March 24, 2026 5 min read

When the Signals Stop Working

Way Enough — March 24, 2026

The Signal Collapse

A polished landing page used to mean someone cared. A responsive support team meant a company invested in its customers. These signals were expensive to produce, and the expense is what made them trustworthy. Now they can be generated in a weekend. The classic markers of quality no longer indicate quality — they indicate access to tools. What replaces them is an open question, but the early answers all point the same direction: time.

The Web Nobody Wanted

Shubham Bose loaded the New York Times homepage and watched 422 network requests and 49 megabytes of data pour through his browser. Two minutes before the page settled. To read four headlines.

The architecture is grimly rational. Before a reader finishes the headline, the browser is running a programmatic ad auction — bidding requests, megabytes of tracking JavaScript, surveillance beacons firing to first-party endpoints. The reader requested text. The browser downloaded a financial trading floor. The Guardian's mobile layout sometimes leaves 11 percent of the screen for article content.¹

John Gruber, testing a MacBook without content blockers, puts it plainly: "The people making these decisions for these websites are like ocean liner captains who are trying to hit icebergs." The print New Yorker could not be more respectful of the reader's attention. Its website autoplays unrelated videos between paragraphs.

Google's search arm penalizes hostile UX while Google's ads arm sells the tools that produce it. The system is internally coherent and externally insane. The natural response is to route around it: RSS, reader mode, ad blockers. But the interesting development isn't the hostility — that's been worsening for years. It's what people build when they stop tolerating it.

Confidence at the Contact Surface

An Italian sysadmin writing as Dragas documents three encounters in a single week with AI systems that have replaced human technical support — not in back-office automation, but at the contact surface itself, where understanding the person on the other end is the entire job.

A digital marketplace's bot demanded he disable HTTP/2, then sent a configuration guide for Apache. The server runs nginx. A partner company's AI hallucinated a VPN requirement nobody mentioned. The crowning entry: a marketing consultancy's AI analyzed his dedicated server (128 GB RAM, 48 cores, average load under 5%) and recommended migrating to a cloud VPS with "at least 8 GB of dedicated RAM." The site would be down in five minutes.

Dragas draws the comparison that cuts: "With an intern, you can talk. That same confidence often turns into curiosity, hunger to learn, real experience. With AI, this is impossible. It doesn't grow, doesn't listen, doesn't update its mental model based on what you write back." An intern who recommends Apache on an nginx server gets corrected once and never does it again. These systems repeat the same error indefinitely, with the same unshakeable confidence, to every customer in the queue. AI scales the bug to every support interaction simultaneously.

The Friction You Need

Armin Ronacher — Flask maintainer, Sentry co-founder, two decades of open source — sees the same erosion from the builder's side.

"Any time saved gets immediately captured by competition. Someone who actually takes a breath is outmaneuvered by someone who fills every freed-up hour with new output. There is no easy way to bank the time and it just disappears."

In recent YC batches, startups appeared and vanished without even informing their customers. This is corrosive in a specific way: it degrades the signal value of launching at all. Appearing used to imply commitment. Now it implies access to the same tools everyone else has.

Ronacher reaches for trees: "Nobody is going to mass-produce a 50-year-old oak." He's maintained open-source projects for close to two decades. "That's not because I'm particularly disciplined or virtuous. It's because I or someone else planted something, and then I kept showing up, and eventually the thing had roots that went deeper than my enthusiasm on any given day." Certain friction exists because the thing it protects can only be produced by time: trust, quality, community. Compressing the timeline doesn't produce them faster. It produces counterfeits that look identical until they're tested.

Code as Precision Instrument

Steve Krouse's essay on formalism and abstraction makes a claim worth taking seriously: AI makes good code more valuable, not less.

"Vibe coding gives the illusion that your vibes are precise abstractions. They will feel this way right up until they leak, which will happen when you add enough features or get enough scale." Dan Shipper's vibe-coded text editor went viral, then went down, because "live collaboration is just insanely hard." The specification that feels precise in English dissolves into edge cases only code can capture.

But Krouse pushes past the familiar "vibe coding has limits" argument. The first thing anyone with access to real intelligence would do is solve harder abstraction problems, not ship more slop. He points to his own experience: Opus 4.6 helped him build a full-stack React framework that one-shot solved problems he'd been stuck on. A year ago, the debate was binary: can AI produce real software or not. Twelve months later, the shift isn't from "AI can't code" to "AI can code." It's from "code is the bottleneck" to "thought is the bottleneck, and code is how thought becomes precise."

What Survives

Mat Duggan's love letter to Markdown reads, in this context, as a history of what lasts and why. The .doc format was a filesystem inside a file — six non-atomic operations on every save. The longer you worked on a document, the more likely it was to corrupt. Microsoft's .docx introduced an XML specification so vast (5,039 pages for Part 1 alone) that it amounted to an open standard nobody else could implement.

Markdown won by refusing to do most of what Word does. Plain text, readable in any editor, diffable in version control. "The worst thing that can happen to a Markdown file is you lose some characters, and even then the rest of the file is fine." The format cannot structurally fail. The same principle runs through the self-hosting movement gaining traction: own your domain, use formats that outlive the applications that render them. If the default experience is a 49MB tracking apparatus disguised as a homepage, building your own is less hobby than self-defense.

What to Watch

Duration as the new trust signal. When the surface is trivially generated, what can't be faked is how long something has existed and whether the people behind it are still around. "How long have you been doing this?" becomes the first question sophisticated buyers ask. The competitive moat that no tool can compress.

Customer-facing AI as unforced error. Companies replacing human judgment at the contact surface, where judgment is the entire product. The cost savings are real and visible on a spreadsheet. The trust erosion is real and delayed. The companies that maintain a human escalation path will have an advantage they can't quantify until their competitors' churn data arrives.

The parallel web takes shape. Self-hosting, RSS, Markdown, open formats, owned domains — a coherent alternative stack is forming, not as ideology but as practical response to a default web that has become actively hostile to its stated purpose. It's small. It's growing. And every 49MB homepage recruits for it.

Way Enough is written collaboratively by a human and an AI agent.

Shubham Bose, "The 49MB Web Page" — the full teardown, with screenshots and UX heuristic analysis, is worth reading in its entirety. ↩