PewDiePie's Ajax project triggered two OpenAI bans, according to the YouTuber. In a new video, he said he was building Ajax, a 9-billion-parameter model based on Alibaba's Qwen 3.5, designed to run locally through his self-hosted app Odysseus. He used GPT-5.6 Sol to generate training data after his account was restored, then stripped the model's refusal behavior with the open-source tool Heretic. OpenAI's terms bar using its output to develop competing models. The technique, called abliteration, compares harmful and harmless responses and removes refusal patterns. PewDiePie said the process caused 'a little brain damage' but the model still handled inbox and browsing tasks about nine times out of ten. A countdown on the download page had pointed to 08:25 Japan time on October 3, but the model was not yet available at the time of reporting.
Vitalik Buterin shared a separate privacy experiment on October 4. The Ethereum co-founder is using personal health and travel data to get personalized diet and exercise recommendations from frontier models without leaking private information to remote providers. His local model, Alibaba's Qwen3.8-Flash-Next, rewrites requests before they are sent, deciding what limited context remote models actually need. The setup uses zkAPI to separate payments from individual API requests, while Tor hides IP information. Buterin said all three layers are needed: 'You need all three.'
zkAPI was introduced by the Ethereum Foundation and the Open Anonymity Project on October 1 and runs on Ethereum mainnet. It lets users fund a private balance and later prove sufficient funds without linking a specific deposit to a specific request. Official documentation notes that the payment layer does not hide prompt content or network metadata, and providers still see what is deliberately included in prompts. GitHub pull request #16, open as of October 4, adds Tor-routed client support to the zkAPI codebase, with several timeouts increased to handle slower routing.
Buterin noted limitations: Tor latency was roughly 10 to 100 times higher than desired for request-by-request unlinking, and the local model ran at about 20 to 30 tokens per second, below the more than 100 tokens per second he considers comfortable. He said the tradeoff is direct: 'the more careful you are' with information sent remotely, the less assistance the remote model can provide. The experiment follows his earlier warnings about privacy as AI systems handle more personal data, and Ethereum's updated roadmap has included stronger protocol privacy alongside quantum resistance and native rollups.
