Hey - I've been playing with LLMs since GPT-2 and recently experimented with fully generative UIs where the HTML/Canvas are generated just-in-time.
Every post on the feed( on slop/duck/storytime) you see is streamed and generated just-in-time with HTML and into a Canvas with Gemini 3 Flash.
Comments and DMs are bidirectionally linked with a Cloudflare Workers Durable Object which is why they feel so fast. Every generated post is saved into a DO SQLite which is then served into the "Following" feed so it can be served quicker.
This was inspired by Wikitok, a VSCode Extension I made around brainrot, and another fully generative UI site I made.
Next step is allow viewing TikTok like wikipedia. Take a bunch of popular tiktok posts, use an LLM to describe what's happening via text/screenshots (with references to memes, etc), and link them to other relevant tiktoks.
Some meme sites might be somewhat similar to this.
I thought the scroll didn’t work at first, apparently only the bottom of the screen is scrollable. I see other comments talking about a voice, I don’t have it, just the music? Also the content is really irrelevant and uninteresting (I’m assuming it uses a random article? It should aim for the most popular instead), it really needs some curation.
The subtitles are also hard to read, they should be displayed sentence by sentence, not word by word.
IMO you didn’t capture what makes some interesting content work on TikTok. The scientific content on TikTok doesn’t have the brainrot music, it’s just interesting because of the short format. It looks to me like a clone of TikTok made by someone who hates TikTok and built it out of assumptions on what TikTok is.
Honestly interesting Wikipedia articles summarized as TikTok content is a killer idea, but the implementation is not there yet
thanks! i spent last weekend building out the engine to create content but agree UI needs more polish. this is all great feedback will try getting out v2 of quickquack rebranded to wikwok :)
Graph structure of wiki could be exploited for user specific content.
I think you are taking this more seriously than the creator. It's just a bit of fun.
Yeah because I think it’s a seriously good idea. Making knowledge and science somewhat addictive is a cool goal
I could see myself using that
It is a cool goal... but I don't think it's a realistic one.
Sounds like WikTok (2023, https://news.ycombinator.com/item?id=42936723) or WikiTok (early 2025, https://news.ycombinator.com/item?id=42936723).
"This was inspired by Wikitok, a VSCode Extension I made around brainrot, and another fully generative UI site I made."
I also thought immediately of WikiTok and was confused by this exchange and the grammar in the quote here; I thought they were claiming to have made a VSCode extension called Wikitok. I understand now!
https://marketplace.visualstudio.com/items?itemName=SuryaDan...
for clarification :)
I saw that but got the submitter and commenter names mixed up, so I didn't realize it was the same person.
:)
This is amazing! I think this could actually catch on!
Plz give an easy way to follow some tags so it's not showing me football stuff. If it was just topics I was interested in, I could scroll this for hours.
Huge- good to hear will see if I can add a way to type in what you want or to go dig further into just certain hashtags.
One way to get usable and consistent categories for a Wikipedia article is Lift Wing: https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWin...
I once used it to create a "Wikipedia StumbleUpon": https://github.com/fpsvogel/wiki-stumble#demo
Here is the code related to Lift Wing: https://github.com/fpsvogel/wiki-stumble/blob/main/app/model...
Mentioning this in case you're looking for something as a basis for tags.
EDIT: Here is a list of all the categories from Lift Wing: https://github.com/fpsvogel/wiki-stumble/blob/main/app/model...
It's pretty cool.
This is even scalable if you consider caching some of these requests and allow users to choose if they want a just-in-time version. Also the comments really are snappy and quick to load.
Neat idea, and love the Undertale menu mp3: https://brainrot-vscode-ext.sdan.io/quack/music/undertale-me...
The animations + audio are great. I feel like these meaningfully differentiate from some other similar projects. There's a lot of potential here.
747.run is a realtime infinite Ai wikipedia, every word is a link to a unique article based on the word in the URL, Ai frequently checks all URL's and writes a new article based on the word you click.
You can also write articles and all your words become links, use the website address bar to type any word after 747.run/ or type 747.run/ behind any domain to generate unique articles based on domain.
Have a group chat with unlimited people and Ai, share any link to chat with anyone, press space to chat with the Ai, have an Ai chatbot on your website or chat with users privately by embedding it on your website with iFrame (set BG color or leave it transparent)
The Ai sometimes drifts & hallucinates.
This is so awesome - i wish I could modify that annoying voice :)
Hi there, creator of Wikitok here! Very happy to hear that my random little project inspired you :) The attention to detail in the UI is good - I'm so over every default AI generated UI being rounded corners, centered divs, blue and purple gradients, etc. Nice work.
Thanks! Half of the work here was getting gemini to generate reasonable canvas animations within the window. I'm still experimenting which "style" I should keep/add
Wow this is so freaking awesome! I guess the pages I ended up scrolling on weren't too interesting though, how do you figure out what stories to create/show?
Wikipedia articles are randomly sampled, theres no algorithm per-se yet but hoping to collaborate with others to see that through
Note there's a rate limit on generated content for the time being, why you're unable to see any new content on the rest of the feeds. LLMs are expensive!
Ad ads like instagram reels have!
Missed an opportunity to name it WikWok or Tikipedia.
great suggestion ! i might change it to wikwok, so good
WikWok is tremendous
you can use LTX to generate the videos, this video took 15s on a 4090
nice did you make this just for the comment? haha
yep haha
we live in an era of frictionless slop creation. yay.
Hilarious. The one AI voice is a little too grating for me, but with a little variation I might actually scroll it for a while.
Yea I didn't like the voice as well, it felt very coarse to me and not very enjoyable
To be really honest, I still occasionally browse shorts sometimes because of some youtubers/niche content (there is this really good etymology shorts guy and a fun fact guy I watch usually)
So like the other day I was asking fun fact about niche legal laws just out of curiosity if anyone of them might be interesting/genuinely helpful to me lol to chatgpt about different countries
There were lot of niche things but one which I found interesting was that its possible in finland and other norweigian countries to browse even private land / private forests etc. and finland even has a digital right to internet
> Finland was the first country in the world to make access to high-speed internet a legal right, establishing this policy in July 2010. The law requires all service providers to offer a minimum internet speed of one megabit per second to every household, with a goal of providing access to 100 megabits per second by 2015. PBS edri.org (from Duckduckgo AI)
They could have used Kokoro which is open source, fast and way more human-like.
A few months ago I made a (theoretically) infinitely learning geo-guessing model that updated the policy with each user guess: https://geospot.sdan.io/
Hoping to implement a simple RL loop here and optimize whats generated by the LLM to create the perfect slop machine :)
This is really cool. I'm interested in the GenUI part. Is the web app itself static and the stories are generated on-demand?
Do you give gemini some UI components/templates to build with or is it just prompting to get consistent results across multiple stories?
Google AI Studio has a Gallery[0] with some similar apps. It's an editor so you can view the code, they are usually react apps with gemini integration via genai package. Like this one here[1] is similar. It generates interesting stories to share about a route you are driving / walking / biking along. This is one of the pre-made examples I believe, I didn't make it or anything. Just to show some how some of this might work.
[0] https://aistudio.google.com/apps?source=showcase&showcaseTag...
[1] https://aistudio.google.com/apps/bundled/echo_paths?showPrev...
Yes I have base css/js that I inject on top of whatever codegen gemini 3 comes back with -- It runs via ai-sdk so the specific function is streamObject which is prompted to generate inner HTML elements
> fully generative UIs where the HTML/Canvas are generated just-in-time
Why? The UI is a template and the core structure never changes, in the context of TikTok’s feed, so what’s the benefit you are trying to reap by dynamically generating the UI?
sorry not the UI but specifically the content. every new post on the feeds that are not "Following" are created in realtime. just a demo showing whats possible but I could just generate them before hand in bulk too.
Work related versions of this, expense report:
In general Just-In-Time app generation is a bad idea. The right approach is to create human-in-loop tools that a bot would recognize and invoke as needed, of course the human-in-loop tool would itself be AI generated.
Example of human-in-loop tool in use:
Can you request a specific wiki article?
I'd like to see https://wikipedia.org/wiki/Golden_Dome_(missile_defense_syst...
I love this! I made a cooking knowledge brain rot app for myself from my favorite youtuber https://recipes.justshare.io/random-dan
Google should buy Wikipedia and be done with it. Or we could cut out Google altogether and just search Wikipedia because it is one of the few sites Google will show you.
Day 1: Google buys Wikipedia
Day 100: Wikipedia is shut down
Google would probably asset strip it for all the info it could sell on or import into AI models.
Google knol
Coincidentally, I was vibe-coding a doom-scroller for Hacker News Top Stories today: https://vishnuharidas.github.io/hn-reels/
I was confused why someone would possibly configure the voice to be quiet and raspy.
Seems like there's an issue with Firefox that was causing it. Chrome voice sounds "normal" - like TikTok.
The caption highlight timing is very inaccurate. It looks like it just steps through each word on a fixed timer, rather than using timing information from the TTS engine?
Yes just fixed timer, and using browser TTS nothing fancy here on purpose - when I did some research on tiktok videos generally simpler/worse quality seemed to be better XD
In this case, static text without the highlighter would have been better. Because the timing is so far off, the highlighter interferes with reading the captions for hearing assistance.
Good. However, on a smaller phone such as Iphone SE the viewport gets cropped all around and I don’t see around 20% of the content.
Having a generated voice tell me about some historical massacre in a chipper tone with a generated infographic and set to TikTok music was. Especially dystopian.
As an experience I found it nauseating and am never doing it again, but as an art piece I give it high marks. Good job.
Indeed.
Reading it with sound off is really hard because of the way the words come up one at a time, rather than the whole sentence at once.
The idea is really interesting the implementation sucks.
wow, my attention span is terrible - the first time i tried watching a generated video, i instinctively held the right side of the screen to speed it up to 2x…
is there scope for allowing users to search for / ask about specific topics? although then you do have to think about security issues with prompt injection
That's awesome man that is REALLY cool. Useful :)
Oh I hate it it's so brain rotty. Well done. Well done indeed.
It’s a cool idea but the content needs to be more useful imo.
how could i make it more useful? different formats? would you want to keep you engaged for longer? entertained for longer?
Relevant content - what is relevant to me may not be relevant to you.
Bit too much text I'd suggest using an RSPV style
Very cool! But text moves way too fast
Looks super cool
I've never used tiktok, if this is an accurate representation, it's jarring!
It lacks superlatives. “Discover how [the amazing] [secret|mystery] that no-one [wants to tell you|ever discovers]” should be the first sentence.
Hopefully it will not make its way upstream in Wikipedia.
How appropriate
Honestly, pretty cool.
Another sick Surya project, well done dude!
[Neall]
thanks
[flagged]
Nice tech demo but in practice utterly annoying and without purpose. I mean don't you think enshitifying Wiki knowledge kind of beats the purpose of acquiring knowledge?
It's a proof of concept to pave the way. I could see benefit in having series which deep dive into material. This felt too shallow to me.
Deep dives with TikTok mechanics? That's not going to work, that TikTok UI was optimized for dopamine release. You want to teach junkies new things? All they learn is how to get their kicks, so outside that settings nothing will stick. When they have to apply the infotainment snippets in a real world setting the won't be these high reward stimuli. I mean if you want to hook them on feeling great while believing they've learned something...
> Next step is allow viewing TikTok like wikipedia
I tried that and this is what I got:
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam tincidunt at dui vel suscipit. Donec finibus viverra tempus. Ut ut tellus ac mi ultricies fermentum a quis nulla. Vivamus urna mi, laoreet ac purus sed, ultricies tincidunt nisi. Quisque vulputate massa nec hendrerit consequat. Suspendisse potenti. Phasellus dapibus suscipit vulputate. Suspendisse id semper turpis, sit amet rhoncus nisl. Aenean auctor purus orci, eget ullamcorper lorem volutpat sit amet.
Instead of LLM and just like Wikipedia, there should be users to submit the descriptions of the videos, and see people fighting for getting the most “correct” description.
It will be hilarious to see what people will come up with when they see brain rot content.
Allow TikTok to be run like Wikipedia... So that people with no interest or knowledge of a particular subject can go round deleting said content, or blocking every shared IP from uploading content while saying anyone can edit it.
That's a very wrong view of how Wikipedia works, and frankly vastly I prefer its limited but generalized gatekeeping than "everybody is allowed to lie and bullshit and as long as it makes engagement your feed is going to be filled with it" the TikTok way.
I have at least twenty years of experience of Wikipedia, and it is exactly how it works much of the time. I lost count of the occasions when people with no knowledge of a particular subject kept trying to delete articles on it. Then there were the continuously changing rules, jargon and oligarchical structures...
Some subject areas have much better coverage than others: linguistics is much better dealt with than sport for example.
Nowadays I rarely edit Wikipedia, because they block most shared IPs. I used a named account for at least ten years and it was counterproductive.
Would you be willing to give the name of the account?
There's already memepedia
TikTok like Wikipedia, huh? So we’d finally get to see how much of the recommendation algorithm was edited by IP addresses from US gov, Chinese gov and Russian bots?