3 ms·
The parallels with AlphaZero are not so easy. The problem is you need some sort of arbiter of who has "won" a conversation but if the arbiter is just another t
by whimsicalism 4y ago
The parallels with AlphaZero are not so easy.
The problem is you need some sort of arbiter of who has "won" a conversation but if the arbiter is just another transformer emitting a score, the models will compete to match the incomplete picture of reasoning given by the arbiter.
- sebzim4500 4y agoWe have that though, just train on the buzzfeed articles that get the most attention or the tweets that get the most likes, etc.