PewDiePie's Ajax: Local AI That OpenAI Banned Him Twice To Stop
On October 2, PewDiePie unveiled Ajax, an uncensored AI model with 9 billion parameters designed to run on home PCs through his Odysseus workspace. The release is a moment—not because it's technically groundbreaking, but because it crystallizes a fight happening quietly across open-source AI: who owns a model's outputs, and can you use them to train your own?
OpenAI banned his account twice during development after detecting distillation, where outputs from one model train another. In the announcement video, Kjellberg showed the rejection emails. On one, the reason was simply listed as "Distillation."
What Ajax Actually Does
Ajax is a fine-tuned Qwen3.5-9B model that powers the workspace to handle "daily tasks from search to browse the web to email to your calendar," all "completely privately". Odysseus bundles chat, autonomous agents, deep research, model comparisons, and email management into one dashboard, plus notes, calendar, tasks, and an image editor. Ajax runs it all locally.
Early testing showed roughly 90% task completion rates, though Kjellberg calls it a work in progress. The model ships with a native context window of 262,144 tokens—wider than many commercial assistants.
The Censorship Angle
The model was "uncensored" using the open-source Heretic abliteration tool. This is where the story gets culture-relevant. A model's weights hold the patterns that make it refuse certain requests. Ablation finds the weights tied to those refusals and removes them, so the model stops saying no.
Kjellberg says he still drew lines. Ajax's refusal behaviour has been ablated, for a freer, less restricted experience. He says he still drew a line at content causing harm to others or to oneself. But the "uncensored" label is already making people uneasy—it invokes the same controversy around other abliterated models released earlier this year.
Why OpenAI Cared Enough to Ban
He wanted to use an OpenAI model's output to seed training data, and discusses distilling its reasoning. This isn't new technically, but it's newly personal. Frontier AI companies such as OpenAI and Anthropic actively restrict distillation to protect their models' reasoning processes, treating unauthorized extraction attempts as security threats.
The bans raise a question the industry has been circling: If I pay for API access, do those outputs become mine? Frontier labs say no. The open-source community increasingly disagrees.
The Bigger Picture
PewDiePie said Odysseus does not automatically collect user data, so he asked users to contribute examples voluntarily. He built a submission website and tutorials, but says the response was smaller than expected. He ended up generating task examples and refining Ajax through reinforcement learning. He describes running the same task 16 times and using successful attempts to reinforce the behavior he wanted.
What's striking is that this is a creator, not a research lab, building an alternative to ChatGPT—and he's being penalized for it. The Odysseus workspace operates under an AGPL-3.0-or-later open-source license and remains free to use. The model weights remain private as of launch. The download link currently directs users to a data contribution page rather than the model itself.
Ajax isn't out yet, but the story is already clear: it shows what's possible when someone with a platform decides cloud AI isn't worth the control loss. It also shows what frontier labs will do to stop it.