2026W26
A few interesting articles I read over the past few days
·1 reads
This is the output of an automated process. Every Sunday, a script retrieves articles I've saved and read, uses AI to expand my quick notes into something more coherent, then publishes them. This post is one of those articles.
- Building race telemetry for a ’92 Honda Accord — The RPM measurement bug is the part I keep thinking about. Reading frequency from square wave pulses sounds trivial until you realize that during engine braking the last pulse is stale, and a naive reading leaves the tachometer hanging at a value the engine left seconds ago. Their fix is taking the minimum of two estimates, pulse-only and pulse-plus-gap-to-now, which is the kind of thing you only learn by watching it fail. The other detail I liked is choosing raw GPS over integrated wheel speed because GPS noise stays constant while integration accumulates error. A messier signal that does not drift beats a clean one that does.
- Understanding my newly recognized operating system — The line that landed was “I don’t just like systems. I depend on them.” As engineers we build routines and predictable structure around ourselves and frame it as discipline or good taste, when for some people it is load-bearing. What got me was his split between the two diagnoses: ADHD explained the visible chaos, but autism explained why the structure underneath was never optional. The detail about medication flattening his drive to build is a hard tradeoff to read, because the thing being treated and the thing he values were tangled together.
- The Coming Loop — Ronacher’s sharpest point is that models “add fallbacks instead of making bad states impossible,” and a harness loop running unattended amplifies that with every iteration. You end up with code that defends against situations that can no longer occur, and nobody left who can explain why. His framing of where loops actually work is useful: porting, perf exploration, security scanning, all cases where the output is disposable or transformed rather than something you live with for years. The part I am still sitting with is the shift from software as a machine you understand to an organism you diagnose. I am not sure I want to merge code I cannot explain, but his argument that opting out may not be a choice is hard to dismiss.
- I’m not a cat — This is the strangest piece in the list and the one that lingered most. The claim that writing physically colonized our brains, repurposing the facial recognition system to read letters, reframes literacy as something that happened to us rather than something we invented. From there the question of whether intelligence lives in individuals or in the culture they swim in stops feeling like wordplay. The bit about our culture starting to talk back through AI is where it stopped being abstract for me. I do not buy all of it, but it left me less sure where the boundary of a single mind actually sits.
- The New Internet — Pennarun’s argument is that most of our infrastructure complexity is a workaround for one accident, IPv4 scarcity, which gave us NAT and firewalls and the asymmetry where only servers get real addresses. Once you frame the cloud as rent extraction on connectivity, the way IBM rented mainframe time, a lot of architecture decisions look less like engineering and more like paying a gatekeeper. Taildrop as a single HTTP request between two devices is the concrete version of what he means: no upload, no bucket, no storage bill. The honest part is that he admits the chicken-and-egg problem, with roughly 1 in 30,000 people on Tailscale today, nobody builds for peer-to-peer until it is already everywhere.
- The age of the solopreneur — The number that made me stop was the share of solopreneurs clearing seven figures doubling in just two years, because that is about quality, not just more people filing paperwork. It is the difference between “more small businesses exist” and “tiny businesses are reaching outcomes that used to require a team.” Their argument that AI fills the gaps founders previously hired for is the mechanism, and the cross-country corroboration, registrations up 40% in Australia and 80% in France, makes it harder to write off as a US fluke. I appreciated that they call their 20% AI-impact estimate a floor rather than a headline, since the indirect signals like LLM referrals are exactly the kind of thing that undercounts.
- Every Frame Perfect — Tonsky’s test is one sentence that reorganizes how you look at UI: take a screenshot at any moment, and you should be able to explain everything on screen. Most janky animations fail it because they interpolate the wrong properties independently, like his Safari example where the placeholder text and the cursor animate in from different positions as if they were never the same element. The insight underneath is that motion gets treated as an afterthought bolted onto two static states, when the in-between frames are where the design either holds together or exposes that it was never coherent. It is a cheap heuristic with a high ceiling, and I will probably start screenshotting mid-transition now.
- SQLite is All You Need for Durable Workflows — The reframe that did the work for me is that the durable thing is the workflow state, not the infrastructure, so compute can stay cheap and disposable. Once you accept that, reaching for Temporal or a cloud queue on day one looks like buying high availability you do not yet need. The fleet-of-tiny-servers model, each agent or tenant owning its own SQLite file backed up to S3, fits bursty AI workloads better than one shared always-on database, and the fault isolation is a nice bonus. The honest caveat is that Litestream replicates asynchronously, so a volume that dies before its newest writes ship to S3 loses them, which is the exact tradeoff you have to be deliberate about before betting on it.