Great idea! The scrolling is a bit broken though. I can't scroll down the text on my trackpad without it going to the next article, and also it just a lot of articles with a tiny movement of the trackpad.
Otherwise it's a really fun idea! Can I suggest you also scrape from https://www.medrxiv.org/? This is where a lot of medical research preprints, not arxiv
It looks great. But it's missing one critical feature of "fast-food" apps like TikTok: the content is not easily digestible. Which is understandable, because scientific papers are dense.
Maybe a good idea would be to parse the abstract through an LLM to make it more understandable (maybe caching the results so it's not expensive)? Maybe also using some standard style, like starting with a couple of "dumbed-down" sentences of the article for the non-expert, and progressively explaining better.
Please don't follow the original comments suggestion; I feel that "easily digestible" is not compatible with what makes the idea shine in the first place.
Your suggestion delegating such functionality to a local LLM is quite nice as a choice but adding it as a core functionality is quite antithetical to leverage the arXiv part, without which everything reverts back to a bland and generic whateverTok format.
Although the suggestion seems to be aware of the fact and provides both a good reasoning and a quite good solution (progressively deepening explanations), the implicit information and nuance lost in a summary by an unreliable LLM would undeniably turn this from a useful and interesting idea to a cool party trick no one uses for more than 5 minutes.
Thanks! I'll text you here when I add the feature. It wouldn't be core, so I think that having two modes where you can read easily papers with LLms or not will be of great help.
Funny to think this is actually kind of similar to how ByteDance got their start. One of their first apps is called "Today's headlines", which is basically the TikTok format, but instead of short videos you are scrolling news article.
Not sure why the author didn't say this, but that's _by default_. There's a search in the top right and you can do like `astro-ph` and it'll return papers from it, afaict by skimming the source.
IMO it should show that it's searching cs / ai as prefilled search terms.
This is legitimately useful! How does it track what a user has already viewed? Is it just showing articles from the current day in order? If I get to a day I've viewed previously, will it start showing me articles that I've already viewed once? I've been hesitant to touch my filter search term since I don't want to start over.
More customization to what papers it shows you could be very useful. You could keep track of user likes and employ a recommendation algorithm. Or better yet, the user could provide a custom prompt, which an LLM would use to filter and/or recommend papers.
Yes definitely not. But I think this could be a legitimately good use case if it were implementable. Filtering based on keywords will throw away papers of interest, while a recommendation algorithm wouldn't give you much control over content. But a language model could probably do a decent job of ranking a days worth of papers based on relevance to a short description of your research interests.
I'm in a field where there are 50+ postings a day, but only 5-10% are relevant to my focus. A good filter would save me a lot of tedium.
But OP says it wouldn't be sustainable to implement, so that's that. Maybe will try this myself and see how it goes.
Exactly, I think that in this case using ML models that are experts in papers can be more useful and that type of models can run locally so it's the best option.
I think you need AI to generate a video of a Gen-Z person talking about each, holding a little microphone. The paper could be the green-screened background.
Great idea! The scrolling is a bit broken though. I can't scroll down the text on my trackpad without it going to the next article, and also it just a lot of articles with a tiny movement of the trackpad.
Otherwise it's a really fun idea! Can I suggest you also scrape from https://www.medrxiv.org/? This is where a lot of medical research preprints, not arxiv
I added now the MedRxiv API, you can try it out and tell me if you notice something weird.
I saw that medrxiv has an API so I'll add it too!
About the behaviour of the scroll it's a bit buggy haha but I'll try to fix that, thanks for the feedback.
[dead]
It looks great. But it's missing one critical feature of "fast-food" apps like TikTok: the content is not easily digestible. Which is understandable, because scientific papers are dense.
Maybe a good idea would be to parse the abstract through an LLM to make it more understandable (maybe caching the results so it's not expensive)? Maybe also using some standard style, like starting with a couple of "dumbed-down" sentences of the article for the non-expert, and progressively explaining better.
Mm maybe running a local LLM for doing it would be a great idea, I'll try to do that and if it doesn't work well, I'll consider doing it with an API.
Please don't follow the original comments suggestion; I feel that "easily digestible" is not compatible with what makes the idea shine in the first place. Your suggestion delegating such functionality to a local LLM is quite nice as a choice but adding it as a core functionality is quite antithetical to leverage the arXiv part, without which everything reverts back to a bland and generic whateverTok format.
Although the suggestion seems to be aware of the fact and provides both a good reasoning and a quite good solution (progressively deepening explanations), the implicit information and nuance lost in a summary by an unreliable LLM would undeniably turn this from a useful and interesting idea to a cool party trick no one uses for more than 5 minutes.
Yes I strongly agree with this. I also want to read the original abstracts.
Thanks for the feedback, I think that it would be like having 2 modes with a toggle: unhingered summaries from local LLMs or the real summaries.
Cool :) I am a scientist, so having an easier way to parse the abstracts would be most welcome. Keep up the good work.
Thanks! I'll text you here when I add the feature. It wouldn't be core, so I think that having two modes where you can read easily papers with LLms or not will be of great help.
Funny to think this is actually kind of similar to how ByteDance got their start. One of their first apps is called "Today's headlines", which is basically the TikTok format, but instead of short videos you are scrolling news article.
Wow, I didn't know that.. It's really ironic
Recent and related:
WikiTok - https://news.ycombinator.com/item?id=42936723 - Feb 2025 (216 comments)
Is it filtered to only talk about AI or all of current science interest is about AI ?
Not sure why the author didn't say this, but that's _by default_. There's a search in the top right and you can do like `astro-ph` and it'll return papers from it, afaict by skimming the source.
IMO it should show that it's searching cs / ai as prefilled search terms.
Added!
Noted, I'll add it now
Yeah, I filtered it mandatorily about AI or computer science
BRB going to create XTokTok so I can flip through TikTok for X clones.
Eventually WaylandTok will be ready.
That has potential..
This is legitimately useful! How does it track what a user has already viewed? Is it just showing articles from the current day in order? If I get to a day I've viewed previously, will it start showing me articles that I've already viewed once? I've been hesitant to touch my filter search term since I don't want to start over.
More customization to what papers it shows you could be very useful. You could keep track of user likes and employ a recommendation algorithm. Or better yet, the user could provide a custom prompt, which an LLM would use to filter and/or recommend papers.
Thanks! I don't track anything; it just shows sorted papers from the current date.
Maybe I'll implement a basic recommendation algorithm locally, because right now, an LLM implementation wouldn't be sustainable.
> an LLM implementation wouldn't be sustainable
Also not the right tool for the job. Not everything has to be a LLM.
Yes definitely not. But I think this could be a legitimately good use case if it were implementable. Filtering based on keywords will throw away papers of interest, while a recommendation algorithm wouldn't give you much control over content. But a language model could probably do a decent job of ranking a days worth of papers based on relevance to a short description of your research interests.
I'm in a field where there are 50+ postings a day, but only 5-10% are relevant to my focus. A good filter would save me a lot of tedium.
But OP says it wouldn't be sustainable to implement, so that's that. Maybe will try this myself and see how it goes.
Exactly, I think that in this case using ML models that are experts in papers can be more useful and that type of models can run locally so it's the best option.
I think you need AI to generate a video of a Gen-Z person talking about each, holding a little microphone. The paper could be the green-screened background.
You mean like this?
https://pdftobrainrot.org/
This website has some of the worst mobile UI I’ve ever seen — blasted by ads, and full screen subscription model — before I could even read the pitch.
That might be hilarious.
(I won't know for sure until I can find someone to explain the zoomer lingo.)
Just by imagining it, it would be a lot of fun haha.
Could you please consider adding video generation? In TikTok style, but with content from Arxiv
That would be cool, I'll consider it!
I like this. Now i want a self hosted version i can point at a dozen sources mixed together.
Pretty weird UX to say "no papers found" before suddenly showing me a paper.
I'm rolling a new version up now, it'll be live in 2 minutes.
I'll try to improve that in a moment.
Can you show number of references as though they were number of favorites or likes?
Added! Try it out.
Yeah, I'll add it quickly.
Feels like XTok is imminent: Browser extension that makes any site scrollable
Hahaha surely one is building it right now
Out of curiosity (since you turned this around so fast) -- I saw that the other author mostly used Claude/Cursor, did you do the same?
Yeah, I used Copilot instead of Cursor.
You can add Show HN to the title, too.
It looks good, though when I scroll up once on my trackpad it scrolls past 3-4 articles at a time.
Thanks for the feedback! I'll add Show HN then.
Might as well throw together HackerTok.
I'll add it as a source too.
Add swipe feature next!
I think I have that now covered, you can try it out and tell me if it works if you want.
That would be nice, I'll add it to the roadmap.