MIRI's Rob Bensinger clarifies that Eliezer Yudkowsky's recursive self-improvement scenario can span months or years rather than occurring instantly

@robbensinger @croissanthology @willdepue The latter is not the intelligence explosion / foom frame.

@robbensinger @croissanthology @willdepue Oh I misread, I thought you were denying the extremely short timescales, but you were taking issue with the without limit.

Yes he often acknowledges eventually some kind of eventual physical limit, for example, but I think diminishing returns on investment is more pertinent.

@xlr8harder @croissanthology @willdepue "ML systems are friendly by default" at least seems like a coherent scenario to me; "AI can never advance scientific frontiers fast enough to outstrip humans' ability to monitor and understand everything" seems a lot harder to imagine.