PewDiePie spent most of an October 1 video walking viewers through Ajax, the small local AI model he has built to run inside his self-hosted Odysseus workspace. But the line that travelled furthest was an aside: OpenAI, he said, had banned him twice while he was making it.
The detail spread quickly on X because it sounded like a celebrity squabble. It is better understood as a real-world stress test of rules that have so far been debated almost entirely in the context of Chinese AI labs. The method he described — feeding a top OpenAI model's output into a fine-tune of an Alibaba model — has the same basic shape as the practice US agencies accused six Chinese companies of running at industrial scale only a month earlier. That overlap is what makes the story more than a curiosity. The systems built to catch distillation see patterns of behaviour, not people, and the response a suspected distiller receives can depend less on who they are than on which enforcement policy the provider happens to follow.
Key Facts at a Glance
- The model: Ajax is a 9-billion-parameter model fine-tuned from Alibaba's Qwen 3.5 9B, built to handle everyday agent tasks inside the Odysseus workspace — browsing the web, running private searches, and managing email and calendars without sending data to a cloud provider.
- The training: It was first trained on successful Odysseus task traces, then refined over weeks of GRPO reinforcement learning on tool-use workflows. The full-precision build needs roughly 22GB of video memory, with an FP8 version at around 11GB and smaller quantised builds planned.
- The bans: PewDiePie said OpenAI banned him twice during development and linked the bans to attempts to distill reasoning tokens from OpenAI's models, which the company's consumer terms prohibit by barring programmatic extraction of output. OpenAI has not commented publicly on either case.
- The disputed claim: Posts on X alleged he used an exploit to decrypt hidden reasoning tokens. No source has substantiated that, and he has not said so himself.
- The launch: Ajax is scheduled to go live on October 3, 2026 at 08:25 Japan time — 7:25 pm Eastern on October 2. No licence and no independent benchmark had been published at the time of writing.
- The wider context: On September 8, the NSA, CISA and FBI published a joint advisory accusing DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of distilling US frontier models, and urged providers to watch for jailbreak prompts aimed at extracting chain-of-thought reasoning.
What PewDiePie Is Actually Releasing
Ajax extends a project that began with Odysseus, the free workspace PewDiePie released at the end of May as a self-hosted alternative to handing personal data to large technology companies. Where Odysseus provides the scaffolding, Ajax is meant to be the reasoning engine inside it — a model small enough to sit on consumer hardware but capable enough to run multi-step agent tasks locally.
The choice of base model is telling. Alibaba's Qwen family has become the default starting point for developers who want a capable open-weight model they can fine-tune without negotiating a licence, and Qwen 3.5 9B fits comfortably inside the memory budget of a high-end consumer GPU. Fine-tuning it on recorded task traces, then running reinforcement learning against tool-use workflows, is a well-worn recipe in the open-model community. What is less common is a creator with an audience in the tens of millions narrating the process step by step.
Why OpenAI Banned Him, Twice
In the video, PewDiePie said OpenAI banned him twice while building the model. Reports have identified the frontier system involved as Sol, the flagship OpenAI released in July as part of its GPT-5.6 family. The first ban was reversed on appeal before a second followed, according to accounts circulating after the video went live. OpenAI has not commented publicly on either ban.
The reasoning tokens matter because OpenAI does not show its models' full chain of thought to users, and that hidden reasoning is precisely what makes a large model valuable as a teacher for a smaller one. A model's visible answer teaches you what to say; its reasoning teaches you how to think through a problem, which is far more transferable and far more valuable in a training set.
One widely shared post claimed he had used an exploit to decrypt those tokens. Such a technique does exist: researchers disclosed in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed into a weaker model and read back in plain text, a flaw the providers have since patched on their servers. But no source has shown that PewDiePie used that method, and he has not claimed to.
A Pipeline the US Government Has Already Described
The September advisory from the NSA, CISA and FBI named six Chinese companies and described the signals providers should watch for, among them jailbreak prompts designed to extract chain-of-thought reasoning. Taking a leading American model's reasoning and using it to improve a fine-tune of a Chinese open-weight model is, in outline, exactly the pipeline that document describes. It is also roughly what PewDiePie says he was attempting from a home GPU rig in Japan.
The scale is obviously not comparable. Nothing suggests a single creator building a 9B model for his own workspace poses the kind of risk the advisory was written about, and a hobbyist's fine-tune is not an industrial distillation operation. The point is narrower and more uncomfortable: a provider's detection system does not see scale or intent at the moment it flags an account. It sees request patterns. And the pattern of someone pulling reasoning out of a frontier model to train another model looks much the same whether that someone runs a lab in Beijing or a YouTube channel in Japan.
That same characteristic means the signals can also catch ordinary developers whose usage happens to resemble a distiller's — heavy programmatic querying, unusual prompt structures, high-volume reasoning requests. False positives are not a hypothetical edge case. They are a structural feature of behavioural detection.
The Ban Was the Visible Option
The most consequential part of the advisory was not the list of names. It was the recommendation that providers respond to suspected distillation by quietly degrading or altering the answers they serve, rather than blocking accounts outright, and that they avoid telling suspected users this is happening.
PewDiePie received the opposite treatment. A ban is the one response a user can see, appeal and talk about in a video watched by millions. Had OpenAI applied the recommended approach, he might simply have received weaker outputs, trained Ajax on them, and never known why the results were worse than expected. A famous creator can turn a ban into a story. An ordinary developer whose answers are silently degraded has no ban to point to, no appeal to file, and no evidence that anything happened at all. The quieter policy is also the one that leaves no trace for anyone to contest.
Where the Hypocrisy Argument Breaks Down
Much of the reaction on X took his side on principle, with one popular post arguing that companies which trained their models on the public's data are in a weak position to object when a user takes something back. The argument resonates because the legality of training on scraped data is still being fought over in courts across several jurisdictions.
But it conflates two different disputes. PewDiePie agreed to OpenAI's terms when he opened an account, and those terms bar programmatic extraction of output. The people whose writing ended up in training sets never signed anything. Enforcing a contract a user accepted is a different matter from the unresolved question of what AI companies owed the authors of their training data — even if both arguments are ultimately about who gets to learn from whose work.
Removing Refusals Is the Part Worth Watching
The less discussed section of the video concerns what happened after training. Ajax has had its refusal behaviour stripped out using Heretic, an open-source tool that applies directional ablation to weaken the internal patterns associated with declining requests. PewDiePie compared the process to brain surgery and said he drew his own line at harm to other people or to the user, keeping some safeguards in place. Ablation, however, can also degrade output quality, and there is no independent evaluation yet of what was lost alongside the refusals.
Once Ajax's weights are public, his particular line becomes a default rather than a limit. Anyone who downloads an open-weight model can run the same class of tool again and remove whatever restrictions remain. That is the central tension of open releases: the creator's ethical boundary ships as a starting point that the next person is free to discard.
What to Check When Ajax Goes Live
The launch will settle several questions the video left open, and these are the details worth looking for on day one:
- Whether Ajax ships under a stated licence, since the Qwen base model uses Apache 2.0 but no terms for Ajax itself have been announced.
- Whether PewDiePie discloses how much of the final training data came from OpenAI outputs, given that he has described them only as part of a seed-data process alongside Odysseus traces and synthetic data.
- Whether independent testers can reproduce the roughly nine-in-ten task completion rate he reported internally.
- Whether OpenAI says anything about why the account was banned, which would show whether the bans targeted reasoning extraction specifically or automated use in general.
Frequently Asked Questions
Why did OpenAI ban PewDiePie?
PewDiePie said OpenAI banned him twice while he was building Ajax, and he linked the bans to attempts to distill reasoning tokens from OpenAI's models, which the company's terms prohibit. OpenAI has not commented on the case.
What is PewDiePie's Ajax model?
Ajax is a 9-billion-parameter local model fine-tuned from Alibaba's Qwen 3.5 9B to run agent tasks such as browsing, search, email and calendar management inside the self-hosted Odysseus workspace.
When is Ajax released?
Ajax is scheduled to go live on October 3, 2026 at 08:25 Japan time, which is 7:25 pm Eastern on October 2.
What is AI model distillation?
Distillation is the practice of training a smaller model on the outputs of a larger one so that it inherits some of that model's abilities. Frontier labs prohibit it when the outputs come from their systems without permission.
Did PewDiePie use an exploit to decrypt OpenAI's reasoning tokens?
There is no evidence that he did. The claim originates in posts on X rather than from PewDiePie himself, even though researchers disclosed in August that encrypted reasoning blocks could be extracted through a technique providers have since patched.
The story will be judged less on what it says about one creator than on what it reveals about enforcement: a rule written to deter industrial-scale copying, applied through detection that cannot distinguish a lab from a hobbyist, and a recommended remedy that leaves the person affected with no way of knowing it happened.
Source: Memeburn News