4 ms·
We understand the training algorithm, but not the network that was trained. There’s the whole interpretability sub-discipline which has limited but interesting
by frabcus 3y ago
We understand the training algorithm, but not the network that was trained.
There’s the whole interpretability sub-discipline which has limited but interesting results. Those just indicate even more that we really have no idea how something like GPT-4 works inside its trained model.