Intentionality is the Spirit of the Work
As I have been thinking a lot about how to build my notetaking system, a key subject of ponderance have been on the limits of the use of automation, and in particular, language models. In particular, what should it be allowed to do, and what should it not.
This is especially important since, as I’ve previously written, automation offsets the need for discipline. Another important idea is that of discipline, or the lack thereof, being the harbinger of rot – the notes repository of one who lacks discipline is more likely to be unorganised, and more importantly, unorganisable, as they would likely have not even curated sufficient context and metadata for any form of automation to organise it for them with reasonably acceptable quality (it is given by the premise that such a person would also not be inclined to organise it themselves).
Drawing a Line
Instinctually, I’ve been driven to be more comfortable with the use of, say, embedding models to calculate the semantic similarity of the content, but not with any form of agentic mechanisms to explore, or god forbid, write the contents of my notes. The question, then, is why? What is the precise line which separates the two?
This question serves a twin purpose: that is,
- it is the falsification criteria of my aformentioned hypothesis; if a precise discriminant cannot be found, then it is but a mere gut feeling, and
- it will be a necessary guideline for the implementation of my system (when I finish building its building blocks and finally get around to doing so™️).
The Machine Does Not Intend
I’ve recently came across an interesting piece of computing history from Simon Willison’s Weblog: a page from an internal IBM training which says
A COMPUTER CAN NEVER BE HELD ACCOUNTABLE
THEREFORE A COMPUTER MUST NEVER MAKE MANAGEMENT DECISION
This begs the question, then: why can’t one hold the computer accountable? After all, despite its many, many faults, computers are now capable of performing actions derived from some, albeit limited, form of intelligence.
I think the answer lies in the computer’s lack of intention. For some – arguably most – definitions of what constitutes intention, computers cannot exhibit it. Insofar as any technological advancements have brought us, intention remains the sole domain of humans.
This captures with it the heart of the discrimination at the thesis of this post: the use of agentic language models to explore or write your notes is a misguided attempt to automate intention. This is something that cannot – and, I’d argue, should not – be automated away.
This is, admittedly, an unsatisfactory answer. It leaves the definition vague, abstract, and largely in the realm of highfalutin philosophy, rather than a formal, rigorous construction.
Don’t Automate Away Your Intentions
When one, say, runs Claude Code on their Obsidian vault and asks it to collect information on some topic, they are leaving the computer to fill in the dots. Natural language is impreciseSee Dijkstra’s writing, ‘On the foolishness of “natural language programming”’.; consequently, any instructions provided to the agent will contain some gaps with respect to one’s true intentions. It is then left to the agent to make assumptions to bridge these gaps by itself.
A second loss of intent would be on the decisionmaking regarding the exploration path taken while gathering the prerequisite information. It is up to the agent to decide where to go, and what, among its findings, to include and what to exclude. Not only does this lead to a loss of control on the human’s end on the final output, but it deprives from one the opportunity for a serendipitous encounter with non-obvious, interesting connections between ideas (for some definition of interesting).It is, after all, one of the key strengths of atomic, relational note-taking systems, such as the Zettelkasten.
I don’t suppose that it would be necessary to provide an exposition on why this also applies to allowing the language model to write in your notes, as the mapping should be self-explanatory and literal.
This contrasts with “simpler” technologies such as embedding-based similarity measurement as the latter is more specific in its use; all it is able to do is evaluate the semantic similarity of two texts. This, therefore, leaves the user alone to define the intent.
Coda
As with most things in life, moderation is key. I think that with technology in particular, moderation is defined by recognising the precise limits in what it can and cannot do. In other words, as Feynman once said:
The first principle is that you must not fool yourself – and you are the easiest person to fool.
By dismissing new technology without so much as a glance at it, you are fooling yourself about what it can do. Yet by subscribing to popular hype without a healthy dose of critical evaluation, you are fooling yourself about what it cannot do.
You must not fool yourself.