Felix Kjellberg, the globally recognized YouTuber known as PewDiePie, has found himself at the center of a brewing conflict with OpenAI. In a recent video update, the content creator revealed that his attempts to build "Ajax," a small-scale, locally hosted artificial intelligence model designed for personal productivity and web browsing, led to his account being banned by OpenAI on two separate occasions. The saga underscores the growing tension between the closed-source policies of major AI labs and the rising movement of developers seeking to bring powerful AI capabilities directly to the user’s desktop.
Ajax is categorized by Kjellberg as a "fine-tuned" AI model. In technical terms, this means that an existing, pre-trained base model—in this case, Alibaba’s open-source Qwen 3.5—has been subjected to additional, specialized training to optimize it for specific tasks. Ajax is designed to operate within "Odysseus," an open-source, self-hosted AI application that Kjellberg launched in June. The core philosophy behind this project is decentralization and privacy: by running the model on one’s own hardware rather than relying on cloud-based servers, users can keep sensitive information, such as emails and personal calendar data, within their own digital environment. Furthermore, this approach eliminates recurring subscription fees, shifting the burden of computation onto the user’s own machine.
The model is currently designed as a 9-billion-parameter system. Parameters represent the adjustable variables within a neural network that essentially constitute the "brain" of the AI; generally, a higher parameter count correlates with greater complexity and reasoning capability. While the project is still in development and pending an official release, Kjellberg intends to make it available through his personal website once he is satisfied with its performance and stability.
The Conflict Over Model Distillation and Reasoning Tokens
The friction with OpenAI appears to stem from Kjellberg’s methodology in refining Ajax. He aimed to perform "distillation" using OpenAI’s GPT-5.6 Sol—the company’s flagship model released on July 9. Distillation is a technique where a smaller, more efficient AI model is trained to mimic the outputs of a larger, more sophisticated "genius" model.
The specific goal of this process was to extract Sol’s "reasoning tokens." These are the segments of text a model generates on a hidden "scratchpad" while it deliberates on a complex problem before delivering a final answer. Typically, OpenAI keeps these internal reasoning logs encrypted and inaccessible to users. Kjellberg noted in his video that he discovered research that outlines a method to crack these logs using OpenAI’s own API. He emphasized that he had no intention of hacking or compromising OpenAI’s infrastructure, but the platform’s automated moderation systems flagged his activity regardless.
The first ban occurred during this experimentation phase. Kjellberg disputed the action, arguing that it was unjustified, and his access was eventually restored. However, the respite was short-lived. A second ban was triggered when Kjellberg used Sol to generate "seed data"—the foundational examples used to train a model. OpenAI’s terms of use explicitly prohibit the use of its outputs to develop, train, or improve competing AI models. Consequently, the company enforced its policy, leading to the second suspension of his account.
This incident highlights a broader industry controversy regarding the privacy of internal AI reasoning. In August, researchers demonstrated that encrypted reasoning blocks from major models—including those from OpenAI, Anthropic, and Google—could potentially be replayed to "sister" models to reveal the underlying plain-text thought processes. By analyzing public code repositories, researchers claimed to have decoded over 300,000 of these reasoning sequences. Subsequently, on September 30, OpenAI reported that it had disrupted a campaign involving individuals associated with Moonshot AI, the developers behind the Kimi chatbot, who were allegedly attempting to extract hidden reasoning data. OpenAI stated that it had closed the vulnerability that allowed for the replay of encrypted reasoning logs.
Removing Constraints: The ‘Abliteration’ Process
Beyond the challenges of model development, Kjellberg has been vocal about his desire to create an "uncensored" AI experience. To achieve this, he utilized an open-source tool called Heretic. This tool is designed to identify and remove the built-in refusal mechanisms that cause many modern AI assistants to decline certain prompts or sensitive requests.
This process, often referred to in the AI community as "abliteration," acts as a form of surgical intervention on the model’s weights. By comparing how the AI responds to harmless versus restricted prompts, the tool isolates the behavioral patterns that trigger refusals and suppresses them. Kjellberg candidly admitted that the process is not perfect, noting that Ajax suffered "a little brain damage" as a result of the procedure. Despite this, he claims the model remains highly effective for practical tasks like managing inboxes or browsing the web, succeeding in approximately nine out of ten attempts.
To further refine the model’s performance, Kjellberg has been employing a method known as GRPO (Group Relative Policy Optimization). In this setup, the model attempts a single task 16 times in succession, allowing it to compare its outputs and learn from the most successful approaches. While the model is intentionally designed to be less restrictive, Kjellberg stated that he drew a clear line at preventing the generation of content that could cause physical or psychological harm. He noted that he had consulted with legal counsel to ensure that Ajax would not be utilized to provide dangerous or actionable instructions that could facilitate illegal or harmful activities.
The Pitch for Localized AI
Kjellberg’s advocacy for smaller, local models is rooted in both economic and environmental concerns. He pointed out the staggering resource requirements of "frontier" models—large-scale AI systems that typically feature trillions of parameters. According to his calculations, running a single instance of such a model would require the computing power of nearly 30 high-end machines and an energy consumption profile equivalent to that of 150 residential homes.
By contrast, Ajax represents his effort to create a more sustainable and accessible alternative. He views the current iteration as merely "version one," noting that he still needs to re-run the ablation process, compress (quantize) the model to improve efficiency, and perform rigorous benchmarking to ensure that the "surgery" has not introduced critical errors into the model’s logic.
For followers of the project, the wait for a public release continues. While an initial countdown on his website once pointed toward an early October launch, that timeline has since been removed. As the landscape of AI development shifts toward more specialized, efficient, and locally hosted solutions, Kjellberg’s experiment with Ajax serves as a high-profile case study in the technical and regulatory hurdles faced by those attempting to challenge the dominance of centralized, closed-source AI providers. For now, the developer remains focused on refining the model, balancing his desire for an open, uninhibited AI tool with the practical realities of software engineering and the stringent terms of service set by industry giants.