Mozilla AI at Internet Archive Europe: Owning Your AI Stack
On 25 June, Internet Archive Europe (IAE) hosted Davide Eynard and Thomas (toto) Bille from Mozilla AI at our Amsterdam space. The afternoon covered a question that sits close to the heart of what we do: when you depend on infrastructure you don’t control, what do you actually own?
Ada, Zangemann, and the case for tinkering
Davide opened with a children’s book: Ada & Zangemann. Ada is a girl who lives in a dumpster, salvages broken hardware, and builds things entirely her own. Zangemann builds beautiful, polished technology that no one else can modify or adapt. The clash between them drives the story, but Davide used it as a frame for something more immediate: most people’s relationship with AI today looks a lot more like Zangemann’s world than Ada’s. You use what you’re given, on the terms it’s offered, for as long as the provider decides to keep it available.
This has nothing to do with technology: it’s a power relationship. And that logic applies to memory institutions every bit as much as to individual developers.
The trade-offs are concrete
To make the point, Davide rebuilt a small tool he’d originally written by hand more than twenty years ago: a script to extract train timetables from a website too clunky to use directly. He produced the rebuilt version with an AI coding assistant, and it worked. But the result lived on someone else’s platform, not his own machine. And unlike the original, which taught him Perl and regular expressions he used for years afterward, this one taught him nothing. Convenience and ownership turned out not to be the same thing.
Search as activity, not action
One of the sharpest distinctions in Davide’s talk was between search as an action and search as an activity. When you type a question into a box and accept the answer, that’s an action. When you use an agent to follow a thread, evaluate what it finds, redirect it when it goes wrong, and build toward a conclusion over time, that’s an activity. The difference matters because the second approach keeps you in the loop. You catch mistakes. You steer.
The clearest example: Davide used an agent to track down the original source of a widely cited Bill Gates quote. The agent searched, hit dead links, found partial copies, and eventually installed a subtitle extraction tool autonomously, downloaded a YouTube video, and identified the exact moment Gates said the thing. Along the way, it returned to the Wayback Machine around a dozen times, working through broken URLs until it found a usable copy.
It was a quiet illustration of something IAE says often: preserved web history is not a nostalgia project. It is working infrastructure. AI systems now depend on it to check what is actually true.
A second demo used a locally run open source model to search the Rijksmuseum’s digital collection for images connected to alchemy and the pursuit of knowledge, producing usable results entirely on Davide’s own hardware. No cloud service, no rented compute, no data leaving the machine.
Otari: making ownership practical
Thomas (toto) Bille followed with a look at the infrastructure side of the problem. Mozilla AI has built Otari as an open-source LLM gateway: a single control plane for all your interactions with language models, whether you run a local model on your own server or route through a commercial API.
The features that generated the most discussion were practical ones. Budget controls granular enough to cap spending per user, per model, or per team. Guardrails that strip personal or sensitive data before it reaches any external provider. A federated router in development that will recommend which model to use based on real usage patterns across the community, with options to prioritise cost, quality, or energy use. Otari is fully self-hostable: if you don’t want your data passing through Mozilla AI’s servers, you don’t have to. The code is on GitHub.
The honest question
Someone in the room asked how Mozilla AI intends to stay financially viable if the tools are free and open source. Toto’s answer was direct: revenue from the hosted version, income from enterprise integration work, and a long-term bet on the community. It’s the same tension that runs through almost every public-interest digital project, including IAE’s own. There’s no clean resolution, but naming it honestly matters.
Why this conversation belongs here
The afternoon drew developers, researchers, and people from across the cultural heritage sector. The discussion after the talks ran for close to an hour.
What connected the room wasn’t a shared technical interest in language models. It was a shared unease about dependency. Memory institutions know what it looks like when access to knowledge sits on infrastructure you don’t own and can’t influence. IAE has spent years arguing that the rights archives have always held offline must be protected online too. The same holds for one layer up, to the AI systems now sitting on top of those archives.
Explore Mozilla AI’s open-source tools, including AnyAgent, AnyLLM, LlamaFile, and Otari, at mozilla.ai and github.com/mozilla-ai.
You can watch a replay of the presentation and conversation on Archive.org and see the slides online.
Mozilla AI at Internet Archive Europe: Owning Your AI Stack Read Post »