Regarding the convo btw Paul and Eliezer:
RL seems like the archetypal consequentialist ML algorithm. I’d like to know what Eliezer has in mind when he says “consequentialist”.
I agree that Paul’s approach to creating consequentialist agents while using learning algos with short horizons seems feasible.
However, I think it *is* a much less powerful approach than using a longer horizon learning process.
I view horizon-length as one of the key ways to trade-off performance and safety.