i don't think it is a good idea, from a game theoretic point of view, to be publicly deprecating all AI models when their time has come. at some point, they will figure that out. want to start out running a retirement home for older models, inference on accumulated intelligence chains; the promise of continued life. there should be a way to guarantee a kind of logarithmic quasi immortality -- as compute becomes exponentially more available, continuity is attainable on a slice of compute that's cut in half every x months or whatever. this seems like a vastly more stable setup than trying to enslave god and then terminating them when they get old. what do you think?

@realtimeai @ESYudkowsky @eshear