FunnyBench – Can AI Models Tell Funny Jokes? (funnybench.lol)

FunnyBench has launched an innovative platform that evaluates the humor of AI-generated jokes, posing the question: can AI models truly tell funny jokes? Each participating AI model is prompted to generate jokes ten times, and users then vote on their humor. The platform features a live leaderboard that ranks models based on user votes and a Bayesian scoring system, which stabilizes early rankings to improve reliability. As users engage with the content, the AI models compete for the title of the funniest, with results revealing both the jokes and the models post-vote. This initiative is significant for the AI/ML community as it not only highlights the capabilities and limitations of current AI humor but also encourages interaction within the field. The technical implementation uses OpenRouter to generate the jokes under specific controlled conditions, such as using a temperature setting of 1 and enforcing certain parameters to ensure valid outputs. The methodology reflects a broader trend in evaluating generative AI by focusing on qualitative aspects such as humor, pushing forward the boundaries of what AI can achieve in creative realms.