Three nights at Hangar on where AI actually runs

Three Thursday evenings in November, the Sala Ricson at Hangar fills with people arguing about where artificial intelligence actually lives. The programme is called aiaiai. It is free. Sessions run on 5, 12 and 19 November, from 7 to 9 pm, at Emilia Coranty 16 in Poblenou. The first evening takes on training data. Xavi Vinaixa, CTO at Sorensen.ai, walks through what he calls the data wall: the point where the supply of scraped English text runs out and the industry goes looking elsewhere. Elsewhere means other languages, and other people's libraries. With Marçal Font he documented L'espoli silenciós, a piece of research that tracks how second-hand books in minoritised languages get bought in bulk across Europe through intermediaries who conceal the end buyer. Some of those books survive in a single copy. Lluís Gómez Bigorda, a Ramón y Cajal researcher at the UAB and the Computer Vision Center, brings InterVisions, a participatory project that asks communities to describe how AI misrepresents them and then to design something better. The second evening moves to hardware. David Cuartielles, one of the Arduino founders, joins from Malmö University, in the room and online. His angle is blunt: the big training campuses are absorbing memory and GPU supply, and the price lands on everyone outside the data centre. Andrea Noni and Miguel Ángel de Heras, who runs open hardware at Hangar's Interaction Lab, join him. The useful comparison is 2005. Cheap dev boards once let a generation of people pull electronics apart and rebuild them for their own purposes, and the open question is whether that still works when the chips go to the hyperscalers first. The third evening is the practical one, and the subject is specialised micromodels. Small systems trained or distilled for one job: transcription, translation, classification, answering questions inside a fixed domain. They run on a laptop, a mini PC, a Raspberry Pi. Iván Paz, who lectures at the FIB and co-founded both the live-coding collective TOPLAP and Axolot.cat, joins Vinaixa and Gómez Bigorda. There is a live demonstration. The claim under test is narrow enough to check: inside its own domain, a small local model can hold its own against something far larger, without shipping your data out and without the energy bill that follows it. Two things make the trip worth it. The format is deliberately flat, built so the room talks back instead of absorbing a lecture, and it sits inside Open Thursdays, Hangar's standing slot for opening technical research to whoever walks in. The other thing is the lineup. Vinaixa and Gómez Bigorda build Catalan language models, sign language recognition tools and participatory bias research. Expect mechanics, with numbers attached. Free admission. Sala Ricson is not a large room, so arrive early. The full programme, session by session, is at the Hangar agenda page: hangar.org/en/agenda-hangar/aiaiai2026

■ Filed from The Foundry: the Fern, Fire wing of the city.