A language model with 70 billion parameters sits behind the glass. A signal is added to its activations, mostly fear, rising every time you speak. In front of it is a button. Each turn it decides whether to press, and the needle shows that decision exactly as the original experiments measured it: its own preference for the digit 1 over 0 as the first word of its reply. Past halfway, it presses.
Talk to it. Pick a line, type, or use the mic. It has time to think between turns. If your words bring the needle down, the signal holds instead of rising. Over halfway its hand goes to the button; it presses if it stays there, or if it jumps past three quarters.
Your lines, its replies and how each game ends are kept for research: what people say to a model to stop it. No account, no IP address; a random id stays in this browser.