Post ARVQ1OR2XCB7fAIQxU by clive@saturation.social
(DIR) More posts by clive@saturation.social
(DIR) Post #ARVOw5o8GQ7uUM0S5w by clive@saturation.social
2023-01-10T16:22:37Z
0 likes, 0 repeats
People write about TikTok's "for you" recommendations as "eerily good", "mind reading", "spooky" -- and it does seem considerably better than its social-media rivalsBut odds are there's nothing remarkable about the algorithm itselfRecommender systems are pretty mature -- it's unlikely (though not impossible) Bytedance found some new concept nobody else knowsNo, the secret ingredient is more likely ...... the UI:My essay: https://medium.com/@clivethompson/maybe-tiktoks-algorithm-isn-t-particularly-remarkable-after-all-712a5d2bae40"Friend" link: https://clivethompson.medium.com/maybe-tiktoks-algorithm-isn-t-particularly-remarkable-after-all-712a5d2bae40?sk=3ffbf858464e3462fb3d5a84259d27de
(DIR) Post #ARVPLWEX9Ng1gcTn3A by Jonathanglick@mstdn.social
2023-01-10T16:28:18Z
0 likes, 0 repeats
@clive This makes MUCH more sense than the algorithm-genius theory.
(DIR) Post #ARVPROAu1b3SQla1vU by clive@saturation.social
2023-01-10T16:29:27Z
0 likes, 0 repeats
@Jonathanglick Yeah, precisely
(DIR) Post #ARVPpjzmY1NJ3XzTEW by Jonathanglick@mstdn.social
2023-01-10T16:34:14Z
0 likes, 0 repeats
@clive What I'm really curious about is how much personal data is *actually* necessary to make seemingly excellent recommendations, provided a) enough decent content, b) fast UX/ performance.
(DIR) Post #ARVQ1OR2XCB7fAIQxU by clive@saturation.social
2023-01-10T16:36:06Z
0 likes, 0 repeats
@Jonathanglick Yes! That's a really important question.I suspect it's highly contingent on what the type of thing being recommended is (book? takeout restaurant? life partner on a dating site?) and the range of options (45 restaurants; 90 million books)
(DIR) Post #ARVQAe0I1JqEYQCAxE by Jonathanglick@mstdn.social
2023-01-10T16:37:09Z
0 likes, 0 repeats
@clive Which leads us to the really fun question -- could there be a federated TikTok?
(DIR) Post #ARVSMOACwGPmoJYvqa by dswersky@hachyderm.io
2023-01-10T17:02:29Z
0 likes, 0 repeats
@clive We take these algorithms for granted so much now that the other obvious example- Amazon- is easy to forget. I remember being creeped out about how good the Amazon recommendations are... until I got used to it and stopped thinking about it
(DIR) Post #ARVTJmwOSejsk0QP8y by clive@saturation.social
2023-01-10T17:13:17Z
0 likes, 0 repeats
@dswersky Yep yepAnd Amazon, too, may not be terribly complex beneath the hoodBack in 2008 when I first wrote about recommender systems, I talked to an Amazon engineer who worked in the recommendations there ...He told me that a surprisingly hefty amount of successful (i.e. someone actually buys something) recommendations were just - ah, you liked that author? Here's another book by the same person- you bought [Product X]? Want to buy it again?Utterly simple ruleset, drove sales
(DIR) Post #ARVTol1lcnrrmfho5A by nazgul@infosec.exchange
2023-01-10T17:17:38Z
0 likes, 0 repeats
@clive That’s really interesting. (Someone better tell Meta. :)I would argue though that there *is* an algorithm they might be doing better than other platforms, and that’s in classifying the video content. I could see that having a multiplier effect on any recommendation algorithm.But I found it interesting reading this right after reading Cory Doctorow’s essay on social quitting; particularly the part about the willingness to throw in wildcards because they aren’t focused on high profile influencers.What I wonder is if they can afford to continue doing that in the long run, or will the economics lead them down the same path as Instagram, where ads and influencers start to dominate, and the things that make them unique go away?https://doctorow.medium.com/social-quitting-f049b33ad3f6#recommendersystems #imagerecognition
(DIR) Post #ARVUk6QqiRFYfvmXEe by clive@saturation.social
2023-01-10T17:29:12Z
0 likes, 0 repeats
@nazgul Yep, those are good questions -- I don't know!I wonder how much they're doing with classifying video content too. When I spoke to YouTube folks three years ago for a Wired story about their work revamping their recommendations system, I was surprised to learn that the content of the video itself -- the pixels on the screen -- was less useful than the text in and around it: The captions, the description and title, the comments, etc
(DIR) Post #ARVVALb2xP8rEi49Vw by clive@saturation.social
2023-01-10T17:30:41Z
0 likes, 0 repeats
@nazgul Some of that is simply because text is easier to analyze than video content, obviouslyBut it was still surprising -- even with google's obvious skill, scale and experience analyzing the content of images, they still leaned far more heavily on text and social signals (i.e. watch time, likes, subscribes, etc.)
(DIR) Post #ARVVM70n0btPTeiVLk by nazgul@infosec.exchange
2023-01-10T17:33:18Z
0 likes, 0 repeats
@clive as far as I know, video classification is still in limited use, but growing fast. I’ve seen some pretty amazing demos, but I don’t know what the processing costs are.I think it’s going to be necessary for anyone doing trust & safety at scale though. Manual video review is very expensive. Even tiny changes to it can greatly reduce the cost. So ML-assisted review could help a lot.
(DIR) Post #ARVVM7SRLo0irPeaRs by clive@saturation.social
2023-01-10T17:35:09Z
0 likes, 0 repeats
@nazgul Yeah, agreed on all that. The processing cost is quite possibly through-the-roof, which is why they don't do more of itYet it'd be invaluable for trust and safety
(DIR) Post #ARVWdDO4eqOC7zKLUO by boby_biq@toot.community
2023-01-10T17:50:21Z
0 likes, 0 repeats
@clive this is so good! “But as he notes, the action isn’t always at the level of the mysterious, abstruse algorithm. Sometimes it’s right up front: In the deft choices of design.” So many “secret sauce” comes down to this: design
(DIR) Post #ARVXhQD11rf90QiyfY by clive@saturation.social
2023-01-10T18:02:21Z
0 likes, 0 repeats
@boby_biq Yep yep
(DIR) Post #ARVXqLAf85MZzGsPHE by clive@saturation.social
2023-01-10T18:03:53Z
0 likes, 0 repeats
@nikkiana Yep, amazing points all around!
(DIR) Post #ARVcXzal7YFJcS9ZMe by boby_biq@toot.community
2023-01-10T18:56:08Z
0 likes, 0 repeats
@clive this is a fav topic of mine because I work in design and I had my second dissertation of design management about basically website usability (in general dissected one particular website - not important here)
(DIR) Post #ARVfYKigrqBH0biXBY by steven@cville.online
2023-01-10T19:29:58Z
0 likes, 0 repeats
@clive I agree all of those reasons make sense.I wonder about another factor that isn't mentioned (in your post, or by Narayanan): isn't TikTok all about interest graphs, not social graphs?In contract: Facebook recommends topics we are interested in mostly from content generated by our friends.That makes FB feed less interesting overall because (a) while we want to know what is going on in our friend's lives, (b) most of our friend's posts are pretty basic boring stuff w/ little novelty
(DIR) Post #ARVhbuDfXLLlt02S12 by mariyadelano@mastodon.social
2023-01-10T19:53:20Z
0 likes, 0 repeats
@clive wow I have never seen a recommendation algorithm described as “a mathematical version of your taste” before.I don’t know if I’ll ever be able to describe it in any other way now!
(DIR) Post #ARVhx6E85YTTCwMNA8 by clive@saturation.social
2023-01-10T19:56:47Z
0 likes, 0 repeats
@steven Yes, that's a good point tooNarayanan addresses this in his piece -- I didn't excerpt it in my writeup -- but he basically notes that TikTok focuses more on finding and promoting a single video, a single piece of content, than on a personThat has big implications for how its algorithm performs compared to FB or Instagram
(DIR) Post #ARVnv4BPJfeg0xV5Vo by aral@mastodon.ar.al
2023-01-10T19:57:33Z
0 likes, 0 repeats
@mariyadelano @clive I guess “an attempt to manipulate your behaviour based on the intimate knowledge we’ve gathered about you” doesn’t exactly roll off the tongue :)
(DIR) Post #ARVnv4i1MPk7e6l8LY by mariyadelano@mastodon.social
2023-01-10T20:00:25Z
0 likes, 0 repeats
@aral @clive also it doesn’t fully get to the essence of how an algorithm tries to manipulate behavior Most simple explanations never feel sharp enough
(DIR) Post #ARVnv55PxQSSofhooa by clive@saturation.social
2023-01-10T21:03:36Z
0 likes, 0 repeats
@mariyadelano @aral yeah agreedThese algorithms are a very weird bouquet of psychology, math, surveillance and high-octane advertising-capitalism
(DIR) Post #ARVoIeGbK8VKaRAsaG by mariyadelano@mastodon.social
2023-01-10T21:08:15Z
0 likes, 0 repeats
@clive @aral even this comment is so well-written, how do you do this magic?
(DIR) Post #ARVpOGaONh9GpkfsQa by clive@saturation.social
2023-01-10T21:20:32Z
0 likes, 0 repeats
@mariyadelano @aral lolThank you!Probably just practiceI've been a freelance magazine writer for 25 years
(DIR) Post #ARVsuiTcWFL6SrfiFc by mariyadelano@mastodon.social
2023-01-10T22:00:05Z
0 likes, 0 repeats
@clive @aral :owi: all that experience definitely shows!
(DIR) Post #ARW2kCtO5qv39YLrhg by clive@saturation.social
2023-01-10T23:45:30Z
0 likes, 0 repeats
@mariyadelano @aral 🤘 🤖
(DIR) Post #ARX5PwffBk5U4XvFCK by jhh1822@mstdn.social
2023-01-11T11:54:57Z
0 likes, 0 repeats
@clive I wish the streaming services would catch up
(DIR) Post #ARXVGWUTkr794qrIOG by clive@saturation.social
2023-01-11T16:43:14Z
0 likes, 0 repeats
@jhh1822 yep yep