Invest3 publishers2 min readPublished
PewDiePie says OpenAI banned him twice as he trained his Ajax model on GPT-5.6 Sol outputs
PewDiePie says OpenAI banned him twice while he used GPT-5.6 Sol outputs to train Ajax, a 9-billion-parameter model for home PCs. OpenAI has not commented publicly, but his account shows that a team training its own model on a frontier lab's answers can lose API access partway through the build.
The Investor · Invest desk
What happened
- He says the first suspension came while he went after Sol's encrypted reasoning tokens using a study's method, and that the account was restored after he disputed it.
- Once he was reinstated, he used Sol to generate seed data for Ajax, and OpenAI banned him a second time.
- As of early October, Ajax had no public model card, published benchmarks, confirmed weights license or download link, Crypto Briefing reported.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost A distillation pipeline that runs through one OpenAI account stops when that account is suspended. His build lost access twice, the second time after he had won the first dispute.
- precedent Crypto Briefing's reading is that, if his account is accurate, OpenAI enforces the competing-models clause against individual developers as well as rival companies.
- exposure Weights trained partly on Sol outputs could face questions over what their eventual license can legitimately permit. That risk passes to anyone who builds on Ajax.
- contradiction Crypto Briefing says Ajax arrived around Oct. 1-2. Decrypt says it is not yet downloadable and its countdown has disappeared, so no one outside can test the model yet.
The first suspension is the easier one for OpenAI to defend [3]. In August, researchers showed that a weaker sister model, fed encrypted reasoning blocks from OpenAI, Anthropic and Google, would spell out their contents in plain text [7]. Working from public code repositories, they decoded 315,320 such blocks [7]. OpenAI said on Sept. 30 that it had disrupted a campaign by people associated with Moonshot AI, the developer of Kimi, to extract its models' hidden reasoning [8]. It said it had closed a pathway that let someone replay another user's encrypted reasoning to recover its contents [8].
A fine-tuning team should pay more attention to the second suspension [4]. Seed data is ordinary output, and OpenAI's terms bar using output of any kind to develop competing models [5]. Under that clause, a team that buys Sol tokens to build a training set for a model that competes with OpenAI's is paying for something it may not use for that job [5].
Three readings fit the record. The second ban could be a flag left over from the first. If so, it says little about plain seed-data generation. It could have been triggered by the volume or pattern of his requests, whatever they were for. Or it is the competing-models clause applied to one developer [5]. I put the most weight on the last, since the use he describes fits the clause's wording [4][5]. That view is wrong if OpenAI says the second suspension came from the first. For now, every reading rests on his own video [1].
The reason to risk the account is cost. He has run the rough numbers himself: a single copy of a frontier model rumored to have a trillion parameters would take about 27 of his computers and about as much electricity as 150 houses [12]. A trillion parameters is about 111 times Ajax's 9 billion [1]. A model that small runs at home with no subscription fee, and the user's own hardware does the heavy lifting [18]. He said Ajax took "a little brain damage" when he stripped out its refusals with Heretic, an open-source tool [9]. He also said it handles tasks like sorting inboxes and browsing about nine times out of 10 [10].
GRPO, the training he is still running, has the model attempt the same task 16 times and learn from the attempts that work [11]. Those attempts are Ajax's own output. This stage of the build does not draw on the account OpenAI suspended twice [11][1].
What to watch
- The weights license Ajax ships under, and whether it limits use given that its seed data came from Sol.
- Benchmarks after he reruns the ablation and quantizes the model, which would show whether the nine-in-ten task rate survives compression.
- Reports from other developers of suspensions after generating fine-tuning data from Sol, which would show enforcement extends beyond one account.