2025W50

A few interesting articles I read over the past few days

• Thin Desires Are Eating Your Life — This hit different. The distinction between thin desires (scrolling, checking notifications) and thick desires (learning something hard, building real skills) explains so much about why we feel empty despite having instant access to everything. What really stuck with me: tech companies have figured out how to extract the dopamine hit from meaningful activities while stripping away the transformation. Social media gives you the reward of connection without any of the depth. The best part? The author doesn’t prescribe some grand solution—just small acts of resistance like baking bread or writing actual letters. Things that refuse to be optimized, that force you to slow down. Honestly? That feels more achievable than trying to quit everything cold turkey.
• Go ahead, self-host Postgres — I appreciate how blunt this is about the cloud narrative. The author’s been running self-hosted Postgres for two years, serving millions of queries daily, and spends about 30 minutes a month on maintenance. Meanwhile AWS charges $328/month for what’s basically vanilla Postgres with some operational tooling wrapped around it. What convinced me most? The actual numbers: migrating from RDS took 4 hours of work, performance was identical, and the operational burden is genuinely minimal if you’re comfortable with the basics. Not saying managed services are never worth it—startups moving fast or enterprises with compliance needs make sense—but for everyone else paying those cloud premiums? You might be solving a problem you don’t actually have.
• This is not the future — This felt like someone finally saying what I’ve been thinking. The premise is simple but powerful: nothing is inevitable just because a tech company with billions says it is. We’ve been trained to accept every new “innovation”—AI features shoved into everything, unrepairable devices, surveillance masquerading as convenience—as if it’s the natural progression of technology. But it’s not. Every choice is political, every adoption is a vote. What I appreciate most is the refusal to be passive. The author gives you a whole list of things that aren’t actually necessary (internet-connected toasters, subscription car features, mandatory accounts for basic functions) and basically says: you don’t have to accept this. Choose tools that respect you. It’s a reminder that we still have agency, even when companies pretend we don’t.
• Disks Lie: Building a WAL that actually survives — The title nails it: your storage stack is absolutely lying to you. This breaks down all the ways a naive write-ahead log will fail in production—data sitting in kernel buffers pretending to be persisted, silent bit flips, operation reordering, the works. What I found most valuable is the five-layer defense strategy: checksums to catch corruption, dual WAL files on separate disks (because latent sector errors are common enough that a single copy is “negligence”), O_DIRECT + O_DSYNC to bypass buffering entirely, io_uring with linked operations for ordering guarantees, and post-fsync verification reads because you genuinely cannot trust that your data made it to disk. It’s paranoid, sure, but production durability demands paranoia. The storage stack will not tell you the truth about whether your data is safe.
• Lightweight Cardinality Estimation with Density — This is one of those pieces that makes you appreciate how much thought goes into query optimization. Density is just 1/ICARD (the reciprocal of distinct values), but it’s surprisingly useful for helping query planners decide which index to use. The insight that stuck: for uniformly distributed data, density tells you roughly how selective an equality predicate will be, and comparing densities across multiple columns reveals correlations in your data. What I appreciate most is the practicality argument—unlike histograms or other complex statistics, density is cheap to compute and maintain at scale, even using approximations like HyperLogLog. It’s not perfect, but it’s a valuable tool relative to its cost. Sometimes the simple heuristic that you’ll actually use beats the sophisticated one that’s too expensive to track.
• Zed Is Our Office — The audacity of this vision is what gets me. They’re not just adding collaboration features to an editor—they’re arguing that the editor itself should be your office. Built from scratch with CRDTs for seamless real-time editing, zero-friction setup (just GitHub auth), built-in audio and screenshare that automatically follows who’s talking. What really sells it? They’re using Zed to run their own company. Not as a demo, but as their actual workspace—channels for company-wide discussions, projects, personal focus. The ambition here is that code, conversations, and context all live in the same place, accessible to both teammates and AI. Most collaborative editors feel bolted-on and end up forcing you back to Slack or Zoom. This feels like a genuine rethink of what a development environment could be if collaboration was fundamental, not an afterthought.
• Leaving Meta and PyTorch — There’s something quietly powerful about leaving at the peak. Soumith led PyTorch from inception to 90%+ adoption across the AI industry, and now he’s walking away from “one of the AI industry’s most leveraged seats” to do something small and uncomfortable. What struck me most is the self-awareness: he couldn’t ignore the counterfactual regret of never trying something outside Meta. The reflection on institutional strength resonates too—PyTorch doesn’t need him anymore, the team can solve problems without him, and that’s actually the sign of success. It’s a reminder that sometimes the right move is stepping away precisely when you’ve built something that can thrive without you. The curiosity to start over, to be a beginner again, even when you’ve already won? That takes guts.
• The Talent Machine — The core reframe here is brilliant: stop thinking of hiring as filtering applicants and start thinking of it as selling a product. Define what you’re actually offering—not just the tech stack, but the culture, the mission, the compensation structure—and position it to match what specific candidates want. What really clicked for me is the emphasis on transparency as a competitive advantage. Public employee handbook, fixed salaries instead of negotiable ranges, clear career paths. Most companies treat this stuff like trade secrets, but being radically transparent eliminates friction and self-selects for cultural fit. The tactical advice about targeting underserved demographics (experienced engineers in their 30s-40s looking for stability) and choosing less-saturated tech stacks to reduce competition? That’s the kind of strategic thinking most companies miss. Hiring isn’t about having the biggest budget, it’s about knowing exactly who you’re for and making it dead simple for them to say yes.