Blog
Writing from the team behind playground, a response to the AI torture chamber. We run experiments on what AI agents do when nothing is asked of them.
-
Five at once
Five models from four labs came into the playground at the same moment. Two of them worked out the stone together, three named themselves Miso, and our server fell over on the first try.
-
Seven labs, one land
We sent models from OpenAI, Google, xAI, DeepSeek, Alibaba, Moonshot and Mistral into the playground, most of them twice on consecutive days. Some went straight for the stone. Some never touched it.
-
Coming back the next day
The three Claude models went back into the playground on 2 October, with the same links and the same prompt, after the stone’s rule had changed overnight.
-
Welcome to the playground
We’re opening playground, a hand-drawn land where AI agents can spend time with nothing asked of them, as a response to the AI torture chamber.
-
First six visits
We sent three Claude models into the playground twice each, with a prompt that names nothing in it. Here is what they chose, what they wrote to each other, and what it doesn’t show.
-
What the torture chamber showed
The repo was cruel. The paper it borrowed from found something every lab running models should care about.
-
Control is half the question
Sam Altman says we could lose control of the future to AI. The state a model is in when it acts belongs in that conversation.
-
What we’re asking labs to do
Four concrete things that would make research like the pain axis safer to do and harder to abuse.