Balance AI With Mind Training • Zettelkasten Method
Balance AI With Mind Training • Zettelkasten Method
The rise of AI threatens to erode our thinking skills. Just as fitness training replaced everyday movement to stay healthy, deliberate mind training keeps the mind sharp, like working a Zettelkasten.
Howdy, Stranger!
Categories
- 3K All Categories
- 152 Research & Reading
- 693 The Zettelkasten Method
- 8 Knowledge Work
- 100 Writing
- 464 Software & Gadgets
- 154 Workflows
- 731 The Archive
- 15 Plug-In Showcase
- 88 Resolved Issues
- 225 Projects Logs and Journals
- 83 Project: Zettelkasten.de
- 53 Critique my Zettel
- 171 Random
- 373 Introduce Yourselves!
Comments
Yes; I have the same temptation. I find after a long back-and-forth discussion with ChatGPT, just to zero in on certain ideas, that I've had a long conversation. ChatGPT knows I have a ZK (from previous questions posed to it), so it offers to write one or several ZK note(s) for me. Asking it to do so would be the height of laziness and entirely contrary to the whole purpose of building a ZK. I know your book will provide some more fodder for this particular discussion.
Infuriating, isn't it? You can tell it not to offer services like that in your prompt. Here is one version I use:
:constraint: avoid proposing next steps or additional questions unless requested.
@tomp
Occasionally I will have long "conversations" with ChatGPT, as I try to get it to be more specific about its answers, or to fill in missing pieces of information. I suspect my past behaviour in this regard has implicitly given ChatGPT permission to propose all sorts of related information for my consideration. But it is useful to have a way to stop that behaviour.
Just for fun (and to see how far ChatGPT would take things), this afternoon I asked it to produce a Zettelkasten-like set of notes on a topic in which I'm currently interested - gut health, prebiotics and probiotics. Following is an index of the main notes that it created. Each note contains 5000 to 10,000 words and is divided into sections, with 10 to 20 sections per note.
I had previously asked ChatGPT about some details of using NotePlan. It assumed that I used NotePlan to store my ZK (which is correct), and so the sections were designated as main headings in the note (using markdown conventions). NotePlan allows me to link to a section heading and not just the note title, so this was a reasonable assumption on its part.
To be clear, I'm not promoting this approach; I'm simply showing how easy it would be for an unmotivated person to do so. I'm fully in agreement with @Sascha and with the principles in his book on Zettelkasten, that one needs to put in the mental energy and focus to create each zettel. In effect, one needs to "earn the right" to progress one's Zettelkasten.
Having said that, I suppose there is an intermediate path one might take when trying to quickly learn about a topic of which one is totally ignorant, which involves some intense brain work on our part combined with some assistance from an AI.
I had a different idea for how to combine a chatbot with a Zettelkasten. I thought that the CB could get trained on an existing ZK, and that it should be especially good at finding non-obvious patterns. The key sticking point is how to prevent the CB from adapting to the practitioner, since otherwise it would become unable raise surprises.
Interesting article in The Atlantic: The People Who Will Thrive in the AI Age. (Sometimes it has a paywall, sometimes it's directly accessible.)
The article imagines a future with extreme cognitive polarization: "Some people will use AI to think more. Other people, maybe most people, will use AI to think less." It also makes a comparison with fitness training: "Modern technology wants to turn you from a mental muscle builder into a mental couch potato." :-)
It recommends some techniques for a more challenging AI usage. For example:
This is doable; I would use a local system myself, but I don't want to splurge on an expensive machine.
LLMs write bland, edgeless prose and invariably dilute whatever point you want to make. It is difficult to overcome the LLM proclivity to temporize, even with custom instructions, which are no guarantee against the model's default smoothing and caution. A criticism of the status quo that doesn't mention or imply a conspiracy will trigger the gratuitous LLM caution against conspiracy theories. This has been my experience with ChatGPT, Claude, and Gemini.
I have some AI slop in my ZK, but I have to pare this down or remove it altogether. I'm preoccupied with validating and repairing note formats that deviate from the specification in the Zettel GitHub repository. Who knows, I might even read my Why create a Zettelkasten.
Zettel GitHub. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
My laptop would overheat if I ran a model for very long.
The chatbot is unable to "know" if its making things up so a prompt cannot completely prevent it. One can, though, insist that it verify or find real references for what it says. That's not perfect either but it's often an improvement. But if you are mainly using it as a sounding board, hallucinations don't matter much.
I don't want a LLM to do any of those things for me. No note writing, no summarizing my notes, no prose for my writing, no producing an outline for me. It could possibly be useful, as I wrote, in finding patterns I hadn't noticed or were different from what I would pick up on, as long as it doesn't adjust too closely to me. Current chatbots, with this attention mechanism, do adjust and that would have to be prevented.
I use a chatbot for web searches when I can't work out how to find something. I also sometimes use one as a sounding board. And I will ask one to tell me the standard or consensus opinion about something. I figure that one of the things they should be able to do well is an average or consensus response.
I think a lawyer could do well to have a chatbot digest a number of possibly relevant cases and then ask it how the present case of interest breaks that pattern. That's what a lawyer needs to know: how his case differs from the precedents. The LLM's are good at detecting patterns.
I did not propose note writing, summarizing, prose generation, or outline generation. My point was narrower: a local model trained on an existing Zettelkasten might be useful for detecting patterns, but generated prose and generated notes are exactly the sort of AI slop I would want to remove.
Zettel GitHub. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
I've been trying out Proton's Lumo recently. Even without special prompting it distinguishes surprisingly clearly between knowledge derived from the model(s) and knowledge pulled from external sources.
Yes. And I find it helpful to go one step further and to define different roles for different kinds of interaction. For example I'm experimenting with a configuration that is optimized for socratic dialog. On the other hand I have configurations that are optimized for topic-specific fact-checking and research.
What I don't find helpful so far are AI agents that interact directly with my zettelkasten.
I agree with @Sascha that it helps to treat a Zettelkasten as "a gym for my mind".
AI coding can help standardize note formats. An AI session produces code to check my Obsidian vault against the note specification in the Zettel repository and can repair some of the notes that deviate from it. For example:
Checked 1145 file(s): 901 valid, 244 invalid. Issue category counts: Count Name ----- ---- 99 missing_see_also 68 filename_id_mismatch 50 h1_mismatch 19 duplicate_h1 14 title_id_mismatch 12 missing_h1 10 missing_references 7 invalid_yaml_header 6 missing_frontmatter_key 5 duplicate_references 5 order_issue 3 invalid_yaml 2 invalid_id 1 h1_not_firstI have test code to ensure the validation and repair code is consistent with the specification.
The repository tests passed:
Zettel GitHub. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
Intellectual discipline can mean many things, but for me it includes resisting the marketing term "Personal Knowledge Management (PKM)" for what is more accurately termed note management.† The term PKM is self-deceptive hype, if you believe it.
An AI may well use the term PKM unless you tell it not to. If you ask an AI to critique a Zettel, it will often fail to distinguish between a note and an essay, and suggest expanding a note until it becomes a self-contained essay. Including matters that properly belong in other notes, assuming they belong in the Zettelkasten at all, defeats the purpose of a Zettelkasten. Perhaps better Zettel critiquing instructions could get around the tendency of AI to zoom past the point of diminishing returns, but lately my Zettel critique GPT has languished for lack of use.
†If you need an acronym for Note Management System, NMS works for me. After all, we're talking about taking notes.
Zettel GitHub. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
the polarization that atlantic piece describes isn't really about whether you use AI, it's about which job you hand it. generation (write the note, summarize, produce the prose) is what atrophies the skill, because it removes the reps that working a zettelkasten is FOR. retrieval and collision do the opposite: surface the note you forgot, flag the two notes that actually contradict, show where a new note bumps into an old one. that hands you more raw material to think about instead of a finished answer to accept.
the local angle upthread is the load-bearing part. the detect-don't-generate use only earns its place if it can run continuously and privately over the whole archive, which basically means local. a cloud model you paste into is already the generation-shaped interaction; a local one quietly watching for contradictions and stale links is the one that keeps the thinking on your side.
so it's less "balance AI with mind training" and more: keep the functions that create reps, refuse the ones that spend them.
To my reading that literally describes balancing AI with mind training.
The distinction is between the "functions that create reps" (mind training) and the refusal of "the ones that spend them" (maintaining balance through intentional selection).
Do you see it differently?
The question is what @Sascha means by "mind training" in his post. Is it a thinking practice that excludes any kind of AI? Is it about balancing using AI with not using AI at all?
Sascha claims that "AI is making us dumber – if we allow it." I found the Atlantic article interesting, because it suggests that AI can make us smarter, if we allow it.
From that perspective it would be more productive to also balance different ways of using AIs: the convenient kind that makes us numb and dumb, and the challenging kind that trains our brains.