PewDiePie Says OpenAI Banned Him Twice While He Built Ajax, a Self‑Hosted Uncensored AI

Preface

Context: This article summarizes recent claims by YouTuber PewDiePie that OpenAI suspended his account twice while he worked on Ajax — a small, self‑hosted AI designed to run on a personal computer. The purpose here is to present the sequence of events, the technical methods he says he used, and the implications of such projects. Objective: Provide a balanced, factual overview that explains what Ajax is claimed to be, how it was built, and why the story has drawn attention from privacy and AI‑safety observers.

Lazy bag

The essentials: PewDiePie claims OpenAI suspended his account twice while he distilled outputs from a flagship model and used open‑source tools to remove built‑in refusals from a fine‑tuned model called Ajax. Ajax is described as a 9‑billion‑parameter model meant to run locally, offering a subscription‑free, self‑hosted alternative that can browse the web and manage tasks. OpenAI has not publicly responded to these specific allegations.

Main Body

The narrative began when PewDiePie, the widely followed content creator, described in a recent video that his OpenAI account was suspended twice during his work on Ajax, an AI he intends to distribute for running on home machines. According to his account, Ajax is a fine‑tuned variant based on Alibaba’s open‑source Qwen 3.5 and is reported to have roughly 9 billion parameters. Fine‑tuning in this context means taking a preexisting model and giving it additional, task‑focused training so it becomes better at a particular set of behaviors — here, acting as a personal assistant inside Odysseus, the self‑hosted application he previously released.

One central technical detail PewDiePie described is a process called distillation. Distillation typically involves using a large, capable model to generate examples that a smaller model then learns from. He said he aimed to distill outputs from a model he refers to as GPT‑5.6 Sol, seeking Sol’s reasoning tokens — the intermediate text a model produces on a private scratchpad while it works through a task. If accessible, such reasoning traces can help a smaller model approximate a larger model’s chain‑of‑thought reasoning. OpenAI protects such internal state, which PewDiePie says motivated his attempts to reproduce or extract useful outputs.

PewDiePie states that his account was first suspended after activities related to studying or attempting to extract reasoning traces through OpenAI’s API. He described finding a published study and public code that demonstrated how encrypted reasoning blocks from several providers might be replayed and decoded by a weaker model. After disputing the suspension, he says his account was restored, but he later used the API again to generate seed data for training — a step he says led to a second ban. OpenAI’s published terms prohibit using its outputs to train competing models, and that policy context explains why such behavior could trigger enforcement.

Another controversial element in PewDiePie’s account is the removal of a model’s built‑in refusals. He describes using an open‑source tool called Heretic to perform what is often called “abliteration” or removal of safety filters. Heretic and similar tools compare model responses to harmful versus benign prompts, learn patterns associated with refusals, and modify the model so it no longer follows those refusal patterns. According to his description, this produces an ‘uncensored’ assistant that will comply with user requests that conventional models decline.

He acknowledged risks: he said he chose to keep refusals that prevent causing direct harm to others or to oneself, and his legal counsel advised him to state that Ajax is not intended to produce dangerous, actionable instructions. PewDiePie also told viewers that the ablation process reduced some capabilities — he described Ajax as having taken “a little brain damage” — but that it remains effective for many intended tasks like email sorting and web browsing about nine times out of ten.

On the training side, he reported using a method he called GRPO, where the model attempts the same task multiple times (he cited 16) and learns from the successful attempts. He also emphasized the practical case for smaller models: cost and energy. He argued that running a rumored trillion‑parameter frontier model would be prohibitively resource‑intensive for individuals, while a compact model like Ajax can run locally without subscription fees, keeping personal data on the user’s hardware rather than on remote servers.

Technically, Ajax is not publicly released at the time of his video. PewDiePie said he still needed to repeat the ablation process, quantize the model to reduce resource requirements, and benchmark it to ensure it continued to perform acceptably. A download page previously displayed a countdown for an October 3 release time; as of the latest reports that timer was removed. Observers note that distributing a locally runnable model with removed refusal behavior raises ethical and safety questions, particularly if it can produce harmful content or actionable illegal instructions.

From a policy perspective, the incident highlights tensions between platform rules, intellectual property and safety provisions, and the desire for local, privacy‑preserving AI. OpenAI’s enforcement actions (as described by PewDiePie) reflect terms intended to prevent the extraction of proprietary capabilities and to limit the development of direct competitors using its outputs. At the same time, proponents of small, open models emphasize user control, data privacy, and reduced ongoing costs.

Independent researchers previously published work showing that encrypted reasoning artifacts from models could be replayed and decoded under some circumstances, and at least one vendor—OpenAI—said it disrupted a campaign to exploit such a pathway. These technical vulnerabilities and the responses to them form a backdrop to the allegations PewDiePie raised. The broader community reaction will likely involve technical auditing of any released model, debate over legal and ethical responsibilities of creators and platforms, and continued attention from developers building self‑hosted AI tools.

In summary, the story as presented by PewDiePie ties together claims about platform enforcement, technical workarounds for model limitations, and arguments for the practicality of smaller, local AI. The claims remain primarily those of the creator; OpenAI has not issued a public statement directly addressing the specific allegations described. If Ajax becomes available, its architecture, training data provenance, and safety mitigations will be central to evaluating both its utility and the potential harms it could enable.

Key Insights Table