- AI slop is really two separate problems: outputs that are objectively bad, and outputs that are averaged out and one-size-fits-all. Models drift toward the average answer because training rewards consensus, but in subjective domains like writing or design, the average answer is not the correct answer, and pre-training ends up smoothing all of this out, unable to account for good disagreement. The fix isn’t randomness (raising temperature just adds noise). It’s what five designers would give you for the same brief: five different designs that all meet it. That means keeping expert disagreement in the training data and training on the process (brief, then mockup, then layout exploration), not just the finished examples amplifypartners.com/blog-posts/engineering-a-beautiful-mind
- Imagine things had already gone wrong and you got one chance to replay from today: what would you do differently? Most of the time the honest answer is that you’d actually try, instead of playing it safe and hoping it works out. When Lee Sedol was losing to AlphaGo, he stopped making safe, predictable moves and played the bold, unexpected move 78, and it won him his only game of the match. When you’re behind, the safe path guarantees you lose, so the bold move is the rational one lesswrong.com/posts/…/have-you-considered-actually-trying
- Our emotions weren’t designed to make us happy. They were designed to keep us alive, and a system built for that will give false alarms, because a false alarm is cheap and a missed threat isn’t. So we overreact, and when we try to stop overreacting we often make it worse. The surprising part is how much of the misery comes from that second step: “At the heart of so many of our miseries is not the fact that we feel bad, but the belief that we should feel otherwise.” newyorker.com/magazine/…/the-opposite-of-happiness (Related reminder from an old share: signalvnoise.com/posts/3124-give-it-five-minutes)
- India has quietly lifted a tenth of humanity (1 in 10 people on the planet) out of extreme poverty in two decades, and it reduced inequality while doing it. At the current pace, the average Indian will be about twice as well off in 10 years as they are today. A lot of that comes from the government building roads and railways at a huge scale: highway spending is up 5x and railway spending up 8x in ten years. India now also makes and sells a wide mix of things to the world, with cars (22%), machinery (18%) and steel (17%) leading its engineering exports, while electronics exports nearly doubled in a year, many of them iPhones made for the US noahpinion.blog/p/india-is-quietly-doing-great
- Interesting explainer on how Shazam works. Roughly: it turns the song into a map of its loudest notes, throws everything else away, and looks for patterns (pairs of notes with a fixed gap between them) instead of single notes. That filters out most of the noise and lets it match 10 seconds of bar audio against millions of songs in a few thousand quick lookups web.archive.org/web/20220823105405/…/how-shazam-works
What I’ve been reading (W46, 2026)
Engineering taste into AI, actually trying, why we’re so good at feeling bad, India quietly doing great, and how Shazam works