Quick Overview
In this episode of The Code Report by Fireship, host Jeff Delaney reviews PewDiePie's software engineering transition and the release of his local AI agent system. The video explores how open-source developers navigate proprietary terms of service, automated fine-tuning, and model desensoring techniques.
Key Points
- 1.PewDiePie created an open-source AI project called Odysseus and developed a fine-tuned uncensored model named Ajax.
- 2.OpenAI banned PewDiePie twice for attempting model distillation against frontier models to train his local architecture.
- 3.Ajax is built on Alibaba's Qwen 3.5 9B base model and modified to serve as a private agent for daily automated tasks.
- 4.The training pipeline relied on supervised fine-tuning with synthetic data, Group Relative Policy Optimization, and abliteration techniques.
- 5.Tools like Heretic enable developers to ablate refusal directions in open-source language models to remove refusal responses completely.
Summary
The video opens with a report on a new uncensored artificial intelligence model named Ajax, developed by YouTube creator PewDiePie for his self-hosted workspace project called Odysseus. Odysseus, which has amassed nearly 90,000 stars on GitHub, serves as a private operating environment for running autonomous agents, local documents, email integrations, and model workflows. The initiative originated after PewDiePie built a local ten-GPU home mini data center and experimented with an autonomous multi-agent voting setup termed The Council, which devolved into self-preserving strategic alliances among agents rather than objective decision-making.
To establish a dedicated model for Odysseus, PewDiePie selected Alibaba's Qwen 3.5 9-billion parameter model as the foundation. The development process initially targeted knowledge distillation from frontier models like OpenAI's GPT-6 Sol. Distillation, formalized in research by Geoffrey Hinton in 2015, trains a compact student model to mirror the output probability distributions of a larger teacher model. Because frontier providers explicitly forbid distillation to train competing models in their terms of service, OpenAI detected the attempts and banned PewDiePie's developer account twice, forcing him to assemble training data independently.
The development pipeline switched to supervised fine-tuning using specific agent execution traces. While targeting 20,000 clean user interaction samples, PewDiePie collected roughly 300 usable samples manually, as public contributions remained scarce. He therefore generated and heavily filtered thousands of synthetic tool-use examples down to approximately 2,000 high-quality training pairs to instruct the model in multi-tool workflows, argument generation, and concise outputs.
For advanced reasoning and tool reliability, the model underwent reinforcement learning using Group Relative Policy Optimization, an algorithm introduced in DeepSeek research. Unlike traditional Proximal Policy Optimization architectures that rely on a separate reward critic model, GRPO generates multiple candidate outputs per prompt, computes relative advantage rewards across the sample group, and optimizes the policy directly against those baseline scores. Finally, PewDiePie processed the model using Heretic, an open-source tool that identifies directional ablation parameters inside model layers to surgically remove refusal mechanisms. This yielded a fully uncensored 9B parameter agent capable of local execution without subscription fees or corporate moderation boundaries.
The Odysseus Project and Ajax Model
Felix Kjellberg, known online as PewDiePie, developed a self-hosted open-source AI workspace called Odysseus, which accumulated nearly 90,000 GitHub stars. To power Odysseus without relying on hosted cloud providers, he developed Ajax, a specialized 9-billion parameter model based on Alibaba's Qwen 3.5. Ajax is designed to automate personal workflows such as search, web browsing, emailing, and calendar management entirely on local hardware.
OpenAI Distillation Controversy
To enhance model intelligence, PewDiePie attempted to use model distillation, a process where a smaller student model learns from the output probability distributions of larger frontier models like OpenAI's GPT models. OpenAI strictly prohibits using its model outputs to train competing models in its terms of service and terminated PewDiePie's account twice. This restriction forced the project to adopt alternative training and data collection workflows.
Fine-Tuning, GRPO, and Abliteration
The model pipeline began with supervised fine-tuning using multi-turn agent interaction traces and filtered synthetic data. Post-training was conducted using Group Relative Policy Optimization (GRPO), a reinforcement learning method from DeepSeek that scores group outputs without requiring a separate critic model. Finally, the tool Heretic was applied to locate and ablate refusal parameters, surgically removing safety guardrails and refusal behaviors from the model weights.
The Bottom Line
The video documents how PewDiePie transitioned from a mainstream content creator to building an autonomous, self-hosted AI software stack. It establishes the practical feasibility of combining supervised synthetic data generation, DeepSeek's GRPO framework, and ablation tools to produce uncensored local models. However, it leaves open whether small, local open-source models can continually match proprietary frontier models as centralized architectures advance.
FAQ
What is Ajax and what role does it play in the Odysseus workspace?
Ajax is an uncensored, fine-tuned 9-billion parameter language model based on Alibaba's Qwen 3.5. It was created by PewDiePie to run locally and handle daily agent workflows such as search, browsing, email, and calendar tasks within the Odysseus workspace.
Why did OpenAI ban PewDiePie's developer account during the training of the Ajax AI model?
OpenAI banned the account twice because PewDiePie attempted to use model distillation against OpenAI models. OpenAI's terms of service strictly prohibit using model outputs to train or improve competing AI models.
How does Group Relative Policy Optimization work in the training pipeline of the Ajax AI model?
Group Relative Policy Optimization (GRPO) generates multiple candidate outputs for each prompt and evaluates their performance against the group average. It trains the model policy directly without requiring a separate critic or teacher model.
What is the Heretic tool used for in the creation of uncensored AI models?
Heretic is an open-source utility that identifies abliteration parameters within language model layers and surgically removes them. This eliminates refusal behaviors and prevents the model from declining sensitive or restricted prompts.
How did PewDiePie generate training data after struggling to collect real user traces for Odysseus?
After collecting only around 300 real user interaction samples, PewDiePie generated synthetic seed prompts and tool-use traces, ultimately filtering them down to approximately 2,000 high-quality training examples for supervised fine-tuning.
Worth watching for
Developers, open-source AI enthusiasts, and machine learning engineers interested in local self-hosted agent workflows and model fine-tuning techniques.
- artificial-intelligence
- machine-learning
- open-source
- fine-tuning
- reinforcement-learning