Doha Debates: Are we heading into a world divided by AI tribes?

[0:08] There's no turning back. AI is everywhere, from maps to shopping. Some even get therapy from [0:14] chatbots. Yet they rarely ask whose expertise those bots reflect. So how will AI shape our [0:20] collective thinking, our politics, our culture? AI models are said to carry the values of those [0:26] who have created them. Some have been called libertarian, some woke, others appear to change [0:32] from day to day or even user to user. I'm Mohamed Hassan and this is the Doha Debates Podcast. [0:37] You should be asking as to what extent AI will solidify and accelerate the dependence that our [0:47] society already has on corporations. I think it's the individual versus him or herself. That's where [0:54] the struggle for freedom will exist in the future. Even though these models are supposed to be open [1:01] for use, they ultimately come embedded with biases and political ideologies. This idea that there's [1:07] this existential race between the US and China to get to super intelligence first. This is like a [1:14] path to planetary suicide. Today's debate, are we heading into a world divided by AI tribes? [1:20] And is this something to celebrate or fear? I'd like to welcome our guests. James Brousseau is [1:33] a professor of philosophy and computer science at Pace University. Robert Wright is the author of The [1:38] God Test, Artificial Intelligence, and Our Coming Cosmic Reckoning. Eleanor Noor is a senior fellow in [1:44] the Asia program at Carnegie Endowment for International Peace. And Evgeny Morozov is a tech theorist and author. [1:50] A warm welcome to you all. Thank you. Happy to be here. The four of you are experts in this topic and me, [1:58] like a lot of people, is just trying to figure out what is going on. So I want us to have a rigorous, [2:03] high level debate. But at the same time, I'm going to interject from time to time with some really dumb [2:07] questions. So please allow me. James, if we can begin with you, can you give us a picture of the [2:15] AI landscape as it exists today? Do you think the direction we're heading in is one of unity or division? [2:23] Well, the landscape is certainly confused. All of us, I suspect, are somewhat confused about what we see [2:28] happening around us. But to my mind, the landscape is full of potential. So it seems to me that there [2:37] are many kinds of positive aspects to the changes we are seeing. There are different ways that we'll [2:41] be able to receive and understand information. So my general outlook is positive. I will say quickly, [2:48] it seems to me that the major threat that we are facing now is one of a kind of homogenization [2:58] that is I'm thinking just of the experience all of us have had. For example, on Netflix, [3:04] when we see a movie and then enjoy it and we get recommended another which is similar and then another [3:09] which is similar. I think the same thing goes for LinkedIn, which is proposing job opportunities for us, [3:16] or for dating websites, which is proposing partners for us. These kinds of tools increasingly force us [3:24] to accept possibilities which correspond with the data that we have accumulated about our own past. [3:32] Right? That's how Netflix produces your movie recommendations. They say, well, what have you [3:36] watched in the past? What have you enjoyed? This is what you're going to get. Now, that can be very [3:42] satisfying and enjoyable. But the risk, and so to answer the question, I've said there's some positive [3:47] things. The risk and what worries me about this technology is the possibility for homogenization [3:55] even with ourselves. That is, it will become impossible for us to escape our own data. The [4:03] preferences and interests that we have established about ourselves will determine what we see when we [4:09] turn on the television set, what jobs we are offered, what partners we might receive. So I think that [4:15] that is the threat. And then finally, I will add, I think that the challenge that waits for us is to [4:22] find ways to use these tools to help us escape who we are, to help us become new and different kinds of [4:30] people. And I think that if we can do that, then I believe that the future will be very positive for [4:36] this technology and for humanity. [4:38] Elena, if AI is creating a homogenous experience of the way that we see the world, the way we interact [4:46] with technology, is that not then going to lead to a kind of unity? [4:50] Yes, potentially. And I think it depends again on the use cases, right? James pointed out some examples. [4:57] I think one of the examples that has a double-edged sword is these accent neutralization programs [5:07] that are being churned out by AI. So for example, you have all these business processing organizations [5:13] outfits in places like the Philippines and in India and Kenya. And on the one hand, they allow for [5:19] greater work opportunities. You know, this idea that there is dignity with work. At the same time, [5:26] through this accent neutralization software, it flattens the accents around the world in order for [5:32] there to be greater efficiency, greater levels of communication, supposedly. But what does it do to [5:40] a person's identity, a person's culture? The way they speak reflects their history, their experiences. [5:46] And once you start flattening that, yes, there is some unity of communication, I suppose. It facilitates [5:52] easy listening and understanding. But at the same time, I think there's a real risk of people losing [5:58] who they are at the risk of facilitating communication and understanding. [6:05] Evgeny, what do you think about this idea that our identities themselves are being shaped by this AI? [6:11] Are we being made to become robots that all think and behave the same way? [6:16] I mean, look, our identities are shaped by almost anything we use, by any kind of technological or [6:24] ideological infrastructure. So, of course, AI would be just like any other technology exercising effects [6:31] upon people who use it. The question is, what kind of effects do we want to exercise upon us in the [6:38] long term? And what kind of options do we still have in terms of determining the path and the trajectory [6:45] in which AI systems would travel? Because if we look at earlier technologies, we see a very [6:53] interesting trend. All of them start as a series of competing ideologies about the good life, [7:01] what it means to live with the airplane, or what it means to live with the car, or what it means to [7:07] live with cable television. Those ideologies, however, all get embedded into socioeconomic systems of [7:15] various kinds, into governance, into corporations, into regulatory bodies. And in some cases, there is some [7:22] resonance where a certain ideology for individualistic driving or living gets attached to a set of [7:30] business interests, and it crowds out all the others. And in some cases, like the United States, [7:36] this car-centric ideology won out. In some cases, like in many European countries, it did not, where you [7:42] still have the idea of public transport. I don't think that AI is any different. We are still very early in this [7:48] development cycle. And there are still relatively open pathways that it can travel, depending on what [7:56] kind of ideology is allowed to win out, and what kind of socioeconomic institutions are in place to be [8:03] able to balance what would otherwise be the dominant ideology, which in our day is one that favors the [8:10] interest of corporations. So whether we manage to balance it out would, to a large extent, determine how [8:17] homogenizing and unifying these technologies would be, because they can have all sorts of effects. [8:23] And it's still up to us to determine what kind of effects we want it to have. [8:28] And Robert, what do you think of this idea of ideologies? Obviously, this is a really astute point [8:34] that the technologists that shape our technologies, their ideologies can and will be imprinted onto these [8:42] systems. Do you think this is what we're seeing already now with AI? [8:46] I think the people who build the AIs can, in principle, use them to shape ideologies through [8:52] the selection of training data, through the selection of the experts who play a role in the post-training [8:58] phase. A lot depends on which magazines of which ideological slant are part of the training data. [9:06] But even beyond that, I think there may be a naturally, you could say, tribalizing effect of AI, much as [9:16] there has been with social media. And for somewhat the same reason, which is that there's an incentive to [9:23] maximize engagement. That's what the corporations want. They want you to keep using their product. We've already seen that in [9:30] in the case of social media. That leads to algorithms that reinforce people's conviction that their [9:37] ideology is right and the other was wrong and feeds them information selectively. And we're already seeing [9:42] the same thing in AI. And if anything, AI can, in principle, do that in a more kind of microscopic, [9:50] fine-grained way. And we've seen cases of people's very specific, narrow beliefs [9:57] being reinforced, even in a pathological way. So people who think maybe they've discovered some [10:03] secret formula, being encouraged by the AI to believe that, even though it's not true, or to [10:12] become more convinced that their conspiracy theories are true. [10:15] AI psychosis, I think, is what we're going to hold it. [10:18] And it's a real danger that it will build kind of cocoons for people to live in that are even narrower [10:25] than our social media cocoons. James, what do you think? [10:28] You know, let me try two different ideologies on for size here. One ideology is the sort of [10:36] Heideggerian early 20th century ideology, which is that what we are looking for as human beings [10:43] is some kind of authenticity. That is some sense of who we are, which is stable and solid, [10:48] in which we can locate and cling to. If we want to go that route, if authenticity is the [10:55] ideology we want to follow, then it seems to me that some of these problems, which you're [10:59] certainly right to point out, Robert, but some of these problems are symptoms of a kind of strength [11:04] that the AI does help us find for ourselves, each one of us, who we are, in some ways that are trivial, [11:11] going back to my example of Netflix movies, but in some ways which are not trivial. For example, [11:15] if the AI consistently finds for me on a dating website, the same kind of partner, then perhaps [11:21] that's who I am. And I should cling to that, hold on to that. So that's one ideology. I think AI can [11:27] serve that ideology. But there's also a different possible ideology, which is that freedom is more [11:33] important than authenticity. And in that case, what we would want our artificial intelligence to do, [11:39] and I personally, I will say I argue for this, what we would want our artificial intelligence to do [11:44] is, going back to your point about languages, is precisely break those kinds of uniformities, [11:51] those kinds of generic realities that AI can sometimes force us into, and help us discover [11:58] new ways of living and existing. So I think that's, those two ideologies are fundamentally, [12:05] it's one way of thinking about this. Fundamentally, what is in play in artificial intelligence is whether [12:11] we want to be who we are, and AI will help us find that, or we want the ability to choose different [12:18] kinds of futures for ourselves, ones that are unknown to who we are now. [12:22] Alina, what do you think? [12:23] Alina Nguyen- So that's a really interesting tension that you brought up. And I think there's [12:28] an overlap as well between authenticity and freedom, because in finding authenticity, you need that [12:34] freedom as well. Whether AI can help us find who we really are, I think is a more questionable proposition, [12:42] simply because a lot of the systems are built in global North countries, and that are sold to us in [12:53] the global majority. And a lot of that data as well, has a bias towards English language, data training [13:03] sets, as well as the ideologies that go with that. So for example, in Indonesia, Stanford Raffles, [13:12] Thomas Stanford Raffles was once the administrator of Java, and he wrote the book, The History of Java. [13:19] Now, I'm willing to bet, I don't speak Javanese, but I'm willing to bet that his translation and [13:24] interpretation of the Javanese people is slightly, maybe even dramatically different from how the [13:31] Javanese would describe themselves. But he, of course, had a different objective to writing the [13:37] history of Java. And it was to tell London at the time that, look, these Javanese people, they are of [13:44] a certain civilization. They are civilized because they have hierarchy in their structure. And this is [13:51] why perhaps we should not impose such a harsh rule over them. And I don't think the Javanese saw themselves [13:57] that way. And so this idea of who we are authentically is partly in that search for freedom. [14:07] But I don't think at this point in time, at this developmental stage, understanding full well, [14:12] Evgeny's point, that we can shape it right now as it is given the bias in the training data and the [14:19] systems of deployment. It's not particularly favorable to people in the global majority to find that [14:25] authenticity when there are narratives about them that have been told by others and that continue [14:30] to float around in the ether. Can I just ask you one quick question? Just a quick thought. Do you think [14:35] AI is fundamentally about societies and communities? Or do you think it's fundamentally about individuals [14:42] in their lives? Because it seems that I was sort of talking about how individuals live. That's a good [14:46] question. And I assume each of you will have a different idea. Right. And I think that you're coming [14:51] at this from how societies are. So I wonder what you think about that, about what AI is, [14:55] where it goes fundamentally. Yeah, my instinct is that it's a tool for society. But obviously, [15:02] the emphasis on whether it is something for individuals or for a community per se is going [15:07] to depend on the context in which it is deployed. Evgeny, what do you think about that? Yeah, no, [15:12] I just want to make a provocation maybe in that I think that AI in itself does nothing. It's an empty [15:19] concept which is up to us to fill with meaning. So in that sense, if you really want to stick to the [15:25] dichotomy of authenticity versus freedom, then we should push it to the ultimate conclusions, [15:30] meaning that we should have the freedom to fill in that concept with the meaning that all of us as [15:36] individuals or communities would like to fill that with. Right. And in that sense, I don't know why [15:42] we would opt for freedom for us as individuals, but we would accept that there is some kind of [15:47] essentialism or authenticity about AI that it should be about individuals or it should be about [15:53] communities. It should not be about anything. It's essentially a very open-ended plastic set [15:59] of technologies, which depending on billions of dollars that are now going to flow into this might [16:06] go one direction or it might go in a completely different direction. We've seen that with cable [16:11] television. We've seen it with the internet. We've seen it with a lot of other technologies where [16:15] there was a lot of rhetoric about how these technologies will develop on their own because [16:20] they have some inherent properties, but the institutional support was not in place. [16:26] And they ended up following the path of least resistance, which was an agenda favorable to [16:31] companies that were well positioned to take advantage of those technologies and monetize them. [16:35] It's all up to us and up to institutions, communities, politics, and political struggles to define [16:41] how all of these ingredients should fit together and who should pay the bills. Should it be individuals [16:46] paying a monthly subscription fee? Should it be some kind of an intermediate institution like we have [16:52] with libraries, which did create very utopian spaces, which suspended copyright law temporarily [16:59] so that people can have wider access to literature? Why aren't we thinking about alternative institutional models? [17:05] Robert, what do you think of this idea of the plasticity of AI? Would you agree with it? [17:10] In principle, AI can be anything we want in a couple of senses. First of all, in principle, [17:17] societies, nations, even the whole world can get together and declare policies and constrain it in [17:24] that way. Moreover, individuals can choose models they like with specific features they like. In principle, [17:32] you could offer consumers a version of the model that will make them mindful, [17:39] which I'd like to be more mindful than I am. It's not easy. It wouldn't surprise me if you have [17:49] different religions offering different kinds of models. In principle, the range is unlimited, but [17:58] I don't think we should act as if it's going to be easy to secure our freedom [18:04] in the face of AI because the makers of the AIs are going to have their own agendas. In the case of [18:11] corporations, again, that's to keep us using the AI. And AIs are going to be better and better at [18:18] manipulating us, at figuring out what we like, what will make us come back to the AI. I mean, there's already [18:27] been a study showing that AIs were better than people at persuading other people to change their [18:35] minds about things. And in particular, the AIs were better if they were given access to personal data [18:42] about the people they were trying to persuade, then they really excelled, okay? And when you remember [18:48] that an AI can, in principle, search your whole social media history, immediately come to know you [18:54] very well, and then set out to persuade you of something, like to keep using it, right? Like [18:59] to push whatever buttons of yours will keep you using it, you know, that's a lot of power. And I [19:05] think if we're going to remain free amid the AI revolution, we're going to have to try consciously [19:12] to cultivate, you know, what some people are calling cognitive sovereignty, you know, and be able to resist [19:21] manipulative tendencies that I think inevitably will be emanating from the technology. [19:27] James, this question of freedom is really key here. How do we know that we can trust these systems [19:32] that we're using? Terrifically provocative. It seems to me that this is a critical point. In the past, [19:39] we have been enslaved. We have been limited. We've been restrained by others, by other people, [19:47] by other institutions. But today, the restraint, the limitation, even in some cases, the slavery, [19:54] sense of nudges, all of that is something that we do to ourselves. We are controlled by our own data, [20:03] our own past. That's what the AI is manipulating. We cannot be nudged, I cannot be nudged by her data, [20:09] or by your data. What the AI uses to control what charity I want to donate to, or what person I want [20:17] to find to date, what the AI uses to control are my own preferences and my own interests. So, it's a [20:23] very different notion of freedom we have now. Freedom is a struggle that I have with myself and with my [20:28] own past, as opposed to what it used to be, a struggle I had with institutions or with others. [20:35] So, then, I think I stand a little bit contrary to your view. It seems to me that we live in, [20:40] we're entering a post-institutional age. And I think that, for the reasons that you were stating, [20:47] everyone can now, or very soon, we'll be able to form their own language model, which is tailored [20:54] to their own preferences and their own aspirations. In the case of language, there are many people working [21:00] now, and as you know, of course, better than I do, many people working now to create new models, [21:05] which do reflect the Javanese language and culture and so on. So, we're going from one general language [21:11] to specific languages. And I think we'll continue moving down toward one person. Each of us will have [21:16] our own large language model with our own preferences, our own interests. And so, that's where... So, I don't [21:23] think it's the individual versus the institution. I think it's the individual versus him or her self. [21:30] That's where the struggle for freedom will exist in the future. [21:33] Evgeny, is AI, in your mind, liberating us from these power structures? [21:38] No, no. First of all, I don't agree that we're entering any post-institutional landscape. I think we [21:44] are actually living in a hyper-institutional landscape. And the institutions that primarily [21:48] dominate our lives are corporations and one matter institution that I would call capital. And the data [21:55] that AI essentially rides on is extracted from me, not because I make these decisions, but because [22:02] there are institutional pressures by banks, by financial companies, by the fact that they need [22:07] to have a credit history, by the fact that our parallel welfare system of some kind expects that [22:13] I need to be using digital services because they're heavily subsidized by venture capital or sovereign [22:19] wealth funds. And it would be naive of us to say that the pressures on a billionaire in Silicon Valley, [22:24] who can be off-grid and send their kids to a school where there are no devices, are the same as on [22:30] someone who is earning minimum wage and has to juggle all these apps to essentially stay afloat. [22:35] Of course, these pressures are very different. And that pressure comes from abstract institutions, [22:40] maybe. Maybe it's the wage that's exercising on the fact that you have a stomach and you need to [22:45] eat. But those pressures are real. They're not fake. So we cannot do anything we want just because there is [22:50] suddenly a series of AI apps that allow us to be hyperproductive. I mean, there are still constraints [22:56] and pressures. So in that sense, I don't see us liberating ourselves from institutions anytime soon. [23:03] But beyond that, again, the question that we should be asking is to what extent AI will solidify and [23:10] accelerate the dependence that our society already has on corporations. I mean, we used to have institutions, [23:18] like museums, libraries, media, and so forth, that had some independence from both the state, at least [23:27] like theoretically, and the corporations. Right now, if you look at what's left of those institutions, [23:33] they by and large survive on donations from tech companies. And many of them have been replaced by [23:39] them entirely. And in that sense, AI will accelerate that process. And we will end up in a world where [23:47] the previous setup, where at least there was some pluralism and the opposing logic of those [23:52] institutions were clashing with each other and allowed some space for the individuals and citizens [23:57] to flourish, will just be oriented towards serving the needs of one sector. And that sector will be big [24:02] tech. I think Evgeny's point is perfectly valid, though, at the end of the day, we're still very [24:07] dependent on only a few actors, primarily within the global north. We're dependent on them for [24:13] infrastructure for standards, technical and policy standards that determine what we can do, what [24:19] types of chips we can use in order to create these much more localized language models, for example. [24:26] So at the end of the day, I think so many of us in the world are ultimately beholden to just a few [24:31] powerful players, corporations essentially, that determine how much sovereignty, how much freedom we [24:37] can exercise, even with the little crumbs that they throw us in terms of, you know, what sort of [24:43] programs we can create with AI, for example. In the nation state of Tuvalu is the first digital nation [24:53] in the world, because it is at the precipice of an existential crisis due to the climate emergency. And a lot of [25:01] the people of Tuvalu have decided they're going to migrate their whole nation state onto the digital [25:07] platform in order to preserve what's remaining of their culture and traditions so that future generations [25:14] can understand who they are in terms of their identity. But ultimately, whether they succeed or not [25:19] in creating a digital nation is dependent on what sort of infrastructure they can access, the kinds of [25:26] submarine cables, the kinds of chips to power the data centers, and the kinds of training data they're [25:32] able to collect themselves. And so this gets to the point that I think both James and Yevgeny are right, [25:39] we're just looking at it at different ends of the spectrum. And do you think this technology can allow [25:45] these nations, sometimes very small nations, to develop their own sense of autonomy, their own sense [25:49] of sovereignty, not only over their future, but also over their past? Yes, yes, absolutely. But then we [25:56] get back to the basic question of, who does all this infrastructure belong to? And what sort of [26:01] standards are they setting for the rest of us in order to access their technologies? We talk about [26:06] open models, Bob, you mentioned that, right? You look at the AI action plan that the White House put out [26:12] earlier this year, it talks a lot about open models. But at the same time, Washington is trying to market [26:20] an American AI stack with American values. So they seem to be completely conflicting. Yet at the same [26:26] time, these are kind of the options that we're left with. There are open models, but we have no access [26:31] to absolutely no transparency of their training data. So we have no idea of the biases, no accountability [26:38] in that sense. And even though these models are supposed to be open for use, they ultimately come [26:45] embedded with biases and political ideologies that many of us may not even be aware of. [26:52] Is there a way to liberate ourselves from these biases or at least be aware of them when we're [26:56] using these systems? Well, I think in terms of national bias, there is. I think the open source or open [27:05] weights models that China in particular is focusing on are not very far at all behind the proprietary [27:15] models that the US is marketing. And if I'm a nation that doesn't want to be dominated by American [27:25] decisions about what a large language model should be and what its training data should be, and I do [27:31] think it's outrageous that they don't even tell us what the training data is, that they don't even [27:35] disclose what journals it did and didn't train on and so on. But if I'm a nation that wants to be [27:45] liberated from that kind of bias, I'm going to find these open source models very appealing because [27:52] you can kind of retrain them to whatever extent you want and actually change the so-called weights, [27:59] which is what determines how they work inside and in effect create a customized model. Now, at the [28:07] individual level, that's more challenging. I mean, most people don't have the wherewithal [28:12] to customize an open source model to them. James, what do you think of that? [28:17] Are these teething issues? I myself believe that the technology is moving so rapidly that working [28:25] outside the system is not practical. And so, in this way, also, I'm post-institutionalist. I believe [28:31] that even if we wanted institutions to be powerful and to be helpful, they could not be. Let me give you [28:37] just a quick example. So, I live half my life in New York City and the other half my life in Trento, [28:42] Italy. I teach AI ethics in both of those two places. But because I'm in Italy, I've seen quite [28:48] a bit of what the Europeans are doing with respect to laws. And one very important law for them was this [28:54] AI Act, which was this 18-month or two-year work to construct this regulating structure that was going to [29:01] control AI finally and at last. And just as they were rolling it out, what happened? ChatGPT appeared. [29:08] So, we have this institutional rule. It's about to be imposed, rolled out, and rolled right back, [29:15] because ChatGPT rendered the whole thing irrelevant. So, it seems to me that the idea of institutions [29:21] or working outside the system will simply not work in a context of this velocity. Instead, [29:28] these kinds of problems that we're facing can be managed or perhaps we can only manage them [29:33] by working inside the system. So, I think that's possible. [29:36] I disagree with everything in this statement in that this is a false dichotomy that you're drawing, [29:43] first of all, between regulation on the one hand and some kind of innovation on the other. First of all, [29:49] the US regulates heavily. Look at all sorts of rules the US imposes on the rest of the world when it [29:56] comes to experts of AI. The US regulates more heavily than anyone else what actually the rest of the [30:03] world is allowed to do with technology. Who has access to chips? Who has access to what kinds of [30:08] chips? The US is not by any means a lesser fair player. Second, Europe is not a good example to follow, [30:17] in part because Europe's own hands have been tied by its dedication to some kind of neoliberal ideology [30:25] which prevented that from building institutions that could balance and counterbalance actually what [30:32] the market does beyond just regulation, beyond just creating laws. You can be creating massive [30:37] public infrastructures that would actually be competing with the tech providers that are private and [30:43] that come from Silicon Valley. Europe has not done that because it has been structurally prohibited [30:48] by its own laws from building up its own tech industry. You're not allowed to do state aid in Europe and [30:55] unfortunately that's how it is. China on the other hand has been regulating heavily and combined it with [31:01] active nurturing of its domestic AI industry and we see the results. It's the only country that actually has [31:08] something to show against the United States. It has a tech industry that actually competes with it and it's [31:15] doing pretty well pursuing the regulatory path that we are told cannot actually work. But beyond that I would [31:21] also be very cautious about thinking that some kind of technocratic fine-tuning utopia where we just [31:29] thinker with the weights of the model is going to save us in part because much still boils down to the [31:35] training data. And the training data is a very expensive project to undertake. And even if you play [31:41] with the Chinese models you will see that despite the fact that they're open source and they're open [31:46] weights, what ends up being is that they still have the same ideological biases that the American models [31:52] because they're trained on the same set of pirated blocks that the Chinese companies got off the same [31:57] torrent sites that the American companies, right? And if you want to have a counter project you really need to [32:03] undertake a massive government-funded effort of going and finding alternative data sets. It might be data in [32:10] your libraries, it might be that you will need to go and digitize millions of books in your own language [32:15] as opposed to relying on the English language books that are traveling around the internet, but you will [32:20] need to have some kind of an effort to do that. And by the way, even the market players who have trained their [32:27] models with that data from torrent websites, they essentially leveraged the volunteering spirit of [32:34] people who have been scanning books, digitizing them and uploading them online. They did not incur those [32:39] costs on their own. But to think that we will be able to solve that through some kind of a technical [32:45] fine tuning of the model, I think is very unrealistic. And in that sense, we do need to think about [32:51] institutions and not just regulations and laws. Regulations and laws can be very useful if they're [32:58] tied to an institutional vision. Elena, what do you think? So James, I would just offer that so many of [33:04] us in the global majority work within constraints and we have no choice but to fine tune because of the [33:10] sheer cost of building your own foundational models that Evgeny has been talking about so much. And I think [33:17] these smaller models, they offer some degree of autonomy, but going back to the earlier points [33:23] about how we would still be dependent on these few large players as a completely unequal market [33:31] concentration of power in terms of who gets what, right? And so, you know, even fundamental concepts [33:38] like time and space, for example, that are embedded in the training data of foundational models, we have no [33:45] eyes on because there's zero transparency. So for indigenous communities, for example, that think [33:51] of time and space as completely inseparable concepts, for them to rely on foundational models with completely [34:01] unclear, untransparent training data is going to change how they even fine tune these models for [34:08] themselves. And so for many of us, we have no choice but to work within these constraints. And so I'm [34:14] sympathetic to Evgeny's argument that perhaps we just need to flip the tables and recreate [34:20] institutions that work for the majority of us. Is AI then a solution to that in your eyes, James? Can AI, [34:28] can we train AI to then liberate itself from these constraints? Right. So one of the projects I'm working [34:34] on is called Curiosity or the Curiosity Engine. And so what we're trying to do with a number of students, myself, [34:41] is we are trying to work with recommenders. So again, what kind of jobs you're recommended, [34:46] what kind of language you might be recommended, and so on. So we're trying to find a way that [34:51] recommendations, which is one important aspect of AI, can lead away from, again, the individual's [34:58] personal data set and interests. And I think that that kind of project could be mimicked on the larger [35:04] social level, find a way to work again within the system to lead away from these the Silicon Valley [35:12] data set. I think that's the challenge. So the question is, how do we address that challenge best? [35:17] I suppose here, we just come down to personal preference. I'm fundamentally an individualist. [35:22] And so I'm going to be prejudiced towards saying that there's an individual path to these solutions. [35:27] Yeah, I certainly agree with you that we should only resort to legislation when that's necessary. [35:34] And a lot of problems you can solve without that. I would like to point to a category of problems [35:41] that are, I think, are going to call for policy solutions. And they're going to call for policy [35:48] solutions at the international level. And I think this ultimately gets back to [35:53] the challenge of the technology being kind of tribalizing at the national level. I mean, [36:01] if you look at something like the threat of biological weapons, okay, there's no doubt [36:10] that these models are going to make it easier for people to do various nefarious things. Individuals and [36:17] small groups of people make biological weapons and so on. That problem is particularly [36:23] queued actually for open source models. And it's why some people think that they shouldn't [36:28] be developed. But in any event, that's not a problem you can solve through national policy. [36:32] I think we'd all agree that preventing a global epidemic caused by a bioweapon is a valid policy [36:38] goal, but it's not a policy goal you can address at the national level. And very broadly, I think, [36:45] if we don't confront the challenge of AI as a cohesive global community, we are in deep trouble. [36:51] And so if it turns out that the technology itself empowers nationalism, that could be a kind of [37:01] paradoxical and very deep problem. And right now, much of Silicon Valley is wed to this idea that [37:07] there's this existential race between the US and China to get to super intelligence first. I think [37:14] this is the absolute worst environment in which to confront the AI revolution. This is like a path [37:21] to planetary suicide. And the dynamic is very strong. The impetus behind it is very strong. I mean, [37:29] the, you know, 10, 20 years ago, kind of China hawkism was not the kind of de facto ideology in [37:40] Washington. It's become that for more than one reason. But lately, one big reason has been that [37:46] so many big Silicon Valley players like OpenAI want Americans to be as afraid as possible of China, [37:53] so that they can say that the American government should not regulate OpenAI or put any constraints [38:00] on it whatsoever, and indeed should subsidize its energy and, and whatever else. [38:04] Now, there are bright sparks in the innovation in terms of the governance models around AI that we're [38:12] seeing, particularly with the indigenous communities, where they are challenging some ideas that are out [38:20] there about what it means to govern AI in a space that prioritizes exploitation and extraction, particularly [38:29] of data, but also other natural resources. And the Maori community in particular have decided that [38:36] they're going to have their own governance models of their own large language models because of the [38:42] history of having been exploited. And I think there are lessons to be drawn here for many of us. [38:48] When we think about these indigenous models that we're seeing growing, does that then present a future [38:53] in your mind of the potential for AI to allow these different communities to create their own tribes, [39:00] their own spaces, that they then can express that autonomy, that sense of sovereignty in? [39:06] Yes. And there are certainly some really positive outcomes that will result from that sort of [39:13] innovation. At the same time, for many multi-ethnic communities and nation states, it does present a bit [39:20] of a risk because it means a fragmentation of the very social fabric of a nation. And I'm thinking [39:26] here of a lot of countries in Southeast Asia, but also in Africa, where, for example, there are local [39:33] language models that are being built in terms of ethnic languages. But the history of these countries, [39:41] as in Indonesia, as in Malaysia, has been built around unifying people through a single language. [39:48] And so when you have this splintering of innovation, of models that are meant to be [39:54] ethical in terms of how they cater to local communities, there is the larger question of what [39:59] does that mean for national unity? And how do we bring those contradicting tensions together? [40:05] This is a point about nationalism, which you brought up before, Robert. How do you see this happening? [40:10] There's obviously a driver now for nationalism and nationalistic pride, especially when it comes to things [40:17] such as the defense industry, which we're seeing a lot of companies in Silicon Valley voluntarily [40:22] stepping into. And there's this race now between the U.S. and China. Do you think that we can be [40:30] getting these models that are reinforcing these identities? And is that a good thing or a bad thing? [40:36] You know, nationalism is always dangerous. And international polarization is always dangerous. [40:43] But I think in a way it's going to become more dangerous than it's ever been by virtue of AI. [40:49] I really think this technology is going to be so powerfully dislocating across a number of dimensions [40:59] and in many different ways call for a concerted international response, that if we don't become [41:06] more of a global community than we've ever been before, we're in really deep trouble. [41:11] James, do you have this concern that AI might entrench us in our already divided lines, whether [41:16] it's national or race? [41:18] Right. So I think it's becoming fairly clear from our discussion so far that I start almost all of [41:24] my thought from the individual level. So I would say two things. I would say one, almost certainly, [41:31] that's true. We're seeing clearly how filter bubbles are forming. We're seeing clearly how polarization [41:37] increases. We're seeing clearly how people are being directed into narrower and narrower affinity [41:41] groups. That's just undeniable on the ground. But again, going back to sort of where we started, [41:46] it seems to me that the answer to that question is one for AI companies to work with individuals to [41:53] help them find new ways of thinking about the world, which would be not moving from one national [41:58] structure to another national structure or from one language to another indigenous language. [42:03] Instead, I think it would mean for individuals to find their own way to a kind of language that they [42:07] might want. For example, perhaps I might decide that I want to learn Italian. I want to live like [42:11] the Italians. Now, I have no Italian past. I have no affinity with the Italians. I have no indigenous [42:17] connection with that. But why can't I become Italian as opposed to becoming, recreating sort of the roots [42:25] I had with my, it happens, my ancestry is Swedish and Danish. And I could go back. There's things called [42:32] lutefisk that we could eat and there's all kinds of specific things we could do. But I could just go [42:37] forward too. I could use artificial intelligence, the kind of tools that it offers to help me gain entrance [42:42] to a different kind of culture. So I think in that way too, this question about nationalities is shattering. [42:50] And instead it will be up for individuals to find their own place and their own destiny in this future. [42:55] And I think one hope generically about kind of the formation of new tribes that are not national [43:01] tribes is that some of them will be first of all benign, you know, not bad and international. They [43:06] will cut across national bounds and so form commonalities of interests within multiple nations. [43:14] And so kind of erode the power of nationalism. I wonder what you think about this path forward, [43:21] you know, this, this potential for division, this potential for autonomy and the potential [43:28] and the reality of the acceleration at which this technology is going. [43:32] Yeah. Well, I, unlike James, do not start from the individual analysis. So I look at the actually [43:40] existing tech companies and not some hypothetical firms out there. The actually existing tech companies [43:49] all go and their CEOs all go and genuflect in front of Donald Trump. They have funded half of his [43:56] White House reconstruction. They are part and parcel of the nationalistic and you might even say [44:03] jingoistic agenda of the United States. So the idea that I would want to have more of this [44:08] companies intermediates my life to help me learn Italian. No, thank you. I became an Italian citizen, [44:14] learned Italian. I'm doing fine. I don't need Sam Alton helping me along the way. By the way, [44:18] all of that is true. So in that sense, no, you know, I just don't know what universe do we live in. I [44:23] mean, these are real actors. Nvidia is soon worth maybe already worth $5 trillion. It's a company that's [44:30] deeply embedded now in Trump's project for redefining the American hegemony and the American empire in the [44:37] world. How can we possibly be making this hypothetical statements about what these companies might do to [44:44] create a more global universe? They are all residents of the United States. We have seen that despite the [44:51] early statements, even during the first Trump administration, they all ended up playing up to [44:57] Donald Trump's tune. I mean, I still remember in 2016 and 17 when people like Sergey Brin were going to San [45:05] Francisco airport to protest against the first ban on the Muslim countries. I mean, now Sergey Brin will [45:12] probably be transporting people to the airport if Donald Trump asked him to. You know, this is where [45:17] we are. These are actually existing tech companies and they're the ones building and rolling out AI. Do I [45:24] want to have this powerful set of infrastructures in their hands or do I want to redistribute control over [45:30] them more widely? Of course, it's the latter. Why would I be not being an American citizen? Why would I want [45:35] them to dominate it and control it and make it part of this somewhat insane Trumpian project? [45:41] And then I want to give you the final word here. When it comes to, you know, you've talked about the need [45:48] for caution, the need to potentially stop, but then there's also the realities of what is actually [45:52] happening in this technology. And in some level, it's very difficult to actually put a stop to it. [45:57] So from your perspective, how do we move forward from here? How do we protect ourselves from the [46:03] worst risks that we already see while also unlocking its potential? I think it's really good to have [46:08] innovation depending on what the particular use cases are and depending on the context in which [46:14] these innovative solutions are deployed. That said, you know, we've talked about kind of the [46:19] communalism and nationalism that AI can agenda. But I think to Yevgeny's point as well, there is a real [46:26] risk of ideological tribalism as well. So on the one hand, yes, you democratize AI for communities that [46:33] are underrepresented, languages that are underrepresented. At the same time, because of the structure of [46:40] the economy that we operate in, that enables this sort of AI innovation at large around the world, [46:48] we are inadvertently sucked into this ideological tribalization of, you know, this sort of [46:55] capitalism being the only way forward in order for us to quote, unquote, innovate. We think of [47:01] innovation, particularly in my part of the world, as an economic driver. But that also means taking into [47:07] account conventional economic theory, where things like the environment and labor are quote, unquote, [47:13] externalities. And I don't think we can afford to fall back on some of these conventional ideas [47:18] anymore, precisely because the sorts of innovation that are being called for are going to require a [47:25] greater lateral sort of thinking. [47:27] Very well put. And before I lose you guys, I have to ask this, and I know it's a personal question, [47:34] but I would love to know what is your LLM of choice? Is there one that you go to? Is there [47:39] an AI tribe that you feel like you already exist in? [47:42] I was ChatGPT, but I'm a believer in Gemini 2.5. [47:46] Hmm. Robert? [47:47] I actually, maybe I should be ashamed of this, but I pay $20 a month to Google and OpenAI and [47:54] Anthropic. [47:55] Elena, what about you? [47:56] I refuse to participate in the subscription as a service model. And so true to my advocacy of [48:03] context mattering, I play around with different models and I pick and choose as to which solution [48:10] is best delivered for my particular context. [48:12] And Evgeny, what about you? [48:13] Okay, so I have to make a confession. So here it will be a real AI Alcoholics Anonymous kind of setting. [48:19] So I have three max Claude accounts for which I'm paying probably $600 or $700 a month. [48:26] I have a Gemini Ultra account and I have GPT Top accounts. So it all ends up to about $1,000 a month. [48:33] Wow. You are the ideal AI user. [48:38] Everything reverses in the last year. [48:41] I mean, I live, like I spend most of my day inside those models. If you really want to innovate, [48:46] somebody has to pay for that. It's not just low-ing fruit that you can just go and get on the cheap. [48:51] And the difference is immense. [48:53] Evgeny, Elena, Robert, and James, I want to thank you for this very lively debate. [48:58] I've personally learned a lot and I hope that you've joined us on this conversation. [49:02] We would love to hear your thoughts, so please share them. [49:05] My name is Mohamed Hassan and this has been the Doha Debates Podcast.