9 okt 2026 · @Waldi
From gaming videos to his own language model
Felix Kjellberg, PewDiePie to his 111 million YouTube subscribers, showed off Ajax on 2 October: an AI assistant that runs entirely on your own computer. Within hours, X was passing around a list of his “character development”. Dropped out of university, biggest YouTuber in the world, books, drawing, philosophy, a move to Japan, Linux, Arch, every Google service gone, learning to code, a room full of GPUs and now his own AI.
The list is funny. It’s also slightly too neat. He didn’t actually train his own LLM from scratch; he fine-tuned an existing one. That doesn’t make the story smaller. It makes it more interesting, because it shows what one determined person with a serious home setup can pull off these days.
What Ajax actually is
Ajax is a fine-tune of Qwen 3.5 9B, an open model from China’s Alibaba. Fine-tuning means taking an existing model and giving it extra training for one specific job. Here, that job is working inside Odysseus, the free, open-source AI workspace Kjellberg released in June. Inside it, Ajax searches the web, sorts your inbox, writes emails and keeps your calendar, without anything going to the cloud.
At 9 billion parameters, Ajax is a dwarf next to the models from OpenAI, Google or Anthropic. That’s precisely the point. Kjellberg’s argument: everyday tasks don’t need a giant cloud model, just a small, specialised one on your own hardware.
He says Ajax v1 gets it right about nine times out of ten on the tasks he built it for. He’s already training version two, letting the model attempt tasks over and over and reinforcing the runs that succeed.
One caveat missing from the enthusiastic posts: Ajax isn’t downloadable yet. There’s no model card and there are no public benchmarks. That nine out of ten is his own measurement. Nothing wrong with that, but it isn’t an independent test.
Banned twice by OpenAI
To make Ajax smarter, Kjellberg wanted it to learn from a much bigger model: GPT-5.6 Sol, the flagship OpenAI released in July. The technique is called distillation. You let a large model generate answers and reasoning, then use that as teaching material for a smaller one. He used Sol’s reasoning chains to create seed data for Ajax.
OpenAI did what OpenAI does. Account deactivated, reason: “Distillation”. After an appeal the account came back; on the next attempt it was closed again. OpenAI’s terms prohibit using its output to build competing models. Worth noting: this is Kjellberg’s version of events. OpenAI hasn’t commented.
His response was dry. AI companies trained their models on huge amounts of public data, so why shouldn’t users take a bit back? In his words: “How dare we, as consumers, take part of their hard work stolen from us?” Legally, he’s on thin ice. But it’s a question the industry hasn’t answered convincingly. Especially since OpenAI said that same week it had shut down a distillation campaign by people linked to China’s Moonshot AI.
Less ‘no’ in the model
Ajax is also deliberately less cautious. Kjellberg used Heretic, an open-source tool that strips the built-in refusals out of a model. His example: the original Qwen model refuses to say what happened in China on 4 June 1989. Ajax doesn’t. He says he kept limits around requests that could help someone harm themselves or others. It wasn’t free: he admits the model took a bit of brain damage in the process.
The interesting question isn’t whether this is allowed. It’s who decides what a model may say. With a cloud model, that’s the provider. With a local model, it’s you, with all the responsibility that comes with it.
What you can do with this
You don’t need to buy a room full of GPUs. Ajax is still an experiment, not a replacement for ChatGPT or Claude. The interesting part is the direction. Small models are getting good enough for specific tasks, run on hardware you own and keep your email and calendar with you. For companies still holding off on AI because customer data can’t go to American servers, that’s not a detail.
If you want to try it yourself, look at Odysseus. That one is available, open source under the MIT licence, and works with other open models while Ajax keeps us waiting.
PewDiePie hasn’t built an OpenAI killer. He’s shown that one person with persistence and a hefty electricity bill can assemble a useful, private assistant. The biggest YouTuber in the world picking a fight with OpenAI is the news. Small, local AI growing up is the story.
