PewDiePie unveils Ajax, an uncensored AI for home PCs, and says OpenAI banned him twice

PewDiePie has shown Ajax, a 9 billion parameter fine-tune of Qwen 3.5 that runs on a home PC inside his Odysseus workspace. He says OpenAI banned him twice over distillation while he built it. Here is what we know and what is still missing.

PewDiePie, the YouTuber whose real name is Felix Kjellberg, has shown off Ajax, a small AI model he trained himself and describes as uncensored. It is built to run on a normal home computer as an always on assistant inside Odysseus, the self hosted AI workspace he released earlier this year. The launch video, posted on October 1, comes with a twist that quickly took over the conversation: he says OpenAI banned his account twice while he was building it, and one of the emails he shows on screen names "distillation" as the reason.

The story is part celebrity tech project, part fight over who gets to learn from whose model. Here is what we know, what is still missing, and why a 9 billion parameter model from a YouTuber is worth your attention.

The Short card: OpenAI banned him twice, over distillation, OpenAI has not commented publicly

The three things you need to know

The model. Ajax is a fine tuned version of Alibaba's open Qwen 3.5 9B. It is made for Odysseus and is meant to handle daily tasks such as web search, browsing, email and calendar on your own machine.
The bans. PewDiePie says OpenAI suspended his account twice while he used one of its models to help create training data. One email he shows cites distillation. OpenAI has not commented publicly.
Uncensored. He removed most of the model's refusals with the open source tool Heretic, but says he drew the line at harm to other people or to yourself. Ajax is not downloadable yet.

What Ajax actually is

Ajax is not a new model trained from scratch. It starts from Qwen 3.5 9B, an open weight model from Alibaba with roughly 9 billion parameters. Parameters are the adjustable numbers inside a neural network, and 9 billion is small by today's standards. That is the point: a model of that size can run on a decent consumer PC instead of a data center.

According to the Ajax page quoted by Tom's Hardware, the model powers Odysseus to handle "daily tasks from search to browse the web to email to your calendar," all "completely privately." Tom's Hardware compares the setup to agent projects such as OpenClaw and Nous Research's Hermes Agent. The idea is an assistant that is always running, has access to your tools, and never sends your inbox to someone else's server.

Fine tuning means taking an existing model and training it further for one specific job. Here the job is being a good agent inside one specific harness. PewDiePie argues in the video that small models trained "for a specific harness" are the smarter path for everyday work, instead of calling a giant frontier model for every small task.

How he built it, and why OpenAI stepped in

To make a small model better, developers often use a bigger model as a teacher. The big model produces answers, examples or reasoning, and the small model learns from them. This is called distillation. Decrypt describes it as training a smaller AI on a bigger one's answers, "like a student copying the class genius's homework."

PewDiePie says he wanted to "distill just a little bit" from one of OpenAI's models. Reports differ on the exact version: Decrypt names GPT-5.6 Sol, while The Frontier notes that the video itself names no version and the transcripts only say "Sol." What he was after, according to Decrypt, were reasoning tokens, the hidden scratchpad text a model writes while it works through a problem. OpenAI keeps that scratchpad encrypted.

Here is the sequence as he tells it:

His account was deactivated the first time. He called the ban unjustified, disputed it, and the account was restored.
He then used the model again "to create my seed data," the starter examples a model learns from.
OpenAI deactivated the account a second time. "How did they even know?" he asks in the video.

The email he shows says the account was deactivated for activity that violated OpenAI's terms and usage policies and cites distillation. It gives no specific examples, Tom's Hardware notes. OpenAI has not explained the two account actions in public, so everything beyond that email comes from his own account of events.

The rules themselves are clear. OpenAI's terms of use bar people from using its output to develop models that compete with OpenAI, and from automatically or programmatically extracting output from its services. Interesting Engineering points out that OpenAI itself describes distillation as a legitimate technique. The conflict is about using it on OpenAI's own models to build something outside OpenAI.

PewDiePie, for his part, is not apologetic. "They worked hard stealing all of our data to make their AI machines," he says in the video, arguing that learning from a model's output is not so different from how the big labs trained on the open web in the first place.

Illustration: a giant machine pours answers into a small machine through a funnel, cut off by a red barrier and padlock

Why the big labs care so much about distillation right now

The timing makes this more than a celebrity spat. On September 30, OpenAI said it had disrupted a coordinated campaign to extract the hidden reasoning of its models, and it attributed a core cluster of that activity to people associated with Moonshot AI, the maker of Kimi. The company also said it closed a pathway that let someone replay another user's encrypted reasoning to recover its contents.

Decrypt adds another data point: researchers showed in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed to a weaker sister model that wrote them out in plain text, and they decoded 315,320 of them from public code repositories.

So when a single user with a large audience says in public that he tried to pull reasoning out of a frontier model to train his own, it lands in the middle of a very active security fight. There is no sign that PewDiePie is connected to any of those campaigns. But it shows how the line between a hobby project and a terms of service violation can be a single script.

Taking the "no" out of the model

The second headline word is "uncensored." To get there, PewDiePie used Heretic, an open source tool that removes a model's built in refusals. Decrypt explains how it works: it compares how the model reacts to harmful and harmless prompts, finds the internal pattern behind its refusals, and cuts that pattern out. The technique is known as abliteration.

The surgery is not clean. PewDiePie says Ajax took "a little brain damage" in the process, but that it still handles tasks like sorting an inbox or browsing about nine times out of 10. He says he chose which refusals to remove and drew the limit at "causing harm to other people or causing harm to yourself." On his lawyer's advice, he adds that Ajax is "not designed to provide dangerous actionable instructions." Tom's Hardware notes that it is unclear where the real limits will be once people can test it.

The Short card: refusals removed with Heretic, about 9 of 10 tasks, not downloadable yet

Small model, big pitch

The most interesting part of the video may be the argument underneath it. PewDiePie's case for small local models is about cost and energy. By his own napkin math, reported by Decrypt, running one copy of a rumored trillion parameter frontier model would take about 27 of his computers and the electricity of roughly 150 houses. "I think it makes no sense to poke that giant beast" for basic tasks, he says.

That is a rough estimate from a creator, not a measurement, and frontier labs serve many users per machine. Still, the direction of the argument matches what a lot of developers are doing: use a small model for routine agent work, keep the giant models for the hard problems.

He is also still training. Decrypt reports that he is using GRPO, a reinforcement learning method in which the model tries the same task 16 times and learns from the attempts that work. Tom's Hardware says he planned two more days of reinforcement learning before rerunning the decensoring step, then quantizing the model, which compresses it to run on less memory, and benchmarking it.

Illustration: a single home computer next to an enormous wall of server racks fed by many houses

What is still missing

As of now, Ajax is a promise more than a product:

No weights yet. The Ajax page first showed a countdown to 08:25 Japan time on October 3. Decrypt reports that the timer is gone, and The Frontier found the page saying the model is "coming soon."
No model card, license or benchmarks. The Frontier found none of these as of October 3. Without them, nobody outside his team can check how capable or how safe Ajax really is.
No word from OpenAI. The bans are documented only through his video and the email he showed.
No community training data. He asked fans to donate training data rather than collect it from Odysseus users, but said nobody had submitted any.

Why it matters

Local AI is going mainstream. When one of the biggest creators on YouTube pitches a private assistant that runs at home, millions of people hear the idea of local, self hosted AI for the first time. That is good for the open model ecosystem around Qwen, Llama style models and tools like Heretic.
Distillation is now a front line. Labs spend billions to train frontier models, and their outputs are the cheapest way to train a competitor. Expect more detection, more bans and more legal language around reasoning tokens, both for state linked groups and for individual hobbyists.
Uncensored models raise real questions. Removing refusals makes an assistant less annoying for normal tasks, but it also removes guardrails that exist for a reason. A well known creator shipping one will push that debate out of niche forums.
Access is conditional on hosted models. As Popular AI notes, a hosted model can be switched off for you by account action. A model on your own machine cannot. That difference will matter more as people hand agents their email and calendars.

Dany's take

I like the core idea here a lot. A small model that runs on my own PC, reads my mail and manages my calendar without sending anything to a cloud is exactly the kind of AI I want for everyday work. If PewDiePie gets millions of people to try local AI, that is a win for everyone who cares about privacy.

But I would not cheer for the way it was built. Pulling training data out of OpenAI after the first ban was a choice, not an accident, and the terms are not hidden. You can think the big labs trained on everyone's data without asking, and I partly agree, and still see that "they did it too" is not a license.

The uncensored part is where I want to see evidence before excitement. "A little brain damage" and "nine times out of 10" are honest admissions, and I respect that. Until there are weights, a license and real benchmarks, Ajax is a very good video, not yet a very good model. When it ships, I will test it and tell you how it does.