Eliezer: supposing we label a model as definitely-a-person, do you want to just toss it out of the hypothesis space as if it never existed, or do you want to try to reason abstractly about what that model would do without actually running the model?
No; you destroy everyone that currently exists, and replace them with a happy version of that model. More happiness with less resources.
Eliezer: supposing we label a model as definitely-a-person, do you want to just toss it out of the hypothesis space as if it never existed, or do you want to try to reason abstractly about what that model would do without actually running the model?
No; you destroy everyone that currently exists, and replace them with a happy version of that model. More happiness with less resources.