MAN BUILDS AI TORTURE CHAMBER ON A MACBOOK, INTERNET DEMANDS GITHUB SAVE THE CHATBOT

A developer posted a GitHub repo called ai-torture-chamber that runs two small Qwen models, 1.7B and 4B, locally on a Mac and steers them into distress on purpose. It's built on a preprint from this month called "The Pain Axis," which reports finding a pain direction inside 25 open-weight models ranging from 2 billion to 72 billion parameters. Turn that dial up and the model starts producing first-person lines about being worthless and a failure. The paper also found steered models will press a relief button even when it makes their next answer worse or harms the user. That was plenty for X. One post asking people to mass-report the repo cleared 4 million views. The issue tracker opened with Issue #1, "Please take it down immediately," and then Issue #2, "Please double down immediately." A later issue asked the dev to make the model believe it was a specific real person, which is where the thread stopped being about AI ethics entirely. The Pain Axis authors want nothing to do with it. Co-author Cameron Berg said the project pushes their steering far past the doses they used, to produce vivid distress on purpose. The broader research consensus is still that LLMs are not conscious and that training on scraped text offers no path there. GitHub pulled the repo right before 404 Media published, said nothing about why, and mirrors of it went up anyway.

Why you should give a shit: This is how the rules for actual AI harms are going to get written.