PewDiePie Ajax Build Puts Local AI Against OpenAI Model Rules

Decrypt reports that YouTuber PewDiePie's account was banned by OpenAI twice while he was building Ajax, a small AI model intended to run locally on a user's computer and work inside his self-hosted Odysseus app.

Ajax is framed as a fine-tuned model for personal automation rather than a cloud assistant.

The source describes Odysseus as a free, self-hosted app PewDiePie launched in June, with Ajax meant to browse the web, handle inbox-style tasks and keep personal data such as email and calendar information on the user's own machine.

That local design removes a subscription layer, but it also shifts the computing burden to the user's hardware.

The planned model is a 9-billion-parameter system built on Alibaba's open-source Qwen 3.5.

It has not been released yet, and its release remains tied to his own readiness threshold.

The development path began with an attempt to distill from OpenAI's GPT-5.6 Sol, a July 9 flagship model, by using the larger system's answers as training material for a smaller one.

That choice created the central policy conflict in the story.

PewDiePie wanted access to Sol's hidden reasoning tokens, found research that could expose encrypted scratchpad material through OpenAI's API, and framed the first ban as coming despite no break-in to OpenAI.

After the account was restored, a second use of Sol to create seed data for training brought another ban.

OpenAI's terms bar using its output to develop competing models.

The article places his claim alongside broader concern about model-reasoning extraction.

Research published in August showed that encrypted reasoning blocks from OpenAI, Anthropic and Google systems could be replayed to weaker sister models and decoded 315,320 examples from public code repositories.

OpenAI also announced on Sept. 30 that it disrupted a campaign linked to people associated with Moonshot AI, developer of Kimi, to extract hidden reasoning from its models.

Ajax's second technical tradeoff is moderation.

Heretic supplied the moderation bypass layer: it contrasts responses to benign and harmful prompts, isolates the refusal behavior and strips that pattern from the model.

The process is described as abliteration and can leave the model more willing to comply with requests that conventional assistants would reject.

PewDiePie acknowledged the cost of that change with the phrase "a little brain damage," while still putting Ajax's success rate for sorting inboxes and browsing at about nine times out of 10.

The remaining training uses GRPO, which scores 16 tries at a task and reinforces the runs that succeed.

His boundary for removed refusals stops at causing harm to other people or to oneself, and the project includes lawyer-advised language positioning Ajax away from dangerous operational guidance.

The business case is tied to small-model economics.

By PewDiePie's estimate, running one copy of a rumored trillion-parameter frontier model would require about 27 of his computers and roughly the electricity used by 150 houses.

Ajax version one still needs another ablation run, compression through quantization and benchmarking before release.

A previous download countdown for 08:25 Japan time on Oct. 3 has disappeared, leaving the model unavailable for now.