This is a special post for quick takes by Altaer. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
For anyone doing empirical alignment research, could you share some key tooling/systems-level challenges you've run into? I'm considering ways I could contribute to this part of the stack and I think any real world experiences people could share would help me come up with a more grounded approach. An example of the kind of response I'm looking for: "Pytorch doesn't support the type of weight-inspection I need to do in a performant enough way/with my model-parallelism setup."