PewDiePie's Homebrew AI Banned Twice by OpenAI, He Says—Now It Triages Your Inbox 🧠

YouTuber PewDiePie said OpenAI banned his account twice while he was building Ajax, a roughly 9-billion-parameter AI model designed to run on a user's own computer and handle tasks such as sorting email and browsing the web. He detailed the project in a new video on his YouTube channel. Ajax is described by PewDiePie as "a fine-tuned AI model" and is built on Alibaba's open-source Qwen 3.5, with a public release pending on his website.

Ajax is intended to live inside Odysseus, a free, self-hosted AI app PewDiePie launched in June. Because the model runs locally, user data such as emails and calendars stays on the user's machine, and there is no subscription fee, though the user's own hardware handles the compute load. PewDiePie said the goal was to distill from GPT-5.1, the flagship model OpenAI released on July 9 — a process that trains a smaller model on a larger one's outputs to capture its reasoning, often called "reasoning tokens." OpenAI keeps that scratchpad encrypted, and PewDiePie said he would never have cracked OpenAI but was banned anyway. He called the ban unjustified, disputed it, and regained access. He said he then used GPT-5.1 again to create seed data and was banned a second time. OpenAI's terms of use bar using its output to develop competing models.

Researchers demonstrated in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed to a weaker sister model that produced the contents in plain text, decoding 315,320 blocks from public code repositories. On Sept. 30, OpenAI said it had disrupted a campaign by people associated with Moonshot AI, the developer of Kimi, to extract its models' hidden reasoning, and closed a pathway that let a user replay another user's encrypted reasoning to recover its contents.

To make Ajax "uncensored" — meaning it answers requests that ChatGPT-style assistants typically refuse — PewDiePie used Heretic, an open-source tool that removes a model's built-in refusals through a process called abliteration, which compares how the model reacts to harmful and harmless prompts, identifies the refusal pattern, and removes it. PewDiePie said the procedure was not clean and that Ajax took "a little brain damage," but that it handles inbox sorting and browsing tasks correctly about nine times out of 10. He added that he is still training it.