Felix "PewDiePie" Kjellberg has released the first version of Ajax, an "uncensored" AI model fine-tuned from Alibaba's Qwen3.5-9B and designed to run entirely on home PCs. The launch came with an unusual disclosure: OpenAI banned his account twice during development, and according to the OpenAI email he shows on camera, one ban was explicitly for "distillation" — using one model's outputs or reasoning to train another.

Ajax powers Odysseus, the self-hosted AI workspace PewDiePie launched in May. The setup resembles OpenClaw and Nous Research's Hermes Agent: an autonomous, always-on assistant that handles "daily tasks from search to browse the web to email to your calendar," all "completely privately," according to the Ajax product page. Because the model runs locally, no data leaves the user's machine — the pitch that anchors the entire project.

The page says Ajax's refusal behavior has been removed for a "freer, less restricted AI experience" and asks users to run it responsibly. The name is a joke, he says — named after the cleaning product, though in Greek myth Ajax was Odysseus' rival.

Two bans, one email

PewDiePie's account was reinstated after the first ban. The second came after he ran the model again "to create my seed data." His reaction: "How did they even know?" The email, as he presents it, offers no specific examples of what he did wrong.

He is candid about the intent. "I wanted to distill just a little bit" to improve the model, he said. He then argued the practice is not fundamentally different from how frontier labs build their own products. "They worked hard stealing all of our data to make their AI machines," he added.

The context matters commercially. OpenAI and Anthropic go to considerable lengths to hide their models' reasoning chains, using encryption and other countermeasures against attempted distillation. OpenAI recently attributed one such campaign to actors linked to Moonshot AI and framed it as a threat. A solo creator doing openly what state-linked actors allegedly do covertly puts the enforcement question in sharp relief: platform bans are currently the main deterrent.

How the 'uncensoring' works

To strip refusals from the base model, PewDiePie used Heretic, an open-source tool that performs automatic "abliteration" — discovering prompts the model refuses, tracing the refusal behavior, and removing it. The risk, as he describes it, is doing "brain damage" by removing too much. He drew the line at prompts involving harm to other people or oneself. On his lawyer's advice, he said Ajax is "not designed to provide dangerous actionable instructions." He cites nuclear weapons as a hard limit, though the real boundaries of the fine-tune remain unclear and unverified.

Small models over trillion-parameter beasts

PewDiePie "strongly believe[s]" in small models trained for a specific harness rather than general-purpose giants. "I think it makes no sense to poke that giant beast" for basic tasks, he said, referring to rumored one-trillion-parameter frontier models. His stated goals are research and a usable model, which for him means small.

He also asked fans to donate training data, calling it superior to harvesting data from Odysseus users directly. Nobody submitted anything, he acknowledged.

What's next

The roadmap is concrete: two more days of reinforcement learning, a rerun of the decensoring step, then quantization and benchmarking before the next release. "I can't wait to release the next version of Ajax and Odysseus," he said, noting Odysseus "has gotten a huge update as well."

The project illustrates a growing pattern in consumer AI: open-weight base models from major labs, community fine-tuning tools like Heretic, and creator distribution channels combining to put less restricted, locally running assistants in the hands of mainstream users. Whether frontier labs' distillation defenses — and their account bans — can keep pace with that pipeline is now an open question.