Dan Romero
@dwr.eth
Why is a TikTok-like algorithm for giving new users a boost hard? 1. Social network vs. social media — we started as a follow-based social *network*. The core audience on the network uses a follow graph to express interests. It's also a social network vs. social media (i.e. television with an algo on phone). TikTok was a broadcast / interest-based algo from the start. 2. Narrow focus with a novel primitive — TikTok was extremely successful with a single category — young people (mostly women) dancing / singing to music with high remix-ability — and it took them several years to expand to their next category (gaming-oriented content for younger men). This existed for multiple years before Meta made the big shift to video on Instagram in August 2020. 3. User acquisition — They spent billions of dollars per year on user acquisition (infamously re-acquiring the same users multiple times). So there was a massive audience (even if not retained) to distribute the videos to (with an at-scale AI algorithm from their parent company Bytedance). So if you boost without the audience, you don't get the same result. 4. Text limitations — videos have a longer shelf life vs. text tends to feel stale after a day. There's also a high degree of unsaid context and in-group value (a typical cast from @six or @gwart has 3 layers to the onion). It's also harder to make a widely interesting text post vs. a video (or image). 5. Text is linked to who you are — a video is more likely to stand on its own (assuming it's visually stimulating) whereas what makes text interesting (on a relative basis) depends on who is saying it. Imagine the average new user coming into the network saying "gm happy to be here" vs. someone with 500K+ followers on Twitter saying the same thing. The engagement is as much about the person as what's being said. (Ironically, this is why anons have a lot of success on text-based social networks -- a lot of work to breakthrough of course -- but once you do no one is judging your opinion based on your credentials but on the merits of how smart / funny you've been).
16 replies
23 recasts
342 reactions
Tony D’Addeo
@deodad
implicit ML signals video 100% of views have a clear signal on engagement in video how many seconds the user spent on it and this ranges from 0 to many nothing like this exists for text so the best you can do is look it actual engagement like likes or comments given that only a small percent of views result in explicit engagement you're discover algorithm needs orders of magnitude more views to get capture the same signal
3 replies
0 recast
8 reactions
0xLuo
@0xluo.eth
Though Threads is a text-based social platform, it gives new users some exposure, likely because its massive user base allows the algorithm to quickly gather engagement signals and decide whether to keep boosting them.
1 reply
0 recast
0 reaction
Dan Romero
@dwr.eth
They have the Instagram graph, cross-app notifications; implied interests and algo. Not replicable for a startup.
0 reply
0 recast
2 reactions
