Not using an LLM, but rather just a plain set of heuristics which can run formatting quickly enough to render while the text is being streamed or edited. The aim is to try to see if I can make text more readable as it's being streamed in.
Uncaught TypeError: The specifier “semfont/analyze” was a bare specifier, but was not remapped to anything. Relative module specifiers must start with “./”, “../” or “/”.
This is a really neat idea. I'm not sure it's actually useful, but it's definitely neat.
And it made me think of the new kid in town: Jev. I wonder if Jev can do this well and quickly? Not that I'm willing to pay the tokens even to find out, but it's another neat thought.
Nifty idea, but the premise seems flawed. This won't help people read, it will help people avoid reading by focusing on which words to skim. It also seems like it would have a problem detecting sarcasm.
The idea is amazing, and… how to say it?; The results are horrible to read in the same way as AI writing when AI writing is bad.†
And I mean this in the best possible way. It's really interesting.
There's this feeling of an intrusive forcing of something. A feeling that things don't… breathe? Feels like one already sifts out this sort of semantic weight when reading and to me the highlighting adds strange janky emotional-processing noise to the text; Uncomfortable and somewhat… dissonant?
It's really, really interesting. I hate how disparaging to the project this comes out!! my apologies for that! I LOVE the project and what it reveals about text comprehension and perception of saliency and rhythm in written text!
Thanks this is such a thoughtful comment!
This project was definitely more to explore the idea than to make it certain that this was the answer.
I think my goal was more so from the perspective of if text is streaming by very fast like it would be for an agent is there anything we can do to make it easier to read.
The answer may genuinely be no but I think if we can get to something where the program highlights things accurately then in theory the reader could skim faster. I remember debaters who have to speak/read very fast highlight and format the sources (https://youtu.be/0FPsEwWT6K0?t=121&is=7fFn3BjKUnJKPXu3) so in theory it may be helpful for high speed cases.
But I definitely agree with your point that it would be annoying if for example a book was formatted in this way haha
Cool, but I hope it doesn't catch on. Otherwise I can already hear managers complaining to people that they have too many red words and not enough green words in their emails/Slack messages/whatever...
This is brilliant and would likely be the most helpful for notes my agent takes for me or my own recorded transcripts. Especially being able to tweak it. I think where I would not want this is when it's done on my writing that's public facing in case it imbues unintentional meaning. But for my own notes, it's a small cost
Seems to require Chrome? In Firefox it's just a text box, nothing happens. Was confused at first about what it was supposed to do, the rest of the page made it clear it was supposed to mess about with whatever text you write.
Trying it in Chrome and seeing the results I'm not sure that it is really that helpful. Has undertones of elementary school grammar lessons highlighting words in sentences.
This implementation is tune-able. One idea that comes to mind is setting up e.g. salience scores for different company-internal lingo and names of internal components and teams, enriching wiki / chats / tickets / logs this way for quicker and more accurate skimming.
Wonderful. When talking about how much of a mistake the proliferation of skittle highlighting in code is, one of my go-to points is that nobody is stupid enough to enskittle prose because identifying syntactic and semantic meaning from monochrome text is a skill we already train our entire lives. Great to see a living example of how this makes everything 10x harder to read!
Something about this irritated me so much that I even felt compelled to comment. If this ever became a mainstream idea, I would just stop using computers altogether. I am smart enough to mark my own emphasis internally, that is why I have a brain, and one of the advantages of using it.
Oh good point I usually list posts on my homepage but for this post I still have it marked as a draft in the repo I use to build the site so its not there I’ll add it.
Maybe I also should make the posts page show both drafts and non drafts so its less confusing. If you cant tell this has been my first week of writing stuff on the blog haha.
other than the fact that the headline is misleading (this isn't what anybody would think of as a font), the concept misses the point of emphasis.
adding emphasis to words in unformatted text draws the reader's attention in ways that are not implied by the text itself.
you cannot figure out which word to highlight with the raw text "I didn't say he stole the money", but the semantics are completely different depending on if/where they exist.
you already have a mental model of surprise. adding visual noise to destroy what is already part of reading just makes things bad. the reason why syntax highlighting is successful is that the semantics are rigid. natural language is explicitly not that.
Slightly different idea, but I think having a very lightweight LLM doing debounced checks to auto correct grammar and reformat (even rephrase) text would be really cool.
This reminds me. I've always been confused by how comic book writers/artists choose when to use which words to put in bold. It's often not what I would have expected and sounds funny if read aloud that way.
This is not a font in any way of the word, this is a typesetter. It would be a good idea for the author to change the title, because right now it's not just clickbaiting, it's flat out lying.
And on an unrelated website note: don't have a mystery toggle. If you want to offer light and dark modes: awesome, but actually say what the toggle's for either through iconography or by using text.
Not using an LLM, but rather just a plain set of heuristics which can run formatting quickly enough to render while the text is being streamed or edited. The aim is to try to see if I can make text more readable as it's being streamed in.
Not working in Firefox. I get this error:
Works in Chrome though.
Here was me thinking it just didn't do anything at all.
yup, just broken for me.
Thanks for calling that out! Someone from here has fixed it: https://github.com/RohanAdwankar/RohanAdwankar.github.io/pul... Thanks Jakub!
This is a really neat idea. I'm not sure it's actually useful, but it's definitely neat.
And it made me think of the new kid in town: Jev. I wonder if Jev can do this well and quickly? Not that I'm willing to pay the tokens even to find out, but it's another neat thought.
Nifty idea, but the premise seems flawed. This won't help people read, it will help people avoid reading by focusing on which words to skim. It also seems like it would have a problem detecting sarcasm.
very very cool
cute. sadly only in english so far.
The idea is amazing, and… how to say it?; The results are horrible to read in the same way as AI writing when AI writing is bad.†
And I mean this in the best possible way. It's really interesting.
There's this feeling of an intrusive forcing of something. A feeling that things don't… breathe? Feels like one already sifts out this sort of semantic weight when reading and to me the highlighting adds strange janky emotional-processing noise to the text; Uncomfortable and somewhat… dissonant?
It's really, really interesting. I hate how disparaging to the project this comes out!! my apologies for that! I LOVE the project and what it reveals about text comprehension and perception of saliency and rhythm in written text!
(† footnote: feels like this to me)
i completely agree with this. it's a very cool idea and also i hope nobody uses it in actual writing
Thanks this is such a thoughtful comment! This project was definitely more to explore the idea than to make it certain that this was the answer. I think my goal was more so from the perspective of if text is streaming by very fast like it would be for an agent is there anything we can do to make it easier to read. The answer may genuinely be no but I think if we can get to something where the program highlights things accurately then in theory the reader could skim faster. I remember debaters who have to speak/read very fast highlight and format the sources (https://youtu.be/0FPsEwWT6K0?t=121&is=7fFn3BjKUnJKPXu3) so in theory it may be helpful for high speed cases. But I definitely agree with your point that it would be annoying if for example a book was formatted in this way haha
i thought the computation was within the font itself
like
https://github.com/RoelN/SansBullshitSans https://typographica.org/typeface-reviews/chartwell/ https://github.com/trishume/numderline
which can be taken rather far
https://www.coderelay.io/fontemon.html
woah i didnt know this was a thing thanks for sharing!!
Cool, but I hope it doesn't catch on. Otherwise I can already hear managers complaining to people that they have too many red words and not enough green words in their emails/Slack messages/whatever...
This is brilliant and would likely be the most helpful for notes my agent takes for me or my own recorded transcripts. Especially being able to tweak it. I think where I would not want this is when it's done on my writing that's public facing in case it imbues unintentional meaning. But for my own notes, it's a small cost
Seems to require Chrome? In Firefox it's just a text box, nothing happens. Was confused at first about what it was supposed to do, the rest of the page made it clear it was supposed to mess about with whatever text you write.
Trying it in Chrome and seeing the results I'm not sure that it is really that helpful. Has undertones of elementary school grammar lessons highlighting words in sentences.
A spyfont?
A weird way to deal with the bloviated nature of AI generated text.
For everyone saying it must only support Chrome... it works fine for me in Safari.
IMO this is a big idea - whether this particular implementation is good or not is not that important.
This implementation is tune-able. One idea that comes to mind is setting up e.g. salience scores for different company-internal lingo and names of internal components and teams, enriching wiki / chats / tickets / logs this way for quicker and more accurate skimming.
That is a brilliant idea. like syntax highlighting for natural languages
Doesn't work on Firefox.
TLDR: not a font just a JS library
Well, at least we're not yet at "a font that reports back to its vendor what you read".
great idea!
I assume it works perfectly in Chrome, the only browser in the world
Wonderful. When talking about how much of a mistake the proliferation of skittle highlighting in code is, one of my go-to points is that nobody is stupid enough to enskittle prose because identifying syntactic and semantic meaning from monochrome text is a skill we already train our entire lives. Great to see a living example of how this makes everything 10x harder to read!
I've long thought that this would be an interesting way to apply syntax highlighting in PDFs typeset by (La)TeX.
This isn't a font, it's just a JS lib that colors text based on semantic analysis. Submission title oversells it by a lot.
NLP font is not real, it can't hurt you
NLP font:
Something about this irritated me so much that I even felt compelled to comment. If this ever became a mainstream idea, I would just stop using computers altogether. I am smart enough to mark my own emphasis internally, that is why I have a brain, and one of the advantages of using it.
- 404 on posts page? https://rohanadwankar.github.io/posts
Oh good point I usually list posts on my homepage but for this post I still have it marked as a draft in the repo I use to build the site so its not there I’ll add it. Maybe I also should make the posts page show both drafts and non drafts so its less confusing. If you cant tell this has been my first week of writing stuff on the blog haha.
other than the fact that the headline is misleading (this isn't what anybody would think of as a font), the concept misses the point of emphasis.
adding emphasis to words in unformatted text draws the reader's attention in ways that are not implied by the text itself.
you cannot figure out which word to highlight with the raw text "I didn't say he stole the money", but the semantics are completely different depending on if/where they exist.
you already have a mental model of surprise. adding visual noise to destroy what is already part of reading just makes things bad. the reason why syntax highlighting is successful is that the semantics are rigid. natural language is explicitly not that.
The title is misleading, but other than that the results were terrible.
Slightly different idea, but I think having a very lightweight LLM doing debounced checks to auto correct grammar and reformat (even rephrase) text would be really cool.
This reminds me. I've always been confused by how comic book writers/artists choose when to use which words to put in bold. It's often not what I would have expected and sounds funny if read aloud that way.
This is not a font in any way of the word, this is a typesetter. It would be a good idea for the author to change the title, because right now it's not just clickbaiting, it's flat out lying.
And on an unrelated website note: don't have a mystery toggle. If you want to offer light and dark modes: awesome, but actually say what the toggle's for either through iconography or by using text.