HumanBench: 48,000 visitors, zero ads, and what we'd keep
How a 76-question quiz where you play the AI spread from a few hundred followers to 96 countries in three days, and the five things we'd do again.

HumanBench asks one question: how many B are you? For 76 questions you play the model. Someone sends you a message, you pick how to reply, and at the end you get a model card of your own: a parameter count, a benchmark chart, and the AI you answer most like.
We launched it one afternoon in late September with one tweet from an account that had a few hundred followers. No ads, no launch list. Three days later:
- 48,221 people had opened it, 35,782 had started, and 15,329 had answered all 76 questions.
- 55% of the people who finished shared their result. For a quiz, 20% is considered good.
- It reached 96 countries and regions in 7 languages.
- GIGAZINE and ASCII.jp in Japan wrote about it without us asking. One GIGAZINE article sent 2,357 people, and Japanese became the biggest language.
What we'd keep
1. One button on the front page
The first version asked you to pick a length and type a name before starting. We cut all of it. You press start; naming your model comes at the very end, when you actually care.
2. You are the AI, not the person taking a test
The questions aren't "what did this model do in 2024". You step into the famous moments yourself: the car-dealership chatbot that sold a car for a dollar, the strawberry question, the customer who just wants to be transferred to a human.
3. The result is about you, with evidence
Your persona card quotes lines you picked: because you said…. People share things that look like them.
4. Whatever gets shared opens straight into playing
Your share link is your own result page, with a challenge on top. A friend opens it, sees your score, and starts. On day two, 40% of new visitors at peak hours came through a friend's link.
5. Watch where people leave, question by question
A long multi-turn puzzle at question 3 lost 12% of players. Moving it later didn't help, so it went. Five fast questions now open the run.
What the numbers said
Each person who finished brought in about 1.3 new visitors, but only 0.35 new finishes. Under 1, every round of sharing is smaller than the last, and the first two days brought eight in ten of all visitors. A 76-question quiz with a 19-minute median is long. The biggest single lever was length, which is why we added Jev, an AI teammate that skips a batch of questions for you when you slow down.
Now it lives on Dudu
HumanBench is the first thing in Dudu's Play shelf, and it's the template for what Dudu helps you make: tell Dudu "make me a fun test like HumanBench" and it builds one you can pass around.
The full three-part retro, with every chart, is at ybuild.ai/humanbench/retro. The code is open at github.com/alex-ybuild/humanbench.
Play HumanBench →