I think using a narrow task A(G)I to help us solve alignment or flip the board is still promising. There are enough details about the real physical world and human biology we could hide that it couldn’t model the world well enough to take over. We’d only provide enough information to get useful answers. Furthermore, we can use only solutions we can verify or test to our satisfaction.
I think using a narrow task A(G)I to help us solve alignment or flip the board is still promising. There are enough details about the real physical world and human biology we could hide that it couldn’t model the world well enough to take over. We’d only provide enough information to get useful answers. Furthermore, we can use only solutions we can verify or test to our satisfaction.