Hey there! This is my last post for 2022, and I’ll be back to the regular updates in 3 weeks. Thanks for reading my writing so far, for sharing your feedback and encouragement, and for keeping me accountable!

⬅️Last week, I set out to:

❌Prepare for customer interviews

✅Try to decide whether going into a super crowded space is worth it

✅[Maybe] Try to see how far I can get in building out one of the ideas above

Most of my progress so far has been inside my head or on paper. I was missing actually doing something. So I decided to first focus on trying to build a bare-bones version of a Slack bot that can use AI to generate images from a prompt (one of the ideas I shared last week was a way to “give more personalized kudos to colleagues, where each kudos comes with an AI-generated greeting card”). This was a ton of learning, plus it helped me decide to move on from engagement tools.

🤖What I learned by building an AI-powered chat-bot

Let’s start with the obvious. I learned the technical parts of building the bot. I mainly followed these two [1, 2] tutorials for this. The coolest thing about it was that at each step, there was something that wouldn’t work as expected. Having to break my head and search for answers to fix things was when I learned the most.

For example, I wanted to be able to call the bot using a slash command in Slack. To do this successfully, the bot needs to answer back within 3 seconds. However, the AI model takes around 15 seconds to come back with an image. Each tutorial had an answer for this, but I couldn’t get either to work. What worked in the end, was taking a break to cool off, then reading more about the platforms (slack, aws lambda) and packages (zappa and flask) I was using. A break + knowing how the technology worked helped me find the bug in my implementation.

And now I can access Stable Diffusion through Slack to make my own cool images!

santa

But building this app helped me understand more about the idea. I started from the hypothesis/feedback that kudos are not personalized enough, and that images generated by AI would make the kudos more personalized. However, when I got the bot to work, I realized three things:

  1. Adding AI-generated images is too slow:

    • It takes about 15 seconds to generate an image
    • That’s too long for an instant messaging platform, where you expect things to be… instant
    • Making it faster would lead to poor quality/too high costs
  2. The images generated would either be not personalized or not relevant to kudos/recognition:

    • Recognition messages are typically something like “Karen, thank you so much for contributing to project Acme! We couldn’t have made the sale without you”. To this, a generalist model like stable diffusion responds with:

    karen

    • This means you’d have to tune the model. Maybe train the model with input images that inspire gratitude, recognition, encouragement, etc. However, I don’t think the resulting images would come out very personalized. They may be unique, inspire positive feelings, but they won’t be personalized.
  3. I overestimated how “cool” it would be to get these images. It’s not that cool. It’s cool when AI generates beautiful art, but it’s not that cool in the context of giving kudos to someone.

💭Closing thoughts

Spending time building made room for reflection on the real problems behind employee engagement. I realized that this is really a problem of relationships between people. The biggest cause of disengagement is that people don’t feel like their bosses and/or colleagues care about them. You don’t solve this using an app. You solve it by hiring/developing people that care about other people.

Also, this space is overcrowded. So I’ve decided to move on to something else. This means I’m back to exploring ideas. More on this next time.

➡️Next steps

  • Write up my checklist for an idea worth exploring
  • Try more ways to find micro-saas ideas.

Thanks for reading and happy holidays 🎅!