3. Evan Hubinger on Takeoff speeds, Risks from learned optimization & Interpretability
The Inside View
00:00
How to Apply More Oftinization Power to a Task?
In training processes, values that are easy to specify will win out systematically. This breaks stratg ceiling, because some som people's values that are l easy to specify wil like systematiclly. And so in one way, we might want a sort of alined training process to do is not the sort of systematically better for some of these sorts of values and other values.
Transcript
Play full episode
