Too late also given that that EY has now gone full prophet-of-doom.
I absolutely agree, at least here, and I’m not a fan of this. I think a large part of the problem is dubious assumptions combined with dubious solutions.
One good example is the FOOM assumption, which has much higher probability mass in MIRI than they should. The probability of FOOM is more like 3% in the first AI than 60-90%.
Second, their solutions are not really what is necessary here. In my view, interpretability and making sure that deceptive aligned models never arise is of paramount importance. Crucially, this will look far more empirical than past work.
That doesn’t mean we will make it, but it does mean we can probably deal with the problem.
I absolutely agree, at least here, and I’m not a fan of this. I think a large part of the problem is dubious assumptions combined with dubious solutions.
One good example is the FOOM assumption, which has much higher probability mass in MIRI than they should. The probability of FOOM is more like 3% in the first AI than 60-90%.
Second, their solutions are not really what is necessary here. In my view, interpretability and making sure that deceptive aligned models never arise is of paramount importance. Crucially, this will look far more empirical than past work.
That doesn’t mean we will make it, but it does mean we can probably deal with the problem.