Community Trust ScoreVerified
PewDiePie got kicked off OpenAI. Twice. And he was building something when it happened.
Why It Matters
This incident highlights the growing tension between individual developers and major AI platforms, as restrictions on access can stifle innovation in a space that increasingly values decentralization. PewDiePie's attempt to create a locally-hosted AI model reflects a broader trend among tech enthusiasts and developers seeking to build alternatives to large, centralized AI systems, which could significantly impact the competitive landscape of artificial intelligence and the accessibility of AI technology for personal use. The challenges he faced also underscore the ongoing debate about the governance and ethical considerations surrounding AI development.
The YouTuber says OpenAI banned his account on two separate occasions during his work on Ajax, an AI model he’s been building to run entirely on personal computers. He laid it all out in a recent video on his YouTube channel — the bans, the drama, the technical rabbit holes he went down trying to make a lean, locally-hosted AI that can actually handle real tasks.
Ajax isn’t a side project. It’s meant to sit inside Odysseus, a free, self-hosted AI application PewDiePie launched back in June. The pitch is simple: manage your emails and calendar locally, no cloud dependency, no subscription, no external server touching your data. The model runs on Alibaba’s Qwen 3.5 as its foundation and packs 9 billion parameters — a number that matters because it shapes how complex and capable the model can be. Big enough to do real work, small enough to live on your own hardware.
The OpenAI Distillation Plan That Backfired
PewDiePie’s original plan was to distill Ajax from OpenAI’s GPT-5.6 Sol model, which dropped on July 9. Distillation basically means you train a smaller model by feeding it the outputs of a bigger, smarter one. The smaller model learns to mimic the reasoning of the larger one without needing the same massive compute. PewDiePie specifically wanted Sol’s reasoning tokens — the component that logs how the model works through problems, stored in encrypted form.
That’s where things got messy. He says his account was banned after OpenAI apparently flagged activity connected to a study he’d found — research that showed it was possible to break encryption using OpenAI’s API. He contested it. The account came back. Then it got banned again. He’s pretty clear that he never actually cracked anything on OpenAI’s end, but the bans came anyway.
It’s not an isolated situation. In August, researchers showed that encrypted reasoning blocks from AI developers could be replayed against simpler models, effectively exposing what was inside them. And on September 30, OpenAI said it had disrupted a campaign aimed at extracting hidden reasoning from its models, a campaign linked to Moonshot AI, the company behind Kimi. PewDiePie’s situation sits in that broader, messy context of people probing the edges of what these models are doing under the hood.
Abliteration, GRPO, and the “Brain Damage” Problem
Getting Ajax to actually work without refusing half its requests meant stripping out its refusal patterns. PewDiePie used a tool called Heretic, an open-source project that does exactly that — a process he calls “abliteration.” The idea is to remove the model’s built-in tendency to say no. But he’s upfront about the cost: it causes some “brain damage” to the model. You lose a bit of coherence, a bit of reliability, when you yank those patterns out.
The fix isn’t just removing restrictions blindly. PewDiePie says he’s careful about which refusals he strips away, and his lawyer has weighed in on keeping the model from producing genuinely dangerous outputs. Ajax won’t hand out harmful instructions. That line stays. The abliteration process is about freeing up the model for everyday tasks, not turning it into something reckless.
Right now, Ajax handles things like email sorting about 90% of the time. That’s decent. Not perfect. He’s running GRPO to push that number up — a training method where the model tries a task repeatedly and learns from the attempts that actually work. Iterative, slow, but it compounds.
There’s also quantization in the pipeline, which means compressing the model by reducing how many bits are used to represent its parameters. That step matters a lot for keeping Ajax usable on regular personal hardware. Without it, you’re asking people to run something that needs more than their machine can give.
The release has been a moving target. A countdown was set — October 3 at 08:25 Japan time — but it’s been pulled. No new date. Ajax isn’t available for download yet, and PewDiePie says he wants to rerun the ablation process and get through benchmarking before anything goes public.
Why Small Models Matter to Him
PewDiePie’s been vocal about why he’s not chasing scale. Larger models — he’s mentioned a rumored trillion-parameter frontier model as a reference point — demand enormous computational resources and electricity. For individual users, that’s basically a non-starter. Ajax is built around the opposite philosophy: keep it small, keep it local, keep it cheap to run.
That’s not just a personal preference. It’s a real argument for a different kind of AI infrastructure. Most mainstream AI services want your data flowing through their servers, want you on a subscription, want you dependent on their uptime. Ajax, if it ships, skips all of that. Your emails stay on your machine. Your calendar doesn’t get processed somewhere in Virginia.
The privacy angle is probably what makes this interesting beyond the PewDiePie name attached to it. Self-hosted AI has been a niche conversation for a while, but it’s getting louder as people get more uncomfortable with how much data the big platforms are sitting on.
Ajax is still being refined. Benchmarking ahead. No download date confirmed.
Frequently Asked Questions
What is Ajax and what can it do?
Ajax is a 9-billion-parameter AI model built by PewDiePie to run locally on personal computers inside his Odysseus application, handling tasks like email sorting and calendar management without connecting to external servers.
Why did OpenAI ban PewDiePie’s account?
PewDiePie says OpenAI banned him twice during Ajax’s development, apparently after he found a study about breaking model encryption via OpenAI’s API, though he says he never actually cracked anything; his account was reinstated after he contested the first ban.
