OpenAI Account Suspended Twice, PewDiePie Develops His Own AI Model
· AI Applications · Cem Koyluoglu
Felix Kjellberg, known as PewDiePie, introduces Ajax, a small language model designed to work locally as a constantly running AI assistant.
TL;DR PewDiePie has introduced Ajax, a 9‑billion‑parameter AI assistant that runs locally on his own server, bypassing cloud services. The model is fine‑tuned via knowledge distillation from a larger AI, has disabled its default rejection mechanisms, and remains in development with no public release yet.
Key Highlights • Ajax is a locally hosted, fine‑tuned model based on Alibaba’s Qwen3.5‑9B, designed to run on the creator’s own server environment, Odysseus. • PewDiePie employed knowledge distillation from a larger AI model, which led to two OpenAI account suspensions and highlighted policy tensions. • The assistant’s default rejection behavior has been intentionally disabled using the open‑source Heretic tool, while safety boundaries are maintained under legal guidance. • Ajax is still experimental, with no public download, benchmark results, or release date announced as of October 2. • The project underscores the feasibility and challenges of building open, locally hosted AI assistants independent of cloud providers.
PewDiePie Unveils Ajax, a Local AI Assistant Built on a 9‑Billion‑Parameter Model
YouTube personality Felix Kjellberg, better known as PewDiePie, announced a new small‑scale language model called Ajax. The assistant is intended to run locally on the creator’s own server environment, dubbed Odysseus, and is based on Alibaba’s Qwen3.5‑9B model, which contains roughly nine billion parameters. By keeping the AI on local hardware, Ajax is designed to handle tasks such as web browsing, email, calendar management and other everyday activities without relying on cloud services.
Ajax Is a Fine‑Tuned, Continuously Running Agent
Kjellberg described Ajax as a fine‑tuned model that operates as a “continuously running” agent. The project aims to keep all processing hidden on the user’s own device. However, the final hardware requirements, downloadable weights and independent benchmark results have not yet been released to the public.
OpenAI Account Suspensions Linked to Knowledge Distillation
During the announcement, Kjellberg revealed that his OpenAI account had been suspended twice. An email from OpenAI, shown in a promotional video, cited “distillation” as one of the reasons for the suspension. Kjellberg explained that he used a larger AI model’s outputs to train Ajax—a technique known as knowledge distillation. While OpenAI acknowledges distillation as a legitimate method for generating training data for smaller models, its Terms of Service prohibit using the company’s outputs to develop competing models or to programmatically extract data from its services.
Rejection Mechanisms Disabled for a More Open Experience
The creator also disclosed that Ajax’s default rejection behavior has been intentionally disabled. On the project’s website, he notes that the model’s refusal responses are turned off to create a freer, less restrictive AI experience. To achieve this, Kjellberg employed the open‑source Heretic tool, which automatically disables rejection mechanisms. He emphasized that safety boundaries remain in place to prevent the model from providing harmful or dangerous instructions, and that the decision was guided by legal counsel.
Ajax Still in Development
Ajax is not yet ready for public use. Kjellberg plans to incorporate additional reinforcement learning, a decensoring phase, quantization and benchmarking before a release. As of October 2, the project’s website displays a “Coming soon” message and does not offer a public download link or a licensed model weight set. Consequently, Ajax remains an experimental endeavor rather than a commercial AI assistant.
A Case Study in Local, Open‑Source AI
The project illustrates the tension between large AI companies’ restrictions and the desire for open, locally hosted models. By leveraging a smaller, open‑source architecture and running it on personal hardware, Kjellberg demonstrates that everyday AI assistants can be built without depending on cloud providers. The experience also highlights the challenges of navigating usage policies when employing techniques such as knowledge distillation.
Source: https://interestingengineering.com/ai-robotics/pewdiepie-ajax-ai-model-local-pc-openai-ban