Reflexive Language Models: Impossibility, Games, and Closure

​I'll present some ideas that have motivated (in particular) Futurarium-affiliated people in the last years in their research about Large Language Models. Can they understand themselves? In some sense, the answer is no, but this impossibility is perhaps quite fertile for creative purposes. I'll discuss a collection of games that can lead us further. Warning: this is work in progress, and as such perhaps an even more experimental talk than usual.

​You can play this game to get an idea: https://xentlabs.ai/

​Clement Hongler sometimes hangs out in the Futurarium and likes to think about originality both abstractly and concretely.