Summary
- PewDiePie revealed in a recent video that OpenAI banned his account on two occasions while he was working on Ajax, a small AI model designed for home computers. The Tech City Authority noted that there was no public response from OpenAI regarding this matter.
- After his account was reinstated, he utilized GPT-5.6 Sol to generate training data and then removed the model's built-in refusals using the open-source tool Heretic.
- Ajax is said to be a 9-billion-parameter model based on Alibaba's Qwen 3.5. A countdown for its release is set to conclude at 08:25 Japan time on October 3.
YouTube star PewDiePie has claimed that OpenAI banned his account twice during the development of Ajax, an AI model intended to run locally on personal computers and assist with web browsing. He shared his experience in a recent video on his YouTube channel.
Ajax is described by PewDiePie as “a fine-tuned AI model,” which indicates that it has undergone additional training for a specific purpose. This purpose involves functioning within Odysseus, the free AI application he launched in June that users can host themselves.
Myriad: Predict Nvidia's peak price.Operating on personal hardware, Ajax ensures that users keep their emails and calendars private, avoiding storage on corporate servers. Moreover, there is no subscription fee, although users must provide the necessary computing power to run the model.
PewDiePie claims Ajax is a 9-billion-parameter model utilizing Alibaba's open-source Qwen 3.5. The parameters are the adjustable components that comprise an AI's functionality, with a higher number generally indicating increased capability. The model is awaiting release and will be available on his website once it is deemed ready.
Details of the Bans
PewDiePie aimed to distill information from GPT-5.6 Sol, the flagship model released by OpenAI on July 9. Distillation refers to the process of training a smaller AI using the answers generated by a larger model, akin to a student imitating a top-performing peer's homework.
The objective was to access Sol's reasoning tokens, which are snippets of text the model generates on a hidden scratchpad while solving problems. OpenAI encrypts this scratchpad.
In his video, PewDiePie mentioned discovering a study that could decrypt this information using OpenAI's own API. He asserted that he never intended to exploit OpenAI's resources, yet he was still banned. He contested the ban, which was later lifted.
Subsequently, he used Sol again to create seed data, initial examples for training the model, but faced a second ban from OpenAI. The company's terms prohibit using its output to develop rival models, which led to his account being banned once more.
Researchers revealed in August that they could replay encrypted reasoning blocks from OpenAI, Anthropic, and Google to a less powerful model that would output them in plain text. They successfully decoded over 315,000 instances from publicly available code repositories.
On September 30, OpenAI announced it had thwarted an effort by individuals linked with Moonshot AI, the creator of Kimi, to extract hidden reasoning from its models. The company stated it had closed a method that allowed users to replay encrypted reasoning from others to retrieve its contents.
Creating an Uncensored AI
To ensure Ajax is uncensored—meaning it responds to requests that typical AI assistants refuse—PewDiePie employed Heretic, an open-source tool designed to eliminate a model's inherent refusals. This tool analyzes the model's responses to harmful and harmless prompts, identifies the reasons behind its refusals, and removes those restrictions.
This process, termed “abliteration,” effectively adjusts the model so it does not need to be jailbroken, allowing it to comply with all requests without hesitation.
PewDiePie noted that this modification is not without its drawbacks, admitting that Ajax experienced “a little brain damage,” yet it performs tasks such as sorting emails and browsing successfully about 90% of the time.
He is continuing to refine the model using GRPO, a technique where the model attempts the same task 16 times to learn from successful attempts.
He specified the criteria for removing refusals, drawing a line at “causing harm to others or to oneself.” His legal advisor advised him to clarify that Ajax is not intended to provide dangerous or actionable instructions.
Advocating for Smaller Models
PewDiePie argues that smaller AI models are more cost-effective and energy-efficient. According to his estimations, a single instance of a rumored trillion-parameter model would require around 27 of his computers and the energy consumption equivalent to approximately 150 homes.
PewDiePie refers to Ajax as version one, stating that he still needs to rerun the ablation process, compress (quantize) it, and benchmark the model to ensure its integrity.
Ajax is not yet available for download. Initially, it featured a countdown timer for 08:25 Japan time on October 3, but that countdown has since been removed.