I disagree with (3) because the Lobian obstacle is just an obstacle to a certain kind of stable self-modification in a particular toy model, and can’t say anything about what kinds of safety guarantees you can have for superintelligences in general.
I disagree with (4) because MIRI hasn’t shown that there are ways to make a superintelligence 90% or more likely (in a subjective Bayesian sense) to be stably friendly, and I don’t expect us to have shown that in another 20 years, and plausibly not ever.
Okay, fair enough. To explain briefly:
I disagree with (3) because the Lobian obstacle is just an obstacle to a certain kind of stable self-modification in a particular toy model, and can’t say anything about what kinds of safety guarantees you can have for superintelligences in general.
I disagree with (4) because MIRI hasn’t shown that there are ways to make a superintelligence 90% or more likely (in a subjective Bayesian sense) to be stably friendly, and I don’t expect us to have shown that in another 20 years, and plausibly not ever.
Thanks! I guess I was unduly optimistic. Comes with being a hopeful but ultimately clueless bystander.