teddit

honest personal take (disclaimer: i'm not an ML expert, i just study and use a lot of social apps):

i think its highly unimaginative and tired to tell new users on a social platform to "reply a lot" in order to get any engagement.

social platforms, for a long time, pinned discovery as a user problem. tiktok changed that. they didn't expect users to earn attention the hard way (hustling to self-promote and get follows).

instead tiktok took on discovery as a platform problem and decoupled distribution from follower count entirely: a content-first, graph-agnostic approach.

on tiktok, a user's responsibility is to make good content and the algo would do the rest based micro interactions and watch behavior. YouTube took a similar approach, but favored creators who mastered their (effort intensive) rules: SEO knowledge, thumbnails, content cadence, etc.

if we're hearing that discovery is a problem, the question should be: whose problem — the user's or the platform's?

take long-form text platform substack: the recommender engine places more of the burden on the platform than the user, and the platform bias is towards quality via a trust graph.

it recommends newsletters based on what you read, introduces editorial curation, and the "recommended by other writers" feature means that small / new writers can get regularly recommended to audiences that a specific writer's taste.

each platform makes a design choice about who should work to be seen. on substack, they believe that good writers should be lifted through network effects and curation.

does this scale easily? no, it scales slowly but is compounding.

so why is this so hard for short-form text-based platforms like twitter, threads, bluesky, or farcaster?

because short-form text is typically low-signal, high-noise. short posts on their own carry very little context or signal. it's easier to produce, but harder to evaluate. as a result, discovery for short-form text platform has relied heavily on follower graphs. 

twitter initially solved this with hashtags (then failed to do it at scale with lists, fleets, and circle). substack partially solved this with its writer graph (trust graph) and categories (topic / semantic clustering).

and i think farcaster's open data, mini apps, and interoperability (zkTLS) can be leveraged for creative solutions here that aren't possible on other platforms. that excites me.