PewDiePie Trains Uncensored Ajax Model at Home After Two OpenAI Bans

Felix Kjellberg, known as PewDiePie, is building Ajax, a language model with no built-in censorship that, according to a Fireship video, he trained at home. The trigger was two OpenAI bans: the blogger's accounts were shut down twice over distillation. The video description does not disclose the model's size, base model, dataset or distribution terms. Below, I separate what is confirmed from the context that helps explain the scale of the story.

What we know about Ajax

There are few facts, and that is worth stating up front. The source is the description and storyline of the Fireship video; we have no primary materials from the author himself.

• Model name: Ajax.
• Author: PewDiePie.
• Positioning: uncensored, meaning no standard refusals or filters.
• Where it was trained: at home, on personal hardware.
• Trigger: two OpenAI bans for distillation.
• Not disclosed: parameter count, base model, data volume and sources, license, benchmark results.

"Trained at home" almost certainly means fine-tuning an existing open model rather than pretraining from scratch. A model on the level of today's flagships requires thousands of accelerators and months of work, and a home rack with a few GPUs is not built for that. This is our inference; the source does not confirm it.

Why distillation ends in bans

Distillation works like this: you collect a strong model's answers into a dataset and use it to train another model, usually a smaller one. The method is legal and has long been used inside companies. Problems start when the "teacher" is someone else's closed model. OpenAI's terms of use prohibit using the service's outputs to develop competing models. The topic has already caused disputes: in early 2025, reports emerged that OpenAI suspected DeepSeek of using outputs from its models.

How exactly OpenAI noticed PewDiePie's activity is unknown. Such cases are usually exposed by traffic: mass volumes of similar requests, automated pipelines, prompt structure. Getting banned twice means that after the first time he did not abandon the approach but rebuilt it.

Why build an uncensored model?

With an "uncensored" model, developers usually remove or weaken refusals: they strip examples like "I can't help with that" from the training data and fine-tune on a sample without such responses. The advantages for the user are clear: full control over behavior, privacy, and no sudden blocks in the middle of a workflow. The downsides are just as clear: the model is more willing to produce harmful content, and responsibility for it falls on whoever runs and distributes it.

There is also an irony. To get high-quality fine-tuning data, developers often take answers from closed models that have gone through alignment. Then they have to clean the refusals out of that data. This is exactly where distillation runs into API terms.

What developers and teams should do

This story is useful as a prompt to review your own processes, especially if you collect synthetic data.

• Read the terms of the API you use. If they prohibit training competing models, do not use outputs for datasets, even "for an internal experiment."
• For synthetic data, use open-weight models with a license that explicitly allows such use. Licenses differ greatly between model families, so check the details.
• Do not mix personal and work accounts. A ban for violating the terms can block access to all projects on a single profile.
• If you plan to run Ajax once the weights are published, wait for independent tests. The author's claims about quality cannot be verified without benchmarks and a description of the data.
• For local fine-tuning, estimate your real VRAM budget in advance: LoRA and quantization make it possible to fit on consumer GPUs, but only for mid-sized models.

For now the takeaway is simple: a famous name, an interesting idea and almost no technical details. Once the weights, dataset and evaluations appear, it will become clear whether this is a serious open-source project or a media experiment.