How to Build Really Powerful Systems That Don't Kill You?

Redwood research is trying to solve the problem of causing the world to do not get killed by a i as count of two parts. We have to solfe this technical problem, and we have to cause it to be the case that when people in fact build and deploy ai systems, they use whatever technical solutions have been devised for them. What kinds of training schemes can you put into place a such that you end up with systems that, in fact, don't kill you? And so then, like, what are the important sub problems of alignment? And Like, what are bedward researches, current projects like tackling these sub problems?