TikTok knew, and I've decided to find that helpful instead of terrifying

I spent a few days buried in personal stuff. I moved, then flew to Switzerland to see family, and it was, to put it gently, a stressful few weeks. When I finally got a few free minutes, I did the responsible thing and opened TikTok to rot my brain for a while. I knew exactly what to expect. My feed is the human version of a LinkedIn cold open: the top five AI tools you have to try this week, a guy explaining why Monday's market drop was a correction and not a crash, a founder mining his near-bankruptcy for a carousel. Tech, finance, founders, the occasional nudge to eat a vegetable.

None of that showed up. The feed had redecorated while I was gone. New creators, different tone, stuff I'd genuinely never been served, and all of it aimed less at what I'm into normally and more at how I was actually doing that week, what I needed to hear. I hadn't liked a thing. Hadn't followed anyone. Hadn't typed a single search. Which left me with two questions, and a slightly uneasy feeling about both. How did it know. And how did it know that fast.

I never changed a thing

Over those few days I gave the system nothing it could see the normal way. No likes. No follows. No searches. No comments. I opened an app and moved my thumb a few times. That was the whole transaction.

And it still didn't feed me more of what I engage with. I engage with tech. By every rule we usually write, it should've buried me in AI tool roundups until I begged it to stop. (Which I do on occasion) Instead it served me something sitting completely outside the profile I've spent years building. It didn't double down on my history. It set it aside.

So what did it actually read? How long you hovered before scrolling. What you watched twice. What you finished. How fast you bailed on the things that didn't fit. What time of night you were doing all this. You're broadcasting the entire time, just by watching, and none of it asks your permission first. I understand all the behavioral inputs, but how fast it adjusted was what got me.

To be fair, it didn't need telepathy for all of it. I'd just moved, then I was suddenly opening the app in Switzerland, on different wifi, at hours I'm normally asleep. Some of the redecorating was probably plain context: new country, new time zone, new 2am. But context explains where I was. It doesn't explain how the feed seemed to know how I was.

A strong signal doesn't need much to go on. A few videos you slowed down on is plenty. The system didn't need a week of me. It needed an evening, less than that - a few minutes.

Recent beat everything it knew

I keep coming back to this one. A multi-year, rock-steady history of clicking the same few topics is about the strongest customer profile anyone could dream of. Most of us would treat it as gospel. The feed treated it as background music.

It made a bet, and a weird one. It decided that a few minutes of right-now behavior told it more about what I wanted than years of consistent interest did. And it was right, which is the really annoying part. Where your head is tonight predicts your next ten minutes better than anything you've loyally clicked for three years. Recency and context beat the profile, and it isn't close.

If you do personalization for a living, that should make you stop. The profile you spent months assembling can get outvoted in an afternoon by whatever someone's doing right now. The history tells you who they are. The last twenty minutes tell you who they are today. Today is the one they act on.

The stuff I'd never pick helped most

The content was good, too. Not good like a slick edit. Good like it was useful to me, what I needed to see, right then, in a way my actual feed has never once pulled off.

Think about what my normal feed is for. It feeds the guy I perform: on top of the tools, ahead of the market, calmly learning from other people's disasters. It's a feed built for my resume. It is not built for a bad week. And in a stretch like that one, the carefully optimized version of my own interests had nothing for me at all. The random stuff the algorithm dragged in, the content I'd never have chosen on purpose and would have flicked straight past a month earlier, was the part that actually helped.

That should poke at anyone whose job is deciding what other people see. We build feeds, and stores, and recommendations around who people want to be. Every so often the thing they need is the exact thing they would never click. I spent years developing RFM-based direct response personas, including intent scores for millions of yearly catalog drops. At least with catalogs you'd hope they 'lay' around long enough for the status quo to set back in. Faster media doesn't have that benefit.

Clarity, or just confirmation

Before I get too misty about my kind and caring algorithm. A lot of what felt like clarity was just the feed agreeing with me. It found content that lined up with how I already saw things, and being agreed with at volume feels an awful lot like insight. "Finally, someone gets it" is a wonderful feeling. It is not the same as being right. I already had a preconceived notion of 'right' the algo just reinforced it. I guess that's how fake news gets started.

A good recommendation engine isn't really in the information business. It's in the keep-you-watching business, and the most watchable thing is usually whatever confirms what you already think. So it hands your own opinion back to you with better lighting and a sympathetic voiceover, and you feel understood. You are part of a community, however fleeting. Sometimes that's real comfort. Sometimes it's a very efficient echo. On a hard week, from the inside, the two are almost impossible to tell apart.

And my normal feed was never above this either. "Why the drop was a correction, not a crash" is also just telling me what I'd like to be true. The machine didn't suddenly start confirming my biases that week. It only changed which bias it was confirming. It was nice while it lasted.

Watch the dwell, not the like

Two things from an evening of accidental research. The first is useful and a little boring: if you build these systems, or sell through them, you're probably still watching the wrong column. The signals people perform are the loud ones. The signals people mean are the quiet ones they never chose to send. The like is the noise. The dwell is the truth.

The second one I'm still chewing on. The same feed that handed me something genuinely useful also handed me my own conclusions back with the lighting fixed, and on a rough week I couldn't always tell the two apart. Genuinely helpful machine. I just wouldn't trust it most on exactly the nights I need it most.