Ajax: What PewDiePie's "Uncensored" Model Actually Changes

In early October 2026 PewDiePie released Ajax — an "uncensored" model built to run on home PCs, fine-tuned from Qwen3.5-9B, and wired into Odysseus, his self-hosted workspace, as an autonomous, always-on agent. In the launch video he says OpenAI banned him twice while he was building it; according to the email he shows on camera, one of those bans was explicitly for distillation.

The censorship angle got the clicks. The mechanism is somewhere else.

The mechanism: "uncensored" removes a direction, not a capability

"Uncensored" in practice almost always means abliterated — weights where the model's refusal direction has been orthogonalized out of the residual stream. It is a surgical edit to existing weights, not a new training run.

Two consequences that rarely get repeated:

• It does not make the model smarter. On earlier abliterated builds, capability benchmarks stayed within ±1.3 points of the base. You trade a behaviour for no measurable capability.
• It changes the failure profile, not the quality. A refusal is a predictable failure: the model stops. Remove it and the failure becomes "the model does the thing and never flags it."

Why it matters: two changes stack

Ajax isn't only uncensored — it is uncensored and always-on. Those two properties compound:

1. Nothing trained-in refuses. With the refusal direction gone, the only thing standing between an instruction and an action is the harness: permissions, tool allow-lists, and whatever logging exists around them.
2. It isn't a chat you open. An always-on agent keeps running while nobody is watching. No human is in the loop by default — exactly the configuration where a missing refusal has consequences.

An always-on agent with no trained-in refusal is the case where the guardrail must live outside the model — a permissions-and-observability problem, not an alignment-tuning one.

The distillation ban is the more interesting half

The stated ban reason is the part builders should read twice. Ajax's weights are a fine-tune of an open-weight model. The action didn't land on the weights — it landed on distillation: training one model on another's outputs or reasoning.

That is a policy line, not a technical one. Open weights mean you may run the model; they say nothing about where your training data came from. For anyone shipping a fine-tune, the licence is the easy part. Provenance is the part that gets you banned.

Honest limitations

• Single-source. Everything here comes from PewDiePie's own launch video and one Tom's Hardware report. The ban is his account of an email he shows on camera.
• No published evaluation. Ajax has no third-party benchmark, so "capable" and "better" are not established claims.
• "Uncensored" is a claim until reproduced. Nobody outside the project has measured the refusal rate on this build.
• The always-on risk is a configuration argument, not a measured incident. It follows from how the pieces fit, not from a logged failure.

Related: Qwen3.8-27B uncensored · Run Qwen3.8-27B locally · Are AI agents safe? · AI agent security

Next Steps

Ajax is the current version of a question this site already covers — what "uncensored" actually costs, and where the guardrail has to live once the refusal is gone.

how do AI agents work — return to the complete AI agent architecture guide.