Humans taking an 'AI benchmark' test — that's the concept behind the trending joke site 'HumanBench.' Launched on September 26th by developer Alex (@Alex_ybuild), it diagnoses what 'B (billions of parameters)' model you would be if you were an AI.
HumanBench is a quiz format consisting of 76 questions and takes approximately 13 minutes to complete. The better your score, the more difficult questions you'll be presented with. It also includes 11 'classic AI blunders.' These include various questions known for tripping up AIs, such as 'How many 'r's are in 'strawberry'?' and 'Which is larger, 9.11 or 9.9?' The test also features the well-known 'If the car wash is 50m away, would you walk or drive?' question, familiar to those in the AI community.
Once the diagnosis is complete, it displays your 'parameter count' as if you were an AI model, whether you're Dense or MoE (Mixture of Experts), your ranking if placed on an AI model performance leaderboard, and your AI 'personality type,' among other results. In fact, many users are sharing their results, such as '235B,' on social media, and it's spreading as a fun way to imagine oneself as an AI model.
Of course, HumanBench is not a real intelligence test. HumanBench itself clearly states, 'It's a joke; parameter count is not brain capacity.' The test is designed to run locally on your device, and answers are not sent to a server.
In the world of generative AI, it's common to compare performance using metrics like 'how many parameters' or 'what score did it get on the benchmark.' HumanBench's humor lies in imagining what would happen if humans were measured by the same logic. If you've been laughing at AI's strange answers, why not give it a try and see if you 'hallucinate' in the same places?
