OpenAI Banned PewDiePie Twice While He Built Ajax, an Uncensored AI That Will Run on Your PC

What it claims

Ajax is version one of a 9-billion-parameter fine-tune of Alibaba's open-source Qwen 3.5, built to run inside Odysseus, the free, self-hosted AI app PewDiePie launched in June. It is meant to handle inbox sorting, calendar work and web browsing on local hardware. PewDiePie says it does those tasks successfully "about nine times out of 10." It is not downloadable. A countdown on the download page pointed to 08:25 Japan time on October 3, but that timer has since been removed. The page gives no new date, only that it will appear on his website when he considers it ready.

How it works

• Seed data: Generated with GPT-5.6 Sol, the flagship model OpenAI released on July 9. He was after Sol's encrypted reasoning tokens and cited a study that recovers them using OpenAI's own API. Both of his account bans happened during this work.
• Refusal removal ("abliteration") with Heretic: The tool compares the model's responses to harmful and harmless prompts, isolates the pattern behind its refusals, and removes it. He says he chose which refusals to strip and kept those covering "causing harm to other people or causing harm to yourself."
• GRPO: The model attempts the same task 16 times and learns from the attempts that succeed. Training is ongoing.

Evidence

The only performance figure is his own informal estimate of roughly 9 in 10 successes on inbox and browsing tasks. The page gives no benchmark, baseline or test set.

The related research mentioned in the piece is more concrete. In August, researchers replayed encrypted reasoning blocks from OpenAI, Anthropic and Google to a weaker sister model, which wrote them out in plain text. They decoded 315,320 blocks taken from public code repositories. On Sept. 30, OpenAI said it had closed the pathway that let one user replay another user's encrypted reasoning.

Limits

• Degraded capability: He says abliteration gave Ajax "a little brain damage."
• Unfinished: He still plans to rerun the ablation, quantize the model and benchmark it "to check that nothing broke." Treat the current model as unvalidated.
• Provenance and terms risk: The training data came from GPT-5.6 Sol output. OpenAI's terms bar using that output to build competing models, and that is the reason given for the second ban. Anyone building on Ajax should weigh this.
• Safety disclaimer: On his lawyer's advice, he states that Ajax is not designed to give dangerous, actionable instructions. The page does not say how that is enforced in a model whose refusals were removed.
• Unknowns: The page states no licence, no hardware requirements beyond "your own hardware does the heavy lifting," and no quantized size.