It is also about User Interfaces (UI). 📱
Last year, I worked on Open-Source integrations with AI-as-a-Service APIs. The goal was to let Dataiku users leverage NLP and Computer Vision models from Amazon, Google & Microsoft with no-code.
A daunting task
20+ APIs, with different input types, architectures, and output formats. I spent months creating a robust framework to send data to the APIs and standardize the results. 🧑💻
Only at the end of the project did I start creating user interfaces. For fun, I downloaded random images and fed them to the APIs.
Some unexpected results came out…
- My Little Pony posters were deemed as sexually explicit. 🎠🔞
- In group photos, not everyone was recognized as a person. 🙈🤪
- AIs were strangely unable to count sheep… 🐑🤯
I kept 2 takeaways from this project:
1. It is difficult to create interfaces to AI models.
2. Even AIs made by tech giants have bugs.
With Giskard, we want to make it easy to create interfaces for humans to inspect AI models. 🕵️