Democratizing AI: Open foundations and shared resources for global impact

Summary

This discussion focused on Switzerland’s initiatives to democratize access to AI research and implementation through international collaboration and open-source development. State Secretary Bernard Maissen opened by explaining Switzerland’s strategic position as a diplomatic bridge-builder in AI governance, highlighting the country’s role in developing the Council of Europe’s AI Convention and plans to host an AI summit in Geneva in 2027. He emphasized Switzerland’s approach of prioritizing quality over scale and fostering public-private collaboration while maintaining ethical AI development standards.

Professor Mennatallah El-Assady presented the Swiss-made LLM, announced as the largest truly open-source language model designed to serve society. This multilingual model, developed through the Swiss National AI Institute combining ETH and EPFL researchers, incorporates over 1,000 languages including Swiss minority languages and is designed to be legally compliant with Swiss law and EU AI Act requirements. The model will be released under an Apache 2 license, making it accessible for research, public, and private sector applications.

Professor Mary-Anne Hartley described ICANN (International Computing and AI Network), a decentralized initiative that provides global access to computing resources, data, and talent for AI research. ICANN operates like “CERN for AI,” focusing on three key areas: climate and agriculture, health and humanitarian action, and education. The network enables international collaboration by connecting researchers with the resources they need to address real-world challenges.

The discussion showcased concrete applications, including medical AI models like Meditron for healthcare applications and Legitron for international humanitarian law, both developed in partnership with the International Committee of the Red Cross. Educational initiatives include interdisciplinary courses combining technical expertise with domain knowledge to solve social challenges. The panelists concluded by calling for increased funding and collaboration to scale these impact-driven research initiatives, emphasizing that academic research can achieve measurable real-world impact when properly supported and networked internationally.

Keypoints

## Major Discussion Points:

– **Switzerland’s AI Diplomacy and International Cooperation**: Switzerland’s strategic positioning as a bridge-builder in AI governance, including hosting an AI summit in Geneva in 2027 and participating in international frameworks like the Council of Europe’s AI Convention and UN Global Digital Compact

– **Swiss-Made Open Source Language Model**: The development of a truly open-source, multilingual large language model by the Swiss National AI Institute (SNAI), designed to be legally compliant, accessible to all sectors, and respectful of data rights and minority languages

– **ICANN (International Computing and AI Network)**: A decentralized international network that democratizes access to AI resources including compute power, data, and talent, operating on principles similar to CERN but focused on AI for social good

– **Concrete Applications and Use Cases**: Real-world implementations across three pillars – climate/agriculture (weather prediction, plant disease detection), health/humanitarian (Meditron medical LLM, Legitron for international humanitarian law), and education (interdisciplinary courses and AI literacy games)

– **International Collaboration and Scaling**: Partnerships with organizations like AI Singapore and the ICRC, emphasizing the need for continued funding, networking, and measurable impact to scale these initiatives globally

## Overall Purpose:

The discussion aimed to showcase Switzerland’s approach to democratizing AI access through open-source models, international partnerships, and concrete applications for social good. The session served as both a presentation of current initiatives and a call for broader collaboration and investment in impact-driven AI research.

## Overall Tone:

The tone was consistently collaborative, optimistic, and forward-looking throughout the discussion. Speakers maintained an enthusiastic and inclusive approach, emphasizing partnership over competition. The conversation had a professional yet accessible quality, with participants actively inviting audience engagement and future collaboration. The tone remained positive from the opening diplomatic remarks through the technical presentations to the closing calls for action, reflecting a shared vision of AI as a tool for global benefit rather than competitive advantage.

Speakers

– **Katharina Frey**: Session moderator/host (also referred to as “Nina” in the transcript)

– **Bernard Maissen**: State Secretary, Switzerland – focuses on diplomatic and policy aspects of AI, international cooperation, and Switzerland’s AI initiatives

– **Mennatallah El-Assady**: Professor – AI researcher working on the Swiss-made LLM project, human-centered AI, and educational applications through the Swiss National AI Institute

– **Mary-Anne Hartley**: Professor – Principal Investigator (PI) for health and humanitarian AI applications, works on medical LLM (Meditron) and humanitarian LLM (Legitron) projects with ICRC

– **Leslie Teo**: Dr., AI Singapore – appeared via pre-recorded video message discussing collaboration between Swiss AI and AI Singapore’s Sea Lion team

Additional speakers:

None identified beyond the provided speakers names list.

# Switzerland’s AI Democratisation Initiative: International Collaboration for Open-Source AI Development

## Introduction and Switzerland’s Strategic Position

This discussion, moderated by Katharina Frey, brought together key stakeholders to present Switzerland’s comprehensive approach to democratising artificial intelligence through international collaboration and open-source development. The session featured State Secretary Bernard Maissen, Professor Mennatallah El-Assady from ETH Zurich, Professor Mary-Anne Hartley from the University of Zurich, and a pre-recorded message from Dr Leslie Teo of AI Singapore.

State Secretary Bernard Maissen opened by positioning Switzerland as a bridge-builder in international AI cooperation, emphasising that Switzerland’s approach prioritises quality over scale. He noted that Switzerland leverages its diplomatic capabilities and strong research infrastructure to facilitate international partnerships rather than competing in what he described as an intensifying AI race among nations.

Maissen highlighted Switzerland’s active participation in developing the Council of Europe’s AI Convention and contributions to the UN Global Digital Compact. He also mentioned Switzerland’s plan to host an AI summit in Geneva in 2027 to foster international dialogue. Switzerland’s philosophy focuses on public-private collaboration while maintaining rigorous ethical AI development standards, recognising that smaller nations can lead through innovation, quality, and diplomatic facilitation rather than scale alone.

## The Swiss-Made Large Language Model Initiative

Professor Mennatallah El-Assady presented the Swiss-made Large Language Model (LLM), describing it as the largest truly open-source language model designed specifically to serve society. Developed through the Swiss National AI Institute (SNAI), which combines researchers from ETH Zurich and EPFL, this initiative represents a fundamental commitment to AI democratisation.

El-Assady emphasised the crucial distinction between “open source” and “open weight” models. Unlike models that merely share final weights, the Swiss-made LLM provides complete documentation of all development steps, enabling anyone to replicate, modify, or host the model independently. This transparency extends to the entire development process, making it genuinely accessible for research, public sector, and private sector applications.

### Multilingual Capabilities and Technical Features

The model incorporates over 1,000 languages, including Swiss minority languages, addressing critical gaps in AI accessibility for diverse linguistic communities. This multilingual approach demonstrates practical benefits for international collaboration and cultural preservation.

The Swiss-made LLM is designed for legal compliance with both Swiss law and EU AI Act requirements, respecting all data sharing and opt-out requirements. The model will be released under an Apache 2 licence, providing maximum accessibility while maintaining appropriate attribution requirements. Both 70 billion parameter and smaller models are being developed to serve different computational needs.

## ICANN: International Computing and AI Network

Professor Mary-Anne Hartley introduced ICANN (International Computing and AI Network) as what she described as “CERN for AI” – a framework for democratising access to AI resources globally. ICANN addresses three critical bottlenecks in AI development: access to computing power, data, and talent.

The network operates on a decentralised model that connects researchers with needed resources while maintaining distributed ownership and control. Rather than creating a centralised authority, ICANN functions as a convening platform that channels competitive instincts constructively.

### Three Pillars of Application

ICANN focuses on three high-impact areas:

**Climate and Agriculture**: Applications include weather prediction systems and plant disease detection tools for agricultural communities globally.

**Health and Humanitarian Action**: Medical AI applications and humanitarian assistance tools with immediate, measurable impact on human welfare.

**Education**: Making AI knowledge accessible and collaborative across different user groups and contexts.

Hartley introduced the concept of “sovereign-ability” – the capacity for communities, organisations, or nations to own and customise AI models for their specific contexts while maintaining participation in collaborative frameworks.

## Practical Applications and Real-World Impact

### Medical AI: Meditron and MOOVE Platform

The discussion showcased concrete applications including Meditron, a medical AI model that incorporates MOOVE (M-O-O-V-E: massive, open, online validation and evaluation platform). This platform allows contextualisation and critical evaluation by end users, recognising that medical AI applications require careful validation and adaptation to local contexts and healthcare systems.

### Humanitarian AI: Legitron

Legitron focuses on international humanitarian law, providing conversational access to legal information about war crimes and humanitarian principles. Developed in partnership with the International Committee of the Red Cross (ICRC), Legitron embeds humanitarian principles of independence, transparency, neutrality, impartiality, and humanity directly into its design.

## International Collaboration Examples

### Swiss-Singapore Partnership

Dr Leslie Teo’s pre-recorded message described the collaboration between Swiss AI researchers and AI Singapore’s Sea Lion team. Teo emphasised that Swiss AI’s multilingual foundation models provide crucial contributions to the global AI ecosystem, enabling regional partners to build upon established capabilities for the 700 million population of Southeast Asia rather than starting from scratch.

This partnership demonstrates how foundation models developed in one context can be adapted and enhanced for different regional needs and linguistic requirements.

## Educational Initiatives

El-Assady described the Human-Centered AI for Social Good course at ETH Zurich, which combines four departments in interdisciplinary teams working on real-world challenges from climate, peace, and health domains. This educational approach emphasises interdisciplinary collaboration and real-world problem-solving rather than purely technical AI education.

Additionally, a collaboration with Data Science Africa will develop an educational card game designed to teach computational thinking and equitable AI principles. El-Assady mentioned working with professors Shira Miner from Kenya and Godliver Ovomugisha from Uganda on these educational initiatives, including concepts like an “AI driver’s license.”

## Infrastructure and Support

The Swiss supercomputer center (CSCS) provides crucial infrastructure support for these initiatives. Hartley emphasised that academic research can achieve measurable impact and scale when properly funded and supported, moving beyond traditional paper publication to real-world implementation.

## Future Directions and Call to Action

The speakers consistently emphasised the need for broader engagement and participation. They highlighted the importance of demonstrating measurable impact to secure continued funding and support, requiring metrics that evaluate real-world outcomes rather than just technical achievements.

Hartley specifically noted the challenge of competing with well-funded private sector initiatives, emphasising the need to support academic research that can match commercial impact while maintaining public benefit focus.

The speakers called for international participation in their initiatives, mentioning online forums and collaboration opportunities for submitting challenges and joining the network. They emphasised that scaling these efforts requires moving beyond the current core group of institutions and countries.

## Conclusion

This discussion presented a coordinated approach to AI democratisation that addresses technical, diplomatic, and practical dimensions simultaneously. Switzerland’s strategy demonstrates how smaller nations can lead in AI development through strategic positioning, international collaboration, and commitment to open principles.

The initiatives presented – the Swiss-made LLM, ICANN network, and various application projects – represent concrete steps toward making AI capabilities more accessible globally while maintaining high standards for ethical development and practical utility. The emphasis on measurable impact and practical applications addresses crucial gaps in AI development, demonstrating that democratised AI access can produce tangible social benefits across healthcare, humanitarian work, education, and climate action.

The speakers’ call for broader engagement and their provision of concrete mechanisms for participation suggest recognition that success depends on expanded international collaboration, continued funding, and moving from proof-of-concept to scaled implementation.

Katharina Frey: Hello, everyone, and thank you so much for the introduction and already announcing everyone and take some work away from me. Thank you for coming here and joining this session on democratizing access to AI. So, I would like to stop immediately into the show and hand over to State Secretary Maissen. Please, Bernhard, could you share some view of why Switzerland thinks democratizing access to AI research and implementation is important?

Bernard Maissen: Yes, thank you. Hello, everybody, dear panelists. Nina, thank you for giving me the floor. In the global technological and economic competition, countries have a strategic interest in developing sovereign AI capacities while also seeking international partnerships. Switzerland is well placed as a small country with diplomatic acumen and strong research and innovation capabilities. On the diplomatic side, Switzerland played an important role in the making of the Council of Europe’s Convention on AI, Human Rights, Democracy, and the Rule of Law, and is currently active at the United Nations level to implement the Global Digital Compact provisions on AI. Since late 2023, numerous high-level AI summits have convened across different nations reflecting the urgency of establishing collaborative frameworks for AI development and regulation. However, mounting geopolitical tensions and intensifying a perception of AI race among leading nations threaten to deepen global polarization. Against the backdrop, political summits focused on AI cooperation become vital for fostering dialogue and preventing fragmentation. Recognizing this critical need, Switzerland will host an AI summit in Geneva in 2027. We are uniquely positioned to facilitate meaningful international cooperation due to our established role as bridge builder in the international community. Yet, all this multilateral talk is not enough if we cannot deliver on cutting-edge sovereign tech through innovative partnerships thanks to our research institutions. This is why we are gathered to be here today. Switzerland aims to set, of course, for AI to prioritization, quality over scale, public-private collaboration, and ethical AI development. Two initiatives will be presented from a Swiss perspective. We will also demonstrate in collaboration with our partners from AI Singapore how dimensioned collaboration can work in practice. The Swiss National AI Institute has the goal to provide a national perspective on AI-based research, education, and innovation. Currently, researchers from ETH and EPFL are working towards a Swiss language foundation model focused on an open and trustworthy AI. The goal is to deploy models in core areas such as healthcare, sustainability, science, and beyond. True to its long-standing humanitarian tradition, the Federal Department of Foreign Affairs, ETH, and EPFL, and international partners launched ICANN, an initiative to create a global infrastructure for accessible compute on AI collaboration. With existing use cases and the institutionalization of the organization, we are seeking partnerships with a wide variety of stakeholders. ICANN aims to ensure that as many voices as possible are heard in AI research and that AI solutions for real-world challenges are perceived to serve global public goods accessible worldwide, for instance, through equitable access to supercomputers. I’m now leaving you to the hands of our panelists and Nina, and I hope we can find ways to work together to ensure that AI is beneficial for all. Thank you very much.

Katharina Frey: Thank you so much, State Secretary, for already laying a bit the ground on exactly these two concrete initiatives, and I think I would like to use the time now to dig a bit deeper. You had mentioned the Swiss-made LLM. I think that’s what it’s called. It was announced yesterday in the media. Professor Mennatallah El-Assady, please tell us a bit more about what’s this Swiss LLM, what you can already say, because I think it’s not public yet, but what can you already share with us, and how can it be the ground layer for more access to AI research?

Mennatallah El-Assady: Thank you, Nina. So, we’re all very excited about the news that came out yesterday about the Swiss-made LLM. This is coming out of the Swiss National AI Institute, SNAI, which basically combines faculty members from ETH and EPFL, and also the Swiss AI Initiative that brings in a lot of different people in a holistic ecosystem in Switzerland. The Swiss-made LLM is an effort from these researchers to try to produce an equitable and accessible language model, and the one that we have worked on is right now the largest truly open language model that is meant for serving society. So, we worked with different principles to ensure that we can share it with the world, and we can share it across different types of communities, whether it’s for the research community, whether it’s for the public and private sectors, or for people who are trying to work on different technologies through this open source principle. The model itself will be open source, not just open weight, and we have documented all of the steps that we have taken, so anyone can take what we have done and replicate it or host it themselves. We have worked on it to be multilingual by design, so we have integrated more than a thousand languages and focused also on integrating maybe underrepresented languages like minority languages in Switzerland as well. We’ve also tried to be legally compliant by principle, so we were looking at data that would basically be suited to train this language model by reading in all of the requirements of how data can be shared, so all of the opt-outs and so on were respected. So, this LLM is designed to respect all of the legal requirements, whether it’s by Swiss law or the EU AI Act, and we will be sharing not just one model but different size models, so a 70 billion parameter model and another smaller one, and the idea here is to allow this to be an effort that basically is accessible to a lot of different frontiers and a lot of different applications, so people can take that and instead of expensively training new models, they can take it, use it, and replicate it. It will be available with an open source license, an Apache 2 license, so it will be accessible for everyone who wants to use this in that open ecosystem. This whole effort is only possible because of the Swiss supercomputer center, so the CSCS has invested in infrastructure way before the hype, so there’s a long-standing effort to actually invest in infrastructure, and there’s also strategic funding coming from the ETH board that allowed us to actually work on this, so it’s a very collaborative, very joint environment in which we’re developing this with the hope to enable applications that would serve society, not just in Switzerland but across the world.

Katharina Frey: Thank you so much, and I think we’re going to dip a bit more into the use cases afterwards, but I would like to first see, now we have this model and we have this collaboration that you mentioned, but Professor Hartley, could you also describe us, I think Ben mentioned before, ICANN, which is kind of the international layer of this Initiative. Could you describe that a bit further? I have to click, I think. Strongly, they told me. Voila. Strongly. There it is. So, Annie, could you please describe us, ICANN, what it is? Maybe it’s also interesting for the audience here how they could join. Please give us a bit the framework and then we will dig into the use cases later.

Mary-Anne Hartley: Yeah, sure. I think what we all saw with the use case over there is what is possible when people work together and when they are convened on purpose, right? And with a specific purpose in mind. It’s so important. And to formalize that around these resources that we do have in Switzerland, we’ve made ICANN, which stands for the International Computing and AI Network. Now, while it has been convened in Switzerland, it is intentionally international. It has a principle of decentralization, democratization of the resources that we have available, often in higher resource settings. And to convene, as you can see here in blue, the mission, to convene people working together and to broaden access to these really essential resources, such as massive scale compute. It’s a very big deal what the Swiss government has done here, but not just in Switzerland, but around the world. And you’ll see that that’s the important word up there is network, where you can see the different partners that we’ve already convened around this, who are also donating their compute to this cause. And data, massive scale data. This data is important to curate. AI is data. We can’t just compute thin air. So it’s important that we convene on data that is purposeful and meaningful and directed. And indeed, we can’t do anything with data and compute if we don’t have the talent. And not just the talent of making models, but the talent of understanding where those models can be used and can be impactful. And the most important thing that we’re convening around is that we’re convening around, as I said before, a purpose. And that purpose, just like kind of CERN for AI, you can think of it like that. CERN over here was very, very impactful to look into, let’s say, the black holes and molecular kind of like particle physics, right? And instead of us doing that, we’re rather looking into, let’s say, black boxes, but really going deeper and having measurable impact as our focus. And here you can see the focus of the goals here in green, where we’ve got, again, these three major foci that are high impact and that allow us to not fragment our efforts. Because, you know, those things are, those things, people want those things. People want compute. They want data. And humans are humans and we do compete. And it’s important to convene people so that we dilute that ability of competition, because this actually can be quite destructive sometimes, and to go beyond our individual egos and to really come together to make something bigger. CERN was just so impactful with that. And so to convene people around these goals of climate and agriculture, for example, huge projects on weather with the data sets that we have here, to convene us around health and humanitarian action. Those are some models that I can speak about. That’s my kind of focus. I’m working on these projects over here for large language models for medicine and humanitarian action. And here, Amanda can speak more about later, maybe about the education poll. And these are high impact areas. Convening people, you often call it a tinder for impact, right? And if only there were more of those. So it’s about making sure that people who know where the impact is supposed to go are convened with the resources that they need. Now, I did use the word donate earlier, right? They donated their resources, donated this, but it’s not just donations. It obviously costs money to run and to employ people, and a lot of people are doing this out of good will. But if we really want this to be impactful, it’s important that we also convene around common funding on impact-based outcomes, and that we can show that impact is worth investing in, and measure it so that we can actually have a return. And I think those are the real values of ICANN, and I really thank the vision that was in Switzerland for convening and hosting this. But again, it’s decentralized, and that relinquishment of power to make sure that you have access is actually very extraordinary. So yeah, thank you.

Katharina Frey: Thank you so much, Annie. And I would like to now go to Singapore. So Dr. Leslie, oh, I managed in once. Dr. Leslie Teo, he couldn’t be here in person, but he pre-recorded the message, and I think it was important for us to already showcase the concrete application and collaboration that has been mentioned by my colleagues before of this Swiss-made LLM with AI Singapore’s work. So if I could ask the colleagues to show the very short video message by Dr. Leslie Teo from AI Singapore, please.

Leslie Teo: Swiss AI’s development of state-of-the-art open foundation models that are multilingual represents a key contribution to the global AI ecosystem. The Sea Lion team is proud to collaborate with Swiss AI to leverage on these open models to build more inclusive AI for Southeast Asia with its 700 million population. This partnership demonstrates how we can all work together to address diverse needs and different communities and languages across the globe.

Katharina Frey: Thank you so much. I hope it already gave you some ideas about this really international collaboration and concrete impact. I would like to use the last 15 minutes that we still have to to talk a bit more about concrete use cases. I think it would be helpful also to the audience if you could dig a bit deeper. You mentioned the three thematic pillars of climate and agriculture, health and humanitarian, and education, and I would like to invite you, probably first Annie, to present some, because you’re also the PI of two of the health and humanitarian pillar use cases. So maybe could you share with us your work on medical LLM and also humanitarian LLM. And before I hand you over the mic, I would just also like to honor two other professors who work in the climate and agricultural pillar. They couldn’t be here today, but it’s Professor Shira Miner from Kenya and he works on a weather prediction model for Kenya and hopefully also Eastern Africa, and Professor Godliver Ovomugisha from Busitema University in Uganda. She works on plant disease detection and I hope that maybe next year we could invite them to present their work. But let’s now dig into medical LLM and ICRC LLM. Annie, please.

Mary-Anne Hartley: Yeah, you know, the message from Dr Leslie there was emblematic of how we are synergizing, right? We make this huge effort of this big model that costs a lot of money to make, a lot of expertise, data, effort, and instead of reinventing the wheel and doing another model and another model, we can take that, it’s purposefully generic, and build on top of it and to make sure that we’re synergizing our efforts and not fragmenting them. Same thing is what we’re doing here. A lot of people are speaking about sovereign AI and this kind of sovereign AI can seem quite selfish, right? It’s kind of opposite to the purpose, but actually it’s sovereign-able, right? It’s how you can make these kinds of models your own and own them yourself and that sovereign-ability is really the spirit of what decentralization we’re trying to achieve. So taking that model, for instance, and now allowing people to, we build on it, we’ve got our own kind of pipeline that’s purposefully generic to medicalize models, right? So if there’s a new base model like this, we can just kind of sprinkle on top this pipeline that allows us to make it really medically focused. So it’s really focused on that and we make sure that we can put it on any base model so that people can own their own base model and medicalize it. So our model is called Meditron and it’s basically designed for people to use it and validate it in their setting. So what’s really important for us again with this principle of sovereign-ability is also contextualization, allowing them to critically engage. We have a platform that goes with it, right, called Move, so that they can move their models closer to their own settings. It’s a M-O-O-V-E, so it’s like massive, open, online validation and evaluation platform and allows us to take any model, critically evaluate it, and engage the end users who will actually be using it so that they can see where it fails, see how their colleagues see where it fails, and build something that is really, really important for AI safety, which is vigilance. And this kind of vigilance can then be fed back into the model and move it closer to the cultural and contextual norms of where those models are going to be used. So moving these models will allow people to own them themselves, which is again this principle of decentralization and sovereign-ability. And then we have another one that we are working on, both of these are with the ICRC, the International Committee for for the Red Cross, and their guidance is just so important about ensuring humanitarian principles are embedded into these models. Independence and transparency, neutrality and impartiality and humanity, of course, are the principles. But this kind of baking it into the model, in the way it’s designed, is important to them and it puts a pressure on how we design these models as well. So putting all of those things together, we made another model where we had Meditron for medicine, right, and we had thousands of doctors around the world are now moving their models toward their context. We also have another one for international humanitarian law, or IHL. It’s the law that dictates war crimes, for example, and it’s something that’s really important for us to understand how models understand this law, but also for us to have a more conversational interface with this kind of tool and to hopefully have better access and validated access to this kind of information, and it’s called Legitron. And now lawyers around the world, convened also by ICRC, will be testing this and exploring how models understand this very important body of information and if it can be used in practice as well.

Katharina Frey: Thank you. I think that explained a bit the concrete applications, and I think one of the other pillars that that’s from IECN is also, I’m sorry I’m losing my voice, is the educational pillar. So Menna, could you give some concrete examples? I’m sorry for my voice.

Mennatallah El-Assady: For the education pillar, I wanted to highlight maybe two different initiatives. One that we started a long time ago where we are collecting different challenges, and this last semester we ran the first iteration of a course at ETH called Human-Centered AI for Social Good, and we focused on climate, peace, and health applications. The way this course was set up is pretty unique. So on the one hand, all of the partners who kind of came to us with different challenges, whether it’s the ICRC, methodological organizations, and others, we kind of assembled all of this and tried to kind of form them into course projects. The other unique aspect is the interdisciplinarity of the way we set it up. So we collaborated across four different departments with professors coming from different perspectives and put together teams of technical people, of domain experts, and others in order to work in interdisciplinary fashion as teams of students working together to solve one of these challenges through human centered design principles. And we produced different applications and different models in order to figure out how we can address these challenges with all of these principles that we can actually assemble from research. So for example, we have models that build on climate modeling where we developed interfaces for counterfactual explainability or trying to answer questions like what if this parameter would be different, or what if my own understanding would form like this and that. We have also other collaborations where we are looking at peacekeeping scenarios and how we can actually address decision-making in these types of very complex tasks, or models where we are looking for the help of AI in healthcare and how we can adjust and personalize the interfaces and the interaction and co-adaptation of them through the interaction. So we just concluded this course. We are very open to collaboration and we’re very welcoming to anyone who is interested to work on us on the next iteration of this course. You’ll find all of the prototypes online at our website at ETH, so if you want to play around with them you’re also welcome to do so. And then the other example that I want to highlight very briefly because it’s just in its starting phase is a collaboration that we will start through Iocane with Data Science Africa. We’ll be working on an educational game to really teach the base principles of what it means to think in a computational manner and to use AI in an equitable and explainable and trustworthy way. So our idea is to develop a card game. We would be working together with researchers from Data Science Africa and also with users across different places, for example here in international Geneva but also in other places, to just try to tailor that to different use cases and user groups. So I’m very excited about this starting soon and we, as I said, we’re very open to also join other initiatives with the effort that we’re doing to make it collaborative and to make it equitable and available for everyone for a social good purpose.

Katharina Frey: Thank you so much. I think you, oh now they pulled out my mic. I’m good again so you can leave it. Thank you. I think you use it, you call it the driver’s license, AI driver’s license. I like the title. So we’re already at the end and I hope you portrayed the message and I hope you received it, Jan, and received it that it’s really a joint effort and we’re open for other joiners to this very different use cases. Could you, because we’ve been here last year and I hope we can be here next year again, so if you could just share like in one, two sentences of what you think, what actions are needed like now so we can scale maybe and present more impactful use cases, quantitatively more impactful use cases next year. Annie, maybe?

Mary-Anne Hartley: Yeah, so just a side note, if you want to see the impact of iCane already, I’m really excited that I’m going to be working with Mennatallah El-Assady and we’re collaborating now because of this network and it’s just so much like the powers combined of all of the people in her lab, all of the people in our lab. We can reach scale and for people to recognize that academia and research can scale and there isn’t this kind of like brush that they tarnish us with where they’re just going to publish papers and they don’t care about implementation. Implementation is science. We should measure impact. There is impact-driven science and for us to value that and to understand that it does happen and you’re not just investing in papers that are published and filling holes in textbooks, that we can and we have direct and concrete and measurable impact and for us to communicate that, I think the next most important step really is to ensure that people who are funding the three guys in the garage that are making the next big LLM, that they understand the power that we can have with our huge network of just wonderful talent, the students, the resources, the data, the power and the respect for also measuring evidence and basing it on that. I think if we can communicate that well and we can get funded to the extent that we are able to scale like that, that’s the only thing that’s really stopping us and I think that’s something that we need to advocate for and hopefully I can feel the tide is changing and hopefully maybe there’s someone in the audience who’s feeling generous. Ah, there we go, we have someone, okay.

Katharina Frey: Fantastic, thank you. Mena, what’s your wish for a call to action?

Mennatallah El-Assady: I can only echo what Annie just said and maybe add to it that we are really, really looking forward to work together and collaborate. Our websites are available, are open, we have forums online where anyone can go in and submit challenges that you would like human-AI partnership to solve. We are open to figuring out how we can synergize with other ecosystems. We just described the very vibrant Swiss ecosystem, the international collaborations that we’ve already established, but I think there’s a lot more opportunity to network and I think it’s really crucial to translate the research that we are doing, the human-centered AI research, the collaborative and co-adaptive work that we are producing from these two institutions and others into practice and to make that available through this giant network that we have. So I’m really, really excited about the next year. Hopefully we’ll be back here presenting other use cases and maybe counting impactful and measurable social good practices that we can actually present as use cases. So I’m really looking forward to that in the next year.

Katharina Frey: Thank you. Bernard, thank you first of all to Switzerland to enabling this initiative, I think that’s important. You opened the floor, I’d like to hand over to you so you can close the floor. Thank you very much.

Bernard Maissen: Well, I think it was very important and interesting to have you here. We’re on the political level, we discuss about principles and now we saw what means to live these principles, concrete examples, and we are really on the good way and I’m looking forward to see you next year and have more examples how it works and we are really looking on the political level to create a framework who lets this collaboration really be alive and so thank you very much. Thank you so much.

Katharina Frey: Thank you so much my panelists for taking the time. Thank you audience to listening and I hope it inspired you. Please reach out. Thank you. Many thanks

Agreement points

International Collaboration and Partnership Models

Speakers

– Bernard Maissen
– Mennatallah El-Assady
– Mary-Anne Hartley
– Leslie Teo
– Katharina Frey

Arguments

Switzerland’s Strategic Role in AI Democratization

Swiss-Made Open Source Language Model Development

International Computing and AI Network (ICANN) Framework

International Collaboration and Partnership Models

Facilitating International AI Collaboration Through Structured Dialogue

Summary

All speakers emphasize the critical importance of international collaboration in AI development, with Switzerland positioned as a bridge-builder facilitating partnerships across nations and institutions. They advocate for structured frameworks that enable resource sharing and joint development efforts.

Topics

Development | Sociocultural

Open Source and Democratized Access to AI

Speakers

– Mennatallah El-Assady
– Mary-Anne Hartley
– Leslie Teo

Arguments

Swiss-Made Open Source Language Model Development

International Computing and AI Network (ICANN) Framework

International Collaboration and Partnership Models

Summary

Speakers consistently advocate for open source approaches to AI development, emphasizing transparency, accessibility, and the ability for diverse communities to build upon and customize AI models for their specific needs and contexts.

Topics

Development | Sociocultural

Focus on Real-World Impact and Practical Applications

Speakers

– Mennatallah El-Assady
– Mary-Anne Hartley
– Katharina Frey

Arguments

Educational Initiatives and Capacity Building

Practical Applications in Healthcare and Humanitarian Work

Scaling and Future Development Needs

Promoting Concrete Use Cases and Practical Applications

Summary

All speakers emphasize moving beyond theoretical discussions to demonstrate tangible, measurable impact across sectors like healthcare, education, climate, and humanitarian work, with focus on evidence-based outcomes.

Topics

Development | Sociocultural | Human rights

Similar viewpoints

Both emphasize Switzerland’s unique diplomatic position and the importance of creating structured platforms for international AI cooperation, with Switzerland serving as a neutral convener and bridge-builder.

Speakers

– Bernard Maissen
– Katharina Frey

Arguments

Switzerland’s Strategic Role in AI Democratization

Facilitating International AI Collaboration Through Structured Dialogue

Topics

Development | Sociocultural

Both advocate for academic research that achieves measurable real-world impact, emphasizing interdisciplinary collaboration and the need for proper funding to scale research beyond traditional academic publishing.

Speakers

– Mennatallah El-Assady
– Mary-Anne Hartley

Arguments

Educational Initiatives and Capacity Building

Scaling and Future Development Needs

Topics

Development | Sociocultural

Both emphasize the power of leveraging shared resources and models to build inclusive AI solutions for diverse global communities, demonstrating how international partnerships can address regional needs effectively.

Speakers

– Mary-Anne Hartley
– Leslie Teo

Arguments

International Computing and AI Network (ICANN) Framework

International Collaboration and Partnership Models

Topics

Development | Sociocultural

Unexpected consensus

Academic Research as Scalable Implementation Vehicle

Speakers

– Mary-Anne Hartley
– Mennatallah El-Assady

Arguments

Scaling and Future Development Needs

Educational Initiatives and Capacity Building

Explanation

There’s unexpected consensus that academic institutions can and should be primary drivers of scalable AI implementation, challenging the typical view that academia only produces theoretical research. Both speakers advocate for impact-driven science with measurable outcomes.

Topics

Development | Economic | Sociocultural

Sovereign AI as Enablement Rather Than Isolation

Speakers

– Mary-Anne Hartley
– Bernard Maissen
– Leslie Teo

Arguments

Practical Applications in Healthcare and Humanitarian Work

Switzerland’s Strategic Role in AI Democratization

International Collaboration and Partnership Models

Explanation

Unexpected consensus emerges around redefining ‘sovereign AI’ not as nationalistic isolation but as ‘sovereign-ability’ – the capacity for communities to own and customize AI models for their contexts while maintaining international collaboration.

Topics

Development | Legal and regulatory | Sociocultural

Overall assessment

Summary

The speakers demonstrate remarkable consensus across multiple dimensions: the necessity of international collaboration, the importance of open source approaches, the focus on practical applications with measurable impact, and the role of academic institutions in scaling AI solutions. Key areas of agreement include Switzerland’s diplomatic role, the value of democratized AI access, and the need for evidence-based outcomes.

Consensus level

Very high level of consensus with complementary rather than competing viewpoints. This strong alignment suggests a mature, collaborative approach to AI democratization that could facilitate effective implementation of the discussed initiatives. The consensus spans technical, diplomatic, and practical dimensions, indicating a well-coordinated strategy for international AI cooperation.

Different viewpoints

Unexpected differences

Overall assessment

Summary

The discussion showed remarkable consensus among all speakers on the fundamental goals of democratizing AI access, international cooperation, and developing ethical AI solutions. No significant disagreements were identified.

Disagreement level

Very low disagreement level. All speakers were aligned on core principles and complemented each other’s perspectives rather than challenging them. This high level of consensus suggests strong foundational agreement on AI democratization approaches, though it may also indicate a need for more diverse viewpoints to surface potential challenges or alternative strategies.

Partial agreements

Similar viewpoints

Both emphasize Switzerland’s unique diplomatic position and the importance of creating structured platforms for international AI cooperation, with Switzerland serving as a neutral convener and bridge-builder.

Speakers

– Bernard Maissen
– Katharina Frey

Arguments

Switzerland’s Strategic Role in AI Democratization

Facilitating International AI Collaboration Through Structured Dialogue

Topics

Development | Sociocultural

Both advocate for academic research that achieves measurable real-world impact, emphasizing interdisciplinary collaboration and the need for proper funding to scale research beyond traditional academic publishing.

Speakers

– Mennatallah El-Assady
– Mary-Anne Hartley

Arguments

Educational Initiatives and Capacity Building

Scaling and Future Development Needs

Topics

Development | Sociocultural

Both emphasize the power of leveraging shared resources and models to build inclusive AI solutions for diverse global communities, demonstrating how international partnerships can address regional needs effectively.

Speakers

– Mary-Anne Hartley
– Leslie Teo

Arguments

International Computing and AI Network (ICANN) Framework

International Collaboration and Partnership Models

Topics

Development | Sociocultural

Key takeaways

Switzerland is positioning itself as a bridge-builder for international AI cooperation, leveraging its diplomatic capabilities and research institutions to democratize AI access

The Swiss-made LLM represents a breakthrough as the largest truly open-source language model, designed with multilingual capabilities (1000+ languages) and legal compliance with Swiss and EU regulations

ICANN (International Computing and AI Network) provides a decentralized framework for sharing compute resources, data, and talent globally, operating like ‘CERN for AI’

Successful international collaboration is demonstrated through the Swiss AI and AI Singapore partnership, showing how open models can be adapted for diverse regional needs

Practical applications are already being developed in three key areas: climate/agriculture, health/humanitarian, and education, with specific models like Meditron for healthcare and Legitron for humanitarian law

Academic research can achieve measurable impact and scale when properly resourced and networked, moving beyond traditional paper publication to real-world implementation

The ‘sovereign-able’ approach allows countries and organizations to build upon open foundation models while maintaining control and customization for their specific contexts

Resolutions and action items

Switzerland will host an AI summit in Geneva in 2027 to foster international dialogue and prevent AI development fragmentation

The Swiss-made LLM will be released with Apache 2 open source license, making it accessible for global use and replication

Continued collaboration between Swiss AI researchers and international partners through ICANN’s platform

Expansion of the Human-Centered AI for Social Good course at ETH with openness to new collaborators

Development of an educational card game in collaboration with Data Science Africa to teach computational thinking and equitable AI principles

Establishment of online forums and platforms where stakeholders can submit challenges for human-AI partnership solutions

Unresolved issues

Funding mechanisms for scaling academic AI research initiatives to achieve broader impact

Specific metrics and methodologies for measuring the social impact of AI applications

Detailed governance structures for ICANN’s decentralized network operations

Long-term sustainability models for maintaining open-source AI infrastructure and resources

Mechanisms for ensuring equitable participation from lower-resource countries and institutions in the global AI ecosystem

Balancing sovereign AI capabilities with international collaboration and resource sharing

Suggested compromises

The ‘sovereign-able’ approach that allows countries to maintain control over their AI capabilities while participating in collaborative frameworks

Combining open-source principles with practical business applications through Apache 2 licensing

Balancing academic research goals with measurable real-world impact through interdisciplinary collaboration

Creating both large-scale (70 billion parameter) and smaller models to accommodate different resource levels and use cases

Establishing ICANN as an international network while being hosted and initially funded by Switzerland

However, mounting geopolitical tensions and intensifying a perception of AI race among leading nations threaten to deepen global polarization. Against the backdrop, political summits focused on AI cooperation become vital for fostering dialogue and preventing fragmentation.

Speaker

Bernard Maissen

Reason

This comment reframes the entire discussion by acknowledging the fundamental tension between national AI competition and global cooperation. It’s insightful because it positions Switzerland’s democratization efforts not just as altruistic goals, but as strategic responses to prevent dangerous fragmentation in AI development.

Impact

This set the stage for the entire discussion by establishing the urgency and geopolitical context. It shifted the conversation from purely technical considerations to understanding AI democratization as a diplomatic and strategic imperative, influencing how subsequent speakers framed their contributions.

The model itself will be open source, not just open weight, and we have documented all of the steps that we have taken, so anyone can take what we have done and replicate it or host it themselves.

Speaker

Mennatallah El-Assady

Reason

This technical distinction between ‘open source’ and ‘open weight’ is crucial and often misunderstood. It demonstrates a deeper commitment to true democratization – not just sharing the final product but enabling complete transparency and replicability of the entire process.

Impact

This comment elevated the technical sophistication of the discussion and established credibility for Switzerland’s democratization claims. It moved the conversation from abstract principles to concrete technical implementations, showing how philosophical commitments translate into practical decisions.

And humans are humans and we do compete. And it’s important to convene people so that we dilute that ability of competition, because this actually can be quite destructive sometimes, and to go beyond our individual egos and to really come together to make something bigger.

Speaker

Mary-Anne Hartley

Reason

This is a remarkably honest acknowledgment of human nature in collaborative efforts. It’s insightful because it recognizes that competition isn’t inherently bad but needs to be channeled constructively, and that successful collaboration requires intentional design to overcome natural competitive instincts.

Impact

This comment brought psychological and organizational realism to the discussion. It shifted the tone from idealistic collaboration rhetoric to practical considerations of how to actually make international cooperation work, influencing how ICANN’s structure and purpose were subsequently explained.

So it’s sovereign-able, right? It’s how you can make these kinds of models your own and own them yourself and that sovereign-ability is really the spirit of what decentralization we’re trying to achieve.

Speaker

Mary-Anne Hartley

Reason

The coining of ‘sovereign-able’ is linguistically creative and conceptually important. It reframes the sovereignty debate from zero-sum nationalism to enabling capability – the ability to be sovereign rather than enforced sovereignty. This addresses a key tension in AI policy discussions.

Impact

This neologism provided a new framework for understanding how democratization and sovereignty can be complementary rather than competing goals. It influenced how the Singapore collaboration was understood – not as dependency but as sovereign-ability in action.

Implementation is science. We should measure impact. There is impact-driven science and for us to value that and to understand that it does happen and you’re not just investing in papers that are published and filling holes in textbooks.

Speaker

Mary-Anne Hartley

Reason

This challenges the traditional academic research model and redefines what constitutes legitimate scientific work. It’s thought-provoking because it argues for a fundamental shift in how we evaluate and fund research, particularly in AI where real-world impact is crucial.

Impact

This comment shifted the discussion toward practical considerations of funding and scaling. It directly addressed potential skepticism about academic research relevance and set up the call-to-action conclusion by arguing that implementation-focused research deserves investment and recognition.

We are really, really looking forward to work together and collaborate… we have forums online where anyone can go in and submit challenges that you would like human-AI partnership to solve.

Speaker

Mennatallah El-Assady

Reason

This transforms the discussion from a presentation about Swiss initiatives to an open invitation for global participation. It’s insightful because it operationalizes the democratization principle by creating concrete mechanisms for inclusion rather than just talking about openness.

Impact

This comment fundamentally changed the audience’s relationship to the discussion – from passive listeners to potential active participants. It provided concrete next steps and transformed the session from informational to actionable, directly supporting the democratization goals being discussed.

Overall assessment

These key comments shaped the discussion by progressively building a comprehensive framework for AI democratization that addresses geopolitical realities, technical implementation, human psychology, conceptual innovation, academic reform, and practical participation. The conversation evolved from high-level diplomatic positioning through technical details to concrete calls for collaboration. The most impactful comments were those that either reframed existing concepts (like ‘sovereign-able’) or acknowledged uncomfortable truths (like human competitive nature), as these provided new ways of thinking about persistent challenges in AI governance and international cooperation. The discussion successfully moved from abstract principles to actionable initiatives, largely due to comments that bridged idealistic goals with practical implementation strategies.

How can we measure and demonstrate the concrete impact of AI research initiatives to secure better funding and support?

Speaker

Mary-Anne Hartley

Explanation

Hartley emphasized the need to communicate that academic research can have direct, measurable impact beyond just publishing papers, and that demonstrating this impact is crucial for securing funding to scale initiatives

How can we expand international collaboration and networking beyond the current Swiss-Singapore partnership?

Speaker

Mennatallah El-Assady

Explanation

El-Assady expressed interest in synergizing with other ecosystems and expanding the network of collaborations to translate research into practice on a larger scale

What specific mechanisms are needed to create effective political frameworks that enable AI collaboration to thrive?

Speaker

Bernard Maissen

Explanation

Maissen indicated the need to develop political frameworks that support and enable the practical collaboration demonstrated in the technical initiatives

How can the educational AI game being developed with Data Science Africa be tailored to different use cases and user groups?

Speaker

Mennatallah El-Assady

Explanation

This represents an ongoing research area for developing educational tools that can be adapted across different contexts and communities

What are the specific outcomes and effectiveness of the Human-Centered AI for Social Good course model for interdisciplinary collaboration?

Speaker

Mennatallah El-Assady

Explanation

While the course was described as completed, there’s an implicit need to evaluate its effectiveness and potentially replicate the model elsewhere

How can the validation and evaluation platform (MOOVE) be scaled to support global medical AI model deployment?

Speaker

Mary-Anne Hartley

Explanation

The platform allows for contextual validation of medical AI models, but scaling this globally requires further research on implementation across different healthcare systems