Image from Anthropic’s research “A global workspace in language models”
I find it strange about myself: my first reaction to this research method was a philosophical one. “Injecting ideas into a mind. What if the subject were a person?” “If the model itself can feel pain, would this count as abuse?”
It also reminded me of Inception, where Cobb plants the idea “this world isn’t real” deep in the subconscious of his wife, Mal. Back in reality the idea keeps growing; she never believes she has woken up, and in the end she takes her own life in order to “wake up”.