Layla Zulfa
Felix Kjellberg, known globally as PewDiePie, has revealed the technical hurdles and regulatory challenges he faced during the creation of Ajax, a specialized small artificial intelligence (AI) model. In a comprehensive video report released on his YouTube channel, the creator detailed how his development process led to two separate bans from OpenAI, the organization behind ChatGPT. Ajax is designed as a fine-tuned, 9-billion-parameter model intended to operate locally on a user’s personal computer, serving as a primary component of his self-hosted AI application, Odysseus.
The development of Ajax represents a growing trend among tech-savvy creators and developers who are seeking to move away from centralized, subscription-based AI services in favor of private, locally-run alternatives. By running on personal hardware, Ajax allows users to integrate their personal data—such as emails, calendars, and private documents—into an AI workflow without transmitting sensitive information to external corporate servers. However, the pursuit of this localized autonomy has brought Kjellberg into direct conflict with the terms of service and security protocols of major AI industry leaders.
The Nature of the OpenAI Bans and Distillation Efforts
The primary conflict between Kjellberg and OpenAI stemmed from a process known as "distillation." In the field of machine learning, distillation involves using a larger, more sophisticated "teacher" model to train a smaller "student" model. Kjellberg attempted to utilize OpenAI’s flagship model, referred to as GPT-5.6 Sol, to improve the reasoning capabilities of Ajax. The goal was to capture the "reasoning tokens"—hidden internal processes the model uses to solve complex problems—to enhance the performance of the much smaller Ajax model.
According to Kjellberg’s account, the first ban occurred after he attempted to access these reasoning blocks. OpenAI typically keeps the "scratchpad" where these tokens are generated encrypted and inaccessible to users. Kjellberg referenced a specific study suggesting that OpenAI’s own Application Programming Interface (API) could be used to decrypt or "crack" these blocks. While he maintained that he did not successfully breach OpenAI’s security, the system’s automated defenses flagged the activity, resulting in an immediate account suspension. Though the account was eventually restored following a dispute, a second ban followed shortly thereafter.
The second suspension was triggered when Kjellberg used GPT-5.6 Sol to generate "seed data"—the foundational examples used to teach a model how to respond to various prompts. OpenAI’s terms of use explicitly prohibit the use of its models’ outputs to develop competing AI products. By using the flagship model’s answers to train Ajax, Kjellberg ran afoul of these proprietary protections. This incident mirrors broader industry tensions; in late September, OpenAI reported disrupting similar campaigns by groups associated with Moonshot AI, the developer of the Kimi model, which were also aimed at extracting hidden reasoning data.
Technical Architecture and the Choice of Qwen 3.5
Ajax is not built from scratch but is instead a fine-tuned version of Alibaba’s open-source model, Qwen 3.5. With 9 billion parameters, Ajax sits in the category of Small Language Models (SLMs). In AI development, parameters are the variables the model learns from its training data, effectively serving as the "synapses" of the digital brain. While frontier models like those from OpenAI or Google often boast hundreds of billions or even trillions of parameters, SLMs like Ajax are optimized for efficiency and specific task performance.
The choice of the Qwen 3.5 base is significant, as the model has gained a reputation for high performance in coding and mathematical reasoning tasks relative to its size. Kjellberg’s objective with Ajax is to create an agent capable of browsing the web and managing local files autonomously. To achieve this, he utilized a training method known as Group Relative Policy Optimization (GRPO). This technique involves the model attempting a single task multiple times—often 16 iterations—and then analyzing which attempts were successful. By reinforcing the successful paths, the model "learns" to perform tasks more reliably without the massive overhead required by larger models.
The Controversy of "Abliteration" and Uncensored AI
A central and controversial feature of the Ajax project is Kjellberg’s pursuit of an "uncensored" model. Most commercial AI assistants, such as ChatGPT or Claude, are programmed with "refusals"—hardcoded safety guardrails that prevent the model from answering requests deemed harmful, biased, or inappropriate. Kjellberg argued that these guardrails can often be overactive, preventing the AI from performing legitimate tasks or providing objective information.
To bypass these restrictions, the development of Ajax involved a technique known as "abliteration." Using an open-source tool called Heretic, the developers analyzed how the model reacted to various prompts to identify the neural patterns associated with refusal. Once identified, these patterns were essentially "cut out" of the model’s architecture. Kjellberg described this as a form of "lobotomy" for the AI, ensuring it would no longer refuse user requests.
However, this surgical removal of safety features is not without consequences. Kjellberg admitted that the process caused "a little brain damage" to the model, slightly degrading its overall coherence or performance in certain areas. Despite this, he claimed that the model remains effective, handling tasks like inbox sorting and web browsing with roughly 90% accuracy.
From a legal and ethical standpoint, Kjellberg emphasized that certain boundaries remain. He stated that he chose which refusals to remove, specifically retaining protections against the AI assisting in self-harm or causing harm to others. His legal counsel reportedly advised him to include clear disclaimers that Ajax is not intended to provide dangerous or actionable instructions, highlighting the complex legal landscape surrounding the distribution of "uncensored" open-source software.
Economic and Environmental Implications of Local AI
Beyond the issues of privacy and censorship, Kjellberg’s push for Ajax is rooted in a critique of the massive energy and financial costs associated with frontier AI models. He presented a "napkin math" analysis comparing the requirements of a trillion-parameter model to those of a local SLM. According to his estimates, a single instance of a major frontier model requires the computing power equivalent to 27 high-end consumer PCs and consumes as much electricity as approximately 150 average households.
By contrast, Ajax is designed to run on a single consumer-grade machine. While the user’s hardware must perform the "heavy lifting," there are no ongoing subscription fees, and the environmental footprint is localized to the user’s own energy consumption. This decentralized approach is seen by many in the tech community as a necessary counterweight to the "AI industrial complex," where a handful of corporations control the most powerful models and the data fed into them.
Timeline and Future Release Prospects
The journey of the Ajax model is closely tied to the Odysseus application, which Kjellberg launched in June as a free, self-hosted platform for AI interaction. Odysseus was intended to be the shell within which Ajax would eventually function as the primary engine.
- June: Launch of Odysseus, the self-hosted AI application.
- July 9: Release of OpenAI’s GPT-5.6 Sol, which Kjellberg attempted to use for distillation.
- August: Researchers demonstrate that encrypted reasoning blocks from major models can be replayed to weaker models to extract plain text, a technique Kjellberg explored.
- September 30: OpenAI officially announces a crackdown on "reasoning extraction" campaigns.
- Early October: Kjellberg releases his video update detailing the bans and the current state of Ajax.
As of early October, Ajax has not yet been released for public download. A countdown timer previously appeared on Kjellberg’s website, indicating a release around 08:25 Japan time on October 3. However, that timer has since been removed. Kjellberg stated that he still needs to perform further tests, including rerunning the ablation process, quantizing the model (a method of compressing the AI to make it run faster on less powerful hardware), and conducting thorough benchmarks to ensure the "brain damage" sustained during training does not render the model unreliable.
Analysis of the Broader AI Landscape
The saga of PewDiePie and Ajax highlights a pivotal moment in the evolution of artificial intelligence. It underscores the tension between "closed" AI systems—maintained by corporations with strict safety protocols and proprietary interests—and the "open" AI movement, which prioritizes transparency, local control, and the removal of centralized filters.
The fact that a single individual with a significant platform can attempt to "distill" knowledge from the world’s most advanced AI models suggests that the "moats" built by companies like OpenAI may be more porous than previously thought. As techniques like abliteration and GRPO become more accessible, the barrier to creating powerful, specialized, and uncensored AI continues to drop.
For the industry, this raises urgent questions regarding safety and accountability. If "uncensored" models become the norm in the open-source community, the traditional methods of AI safety—which rely on centralized oversight—may become obsolete. For users, however, the Ajax project represents a potential future where the power of AI is no longer a rented service, but a personal tool owned and operated entirely on one’s own desk.
