social

Farcaster

The good news is that doing social media algorithms is *hard*, so no one really should be faulting Merkle.

I'm sure you've seen the complaints on X (so much so they released their algorithm, never updated it, and now claim to be Grok LLM powered). You've seen the complaints here on Farcaster. 

There are complaints on Threads too!

We once had only "following", which I guess many did get used to (the Facebook timeline/news feed). We were told algorithms would be better for us, but even facebook hasn't solved this since 2006!

Maybe the only people that cracked algorithmic discovery so far has been TikTok. And no, we haven't seen an update to Monolith since they talked about it. https://github.com/bytedance/monolith

hypothesis: text/image feed algos are "bad" because of a lack of preference revealing signals the algos have to determine what you like. the only reliable signals you can get from these feeds are explicit engagement (reactions/clicks/shares)

watch time is a very strong "interestingness" signal

a simple embedding is sufficient as as means of categorization

it's also kinda interesting that the algos don't need to understand content in order to be effective

I kind of agree with this hypothesis - because of watch time, but you can also post longer, or use threads. There are other signals X uses, like if there's a profile visit, as part of engagement metrics.

Monolith has a batch training stage (tensorflow like), and a online training stage (live feedback). Then the Kafka pipeline is highlighted, and they increase snapshot frequencies. 

They even have an example of using the MovieLens dataset, and a criteo dataset. 

Overall a very good paper to read!

Applied AI, databases, data, vibe coding (sic.), open source, community/devrel, remote. 

itinerant businessperson. globalist.

Kuala Lumpur, Federal Territory of Kuala Lumpur, Malaysia

I kind of agree with this hypothesis - because of watch time, but you can also post longer, or use threads. There are other signals X uses, like if there's a profile visit, as part of engagement metrics.

Monolith has a batch training stage (tensorflow like), and a online training stage (live feedback). Then the Kafka pipeline is highlighted, and they increase snapshot frequencies. 

They even have an example of using the MovieLens dataset, and a criteo dataset. 

Overall a very good paper to read!

https://grouplens.org/datasets/movielens/

Overindexing on an unreliable signal = bad

I think most text based feed algos have also evolved to a point where they over index on time spent (watching / reading) etc on a text content as opposed to just explicit engagement

dev + founder | @vibesengineering.eth @herocast

it works better for full screen videos but it can be a great signal compared to just cast interactions 

we have character limit and video length limit. I def need longer to read because I’m slow, can be noisy attention signal for many casts on screen, model will still learn just take more data vs full screen content

- biases long casts
- it takes less than a second to read most casts
- you can fit many casts on the screen so how do you know which is being read

we can’t do read time for text based as a signal?

DX Engineer | DevRel at OpenZeppelin | Building orbiter.host | stevedylan.dev/links | opinions are my own

yup, and to be clear, protocol wise, we don't have time on screen, this is implemented in the client level

Imo the ad scene has definitely moved from clicks to time on screen, and nothing does it better than enabling addictive behaviors

The good thing about Wallet Connect etc being awful is that at Kiwi I learned to favor revealed signal over stated. The annoying thing is that it makes the p2p replication protocol kinda pointless as upvotes is just one out of very many signals