Want to scrim Science Bowl but don't have anybody to play with? Not to worry - SBReader now has an AI opponent. The AI Opponent comes in three levels, beginner, intermediate, and advanced. Check it out!
How it works:
- Offline, there is a backfill script that, given a problem, finds where each level of difficulty should buzz and what its accuracy should be using an OpenAI API call. This is based on a few exemplars I give the model in context. It stores the schema in a JSON file. The locations are determined by the last 5 words read before the point.
- While the user is playing, the frontend calls into the JSON file for each problem and retrieves the buzz points. The AI mode then buzzes on those points.