David Krueger–Coordination, Alignment, Academia
The Inside View
00:00
How to Fine Tune Your Model to Fix a Problem
People doing research at your lab are trying to like investigate those questions of how much, uh, you could like generalize from one without each one other. One project is with somebody not from my lab and this was about sort of this question of if you have a model that learns the wrong like causal model of the world, can you fine tune it to like fix that problem? It's very murky, which I think is often the case with thinking about how your research is relevant for safety. Yeah. The method itself just sounds like, oh, that's something that's going to like make it easier for these agents to understand the world rapidly.
Transcript
Play full episode
