2025-10-03
I've been thinking a lot about a non-algorithmic, non-infinite scroll "social feed" website. Something that can collate information from blogs, websites, news outlets, even social media posts, but only those places that a person explicitly seeks out.
I previously removed all social media apps (except messaging apps) from my phone - I instead will force myself to go through the friction of opening a site on browser, running through screen time limits, etc. I mostly end up looking at twitter once in a while, youtube as well, but really have been off of instagram, etc.
I used to use a lot of twitter. The scroll, tweet structure, etc makes a lot of sense. I also am on a bunch of different email newsletters that I actually am interested in reading. (some examples are VC newsletters just for fun, or I've been really interested in Car Dealership Guy's newsletter about the car market, idk why)
Personally, I'm not super into the email newsletter model - I don't want more notifications, and I want a better feeling of opting in or opting out.
I like the idea of having to open a specific place to get the internet blasted at me, and then once closed, I can get back to regular life. This, by the way, is why I've been trying to unsubscribe from a lot of automated emails as well. I have to use email for work and jobs and such, I can't delete it.
Infinite scroll is bad - there should always be some cap.
Blogs and explicit writing from others. When people put in the effort to write something and publish it, it shows. Simon Willison's blog, paulg's essays (ha!).
News. News media as a whole has devolved a decent amount, but there are still sources out there that are interesting to me. This can range a lot - I'm very interested in economic news, such as news about rate cuts (stay strong jerome powell), news about politics (preferably not stupid polarization rage bait though). I've been listening to a lot of NPR as a source of news, but there's also just general news sources. Also - sports! Being bay area born and raised, I actively seek out news about bay area sports teams. Some news around it is drivel (lots of the news around the warriors is just engagement bait) but some of it, like Sheng Peng writing about the San Jose Sharks at SJ Hockey Now is just generally great journalism.
There's an interesting note here about how a lot of news outlets will just churn out the same variations of the same news media multiple times. I don't need to see 18 different takes on jonathan kuminga signing a 2 year deal with the warriors, I just need a couple of interesting perspectives on it in total.
Why do I still engage with (read through) twitter even though the majority of the feed is now neoconservative conspiracy theorists, angry pitchfork mobs and bots? Once in a blue moon I'll stumble upon an interesting piece of writing, some new voice that I never knew about. Recently I discovered simon willison's weblog (it's been around since 2002) - he wrote an article titled "Designing agentic loops" that was super relevant to the work I'm doing myself. Part of the draw of algorithmic feeds is the ability to discover something new once in a while. However, the pace of that discovery is just too slow. It's like looking for a needle in a haystack, except the hay is a flaming pile of horseshit.
What happens when you see 18 different media outlets publishing the same thing? You mentally need to make a choice. That sucks. LLMs let us do a lot of summarization, processing, combining - while citing sources.
I explicitly choose some feeds that I know I want to read. For example, atom feed of Simon Willison's blog, RSS of paulg's essays. I explicitly define some news terms that I care about. Economic news, gsw news, sj sharks. I set up some general interests. Tech development. Photography. Art and culture events in nyc.
The system pulls feeds, scrapes news, and ingests social media (maybe I have it hooked up to my own twitter account as an example). It summarizes topics that have multiples of the same article and combines it into a single view.
No comments. Liking/disliking is purely to help tune the LLM layer.
Scrape RSS feeds, social feeds, news sources → LLM processing layer to summarize → recommendation algorithm tuned only on myself → limited item UI → read items go away
Recommendation algo should always balance explicitly subscribed items with new sources. Adding a new source should be easy, similar to following.
Components is probably the wrong word here but this is outlining the major chunks of work needed.
RSS and Atom feeds are easy (standardized). What about scraping news sources, search sources, etc? What about social media?
What is the structure of a reading post that I want to see? Need to define that. How does the prompt work? LLM needs to do some labeling on content.
Define labels used for a simple matrix rec algo. Define specific limits of daily content.
What are engagements with posts. How many posts do I see.
Save items
Share items
Profile
This is not a social app. I don't want other people in my pool, shitting in my algorithmic feed.
This is a reading product. That's it.