News for hackers. Stuff that matters.

sdan

Re:Contrastive Language Models on Thursday September 24, 2026 @04:19AM (#49827766)
I tried running this on a H100 and got 190ms compared to Jev's 170ms. Maybe I set it up wrong?
nice! kind of similar but re-made a scrollable wikipedia a few weeks ago: https://quack.sdan.io/
Re:Show HN: Scroll Wikipedia like TikTok on Saturday January 10, 2026 @07:37PM (#46571516)
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.
Re:Show HN: Scroll Wikipedia like TikTok on Saturday January 10, 2026 @07:35PM (#46571505)
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 :)
Re:Show HN: Scroll Wikipedia like TikTok on Saturday January 10, 2026 @04:00AM (#46564044)
Wikipedia articles are randomly sampled, theres no algorithm per-se yet but hoping to collaborate with others to see that through
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @10:20PM (#46562446)
how could i make it more useful? different formats? would you want to keep you engaged for longer? entertained for longer?
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @09:56PM (#46562300)
great suggestion ! i might change it to wikwok, so good
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @08:51PM (#46561926)
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!
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @08:50PM (#46561922)
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 :)

Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:59PM (#46559975)
thanks
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:59PM (#46559974)
nice did you make this just for the comment? haha
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:58PM (#46559969)
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
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:56PM (#46559948)
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
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:55PM (#46559935)
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.
Re:Show HN: Scroll Wikipedia like TikTok on Friday January 09, 2026 @04:55PM (#46559922)
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

Show HN: Scroll Wikipedia like TikTok

https://quack.sdan.io

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...

( Read More... | 89 comments | 96 points | on HN )
Re:Show HN: I embedded 10M StreetView images on Friday December 26, 2025 @02:42PM (#46395422)
To prevent overfitting
Re:Show HN: I embedded 10M StreetView images on Friday December 26, 2025 @02:41PM (#46395420)
Getting there! training a new model, on thinking machines' tinker api which uses a considerably larger and different architecture.

In the meanwhile I tried at least 100+ variations of trying to train this model to be SOTA most of which led me down getting better data which I believe I do, just needs more tuning still.

Older »