Building Crowdians: From a Tamagotchi AI Chatbot to Version 4

Hi, I’m currently building Crowdians, a feedback platform powered by synthetic AI personas.

Crowdians is actually already on its fourth version.

I want to document how it has changed, what I’ve learned along the way, and where I plan to take it next. Think of it as an ongoing postmortem or a notebook of mistakes and lessons learned.

The first version of Crowdians got around 150 visitors on its launch day, but only three people signed up.

To explain why I started building it, I need to go back to the AI landscape at the time.

Back then, ChatGPT’s overly agreeable responses had become such a common joke that phrases like “Wow, you really nailed the core of it” were turning into memes.

I became interested in why AI systems often seemed so eager to agree with users.

One thing I learned about was RLHF, or Reinforcement Learning from Human Feedback. RLHF is used to align AI responses with human values, preferences, and intentions. However, optimizing for user satisfaction can also contribute to overly agreeable or flattering behavior.

That led me to a question:

What if I could chat with an AI like a Tamagotchi, gradually help it evolve, and let other people teach it when it gave a bad response instead of correcting everything myself?

The first version had two main features:

1. Chatting with a character I had created.
2. Letting other users answer quiz-style questions about what was wrong with the character’s responses. (Sorry, I only have screenshots from the Korean version left.)

At the time, Grok Imagine was available for free within certain usage limits, so I used it to create the characters. I then converted them into spritesheets with Aseprite and used them as assets.

I also made the character users encountered depend on the first action they took after entering the site.

After building the service, I launched it and shared it across several online communities. Around 150 people visited on the first day.

I was excited and kept monitoring the database and server logs, checking whether any errors had occurred or whether the AI API connection had failed.

But the result was very different from what I had expected.

Most people looked around a few pages and then left.

Out of those 150 visitors, only three signed up.

I started asking myself why people were leaving so quickly, and I also asked some of the users directly.

I learned two main things:

First, the AI responses took too long.

Second, users felt that the core feature, giving feedback on the AI’s responses, felt more like work than something fun or rewarding. People had come to interact with a character, but I was asking them to help train it.

That experience taught me that the problem wasn’t simply that the service needed more features. The core experience itself wasn’t compelling enough, or at least the value wasn’t clear to users.

That was when I started thinking about how to change Crowdians.

And that’s how I began building the second version.

To be continued...