PewDiePie builds local AI model Ajax after two reported OpenAI bans
OpenAI reportedly banned the world's biggest YouTuber for distillation. His response: building a local, uncensored AI for consumer hardware.
Models · Source: Interesting Engineering
What happened
Felix Kjellberg, known globally as PewDiePie, is building his own artificial intelligence model. He calls the project Ajax. The initiative started after OpenAI reportedly suspended his account twice. Kjellberg says an email from OpenAI cited distillation as the specific reason for the bans. Distillation is a machine learning technique that involves using outputs from a large teacher model to train a smaller model. OpenAI strictly forbids using its generated data to create competing models. Kjellberg reportedly got his account restored after an appeal, only to be suspended again when he generated seed data for his new project.
Ajax takes a completely different approach from massive cloud systems. It is a 9-billion parameter model based on Alibaba's open-weight Qwen3.5-9B. Kjellberg designed it to run locally on consumer hardware inside a self-hosted software environment called Odysseus. The goal is to create an always-on agent capable of handling web searches, emails, calendars, and daily browsing. By running locally, Ajax processes everything on the user's own machine rather than sending private daily tasks to a corporate cloud provider.
Kjellberg also deliberately stripped away standard safety filters. He used an open-source tool called Heretic to perform automatic abliteration. This process reduces the model's refusal behavior without extensively altering its core capabilities or causing what he called brain damage. He claims to have retained boundaries against self-harm and dangerous actionable instructions on legal advice. The model is not publicly available yet. The project website lists it as coming soon while Kjellberg plans additional reinforcement learning, a second decensoring pass, quantization, and benchmarking.
Key facts
- 9 billion — Parameters in the underlying Alibaba Qwen3.5-9B model
- 2 — Number of times OpenAI reportedly suspended Kjellberg's account
- Heretic — Open-source tool used to reduce the model's refusal behavior
Why it matters
The barrier to entry for capable, agentic artificial intelligence is dropping to zero. You no longer need massive data centers or billions in venture capital to build a useful daily assistant. By fine-tuning a small 9-billion parameter open-weight model, a single independent creator is attempting to replace cloud-based services for complex daily tasks. This proves that local, private artificial intelligence is rapidly becoming a viable alternative to strict API dependency. Builders can now orchestrate powerful local environments that users control entirely.
The tension between giant artificial intelligence labs and independent builders is also escalating rapidly. OpenAI enforcing its terms of service against a massive internet creator highlights the strict control they want over their data exhaust. As frontier models try to lock down their outputs to prevent competitors from catching up, the open-source community will accelerate tools like Heretic to bypass corporate restrictions. The future of artificial intelligence is fracturing into two distinct paths. One path is highly controlled cloud APIs managed by tech giants. The other is a wild west of local, uncensored models running on consumer hardware.
For builders
Local agents bypass cloud privacy risks
Running models locally on consumer hardware keeps sensitive data like emails and calendars off corporate servers. Builders can target privacy-conscious users who refuse to hand their daily workflows to giant labs. However, local execution alone does not close every privacy gap. You must manage permissions strictly, as misconfigured tool servers or network connections can still expose data.
Distillation carries severe platform risks
Using outputs from frontier models to train your own systems violates the terms of service for companies like OpenAI. Founders who rely on this method risk sudden account bans and total loss of access to critical infrastructure. You must secure independent training data or leverage permissive open-weight models to avoid platform de-platforming.
Abliteration enables custom refusal behaviors
Tools like Heretic allow developers to remove standard corporate guardrails from open models without retraining them from scratch. This lets you build specialized tools that do not refuse valid prompts or moralize at your users. The risk falls entirely on the builder to implement custom safety boundaries for dangerous or illegal instructions.
My take
OpenAI trying to gatekeep artificial intelligence by banning people for distillation is a losing battle. When you lock the doors and restrict the data, builders simply go home and build their own house. Local, uncensored models running on consumer hardware will eventually eat the bottom of the cloud API market because users want control, not corporate lectures.
Original reporting: Interesting Engineering. This is my rewrite and opinion.