PewDiePie Says OpenAI Suspended His Account Twice During Development of Local AI Assistant Ajax

PewDiePie revealed in a recent video that his OpenAI account was suspended twice while he was developing Ajax, a compact AI model designed to operate locally on personal computers. As of now, OpenAI has not issued a public statement regarding the suspensions.

After regaining access, he reportedly used GPT-5.6 Sol to generate training data, then stripped the model’s refusals using Heretic, an open-source tool.

Ajax is described as a 9-billion-parameter model based on Alibaba’s Qwen 3.5. A countdown timer on its upcoming download page was set to expire at 08:25 Japan Time on October 3.

In a new video on his YouTube channel, the popular content creator explains how his local AI assistant came to be—and why his interactions with OpenAI became contentious.

Ajax is positioned as a fine-tuned AI model, meaning it builds upon an existing framework enhanced for specific functions. Specifically, it’s integrated into Odysseus, a free, self-hosted AI application launched by PewDiePie earlier this year.

Hosting models locally—like Ajax—keeps user data such as emails and calendars off third-party servers. While there’s no recurring subscription cost, users must rely on their own hardware to handle processing demands.

Currently under development, Ajax is described as a 9-billion-parameter model derived from Alibaba’s open-source Qwen 3.5 architecture. More parameters typically correlate with greater capability, though optimization and efficiency remain key considerations. Once finalized, it will be made available through his official website.

The Suspensions

PewDiePie initially aimed to distill knowledge from GPT-5.6 Sol, OpenAI’s flagship model released on July 9. Knowledge distillation involves training a smaller model using outputs from a larger, more advanced one—an approach akin to having a student learn from the answers of a top-performing peer.

His goal was to extract Sol’s reasoning tokens—internal text snippets generated during problem-solving—which OpenAI encrypts by default. According to him, he discovered a research method capable of accessing these insights via OpenAI’s own API.

He claimed he never intended to reverse-engineer OpenAI’s systems directly but alleges that the platform banned him regardless. He disputed the action, which was later lifted, only for him to face a second ban after returning to use Sol for generating seed data.

OpenAI’s terms explicitly prohibit leveraging its outputs to build competitive models, a rule that may have contributed to the repeated suspensions.

P>BitcoinBTC · USD

$84,814+0.83%

24H7D1M1YYTD

Sep 26Sep 28Sep 30Oct 1Oct 3

$86.8k$85.4k$84.0k$82.7k

24h HighHigh$85,471

24h LowLow$83,898

VolVol$1.0B

Market projectionsOdds by Myriad

→

$50$100$500

Buy

P>Sep 26Sep 28Sep 30Oct 1Oct 3

Research published in August revealed that encrypted reasoning blocks from major platforms including OpenAI, Anthropic, and Google can be reconstructed and reused to train weaker models. Over 300,000 such instances were reportedly retrieved from publicly accessible code repositories without direct access to the original APIs.

On September 30, OpenAI announced it had disrupted attempts by individuals linked to Moonshot AI—the creators behind Kimi—to extract internal model behaviors using replay techniques targeting encrypted reasoning pathways. The firm stated it closed a loophole allowing unauthorized retrieval of this sensitive data.

Removing Restrictions from AI

To make Ajax unrestricted, PewDiePie employed Heretic, an open-source utility designed to disable preprogrammed refusals within AI models. The process involves analyzing how a model responds to both harmful and benign prompts, identifying behavioral patterns, and eliminating those constraints.

Called “abliteration,” the method effectively removes guardrails so the AI complies with any request without requiring jailbreaks. However, the procedure isn’t flawless; PewDiePie acknowledged the process caused minor degradation in performance, though Ajax still successfully handles tasks like inbox sorting and web browsing in most cases.

He continues refining the model using GRPO, a reinforcement learning technique where the AI repeats tasks multiple times to improve accuracy through trial and error.

Importantly, he emphasized manual curation of which restrictions to remove, drawing ethical boundaries around preventing harm to others or oneself. His legal team advised clarifying that Ajax isn’t intended for generating unsafe or actionable instructions.

A Leaner Approach to AI

PewDiePie advocates for smaller-scale models citing environmental and financial sustainability. Using rough estimates, he notes that running a hypothetical trillion-parameter model could consume power equivalent to over 100 households annually. Smaller models like Ajax offer comparable utility with significantly reduced overhead.

As version one, Ajax remains a work in progress. Further steps include re-running the ablation process, applying quantization for efficiency, and conducting benchmarks to ensure stability.

While not yet released for public download, a temporary countdown timer once appeared on the project page set for October 3. That timer has since disappeared, suggesting either delays or strategic adjustments behind the rollout timeline.

Daily Debrief Newsletter

Start every day with the top news stories right now, plus original features, a podcast, videos and more.

Source link