Interpretability is hard.

even when we get it, we have to know what to do with it.

we would have to manipulate the Internal goals of AI.

• Internal goals are based on internal simulations and internal representations.

• We have to manipulate those in order to write hard rules into AI’s.