Poll: Which AI feature would you prefer for your workflow?

tjex

My masters thesis is investigating intersections between the Zettelkasten
method, creative ideation and software tooling.

Part of the thesis is a practical project, where I will extend the open-source
Zettelkasten program I use (zk) with an ML
enabled feature. The philosophy is that the ML tool / feature should not be
"executive", in that it writes text for you or converses with you like a chat
bot / collaborative partner. Instead, it should provide "latent" information
about your vault / notes, that can help reveal useful information to you about
your corpus of notes or relationships between notes. Therefore, acting much
more like a tool, rather than a co-pilot.

I've narrowed down to two potential features, and am really interested to get an
impression from the zk community here as to which feature you would prefer.
Ideally, an explanation of why would be helpful

If you'd prefer that ML just stays out of your vault completely, then no need to
reply! I get it.

The two features are:

Topic Modelling: See which topics your vault deals with, and which notes are
related to those topics (potentially in a network graph format, but at the least
in a list format). This can help answer the question "what notes do I have on
topic x".

Similarity Learning: See which notes are most similar / dissimilar to each
other. And see which notes are most similar / dissimilar to the current note
being edited. There would also be a similarity-strength setting, i.e, 100% very
dissimilar, 0% very similar). This can help answer the question "which notes do
I have that support/appose/contradict this idea?"

Andy

The Mac app DEVONthink, which is one of the apps I use as an interface to my note system, has had a document-similarity ranking feature for a couple of decades that is very useful for suggesting similar notes. It also suggests keywords, which is not topic modelling but is related, and does concordances. Writer Steven Johnson wrote about how it is useful in January 2005 in a blog post and a New York Times article. So, such features that surface latent information have a long track record of being useful to writers, although these days I'm sure you could make something more sophisticated than what has been available for 20+ years in DEVONthink.

EDIT: I actually use the similarity learning feature in DEVONthink but not the keyword suggestion (I apply my own tags manually), so based on experience I would say similarity learning is more important than topic modelling for me, but I've never used true topic modelling, so I don't know empirically how useful it is.

Will

Hi, @tjex. Your thesis sounds interesting. I like your idea that an AI tool should not write text for you. I'm curious why you consider a chat to be in the same category of inappropriate uses of ML as having a tool that writes your notes.

Please explain what you mean by "it should provide "latent" information about your vault / notes." What does latent mean?

I like using AI as a writing and thinking mentor, kind of like a virtual English professor who helps me clarify my thinking.

Topic Modelling with ML would help surface nuanced topics with my Zk that are underdeveloped. Some of these would be worth investigating more fully and some not, but seeing them would indeed be helpful in guiding future zettelkasting

GeoEng51

I don't have a strong feeling either way; both would be useful. But I lean a bit towards "door #2", so have voted for that one.

My reasoning is that between plentiful tagging and structure notes, I have a pretty good idea of my topics (but not a perfect or comprehensive idea). On the other hand, when writing a zettel, I'd love to have some indication of what other zettels in my database are similar or dissimilar or even express an opposing idea, as I could then explore them for links.

zettelsan

If limited to topic modeling and similarity learning, I would think topic modeling would be more helpful in a generic way. The reason being, if you can let the topic model give me document relations based on a topic (which in and of itself can be think of a measure of similarity). That substantially cuts down the number of documents, so then evaluating document similarity among them would be very manageable, even when done manually.

mlbrandt

I picked similarity learning because I suspect it would be more "random" (i.e. there might by more synchronicity?) but I'd like both!

tjex

@Andy . Ah cool, thanks for the pointer to DEVONthink. I'll be sure to look into it, very relevant for the "related work" section!!

@Will said:
I'm curious why you consider a chat to be in the same category of inappropriate uses of ML as having a tool that writes your notes.

Because the chat process is procedurally very similar to a bot writing for you. You ask it questions / converse with it, and it responds back to you with text. So, having a bot think/write "with" you in the context of Zettelkasten can easily leading to the same end product.

What you write, is a product of what you think. So if your thinking is also being guided by a ML model, then it's not too different than if it were to write the product for you in the first place.

The strength "too" in "not too different" is therefore reliant on how frequently we use chat, and how easily and willingly we accept its responses.

Please explain what you mean by "it should provide "latent" information about your vault / notes." What does latent mean?

By "latent" I mean, the information is "there", but not immediately / practically observable to us. I.e, we can't work our way manually through hundreds or thousands of notes and assign them to topics (i.e, tag hell 😅). But a ML model can.

I like using AI as a writing and thinking mentor, kind of like a virtual English professor who helps me clarify my thinking.

I agree. It's a great tool to check where things can be improved. I think the big issue though (talking about use en mass), is that if people increasingly accept its output as "the" answer (and don't think further themselves), we all get lead down an increasingly narrower tunnel. As the models are constantly being trained, and they're starting to be trained now on a lot of their own output which has been peppered through the internet... So their responses could begin to become increasingly generic. More like a logic based chat bot (if / then), rather than a neural network chat bot.

ZettelDistraction

Added to the similarity chorus.