Had a really exciting weekend testing out an experimental inference stack with Glamin piled on top of it for tools and memory. Also, if anyone's interested, I added some updates to Glamin that dramatically improve the C interop.
This one's been hard to explain to people.
I mean, what do you do with a programming language based in higher math alone with entirely new concepts like distance gates for logic that nobody's ever seen before?
Well, you put it into practice, and show a practical implementation based on it.
So I'm using it to build a state machine, which you can let a model sit on, and the agent should be able to record data in rich native vectors, read "memories" and associate different kinds of information, with different types of actions. Those actions can be grouped in "apps" that the system can create or edit by itself, emergently.
The beauty of doing it with distance vectors instead of logic trees... is that even smaller models should be able to do it, because they're all trained on how to use vectors, and usage falls squarely within known parameters.
From there, I wired up the system such, that the model can see into Glamin in real time, and it functions as sort of a state machine. Tools can be triggered with a simple state change, instead of wiring up strings of JSON that need to be reasoned through.
I had a whole article written up, but I realized that the experiment isn't really done yet. I'm at a point where the concept is proven, and I'm building out a practical implementation that you can run locally.