Comment by narrator

If you're going to make something smarter than a person, you got to be convinced that you're only going to be able to understand it on the single training step level and then inductively trust that the rest of it works. We do empirical testing of course with evals, but there's sort of an art to figuring out what is theoretically going to improve eval performance. Trying to fit the meaning of all those weights in your little human brain and working back from there isn't going to work for more than a little slice of the dataset at a time because that's all we can fit in our understanding.