CS CY - Trust by Design Open Source Systems for Secure Digital Infrastructure: The Swiss approach

[applause] Before I talk to um the panel the panelists I want to connect with Switzerland and the poly technic school of loan because uh Oleg Labros Lavroski who is uh a panelist working for aus artificial intelligence the only as I know so far opensource artificial intelligence who can connect with us and particip So if someone can help me here with your infrastructure, I would be >> hi. Perfect. So far away. >> Can you hear me? Okay. >> Yes. Can you hear us? >> Fantastic. Yes. I I had I had been uh having a little bit of trouble in the in the previous panel because of some people seek to speak away from my the microphone. But I hope in this one we will all be well behaved and loud and clear. Thank you. >> So good afternoon dear friends, ladies and gentlemen. In 1993, I was a student at the University of Ber and then I realized that CERN [snorts] made a historic decision. Rather than keeping the worldwide web proprietary, it released it freely to the world. The reason was simple but profound. A global information network could only succeed if it remained open, accessible, and interoperable for everyone. Today, Europe faces, I think, a similar challenge. As digital infrastructure becomes critical to our economies, societies and democracies, the question is no longer only how to innovate but how to ensure that innovation remains trustworthy. Trust is not something that can be added later. It must be designed into the technologies we built from the start. This is the theme of our discussion today. Trust by design. We are fortunate to be joined by expert working on it. From secure networks and its resilient access to knowledge to trustworthy artificial intelligence and privacy preserving digital payments as we've heard before from Leo Schumach. Let me begin by inviting and welcoming each of our panelists. Ole Rose, Keef, Nicola Rosoli, thank you. Um, Emmanuelart and John who is the youngest and will speak about his experience with Albertus please. [applause] So uh my first question and I will stay in here and I'm not going there because I want also to have a eye contact with O this is um please introduce yourself briefly and tell us about your work and project um let's start with you Emmanuel press the button button >> press the button >> only once collect your body. >> Fantastic. Yes. >> So my name is Manuel Ling. I launched 20 years ago a project called Cubix which has for purpose to bring online content offline for people with difficulties to access problem internet and 20 years later I'm here to uh it. >> Go on Nicola. >> All right. [snorts] My name is Nicolari. So I lead design association. We are a nonforprofit organization based in Switzerland and we maintain design protocol. So that means uh open source implementation specifications and then a lot of community. >> Yes. Hi. So I'm the technical community manager for the Swiss National AI Initiative. It's a collaboration between the two technical universities EPFD teach as well as the Swiss supercomputing center in Logano and I am based in uh actually in Bur. So greetings from Bur uh where we have today the Swiss IGF internet governance forum and uh yeah I would have otherwise been happy to be with you in person but we have parallel conferences today >> hopefully next year in Athens. >> Exactly. So, um, George, >> hello. Hello. So, uh, I'm George Bernakis. Difficult last name. Please, uh, excuse the Greek background. Um, so my my background is in military history and security studies. Uh, and I have engaged in research regarding the impacts of AI and large language models on uh, society uh, and the economy. uh and uh my main uh concern and presence here is uh to talk a little bit about my experience with apprentice but also to comment on um the engagement of uh the social sciences in the AI space. >> Thank you all. So then my question to all of you is uh that the title of this panel is trust by design. What does trust mean in the context of your work and how can be embedded into digital infrastructure from start Nicola? All right. So um especially when it comes to new technologies, trust by design means uh that you have to think about the security but also the governance uh part of a technology really from early on and uh possibly really bake it into the technologies and this is really what science does at at the networking layer when it comes to the connectivity. >> Thank you. manual. >> So in TV that's a bit easier because we are not really such a big part of the infrastructure like the replacement of network infrastructure or sales solutions this kind of thing. So the solution is easier. Um and for the rest of the trust uh this is probably one of the biggest assets we have like as a trademark as a project and what I can say is that the fact that we are dealing with open source only open source software and open content in general both of them gives almost by definition a super good uh note in term of trust trustworthiness. Thank you. >> Well, apparus means openness and it's uh you know in the context of my work as a community builder uh very much about opening technology projects to the users and trust to me means not like a static uh concept. It's something very dynamic needs to be continuously maintained through ongoing efforts basically to address user concerns to be an active uh partner to incorporate their feedback and of course to reflect societal values laws and emerging technological capabilities and I mean there are so many areas where an AI project like apparatus you know touches upon things I'll just mention a couple of things we're working on one is explanability ility. Of course, we would like to be we are I think one of the most transparent projects of our kind with a massive amount of documentation going out, but we also have an interpretability track where we are taking specific concepts in the development of LLMs and making those um you know making explanations about them. The other one is around data quality and mitigating bias. We know that there's an enormous amount of problems and issues uh in in understanding how how bias comes into play. This is a massive area of concern. Um things like security and privacy get putting putting boundaries around what people can do with LMS places where LLM should push back. Um and of course human in the loop you know how do we involve researchers and the general public in critical decisions? How do we build feedback mechanisms into the model? Um, one of the main areas of that is of course value alignment, but also things like accessibility, making sure that the system is accessible in multiple languages, that it's intuitive to people from different backgrounds and abilities. Those are just a sampling of many areas. And we believe that the open source approach to large language model development um is is definitely one of the fundamental blocks to a more trustable system. >> Thank you. George, do you have a take as a user of technology? >> Uh, as a user of technology, well, um, what I would say is that for me trust uh in um in open source technologies and in uh AI broadly has to do with um governance uh incentives, which is something very important. It's what kind of incentives do the proprietors of uh AI, the developers of AI, those funding it, what incentives do they have to um make these technologies interact with the public properly? and also uh for regulators to be uh intimately aware of all the ways in which uh these technologies interact with uh the public broadly not just people who know how to use them and how to work with them not just people in tech but everyday people who will interact with LLMs in increasingly uh you know large ways in their daily life. How do we avoid uh the certain pitfalls? how do we investigate uh those uh the problems that could arise from something like this? Uh and I believe that open uh open source and uh open uh research and uh an environment of discussion is the most important way in which we can explore these uh problems. >> Thank you Nico. I have two questions for you. Uh many groundbreaking technologies remain in the research space. What enabled scan to move from academic research into operational development? >> Um, thank you. So scan is a network technology used in critical infrastructure ecosystems and this is notoriously uh hard uh very hard sector to to to produce in innovation because technology life cycles are very long and typically operators of crystal infrastructure are very reluctant to change things. So scan I think here is really a bit of a special case in the sense that um the technology could really move out of 88 sur towards uh for example deploy big deployments like the Swiss interbanking network. So really processing billions hundreds of billions of friends in financial transactions thanks to sign connectivity and this was possible because of a few ingredients. So one of this I think the most fundamental was a really collaboration between a very strong um research uh community in Switzerland and uh potential users. One of the most important one was for example the central bank that is also the regulator when it comes to payments and they have been looking at the resilience of uh the internet network and uh by looking at that and at the same time by having research collaborations with the university then they realized that there is some potential to use some emerging technology like scan and that's that's really how then the early trials of the finance network started and that's really what gave scan its initial commercial impulse and yeah today it's it's really widely deployed in in that use case and it is there are many more that are uh coming to life. Last factor I would say is also that uh I think in Switzerland there is a lot of attention to resilience. Why? because I think uh typically uh there is some risk aversion and that really helps that that the regulator was really looking at even the network which is something that typically we we all forget about we only think about you know the data center and so on. >> Yeah. My second question to you is that Europe talks increasingly about digital sovereignity. Can digital sovereignty be achieved without investing in sovereign network infrastructure? >> That ties back to what I was mentioning that we are talking a lot about the data center. We are talking um uh a lot about governance on on who owns uh you know our supply chains and so on. But we forget the network and the network is really a really key part of any IT supply chain. you know there is no no AI uh no cloud without connectivity and that's something that can scan uh tries to tackle by really giving its users visibility on which uh which telco providers you're going through the network by creating this federation so I think this is really a very important piece that we need to kind of keep keep in mind beyond beyond what sits in in the data center and what sits on the end user side so yes of course Thank you, Emmanuel. I have also two questions for you. The first question is, what lessons can we learn from Hubix about ensuring that knowledge becomes and remains accessible, open and independent? So, the first things to learn is that half of the world population does not have an access to internet. That's something we tend to forget because it's so privacy for us. We have it everywhere here at home in the train everywhere just there. But actually even if problem access internet moves ahead, more and more people have access to it. Uh there is still so many of them who don't have access and even worse we have more and more people having access to broker internet without having free access to proper internet that means the internet access is censor uh we talk here about very important country like uh China that's obvious one but also new countries like Russia where the censorship is very severe they don't have access the Wikipedia. So as Kubix this is we are kind of spin up of the Wikipedia movement. So we are always sensitive about who has access to Wikipedia. So that's was the problem about the solution. Uh obviously opensource software is a technology transfer. So you don't have a first access um problem type by paying everybody can use it and that's I mean that's that's an obvious thing but that's something to say and we we talk a lot about helping what we call the global self uh but uh just making open source software is a very good things to do it and it's for everybody so including the global The second point is the content itself. Uh we are not only a software solution. We maintain a library of thousands of content to be available offline. Wikipedia is one of them but we have hundreds of other content and that's made possible in a legal and practical manner only because there is a strong um I don't like like really the open content movement but a movement to make free content content which can be reused and downloaded freely and redistributed freely. So um free content is very very very important for us and it's very very important for people having difficulties to access to knowledge. >> Thank you. My second question to you is that Kix um has shown that access to knowledge remains possible even when connectivity is limited or unavailable. Why should access to knowledge be considered a strategic component of digital resilience? Maybe this is >> yes that's an easy one. >> So everybody need access to reference documentation to books to curriculum to study. Um maybe the the best point is just for schools for universities that they can just work. If they don't have access to the to this document then they cannot work. Of course this is only part of the challenge. They also need kind of medium for mediators, teachers, librarians to explain to follow the curriculum and to get the job done on transmitting from the uh from the document from the books to the to the students. But if you don't have the reference documentation and I would say if you don't have in a digital manner today everything is digital. Um I I just just just something to remind if you have the document as a book in your library you can give it one student but if you have a the same content as a digital book using kix or any other platform then you can give it to all your students. So the digital form is very important to help to spread the knowledge. So if you want to prepare the next generation to be educated and make educated choices then uh better use cubics and better secure their access to the content. >> Thank you. Uh Ole I have three questions for you. >> Sure >> only because I have only one for uh George >> afterwards. Artificial intelligence is becoming part of Europe's critical digital infrastructure. What role does openness play in making eye systems trustworthy? [clears throat] >> So first of all I think in your introductory remarks you mentioned something like apparatus is a very unique uh project. It's not there are actually prior projects for years before I joined the appertis team I had been using the Euro LLM uh model from kind of a joint venture between the various data science centers across Europe. Um we are connected to organizations really around the world creating um you know large language models really thousands of them are being uploaded to places like hugging face. This is a burgeoning industry, but it's also a place where people can find an amazing amount of creativity and inspiration practically on every topic. Whether you're interested in in health or if you're interested in education, you can find efforts to you know develop new systems around that. But also very very niche things like you know local problems, local dialects, local data sets are also being addressed by these machine learning and large language model communities. So I just want to first point out that this is not something unique and I think that openness is really one of the key things that can engender creativity and at least to perceive the amount of creativity and and interest there is to to use this particular technology. Right? Because if we if we stay only consumers of products and have a very personal interaction with large language models through a chat interface, we do not we are not exposed to what other users are doing. We're not exposed to what other communities are building, right? So it's very important to to at least publish to talk about what we're doing and we have regular scientific exchanges colloquia and I think that's that's really for me this uh this the pillar of that. Um but of course on a you mentioned on a governance level so we're currently going through a review process with the digital public goods alliance who are one of the organizations there you know there are others um who are scrutinizing our work going through the report downloading you know the the code and checking to see if the data sources that we're using are you know really free of uh of of legal or political issues and of course they're not. Of course there are always things you can find and so it's also important to have you know openness as as as a strategic uh instrument in order that regulation can be developed in order that things can be assessed and that there is and that this in in the case of apparatus we try to represent Swiss values right but you know what what is Switzerland it's it's it's it's a message it's a construct you know transparency neutrality these are these are topics we have to defend on a daily basis right and again you know if you try to pack that into a project like apparatus it's it's it's a constantly shifting goal, right? So I think in in that in in that sense you know that we have the Swiss AI initiative um which is also important to say Peris is one of [clears throat] of many dozens of projects within this large you know uh collaborative environment where we democratically govern access to this shared infrastructure which is the supercomputing facility in Lugano where the model gets trained or where other experiments get run also materological simulations happen and all kinds of analytical tasks get done with that with that system um you know we have tens of millions of GPU hours and those are shared across our teams and this we hope that others would also you know understand that principle and apply it to various other kinds of infrastructure you don't have to have a second fastest supercomputer in Europe you can also do this on university infrastructures or federated machine learning systems um at at at a at a at a smaller scale you know so for me that's also something like openness is like an invitation to approach problems and you know even from this level of home labs or amateurs or hackathons you know giving giving people the right to learn and to experiment with the technology and again if they find a way to do that creatively expressively inclusively they should be able to share that with others. >> Thank you. My second question is that there is often a tension between powerful eye systems and transparency. How can open source I help Europe build trust while remaining competitive globally? >> Is that a question for you or for George? >> For you. [laughter] So I mean first of all I think we need to question what AI means and it's it's become a very diluted uh moniker for a whole bunch of things which in many cases are not particularly artificial nor intelligent. So you know I I've I've read the European Union's um report. I think it was it was mentioned um in in the keynote earlier you know the one that was published at the beginning of the month you know and I think we we should not first of all lose the fund view on the fundamentals I I've been working in the area of open data for many years and trying to understand what is really the value of these you know this digital heritage right these these resources like digital archives and statistics and you know engineering analytics that are you know that's dup some cases this all just about duplication of effort. Um and in some cases yes it's questioning you know these aspects of sharing the commons right like in the case of kwix you know you you definitely want people to have access to knowledge but you also need to have some form of scarcity because without you know scarcity there's also it's very hard to for people to see value in in a particular resource right so there's all these variables at play and and these these have already you know been very much part of digitalization strategies and open source strategies so fundamentally We talk about open-source large language models. So that's the kind of open source we're developing. An open source large language model is these model weights, these parameters, these recommendations, but it's also within it doesn't work on its own. It's a component in a system which includes you know the frameworks to you know to run the model to to test it to offer it to users in the form of chats or APIs and and there and it's it's just very it's just a very advanced piece of engineering but it's doesn't mean that it's not something you could put in in a box you know that you could take offline right so I'm also quite excited by the way of maybe combining forces with with projects like uh like Kix to kind of you know make LLM available um on smaller devices. Actually yesterday we released apparatus 1.1 uh the mini version of apparatus which runs on a Raspberry Pi for example. Right. So we are taking also steps in in that direction so things can be more more portable and accessible. >> Right. But >> yeah thank you. I am going to George and then I will come back to you if time allows because I have also other questions for our participants. So George, as a new user of Alberto, what kind of differences do you notice compared to other LLMs? >> So um I have been using LLM for a little while as have all of you. I started using LLM while I was in uh university in 2023 and I did not like them. I did not trust them. I did not enjoy uh my use of them and I noticed two major differences when I started using uh appertus. Um something that always uh bothered me as a broader implication of the use of LLMs was uh the amount of flattery they engage in. They are little sick of fans a lot of the time. And something that I noticed uh about uh Apertus was that first of all it did not want to flatter me. It wanted to uh be straight about what it can do, what it can't do. Hedging something that other LLM often don't do. Uh and that was the main positive thing that I had as a as a takeaway uh when uh using a purchase. And I think um it made me think a little more about um what the day-to-day interaction between uh LLMs and people broadly outside of the tech world uh can do to uh society. Um you've probably heard of stories with uh people uh having psychotic breaks after discussing with AIS uh for a long period of time. uh and that has to do with uh us not necessarily being capable of distinguishing uh an AI LLM as a tool uh and as a potential friend, advisor or confidant. Um it is vital uh for um societies to be able to um investigate very closely how uh millions and millions of people interact with AIS and uh to avoid having them damage our psyche. just to um as a as a last thing here's where I think public funding uh is very important and in combination with uh open-source approaches uh to LLM development because if there isn't oversight uh and incentives beyond just let's maximize how many users uh are involved here by any means necessary uh I believe that LLMs could seriously uh harm a lot of our social relations. So uh I'm very glad uh with my experience with Apertus was I saw those uh guard rails being uh formed in real time and being implemented into the development uh of the LLM from the beginning uh having the appropriate um incentives uh to continue forward. Since we are in Cyprus, I have for all of you a question on Cyprus. Um, if if Cyprus wanted to become a regional hub for trusted digital infrastructure and opensource innovation, what would be the first practical step to do? I would uh definitely recommend having a platform framework somewhere to have collaboration between um between the private sector, the the governmental sector, policy makers and uh the academia. It definitely looks like from previous discussions that the ingredients are there and you need to somehow mix them. But for trusted infrastructure, then the governance part of it is also very very important. Am I live? >> It will be difficult to say something different. Um, >> it's not very different to ask what should be the the the environment to to be uh to help to to create uh companies or to to to push entrepreneurship. It's basically the same the same kind of of thing you need to create the ecosystem. uh at least the most basic things uh like uh the places uh the support uh and of course an economical environment coverizing this that's been with with clients or sponsor uh to help to to to bootstrap your project and move forward because this is really the beginning the most complex thing. >> Thank you. So [snorts] I I think it's quite quite important to to have a kind of a digital innovation office of some kind some agency. That's something that's been topic of course in in Switzerland but also within each Canton. Uh we've seen this emerg the emergence over the past years of new roles like data steward you know people who are really taking responsibility for uh these topics addressing public concerns quite seriously. And I think it would be good for them to have a mandate and a team and some resources to to to to run you know programs and uh you know I'm a little bit biased because we have the you know Swiss IGF here in Burn today but you know working with the internet society you know having having that chapter involved is also a very good step towards that >> and I will have a last question I have a lot of questions of course but my last question to you because I suppose I don't have enough time uh is that um imagine we need again in 2030 this is in few years uh what would success look like what would need to have changed for Europe to say that it has truly built trust by design I start with you >> there's uh an interesting contest um until the end of July any anyone who is a writer can submit a short story in for the Protopian writing award. It's just a little bit advertisement that's uh run through the public AI network. Um so definitely thinking about the future is great. I I think it would be I think we we should I would I would I would maybe think about my children, you know, who are growing up with this technology who are definitely thinking about AI as their companions and and confidence like George said. Um and maybe fulfilling, you know, new needs that are hard for the previous generations to to imagine. Um but I'm not really an oracle, but let's write a story. >> Nice. Emmanuel, what you think? Yes. Yes. So actually openness, transparency is core for trust. It's a cultural things, a cultural switch and I think at the European level it's moving in the right direction. uh open data is a something obvious for the moment and now it it wasn't in the past and uh open source software uh is a key asset for sovereignity for trust uh for Europeans. So this is I think a fight which is going to be won and the next step is obviously for AI like said Ole and then there is still a lot to do so let's see four years >> George do you have a short please because we don't have time >> um so for me success in uh maybe four years time uh looks like a movement from the mentality of move fast and break things. I don't want to break things. I would rather see uh people uh properly taking the steps to harness the uh productive and um and uh transformative aspects of AI that can genuinely help society and not have this blas attitude about oh yeah we're going to just uh enter a development arms race with uh China and the United States and if we uh destroy society In the meantime, that's okay. That's a a different discussion. No, no, no. I think society should be right in the center of the discussion of AI. And uh the I want to see the a much greater involvement of the social sciences in uh open-source uh research and in um and in AI ethics. And that's what success looks like to me. >> Thank you. Nicola, your take >> on my side. community on more on critical infrastructure ecosystems. Um for me success is that in ideas that they are uh built on on open source on technologies that have trust by design and maybe that the next ECB tenders are then uh more open to uh to new technologies because we definitely have the ingredients but today we are not really uh always leveraging what we have. >> So thank you very much. Do we have time for another question or not? No. Do you if you want wish to um ask our guests please go on. We have also a social event afterwards. So we our guest. Thank you very much all. I thank you very much FOR [applause] [applause] I hope that we will be able to organize something next year. um as internet society switcher uh chapter and um NGIO maybe in Athens, who knows? Um we will be there to welcome you. Thank you very much.