PewDiePie says OpenAI banned him twice while building uncensored AI

YouTuber PewDiePie, whose real name is Felix Kjellberg, stated in a recent video that OpenAI banned his account twice while he was developing Ajax, a small AI model intended to run on personal computers. He described Ajax as a "fine-tuned AI model" designed to work within Odysseus, his self-hosted AI application launched in June.

PewDiePie explained that he used OpenAI's GPT-5.6 Sol model, released on July 9, to generate training data. He aimed to extract the model's "reasoning tokens," which are typically encrypted. He claimed he found a method to access these tokens via OpenAI's API, leading to his first ban, which he disputed and was later overturned. Following the account's restoration, he used Sol again for seed data, resulting in a second ban, which he attributed to OpenAI's terms of service prohibiting the use of its output to develop competing models.

To make Ajax uncensored, PewDiePie utilized Heretic, an open-source tool designed to remove an AI model's built-in refusal mechanisms. This process, described as "abliteration," essentially removes the AI's ability to decline requests. PewDiePie acknowledged that this "surgery" may have caused some "brain damage" to the model, but stated that Ajax handles tasks like sorting inboxes and browsing successfully about nine out of ten times. He also mentioned ongoing training with the GRPO method and that he personally set the limits for removals, drawing a line at causing harm to others. His lawyer advised him to state that Ajax is not designed to provide dangerous instructions.

Ajax is reportedly a 9-billion-parameter model built on Alibaba's Qwen 3.5. PewDiePie advocates for smaller AI models due to cost and energy efficiency compared to larger frontier models. A countdown on the download page for Ajax, initially set for October 3, has since been removed.