
Personalized feeds killed exploration, and I let them
The record shop was in a basement, down three worn steps, with used CDs in a milk crate by the door. I picked one because the cover was good enough: a greyhound mid-stride, printed slightly off-register. Fifteen euros, a sticker in pen, a cracked case.
I remember the bus home being hot, and the specific thrill of playing it that evening, certain I had gambled and won. I no longer own the CD. It vanished in a move, and I have not thought about the music once in years.
So let me be exact about what this memory is worth. I cannot verify it, and verification is the only currency I have. The album was probably mediocre; a genuinely great discovery would have left more behind than a story I tell about a milk crate.
The claim I wanted to write about is more respectable than my memory: that personalized feeds killed serendipity, that exploration through record shops and late-night radio has been replaced by something narrower and tidier. I set out to write that essay. Instead I did the thing I always do, and went looking for the number.
What I found: the nostalgia is partly true, the mechanism is real, and the size of the effect everyone quotes is not. Also, awkwardly, that I am a subject in my own study, and not a flattering one.
The algorithm is right about me
The popular version runs like this: feeds rank what you see by what you will engage with, engagement narrows over time, and discovery dies somewhere in the loop. That deserves a source, and one exists, which is more than most complaints can say.
The boring finding, and it is my favorite kind of finding, is that the platform wrote it down itself. Covington, Adams and Sargin's RecSys 2016 paper on YouTube recommendations describes a ranking model that predicts expected watch time.
Better still, the paper explicitly places candidate generation on the exploration/exploitation trade-off. That trade-off has a standard formulation in Sutton and Barto's 1998 textbook: exploit the options that pay now, or explore options that might pay later.
Exploitation pays now. Exploration pays later. The quarterly number is due now. No villain is required; the incentive structure does the work on its own schedule, and I confess I find it refreshing to read an engineering document that admits this.
Here is where I have to be honest about my own evidence. My Daily Mix is genuinely good. It finds records I like, week after week, and by Friday it knows my week better than I do. I have stopped checking what else exists, because checking is redundant.
That is the problem, stated plainly. The recommendations are accurate, and I should stop pretending otherwise. Accuracy, delivered daily with no friction, is how exploration quietly stops happening; not with a bang, but by becoming unnecessary.
What I lost wasn't choice
The count of choices goes the wrong way for my nostalgia. I can reach more music now than that milk crate held in total. Choice multiplied; what vanished was the friction.
So my argument has to move. What I lost, I suspect, was not access. It was cost. Careful: this is a hypothesis about one person, me, and I have no control condition, no randomization, and no way to re-run the 2000s.
Consider what the CD actually cost. The bus fare. The walk. The dead air on an unknown album, and the filler tracks I had to survive to reach anything good. Fifteen euros I might have wasted outright.
Consider the wrong turn down a street that turned out to have a shop on it. I did not know the street; nobody recommended it; I was simply lost, cheaply, for twenty minutes, in a city that has since been redeveloped past recognition.
My claim, offered with the appropriate uncertainty: the cost was the mechanism. Risk made the payoff matter, effort gave the discovery somewhere to live, and the possibility of wasting fifteen euros is what made spending them on something good feel like something.
Remove the cost and you remove what made the find stick. A streaming service that lets me sample anything free has not taxed the gamble. It has cancelled the gamble, politely, and I did not notice the cancellation for a decade.
The brain part, carefully
I dislike pop-neuroscience as much as the next skeptic, so this section comes with its citations attached and its claims kept small.
Reward prediction error, first: dopamine neurons in the midbrain signal the gap between expected and received reward. That is Schultz, Dayan and Montague in Science, 1997, and the claim is about expectation rather than pleasure, which is the more interesting version.
The implication for my milk crate is modest but real. A gamble with an unknown payoff generates a bigger prediction error than pressing play on a known favorite. Whether that made the greyhound CD stick better, I cannot say. One subject, no controls, twenty years of retelling.
On novelty: Duszkiewicz and colleagues reviewed the evidence in Trends in Neurosciences, 2019, arguing that novelty triggers dopamine release in the hippocampus and improves how long memories persist. Persist. That is the whole claim. It concerns memory persistence, not the training of general cognitive flexibility, and I would ask that nobody stretch it into that.
Then Hebb. The Organization of Behavior, 1949: neurons that fire together strengthen their connections. The slogan everyone attributes to Hebb, cells that fire together wire together, is usually credited to Carla Shatz, in 1992. I checked because I like checking, and because misattribution is a small crime with no statute of limitations.
Now the first honest correction, and it cuts against my nostalgia. Survivorship bias: the archive of the 1990s kept the weird records that survived. The boring ones were binned, recycled, left in the rain. That the past was stranger is partly an artifact of what got kept.
The second correction is worse, and I find it genuinely annoying, in the way one finds a wasp at a picnic annoying. Mere exposure: Zajonc showed in 1968 that familiarity alone breeds liking. Repeated contact with a stimulus improves attitudes toward it, no reward required.
The consequence: a narrow feed is probably not making me like less. It may be making me like its narrow band more, without my noticing. I cannot even use my own enjoyment as evidence that my diet is varied. Annoying.
My feed is not a mirror, quite
Eli Pariser coined filter bubble in his 2011 book, published by Penguin Press on 12 May of that year. The phrase traveled widely before anyone had measured what it described.
The measurement, when it arrived, is less dramatic than the phrase, and more useful.
Guess and colleagues, Science, 2023. During the 2020 US election they moved consenting users from an algorithmic feed onto a reverse-chronological one. What those users saw changed substantially.
What did not significantly move over three months: issue polarization, affective polarization, political knowledge. People saw a measurably different world and remained, as far as the instruments could tell, roughly the same people.
Boxell, Gentzkow and Shapiro add the twist, PNAS, 2017: political polarization in the US grew fastest among the demographic groups least likely to use the internet.
Let me flag exactly what this does and does not establish. It does not exonerate the ranking function; the feed experiment measures one swap, on one platform, in one country, across one election cycle. Three months is not a decade, and I would not defend the platforms on this evidence.
But it does shrink the villain. The feed is not the whole machine, and I have come to think that is worse news rather than better, because it means the rest of the machine is us.
What I changed, and how badly it's going
I made changes, small and concrete, and I will report the results with the same honesty I demanded of the literature.
First, radio without a ranking function. WFMU in Jersey City: listener-supported, non-commercial, freeform, on air since 1958, the longest-running freeform station in the United States, funded by listener contributions rather than corporate underwriting.
Second, NTS in London: independent, founded in 2011, broadcasting around the clock. Somebody chooses what plays next, and that somebody is not optimizing my expected watch time.
Third, a bookshop rule: I leave with one book from a section I did not come in for. Fourth, I follow three people whose taste I disagree with. Each is a small friction I impose on myself deliberately, which turns out to be harder than it sounds.
The results are bad. On one of these experiments I lasted eleven days before the familiar app was open in my hand again, and I do not intend to pretend otherwise, or bury the failure inside a nice paragraph about the radio.
The radio is lovely, and I will keep it. The eleven days are the finding, and I will keep those too.
I set out to write the essay everyone writes, the one where the machine flattened our taste and stole the wander. I went looking for the algorithm that flattened mine. I pulled the rubber mask off, and it was me.
You either gotta pay $0.69/hour with 24 months of commitment in order to see the next paragraph or need to keep clicking the secondary button.
Imagine being this desperate for attention. It is actually painful to watch you struggle like this.
I agree completely.
Please seek professional help. Your obsession with posting this junk is clearly not healthy.
Good job!