Great article as always. thank you for breaking it down and including diagrams! (PS, is the Impact description correct in the poster? It seems to be missing its ending.)
“Whenever we anthropomorphize or simplify, we inadvertently legitimize what the tech platforms are actually doing.” Was hoping for a deeper breakdown of the how and why behind this. Some interesting threads in this piece nonetheless!
Here’s a quote from Kevin Munger’s substack iPhoneomenology that Adam had linked in this post: “The fact that we see agency—often cruel but at least human—in the algorithm reveals our fundamental need for interpersonal connection. It lets us imagine someone like Zuck in control, instead of adrift in “the apparatus” like the rest of us. We wish Big Brother were watching. But we might just be alone with our phones.”
This is a great breakdown! I see a few comments mentioning content-aware vs collaborative filtering, so I’ll dive into this a bit. Collaborative filtering (often based on user clustering, like what users with similar tastes also enjoy) is usually very effective, however it falls flat on cold start, which is a scenario where there is no data about an entity because it hasn’t interacted with anything yet and exists in a vacuum. Losing new users due to a poor cold start experience (due to bad or lacking content recommendations) is one of the biggest problems that platforms are trying to solve, so they are highly incentivized to quickly categorize users and content and exit cold start as quickly as possible. If you want to categorize new content with no engagement metrics, you’d probably want to examine what other content it’s most similar to, and you might start by trying to extract key elements like language, pacing, aesthetic, etc and then compute a predicted audience and popularity. Tracking content based features allows for more fine grained user and content profiling, moderation, and so on, but it also adds a lot of complexity to the system, so if it wasn’t adding value it would be removed.
I am alarmingly social media illiterate, but then again I am a baby boomer. The world has shifted completely beneath my feet and I hardly noticed. I mean I know my data is mined and I know about algorithms, but not to the level described in this article. Kids need to start learning this in school as soon as possible.
It’s very worrying to me as a young person how me and my peers use social media so much without any proportionate understanding of the mechanisms behind it. We’re setting up the kids to be manipulated by media like never before.
These makes me think of how, to me, there are other sides of media, say Tiktok, that are so foreign to me that are viewed and liked by millions. Entire celebrities and drama along with trends I never heard of when the algorithm "breaks" and shows me "mainstream" content
Other comment mentions this but I wonder how much their fleet of algos rely on collaborative filtering, which doesn’t inherently require “knowing” anything about what is in the actual video. Those algos in isolation sorta suck for cold users and content though
I think it’s interesting that TikTok actually uses content aware algorithms, I expected it to be mostly content unaware collaborative filtering, but maybe that’s part of TikToks success.
Great article as always. thank you for breaking it down and including diagrams! (PS, is the Impact description correct in the poster? It seems to be missing its ending.)
ah thanks - my publisher is fixing that before posters are printed
“Whenever we anthropomorphize or simplify, we inadvertently legitimize what the tech platforms are actually doing.” Was hoping for a deeper breakdown of the how and why behind this. Some interesting threads in this piece nonetheless!
Here’s a quote from Kevin Munger’s substack iPhoneomenology that Adam had linked in this post: “The fact that we see agency—often cruel but at least human—in the algorithm reveals our fundamental need for interpersonal connection. It lets us imagine someone like Zuck in control, instead of adrift in “the apparatus” like the rest of us. We wish Big Brother were watching. But we might just be alone with our phones.”
This is a great breakdown! I see a few comments mentioning content-aware vs collaborative filtering, so I’ll dive into this a bit. Collaborative filtering (often based on user clustering, like what users with similar tastes also enjoy) is usually very effective, however it falls flat on cold start, which is a scenario where there is no data about an entity because it hasn’t interacted with anything yet and exists in a vacuum. Losing new users due to a poor cold start experience (due to bad or lacking content recommendations) is one of the biggest problems that platforms are trying to solve, so they are highly incentivized to quickly categorize users and content and exit cold start as quickly as possible. If you want to categorize new content with no engagement metrics, you’d probably want to examine what other content it’s most similar to, and you might start by trying to extract key elements like language, pacing, aesthetic, etc and then compute a predicted audience and popularity. Tracking content based features allows for more fine grained user and content profiling, moderation, and so on, but it also adds a lot of complexity to the system, so if it wasn’t adding value it would be removed.
I am alarmingly social media illiterate, but then again I am a baby boomer. The world has shifted completely beneath my feet and I hardly noticed. I mean I know my data is mined and I know about algorithms, but not to the level described in this article. Kids need to start learning this in school as soon as possible.
It’s very worrying to me as a young person how me and my peers use social media so much without any proportionate understanding of the mechanisms behind it. We’re setting up the kids to be manipulated by media like never before.
Regarding the poster, American customers means US only??
yes, unfortunately
These makes me think of how, to me, there are other sides of media, say Tiktok, that are so foreign to me that are viewed and liked by millions. Entire celebrities and drama along with trends I never heard of when the algorithm "breaks" and shows me "mainstream" content
:)
Nice job, nerd
Super informative, great stuff 👏
Other comment mentions this but I wonder how much their fleet of algos rely on collaborative filtering, which doesn’t inherently require “knowing” anything about what is in the actual video. Those algos in isolation sorta suck for cold users and content though
I think it’s interesting that TikTok actually uses content aware algorithms, I expected it to be mostly content unaware collaborative filtering, but maybe that’s part of TikToks success.