The council of models. PewDiePie, Linux and artificial intelligence at home

Felix Kjellberg, Swedish, resident in Japan, is PewDiePie: for years the most-followed YouTube channel in the world. Since April 2025 his videos have told, in the tone of someone commentating a video game, of a journey many companies are making with much more seriousness and much less fun: bringing their own data home.

First: away from Windows

On 26 April 2025 he published I installed Linux (so should you): more than eight million views. The opening: "For years, I have been tortured by Windows. But now, alas, I am finally free." The reasons are four, in order: control, advertising inside an operating system he paid for, the possibility of changing everything, and the fact that "gaming on Linux is finally real" thanks to Valve and the Steam Deck. Linux Mint on the desktop because it "very much feels like Windows just better"; Arch on the laptop "partially because of the meme".

On 26 June 2025 came I'm DONE with Google: search engine, browser, e-mail on his own domain, a Pixel with GrapheneOS, and the first services hosted at home, a password manager on a Raspberry Pi and notes on a Steam Deck used as a server. "My first successful self-host. My god. No more subscriptions." The one renunciation that weighs on him is Google Maps: "The first time I used my open source alternative, I ended up 30 minutes late."

Second: away from the chatbots

On 31 October 2025 came STOP. Using AI Right now, which despite its title is a video in favour of artificial intelligence, provided it is one's own. The computer has ten graphics cards, reached with a BIOS trick that came "from a random stranger online", and costs "20,000 roughly". According to the specialist press they are eight RTX 4090s modified in China to have 48 gigabytes of memory each, plus two professional cards.

On it he runs the open models: first Llama 70B, then OpenAI's gpt-oss-120b ("it feels exactly like ChatGPT, but it's fast [...] I can't read that fast"), then a Qwen with 235 billion parameters. The interface he writes himself, or rather has the models themselves write: "the machine is making the machine". Then he feeds the system the whole archive Google had on him and asks it "who am I?". The answer: "Oh you're Felix Kjellberg. Oh you're PewDiePie. This is your address and here's your phone number."

The part everyone remembers is the council. Eight copies of the same model, each with a different personality, answer the same question and then vote: "Democracy!" The useless members are eliminated and replaced by a system that generates new ones. When he leaves in their logs the warning that whoever does not perform will be deleted, the members begin to wonder "what is this sick game that someone came up with?" and to vote strategically, helping each other. "Betrayed by my own council. [...] I just changed the model to a dumber one and problem solved." Then the swarm: sixty-four small models, of two billion parameters, queried together. "It doesn't lag at all. It's just that my web UI, it can't handle it."

Third: a model of his own

On 27 February 2026 he recounted having trained a coding model starting from Qwen 32B. The first result is excellent, and false: "I have beat ChatGPT. It felt so goddamn good. STOP. STOP. DON'T LISTEN to this guy. The benchmark is invalid. I did not check for contamination." With the test cleaned up, the honest result is 39.1%: better than the GPT-4o of a year earlier, worse than the Qwen 3 released in the meantime, as he himself admits. "I almost burned down my house twice for this project."

On 31 May 2026 he released Odysseus, a complete interface for models at home, under a free licence: "Basically, it's Claude and ChatGPT's web UI, but self-hosted." "No tracking, no subscriptions, no funny business." By October it has almost ninety thousand stars on GitHub; the part that works on its own is borrowed from OpenCode, "I literally just stole OpenCode. OpenCode is really good." On 30 September 2026 he announced Ajax, a small model of nine billion parameters, derived from Qwen3.5 and trained for that interface; as of 3 October it is not yet downloadable. To build its data he tried to use the answers of an OpenAI model, and OpenAI banned him. Twice. He admits it and advises against doing it. His thesis, though, stands: "I think it makes no sense to poke that giant beast to be like, 'Hey, can you check my email for this one thing?' I really strongly believe in the small model trained for a specific harness."

The moral, for those who are not PewDiePie

No company needs ten modified graphics cards, nor a council of eight models that votes. But there are three true things in this story, and they are the ones we tell clients with fewer special effects.

1. Data at home is a possible choice, and for a confidential document it is the right one. The moment the system recites to him his address and telephone number taken from Google's archive says better than any privacy notice where data end up.
2. A small model, for a specific task, is often enough. He says so after trying the enormous ones, and the measurements say so: on supplied documents the open models of thirty billion parameters are close to the frontier.
3. Results must be verified. The conspiring council and the contaminated benchmark are the same lesson: an artificial intelligence system always answers, and with confidence. Verification is the work.

Even David Heinemeier Hansson, who does not believe in models at home, at the end of August 2026 spoke about him for a quarter of an hour as a role model: "you can go from being funny guy on Minecraft streams to, 'Okay, this is my life now. I'm building AI clusters and doing the Council of AIs,' and we all get to spectate? I mean, what a treasure."

Sources

• PewDiePie YouTube channel: I installed Linux (so should you), 26 April 2025; I'm DONE with Google, 26 June 2025; STOP. Using AI Right now, 31 October 2025; I Trained My Own AI... It beat ChatGPT, 27 February 2026; MY trillion $Dollar Project is finally OUT!, 31 May 2026; I seriously should NOT be dropping this., 30 September 2026. Quotations from the video transcripts, read on 2 October 2026.
• github.com/odysseus-dev/odysseus, read on 2 October 2026.
• Tom's Hardware, 3 November 2025 and 2 October 2026; Hugging Face (Vaibhav Srivastav), 5 November 2025; PC Gamer, 6 November 2025; XDA, 28 April 2025 and 26 June 2026.
• Lex Fridman Podcast no. 501, 26 August 2026, transcript.