I should clarify that by “I hope it happens” I mean “I hope we get to experience this conditional on a fixed outcome”, not “this would be good in expectation”.
“Your own version of the internet tailored to you” really undersells it. You program the boundary conditions of a massive multi-agent simulation. You can go down any not-yet-existent rabbit hole (latent multiverse branch). This encompasses all fields of research; it encompasses code. Now replace “you” with “anything that writes”. Web4 would accelerate timelines if not directly foom. A self-programming simulation hallucinated by superhuman AI (GPT-30) is even more of a ticking time bomb than base reality.
… perfectly adjusted to maximize everyone’s enjoyment of time spent on the internet, learned from the users’ individual feedback.
If this is done using RLHF I particularly expect it to end in wireheading / pathological mode collapse / misaligned takeoff.
I should clarify that by “I hope it happens” I mean “I hope we get to experience this conditional on a fixed outcome”, not “this would be good in expectation”.
“Your own version of the internet tailored to you” really undersells it. You program the boundary conditions of a massive multi-agent simulation. You can go down any not-yet-existent rabbit hole (latent multiverse branch). This encompasses all fields of research; it encompasses code. Now replace “you” with “anything that writes”. Web4 would accelerate timelines if not directly foom. A self-programming simulation hallucinated by superhuman AI (GPT-30) is even more of a ticking time bomb than base reality.
If this is done using RLHF I particularly expect it to end in wireheading / pathological mode collapse / misaligned takeoff.