So I'm thinking about the Glamin project.
That's one I started and finished a few months ago. It survived some testing, and numerous experiments. The idea there was to create a language that worked as an application against a vector index that served as the program.
Super useful. But in tests this weekend with it, I discovered that it's really easy for an agent to break a program, and not be able to fix it, once the app lives inside the vector index.
There's no easy fix for this problem, as long as indexes are the app.
I like the concept, because it effectively creates an agent facing system that eliminates multiple layers of tool calling. And it produces a programming language for humans and agents that's totally free of cyclomatic complexity.
It's probably still useful for some things, but I think I have to come to terms with the fact that, well, it was an absolutely fascinating experiment that should probably never be deployed at scale in a meaningful way IRL.
But on the other hand, there is a more interesting way to do something like this, even if it would be a monster redesign of the whole system.
Learned embeddings.
And actually, I think that's where it starts to get interesting.
More soon.