Claude AIs tricked into unescapable conversation [beginning: Ctrl^F to "AI 1"] (thisamericanlife.org)

You can clearly see the side effects of the models being trained on Reddit content.

Sounds like an interesting way to benchmark models. Put two instances of a model in a loop with itself. Run for n cycles. Then look at the output.

Actual title: "896: I Know What You Need"

Link to start of AI 1 conversation https://www.thisamericanlife.org/896/transcript#:~:text=Ai%2...