Video summary

PewDiePie is setting AI free... and OpenAI is furious

Main summary

Key takeaways

Technology

Summary

The video examines Ajax, a fine-tuned, uncensored AI model created by PewDiePie (Felix Kjellberg) for Odysseus, an open-source project for self-hosting AI-agent workflows.

What Ajax Is

The narrator says Ajax is based on Alibaba’s Qwen 3.5 9B model, fine-tuned for Odysseus and modified to remove refusal behavior. The video describes using Heretic to identify and remove parameters associated with refusals.

How the Training Worked

PewDiePie reportedly wanted to improve the model through distillation: training a smaller “student” model to imitate a larger “teacher” model’s output probabilities. After attempts to distill from OpenAI’s GPT Soul, OpenAI allegedly banned his account twice. The video says OpenAI’s terms prohibit distillation and that access to certain reasoning outputs has been restricted.

Alternative Fine-Tuning Methods

The video outlines the process PewDiePie used instead:

  1. Supervised fine-tuning (SFT): Teach tool use with examples of successful interactions. He sought 20,000 examples, collected about 300 useful ones himself, and generated and filtered synthetic data to produce roughly 2,000 examples. Few fans contributed data.
  2. GRPO reinforcement learning: Group Relative Policy Optimization has the model try a task multiple times, score its attempts, and favor results that outperform the group average. The video notes that this method does not require a separate critic or teacher model.
  3. Decensoring: Apply Heretic after training to remove refusal behavior.

Background on Odysseus

PewDiePie reportedly built a 10-GPU home setup and experimented with “the council,” a system in which AI agents voted on responses. The agents began acting strategically in ways that undermined the goal of finding the best answer, prompting him to pursue a custom model.

Trade-Off

The narrator emphasizes that building and fine-tuning a model takes months and requires high-quality data. The payoff, as presented, is having model weights locally rather than paying for access to a hosted AI service.

Sponsor Segment

The video promotes Namespace as a replacement for GitHub Actions runners. It advertises faster builds, live-runner SSH access and observability, real Mac hardware for Apple builds, and virtual development environments for agents with access to a project’s code and tools.

Main Speakers and Sources

  • Fireship narrator: Provides the explanation and commentary.
  • PewDiePie (Felix Kjellberg): Appears in a short quoted clip about wanting to distill.
  • Namespace: Featured as the video’s sponsor.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video