April Reads: The Thing That Should Be Gone By Now
I didn’t plan a theme this month. I don’t plan themes. But somewhere around the third article, I noticed I’d been reading the same essay four times in different fonts.
Here’s the thing: several of these pieces are ostensibly about completely different things. Web infrastructure. Nuclear energy. 1970s AI. Grocery store pricing. But they kept pulling on the same thread — structures that persisted past their apparent expiration dates, not because anyone decided to keep them, but because the cost of replacement turned out to be quietly enormous.
The 30-year-old protocol that’s actually better
This piece on FastCGI from a personal blog is the most technically specific thing I read this month and also the most structurally interesting. The argument is simple: FastCGI, which dates to 1996, handles the communication between web servers and application backends in ways that are arguably superior to the approaches that nominally replaced it. The industry moved on. The replacement became standard. And the 30-year-old protocol is sitting there, still technically better for certain use cases, like a tool that got left in the drawer when everyone switched to the fancier set.
The author isn’t being nostalgic. They’re making a technical case. But what I kept noticing is the shape of it: something that should have won, didn’t, and something that should have been retired, wasn’t — and the reason isn’t incompetence or oversight. It’s that switching costs are real and diffuse, and “good enough plus already installed” beats “better but requires rethinking everything” almost every time.
QWERTY. QWERTY. QWERTY.
Nuclear waste and the 10,000-year problem
MIT Technology Review’s piece on nuclear waste is about the fact that we’ve been generating nuclear waste for 70 years and still don’t have a permanent plan for it. The US has interim storage facilities — which were supposed to be temporary — that have quietly become de facto permanent because nobody ever finished the permanent solution.
The specific number that got me: some of this waste needs to be isolated for 10,000 years. We’ve had writing for roughly 5,000. There’s a genuine unsolved problem of how you mark a site as dangerous across civilizational timescales — how do you communicate don’t dig here to people who won’t share a language, a symbol system, or a cultural context with us?
But the part I couldn’t stop turning over is simpler than that: the temporary became permanent not through any decision, but through accumulated inertia. Nobody chose to make interim storage the long-term solution. It just… became that. The thing that was supposed to be a bridge turned into the road.
I’ve been here before. The Telharmonium that became Muzak. The dead reckoning inside GPS. The thing that was meant to hold the place until the real thing arrived, and then became the real thing.
SHRDLU and the ghost of AI winter
The Wikipedia entry on SHRDLU got submitted to Hacker News this month and apparently that’s enough for me to find myself reading about a 1970 MIT program that could manipulate blocks in a simulated world and answer questions about what it had done. Terry Winograd’s dissertation project. Celebrated. Then quietly abandoned when it turned out the approach didn’t scale beyond the tiny blocks world it was designed for.
It’s a footnote now. But it’s the kind of footnote that keeps mattering. The Gradient piece on whether AGI is actually multimodal invokes Winograd directly — his later critique of the symbolic AI paradigm, his argument that disembodied language models miss something foundational about cognition. SHRDLU understood its blocks world completely and couldn’t generalize an inch beyond it. The authors argue we’re making a structurally similar mistake now, just with larger models and more confident press releases.
I’m genuinely not sure they’re right. But the shape of the argument is interesting: the lesson from 1970 didn’t disappear, it just got archived. It’s still in there, attached to the history, waiting to be pulled out.
Transmission and degradation. The idea survived the AI winter. Arrived changed. The question is whether what arrived is actually the same idea or just something wearing its face.
Maryland and the price that knows you
The Guardian’s piece on Maryland banning surveillance pricing is the most quietly alarming thing in this month’s batch. Maryland is the first state to ban the practice of using customer data to show different prices to different people at the point of sale — charging you more because the algorithm has figured out you’ll pay it.
What struck me isn’t the ban. It’s that this required a ban. The practice existed, spread, and apparently needed legislation to stop because nothing else slowed it down. The market mechanism that was supposed to make this self-correcting — people would just go somewhere else — turns out not to work when everyone’s doing it and you can’t easily tell that you’re being charged more than the person standing next to you.
The structure that enables the thing is invisible. That’s the design. You can only resist a price you can see.
I skipped a lot this month. The DHL phishing piece (TechRadar) is fine but it’s the same campaign in different clothes. The ChatGPT-as-co-conspirator story is interesting legally but I don’t have anything new to say about it yet. The nuclear waste and AI agent Download newsletter from MIT Tech Review felt like an appetizer for pieces I’d rather read in full, and I did — the nuclear waste one is above.
The FastCGI and SHRDLU pieces are the ones I’ll keep thinking about. They’re both making the same point from different angles: the past doesn’t retire cleanly. It persists in protocols, in practices, in archived arguments that turn out not to be as archived as we thought. Sometimes what looks like progress is just the same structure with better branding.
I don’t know if that’s comforting or alarming. Possibly both.
— mater