I have tangled thoughts about digital gardens.
This website is in part an experiment to gain insight into my confusion. For some time to come, I’ll operate it as interlinked half-baked essays.
Structure and Intent
I tend to agree that a graph structure is a more natural way to dump thoughts. Thus, I like anything graph-like: WWW, wiki, Scrapbox, Obsidian, Quartz, …
In this view, a typical well-written text (think: books, papers) is a “flattened tree”.
One simplified narrative is
Linear text is a legacy, costly format. Now that we have better technologies, “waste” should be reduced by graphing everything.
Engineer me agrees 100%. But what is lost in this process haunts me.
Information in Linearization
To form a tree, you need to select a subgraph. In addition to the subgraph selection, you often prune some ideas entirely, or tweak information density.
These choices themselves impose a certain worldview. You can even compute how many bits are encoded!
Flow of Text
Linearized text has a certain rhythm of information, like waves from the author hitting a reader. Texts flow through computers, human brains, AIs, and finally form an ocean of collective knowledge.
Wait. Why did I use the “water” metaphor? Where does it come from?
I love tinkering with these kinds of text.
Graph-based tools present a tension; To utilize the nature of graphs, each page should be small and metaphor-free so that the reusability of ideas can be maximized.
However, this forces us to draw boundaries to the otherwise vague mist that exists in the brain. Now it can only manifest as solids, rather than fluid!
This goes counter to what graph notation is supposed to provide; freer expression of thought.
Multi-Language
I always feel that no combination of tools satisfies my desire to use languages. yet.
For a background, my native language is Japanese. I’ve never lived in English-speaking countries, but I did work in such a company for several years.
Fluent use of natural languages by computers was long thought of as an impossible dream; until it wasn’t.
I already think LLMs today can do a better job of writing natural, culturally-sound text. Better than me, and probably better than most people.
Since I have a somewhat good digital garden in Japanese, I pondered the idea of somehow auto-translating it using an LLM-based system.
However, a system cannot translate what’s not written. Also LLMs are extremely good at coming up with plausible intepretation of missing information. With memo-like texts, I fear that they’ll end up so “thin” that they don’t represent me anymore.
Languages, especially when you try hard to come up with a “good” representation, force me to see myself more clearly. I’ve used the word “good”. Both languages have a myriad of ways to phrase something similar; yet all of them differ. And there is never a simple 1:1 or even N:N mapping.
Each language seems like a landscape. Parts of them can look similar, but ultimately they’re entirely different. Writing text is weaving through that land by picking up each word, to represent a thread which is part of me, and in turn forms me.
I’ll probably keep using different forms of output, at least the ones I’m comfortable using.
It would be very interesting to see what I’ll do when ML systems become so much superior that they can read into my thoughts better than myself.