'm totally aligned with you on that.

And then, of course, it's like the fit for purpose thing is also something from a responsible AI perspective we advocate a lot and often falls on deaf ears. You have this AI system, but the problem you're trying to solve is actually not solvable by this AI system. Or it's maybe partially solvable by this AI system, but could there be a better solution here, a simpler, more affordable, more human-centric solution that you're not even thinking about because you have this hammer.

And you are going to use this hammer. And you're very determined to use this hammer rather than thinking maybe a wrench might be more suitable in this circumstance.

KIMBERLY NEVALA: Ah, fighting words right now. You're absolutely right. It's the right tool for the job. And I guess it starts with understanding what the job is to begin with.

So as we wrap up here and with all of the work you've been doing, the learnings that you've been gathering, are there particular perspectives or principles you'd like to leave with the audience and that you'd like to see more centered as we move forward in this brave new world of AI?

HIWOT TESFAYE: Yeah, I think we, I guess I, have a tendency maybe to focus on the challenges and what's not solved and all the maybe not so wise paths that we find ourselves on.

But as a skeptic, despite being a skeptic, I have genuinely been surprised and delighted - it makes me nauseous saying that -but delighted by some of my experiences using AI systems. It has been incredibly helpful in my day-to-day work, I'll say, less about entertainment and just killing time chatting with AI systems. But more so, I'm trying to find information. I'm trying to synthesize information. I've found it to be incredibly helpful.

I've also found it to be incredibly helpful in envisioning other realities for myself from a personal perspective. Of course, when you are maybe not, your perspective and even the way you look is not well represented in image generation systems, you have to find tricks to make it actually represent the thing that you're trying to have it generate. And that has been, of course, not great. But I have found it to be incredibly helpful to envision other realities for myself. It's been fun doing that. I've been working on that with a career coach of mine, actually, which has been really fun. So yes, I think I don't want to harp on the negative only. There is really interesting and positive uses for these technologies.

But I think the last thing I'll say is pulling on the thread that we've been talking about this whole time about multilingual, multicultural AI systems, making these systems work globally. I think there just needs to be more work done to make the business case for this. There is an assumption that there aren't customers willing to spend the capital to actually acquire these services or these tools that work well for their context.

There's an assumption there, and I think there needs to be some work to dispel that assumption, if it's true, in fact, if there are customers out there that are willing to spend the capital. Let's do the research on that. Let's make that business case. So yeah, I think I just want to leave it at that. That could be one of the key things that can help us make these systems available and useful for a broader population.

KIMBERLY NEVALA: I think that's a great note to end on. And I just want to really thank you for your time and thoughtful insights today as well.

HIWOT TESFAYE: Oh, thank you, Kimberly.

KIMBERLY NEVALA: This has been fun. We will continue to follow your work and we will let people know where to find you in the show notes as well. And hopefully, we can get you back in the future and see if we are gaining ground in that vision of truly inclusive technology.

HIWOT TESFAYE: Fingers crossed. We'll be doing the work.

KIMBERLY NEVALA: Fingers crossed.

HIWOT TESFAYE: Thanks, Kimberly.
[MUSIC PLAYING]

KIMBERLY NEVALA: To continue learning from thinkers, doers, and advocates like Hiwot subscribe to Pondering AI now. You can find us wherever you listen to your podcasts and also on YouTube. In addition, if you have comments, questions, or guest suggestions, please let us know at PonderingAI@SAS.com.