# TikTok Algorithm Anatomy A crafted machine that shows how a For You feed is chosen: eligibility, fast two\-tower retrieval, prediction scores, a diversity mix, instant play and learning from reactions By aqilirwan Topic: Computing, AI & Electronics [Open the interactive Jig](https://ask.jig.com/j/WgaHrUEkvldk) ## Transcript ### 1. The For You machine Every For You feed is different\. This little machine is a model of how TikTok picks your next video: it estimates which videos you are most likely to enjoy, then plays them\. Drag the empty paper to turn it around\. The parts are a playful picture\. The real system is software running on many computers\. ### 2. Only eligible videos New videos drop in at the gate\. TikTok says some are never recommended: spam, duplicates, videos still under review, ones you have already seen, and content not suitable for a general audience\. Those fall into the bin\. Eligible just means allowed in\. Nothing has been chosen for you yet\. ### 3. Fast retrieval: two towers The catalogue holds an enormous number of videos, far too many to check one by one\. A common trick is two towers: one turns you into a list of numbers, the other does the same for each video, and close matches are grabbed in a flash\. This is how typical big recommenders work\. TikTok has not published its exact design\. A first shortlist is often around a thousand videos\. ### 4. Predicting what you will enjoy Next, a model gives every candidate a prediction score: how likely you are to be interested\. It uses what you watch, like, share and comment on, plus each video's captions, sounds and hashtags\. The highest score wins\. Watching a longer video to the end is a strong signal\. Being in the same country as the creator is a weak one\. ### 5. Mixing it up A feed of only top scores could get boring\. TikTok mixes in different kinds of videos, and it tries not to show two in a row with the same sound or from the same creator\. The turntable swaps in a surprise\. A surprise video can help you discover something new to love\. ### 6. Instant play Apps like TikTok fetch the first part of the next few videos before you reach them\. They wait behind the phone, so when you swipe, the next video is already there and starts right away\. Studies of TikTok\-style apps found that a lot of this preloaded video is never watched, because people swipe past it\. ### 7. Your reactions are signals Every swipe, like, share, comment, follow or Not interested tells the machine something, and so does watching to the end\. Press Swipe, or tap the phone, and watch your reaction travel back along the rail\. TikTok says follower counts and a creator's past hits are not direct factors\. ### 8. Learning while you watch ByteDance, the company behind TikTok, published a system called Monolith that keeps learning from new reactions while it serves videos\. Its quick parts update about once a minute, so a feed can adapt fast\. Here one turn of the clock stands for about a minute\. Monolith's bigger, slower parts update about once a day\. ### 9. You are in charge too The machine predicts; it cannot read your mind\. You can tap Not interested, filter out words and hashtags, or refresh your feed to start fresh\. Then the whole loop runs again, one video at a time\. Look for Why this video in the share panel to see some reasons a video was picked\.