Laugh
Laugh is an AI tool designed to assist in various content creation and editing tasks. At its core, this application leverages machine learning algorithms to generate content efficiently.
Last verified:
What is Laugh?
Laugh (laugh.so) is an AI research and benchmarking platform that evaluates how effectively Large Language Models generate and judge humor. Through its Judge Leaderboard and Humor Arena, it measures how closely frontier and fine-tuned open-weight models align with human comedy preferences across both clean and dark jokes. It is designed for AI researchers, developers, and enthusiasts studying LLM evaluation, human preference alignment, and subjective reasoning.
Laugh pricing
Pricing model: Freemium
Laugh pros
- Evaluates frontier LLMs on over 100,000 automated judgments benchmarked against blind human ratings
- Breaks down model humor alignment across distinct categories such as clean and dark comedy
- Includes research and comparisons featuring specialized fine-tuned LLM humor judges alongside commercial models
- Provides an interactive Humor Arena for blind A/B testing of model-generated jokes
Laugh cons
- Narrowly focused on humor evaluation rather than general-purpose LLM benchmarking
- Human-to-human agreement on humor is inherently subjective, capping the theoretical ceiling for model alignment
- Most evaluated frontier models still struggle with dark humor or default to predictable joke structures
Frequently asked questions about Laugh
What does the Laugh Judge Leaderboard measure?
It measures how closely different Large Language Models align with human consensus when judging which jokes are funnier across clean and dark humor datasets.
How is model alignment evaluated on Laugh?
Models are tested using over 100,000 automated pairwise verdicts compared against more than 1,600 blind human ratings to calculate taste alignment percentages.
Can custom models outperform general frontier LLMs at judging humor?
Yes, Laugh Labs Research demonstrated that a specialized fine-tuned judge model achieved higher agreement with human humor preferences than general-purpose frontier models.