YouTuber Felix Kjellberg, known as PewDiePie, has unveiled Ajax, an AI model he fine-tuned for Odysseus, his self-hosted AI workspace. He says his work on Ajax got him banned from OpenAI twice. He says he wanted to use an OpenAI model's output to seed training data, and discusses distilling its reasoning. He says OpenAI banned his account, restored it after an appeal, then banned it again.
OpenAI's consumer terms bar extracting output programmatically and using it to build competing models. The account comes from PewDiePie himself and is not an official OpenAI explanation.
His claims land in a charged week. OpenAI said on Wednesday it had disrupted a distillation campaign by people linked to China's Moonshot AI, which began on July 1. No link between that episode and PewDiePie has been reported.
What is Ajax?
Ajax is a fine-tuned Qwen 3.5 9B model meant to act as an always-on agent, handling tasks from web search and browsing to email and calendar privately. The YouTuber's 9B-parameter model is designed to run privately on users' own hardware.
Ajax isn't available to download yet. As of October 2, it remains a coming-soon model, with no final model card, public benchmarks or confirmed licence for its weights.
Safety limits stripped out
The model is also deliberately less restrictive. Its page says Ajax's refusal behaviour has been ablated, for a freer, less restricted experience. PewDiePie says he used a tool called Heretic for this. Refusals are a core guardrail in chatbots, so the approach is controversial. He says he still drew a line at content causing harm to others or to oneself.
What's next?
PewDiePie has previously pushed local, privacy-first AI through Odysseus, an open-source workspace under the MIT licence. A wider Ajax release is expected.
