Want to point out that these one to two sentence summaries are not abstracts. They lack the structure and content of actual abstracts. Also how can they be community-written if they are written by AI (edit: disregard this, I'm an idiot)? Also also, this should be Show HN right?
Which is really what makes them useful. Abstracts are great at what they try to be, but the typical abstract requires a lot of attention to actually read. These summaries lend themselves to quickly scanning over a list of papers to figure out which ones you want to read the abstract of. They are somewhere between summaries and better titles.
“The Net’s interactivity gives us powerful new tools for finding information, expressing ourselves, and conversing with others. It also turns us into lab rats constantly pressing levers to get tiny pellets of social or intellectual nourishment.”
― Nicholas G. Carr, The Shallows: What the Internet is Doing to Our Brains
If you're a fan of tldr-ai, you might also like my site, EmergentMind.com, which does something similar: it surfaces trending AI papers based on social media engagement (including HackerNews upvotes!), then summarizes those papers using GPT-4 (a bullet point summary + detailed writeup based on the actual content of the paper), and highlights discussions on HN, Reddit, YouTube, GitHub, and X about that paper.
I don't want to highjack this launch post (we definitely need more tools in this space!), just wanted to share my tool for anyone interested since it's related. Feedback welcome: matt@emergentmind.com.
I'm confused. It's supposed to be human-written abstracts of papers relating to "AI"? Or AI written abstracts for science papers in general?
The examples seem to all be machine learning related, but tags include "artificial intelligence" and "machine learning" - if all papers are on those topics are these tags needed?
It started as a TL;DRs platform for research papers on AI, ideated by Francois Fleuret on X, but when I started coding this (and of course already bought a domain :p), I realized that there's no reason to limit the content to AI papers. I'm going to add filters for each subject, to make it more useful for a different subjects
- Most of the abstracts are written by the author of the paper, so might not be as unbiased as an actual "community-written" abstract.
- There's no stated guidelines for the "community-written" abstract e.g. should it be less biased than the original abstract, should be shorter than the original, should it be more accessible to a less AI crowd or all of the above.
- There's no way to upvote/downvote some abstracts e.g. the "attention is all you need" paper has two abstracts and one of them is clearly worse than the other.
Hello! You are on point, building community to write the abstracts, while also setting up relevant guidelines are my main focus for now.
I will think of how to communicate guidelines and expectations to the content in a clear way. Thank you!
Upvote/downvote is available for logged users. As you pointed out, it's not visible when you're not logged in yet, so it would make sense to show score and buttons to anonymous users as well
Can we change the title to something like "Show HN: 'TL;DR AI', community-written abstracts for research papers"?
The name of the project, "TL;DR AI", seems to be causing grammatical confusion. It's easy to read the current/original title as "TL;DR, AI-generated abstracts for research papers". I feel lots of comments are responding with that understanding.
It seem to be a go-to choice platform for communities, so let's try it. You can request features and changes and I post updates about ongoing development
The problem with this kind of tool is the shallowness of knowledge distillation. Content on the internet should enlighten us, not feed us with 140-character TL;DR abstracts, either from humans or AI-generated. This is a tool for exacerbating FOMO, not an actual tool for learning. What I want to see is a tool that teaches me something like GeoHot's 8-hour streams or Karpathy's tinygrad video.
I think it’s an interesting idea. It could use a documentation page with some clear examples and some guidelines. Don’t use undefined acronyms, try to make it self contained, etc.
For example, the tldr(s) “attention is all you need” don’t actually says what attention is or does, which makes the utility of the tldr sort of limited if you don’t already know what attention is. That might be fine depending on the audience for the tldr, but for a general audience IMO it would be better to write more accessibly, or maybe hyperlink to Wikipedia or add footnote definitions for well-known concepts to balance conciseness and discoverability
Thinking about the hyperlink thing more, I think I would revise my suggestion to avoid undefined acronyms - I think you could explicitly encourage them, but only if they are tagged and defined and linked to the originating paper so that experts can skim and newcomers to the field can get context with an expandable html footnote or something.
That way you don’t have a zillion tldrs trying to explain what NeRFs are and do, they all just tag NeRF and link to the canonical tldr for NeRF
That would make sense. I think how to approach it, so we can have collectively built definitions of concept like NeRF. Maybe we can treat them in a similar way to TL;DRs - everyone can write their own version and we can vote on the best one?
Another approach could be some collective improving the definition, similarly to how Wikipedia contributors work
Yeah that sounds awesome to me. It addresses one of the shortcomings of the conventional abstract style which is that nothing is cross referenced - most journals still disallow citations in the abstract
A feature like this would be great for discovering new topics. Right now my approach is to find an interesting paper, and if I need to dig into underlying concepts to really evaluate whether to read in detail then I skip to the lit review and hope for some good bread crumbs, or maybe jump to a fresh GScholar query to do some depth-first-ish graph traversal
It would be really nice to just surf some linked tldrs with high level descriptions of papers and concepts instead!
Laughed hard on that :D Sorry if the wording is not great, I'm neither a native English speaker nor good at copy-writing. I worry the title can't be changed anymore anyway
If we need AI to summarize research papers, we have lost. To me, this is more evidence that research now is mainly conducted for the sake of producing research, rather than producing knowledge that actually improves life in some way. Even if we look at scientific knowledge merely for the sake of intellectual amusement, this sort of research has gone far past that: extremely specialized papers are really not that interesting except in regard to playing the gmae of research.
And I'm speaking from the perspective of a pure mathematician who has read the most specialized an abstract papers and has published several. Yes, you can "get interested" in them simply because they're part of the social game in academia, but for the vast majority of people, even with the capability to understand them, they are far past the point of stimulating natural curiosity.
This is all rather ridiculous, because we are devoting so much time and resources to do something so meaningless. It would all collapse on itself it it weren't for the incremental and short-term economic advantages brought on by 1% of the research, which is mainly in turn about developing unsustainable ways of life.
Presumably you read the summaries to find promising papers where you will read the abstract (which you then follow up by reading the figures, conclusion, discussion, introduction, then the rest, bailing out once you think this isn't a paper you want to read further).
If at any point you figure out the "community abstract" was misleading you can come back and downvote it.
The only issue is if community abstracts get published that make good papers unappealing to read, or if the majority of votes come from people who don't actually engage with the paper
Here's how I picture this, in an optimistic scenario: Once the number of abstracts grows up, we vote up and down on them, so at the end you get the best abstract for each paper, which is supposed to be informative enough to give you a gist of what authors have learned from their research. It should be enough to let you keep an eye on what's going on in research world in real time and also serve as a filter on what paper is worth reading next
This is a grand vision of course, right now we have 9 abstracts, but hey, it's still a good starting point! :p
Better than human is good enough. Also, you will be limited by the texts themselves, which can oversell or just make wrong claims.
The main advantage I am looking forward to is scaling. As a human there is only so much complex text I can read in 24h. If I can just ask my computer to skim 200 articles and tell me which ones seem worth really digging into that would be awesome.
Yes! After finishing the most urgent features like sorting, filtering and fixing some bugs, I'd like to make it easy to subscribe to a chosen subject to get a relevant papers to your mailbox
When I was writing papers in grad school the abstract was the part I worked on most closely. This is not because of emotional stakes, but because communicating something subtle in few words is hard. IMO, it is the least applicable section of a paper for AI generated content.
Hello, much thanks for that. AI-generated abstracts might help with filling the platform with a content and as you mentioned, it might be less biased and more objective. This is something people on X also discussed below the launch post [0] and I think you might be right that it's something to do in the near future
The holy grail would be generating TL;DRs with a model trained on the best human-written abstracts. Although - there should be a way to hide machine-generated content, for your convenience, since not everyone might be interested in this content
No. The holy grail would be TL;DRs written by the actual author, at the time of writing the actual article. No one can better know what the important points are and how to best highlight them. The work of producing a good abstract is a tiny fraction of tiny fraction of the work involved in producing a good article.
Community-written abstracts also have their merits too. The author's perspective might be limited to their area of expertise and is therefore biased. That perspective can be normalized by the more nuanced interpretations from the community.
I feel like this is basically already what people who promote their papers on twitter are doing. Other than Twitter or LinkedIn or something (I don’t really use either) I’m not sure there’s really a place for researchers to actually post such a summary where it would be useful, unless everyone starts using this site I guess.
Even then, I feel like my time as a researcher would better be spent writing a better abstract, so that this sort of tldr could just be a bolded first or last sentence of a conventional abstract… but I’d be very interested in what other people see as the important or most interesting aspect of my paper
Along these lines, I’ve recently taken to using revtex structured abstracts, which are basically a series of topical tldrs covering context, purpose, methods, results, and conclusions. Maybe I’ll experiment with adding a global tldr at the top
― Nicholas G. Carr, The Shallows: What the Internet is Doing to Our Brains
I don't want to highjack this launch post (we definitely need more tools in this space!), just wanted to share my tool for anyone interested since it's related. Feedback welcome: matt@emergentmind.com.
The examples seem to all be machine learning related, but tags include "artificial intelligence" and "machine learning" - if all papers are on those topics are these tags needed?
It started as a TL;DRs platform for research papers on AI, ideated by Francois Fleuret on X, but when I started coding this (and of course already bought a domain :p), I realized that there's no reason to limit the content to AI papers. I'm going to add filters for each subject, to make it more useful for a different subjects
- Looks like there's just 7 abstracts right now.
- Most of the abstracts are written by the author of the paper, so might not be as unbiased as an actual "community-written" abstract.
- There's no stated guidelines for the "community-written" abstract e.g. should it be less biased than the original abstract, should be shorter than the original, should it be more accessible to a less AI crowd or all of the above.
- There's no way to upvote/downvote some abstracts e.g. the "attention is all you need" paper has two abstracts and one of them is clearly worse than the other.
I will think of how to communicate guidelines and expectations to the content in a clear way. Thank you!
Upvote/downvote is available for logged users. As you pointed out, it's not visible when you're not logged in yet, so it would make sense to show score and buttons to anonymous users as well
The name of the project, "TL;DR AI", seems to be causing grammatical confusion. It's easy to read the current/original title as "TL;DR, AI-generated abstracts for research papers". I feel lots of comments are responding with that understanding.
It seem to be a go-to choice platform for communities, so let's try it. You can request features and changes and I post updates about ongoing development
[1] https://github.com/zulip/zulip
1. Each paper and its abstract as a post 2. Up/down vote arrows and a number 3. Comment section
For example, the tldr(s) “attention is all you need” don’t actually says what attention is or does, which makes the utility of the tldr sort of limited if you don’t already know what attention is. That might be fine depending on the audience for the tldr, but for a general audience IMO it would be better to write more accessibly, or maybe hyperlink to Wikipedia or add footnote definitions for well-known concepts to balance conciseness and discoverability
Thinking about the hyperlink thing more, I think I would revise my suggestion to avoid undefined acronyms - I think you could explicitly encourage them, but only if they are tagged and defined and linked to the originating paper so that experts can skim and newcomers to the field can get context with an expandable html footnote or something.
That way you don’t have a zillion tldrs trying to explain what NeRFs are and do, they all just tag NeRF and link to the canonical tldr for NeRF
Another approach could be some collective improving the definition, similarly to how Wikipedia contributors work
What do you think?
A feature like this would be great for discovering new topics. Right now my approach is to find an interesting paper, and if I need to dig into underlying concepts to really evaluate whether to read in detail then I skip to the lit review and hope for some good bread crumbs, or maybe jump to a fresh GScholar query to do some depth-first-ish graph traversal
It would be really nice to just surf some linked tldrs with high level descriptions of papers and concepts instead!
- platform is open to abstracts (TL;DRs) for all subjects of research
- AI-generated TL;DRS will be soon integrated to the platform (initially as a fallback when there is no human-written abstract yet)
- naming is the hardest thing in computer science :D
Thank you so much for the feedback and the insight, I can't wait to provide the next updates on the progress to you
In the meantime, feel invited to join Discord channel https://discord.com/invite/AJ9YCcqD and X https://twitter.com/keep_FOMO_away
And I'm speaking from the perspective of a pure mathematician who has read the most specialized an abstract papers and has published several. Yes, you can "get interested" in them simply because they're part of the social game in academia, but for the vast majority of people, even with the capability to understand them, they are far past the point of stimulating natural curiosity.
This is all rather ridiculous, because we are devoting so much time and resources to do something so meaningless. It would all collapse on itself it it weren't for the incremental and short-term economic advantages brought on by 1% of the research, which is mainly in turn about developing unsustainable ways of life.
- they don't have to add anything that wasn't already expressed in the paper, limiting the need for creativity
- they can use boring and formulaic language because that's how technical writing is done, limiting the need for creativity
- they can give a more neutral overview of the research since they aren't an involved entity with emotional stake
But how could we know they add nothing or missed something important like a not?
>they can give a more neutral overview of the research since they aren't an involved entity with emotional stake
Biased input = biased output.
AI can't help with that.
Probably by reading it.
Sounds like more work not less.
If at any point you figure out the "community abstract" was misleading you can come back and downvote it.
The only issue is if community abstracts get published that make good papers unappealing to read, or if the majority of votes come from people who don't actually engage with the paper
This is a grand vision of course, right now we have 9 abstracts, but hey, it's still a good starting point! :p
The main advantage I am looking forward to is scaling. As a human there is only so much complex text I can read in 24h. If I can just ask my computer to skim 200 articles and tell me which ones seem worth really digging into that would be awesome.
[0] https://twitter.com/jedmaczan/status/1784889300452077796
Author said so, here: https://news.ycombinator.com/item?id=40208511
Even then, I feel like my time as a researcher would better be spent writing a better abstract, so that this sort of tldr could just be a bolded first or last sentence of a conventional abstract… but I’d be very interested in what other people see as the important or most interesting aspect of my paper
Along these lines, I’ve recently taken to using revtex structured abstracts, which are basically a series of topical tldrs covering context, purpose, methods, results, and conclusions. Maybe I’ll experiment with adding a global tldr at the top
https://journals.aps.org/prc/edannounce/PhysRevC.84.030001