Week 11: The feed — recommendation as prediction

Why a recommendation feed is autocomplete for your attention rather than your sentence, and why that reframing changes what counts as a good guess

10 May 2027

Every system this course has covered predicts a specific, checkable thing: the word you meant, the pixels that belong in a gap, the reply that fits a message. A recommendation feed predicts something looser and, in a specific sense, easier — not “what did you mean” but “what will you engage with” — and optimising for the easier target has turned out to produce some of the strangest behaviour in this entire course.

A word predictor that’s wrong gets corrected and moves on; a feed that’s wrong in a specific, exploitable direction gets rewarded, because “wrong” for a feed means “less engaging,” and the feed’s only signal for what’s engaging is what you already clicked. That’s a prediction system with its own success metric quietly reshaping the thing it’s predicting — you become slightly more like the audience the feed has already decided you are, which makes the feed’s next prediction slightly more accurate, which is treated as the system doing its job well. Nothing about this requires the feed to be malicious; it only requires it to be good at the metric it was actually given.

Outline#

• recommendation as a prediction problem, and why the target it optimises for is easier to satisfy than to want satisfied
• feedback loops: how a feed’s accuracy and a feed’s influence become the same measurement
• what “for you” is actually claiming, compared to what it’s actually doing
• what the lab’s reverse-engineering exercise is trying to surface