As you probably know, Gwern.net’s solution is to use sidenotes, falling back to footnotes.
My current overall approach for writing right now is to:
footnotes/sidenotes are intended primarily for a mix of “f”/”t”: further information relevant to a reader particularly interested in a specific passage or claim, but which the majority of readers would want to skip, and which doesn’t pertain to a specific citation.
This reduces the cognitive load to simply, “am I interested enough in this section that I want to look at the sidenote beyond a near-subconscious saccade to skim keywords/formatting?”
in cases where the ‘further niche information’ is practically a mini-essay, I have been trying to avoid writing those and fixing up old instances. Usually, they should be factored out to a section or page, or much of the material (especially blockquotes) moved into an annotation.
In cases where those are not really relevant, we have ‘collapse’ functionality, where arbitrary sections or ranges can be ‘collapsed’ into a button to expand them, with an optional short description. They’re sort of like an anonymous inline-only footnote, if that makes sense to you. (We use this heavily for authors in popups: instead of either showing the entire list, which might be literally several screens of hundreds of authors, or arbitrarily truncating at n characters, we just show the first few authors, and collapse the rest ‘and so on’.)
specific citations: I hyperlink citations to the fulltext, with an abbreviated citation format to make it much more readable when there’s a lot of them, using subscripts.
Generally, no more bibliographical information is necessary than just ‘Foo et al 2024’, and that reads very nicely when written subscripted as ‘Foo...2024’. All additional information, such as the full title/author, the abstract and key excerpts, commentary etc, can be provided in the popup for immediate reference.
This removes your “c” example, as well as handling a chunk of “f”.
‘d’ definitions: I find definitions either ought to be inlined (where I use a Wikipedia-like convention of bolding the newly-introduced term, followed by its definition, possibly in a parentheses), or have a relevant hyperlink, like the specific citation that introduced that term or just a Wikipedia article. No sidenote/footnote apparatus necessary.
(You might think the link is still obtrusive, but my belief is that the link-icons make it less of a cognitive burden. If you see some technical term followed by the familiar ‘W’ Wikipedia or chi Arxiv link-icon, what else could it be but a definition? And so you can easily ignore it if you feel adequately familiar with the term.)
Your taxonomy of c/d/f/t sounds like it could be somewhat cumbersome (who wants to mark up every single footnote they ever write with at least 4 classes?), but it also could be fully automated, I’d note. Just let the author write footnotes however they want, then LLM-classify them and mark up with different symbols. You could also use the LLM to summarize them (this is something I’ve been experimenting with for margin notes and labeling clusters of annotations).
For example, you could simply number ‘c’, indicating with the cold mechanical enumeration that it’s not important (because just citation); replace ‘d’ with a little dictionary icon postfixed to the term being defined; replace ‘f’ with a 1–3-word summary by the LLM which is grayed out or shrunk to indicate it’s a tangent (with of course a possible manual override by the author if they decide the LLM suggestion is bad); and markup ‘t’ with some ellipsis symbol like ‘...’ to indicate that it’s some very unimportant expansion of the current thing.
This is doable mechanically with 0 author effort (very important for adoption! even if you think you are personally willing to bear the burden of this extra overhead, you really don’t want to and will shy away from writing because of it) because whatever document format you are using will support reliably extracting footnotes from the AST (even TeX ought to) & feed into a LLM & adding some markup class which your HTML/CSS/JS can use, and I think my GUI suggestions should be intuitive: numbers=citations already, the summaries would be self-explanatory, while dictionary icons & ‘...’ are semi-novel but a reader will guess them on the first instance and remember them after an interaction confirms it.
(The idea of a 1-word footnote anchor, using LLMs to do the heavy lifting of picking a key word, is interesting enough I may give it a try myself.)
As you probably know, Gwern.net’s solution is to use sidenotes, falling back to footnotes.
My current overall approach for writing right now is to:
footnotes/sidenotes are intended primarily for a mix of “f”/”t”: further information relevant to a reader particularly interested in a specific passage or claim, but which the majority of readers would want to skip, and which doesn’t pertain to a specific citation.
This reduces the cognitive load to simply, “am I interested enough in this section that I want to look at the sidenote beyond a near-subconscious saccade to skim keywords/formatting?”
in cases where the ‘further niche information’ is practically a mini-essay, I have been trying to avoid writing those and fixing up old instances. Usually, they should be factored out to a section or page, or much of the material (especially blockquotes) moved into an annotation.
In cases where those are not really relevant, we have ‘collapse’ functionality, where arbitrary sections or ranges can be ‘collapsed’ into a button to expand them, with an optional short description. They’re sort of like an anonymous inline-only footnote, if that makes sense to you. (We use this heavily for authors in popups: instead of either showing the entire list, which might be literally several screens of hundreds of authors, or arbitrarily truncating at n characters, we just show the first few authors, and collapse the rest ‘and so on’.)
specific citations: I hyperlink citations to the fulltext, with an abbreviated citation format to make it much more readable when there’s a lot of them, using subscripts.
Generally, no more bibliographical information is necessary than just ‘Foo et al 2024’, and that reads very nicely when written subscripted as ‘Foo...2024’. All additional information, such as the full title/author, the abstract and key excerpts, commentary etc, can be provided in the popup for immediate reference.
This removes your “c” example, as well as handling a chunk of “f”.
‘d’ definitions: I find definitions either ought to be inlined (where I use a Wikipedia-like convention of bolding the newly-introduced term, followed by its definition, possibly in a parentheses), or have a relevant hyperlink, like the specific citation that introduced that term or just a Wikipedia article. No sidenote/footnote apparatus necessary.
(You might think the link is still obtrusive, but my belief is that the link-icons make it less of a cognitive burden. If you see some technical term followed by the familiar ‘W’ Wikipedia or chi Arxiv link-icon, what else could it be but a definition? And so you can easily ignore it if you feel adequately familiar with the term.)
(We also have a few relevant features, like popups which highlight ranges, added for the detailed line-by-line analysis of “Suzanne Delage”, and within-section display of backlinks and backlinks context snippets, which are relevant here.)
Your taxonomy of c/d/f/t sounds like it could be somewhat cumbersome (who wants to mark up every single footnote they ever write with at least 4 classes?), but it also could be fully automated, I’d note. Just let the author write footnotes however they want, then LLM-classify them and mark up with different symbols. You could also use the LLM to summarize them (this is something I’ve been experimenting with for margin notes and labeling clusters of annotations).
For example, you could simply number ‘c’, indicating with the cold mechanical enumeration that it’s not important (because just citation); replace ‘d’ with a little dictionary icon postfixed to the term being defined; replace ‘f’ with a 1–3-word summary by the LLM which is grayed out or shrunk to indicate it’s a tangent (with of course a possible manual override by the author if they decide the LLM suggestion is bad); and markup ‘t’ with some ellipsis symbol like ‘...’ to indicate that it’s some very unimportant expansion of the current thing.
This is doable mechanically with 0 author effort (very important for adoption! even if you think you are personally willing to bear the burden of this extra overhead, you really don’t want to and will shy away from writing because of it) because whatever document format you are using will support reliably extracting footnotes from the AST (even TeX ought to) & feed into a LLM & adding some markup class which your HTML/CSS/JS can use, and I think my GUI suggestions should be intuitive: numbers=citations already, the summaries would be self-explanatory, while dictionary icons & ‘...’ are semi-novel but a reader will guess them on the first instance and remember them after an interaction confirms it.
(The idea of a 1-word footnote anchor, using LLMs to do the heavy lifting of picking a key word, is interesting enough I may give it a try myself.)