3 ms·
Yes, I do think that is to get more training data... But my impression is that continual training is difficult, so it makes more sense to get a big chunk of it
by drdeca 4y ago
Yes, I do think that is to get more training data...
But my impression is that continual training is difficult, so it makes more sense to get a big chunk of it and then do a big chunk of training on that? But maybe they have done that at least once already with this new data?
- rrrrrrrrrrrryan 4y agoFrom my understanding, they have two networks: One to generate responses (with some randomness), and another network to predict whether or not that response will get upvoted or downvoted. The human upvotes and downvotes obviously train the second network. It's possible they iterate these two networks independently, but I'm not sure.