Two other theories that could explain why OpenAI's models do so well:
1. They generate chess games from chess engine self play and add that to the training data (similar to the already-stated theory about their training data).
2. They have added chess reinforcement learning to the training at some stage, and actually got it to work (but not very well).
1. They generate chess games from chess engine self play and add that to the training data (similar to the already-stated theory about their training data).
2. They have added chess reinforcement learning to the training at some stage, and actually got it to work (but not very well).