
3. Evan Hubinger on Takeoff speeds, Risks from learned optimization & Interpretability
The Inside View
00:00
Using Ai Thay, I'm a Liman, and I'm Not Sure if It's Alined, or Not.
Aliman is like, useful for almost anything you want to use your a iy for. But this relies on there being some separability between the aliment parts and the non aliment parts - which seems really unlikely. And especially again, when we're in a situation where these sort of trainn ones have a truly astronomical cost. We're not going to be able to replicate randics because they are so expensive. It doesn't really seem much different than just, like, you know. Wa wat ye're mentioning.
Transcript
Play full episode