FlowMaster.ai: Experiment in chat-controlled apps

April 06, 2023

Back in December, I was excited about interactively working with LLMs to collaboratively create together. Github Copilot is very helpful inside of Visual Studio, but it would often give me code suggestions that weren’t really aimed at what I wanted to accomplish.

I’d done some interactive experiments with GPT-3, e.g. I gave it an engineering interview and saw it could carry on an extended conversation and actually build code as I gave it instructions. This interaction was so much more fruitful than the “advanced auto-complete” that is Copilot.

ChatGPT had just come out as well and was getting a lot of attention, but you couldn’t actually build with it - just ask for advice. It was also surprisingly good at writing rap lyrics, and I had used it with my friend’s kids recently to put together some rap songs with them, which was very fun.

FlowMaster.ai was the result. I built a rap editor along with a chat interface in which the AI knows how to interact with the editor. A human user can ask for inspiration, for feedback, or just ask the AI to make changes to the lyrics directly.

Example: Weekend Coding Jam - a rap by FlowMaster

These days, this interaction paradigm is becoming more common and I am still excited about it - I want to code not just with a smart editor, but I want to pair program with Linus Torvalds or Bill Gates (though perhaps a friendlier version of these two…). I want to write song lyrics with Eminem coaching me and offering suggestions and feedback.

The next frontier for me is agents that have larger context and can perform multi-step actions.



m [at] mpdaugherty.com