Why do I get one video that has nothing to do with me?

Deliberate exploration. A system that only serves what it already believes about you can never discover it was wrong. A small fraction of slots are spent on deliberate guesses, which is also how the feed escapes a rut it put you in.

Is it reading my microphone?

It does not need to. People with similar watch patterns to yours already watched the thing you are about to be shown. Collaborative filtering feels like surveillance because it is accurate, not because it is listening.

Is optimizing for watch time the same as showing me what I want?

Only where the two agree. Watch time measures what held you, not what you would have chosen. Anything that reliably stops a thumb - outrage, cliffhangers, mild anxiety - scores well on the metric while scoring badly on your actual preferences.

Doodle 10

For You

Q: Does the algorithm use my likes?

Barely. A like is one bit, given rarely and deliberately. Watch time is continuous, given constantly, and given without deciding to. Completion rate and rewatches carry far more weight than the button you press.

The feed does not learn what you like.
It learns what you don't stop.

Everyone has a theory about the algorithm. It listens to your microphone. It knows you better than your friends. It is showing you this because of something you said out loud last Tuesday. The truth is duller and stranger than that, and you can watch it happen to you in about forty seconds further down this page.

jacobrakai · Network & System Administrator

It is not a feed. It is one decision, repeated.

There is no list being scrolled. There is a slot being filled.

A timeline is a list someone else ordered. A recommendation feed is not that. Every time your thumb moves, the system answers a single question — out of several million candidates, which one goes in this slot, for this person, right now — and then it throws the answer away and asks again.

That difference is the whole thing. A list can be finished. A slot cannot. And because the question is asked fresh every time, the system gets a fresh piece of evidence about you every time. Roughly two hundred of them in a ten-minute sitting.

   candidates          filter            rank            slot
  ~ millions   -->    ~ 1,000    -->    ~ 10     -->     [ 1 ]
                    cheap match      expensive        your screen
                    (region, lang,    model, per
                     recency, tags)    candidate

Two stages, for cost reasons. A cheap pass throws away almost everything; an expensive model only ever scores the few hundred that survived.

What it actually measures

Almost none of it is the button you press.

People assume the like button drives the feed. It barely registers. A like is one bit, given rarely, and given deliberately — which means it tells the system what you are willing to be seen endorsing.

Watch time is continuous, given constantly, and given without deciding to. You cannot curate it. That is exactly why it is worth more.

completion ratewatched to the end

rewatchlooped it without meaning to

watch timeseconds before the swipe

share / saverare, but strong when it happens

commentengagement, not necessarily approval

likeone bit, deliberate, easy to fake

Illustrative weights — the real ones are proprietary and change constantly. The ordering is the durable part, and it is the part people get backwards.

Watch it learn you

Swipe. It has no idea who you are yet. Give it eight tries.

now playing

Tap start

What it thinks you are

swipes: 0 · confidence: none

Nothing yet. With no history, the first few are near-random — the cold start problem. It is guessing, and it knows it is guessing.

Runs entirely in your browser. Nothing is sent anywhere, nothing is stored, and reloading wipes it. Five made-up dimensions instead of the several hundred a real system uses — but the mechanism is the mechanism.

Why you get one video that makes no sense

It is not a bug, and it is not the algorithm losing the thread.

A system that only ever serves what it already believes about you can never find out that it was wrong. If it decides on Tuesday that you are a woodworking person, and it only shows you woodworking, then every piece of evidence it collects from then on is about woodworking. The belief becomes unfalsifiable.

So a slice of slots — single digits, typically — is spent deliberately off-target. In the demo above that is the swipe that arrives from nowhere. It is the only way out of a rut the system put you in, and it is why a feed can still surprise you on year three.

  exploit  ################################.....  ~92%   what it thinks you want
  explore  ###.................................    ~8%   a deliberate guess

           the 8% is where new interests come from,
           and where the system finds out it was wrong

The formal name is the explore/exploit tradeoff. Every recommender has to pick a ratio, and the ratio is a product decision, not a mathematical one.

The part that closes

You are not being measured against who you are. You are being measured against who the feed already made you.

Here is the loop. It shows you something. You react. It updates. The next thing it shows you is chosen using that update — so your next reaction is to a menu it already narrowed. Every observation it makes about you is an observation about a person it has been shaping.

Nobody designed that as manipulation. It falls out of the arithmetic. A model trained on its own outputs drifts, and here the model's output is your input.

      +---------------------------------------------+
      |                                             |
      v                                             |
  [ it shows ] --> [ you react ] --> [ it updates ]-+
      ^
      |
   the menu was already narrowed by the last pass

The arrow that matters is the one going back. Without it this is a survey; with it, it is a feedback system.

The honest version

None of the above is a conspiracy, and the engineers who built it are not villains. The system does exactly what it says: it predicts what will hold your attention, and it is extremely good at it.

The problem is the gap between what holds you and what you would have chosen. Those overlap a lot, which is why the feed is genuinely good. But they are not the same, and everything living in the gap — outrage, cliffhangers, low-grade anxiety, the thing you would be embarrassed to have watched for forty minutes — scores beautifully on the metric while scoring badly on you.

The metric cannot tell the difference. That is not a flaw in the model. It is the model working, on a target that was never quite the right one.

The feed is not showing you what you love. It is showing you what you don't leave.