Eric Michaud on scaling, grokking and quantum interpretability
The Inside View
00:00
The Risks of Advanced AI Systems
Part of the risk, maybe, from advanced AI systems is that it arises from the fact that we maybe don't understand the internals of the models very well. I'm just generally hopeful that if we had a much more rigorous understanding of neural networks, that we would be in a much better position to avoid some sort of worst-case alignment failures. It's a little bit risky, because you also might discover something that accelerates things. But on net, it feels really potentially quite good to have a better understanding of the systems.
Transcript
Play full episode
