I would actually bring a parallel to the device driver-firmware blob split that’s common with hardware support in Linux. While the code needed to run inference with a model is straightforward and several open source versions exist already, the model itself is a bunch of tensors whose behaviour we don’t have any visibility into. Bias is less a problem of the inference code and more an issue with the data it was trained on
I can suggest LogSeq as a nice alternative for Obsidian. Notes are all in Markdown too!