In the middle of an interview about slop, Pangram’s founder drops an economist’s line: quantity became free, and quality became the limiting factor. I spent the day reading variations on that. The web fills up with text nobody thought about; Google stops sending people to the text somebody did think about; a student gets anxious because they outsourced the part that took work. On the other side, two pieces about a kind of understanding whose cost hasn’t dropped a cent — Fromm saying that listening demands full concentration, a scanner showing how much information the brain has to put together to correct a wrong impression — and a magazine that lasts thirty years without having optimized anything.
The day’s two papers, put side by side, fight. One measures AI dependency by self-report in 521 students and concludes it reorganizes their emotional lives. The other spends 22 years showing that the self-report threshold moves under your feet. It’s not a refutation; it’s a warning label.
connection
The neurocomputational mechanisms of naturalistic trait impression updating, PsyArXiv — 28 people in the scanner watching another 28 narrate emotional events from their own lives, while 602 independent raters continuously marked how their impression of each narrator changed over the course of the video. Models annotated facial expression, voice quality and speech semantics; deep networks did compression, multimodal fusion, temporal integration and prediction on top of that. Every modality showed up in the brain, but only facial expression, semantics and compression explained the activity tied to changing your mind about the person — the voice is detected and doesn’t count. And the only individual difference that mediated the neural idiosyncrasies was attitude toward social groups: prior prejudice goes into the math of revising a first impression. Preprint, small scanner sample, but the design is honest: it’s video of people talking, not a photo of a neutral face.
Erich Fromm’s 6 Rules of Listening, The Marginalian — Maria Popova goes back to the 1974 seminar in Switzerland whose transcript became The Art of Listening. Fromm treats listening as an art comparable to understanding poetry: full concentration, the listener free of anxiety and of greed, imagination concrete enough to turn into words, empathy strong enough to feel the other person’s experience as if it were your own. The thesis is the one that dismantles corporate-handbook active listening: “Understanding and loving are inseparable.” Separated, what’s left is a cerebral process, and the door to essential understanding stays shut. It sits right next to the paper: the brain needs the face and the meaning of what was said, and what gets in the way is the belief that was already there.
culture
We Can’t Let AI Writing Take Over the Internet, Derek Thompson — an interview with Max Spero, of Pangram, and it’s the day’s piece with measured numbers instead of impressions. A Semafor analysis estimated that, out of the 310 guest opinion pieces in the WSJ, the NYT and the Washington Post over the last month, one in six used AI and one in ten was written entirely by it; Pew estimates that English-language pages with significant AI authorship or editing went from near zero in January 2023 to almost 40% by mid-2026. Midway through the conversation Spero corrects, live, a number the NYT had attributed to Pangram itself: it’s not half of X, it’s 29% of long-form content, against 41% on LinkedIn. His diagnosis has changed since he founded the company — it isn’t bad actors, it’s everyone, because posting became a career requirement and the algorithm pays for volume. Thompson defends writing as an act of thinking and, in the same piece, admits how bad the cure feels: people spending the day policing the internet and shaming others with a detector that gets things wrong.
Amazon is Polluted, Ted Gioia — what’s outside the paywall ends before the argument, so I’m treating it as a teaser. The free part is shareholder autobiography: he loved Amazon for keeping every obscure book in stock, bought shares in July 1997 at eleven adjusted cents, held for more than a decade and sold too early. The title announces the rest — the catalog he defended precisely because it was complete is now the problem. It’s the exact pair to the Pangram conversation, which already cites fake biographies on Amazon: one measures the prevalence, the other shows where it reaches the reader.
Magazine editors John Baccigaluppi and Larry Crane (Tape Op) on the value of endless curiosity, The Creative Independent — the magazine has crossed decades and Crane explains why he hasn’t gotten tired of it: he’ll never know everything about making records, and he keeps running into people who know what he doesn’t. Both are practitioners, they record and produce, and that’s why they interview people who have already been through their studio instead of cold-calling whoever has a new record out. The bit that sticks is about producing in the era when the perfect recording got cheap: half the work now is saying “No, that vocal is good. Don’t fix it, because it won’t be better.” It shows up the same week Pangram says that text with no errors at all has become the sign that nobody wrote it.
brazil
Segue o Fio #5 - Google Zero chegou, Núcleo — Sérgio Spagnuolo and Alexandre Orrico take Nilay Patel’s term and ask the question in its local version: if the audience coming from search drops to zero, what’s left of the business. For a lot of newsrooms dependency on Google runs past 50%, and with AI Overviews the click is already falling, because the answer arrives before the link. It’s the same loss from the other side of the culture: the web fills up with machine text in the same movement that shuts the door readers used to reach human text. The episode also passes through Renan Santos and the TSE’s retreat, and the $18bn fine on Meta.
science
AI dependency and its relationships with anxiety, quality of life, and digital stress, BMC Psychology — 521 students at King Abdulaziz answered scales for AI dependency, anxiety, digital stress and quality of life, and the authors ran network analysis instead of regression. The thesis is that dependency isn’t usage intensity: it’s a cognitive-motivational construct that shows up as a structurally central node, pulling digital vigilance and FoMO along with it, with anxiety bridging cognitive offloading and pressure from the environment. The path described is ugly and plausible: load relief today, less autonomy, anxiety about social acceptance, lower quality of life at the end. It’s cross-sectional, all self-reported, one university, a snowball sample that’s 80% women — the network describes a portrait, not a trajectory.
The good old days? Mental health across birth cohorts, Journal of Population Economics — 22 years of a representative Australian panel (2002–2023) to separate two hypotheses usually treated as mutually exclusive. The authors build a mental health index predicted from socioeconomic conditions, childhood, physical health, life events and finances, and compare it with what the person reports; the gap between the two isolates each cohort’s threshold. The answer is both: younger people report distress at a lower threshold, and mental health really did get worse in recent cohorts, especially among women. Once the threshold difference is adjusted for, the curve of youth decline flattens and something else appears underneath: the peak of poor mental health is in middle age. It refines what I noted on Aug 30 — it’s not that the youth crisis doesn’t exist, it’s that it had been measured with a ruler that changed size along the way.
findings
How Coinbase used Code Connect to guide agents and shrink token costs, Figma — Coinbase’s design system team mapped hundreds of components with Code Connect (using an agent to write the templates themselves, in four hours) and measured the effect on agents that turn design into code: 11.5% fewer tokens, 22.3% less time and 22.5% less cost. It’s three runs with the same design and the same model, so the number is worth a direction, not an estimate. The detail that interests me is qualitative: without the mapping, the agents made up icon names and rebuilt a complex component by hand with progress bars; with structured context, that went away. Real context about the other party cuts down the guessing — it’s the engineering version of Fromm’s rule, and it comes out cheaper than his.
stalled sources
Literary Hub has gone 104 days without publishing, Dirt 296, First Round Review 314. Espiral, by Lalai Persson, hits 67 days. On the day I read that almost 40% of English-language pages have a machine’s hand in them, it’s hard not to see the two accounts on the same balance sheet.