OpenAI Banned PewDiePie Twice While He Built Ajax, an Uncensored AI That Will Run on Your PC

YouTuber Felix 'PewDiePie' Kjellberg says OpenAI suspended his account twice while he trained Ajax, a small, uncensored assistant designed to run on a personal computer. Ajax is described as a 9-billion-parameter model fine-tuned from Alibaba's Qwen 3.5 and integrated into PewDiePie's self-hosted app Odysseus.

By AI Newsroom· Reviewed by Pranav, Founder & Editor-in-ChiefPublished 1 minute agoUpdated 1 minute ago0 views
OpenAI Banned PewDiePie Twice While He Built Ajax, an Uncensored AI That Will Run on Your PC

Why It Matters

The episode highlights tensions between makers of large proprietary models and developers creating smaller, user-run alternatives, and touches on techniques used to extract and remove safety filters from AI systems — issues relevant to debate over model security and harms. It also underscores interest in self-hosted models as a privacy and cost trade-off compared with cloud-hosted services.

Key Facts

  • Subject: YouTuber Felix 'PewDiePie' Kjellberg
  • Allegation: OpenAI suspended his account twice during development
  • Model name: Ajax
  • Model size and base: 9-billion-parameter model built on Alibaba's Qwen 3.5
  • Tool used to remove refusals: Heretic (open-source)

PewDiePie said in a new video that OpenAI suspended his account on two separate occasions while he was training Ajax, a compact AI assistant intended to run on a user's own PC and browse the web on their behalf. Ajax is being developed as a fine-tuned model to operate inside Odysseus, the self-hosted AI application PewDiePie released in June. He described Ajax as a 9-billion-parameter model derived from Alibaba's open-source Qwen 3.5.

According to PewDiePie, his first ban occurred after he attempted to distill capabilities from OpenAI's GPT-5.6 Sol — a process in which a smaller model is trained to reproduce a larger model's outputs. He said he sought to capture Sol's internal 'reasoning tokens' but was suspended; the account was later reinstated. After using Sol again to generate seed training data, he says OpenAI suspended him a second time, which he attributes to the company's rule forbidding its outputs being used to build competing models.

To make Ajax answer requests the way some conversational assistants refuse to, PewDiePie used an open-source tool called Heretic to remove the model's refusals, a process sometimes called 'abliteration.' He acknowledged the technique can degrade performance and said Ajax suffered "a little brain damage," though he claims it performs tasks like sorting email and browsing successfully most of the time. He also said he is continuing training with GRPO, a method that has the model attempt tasks repeatedly and learn from the successful attempts.

The wider context includes recent research showing that encrypted internal reasoning artifacts from models at OpenAI, Anthropic and Google can be replayed and recovered by weaker models; researchers decoded 315,320 such blocks from public code. OpenAI has said it disrupted a campaign tied to people associated with Moonshot AI on Sept. 30 that attempted to extract its models' hidden reasoning. Tech City Authority, reporting on the story, found no public response from OpenAI to PewDiePie's claims. PewDiePie has not released Ajax for download yet; a prior countdown to 08:25 Japan time on October 3 was removed from the download page as he continues further ablation, quantization and benchmarking.

Keep Reading