Note to the reader

This is simultaneously a critique of pure text as a medium for content, and a experiment to fix it. The key idea is that if we combine pure text with a parallel means of apprehending and navigating it (experimentally instantiated with the graphs on the right), the reader is empowered further. There is a request for feedback at the end of this post.

A region of text is selected by hovering over it, or by scrolling. The top graph shows nearby logical connections for the local region of text. The bottom graph shows how the local region connects to the main point of the post. Grey nodes are nodes which are not explicitly stated in the text.

Text is a channel for rhetoric, let’s add one for logic

Text should not be left alone

Many things-worth-knowing are put into text instead of conversations and experience.

This is unfortunate, because getting them back out of text is hard

Here are the core problems with pure text, as a medium for communication:

[rhetoric-logic confusion]
Evolution put rhetoric and logic in the same room of the mind
.
This has the side-effect of making us extremely vulnerable to believing the wrong things
.
Bad logic becomes persuasive through good rhetoric
.
Good logic stays undiscovered through bad rhetoric
.
The linearity of text makes this worse
,
by forefronting rhetoric (which is increased in magnitude with the building of every subsequent sentence)
at the expense of
backgrounding logic (whose structure is hidden).
[apprehending structure]
The structure of an argument is important for analyzing it.

It is difficult to immediately see that structure
, though,
because text is linear - one statement follows another
.
Imagine two adjacent sentences that are actually structurally parallel, and together stand as premises which hold up the conclusion of a third sentence. Let us concretize this with a familiar example: “Socrates is a man. All men are mortal. Socrates is mortal.”

Nothing about the form of text gives the first and second sentences the clothes of premises
,
nor the third sentence the mark of conclusion
.
Nothing about the form of text even makes the link from the first sentence to the third
,
nor the second to the third.

At most, one could argue that a 2 —> 3 link is implied by adjacency
.
This is true, but a 1 —> 2 link is also implied and is wrong. So link-implication-by-adjacency is not reliable.
We need better, especially for more complex argumentation.
[exploration]
The author's imposed order on text is hard to break
.
When I read an interesting point, I might want to see how it relates to another particular point
.
Perhaps I read about how London taxi drivers developed larger hippocampi to accommodate the spatial knowledge of car-traversable London, and want to know whether this big brain gain came at the expense of some other cognitive abilities
.
I will be at the text's mercy
.
My desire is only reliably fulfilled if that relation just so happens to be the next sentence
.
Else I need to keep reading, and hope that I will not forget my query (and hope that the query will be eventually answered).
[high-level understanding]
It’s difficult to get a sense of the key points of a text
.
When reading scientific papers, common advice is to jump to the abstract, then conclusion, and then check any interesting figures before reading the paper in its entirety
.
In speed reading-esque circles, common advice is to read the first and then last sentences of each paragraph on a first pass
.
These are both abuses on the form of text, which is inherently linear (read it start-to-finish) thus non-interactive (no jumping around!)
.
Yet by treating the text as such, you take up the reins of control over your interaction
. You decide where to go, and thus your understanding is now your active responsibility instead of being delegated solely to the author.
But with the current incarnation, text does not like this use, and so the form of text itself struggles against you.
[search]
It’s hard to find the exact information wanted from text
.
For example, you may know that David Hume, in his Enquiry Concerning Human Understanding, outlines base rules that our minds use to combine pre-known concepts into new ones. Based on this, imagine that you want to know exactly what these base rules are.
The text will give you a hard time in your search.
You can’t use a quick visual inspection - the table of contents isn’t helpful here, and there are too many pages to flip through
.
You can’t even do a Find on your PDF viewer - the particular terms he uses to express this idea are not the same as the ones I have used

Side note about [search]: Of all the flaws of text mentioned, this is the closest to being solved.

We now have vector databases that allow text to be semantically searched (search-by-meaning) instead of syntactically searched (search-by-exact-text). The underlying mechanism is to convert your search query to an eerie mathematically-represented understanding of the meaning of the query, and compare its numeric distance to similarly-represented chunks that encompass the entire underlying text.

It’s not mainstream yet, but I suspect that books will start including search tools to increase their interactivity soon

There’s probably profitable startup idea there - an LLM that unlocks knowledge bases when the book purchase is verified

Personally, I find this monetization to be disgusting as it is a movement towards barriers to knowledge.

This post details an experimental attempt to fix many of the problems of text, while also demonstrating the experiment. The tech is there. It’s been there for a while. But the idea and its adoption do not yet exist.

On the utility of rhetoric

At its core, all of language is just gesticulating vaguely at concepts and things which are too big for us to hold losslessly in our tiny brains.

The trick to good communication is thus to make pretty signposts and point them in the roughly-correct direction for the particular traveller you’re trying to orient.

Seen from this lens, rhetoric is a wonderful pathogen with which to infect readers with understanding.

[memory] A well-turned phrase makes things memorable. Personally, it’s like the fog I’ve been wading on suddenly settles into solid ground. There is joy and relief, and a deep (though often erroneous or illusory) feeling of knowing.
[non-conscious] Rhetoric speaks to the secret parts of the self in the hidden society of the mind. Perhaps more important than speaking to conscious-me, rhetoric moves the core actuators of my being. There is a difference between knowing something in the brain, knowing something in the heart, and knowing something in the bones. Rhetoric is the unifying politician that stirs them all. This is good, but unfortunately it also blurs their boundaries. A thing which should only be heart-known could seep into the brain and bones.

Understanding, however, is not the same as truth-seeking. You can certainly understand things that are false, and there is often value in doing so. The problem comes from believing that false things are true.

Like a bad roommate, rhetoric gets logic drunk in the text and then leaves it alone to find its own way home.

Thus our solution should preserve the life-of-the-partyness of rhetoric, while also keeping logic sober.

The solution: A logic-forward channel to complement text’s rhetoric-forward channel

The solution is simple:

let rhetoric have its stage in text

Meanwhile, we assemble a parallel arena for logic to perform, and place it visibly at text’s side

These are different views on the same content, but they are synced.

The reader may flit between them as they see fit.

The writer may use the logic-forward channel to extend communication beyond text

by making the implicit into the explicit

, and

drawing links that are not easily seen through linear presentation.

Rhetoric is simple and requires little to perform. Logic, despite being the quiet one, is more of a diva. Let us investigate her needs.

The building blocks of logic’s stage

To show logic is to show structure. Thus we require knowledge of what is allowable in the structure, and a way of marking up statements to fit into that structure.

The particulars of what is allowable in the structure is a work-in-progress. I welcome any ideas for improvement. For the work you are currently reading, the basics of the structure are as follows:

Nodes, which are boxes that each represent a particular point which the content makes.
Directed links, which are arrows that relate nodes to each other. These are the workhorse of the structure.

Links have different types. The types currently used are as follows:

claim

analysis

artefact

constraint

part

example

extension

context

Thus if Node A points to Node B with the link type of claim, then we know that A is a premise which supports the conclusion B.

Marking the Main Conclusion

I have marked the main conclusion of this essay. This serves as a focal point for all other nodes.

The advantage is that it is easy to show, from any point of the text, how this links to the overarching point-of-interest (the main conclusion).

The disadvantage is that, in fact, there may be multiple overarching points-of-interest.

A future extension should thus generalize the concept of the “main conclusion” to multiple overarching points-of-interest (main conclusions).

Levels of detail through keypoints

For years, video games have been cheating computational demands by using lower-resolution (computationally cheaper) elements for things that are far away.

The player won’t notice or care yet, anyway - these far things are not in their immediate vicinity.

Similarly, this text uses keypoints in its High Level view in order to show how the point leads to the main conclusion.

There are interesting extensions here:

[recursive node groups] Should we allow particular nodes to be grouped together, and for these groups to themselves form higher-order node groups? This would allow for multiple levels of detail, but would cost complexity on all fronts (UI, code, the author's markup time).

[multiclassing] Should a node be allowed to have >1 node group? How would this affect the representation to the user?

[redundancy] Should multiple parts of the text be allowed to be associated with the same node? Written text tends to have a good degree of repetition.

Note: in practice, levels of detail techniques in video games often break down in jarring ways, e.g. distracting the player by having more details suddenly pop in.

But these are, I believe, implementation issues or hardware limitations, rather than problems with the metaphor.

Some insights from marking up the logic in this post

The process was painful, but this is likely fixable

Here were the steps:

Write the post on Notion
Export to markdown, generate HTML from the markdown
Specify the graph details in the JavaScript, and couple the HTML to particular nodes.
Discover weak aspects of the post and structure
Edit and repeat from step 3

The visualization code was a one-time cost that is usable for any future posts, so I won’t discuss it.

Feel free to check out the code on GitHub, but be warned that it’s hacky.

The pain comes from the process of marking up the text to extract the logic graph.

This is fixable - I could create a custom GUI for making the graph.

The process helped me see my own thinking more clearly

Writing, of course, always clarifies thought.

But that clarification often comes with an illusory sense of certainty, that when tested, falls apart.

Writing too often helps me believe what I think, without ensuring that I know it.

In this way, the logical layer was a stress-test for those beliefs. This is the second post I have written in this format. Like the first, I immediately saw weaknesses in my own argumentation and thought.

Thus the act of logically marking-up my beliefs was a nice complement to writing them down.

The logical markup is still subjective

Prior to starting, I believed that the outcome would be formal and unambiguous. This is not the case. Even with the framework fixed, my subjectivity rests within the following choices:

[node text choice] At first, I believed that nodes should be fully self-contained and unambiguous. However, this made nodes too long, which limited their interface utility. Node 41, for example, was originally "Showing the path from the current point in content to the main conclusion makes text more understandable", which was too much. It is now "Showing path to forcal point increases understandability". Additionally, the drive to disambiguate led to too much local redundancy (by repeating context). In the end, I removed much repeated context. When the reader is in its local neighborhood, the context is clear. Much like making a geographical map, while we want the map of the text's logic to be as detailed as possible, we still want it to metaphorically fit in the navigator's pocket.
[link type choice] The problem here is multiple competing alternatives. For instance, is Node 61 ("Logically marking up this text helped clarify my thoughts") a claim for Node 54 ("Some insights from marking up the logic in this post"), an artefact or an analysis? All of these choices are fine, and so my decision to actualize one over the other adds a flavor to the logical structure.

We may see this subjectivity as an anologue to rhetoric. Neither channel is pure logic, nor pure rhetoric. But each channel deals the proportions differently.

The detail in the logical markup can be varied as I see fit

At points where I feel that things require less explanation, or points where I simply feel like marking up the logic would be too taxing, I can choose to use a single node instead of decomposing the details into a larger structure. This is compelling, because as the shepherd of meaning, I know that I do not need to give equal care to each of my flock's members. This paragraph is itself an example of variation (only one node represents it in its entierity).

Much of the logic was implicit

The connections and conclusions I want the reader to draw are often not associated with any part of the text at all. They are implied, and this often surprised me as I thought I was being clear.

I have marked these nodes in light gray.

As an example, the main point of [rhetoric-logic confusion] was that "Text obfuscates logic through rhetoric" (Node 912). This is fairly clear from reading the text, yet it is never stated explicitly.

Additional Thoughts

Structural comments are a promising idea

Comment threads are already a fantastic system for interaction.

Adding the logical channel enables even more specificity when discussing content.

This is an extension for the future, but one that is easy to imagine.

It would be far more fruitful to point with specificity to the part of the content or to the part of the relationship between two nodes when commenting on it.

This would aid comment writers in honing their thoughts

, and

readers in apprehending those thoughts

One can also imagine that this specifity unlocks a new way of discovering comments

similar to how YouTube might replay comments made at a particular timestamp

a user might see associated comments when they are at the relevant node.

Why not automatically extract the logical graph from text?

It would be obviously-useful to have any text that you are reading represented in the format that I am proposing.

At worst, it can't hurt (just ignore the logic channel).

I believe that AI is advanced enough to generate knowledge maps of a given text.

In fact, it might even do a better and faster job than I can in most cases.

This is highly useful for other-people’s-writing, and I might build something for it.

However, I am currently against walking the route of automatically extracting knowledge maps for my own writing. Here’s why:

There is something special about me marking up my own text.

I believe this has two aspects:

[better self-critique] It deepens my own understanding of my text. I see flaws and blemishes that I otherwise would have missed.
[more interpretation control] It allows me to collapse the ambiguity of my text into what I want.

The logical channel provides more firewood for the interpretation engine in the reader’s mind. This firewood decreases ambiguity, and the reader will better arrive at what I wish to deliver.

In order for an automated process to extract my intent from textual ambiguities, it has to know me extremely well.

I do not trust any automated process to know me well enough
, particularly since
I am ever-evolving (and thus a shifting target for it to reach)
and
have not yet given any process enough of my data to derive me from.

Problems and Limitations

Problems that I believe are solvable

[interface jankiness] The interface has issues.

Sometimes, I accidentally hover over text that I don’t mean to, and the whole graph changes.

Also, transitions between graphs are not fluid.

Also, I’m not sure if these are... I could go on.

These are UI issues which, while decently interesting, are ultimately solvable.

I will not address solutions in this post, but if you have thoughts, please do share them.

Problems that may not be solvable

[appeal] This structure appeals to me.

It’s similar to how I want my brain to work when reading text.

But this appeal might not generalize to other people
-
people are strange and have different processes for arriving at what seems like the same end result.

We even see this in the procedure for counting numbers - the end result is the same across people, but the process is different.

Here is Richard Feynman describing how he can’t count and speak at the same time but can count and read, whereas a friend of his, who clearly has a different internal process for counting, is the opposite.
[adoption] The graph is new technology, and even if better, new technology has no guarantee of adoption.

A close, intelligent friend who read an earlier draft of this idea ignored the logical channel entirely, even when specifically asked to read the post to critique the logical channel.

If this is too difficult to adopt, it would suck.

Given that this works well for me, I’d like other people to benefit from using it too.

But the most important aim of this post is to build better technology for myself, so it is still a useful attempt.

Limitations

[scope] It’s a fool who earnestly tries to build a knowledge map out of a poem.

This supplement to text is not meant for every text.

Some of the most perspective-shaping text I have read has had beauty as its central point, instead of sense.

These texts gifted me with psychological imagery and feelings that became part of the inner mental language through which I grapple with the world.

Some text is a puzzle

the zen koan is ambiguous on purpose because it is a journey meant to transform you along the way.

To increase clarity at the outset is to rob you of the procedure for self-betterment.

Elucidating the logical structure of such texts is not only useless, but detracts from the purpose they serve.

End

This is part of a grander project I have, to create modern tools to increase my agency in making sense of the world.

Technology is evolving faster than useful information is compounding

, and

the previously untraversable terrain may now be tamed

Fundamentally, I want to live in good company.

And good company is friends with both rhetoric and logic.

Additional: Some nice related things

Ideas

Links to internal content
, such as
Wikipedia
, or
Obsidian
(e.g. see
Andy Matuschak’s notes
or
this wonderful resource on mental health
)
Semantic Web - a vision to make the web more computer-understandable on a more granular level that was expressed by some web developer
.
This would unlock better representation of the information of the internet, possibly even in the way that I’m trying to create for myself.

It’s a cool idea, but one that never really took off
, and is mostly made obsolete by LLMs anyway.

Readings

Plato’s Gorgias
. A very enjoyable Plato dialogue, where Socrates gets tense while
debating rhetoricians who claim that rhetoric is the most important skill of all
(e.g. what use is it to be a doctor, when no one believes that your cure will solve their ail?).
Leibniz's The True Method
.
A brief essay where a father of calculus outlines his vision for a method of expression that would lead to universal agreement. Disagreements would thus be solved with a simple procedure: "let us calculate, sir". Of course, this vision is as-of-yet unachieved.

Meta: Request for Feedback

The feedback I'm looking for is as follows:

[core idea] Do you think the core idea of some logical channel for info display is promising, or do you think it's not worthwhile?

[current utility] Does the current logical channel help you understand the text better?

[reader interface] Assuming we have the logical markup of a text, how can we use that to help readers grasp the parts of the text they're interested in faster? How should we display the information?

[author interface] What would an ideal author interface look like, for marking up text while writing it?

Please do send me any comments you might have. Even high-level impressions from incomplete readings are valuable.