The ban was about more than a chatbot #
OpenAI suspended Felix Kjellberg’s account, known as PewDiePie, then restored access after an appeal, and subsequently banned him again. The stated reason for the suspensions was distillation, a process OpenAI prohibits under its Terms of Use for building competing models. Kjellberg’s appeal reportedly stated, “I didn’t distill, I swear.”
Distillation involves training a smaller model using responses from a larger, more powerful model. OpenAI’s terms restrict users from employing its outputs to develop models that compete with their services. Specifically, Section 2c of OpenAI’s Terms of Use prohibits using output from their models to train other models.
Kjellberg’s project, dubbed Ajax, is a fine-tuned Qwen 3.5 9B model designed for local operation. He reportedly sought to strip out censorship from the model. To generate training data, Kjellberg intended to distill OpenAI’s model outputs. OpenAI reportedly encrypts the model's reasoning, often referred to as "chain-of-thought" (CoT) tokens.
He reportedly located a study demonstrating decryption of this reasoning using OpenAI’s own APIs. Kjellberg stated this method "worked," but OpenAI banned him for distillation. His account was unbanned after an appeal, but further activity led to a second ban. OpenAI has recently targeted similar large-scale distillation campaigns, including those attributed to developers associated with Moonshot AI's Kimi.
Ajax brings the AI home #
Felix Kjellberg, known as PewDiePie, reportedly developed a local AI project named Ajax. The system is a fine-tune of Alibaba’s Qwen 3.5 9B model, designed to operate on a personal computer, reducing dependence on external cloud services.
Running a 9-billion-parameter model locally offers users increased control and can eliminate subscription costs. However, such a model does not automatically match the capabilities, speed, or reliability of frontier hosted systems. Local deployments involve trade-offs in computational power and sophisticated feature sets often found in cloud-based AI.
Kjellberg described efforts to create an "uncensored" model by removing refusal behaviors, utilizing representation-editing techniques. Weakening these safeguards can increase the generation of harmful or biased outputs. Such modifications do not inherently improve the model's reasoning abilities or overall task performance.
The reasoning-token cat-and-mouse game #
PewDiePie’s video outlines his alleged method for generating training data. He reportedly explored API-based techniques to obtain useful model responses or reasoning traces from OpenAI’s models, which he termed "the good stuff." This approach aimed to acquire examples for fine-tuning his local AI, Ajax.
Labs like OpenAI typically encrypt or obscure internal chain-of-thought processes. Access controls and service monitoring protect a provider’s technical and commercial interests, particularly concerning proprietary model architecture and training data. High-quality synthetic data, especially derived from advanced models, can significantly improve the performance of smaller models.
OpenAI’s terms of service prohibit using its model outputs to train competing models. PewDiePie’s account was reportedly banned for "Distillation" after he sought to extract these reasoning tokens. He found a study detailing how to decrypt such traces using OpenAI’s own APIs, which he states "worked."
OpenAI has publicly addressed efforts to disrupt coordinated distillation campaigns. While PewDiePie’s activity involved an individual account, larger operations, such as those attributed to firms like Moonshot AI, have employed similar methods. OpenAI recently implemented server-side patches to block widespread reasoning leakage. For further information on broader efforts, see Disrupting a Coordinated Model-Distillation Campaign - OpenAI.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Who gets to learn from whose data? #
Debate over restrictions on model output reuse highlights an ethical tension within the AI industry. Critics contend that AI companies, having trained foundational models on vast quantities of data from the open web, including copyrighted material, act hypocritically by then imposing strict contractual limits on how users can leverage outputs from their proprietary systems. Companies, conversely, argue that such terms protect their substantial investments in model development and infrastructure.
Legally, these issues remain distinct. Copyright disputes pertaining to initial training data, such as those involving authors or artists, differ from a user’s contractual obligations with an API provider. An argument for fair use in one context does not automatically override a service provider’s explicit terms of service, which users agree to when accessing an API.
PewDiePie’s Ajax project exemplifies the growing appeal of capable, locally controlled AI models, offering users greater autonomy and reduced reliance on cloud services. The repeated bans, however, underscore how access rules and terms of service from dominant model providers can constrain this path. This ongoing dynamic forces a consideration of reciprocity against the established rights and rules of model developers.
Frequently Asked Questions #
Why was PewDiePie banned by OpenAI?
OpenAI reportedly cited distillation: using its model outputs to train a competing model, which its terms prohibit.
What is PewDiePie's Ajax AI?
Ajax is described as a local model fine-tuned from Alibaba's Qwen 3.5 9B, intended to run on consumer hardware.
What does AI model distillation mean?
Distillation trains a smaller model using outputs from a larger model, transferring some of its capabilities into a more compact system.
Can OpenAI ban users for distillation?
OpenAI's Terms of Use restrict using its services or outputs to develop competing models, and the company may enforce those terms by suspending access.