5 ms·
Well, Go is a pretty complex game, and AlphaGo RL’d its way to excellence just by playing the game like you describe.
by robotresearcher 12d ago
Well, Go is a pretty complex game, and AlphaGo RL’d its way to excellence just by playing the game like you describe.
- bananaflag 12d agoBy training. When we access the API, we don't get to train the model, we just do inference on the already trained model.
- robotresearcher 12d agoOh, I see. That’s a very different requirement. It’s not a technical limitation but a product decision to not allow training. An advantage of properly open source models is that you can train and tune them. It’s an interesting challenge though. I might start to tackle it by having the model write its own tool program(s) to play the game. It’s possible that the model could choose that strategy itself from a high level prompt alone.