Wikimania 2026 lightning talks on open AI, Incubator and open source: transcript

Key points

• The first speaker says AI training needs large, good quality, diverse data under respected licences, and that training AI on AI-generated content causes model poisoning.
• He contrasts closed, open-weights and fully open source models, and presents Luciole, a model released two weeks earlier in 1B, 8B and 23B sizes.
• Nikolas KHF praises Incubator's automatic categories, local admin rights and simple guides.
• He suggests a project checklist, a more detailed guidebook for new communities, and equal ground for new contributors in language committee approvals.
• Valerio Bozzolan presents Wikimedia CH's technical strategy, asking that commissioned software use approved open source licences and be compatible with Wikimedia Toolforge and Cloud.
• In Q&A, the Open LLM France speaker says the project has French public funding and works alongside other European open LLM initiatives.

Chapters

1. 0:00AI training data, model poisoning and when AI helps
2. 3:06Closed, open-weights and fully open models
3. 4:33Frugal models and the Luciole release
4. 6:46Nikolas introduces Wikimedia Incubator
5. 7:50What already works well in Incubator
6. 9:16Proposed Incubator improvements
7. 12:25Wikimedia CH technical strategy introduction
8. 16:09Open source requirements and open decisions
9. 19:01Q&A on Open LLM France funding and collaboration

Full transcript

0:00We talk about temperature, which means that we add random stuff, so we get hallucinations. Next, please. So, why does AI need data? Because everything about AI is about learning and training models with data sources. What's important is that we have big volume, good quality, diversity, and of course that we respect the license. This last one is very important. I don't know if you just saw this morning, Entropic is paying like 1.5 billion dollars because they forgot about this one. Next.

0:50Wikipedia fed the whole AI revolution. That's why I'm talking about this today. Wikipedia is a galaxy of projects. We know Wikipedia, but there's also Wikisource, Wikimedia Commons, where you can find very, very good pictures from Wikiportraits. Wikidata, Wiktionary, Wikinews, Wikibook, Wikiquotes, all these data are very well organized, and that's very important for AI and training models.

1:21Next. So, just one word about model poisoning. When you use AI to create new content, if you use that new content to train again AI, you will do what we call model poisoning, because as you loop on data and you train on stuff that was generated by AI, the content quality will get crap, basically.

1:51So that's a huge problem. Next, please. So why do we need AI at all? Because if we use AI to fill, let's say Wikipedia, and we use Wikipedia to train the AI, at some points we'll get very bad content everywhere, including Wikipedia. So, on the fr wikipedia website, there is a notice that basically tells you don't use AI.

2:30So, but at some point, I guess, who is using AI here in LLM? Okay. So, vast majority, of course. I use AI as well. So there are some use cases where it makes sense to correct, to translate, to review, to summarize, to structure, to check, like if you have some data that are not organized but you want to make a CSV file or whichever, detect and verify.

3:06So there are some use cases where AI and LLMs make sense and won't degrade the quality. Next please. So now a little word about three kinds of AI models. We are used to what we call closed models. If you use ChatGPT or Cloud, you don't see anything. You are just using either an interface or an API, but you don't know the first thing about what's behind.

3:37You don't know the models, you don't know the weights, you don't know everything. Open weights gives you the ability to download the model and to use it locally because you have all the weights but you don't know the data set, you don't know how it's being trained and you don't know the biases. At Linagora we've been working for years on fully open source end-to-end AI models which means everything is transparent and you get the full data sets there is no secret source the source code is published next please And I'm almost done. So we do that at LLM, Open LLM France, which is a set of companies working on this. We use French data set, but not only, of course. Everything is open license, so you can use it and modify it and distribute it. Next, please.

4:33just one word about frugality because we know that AI consumes a lot of water so just so that you have an idea we don't know the real numbers for the closed models because they don't disclose it but this 1 billion parameters can fit within your cell phone 23 can fit into your laptop and it will do translation and etc. Next please. Luciole is a model that we published two weeks ago so it's brand new and we published it 1B, 8B, 23B. You can download it and use it for yourselves. Next please. You can see that we tried to have the most diverse data I said possible so that it's not English only, but it's also French, German, Italian, etc.

5:34And we compare to the only closed model that we know anything about, which is LAMA2. Next, please. This is how we've been using the Wikimedia project so far. And, yeah, we use only 1.4% of the whole data set, but it's the quality that matters here. Next, please.

6:08so I would love you to get in touch with me so feel free to contact me to talk to me if you see me wandering around or trying to plug a cable or anything there's an email address here if you want to get in touch with OpenLLM France and that's it for me thank you very much all right we're gonna move right along to our next speaker Nicholas because do you have slides okay perfect Olivia has your slides okay Should I?

6:46You can climb it, yeah! So, while we wait, I think I can introduce myself first. So, this is my first time in France, in Paris, outside of Indonesia also, I guess. And I'm from Indonesia, okay, thank you. Okay, okay, can I start now? Okay.

7:17Okay, so hello everyone, I'm Nicholas KHF, and I'm usually active in Indonesian Wikisource and also Wikimedia Incubator, where with my community I usually develop Wikipedia and also Wikipedia and while we're doing that, I feel like we can do something with the Wikimedia Incubator that we can try to improve that, basically, so please next. Wikimedia Incubator is basically the test project that we can do.

7:50We can try to test the project that we want to create for new languages, something like that, and we try to gather the community. We can do that in the Wikimedia Incubator. Next, please. So, in Wikimedia Incubator, there are several things that are already good in here, and I can summarize it into three points in here. Next, please. The first one will be a stable system where in Incubator we already have the auto-category session where we can have the category that created automatically for each project and also each language code in here, which means that we can have the statistic on how much article that we have created, how much we have grown, etc. for that.

8:33and because it is automatic means that we do not need the newcomer to write the code category by themselves and that will make it less prone to error. The second one is test admin rights where the administrator and incubator are local which mean that if I have the administrator rights for sasa.nist Wikipedia only I cannot do any administrator tools on the sasa.nist dictionary, which is good because community can have, we can manage their own project and it will not have any risk to break the entire incubator, like entire project for that.

9:16And the last one is it has a simple guide and policy where if you read the manual, you can see that newcomer can read the manual and try to create their own new project. But I think it's just a few hours, so it is really good for that. Next, please. But as the title suggests, there are several things that we can still improve. In here, I... Let's summarize it into three. Next, please. The first one will be have a project checklist. So far, in Incubator, we don't really have a very clear project checklist for what we need to have, what we need to do for the Incubator.

9:57So, for example, in here, we can have a project, for example, if we want to create a new Wikipedia language, maybe. we need to have maybe translate 1,000 articles that every Wikipedia should have, and then we need to have minimum three active volunteers for the whole year, for example, and etc. for that. And then we can combine this checklist with qualitative assessment, such as maybe from that 1,000 articles, we can have at least maybe 10% or 20% that have a good quality article.

10:28So we still can have a qualitative assessment there. Next. Okay, the second one is more detailed guidebook. So after we create the Wikipedia project, at first I was like, what should I do after that? What should I do next for that? And maybe we can try to do that. Maybe we can try to create a detailed guidebook. To the new community on how to start it, tips on starting maybe, tips on gathering community, tips on growing the community, and then tips on which article should be prioritized, such as 1,000 article that every Wikipedia should have, or maybe for the week's journey, maybe we can have the Swedish list, something like that.

11:13Okay, and then next. And the last one is equal ground for new contributors because each language project should be approved by language committee first, right? And on the transference section, it says that the committee especially interest in hearing those who propose new wiki in the past and got either refused or approved, which mean that the past professor have like more interest by the language committee and this bias against new competitors may make them less enthusiastic. So hopefully next time we can have an equal ground for new contributors. And next, I think we can go next again.

11:54Yeah, I can go next. Yes, this one. So probably after we tried to improve the wiki media incubator, we can have a little bit more efficient ovens that can maybe have a more quality for the foods right but you know that the quality of the food itself are more influenced By the ingredients and also recipe, which means that the quality or the capacity of the volunteers itself will be more important than the oven.

12:25So we can still try to improve both of the ovens and also the volunteers. I think that's all. Thank you very much. Thank you, Nicholas. And next we have Valerie on open source. Valerie? Valerio. Oh, Valerio. I'm sorry. It's been a busy morning. I'm hungry. Who's hungry? Thank you so much, we are loading the screen, yes, I never leave my backpack unattended, never leave your backpack unattended, somebody will install Linux on your computer if you leave your backpack unattended, can you open please this tab, yes, no?

13:28We can close this? Yes, thanks. Okay, not really. What's happening here? Okay, nothing works. Okay, Wikimedia CH technical strategy. I'm here as volunteer today. I work for Wikimedia CH since some years, but today I'm volunteer presenting my work with this organization because it affects volunteers.

14:03and I wanted as volunteer to have something that positively affects also me when I'm talking with an organization. So please raise your hand if you use a Google Form or Google Docs or Google something in the last day. That's what we tend to do, but we cannot assume that every single volunteer is a customer of Microsoft, is a customer of Google, is a customer of this kind of giant companies producing closed source software just to join our movement so this is our potential issue let's say so our My goal was to produce a document which was easily copy and passable for other organizations.

14:54So it's so meta. The meme, it's so meta, is true. So the core thing of having a technological strategy for an organization, it can be small or medium, is making sure that your collaboration, like professional collaboration, when you have a budget as an organization and you want to enable a project, you create a new project, create a new software project, so you expect somebody else producing software for you, it's important to have somehow very clear requirements for your organization to be sure that the software is open so that you can use the software because sometimes it's not so clear because we are in an hackathon in this very moment.

15:36It was five seconds ago in the hackathon room, and we have seen a lot of incredible AI-generated projects which are source available but are closed source because I have not an open license so it's people for people it's too much complicated to understand what is open source but we are Wikimedians we know what is open content So we should be very opinionated about what is open content and so open source, because there is no open content without open sources.

16:09So that is our historical context was the COVID indeed, where in that moment, every single school of the planet was adopting, massively adopting closet source software offered for free, you know, so every single children is now a customer of these organizations now because of our choice. so it was easier at the time and now we will see the results if this next the generations will be more active or more passively ingesting technologies without knowing how to improve it so maybe we can change this trend maybe we we can just maybe take the best we can for example in Switzerland we have many data centers focusing and supporting open source software maybe we have many billionaires acquiring big social networks and then then potentially opening it or not we will see but we must have clear announcements that if you want to make it open source it must be an open source

17:14approved license not just open blah blah And so, again, this is just a, if my phone is not locked by monsters, the first point is technologies should support free knowledge, which is something obvious, but is it really, for example, sometimes, I really recommend to have clear directions to support your free content, because indeed, if you create a software which is not compatible with Wikimedia Tool Forge, It will never be a tool.

17:50If we have a software which is not compatible with Wikimedia Cloud, it will never be used by the community. So it's important to have this kind of very clear requirements because it's easier. Also, I would just suggest to don't continue this discussion about security through obscurity Because it's not that if I don't know where I leave my key of my house, then it's safer.

18:21No, my house should be well designed in a way that whatever person, whoever in the planet knows how is designed my house, how is designed Wikipedia, but nobody can enter Wikipedia and get my data. So we tend to do different things while thinking about real things, but Wikipedia? It teaches that open standards are more safe. So I encourage you to do this more documentation, more open documentation, open standards, open everything, because this is the only way to have a safe system, easily auditable for external security experts, and not just by us.

19:01and so again also we have no time really but I encourage you to also open decisions so when you make a decisions don't just communicate these final decisions have open meetings with your community and help us help them in be informed and involved and listen them also in technology thank you does anybody have any questions for our speakers okay two minutes for questions we have one in the back Hi, my name is Ulrike. I have a question for the Open LLM, Open LLM France, and how does this fit into, does this align to any political strategy on the French government side or the European Union side, and do you have funding for it or not?

20:14I think that's that's for me right yeah of course we we do have funding from the French government from the BPI banque public investissement and we are not working alone on this and many other company help us the idea is to create a common because everything about Wikipedia is to have to share knowledge and to share tools and to have something in common so that's what we're doing here So, the only thing that I can say is that yes, there is public money involved, so we are building a public tool. Anyone can join. There is a discourse server if you want to talk with people building it like right now and that does that answer your questions

21:22Maybe a follow-up. I was at the Wikimedia Futures Lab around Wikimedia Europe General Assembly and there was a person from Open Euro LLM at the Charles University in Prague. They're working on something similar but broader scope and more languages. Do you have any insight on how this Open LLM ecosystem works? like do you collaborate in some way or another or are you working completely on your own are there conferences in this direction so i'm i'm not working actively on these so i won't be able to answer very specifically but um i don't think that this has any anything to do with the openlm france team and i think that's a good thing that there are several initiatives all around Europe and the world working on building AI tools for everyone and of

22:22course we share information and ideas and the idea of building open source like real open source with a logiciel libre free license makes sense because because it allows to go faster. And you don't have to be like married to work together. All right. Unfortunately, that's all the time we have for today.

22:57Let's thank our speakers. Thank you very much.

Source and licence

Recording: File:Wikimania 2026 – AI & Governance Lightning Talk Series.webm by Benjamin Bellamy, Valerio Bozzolan, Julie Hunter, NikolasKHF (2026-07-22), via Wikimedia Commons. Licence: CC BY-SA 4.0. The transcript, summary and chapters on this page are shared under the same licence (CC BY-SA 4.0).

The transcript is machine speech-to-text by the TidyTools Audio & video transcriber; the summary, key points and chapters were written by Claude from that transcript. Both may contain errors: check the recording before quoting it.