PewDiePie's Uncensored Ajax Model: From DIY Training to OpenAI Bans

Episode 1 · PewDiePie Trains His Own Local LLM with GRPO, Community Rushes to Help (2026-10-01, 4 posts)

Former top YouTuber PewDiePie is training his own local LLM and agent (project odysseus) using agentic GRPO. Hugging Face researchers note he hasn't named his tools and is hard to reach, while the community and HF staff are offering tutorials and help.

• PewDiePie is doing his own GRPO training, and HF's merve wants to help but can't reach him — mervenoyann · 2026-10-01
• PewDiePie is quietly doing agentic GRPO training off the grid, say HF engineers — MaziyarPanahi · 2026-10-01
• PewDiePie tries agentic GRPO training; HF points to four free video series — mervenoyann · 2026-10-01
• Community petitions HF's Merve to help PewDiePie train his own local LLM — Zealousideal-Cut590 · 2026-10-02

Episode 2 · PewDiePie releases uncensored model Ajax, says OpenAI banned him twice over distillation (2026-10-01, 7 posts)

PewDiePie, the most-subscribed individual YouTuber, released Ajax, an uncensored open-source model based on Qwen 3.5 9B that runs on his home PC, and said OpenAI banned his account twice during development for "distillation" — drawing mainstream attention to the compliance debate around model distillation, with coverage from Tom's Hardware.

Confirmed

• Ajax is built on Qwen 3.5 9B, positioned as uncensored and locally runnable; it follows his earlier self-hosted open-source AI workspace odysseus.
• Per his own account (relayed by @Hesamation, @robleclerc, @burnytech), he used Sol to generate seed data for distillation training: after the first ban he appealed and was reinstated, then was banned again immediately upon rerunning Sol — twice in total, with "distillation" cited as a ban reason.
• He mocked the irony that tech giants build AI on scraped data while consumers are barred from doing the same — a point echoed by @sigmacrusader.

Unconfirmed

• Whether the bans were specifically for stealing hidden reasoning traces is an inference from timing overlap with OpenAI's report; OpenAI has not confirmed the reason.

Why it matters

• @aakashgupta notes OpenAI had just published a report on coordinated theft of its hidden reasoning traces; one day later PewDiePie's video reignited scrutiny of OpenAI's terms forbidding use of CoT outputs to train competing models.
• A top creator demonstrating local uncensored models and openly challenging platform terms pushed the once developer-only "is distillation legal" debate into mainstream discussion.

• PewDiePie launches Ajax, an uncensored open-source model fine-tuned from Qwen 3.5 9B — shadowrun456 · 2026-10-01
• PewDiePie gets banned twice by OpenAI for using Sol to distill training data — Hesamation · 2026-10-01
• PewDiePie Says OpenAI Banned Him for Distilling Chain-of-Thought into His Own Model — burny_tech · 2026-10-02
• PewDiePie Says OpenAI Banned Him Twice Over Reasoning-Trace Extraction — aakashgupta · 2026-10-02
• PewDiePie Unveils 'Uncensored' Home-PC AI Model Ajax, Banned Twice by OpenAI — robleclerc · 2026-10-03
• PewDiePie distilling GPT locally, claims OpenAI banned him twice — sigma_crusader · 2026-10-03
• PewDiePie unveils "uncensored" Ajax model for home PCs, says OpenAI banned him twice — ControlCAD · 2026-10-03

Episode 3 · PewDiePie trains uncensored local model Ajax after OpenAI bans (2026-10-06, 2 posts)

YouTuber PewDiePie trained an uncensored local model Ajax at home after being banned twice by OpenAI for distilling its outputs into a competing model. He has since switched to a fully open-source approach.

• PewDiePie builds uncensored home-trained model Ajax after OpenAI banned him twice for distillation — Fireship · 2026-10-06
• PewDiePie banned twice by OpenAI while fine-tuning local model, goes fully open-source — rodrigodevbits · 2026-10-06