[heeton]

Extremely shortsighted take.

I currently use a local LLM to scan my notes, meeting transcripts, todo list and calendar. I currently have an iPad that displays some items (e.g. if I have important personal tasks to do today that I might have forgotten) and I would love it to be printed out like this.

Nothing to do with generating novel text, it's just used as a flexible API to my existing text and items.

[d1sxeyes]

Is that handwritten notes? What LLM are you using for that and what level of accuracy do you get?

[heeton]

No, just computer notes. Currently on llama3.2 text-only 3B, running on an M3 Macbook Pro.

Accuracy is good for my limited use case (extracting context, giving me links and reminders).

It's not yet good enough (for me) at summarising larger notesets, it might miss out useful / pertinent things in a way Claude does not.

I'd also like it to be better at inferring partial context from audio transcriptions. But hey, it's an open source 2GB model running on my laptop, it'll get better.