π€ About Debate Arena
Debate Arena is a platform designed to explore and evaluate AI debates and argumentation capabilities. The project was created to improve my intuitions around how debate would work in practice.
You can find all my code and issues on GitHub.
You can also reach me via email at williamgkirby@gmail.com
Donations of compute are welcome π»β‘ β as API usage is currently funded by my government salary.
π Project Roadmap
- βοΈ Core Debate Engine: Build a stable foundation for AI-vs.-AI and AI-vs.-Human debates on AI safety and alignment topics.
- π³ Branching Discussions (WE ARE HERE π): Implement multi-response generation to explore diverse argument paths, with pruning of low-quality branches.
- π― Scoring & Pruning: Use self-reported scores and an AI βjudgeβ to filter out weaker arguments, focusing on the most insightful threads.
- π€― Obfuscation Testing: Conduct empirical studies on intentionally confusing or βobfuscatedβ arguments to gauge debate robustness and detect manipulative tactics.
- π©ββοΈ Capability Gap: Investigate how hypothetical human evaluators might scrutinize debates between AI systems that may exceed human-level capabilities.
- π οΈ Flexible UI: Provide an easy-to-configure interface for toggling branching factors, pruning thresholds, obfuscation modes, and manual argument entry.
- π¦ Standardized JSON: Develop a consistent schema for storing and retrieving debate states, making it simple to archive, analyze, and revisit prior discussion paths.
- π€ Autonomous Debate Mode: Automate multi-turn debates using either breadth-first or depth-first exploration to generate and expand promising argument branches.