Today I've been working on splitting out more back-end functionality from the Qt-built Humbug front end. I've taken the latest DeepSeek v4 for a spin while doing this.
Separating the backend
I started separating out the backend of Humbug a while ago but I want to get much closer to doing this now. There were 3 interesting problems to solve today.
The first was to split the System AI tool out and make it front-end agnostic. This did mean adding a few more things into the context registry than I'd originally wanted because the registry now needs to understand the idea of context positioning, but this then allowed the AI tool to manipulate contexts and have the UI simply react.
The second was to make the editor context Qt-unaware. This involved creating an abstract editor concept that is just a list of lines and then have the Qt editor react to changes in that.
The final change was to make the column manipulations happen in the context registry too and have the UI react to them.
Still quite a lot more work to do in this area, but we're moving in the direction of having the UI be able to run entirely disconnected from the backend (just using a network link).
DeepSeek v4 Flash
I thought I'd try DeepSeek v4 Flash on Ollama cloud and throw it at the last of the changes above. It aced the test, but at staggering speed. It's way faster than any other model I've seen, but also did incredibly well at the agentic workflow required. Unlike Claude and GLM, I didn't have to remind it of things as we went along.
It didn't completely do things right as I had to spot a couple of methods that were now unused, but it was impressive.
What was completely new, however, was that it would launch a task and then think about something else while the task ran!
