WS #208 Democratising Access to AI with Open Source LLMs

Summary

This discussion focused on democratizing access to AI through open-source large language models (LLMs). Panelists explored how open-sourcing can influence innovation rates in the AI industry and prevent monopolization by large entities. They highlighted the potential of open-source LLMs to foster collaboration, address local needs, and empower smaller economies and the Global South.

Key points included the importance of truly open-source models that allow free use, modification, and redistribution. Panelists discussed the challenges of building open-source AI infrastructure, particularly for developing countries, including the need for computing power, technical expertise, and high-quality data. The discussion touched on initiatives in countries like the Dominican Republic and Brazil to develop localized AI models that reflect cultural nuances and languages.

Participants debated the role of regulation versus open-source approaches in addressing monopolies and ensuring equitable AI development. Some argued for hard regulation to manage competition and protect data sovereignty, while others emphasized the potential of open collaboration and shared resources.

The conversation also covered the risks associated with open-sourcing, such as potential misuse and reduced incentives for large-scale investments. Panelists stressed the need for governance structures, ethical considerations, and investment in local capacity building to mitigate these risks. The discussion concluded with calls for trust, collaboration, and a focus on inclusive AI development that serves the public good and represents diverse populations.

Keypoints

Major discussion points:

– The role of open source in democratizing access to AI and large language models

– Challenges and opportunities for developing countries in leveraging open source AI

– The need for infrastructure, computing power, and data to support open source AI development

– Concerns about monopolization of AI by large tech companies and how open source can help address this

– Cultural and linguistic representation in AI models, especially for underrepresented regions

The overall purpose of the discussion was to explore how open source approaches to AI and large language models can promote more equitable access and development of these technologies, especially for developing countries and underrepresented groups. The panelists aimed to highlight both the potential benefits and challenges of open source AI.

The tone of the discussion was generally optimistic about the potential of open source AI to democratize access, but also realistic about the significant challenges involved, especially for developing countries. There was a mix of idealism about open collaboration and pragmatism about the resources required. Toward the end, some panelists expressed a more cautious view about the need for regulation in addition to open source approaches.

Speakers

– Ihita Gangavarapu: Coordinator of India Youth IGF, works in cybersecurity domain in India

– Daniele Turra: Private Sector, Western European and Others Group (WEOG)

– Melissa Muñoz Suro: Director of Innovation at the Government Office of ICTs in the Dominican Republic, GRULAC

– Bianca Kremer: Civil Society, GRULAC

– Abraham Fifi Selby: Technical Community, African Group

Additional speakers:

– Yug Desai: Online moderator from South Asian University

– Purnima Tiwari: Rapporteur for the session

– Audience

Expanded Summary: Democratising Access to AI through Open-Source Large Language Models

This discussion explored the potential of open-source large language models (LLMs) to democratise access to artificial intelligence (AI), with a particular focus on fostering innovation and empowering smaller economies and the Global South. The panel, comprising experts from diverse backgrounds and regions, delved into the opportunities and challenges presented by open-source AI, as well as the implications for governance and regulation.

Understanding Open-Source AI

Daniele Turra of ISA Digital Consulting provided a foundational explanation of open-source software and its four freedoms: the freedom to use, study, modify, and redistribute the software. He emphasized that truly open-source AI models should adhere to these principles, allowing for free use, modification, and redistribution. This context set the stage for discussing the potential and challenges of open-source AI.

Benefits and Potential of Open-Source AI

The panelists broadly agreed on the positive impact of open-source AI on innovation and accessibility. Ihita Gangavarapu, coordinator of India Youth IGF, emphasized that open-source enables broader access and participation in AI development. Daniele Turra noted that open-source models can reduce costs and foster innovation. Melissa Muñoz Suro, Director of Innovation at the Government Office of ICTs in the Dominican Republic, highlighted the potential for customization to meet local needs and languages.

Abraham Fifi Selby, an expert in AI development in the Global South, argued that open-source approaches can level the playing field for regions with limited resources. He stressed the importance of multilingualism and local policy development in addressing African needs. Bianca Kremer, a researcher and activist from Brazil, added that open-source can help address biases in AI models, contributing to more inclusive and representative technologies.

Challenges and Limitations

Despite the optimism, significant challenges were acknowledged. Daniele Turra pointed out that substantial computing resources, such as GPU clusters, are still required to train large models, which can be a barrier for many organizations and regions. Melissa Muñoz Suro and Abraham Fifi Selby both highlighted the lack of infrastructure and expertise in developing countries as major hurdles.

Melissa Muñoz Suro drew attention to the ongoing costs of maintaining and scaling open-source systems. The need for high-quality local data to improve models was emphasized by Selby. These challenges underscore the complexity of implementing open-source AI solutions, especially in resource-constrained environments.

Governance and Regulation

The discussion revealed differing opinions on regulating open-source AI development. Daniele Turra stressed the need for clear definitions and licensing of truly open-source models. In contrast, Bianca Kremer called for hard regulation to address competition issues, suggesting a more interventionist approach.

Melissa Muñoz Suro emphasized the importance of data sovereignty and local control of AI systems. Abraham Fifi Selby proposed exploring public-private partnerships to support open AI development. Daniele Turra suggested a “computing tax” and partnerships with civil society organizations as potential governance structures.

Practical Applications and Cultural Context

Melissa Muñoz Suro shared insights about the development of ‘Taina’, an AI system in the Dominican Republic designed to reflect local culture and language. This project exemplifies the potential for open-source AI to be tailored to local needs while respecting cultural nuances. Melissa Muñoz Suro detailed how Taina was developed using open-source tools and local data to create a culturally relevant AI assistant.

Bianca Kremer provided examples from Brazil, including the Tucano and Maritaca AI projects, which demonstrate successful open-source AI development in Portuguese. She also highlighted the issue of algorithmic racism, using an example of how Chat GPT associated the term “favela” with negative connotations, underscoring the importance of addressing bias in AI models.

Abraham Fifi Selby offered perspective on the African context, highlighting how open-source AI systems are enabling young innovators to develop solutions at a lower cost, despite challenges in funding and infrastructure.

Audience Engagement and Unresolved Issues

The discussion included audience questions, particularly regarding competition and monopolies in AI development. This led to a broader conversation about balancing open collaboration with the need for regulation in AI governance.

Other unresolved issues highlighted include:

1. Effectively distributing computing power for open-source AI development

2. Ensuring cultural nuances from underrepresented regions are included in AI models

3. Creating sustainable funding mechanisms for open-source AI in developing countries

Conclusion

In their final remarks, panelists reiterated the transformative potential of open-source AI in democratizing access to technology and fostering innovation, particularly in developing regions. They emphasized the need for continued collaboration, investment in local capacity building, and addressing both technical and socio-economic challenges to realize the full potential of open-source AI for inclusive global development.

Ihita Gangavarapu: All right. Hi everyone. Good morning. Welcome to our session. I’m the coordinator of India Youth IGF and I also work in the cybersecurity domain back in India. So our session is titled democratizing access to AI with open source LLMs, large language models. It’s a 60 minute session where we have ample time for audience interaction. So when we talk about, we have, although before we start, we have a few speakers offline, but we do have a few speakers online, including our online moderator. So when we talk about democratizing access to AI, we are talking about making sure that artificial intelligence technologies and resources are accessible to a broad range of people, not just to the large corporations, governments, or the highly skilled participants. And the goal is to ensure that everybody is empowered, even the small businesses, educators, researchers, and organizations from all backgrounds, all economies, and they benefit from AI. The development and dissemination of AI, particularly the large language models, are increasingly dominated by major technology companies right now. And that does raise certain critical issues around access, control, and equity. Now with proprietary models that are accelerating innovation economic gain for some, they are also risking consolidation of power and limiting the technological diversity. So when we speak of open sourcing LLMs, we are looking at it creating a pathway to democratize AI, potentially reducing the costs and fostering innovation by enabling more and more stakeholders to participate in the development of AI. So today’s discussion is going to be focusing on the strategic, economic, and social implications of open sourcing. open sourcing AI, LLMs particularly, and the potential to counteract monopolistic controls and encourage a broader distribution of technological and economic benefits. So before I start, I’d like to introduce you to our panel. I am Ahitha, but I’m joined by Daniel A, who is working with ISA Digital Consulting. I also have Abraham from Payag. Online, we have Yug Desai, who’s the moderator, online moderator from the South Asian University. We have Purnima Tiwari, who is the rapporteur for our session, as well as our speaker, Melissa, joining us remotely, who’s working in the innovation cabinet, Dominican Republic. Thank you all for joining us, and I now start off with our discussion. Very first policy question is to Daniel A. How does open sourcing influence innovation rates within the AI industry? What are the long-term implications of open source AI on the structure of the tech industry itself?

Daniele Turra: Thank you so much, Ahitha, for presenting me today. I’m so glad to be here to discuss this very important topic. Everybody right now is talking about AI. Open source has been around for a very, very long time, and the narrative is being, in a way, of course, influenced by these large big tech giants that we have just mentioned. But open source has, in a way, a different history, especially free and open source software. So before getting into the specific industry implications, I would like to spend a minute to just introduce, once again, the concept of open source. And we can just start by saying that open source was a philosophy that was first, in a way, brought forward by Richard Salmon and other important scholars in the United States. States that believed that open source should mean, of course, sharing the code, but they also tightly related with the concept of freedom. So free, not as in beer, as they say, but free as in freedom. That is, freedom of speech, but especially the four freedoms that define the core ideas of open source. There are the freedom to use code, the freedom to study code, to redistribute it, and to modify it. So there should not be any large or small actor entitled to, in a way, own strong intellectual property on that code. And this, of course, is an idea that can benefit so many actors, from the smaller to the larger, but in a way, can enable others to join the industry as well. So when it comes to the AI context, I think I had a few slides. I don’t know if tech support can put them on. But we are talking about specific solutions and software that are always created by two different parts. So there is not only a general idea of open source software. We are talking about models and weights. So when you produce a model, there are mainly two files. One is about the weights, and one is about the model itself. So based on this, the Open Source Initiative published the free and open source idea that defines both models and weights to be fully open source. So anybody that can access a truly open source LLM has access to both the model and the weights that is also the result from the training of the data set used to train that model. Then on the other side of the spectrum at the very opposite side we have the idea of a fully closed model you know when you are just maybe accessing it through APIs or anything like that but it’s already in production it’s not something that can be in a way you know inspected modified or distributed right and in between models that have been defined as open weights where you know there are licensing where you are as a researcher maybe allowed to explore either the model or the weights again but are not really entitled to use it as for commercial purposes and of course most often you’re not even allowed to you know redistribute it itself so this of course creates situations in which not everyone can actually benefit from those those models and the again I would like to stress that the only definition that is truly compliant with open source as we free free and open source software as we know it is the one that embodies all four freedoms as defined by the thinkers of the free and open source thought so I don’t know again if we have lights I don’t know about my timing right now but again we can think next slide please yeah here you can see the the frameworks I was talking about different licenses and not all providers actually have the same models and the same licensing for the models they provide so in this other slide that is AI as a service stack. I would like to bring the focus again on the components that are needed to build an actual AI solution from end to end. So a few scholars traced some comparisons between the cloud computing model. And at each of these layers, open source software can always be employed. So we should also ask ourselves, are we actively as private entities or public entities and so on, are we really entitled to have something that is truly closed source, even if we are employing so, so many community efforts coming from the entire open source community? And so in a way, this can be some food for thought, thinking about the different steps that are in implementing actively AI solutions. And when it comes to actively industry impacts, we can also think of all the open source software that goes both at the AI software services training and fine tuning the models, down to the actual infrastructure that is needed to have the computing power to have those solutions actively built. Because in the end, as the last slide shows it, please change the slide to the last one. Um, last slide, please. OK, the last one. There is a supply chain that starts from data collection to data storage, data preparation, algorithm training, application development. And at each of these stages, the entire idea of having community-supporting solutions that are open source can be something that can benefit also the private sector. So again, this is an entire invitation to think in terms of who builds the software and all the different steps that are needed to get to it and how technologies that are already around there can actively help in achieving truly, truly open source models. Thank you.

Ihita Gangavarapu: Thank you so much for your points. I would actually bring in Melissa here, who’s joining us remotely, to answer the same question. How does open sourcing influence AI innovation in the entire industry? And what are the long-term implications? Melissa, over to you.

Melissa Muñoz Suro: OK, can you hear me well? Yes. Perfect. Good morning, everyone. So yeah, I’m going to start. My name is Melissa Munoz and I work as the Director of Innovation at the Government Office of ICTs here in the Dominican Republic. And basically what we do in OPTIC is using technology to improve lives and make everyday interactions with government more efficient, inclusive, and even more enjoyable. I wanted to answer this question, illustrating it with a case of what we are doing here in OPTIC. And one of the most exciting ways that we are doing this through our national AI strategy, in which a big part of that vision is Taina. That basically is an open source AI system that in the future will make the government services faster, smarter, and even more personal. That’s what we are trying to do. Taina isn’t ready yet. Right now we are focusing on laying the groundwork with a project called Ciudadania. Open source technology plays a key role in this project because in a word it opens the door for more collaboration and innovation. We are building a strong foundation for Taina by collecting and organizing the data that we need basically to make it work. And how does it work? Well, we are collecting data from existing government systems like PuntoGov that is in personal service points and from online service platform GovDo and the 462 point line. And these systems let us possibly gather insights of how citizens interact with public services. We have also set up specific interaction points where people can actively contribute to the data. Things like how they phrase requests, questions. And this isn’t about collecting personal information at all. It is about understanding the way Dominicans communicate so the AI reflects our culture and our language. And this is a collaboration between government, citizens, and local universities. The universities help us basically to ensure that the data is… accurate, well-structured, and aligned with privacy standards. What is interesting about it is how open-source doesn’t just fool innovation itself, but it also shapes the structure of the tech industry, especially in smaller economies like DDR. And by using open-source frameworks, what we are breaking is the dominoes of the big tech companies. Instead of relying on their tools, we are creating solutions tied to the Dominican Republic specific needs. For us, that means building systems that understand Dominican and Spanish, that is different to all Spanish, and reflect our culture, solve our local challenges, but the potential doesn’t stop there. Specifically, open-source means that other Spanish-speaking countries can learn from what we’re doing, and that’s what we’re trying to do, to escalate this regionally. And Ciudadanía could inspire similar projects in the region, fostering cooperation, creating a shared path towards a more inclusive AI development. And open-source isn’t just about influencing innovation rates, it is about basically fundamentally reshaping how technology serves people, that’s what we think in DDR. And our current work with Ciudadanía and our vision for TIE at the end shows how open-source principles can empower governments to, and also engage citizens, and create opportunities for smaller economies to thrive. I got to believe that technology should make life simpler, that’s ultimately happier, and open-source is a key tool to achieve this vision and create a more inclusive and accessible world, and people-centered tech industry. Thank you.

Ihita Gangavarapu: Thank you so much for your points, Melissa, you also highlighted on certain initiatives. I would now, first, I mean, before I hand over the question to Abraham, I actually would like to introduce Bianca, who’s joined us, thanks for joining, she’s from CTSFGV Brazil. So the same question applies to you, Abraham, and after which we would like to take a comment or a question from the audience, before we move on to the second policy question.

Abraham Fifi Selby: All right, thank you very much for the session, and I’m very happy to join this panel. I’m from the Global South, especially in the African context, so I will be speaking based on that context, and we understand how open source in terms of a large language models can help us. When we say the impact of innovation rates, I’m going to highlight some few points, then we discuss about see how it is going to help the African context, the Global South context. In terms of influencing innovation rates, what we see in Africa is that we have a democratization of AI development, which means that there is a very low cost in terms of using the open source system. In Africa, getting funding for startup researchers in terms of developing AI systems is very hard, and we don’t have large systems, large data centers, so investments have to go through before we get that. This open source AI system is helping young people to bring out innovation because they can tap on such systems at a very low cost or very low rate so that they can improve upon development. Let me also go into fostering collaborations. Basically, the world is changing, and everything also evolves around technology is now moving into AI. We cannot focus on advanced countries and also not look at the local countries, so at least open source systems is helping the people from the Global South to make sure that at least they also provide a source of data which can improve the other regions and which the other regions can also tap upon. In Africa now, there are some countries creating a policy document, so let’s say startups, business people who need to get information about which policy they are doing this kind of policy. I need this policy, which does this law, policy, regulating that. Now these AI systems, people are feeding in data on it. They’re connected to open source AI systems, AI tools, which also helps. So if someone is in Europe, want to get an idea of business or some regulation policies in Africa, because there is a collaboration between Africa and Europe, they can be able to get and also ACI and other American countries. Let’s also look at addressing local needs. This AI tools that we use, basically in terms of, let me come into multilingualism, we have languages that people may even seem not to understand because of the evolving in Africa. Now collaborating this AI tools in terms of addressing local needs means that there are some needs that it may be in the context of global South and Africa, which may not be a need in Europe. So whilst we are talking about this open source system, it helps us to understand that we need to localize some kind of documentation, some kind of data, which will help us in the global South context. Let me also move into assessing some kind of investment that we need in Africa. This large language models helps so much to understand the context that there is more investment in there, but there is not much investment in Africa. So Africa is really very happy that at least we can tap this open source AI system from advanced countries to also improve on our livelihood. Let me go to some long-term implementations, which will help the structure of the tech industry. I’m looking at the tech industry around the globe and also specifically moving into the African context. There is a growth of local AI ecosystem in Africa. Africa, because now we are now tapping. Despite all this in the policy implementation, ethics has been a very difficult time, because I know Europe and some other countries have developed their AI strategy documented. Africa is still struggling in that. So if there are some ethics that gives maybe global concerns, licensing people, helping people to tap into this open source, I think it’s a very good way that we can enhance the rate of innovation. And also, monopolization. We see big entities having the data source and everything about open AI instance. Every system is now tapping from them. They have the large language model. They have everything, like the child GPT and other stuff. These AI tools is moving, and Africa is lagging behind. Why? Because we feel that if governments have upper hand on these AI tools, they can develop on their own. But we need to also connect to other sources and other open. So we need more investments to the large language models. We need more collaboration on that. And it will also reduce the monopolization, because startups can also build their own AI models that can also support development in Africa context. And the last thing is that in capacity building. We see so many AI labs that I’m very even happy about it, where my other speakers also talk about in their respective countries. But the global South, where is it? We are not getting it that. We are not getting it that the schools, the academic, are not also moving. So we must also invest much in our academia, bringing capacity building of AI models within that context. And I will leave my other colleagues to talk. Then I will come back later with some other point.

Ihita Gangavarapu: Thank you. Great points, Abram. And I think Melissa, as well as you, have spoken about addressing the local needs with respect to cultural languages as well. What we’d like to do now is hand it over to you all. Within a minute, if you have. a comment or a question. We will of course also be having a Q&A round towards the end of the panel discussion, but is there any comment or a question, we’ll be happy to hear from you. Yes. Meanwhile, Yuk, if there’s any comment or a question in the chat or from the online participants, please let us know.

Audience: Thank you for this. My name is Lina and I work with the Council on Tech and Social Cohesion and Search for Common Ground, it’s a peace building organization. So we work in conflict affected contexts and we are trying to deploy AI to build trust and collaboration. And there’s two challenges that I see. So you just mentioned ethics. So we’re actually trying to build things on top of the commercial models. And then there’s still an ethical question about where the data goes and how much we can be sure whether or not that data is being used to train those models. And the second is, you just said Africa needs to build its own models, but the resources required to actually have built these models is because they had billions of dollars in investment, including from the Saudi government and many other places to make this an enterprise that will dominate the market and which will end up becoming a major revenue builder. So there’s something a little bit almost naive about this idea that we can compete. There’s no competition unless you regulate the monopolies here. Otherwise, there will be no. So it’s kind of two questions. Thank you.

Abraham Fifi Selby: you’d like to answer. Yeah, I agree with you 100%. There is no competition in terms of this. That’s why I was even emphasizing that we have to foster collaboration. Because we need the global north data, and the global north also need African data. Despite that, we also have to encourage our government and member states in global south to also have some investment, as you said, in the infrastructure. What I want to address is that the building of our localness is where I was much emphasizing that. Let’s say the global south cannot come and build the data sources that we need locally. So let’s say we have some language models in Africa, Swahili, let’s say Arabic, French, and other stuff, Portuguese. All these things address localness in some African country context. So if we Africans are not making it up to also build some context that can connect to the large open source systems, we may be lacking behind. And this is the way that ethics comes in, that we must copy from what the global north is doing and build upon on ourself. But we can always not rely on the global north from the global south perspective. We must understand that we need to also build our own data models that we can use it to share with the other continent that can improve our AI development and AI policy. So this is how I was addressing that context. And I hope I have answered your question well, because you made a very clear point based on the investment in ethics, investment in infrastructure. And I really agree with you, because Africa is lacking because we don’t have that infrastructure. and we must all rely on the Global North. But in relying on Global North, we must also contribute to the Global North perspective by providing data that address local needs so that everyone can tap in information when they need according to the AI development. Thank you.

Ihita Gangavarapu: Is it okay if we just pick the question in 15 minutes? We have, yeah, we have, we’d move on to the second set of the panel discussion and then we’ll pick it up again. Well, the question I wanna ask is directly related to the monopoly competition issue. So I’d- Please go ahead.

Audience: Okay, so I’m disagreeing with the last question. I thought the whole point of open source was that it was open and that you essentially are sharing. So if you’ve developed a model, you’ve made all that capital investment and if it’s open source, which I understand that meta is halfway there or maybe full, maybe the first presenter would like to comment on this. But the point is you can have access to that. You don’t have to have the investment, you can use it. And regulation presumes that they are monopolies and you’re going to regulate how they, I don’t know how they sell it, which to me does not distribute the knowledge, does not distribute the capability. It is much better to go open source than it is to go through regulation. So I’m confused about how you’re approaching this issue.

Ihita Gangavarapu: So our next question is actually on monopolies, but maybe Daniel, if you want to just keep it under a minute to address this and we can pick up the discussion in a few minutes again.

Daniele Turra: Sure, actually, I was about to introduce some of the points that might help in that sense in this following question. But in a way, I agree on the fact that open source is a tool, is a philosophy that can help. in not really not systematically regulating monopolies but for sure sharing the knowledge and giving the opportunities to other actors to you know again get the skills and not being you know blocked by specific intellectual properties on that. So this is a way to do it again and Meta is again as you said halfway there. They have Lama as an open weight model so it’s it’s also not always some flavors of it are not commercially available. They are available for for example researchers to use it but only in some contexts the models also from other providers are allowed to be used in a commercial context. But again let’s always think about the skills and the resources needed to build those models and if some actors are really you know should be entitled really put a fully open source definition on that because there is an entire supply chain and I would like to avoid some let’s say open washing of things. We need to you know really categorize things and call them with the right name. Hope that was a you know answer could you know answer some some of your doubts. Thank you.

Ihita Gangavarapu: All right thank you. Thank you so much for your question as well. I would now like to request you can you confirm you can hear me? Yes I can hear you. Perfect so you will be taking care of the second segment of the discussion so I hand it over to you.

Audience: Yeah before that I think there was a hand raised in the online audience so I would like to pass the floor to Raj Jahan, if he has anything to add on the first policy question related to innovation and open source. Raj, are you there? Okay, it seems he’s not able to speak right now, so let’s, let’s jump to the second question and we’ve already sort of gotten into it already. So, the second policy question is, in which ways can open source models prevent a few large entities from monopolizing the AI landscape and what governance structures could be necessary to manage this? And to answer this question, I first want to pass it on to Bianca, Bianca, if you can share your thoughts on this question.

Bianca Kremer: Hi, everybody hears me? First of all, I’d like to apologize for the delay and other procedures, we’re in workshops room one, so thank you so much for the invitation. I’m sorry for all the logistic trouble, but I’d like to introduce myself, I’m Bianca Kremer, I’m a researcher and also an activist from Brazil. I work with AI and law, especially in the topics of discrimination, specific topics of racial discrimination in Brazil. And for this panel, I have, I have been questioning or proposing three specific questions that could address the topic in a way we can be, this panel could be a food for thought for us to exchange a little bit on these topics. So unfortunately I lost the first part of Daniel’s presentation, but I observed that he could bring a good opportunity for us to understand the supply chain of LLM. And I have some observations about what is happening in Brazil from your perspective and as well as you, Abraham. So the first question is, I’d like to take some steps back so we can move forward on the topic. The first one is, what actually are we talking about when we talk about open source LLMs? Otherwise, if we don’t address these topics, we will have some misunderstandings on questions about these subjects. So we have the difference between the closed source LLMs and open source LLMs. This is the first question that we will address. The second one is, what qualifies as an open source LLM? This is actually really important for us to address the challenges we will face on this. And the third one is, where are specific concrete cases where we can find possibilities with open source LLMs? And after that, I will bring some experiences from Brazil that we have been experiencing on open source platforms. And in a way, we can address the competition problem as well, being developed by universities in our country. It’s hard to do this, but I think we have been addressing. So what are we talking about when we talk about these open source LLMs? We are talking about these large language models as AI models that they are publicly accessible. OK, I think we have been talking about it. But what it means is that the source code and training data are made available to the public. And what happens? It allows us, not only the developers, but also us researchers and organizations from civil society as well, to freely use, modify and improve it, like the models, in their own purpose. So, this is something we have been, when we talk about the ratio, gender bias, for example, when we have open sources models, we can have an opportunity for us not only to improve it but make it better in a way that companies and business models are not interested in achieving due to economic purposes. Okay? So, what differences, how do these open sources differ from the other counterparts? We have the closed source LLMs, they are developed and maintained by companies like, as we have been saying, OpenAI, Entropic, Google, and they are typically proprietary. What it means is that you cannot access the underlying architecture of data that model was trained on. The other open source LLM, on the other hand, they are models that are free to download, they are free to modify, and they are free to be adapted. So, these projects have been instrumental, actually, in making models available to the public in order to also, not only, but also address social problems that we have been facing in the development of these technologies in some certain societies. Since we have been talking about global south, for example, in Brazil, we had a case, a concrete case that I would like to share with you about a deputy called Renata Souza, which she wrote on Chat DPT, for example, the word favela. Favela is a community, a poor community in Brazil, I don’t know if it has been heard about before, but she wrote a black woman in favela, and what happened is that an image was generated about a black woman holding a gun, pointing up. She didn’t write anything about guns, but it happened. So, it was a case of what we have been studying. the last 10 years what we call algorithmic racism in platforms in Brazil, for example. It’s a case, concrete case about how these generative AI technologies have been developing. Not talking about the gender bias we all know, when you have been written two years ago, name 10 philosophers and they were all European and white men. And then you say there are no women and they are always white women, European or North American. Also always white. So these are some bias we have been facing in the usage of these platforms that open source, for example, could be open to address and also to modify the model and addressing topics of solving some problems of bias. Not all of them, because when you have algorithms you always have bias, but some of them. So this is something I don’t want to talk too much, but just to address the topic of what we have been talking about and after we can open for questions and things. But I would like to also exemplify with the cases of Tucano activity and also Maritaca AI. Tucano and Maritaca are two birds from Brazil. We have several birds in our region, especially Rio de Janeiro where I’m from. So they are both birds and these are projects from public universities in Brazil developing open source technologies and we have been very successful in developing these technologies in Portuguese. And this is the third part that I want to have my remarks so I can hear my other colleagues, but just to not only clarify but also make an imagery of what we have been talking about. I am from CTS, Center for Technology and Society University in Brazil, but also I am a member, a board member of the Brazilian Internet Steering Committee, a political party. BDS that has said that the internet is the place for Ù¾ Pulling makes a perfect thing. With the information from the participants it’s a political position. This year we have a forum held in Cape Verde Aprica and a lot of members from the community are here to talk about the importance of the internet in a broader community, in a broader perspective. And it’s even more dramatic for us because when you go to African countries, for example, they speak Creole among them others much more than Portuguese. So this is something I have been talking about. It’s also a matter of sovereignty. So I would like to thank you for the opportunity and I keep myself open for questions and to exchange with my colleagues. Yoke, over to you.

Ihita Gangavarapu: Yeah, maybe I’ll take it from here. I think we have time for one or two more questions and then we’ll take it over, Yoke, since we cannot hear you. So we, I would like to actually pose the same question to Daniel. And I want to understand in what ways can open-source models prevent a large entity from monopolising the AI landscape.

Daniele Turra: Thank you for the question. I think the answer is that technology licensing itself cannot really alone prevent a large activity from taking over in that sense. As I stressed earlier, I believe that we have to, in a way, protect the actual and correct definition of open source and software. But of course, software sharing practices and business can help, as one of the men here in the audience actually started to pointing to in terms of monopolies. So when we’re talking about licensing, I believe that open source is ideal, but also open parameters models can be a good way to achieve that sharing of knowledge in the larger ecosystem and be a big boost in this type of ecosystem, even for the global south. Some of these are not completely open source, but still an important role. But again, I would like to stress about the resources. I think that having, for example, a publicly managed infrastructure could be something that can help us. Just like the man in the audience actually started to pointing out, these large models are developed by big companies with private money, but it doesn’t mean that we cannot, in a way, benefit from those altogether. So not all businesses, especially SMEs, can employ these models. The global south is poor in terms of computing power and therefore does not have enough power to train these models. So the fundamental infrastructure there is lacking, and in that sense, this makes all these models not inclusive from the real beginning, because we are not including researchers and civil society organizations that could provide good input also in the sharing and the management of that infrastructure. So in that sense, we could also think of in terms of building the models and running the models in production. And in general, in both cases, I would say that… Now, one important thing that I would like to bring as a proposal is to have large cloud businesses that have the computing power offer this capacity for free when it comes to develop truly open source models that can be, in a way, published as open source and true open source. So, for example, the allocation of that computing power could be managed by a law, by, for example, paying, let’s call it a computing tax of some sort, or maybe partnership with some civil society organizations that work for coordinating the freedom of open source software production. Let’s think in terms, for example, the Eclipse Foundation or the Python Foundation. They supervise a lot of efforts in the open source community. I think we can do something very similar, including civil society organization in the production of open source. There is a lot of food for thought here. I’ve seen a few, I’ve raised a few eyebrows probably. But again, this is new for everyone. So I think if we get a look at how open source community works, we can get a few good inputs on how to develop the new open source model for the future. Thank you.

Ihita Gangavarapu: Very well answered, actually. Now, I’d like to request Melissa, who’s joining us online, for your comments on this question, if you could kindly keep it under three minutes, please. Thank you. I think we’re facing some issues. We can’t hear the online participants. Seems they cannot hear the online. Yeah, I think you’re all audible now. Please, Melissa, over to you. Can you hear me well? Yes. Okay, perfect.

Melissa Muñoz Suro: So basically, building on what I was mentioning earlier about our national AI strategy back in the Dominican Republic, one of its core principles basically is achieving technology and data sovereignty. This may ensure that the tools, systems, and data we create remain under national control, protecting both our public access and privacy for our citizens. And that’s why we choose to develop, in this case, our RELM, to not only anti-U.N. title from scratch, using our personal framework, but platforms like, for example, GPT-4 are robust. They came with significant recent dependency and data exposure. Open source allows us to design systems that align with our national priorities and values, ensuring independence and security in managing these technologies. Open source AI models can reduce reliance on external corporations, enabling nations to build systems tailored to the needs while fostering regional cooperation, mainly by using open frameworks which retain control over our tools and data, preventing external exploitation, ensuring that technology serves, actually, for public interest. However, open source is not without challenges. That’s something important I wanted to mention before. Building and developing these systems require more than just access to code. It demands robust infrastructure, technical expertise, high-quality data. And these are areas where developing countries like mine must focus to ensure successful implementation. One of the biggest challenges we’re facing right now with open source AI is having the right tools to make it work. These models need powerful computers to run what we call GPU clusters, and they don’t come cheap. For countries like the Dominican Republic, it’s hard to justify spending so much money on equipment when there are so many other priorities, like education, health, I don’t know, poverty. We like external services where the infrastructure is already handled for you. And with open source, we’re the ones who have to set it up, maintain it, and make sure it works. That’s something good to have in mind. Another big issue is getting the models to perform how we need them to. Open source models don’t come ready to solve… every problem, they are like a black canvas. So you need to put in the effort to fine tune them, to teach them how to understand specific techniques. And that takes time, expertise, and yes, more money that we, of course, don’t have in the global South as the one fellow was speaking before. And there is a blurring also with the data, a lot of data we have in the government systems, messy, it’s all scattered, and it’s not always useful. For a project in Surinam, for example, we have to work hard to clean it up, this data, and combine it with new information. We collect it from different government platforms. I was first in Surinam, and we have also set up places where citizens can share how they talk and ask questions so that they can build and truly understand our culture and language. Finally, the cost of keeping everything running. Open source sounds great, but because you are not paying someone else every time you use it, but the truth is, it’s still expensive to keep systems working over long-term. You have to upgrade the hardware, phase issues, make sure they can handle more users as it grows. And the reality is, open source AI isn’t something you just turn on and forget about it. It needs investment, planning, teamwork, and that’s why we are looking for partnership in the DR with other countries and trying to make it regional with international organizations too. And so we can share resources and make open source AI a solution that works for everyone. And that’s it on my part. Thank you.

Ihita Gangavarapu: So much, Melissa. All your points have been noted. I would now, given that we are a little short on time, we would like to open the floor for all of you for our interaction. But before that, I’d like to pose a question to all of you. What specific risks do you see open sourcing pose? And such as it could be the increased potential for misuse or reduced incentives for large scale investments in AI research. And how can these risks be mitigated? while still promoting open development and harnessing the opportunities. The floor is yours. Do we have anyone who would like to add a comment or a question? Yes, please.

Audience: Is it working now? Yes, perfect. Hi. Thank you very much for your panel and the interesting discussion that you were having. It’s not an answer to your question, but more of a comment or a question to you. I wanted to ask you more concretely about what are the enablers of open source LLM, so open source AI, because you were touching on the competition issue. And we see thus far from a market incentive, Facebook goes some extent towards open source, but they are still lacking. And they have little incentive to actually open the model fully to full reuse. Just for these large, big tech companies, there’s no real incentive to do it. And their models might have the ability to outperform open source models for a time. So we’re talking about this issue of smaller languages than English. So I would guess there would need to be some kind of common data sets, open data sets. Daniela was touching on the issue of how to distribute computing power. So I was wondering maybe if you could make concrete recommendations of what would we need to build, and how could we do so to actually set up a system in which open source LLM, so open source AI, can thrive. Thank you very much.

Ihita Gangavarapu: That’s a very detailed question. So maybe could we take one more comment or a question before we let the panelists answer them? And, Joerg, if you have anything from the online participants, feel free to unmute.

Audience: We had a question a while back, and I think this was related to what Abraham was saying. How do we ensure cultural nuances from Africa are included in these models, as well as sovereignty is maintained? That’s all I have online.

Ihita Gangavarapu: OK, OK, perfect. Thank you. I think for the second question, Abraham did answer quite a bit. But what I’ll do is, if we can have any comments on the first question. question, please, from the panel. Would you like to go?

Daniele Turra: Yeah, I’ll try to be very brief. So one key difference that we can see in open LLMs when it comes to their nature than other types of open source software is the fact that it needs lots of computing power to train. This is not the same for other products that are iteratively developed over years and years. And in a sense, can become much without specific compute power needed. But when it comes to training, we need GPU infrastructure. And that thing, again, doesn’t come cheap. So actual proposals, as I was saying, is, again, a way to make sure that who has the resources either be a private sector institution or a public sector research center, whatever it is. They have mechanisms in their governance systems to make sure they can allocate at least a percentage of that power to the development of those unrepresented, for example, communities or languages. I don’t know. I might have brought a few proposals earlier during my comments. But again, the general takeaway message of like you folks here to have is this. Let’s try to redistribute and better share that computing power.

Abraham Fifi Selby: OK. All right. Thank you. So I like the question that I’m asking about data sovereignty in Africa. The context is that in Africa now, we are now growing in terms of the digital landscape and economy. And data sovereignty is something that we cannot leave out because of how data is stored, data is collected, because there is a gaps related to the policies and regulations. the data And in terms of building in the global south context, there must be an established funding mechanism that can help in terms of grants, private or public partnerships, and also an investment that supports open AI or open source AI development within the global south by connecting researchers and innovators to the advanced countries on the global north, which can also bring that knowledge back to the global south for development. So this is what I would say. And it’s a very useful section that I really appreciate the experts and the questions that have been asked. And we can all build that together in that context.

Ihita Gangavarapu: Thank you very much. Bianca, over to you, please.

Bianca Kremer: Thank you. Very briefly, I’m less optimistic. I do believe that to address competition and the advance of the economics of these digital platforms as we are facing, we need hard law. So we need regulation. Of course, I’m from law. Maybe I’m biased. But in Brazil, we have just discussed the AI bill and the data protection laws for the last six years. And I do believe that when we have regulation on these topics and the participation of government on the free development and industrialization or deindustrialization of our countries for our participation in economy, if we don’t rely on regulation, hard regulation on these topics, we won’t move forward in our own development, not only as countries, but also as economic partners in the global south, for example. So this is why I’m not that optimistic. I do believe we need more enforcement in terms of legal participation on this process.

Ihita Gangavarapu: All right. Thank you, Bianca. Melissa, if you can hear us, your closing remark, please.

Melissa Muñoz Suro: Can you hear me well? Okay. Well, I think we should focus on trust and collaboration, as Bianca was saying, as the foundation for employee and I. This means prioritizing data ethics and being transparent about where the data comes from, how it is used, and ensuring it serves the public good at the end. And we also need to invest in local capacity. I think that’s the most important message that I can leave here on this panel. Our partnership with universities, research institutions, to develop talent and create a culturally relevant data set that truly represents the whole population, in the case of governments, for example, and also an invitation to invest in inter-regional collaboration to building resources and infrastructure to make AI accessible for all. Well, and inclusive AI should be at the center of our efforts, ensuring that technology works for people, builds trust, and attracts sustainable investment at the end. That’s my final thought. Thank you.

Ihita Gangavarapu: With this, we come to the end of the session. Thank you all for joining us. And when we talk about democratizing access to AI, there are a spectrum of concerns that come into play, and many of which our panelists have highlighted. So thank you very much for your inputs, and thank you all for joining us. And we hope that you carry forward these deliberations and come up with great recommendations in the halls of IGF and after. Thank you. Thank you. Thank you.

Agreement Points

Open source AI enables broader access and innovation

Ihita Gangavarapu

Daniele Turra

Melissa Muñoz Suro

Abraham Fifi Selby

Open source enables broader access and participation in AI development

Open source models can reduce costs and foster innovation

Open source allows customization for local needs and languages

Open source democratizes AI development in regions with limited resources

The speakers agree that open source AI models promote wider access to AI technologies, foster innovation, and allow for customization to meet local needs, particularly benefiting regions with limited resources.

Challenges in implementing open source AI in developing countries

Melissa Muñoz Suro

Abraham Fifi Selby

Lack of infrastructure and expertise in developing countries

Ongoing costs of maintaining and scaling open source systems

Need for high-quality local data to improve models

Both speakers highlight the challenges faced by developing countries in implementing open source AI, including limited infrastructure, lack of expertise, and the need for high-quality local data.

Similar Viewpoints

Both speakers emphasize the importance of investing in local capacity building and fostering regional collaboration to advance AI capabilities in developing regions.

Melissa Muñoz Suro

Abraham Fifi Selby

Invest in local capacity building and talent development

Foster regional collaboration to share resources

Both speakers stress the importance of addressing biases in AI models and ensuring cultural relevance and ethical considerations in AI development.

Bianca Kremer

Melissa Muñoz Suro

Open source can help address biases in AI models

Ensure cultural relevance and representation in datasets

Prioritize data ethics and transparency

Unexpected Consensus

Need for regulation in open source AI development

Bianca Kremer

Melissa Muñoz Suro

Call for hard regulation to address competition issues

Importance of data sovereignty and local control of AI systems

Despite the general focus on the benefits of open source AI, both speakers unexpectedly agree on the need for some form of regulation or control to ensure fair competition and data sovereignty.

Overall Assessment

Summary

The speakers generally agree on the benefits of open source AI in democratizing access, fostering innovation, and addressing local needs. They also recognize common challenges in implementing open source AI in developing regions, including infrastructure limitations and the need for local capacity building.

Consensus level

There is a moderate to high level of consensus among the speakers on the main benefits and challenges of open source AI. This consensus suggests a shared understanding of the potential of open source AI to address global inequalities in AI development, while also acknowledging the practical difficulties in implementation. The agreement on these points implies a need for collaborative efforts and targeted investments to fully realize the potential of open source AI, particularly in developing regions.

Different Viewpoints

Role of regulation in open source AI development

Daniele Turra

Bianca Kremer

Need for clear definitions and licensing of truly open source models

Call for hard regulation to address competition issues

While Daniele Turra emphasizes the importance of clear definitions and licensing for open source models, Bianca Kremer argues for strong legal regulation to address competition issues in the AI industry.

Unexpected Differences

Overall Assessment

summary

The main areas of disagreement revolve around the role of regulation, the approach to addressing resource limitations, and the balance between open source benefits and implementation challenges.

difference_level

The level of disagreement among the speakers is moderate. While there is general agreement on the potential benefits of open source AI, there are differing perspectives on how to implement and regulate it effectively. These differences highlight the complexity of democratizing AI access across diverse global contexts and the need for nuanced approaches that consider both technological and socio-economic factors.

Partial Agreements

All speakers agree on the potential of open source AI to democratize access, but they differ on how to address the challenges of limited resources and infrastructure in developing countries.

Daniele Turra

Melissa Muñoz Suro

Abraham Fifi Selby

Significant computing resources still required to train large models

Lack of infrastructure and expertise in developing countries

Open source democratizes AI development in regions with limited resources

Similar Viewpoints

Both speakers emphasize the importance of investing in local capacity building and fostering regional collaboration to advance AI capabilities in developing regions.

Melissa Muñoz Suro

Abraham Fifi Selby

Invest in local capacity building and talent development

Foster regional collaboration to share resources

Both speakers stress the importance of addressing biases in AI models and ensuring cultural relevance and ethical considerations in AI development.

Bianca Kremer

Melissa Muñoz Suro

Open source can help address biases in AI models

Ensure cultural relevance and representation in datasets

Prioritize data ethics and transparency

Key Takeaways

Open source AI models can democratize access and foster innovation, especially in developing regions

Open source enables customization for local needs and languages, helping address biases

Significant challenges remain around computing resources, infrastructure, and expertise for open source AI in developing countries

There is debate over whether regulation or open collaboration is the best path forward for inclusive AI development

Investing in local capacity building and regional collaboration is crucial for open source AI to benefit the Global South

Resolutions and Action Items

Invest in local capacity building and talent development for AI in developing countries

Foster regional and international collaboration to share AI resources and knowledge

Prioritize data ethics and transparency in AI development

Ensure cultural relevance and representation in AI training datasets

Unresolved Issues

How to effectively distribute computing power for open source AI development

How to balance open collaboration with the need for regulation in AI governance

How to ensure cultural nuances from underrepresented regions are included in AI models

How to create sustainable funding mechanisms for open source AI in developing countries

Suggested Compromises

Large tech companies could allocate a percentage of their computing power to develop AI for underrepresented communities

Combine open source collaboration with some level of government regulation and public-private partnerships

Develop shared, open datasets that include diverse cultural and linguistic information

Open source has, in a way, a different history, especially free and open source software. … There are the freedom to use code, the freedom to study code, to redistribute it, and to modify it.

speaker

Daniele Turra

reason

This comment provides important historical context and defines key principles of open source, setting the foundation for the discussion.

impact

It framed the conversation around the core values and goals of open source, influencing how participants approached the topic of open source AI models.

We are building a strong foundation for Taina by collecting and organizing the data that we need basically to make it work. … This isn’t about collecting personal information at all. It is about understanding the way Dominicans communicate so the AI reflects our culture and our language.

speaker

Melissa Muñoz Suro

reason

This comment highlights a concrete example of how open source AI can be tailored to local needs and cultural contexts.

impact

It shifted the discussion towards practical applications and challenges of implementing open source AI in specific cultural contexts, especially in developing countries.

In Africa, getting funding for startup researchers in terms of developing AI systems is very hard, and we don’t have large systems, large data centers, so investments have to go through before we get that. This open source AI system is helping young people to bring out innovation because they can tap on such systems at a very low cost or very low rate so that they can improve upon development.

speaker

Abraham Fifi Selby

reason

This comment brings attention to the unique challenges and opportunities that open source AI presents for developing regions.

impact

It broadened the conversation to include perspectives from the Global South and highlighted the potential of open source AI to democratize access to technology.

So, if you’ve developed a model, you’ve made all that capital investment and if it’s open source, which I understand that meta is halfway there or maybe full, maybe the first presenter would like to comment on this. But the point is you can have access to that. You don’t have to have the investment, you can use it.

speaker

Audience member

reason

This comment challenges the notion that open source necessarily requires massive investment from all parties and highlights the collaborative nature of open source.

impact

It sparked a discussion about the true nature of open source and how it can be leveraged even by those without significant resources.

Building and developing these systems require more than just access to code. It demands robust infrastructure, technical expertise, high-quality data. And these are areas where developing countries like mine must focus to ensure successful implementation.

speaker

Melissa Muñoz Suro

reason

This comment provides a reality check on the challenges of implementing open source AI, especially in developing countries.

impact

It deepened the discussion by highlighting the complexities beyond just having access to open source code, leading to a more nuanced understanding of what’s needed for successful implementation.

Overall Assessment

These key comments shaped the discussion by broadening its scope from theoretical principles of open source to practical challenges and opportunities in diverse global contexts. They highlighted the potential of open source AI to democratize access to technology while also acknowledging the significant hurdles, especially for developing countries. The discussion evolved from defining open source to exploring its real-world implications, cultural adaptations, and the need for supporting infrastructure and expertise. This led to a more comprehensive and nuanced dialogue about the role of open source in AI development globally.

How can we ensure cultural nuances from Africa are included in AI models while maintaining sovereignty?

speaker

Online participant (via Joerg)

explanation

This is important to ensure AI models are culturally relevant and don’t perpetuate biases against underrepresented groups.

What are the specific enablers of open source LLMs?

speaker

Audience member

explanation

Understanding these enablers is crucial for creating an environment where open source AI can thrive and compete with proprietary models.

How can we create common open datasets, especially for smaller languages?

speaker

Audience member

explanation

This is necessary to improve AI model performance for less-represented languages and cultures.

What concrete recommendations can be made for building systems to support open source AI?

speaker

Audience member

explanation

Practical guidance is needed to implement and support open source AI initiatives effectively.

How can computing power be distributed more equitably for AI development?

speaker

Daniele Turra

explanation

Addressing the disparity in access to computing resources is crucial for democratizing AI development.

What governance structures are necessary to manage open source models and prevent monopolization?

speaker

Ihita Gangavarapu (moderator)

explanation

This is important to ensure fair and equitable development and use of AI technologies.

How can we mitigate the risks posed by open sourcing AI, such as potential misuse or reduced incentives for large-scale investments?

speaker

Ihita Gangavarapu (moderator)

explanation

Addressing these risks is crucial for the responsible development and deployment of open source AI.

What funding mechanisms can be established to support open source AI development in the Global South?

speaker

Abraham Fifi Selby

explanation

This is necessary to ensure equitable participation in AI development from developing countries.