AGI Can Be Safe
Data Skeptic
00:00
The Importance of Observing the Behavior of an Agent
When a human is considering whether or not they should hit that safety stop button obviously they can look at the behavior of the agent. Should they also be observing something under the hood maybe getting a dashboard of metrics coming out of the safety layer? Yeah I think ideally they would. If you imagine a weapon system on the battlefield which is also has a non opaque model inside which is a bit like the chess board of the go board you would see in an alpha zero system. You can actually look inside that thing and see exactly that it's going to make you in five moves. This is all nice what we're seeing now is current generation technology does not allow you to look deep inside if it's
Transcript
Play full episode
