Re: Status of Voilib and building AI tools

Last week, a Voilib user reached out to me with two questions: how the project is currently progressing, and whether I could recommend other AI-based tools that might support his learning journey.

I’d like to share part of my reply here, since I believe it could be useful for others as well:

I’m glad to hear Voilib was useful for you. I have to admit that keeping it updated isn’t one of my main priorities these days. It’s a project that’s a bit too big to simply self-host on a Raspberry Pi and check in on every few weeks. It’s harder to maintain when the number of channels grows: the transcription process uses a lot of CPU, and the embeddings database grows really quickly. Still, building it end-to-end was a lot of fun and a huge learning experience.

If I had to rebuild it today, I would try to simplify every component much more. For example, using sqlite-vec could be a nice way to store embeddings without relying on an additional service. And without any doubt, I’d make heavy use of Simon Willison’s excellent llm library (https://github.com/simonw/llm) as it handles embeddings, tool calls, and supports any LLM provider. Both in a CLI and as a Python library.

In fact, that’s what I recommend to everyone: explore everything Simon has built on top of that library. It’s impressive, and you can learn how agents really work without all the unnecessary complexity others often add. Write a for loop using tool calls and that library and you will have built your own agent.

Good luck on your learning journey!

Best, Pablo

And that's it. AI hype or AI winter, just start small (really small), build things, and you’ll see what’s signal and what’s just noise.