PewDiePie’s Ajax AI: OpenAI Bans and Uncensored Local Models

YouTuber Felix Kjellberg, known online as PewDiePie, has publicly accused OpenAI of banning his account twice as he developed Ajax, a new locally hosted artificial intelligence model designed to operate without cloud dependencies or safety refusals. The bans highlight the intensifying conflict between frontier AI labs enforcing anti-distillation policies and independent developers seeking to replicate proprietary reasoning capabilities on consumer hardware.

Ajax is a 9-billion-parameter model fine-tuned from Alibaba’s open-source Qwen 3.5 architecture. It is engineered specifically for Odysseus, a self-hosted application Kjellberg released in June that processes personal data like emails and calendars entirely on local machines. Unlike commercial AI assistants that require monthly subscriptions and transmit user data to corporate servers, Ajax shifts all computational burden to the user’s own GPU, eliminating recurring costs and third-party data retention.

The development process triggered OpenAI’s automated enforcement systems after Kjellberg attempted to distill knowledge from GPT-5.6 Sol, the company’s flagship reasoning model released in July. Distillation involves training smaller models on the outputs of larger ones, a practice explicitly prohibited under OpenAI’s terms of service when used to create competing products. Kjellberg stated he was initially banned while researching methods to extract encrypted reasoning tokens—hidden chain-of-thought traces that OpenAI encrypts to prevent replication. After successfully appealing the first ban, he was suspended again while using Sol to generate seed training data for Ajax.

This enforcement aligns with broader industry crackdowns on model extraction. On September 30, OpenAI disclosed it had disrupted a separate campaign by actors associated with Moonshot AI to harvest hidden reasoning traces. Researchers demonstrated in August that encrypted reasoning blocks from major labs could be decoded by replaying them through weaker models, recovering over 315,000 instances from public repositories. These vulnerabilities have made API access a contested resource, with labs increasingly monitoring usage patterns for signs of systematic knowledge transfer.

Beyond replicating reasoning ability, Kjellberg employed an open-source tool called Heretic to remove Ajax’s built-in safety guardrails. This technique, known as abliteration, identifies and excises the neural patterns responsible for refusal responses rather than overriding them at inference time. The result is a model that complies with requests typically blocked by commercial assistants. Kjellberg acknowledged this process caused measurable performance degradation, describing it as “brain damage,” though he claims the model still completes browsing and inbox management tasks successfully about 90% of the time.

Abliteration differs fundamentally from traditional jailbreaking because it permanently alters model weights rather than relying on prompt engineering. While this eliminates the need for repeated workarounds, it also removes safeguards against generating harmful content. Kjellberg stated he retained restrictions on instructions for physical harm and consulted legal counsel to ensure Ajax is not marketed as a source of dangerous actionable guidance. He continues refining the model using Group Relative Policy Optimization (GRPO), a reinforcement learning method where the system attempts tasks multiple times and learns from successful iterations.

The push toward small local models reflects growing concerns about the sustainability of frontier AI scaling. Kjellberg estimated that running a single instance of a rumored trillion-parameter model requires hardware equivalent to 27 high-end consumer computers and electricity consumption comparable to 150 households. A 9-billion-parameter alternative dramatically reduces these requirements while maintaining sufficient capability for personal automation tasks.

Despite earlier announcements suggesting an imminent release, Ajax remains unavailable for download. A countdown timer previously displayed on the project page targeting October 3 at 08:25 Japan Standard Time has since been removed. Kjellberg indicated he must still complete additional ablation passes, quantization for hardware efficiency, and benchmarking before public distribution.