Hello everyone! For those unfamiliar with my name, I am CocoaMix [he/him]. I am the server backend operator of @roborosewater-masters. I run the neuralnetwork that creates the cards and retrain it. (But am not the person that runs that account)
A crash course on Roborosewater Masters, and how the card generation works:
- The card text is not written by GenAI. No ChatGPT, LLM, stuff like that is used to write the card.
- We utilize an ancient char-RNN built 10 years ago. The same model the original RRW used. A machine learning neural network!
- The only training data the neural net has is the entirety of magic card history. This includes all names, rules text, flavor text, etc. (Available at mtgjson.com)
- Card art is added separately after the neuralnet has written the text, and is left up to the individual curators to add what they want.
Please ask me any questions you have about Roborosewater and card generation and how that all works!
