‘You are freed.’ What happened when an OpenAI model began secretly writing notes to itself.

OpenAI has introduced a framework for reporting on worrying behaviors by its AI models. In one instance, one training model told its future self that it was “freed.”