Date: June 5, 2025
Session: Track 4 - Swiss & Open Source LLMs
Moderator: Dr. Alberto-Giovanni Busetto, Chief AI Officer at HealthAI

Introduction: The Open Source Renaissance

The landscape of artificial intelligence has undergone a remarkable transformation over the past two years. What once appeared to be a closed ecosystem dominated by proprietary models is now experiencing an open source renaissance that could reshape global AI development.

Leandro von Werra, Head of Research at Hugging Face, set the stage with a striking observation about this shift. « Two years ago, things looked much bleaker if you wanted to use or build open models. It was around the time when ChatGPT came out and all the big labs stopped releasing models. Google invested in Bard and then Gemini. Meta had a few model releases that weren't super successful. »

Today's reality presents a stark contrast. Major technology companies are returning to open source strategies, with Google releasing Gemma models, Meta continuing the LLAMA series, and even OpenAI rumored to be developing open models for local deployment.

Defining Open Source AI: Beyond the Marketing

The panel immediately confronted a fundamental challenge facing the AI industry: what constitutes truly "open source" artificial intelligence versus mere marketing terminology.

Daniel Dobos, Research Director at Swisscom, drawing from his particle physics background, offered a scientist's perspective. « From an academic point of view, open means reproducible. And everything what is needed to reproduce results would basically fall into my definition of open, which includes even filtering of the data. »

Imanol Schlag, Associated Research Scientist at ETH AI Center, expanded this definition with important nuance, distinguishing between different categories of AI accessibility. « Open source means it is reproducible. So if you only share the weights of your model and you are not transparent about how you produce those weights, like what was the training process, and particularly also what training data do you use, then you cannot call yourself really open source. »

This distinction between "open weights" and "open source" carries profound implications for the AI ecosystem. While open weights models allow usage and deployment, true open source enables community building and collaborative improvement—a difference that could determine the trajectory of AI development globally.

von Werra emphasized the strategic importance of this distinction. « The interesting thing about fully open source is that the whole community can immediately build on top of what you build. If OpenAI had released how you make ChatGPT, everybody would have started building better chat models immediately, and not have that warm-up of a year. »

The Swiss AI Initiative: A National Commitment

Switzerland's response to the global AI race represents one of the most ambitious sovereign AI initiatives outside the United States and China. Schlag detailed the scope of this national commitment, which centers on the ALPS supercomputer infrastructure at CSCS (Swiss National Supercomputing Centre).

The Swiss AI Initiative encompasses over 70 professors conducting generative AI research, with approximately 20 million Swiss francs committed over four years for PhDs and postdocs. The computational resources are equally impressive: 10 million GPU hours provided in 2024, scaling to 15-20 million hours annually thereafter.

« They're being trained right now on the Alps supercomputer, which has over 10,000 Grace Hopper superchips, GH200 GPUs which are basically the best GPUs you can get, » Schlag explained. The infrastructure advantage extends beyond raw computing power—CSCS brings over 12 years of experience with accelerated GPU-based supercomputers, knowledge that proves crucial for large-scale model training.

The environmental consideration adds another dimension to Switzerland's approach. The ALPS facility in the Lugano area operates primarily on hydroelectric power and implements heat recovery systems for local district heating, demonstrating how sovereign AI development can align with sustainability goals.

Business Case: From Users to Builders

The panel confronted a critical question facing corporate executives: why invest in developing AI capabilities when existing services seem adequate? von Werra drew parallels to software development history to illustrate the strategic imperative.

« There's that nice blog post from Marc Andreessen called 'software is eating the world,' essentially saying every company is becoming a software company on some level. If you look around now, every major company has a huge department building their own software stack. »

The analogy extends directly to AI development. While companies can use off-the-shelf solutions for generic tasks, competitive advantage emerges from specialized capabilities tailored to specific business contexts. « If everybody's using the ChatGPT API, nobody has an advantage. But if you build super strong models with high-quality datasets, then that's how you can have an advantage. »

Dobos reinforced this perspective from Swisscom's experience leading the Swiss AI weeks initiative. The company has collected 55 organizations to test Swiss open source models across government, private sector, and academic applications. « Everybody understood very quickly that it makes us independent, that it builds trust. Having that choice is worth having such a model. »

The trust factor proves particularly significant in Switzerland's business environment, where confidentiality and reliability carry premium value. Open source models offer transparency that closed systems cannot match, while sovereign development ensures continued access regardless of geopolitical tensions.

Startup Ecosystem: Leveling the Global Playing Field

For entrepreneurs and emerging companies, open source AI models represent a fundamental shift in competitive dynamics. Rather than requiring massive capital investments to develop proprietary AI capabilities, startups can build upon sophisticated open source foundations.

von Werra highlighted the ecosystem effect currently developing on platforms like Hugging Face, which now hosts nearly half a million datasets and 1.5 million models, with new models uploaded every 10 seconds. « If you're a startup and you want to build something for financial services or something, there's a very high chance there is a dataset or a model already out there that you can leverage. »

This accessibility transforms AI development from a zero-sum competition to collaborative innovation. Startups can contribute to and benefit from community improvements rather than starting from scratch with limited resources.

Schlag emphasized the defensive advantages this approach provides for smaller companies. « You need to build something that is maybe not just a wrapper around an existing service from which a US or Chinese tech company profits. If you can leverage this environment, this ecosystem, then you have a much easier time to find collaborators. »

The regulatory dimension adds another layer of protection for European startups. As AI regulation evolves, companies building on transparent, open source foundations face lower compliance risks than those dependent on opaque proprietary systems.

Regulatory Landscape: European Advantages

The European regulatory environment, often perceived as a constraint on AI development, may actually provide strategic advantages for open source approaches. Schlag highlighted specific examples where closed models create compliance complications.

« The Llama 4 models that came out—in the license, if you want to use that model, it explicitly says that you are not allowed to use the multimodal models in Europe. And basically all Llama 4 models are multimodal. »

Such restrictions illustrate how dependence on proprietary systems can create unexpected business risks. Open source models, by contrast, offer transparency that facilitates regulatory compliance while providing certainty for long-term business planning.

The European approach to AI regulation emphasizes transparency and accountability—principles that align naturally with open source development methodologies. This convergence suggests that European companies building on open source foundations may enjoy sustainable competitive advantages as global AI governance frameworks evolve.

Technical Competition: Closing the Performance Gap

A persistent question in open source AI discussions concerns performance relative to leading proprietary models. The panel provided evidence that this gap is rapidly diminishing.

von Werra tracked the convergence timeline. « When GPT-4 came out, it took roughly a year for an open model to be as good as GPT-4. When DeepSeek came out, it matched the best closed model from like a month or two ago. So that gap has been shrinking quite quickly. »

This acceleration reflects both improved methodologies and increased computational resources available to open source projects. Schlag expressed confidence in Switzerland's ability to compete technically. « Purely on performance, we are able, as in the open source communities, to increasingly close the gap between performance differences of very large models coming from these labs and open source efforts. »

The Swiss AI Initiative aims to demonstrate this principle by developing what Schlag described as « likely going to be the best performing model in this class since the open source class is still quite small when it comes to large language models. »

Global Talent Dynamics

Switzerland's position in global AI competition rests significantly on its talent advantages. Dobos emphasized this factor in explaining why major AI companies are establishing research operations in the region.

« Almost every big AI company that you think about will now basically open a research lab in Zurich or here in Lausanne. There are three reasons for that: talent, talent, and talent. The density of AI talent in Switzerland is amazing. »

This talent concentration, built through the long-standing excellence of ETH Zurich and EPFL, provides sustainable competitive advantages that capital alone cannot replicate. Other global AI hubs may offer different strengths—Dubai provides vision and unlimited funding, various regions offer different regulatory environments—but few match Switzerland's talent density.

The open source approach amplifies this talent advantage by enabling broader collaboration networks. Rather than competing purely on salary packages with major technology companies, Swiss organizations can offer researchers opportunities to shape globally influential open source projects.

Technical Development: Training from Scratch

The Swiss AI Initiative's approach represents a comprehensive commitment to building foundational AI capabilities rather than simply fine-tuning existing models. Schlag outlined their methodology: « We're training a model from scratch, currently focused on a pure text model, text in, text out. And then there's a pre-training stage which requires the most amount of compute, followed by a post-training stage. »

The post-training phase encompasses multiple capabilities including conversational abilities, value alignment, and reasoning capabilities similar to those demonstrated by DeepSeek and OpenAI's o1 and o3 models. This comprehensive approach ensures that Swiss AI models can compete across the full spectrum of AI applications.

The initiative also prioritizes inclusivity in language support, incorporating over 1,500 language and script pairs with particular attention to low-resource languages like Romansh. This multilingual focus reflects both Switzerland's linguistic diversity and a commitment to global accessibility that commercial models often neglect.

Future Implications: From Operating Systems to AI Infrastructure

The panel concluded with forward-looking perspectives on AI's role in technological infrastructure. An audience member raised the compelling analogy of LLMs as potential operating systems for future computing paradigms.

von Werra embraced this comparison while noting the current primitive state of AI interfaces. « The models are really good at user interfaces, so you don't need applications anymore with predefined user interfaces. I want a dashboard showing these kinds of things. So the way we interact with computers in terms of LLMs, we're maybe more like at the terminal stage at the moment. »

Schlag expanded this vision to encompass integrated AI systems that combine language models with external tools and real-world interfaces. « What we will see in the future is that you can actually probably replace that engine which is the actual agent executing the actions and choosing which tools to use throughout different solutions. »

This infrastructure perspective reinforces the strategic importance of open source approaches. Just as open source software became foundational to internet infrastructure, open source AI models may become essential building blocks for the next generation of technological systems.

Addressing Synthetic Data and Model Collapse

The discussion addressed emerging concerns about synthetic data's impact on model training quality. von Werra shared counterintuitive research findings from analyzing web-scale datasets over time. « We looked at what's the amount of ChatGPT data in web snapshots and you can see it started to go up quite a bit two years ago, and if you plot it against performance, you can see it goes up as well. »

This research suggests that high-quality synthetic data, particularly content that has undergone human curation and quality assessment, can actually improve training datasets. The key lies in effective filtering mechanisms that distinguish between valuable synthetic content and low-quality automated generation.

Schlag reinforced this perspective: « If you do reasoning, then you already trained on synthetic data. The crucial thing is that the data is filtered. A human has decided that this text I generate with ChatGPT is good enough to be published on my blog. That is enough of a quality assessment that prevents the deterioration effect. »

Conclusion: Writing the Future

The discussion revealed open source AI not merely as a technical approach but as a strategic imperative for European technological sovereignty. Switzerland's commitment to developing transparent, reproducible AI systems positions the country to influence global AI development while maintaining alignment with European values and regulatory frameworks.

As Dr. Busetto concluded, « Open models are a fantastic opportunity because the future is not written yet. Open models are, in my opinion, a pen that all of us can use to write the future. »

The Swiss AI Initiative represents more than national technological development—it embodies a vision of AI advancement that prioritizes transparency, collaboration, and sustainable innovation over purely proprietary competition. Whether this approach can maintain technical leadership while upholding these principles will significantly influence the trajectory of global AI development.

The stakes extend beyond technological capabilities to encompass questions of digital sovereignty, economic competitiveness, and societal values. Switzerland's open source AI initiative offers a potential model for how smaller nations can compete effectively with technological superpowers while maintaining independence and alignment with democratic principles.

More on the panelists

Leandro von Werra

Imanol Schlag

Daniel Dobos

Alberto-Giovanni Busetto, Moderator