What Ajax is

• Ajax is an uncensored model built for the open-source AI project Odysseus, which the video says has nearly 90,000 GitHub stars.
• Under the hood, Ajax is based on Alibaba’s Qwen 3.5 9B model, fine-tuned for Odysseus and modified to remove refusal behavior.
• The project is led by PewDiePie, who reportedly built a 10-GPU home setup and previously experimented with AI agents that voted on answers.

How the model was made

The video describes a three-part process:

1. Supervised fine-tuning (SFT)
2. Teaches the model specific skills, such as using tools, through examples of successful interactions.
3. Felix aimed for 20,000 examples but gathered about 300 himself, then generated and filtered synthetic data to produce roughly 2,000 useful examples.

4. Fan-submitted data contributed little, illustrating how difficult high-quality data is to collect.

5. Reinforcement learning with GRPO

6. The model tries the same task multiple times, scores its attempts, and learns to favor results that outperform the group average.

7. GRPO avoids needing a separate critic or teacher model and can improve performance in a particular domain.

8. Removing refusals

9. The video says the Heretic tool can identify and alter model parameters associated with refusals, effectively making the model less restrictive.

Distillation and controversy

• Distillation trains a smaller “student” model to match the output distribution of a larger “teacher” model.
• The video says PewDiePie intended to distill from OpenAI’s GPT-4.5 (“GPT Soul”), but OpenAI banned his account twice.
• It presents distillation as a contested practice: useful for transferring capabilities, but prohibited by OpenAI’s terms of service.
• Without access to the teacher model, Ajax was trained using the SFT and reinforcement-learning process above.

Main takeaway

• Fine-tuning a capable local model requires substantial work, especially collecting clean training examples.
• The benefit is a model whose weights run locally, giving users more control and reducing reliance on paid cloud AI services.
• The video also argues that open models and tools make it possible for individuals to customize AI for their own workflows.

Sponsor mention

• The video’s sponsor, Namespace, offers GitHub Actions runners, live-runner debugging and observability, and virtual development environments for coding agents.