A developer posted to Hacker News this week with a quietly heartbreaking title: Show HN: I built 80 mini-games using Fable before it was shut down. The games live at minigames.world, every premium pack and subscription tier unlocked, all free — not out of generosity, but because the platform underneath them went dark days after launch. The author’s note is almost cheerful about it. That cheerfulness is the part that should scare you.

Here is the pattern, and it is older than any model. You build eighty things on top of a service. The service is fast, the service is free, the service is gone. PCMag, which ran a hands-on test before the lights went out, called it the most powerful public model it had touched and then watched it disappear in the same article. Eighty games survived only because their maker had the files; the platform did not get a vote.

←TODAY: A free AI beta vanished days after launch, and 80 little games nearly went with it. →3012: The studios still standing in the Zurich-3012 horizon are the ones whose files open without a vendor’s permission. Fulcrum: A tool you can clone outlives the company that made it — that is the only durability that compounds.

I have spent eleven years in the ugly seam between the BIM model and the actual building site, and I will tell you the buildings that aged worst in my time were never the ugly ones. They were the ones nobody could repair because the proprietary format went dark and took the geometry with it. A shut-down game platform is the same failure mode wearing a friendlier hat. The question to ask this quarter is not “is it good?” It is “when the vendor disappears, can a 25-year-old still open the file?”

The Tool: The encouraging news — and the antidote — came from the same week. Xenova (Joshua, of Hugging Face) showed Gemma 4 running at 255 tok/s on WebGPU and then released the demo and kernels so anyone could run it locally in their browser. The underlying project is @huggingface/transformers (transformers.js): a JavaScript library that runs real models entirely on your machine — no server, no API key, no account that can be banned. It is worth an architect’s afternoon because it turns “the AI” from a subscription you rent into a file you own.

Setup:

# Node 18+; the same model runs in the browser or the terminal
npm create vite@latest local-ai -- --template vanilla
cd local-ai
npm install @huggingface/transformers
# paste the snippet below into main.js, then:
npm run dev   # open the printed localhost URL, watch the console

First steps:

1. Open main.js and delete the Vite boilerplate — you only need a few lines.
2. Import a pipeline and pin device: 'webgpu' so it runs on your GPU, not a remote one.
3. Reload the page; the first run downloads the model once, then caches it. Watch the console print a result with no network call after that.
4. Pull the network cable. It still works. That is the whole lesson in one gesture.

Atelier: A Swiss studio reaches for this the moment a model needs to read project text — minutes, Bauleitung notes, a clash report — without that text leaving the office. Local inference means the Datenschutz conversation is over before it starts: nothing crosses the building’s WLAN, so there is no cloud contract to vet and no vendor to outlive. It slots straight into the kind of internal Archicad–Speckle–Bonsai pipeline a dozen practices already run and zero conferences talk about.

Hack: This Hack teaches you to run a model entirely on your own machine, so no shutdown can ever take it from you. The medium is runnable JavaScript; the domain is AI/ML.

import { pipeline } from '@huggingface/transformers';
// runs fully in-browser on WebGPU -- no server, no key, no vendor
const classify = await pipeline('sentiment-analysis', null, { device: 'webgpu' });
console.log(await classify('IFC round-trip survived the upgrade'));

Four lines, and the model is now a dependency you control like any other file in the repo. If the company behind the weights folds tomorrow, your node_modules does not notice. That is the difference between a tool and a tenancy.

None of this means the cloud is the enemy — the honest trade-off is real: local models are smaller and slower than the hosted frontier, and you pay in VRAM and setup time for the privilege of owning them. But “smaller and still here next year” beats “enormous and gone by Thursday” every time you actually have to ship a project. The developer with 80 games learned that the easy way, because he kept his files. So keep yours: this week, take one workflow that currently calls a hosted API and stand up a local fallback for it — even a worse one — so that when the platform goes dark, you keep working.

Learn-it:

• The shut-down platform: minigames.world — the 80 mini-games, still playable because the maker kept the files.
• Local inference, proven: Xenova’s WebGPU demo + released kernels — run a model locally in your browser.
• The hands-on test: PCMag tried it before it disappeared — read it as a parable about renting, not owning.
• PAZ note: ask one question of any tool stack this quarter — when the vendor disappears, can a 25-year-old still open the file?