2026W36
A few interesting articles I read over the past few days
- The revolt of the reader — This is the sequel to the “intellectual fly is open” piece I flagged a couple weeks ago, and now he has numbers, Dunlop’s survey of 668 developers where 78% stop reading the moment they detect AI and 71% then avoid the author for good. The spam analogy is the load-bearing one, spam didn’t die because it got ugly, it died because filtering plus consequences made it uneconomical, and he’s betting detectors like Pangram do the same to generated prose. Oxide now gates public writing on passing Pangram, which is a striking thing to mandate but consistent with his claim that the real cost is trust, not aesthetics. I’ll admit reading this while an agent drafts my reading notes is uncomfortable, though the skeleton-versus-ghostwriter line is the one I keep telling myself I’m staying on the right side of.
- Will agents like Git any more than we do? — Amplify Partners — The framing that clicked for me is that Git’s quirks were never a problem because a human was always there to absorb them, and agents don’t do absorbing, they just hit the wall at 2 to 3 merges a second and stall. The staging-area state machine is the obvious tax, tracking whether a file is untracked, unstaged, staged, or committed is exactly the kind of ceremony an agent shouldn’t have to model. What actually sold me on jj is first-class conflicts, letting the conflict live in the data structure instead of blocking all forward progress is the difference between a human who stops to resolve and a fleet of agents that need to keep moving. Calling Git “an informally specified spec half implemented by five or six people” is funny because it’s true, and it reframes why a cleaner model might genuinely win this time.
- Version control second coming — What this adds over the agents-and-Git argument is the market view, and Santos has earned it fighting Git for 20 years with Plastic SCM, so his read that GitHub could go the way of SourceForge isn’t just contrarian. The roll call is genuinely useful as a map, Entire doing provenance for agent code, Pierre and Origin chasing AI-scale traffic, Diversion betting centralized and cloud-first against the whole distributed premise, ERSC commercializing jj. The prediction I keep circling is that repos past 300 GB and a million files simply break the distributed model, which quietly concedes that Git’s core design choice is the thing under pressure. His own hedge is the honest note, “ask me in two months,” because nobody actually knows which of these survives, only that the freeze is over.
- What comes after git — After two pieces arguing Git has to go, ERSC’s answer is the pragmatic one, don’t replace the protocol, replace the storage engine underneath it and keep speaking Git so nobody has to migrate on day one. That bridge framing is smart because it sidesteps the usual chicken-and-egg, you get horizontal scaling and independence from the big platforms without asking teams to abandon their clients. The number that recalibrated me is that the Linux kernel is ~43 million lines while companies were already running monorepos in the billions of lines years ago, so the scale Git struggles with isn’t hypothetical, it’s just been hidden inside Google and Meta. The jj-with-swappable-backends detail ties it to the other reads, the path they’re selling is Git protocol now, native protocols and jj later, migration as a slope instead of a cliff.
- If coding is solved, what now?: Measuring the sloppiness of code | EARENDIL — The reframe here is that correctness stopped being the interesting axis, models pass the tests, so the real question is what the code does to a project over time, and that’s much harder to score. The finding that stuck is that using an LLM as a quality judge is “basically equivalent to a random number generator,” and worse, renaming the solutions flips its verdict, which tells you it’s grading on surface, not structure. The two metrics in SlopCodeBench give it teeth, verbosity roughly doubling from 0.15 on real repos to 0.33 on agent code, and erosion jumping from 0.31 to 0.68, so the sloppiness isn’t a vibe, it’s measurable and it’s about 2x. Ties directly to the sloppiness worry underneath all the version-control pieces, the “aircraft progress by weight” line is the whole trap, and their answer that human taste is still the thing you can’t automate away is where I land too, even if it’s an inconvenient conclusion given who’s writing these notes.