Today I ported Peter Norvig's famous Sudoku solver. 100 lines in Python turns into 200 lines in Lua. But something is starting to click in my head about constraint propagation.

egalite-ai -- prolog-lite thingy that does pattern matching in JS

github.com/uprun/egalite-ai

Have been trying to build out and share more demos of practical uses of agents these last few weeks using the ai framework I've been building, and all my past demos have involved calls to LLMs, but this latest one is just API calls between different services but using the agent patterns.

Was just struck by how different and much simpler the design for this came out using an agent framework than if i had built it a few years ago before diving in to building AI apps.

I'm starting to wonder if a pattern we'll start to see with all these agents people are building is to use an LLM in certain places at first (or just an old-school pre-llm implementation), but then over time replace that call over to something more deterministic or at least specialized once something stable comes out. Almost having to have well-factored designs up front because you're guaranteed to need to swap components out as better, cheaper, or faster ones come along.

I have redone my UI concepts. A dockable panel now is "just" a notebook. This works out much better. Before I was adding lots boilerplate to slice a top level notebook into pieces that were manually placed. It was hard to program and very fragile. Furthermore, it was hard to navigate the module structure of notebooks.

This new way is much better. A whole lot of concepts are removed, but also, I think notebooks tend to explain what they do within the notebook, so you end up learning about the orthogonal components in a more approachable way, plus the whole concept of a module structure manifest to the UI in a way that encourages learning.

The entire concept of a top level notebook kind of disappears altogether as well, which is closer to what the serialized and runtime format is. You can have circular imports in Observable, there is no "top" at the runtime level, its a graph of reactively cooperating modules. The UI now reflect that properly.

I've mapped the nested docker layout structure to the hash URL so the view layout can be saved and hyperlinked, and in HTTP terms is a pointer to a specific part of the cooperating module graph. I will add the ability to set a default layout if no hash is found, and that kinda servers the purpose of pointing to what the intended purpose of the assembly is (a role previously fulfilled by having a single top level notebook).

removed all LLMs from my laptop, I better download a snapshot of Wikipedia