LLM Wikis are having a moment, and my biggest worry about using them is that they cut down on germane load in PKM work.
Back in the early days of the ChatGPT craze, something about it felt off to me, so I wrote a newsletter piece: The One Skill You Must Never Outsource to AI – Summarizing. Generalizing a bit further, though, the underlying issue seems to be this reduction of germane load.
The LLM wiki I'm targeting in this post isn't the kind used as long-term memory for an AI agent — I mean people pairing it with Obsidian and using their PKM as an LLM wiki. To give it a name, it's the problem you run into when you do Vibe Noting.
Germane load in PKM
There's a germane load that arises when you use a PKM. Stripping that load away does the user very little good. And certain patterns of AI use do exactly that — they take germane load away from the user.
It's like doing a squat where I'm holding the barbell at the bottom, the coach lifts it all the way up for me, and then I shoulder the bar again. The process is the entire point, yet this drops the process and keeps only the outcome. I hear people these days skip exercise altogether and just take Mounjaro or Wegovy — it gives me the same feeling, and I find it off-putting.
The germane load that arises in PKM use comes in the following kinds.
First, linking. You can't do PKM without links. The act of connecting one note to another is an important germane load, because it lines up exactly with what cognitive psychology calls elaboration. The question "what is this note related to?" pulls my existing notes out of long-term memory, sets off an encoding process along the way, and lets me hold the note I'm writing right now in memory far more solidly. And forming schemas by connecting notes from different contexts is itself a key part of how insight emerges in PKM.
But linking notes purely according to the AI's recommendations reduces that germane load. The process is easy and fast, so building an interconnected network of notes is easy. Building a graph view worth showing off online is easy. Yet since those concepts were never actually connected inside my head, the result is no different from material sitting out on the internet. They're my notes, but they aren't really mine.
Second, extracting core concepts. You could also call this summarizing. On this point, The One Skill You Must Never Outsource to AI – Summarizing may be worth a look.
How to deal with it
The first remedy is accepting desirable difficulty. This is a mindset-level approach: don't dodge the parts of PKM that feel hard and irritating — accept them. Writing and maintaining notes is unmistakably a high-cognitive-load process. But if you lean on AI to escape that pain, you end up building a PKM that's effectively useless. So start by admitting that the pain of maintenance is unavoidable. (Which of course doesn't mean we should accept extraneous cognitive load.)
The second remedy is navigating your notes by digging further and further into the linked ones. (Keeping Yourself Looking at Your Notes) For a PKM to function as a second brain, it has to become an add-on to my actual brain — the first brain. The wider the gap between my PKM and my brain grows, the more useless a second brain it becomes.
To work out where that gap can be narrowed, I want to borrow a few elements from the extended mind thesis. For an external tool to become an extension of my mind, it must be (1) accessible at any time and (2) trusted. I read "accessible at any time" as going beyond simply carrying a smartphone around so I can open Obsidian — it means being able to reach the note I want, quickly, whenever I want it. Looking often at the notes I've made, whether by AI or by my own hand, makes reaching the note I want faster. In the process my skill at finding notes grows, and I also restructure the notes so they're easier to find. And looking at notes increases fluency, which in turn increases trust itself. 1
The third is Progressive Summarization. For convenience's sake, it'll be hard to avoid making notes with AI entirely going forward. But once they're made, you can generate germane load through progressive summarization. Summarizing must not take the form of merely deleting unnecessary sentences. You have to paraphrase it back in your own words. For instance, delete an entire paragraph and write that paragraph's core in a single sentence of your own.
Footnotes
-
There's a well-known problem where simply reading a text often increases fluency without actually increasing understanding. Here, though, I want to take the upside of that — using increased fluency to secure trust. ↩