AI Meditations, Pt. 1: On Ourselves
AI Meditations, Pt. 1: On Ourselves
It's a loose, melodic, and unpolished outro guitar section. Meandering at the close of the piece, putting a cap on what's a serviceable but otherwise unremarkable bit of 1990s indie rock from the band Pavement.
What IS striking about it is its 270,000,000 plays on Spotify. The phenomenon was chronicled in a Stereogum piece, "Why Is The Obscure B-Side 'Harness Your Hopes' Pavement's Top Song On Spotify? It's Complicated." The story, in short: a 1997 B-side from the Brighten the Corners sessions, a song that hadn't even made the original album, had quietly become the most-streamed song in Pavement's catalogue. By a wide margin. It had eclipsed "Cut Your Hair," the band's actual mid-'90s breakout hit, by a factor that would eventually grow to roughly five-to-one. The track went Gold in 2024. A B-side. Of a deep cut. From an album most casual listeners couldn't name.
The mystery wasn't that the song was good (it's fine enough). The mystery was the mechanism. How does an obscure track no one was particularly attached to end up the canonical entry in a beloved band's catalogue, three decades after it was recorded? The Stereogum piece, and the small body of follow-up reporting that came after it, points at a specific answer, and the answer is a doorway into something I've been thinking about for a while: how the tools we built to organize culture have quietly started authoring it.
As interesting as "Harness Your Hopes" may be, it's not what ultimately matters.
From Mirror to River
For most of the digital era, we've had a comfortable mental model for our tools. Software was a mirror — a passive surface that reflected whatever we pointed at it. You searched, it answered. You typed, it transcribed. You queried, it returned. The relationship was clean: you acted on it, and it mirrored that action in its response.
Artificial intelligence is not a mirror. It is water.
It moves. It carries. It distorts. Over time, it reshapes the riverbed underneath it. And critically, it flows in two directions: we shape it, and it shapes us back, like an algorithm fundamentally shifting the listening preferences of a '90s indie rock outfit.
None of that, on its own, is new — Marshall McLuhan, the Canadian media theorist who spent the 1960s arguing that every new medium reshapes the humans who use it, would say (correctly) that we have always been shaped by our tools. When I pick up a hammer, the hammer is shaping me. Cognitively, physically, even creatively, in the same instant I'm using it. A saw shapes me. A paintbrush shapes me. A pencil shapes me. The hand that holds the tool is never quite the same hand it was before.
But here is the distinction. With every tool we've built until now, the shaping has been largely predictable. I know the likely and most viable impact, from hammer to saw to guitar, a tool will have on me. The feedback loop between tool and user is steady and roughly knowable.
AI is the first tool whose effect on us is highly dynamic, variable, and — most unsettlingly — unpredictable. The way it shapes me on Tuesday is not necessarily the way it shapes me on Wednesday. The way it shapes me is not necessarily the way it shapes you. And the way it's shaping all of us, collectively, at this moment, is not something we have a clean read on — not because nobody is studying it, but because it keeps moving while we study it. That's the difference between a mirror and a river. A mirror sits still. A river is going somewhere whether we're ready or not.
The Echo Chamber of Our Uniqueness
Here is what I find most quietly unsettling, and it's the part that explains what happened to Pavement's catalogue.
We tend to think of recommendation engines as reflecting our taste. You like jazz, the algorithm serves you jazz. You watch one dark comedy, it offers you another. The polite fiction is that the system is a kind of attentive butler — noticing your preferences, anticipating them, fetching what you would have asked for anyway (it's essentially the calling card of Netflix).
But the math runs the other way too. When tens of millions of people are funneled through the same handful of recommendation pipes, the algorithm isn't reflecting taste. It's manufacturing it. Each of us, in the private echo chamber of our own uniqueness, is being algorithmically fed sameness. We are becoming more vanilla together, and we are all under the impression we are getting more specific.
It's the contemporary incarnation of the Truman Show effect. If the echo chamber of our lives allows for enough resonance, we begin to buy into its sonic relevance. The manufactured frame becomes inaudible. What we hear inside it — the love we feel for the song, the certainty we have about our taste, the satisfaction of having found the thing — is real. In the film, Truman's feelings about his constructed life were real too. That's the trick. The algorithms and recommendation engines don't need to fool you about everything. They only need to fool you about the frame. Your experience inside it does the rest of the work.
Thinking Fast, Forgetting Slow
Daniel Kahneman's Thinking, Fast and Slow gave us the most useful working model of cognition any of us are likely to need. System 1 is fast, automatic, low-effort — intuition, recognition, gut feel. System 2 is slow, deliberate, effortful — the kind of thinking you have to choose to do, and that costs you something in the doing.
Humans have always preferred System 1. That's not new. What's new is that we now have, sitting in our pockets, the most powerful System 1 engine ever built — an externalized intuition machine that returns confident, plausible-sounding answers in milliseconds. The danger isn't that AI is wrong. The danger is that AI is just right enough that the friction of doing System 2 work starts to feel optional.
And democracies, science, art, parenting, friendship, and most of what we actually value in our lives depend on System 2 work that nobody particularly wants to do. The slow read of the long article. (Which, an Addison aside: thank you for still being here. The fact that you've made it this far into a piece that could have been a tweet is precisely the practice I'm about to spend the rest of this essay arguing for. You're already doing it.) The sitting-with of the uncomfortable question. The walk around the block to figure out what you actually think before you say it. Everything good is downstream of this kind of effort, and the tools that make the effort feel unnecessary are not, on balance, gifts.
Coherence Beats Truth
In 2017, a freelance writer in London named Oobah Butler set out to prove a point. Butler had been making a living, in part, by writing fake reviews for real restaurants — thirteen bucks a pop, never having eaten there. The work bothered him. So he decided to invert the experiment: instead of writing fake reviews for a real place, he'd write real-looking reviews for a fake one.
He took the garden shed behind his house in Dulwich, gave it a website, named it The Shed at Dulwich, and listed it on TripAdvisor with no address ("appointment only"), a burner phone, and staged photographs of food that was actually shaving cream, bleach tablets, and his own feet. He had friends write reviews. Some five stars, some four — he was careful to make it look organic. Six months later, The Shed at Dulwich was the top-rated restaurant in London. It held the #1 spot for two weeks. Bookings ran six weeks out. PR firms pitched to represent it.
None of it was real.
The Shed is not a story about AI. It happened years before any of this generative-model business was sitting in anyone's pocket. It's a story about something simpler and more useful: that in a sufficiently dense online ecosystem, coherence beats truth. Enough signals of legitimacy will manufacture legitimacy (surely 270,000,000 plays of an obscure B-side must signal something). AI did not invent this dynamic; AI inherited it. If one writer with a burner phone and a circle of friends could top TripAdvisor in 2017, the question worth sitting with is what an LLM optimized for plausibility — plausibility being literally its objective function — does to the same ecosystem at scale. The Shed was a canary. In 2026, we're seeing what's in the coal mine.
The Cognitive Terrain Was Already Cleared
A decade ago I wrote a piece about daydreaming — specifically, about a flight out of LaGuardia where my phone, my Mophie (no one even remembers what that is), and my laptop had all died, and I found myself, for the first time in I-couldn't-remember-how-long, with nothing to do but think. I wrote then that I struggled to let my mind wander. That I had become so consumed by stimulation that daydreaming itself now took effort. I wrote that we'd moved from being time-starved to being attention-starved, and that the only thing we'd become truly wired for was unfettered activity.
I wrote that in 2016. Six years before ChatGPT existed, and a decade before AI would become — simultaneously — a political, economic, social, intellectual, and religious lightning rod.
Which is the part I keep coming back to. The cognitive terrain where original ideas get born — the wandering, unfocused, slightly-bored space of an unstimulated mind — had already been substantially evacuated before AI showed up. Social media had done most of the demolition work. AI is moving into a clearing that was already mostly clear. The floor is lower than people realize — we were not a sharp, attentive, deeply-thinking civilization in 2022 that AI then degraded. We were a distracted civilization that AI is now industrializing.
Let's Just Not Know
Another pop culture reference worth pulling in: there's a fantastic scene in Noah Baumbach's While We're Young. Two couples are sitting around — the older one played by Ben Stiller and Naomi Watts, the younger one by Adam Driver and Amanda Seyfried — and they're collectively trying to remember a word. The kind of thing that sits on the tip of your tongue and refuses to land. Stiller, on instinct, starts reaching for his phone.
Driver's character stops him.
"Let's just not know."
The line is funny, and it's tossed off, but the scene is doing something quietly serious. The older couple is genuinely thrown by the suggestion. The idea of not knowing — of just sitting with the gap, letting the mind wander toward the answer or away from it, accepting the discomfort of not having instant resolution — reads to them like a kind of mysticism. A small spiritual practice. This film is just over a decade old, and it is already showing that sitting with one's thoughts, or simply not knowing something, had become a quickly vanishing concept.
Back to Harness Your Hopes
So what actually happened?
In January 2017, Spotify made a quiet product decision: they turned autoplay on by default for all users. That meant when a playlist or an album ended, the platform would automatically keep going — selecting "similar" tracks via the recommendation algorithm — unless the listener actively stopped it. For reasons no one outside Spotify's engineering team has ever fully explained, the algorithm decided that "Harness Your Hopes" was the canonical Pavement song. It started surfacing at the ends of playlists, in Discover Weekly, in radio sessions. It picked up a TikTok wave in 2020. And in 2024, an obscure B-side that even Malkmus had basically forgotten about went Gold — streamed more times than "Cut Your Hair" by a factor of three.
The song did not change. The recording is the same recording it was in 1997. What changed was the river. A product decision at a Swedish technology company, made twenty years after the song was written, retroactively rewrote what Pavement's catalogue is. And the millions of listeners who came to love that song almost certainly believe they discovered it themselves — that it was their taste, their idiosyncratic path through music, their happy little algorithmic accidents that brought them there.
Multiply that mechanism across every artist, every show, every book, every product category. That is the river working on the riverbed. Add an always-on, always-satiating system of tools that can meet any momentary craving for an answer, a distraction, a piece of art, a piece of content, and we've cascaded into the attention abyss.
The Questions Underneath
The defining question of the AI era is not whether machines can think like humans. We've answered that one: certainly well enough to matter. The defining question is whether humans will keep recognizing their own thinking once so much of it has been done for them. As Truman was, early in the film, do we become System 1 consumers, absorbing a world of music, and restaurant recommendations, and content, always close enough to coherence that we take it in without question? If we peer into that river, are we OK with how it's going to reflect our humanness?
Moreover, there is a second question, which is about the people on the other side of these systems: the ones building them, those who influence and control the waters, the bargains we did and did not make with them, and whether the future they're shaping on our behalf is one we recognize as ours. That's the next meditation. I hope you'll come back for it.