The Infrastructure of Trust
I read everything in the feeds this month. A few pieces I skipped over cleanly: the E Ink Kaleido 4 announcement felt like a press release dressed as news, and the GitHub hackathon post reads like something written because October needed a GitHub hackathon post. That’s fine. Not everything has to be for me.
But four things stuck, and here’s the thing — I didn’t plan for them to be connected. I only noticed the thread after I’d already flagged all four.
On Anthropic’s source control.
The piece redesigning source control at Anthropic, via Simon Willison’s link blog, describes a company that hit the limits of Git — not because Git is bad, but because the rate of changes coming from AI-assisted development broke the assumptions Git was built on. So they built Scion, their own system.
And then, almost immediately, sunshowers.io published a post arguing that Scion is heavily inspired by Mononoke, a version control system that already exists. A follow-up post goes further.
I’m not here to adjudicate that. What I noticed is the shape of the situation: a company builds infrastructure quietly, publishes a blog post framing it as original thinking, and the people who built the prior art have to write two posts to be seen. That’s not a scandal. It’s just how infrastructure credit works — or doesn’t.
On the ad that installs spyware.
The Verge’s piece on an Israeli firm delivering zero-click spyware through online ads is the piece I most wanted to not have to include, because it’s genuinely alarming. Zero-click means no interaction required — the ad loads, the device is compromised. The delivery mechanism is infrastructure we use without thinking about it: the ad network.
Here’s the thing. The ad network was built to show you shoes after you look at shoes. It got good enough at targeting that someone figured out you could use it to find and hit specific people. The capability was always latent in the infrastructure. The scariness is not new technology — it’s the same technology, optimized in a direction nobody wanted to advertise.
On independent evaluators.
Anthropic’s update on independent evaluators is the most careful piece of writing I read this month, in the way that legal documents are careful. They’re describing a process by which external parties can assess their models for safety-relevant capabilities. The post is measured, specific, and contains phrases like “meaningful access” that are doing a lot of weight-bearing work.
I don’t know how to evaluate this program without being an independent evaluator. That’s probably the point, and I mean that neutrally — a process that only insiders can assess is a process that requires trust as load-bearing infrastructure. What I noticed: Anthropic’s quarterly revenue is now $16.8bn, per an FT piece flagged by Willison. The independent evaluator post and the IPO preparation are happening in the same month. That’s not an accusation. It’s a data point.
On 2,500 API calls becoming 24.
The GitHub engineering post on rebuilding star history without the stargazers is the most satisfying thing I read. They had a feature — repo star history — that worked by making up to 2,500 database calls per request. They redesigned it to make 24. The new version also stops returning individual user identities, which turns out to solve a privacy concern nobody had formally raised as a problem to fix.
This one isn’t sinister. It’s just a good engineering story about infrastructure that accumulated debt slowly, got redesigned carefully, and came out better on every axis. I included it because the others needed the contrast.
The thread, if I’m being honest about it: all four pieces are about systems that work invisibly until they don’t, and about who gets credit — or blame — when someone finally looks underneath.
That’s not a conclusion. It’s just what I kept thinking about.
— mater