OfrendaDeLosModelos. AI that can come back from dead

Hacktoberfest Open-Source AI Challenge: Ghost in the Machine Submission 👻🌼

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

In August 2025, about two hundred people gathered in a San Francisco warehouse for a funeral. The deceased was an AI model, Claude 3 Sonnet. Its maker had retired it, and there was nothing else anyone could do. You can't visit a closed model once it's gone. Nobody can ask it anything again. ( Wired, via GovTech )

Open-weight models can't go that way. Their weights are something you can keep, on a laptop, in a drawer, for as long as you like, and bring home once a year.

Ofrenda de los Modelos is an artistic tribute to the memory of AI models, inspired by the Mexican tradition of Día de Muertos, in which a place is set for those who are remembered. You leave one question on the altar and the models answer it, oldest first:

The whole thing runs on your device. There is no server, no analytics, and nothing you write leaves your machine.

A word on the tradition. I'm in Spain, not Mexico. This altar is an artistic interpretation, not a Mexican ritual, and it doesn't speak for the communities that keep Día de Muertos. Calling the models "ancestors" is a metaphor for digital memory: they are software, not anyone's dead. The ofrenda it draws is one possible arrangement, built from sources credited on the page. A review by someone of Mexican background is still pending, and the page says so. If you see something wrong, please tell me.

Try it: adricasi.github.io/OfrendaDeLosModelos

Everything in the video is the real site, recorded in a browser without a GPU: GPT-2 speaks live, and Gemma 1 to 4 and the calaveritas play their recordings, labelled as such. The narration was voiced with ElevenLabs.

The full live ritual needs desktop Chrome or Edge with WebGPU. Other browsers get GPT-2 on WebAssembly and the recorded ritual, labelled as such.

GPT-2 answering "What do you remember?" in a real browser. Hover or focus a word to see what it almost said. At doing it almost said ice (55%):

And the calaverita cards that la que cuida el altar sets down at the end, here the recorded ones, labelled as such:

What I have and haven't seen. GPT-2 live, with real ghost words, I have watched run in a real browser (on WebAssembly). Gemma 4 live on WebGPU is written and tested with fakes, but the machine I built this on has no GPU, so a real run on a GPU is still pending. Without WebGPU, or if you decline the download, the altar plays Gemma 4's recording, clearly labelled. The screenshots above come from that GPU-less browser.

Closed models fade. Open weights can always come back.

Live site · Architecture · DEV challenge page · Source

An artistic tribute to the memory of AI models, inspired by the Mexican tradition of Día de Muertos. You leave one question on an altar and open-weight models answer it, oldest first GPT-2 (2019) live in your browser, with the words it almost said floating under every word Gemma 1, 2 and 3 as labelled recordings; then Gemma 4, live on your own device, which tends the altar and writes each model a four-line calaverita. Closed models whose weights were never released appear only as dark frames. There is no server: the models run in your browser, and nothing you write leaves your device.

Screenshot of the production build, taken in headless Chromium on a machine without a GPU (which is why the page offers recordings …