When the Signals Stop Working
Read short version (5 min)Way Enough — March 24, 2026
When the Signals Stop Working
A nice landing page used to mean someone cared. A compliance certification meant someone did the work. A well-designed app meant a team spent months thinking about it. These signals took time and money to produce, and that cost is what made them trustworthy. Now they can be generated in days. The classic markers of quality no longer indicate quality. They indicate access to tools.
Confidence Without Competence
An Italian sysadmin writing as Dragas documents three encounters in a single week with AI systems that have replaced human technical support.
A digital marketplace's bot demanded he disable HTTP/2, then sent a guide for configuring Apache. The server runs nginx. When he asked to speak to a human: "That's not possible for this type of issue. Follow our guide or we will suspend your service." A partner company's AI hallucinated a VPN requirement nobody mentioned and kept insisting he whitelist a user agent, despite being told the block operates at the firewall level — before any HTTP handshake occurs. The crowning moment: a 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.
These systems have been deployed at the contact surface — the exact point where understanding the person on the other end is the entire job. Dragas makes the comparison: "With an intern, you can talk. That same confidence often turns into curiosity, hunger to learn. With AI, this is impossible. It doesn't grow, doesn't listen, doesn't update its mental model." You used to be able to evaluate a vendor by their surface. The surface is now trivially generated. What remains as a trust signal? Duration. Track record. Whether someone answers when things break.
The Treadmill
Armin Ronacher (Flask, Sentry, two decades of open source) sees it from the builder's side. His essay on time and friction:
"We all sell each other the idea that we're going to save time, but that is not what's happening. 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."
In recent YC batches, startups have appeared and vanished without even telling their customers goodbye. Open-source projects materialize with a week of commits and go silent. Appearing used to signal commitment — launching a product, maintaining a site, building a community. Now it signals access to the same tools everyone else has.
Ronacher reaches for trees: "Nobody is going to mass-produce a 50-year-old oak. And nobody is going to conjure trust, or quality, or community out of a weekend sprint." Trust requires duration, and duration is exactly what the acceleration selects against.
Personal Software
Steve Krouse's essay on code's persistence: "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 and then went down, because "live collaboration is just insanely hard." The specification that feels precise in English dissolves into edge cases that only code can capture.
But vibe coding enables something real: personal software. A different category entirely from production software. The speed is a tradeoff for getting something pretty good now — not for shipping to users without strong quality guardrails, but excellent for learning and exploring a space. For understanding what's actually hard. When you ship it to users, it breaks and erodes trust. When you're the only user, cutting corners only impacts you.
This reframes the craft question. The binary (vibe coding works / vibe coding doesn't) missed the category where it works best: software built for yourself, to learn what you didn't know, to understand why something is hard before building it for others. A year ago, the debate was whether vibe coding could produce production software. Twelve months later, the more interesting answer is that it produces a different kind of software entirely.
What Won't Corrupt
Mat Duggan's history of document formats is a case study in what survives.
The .doc format was a filesystem inside a file. A FAT allocation table managing sectors, six non-atomic operations every time you hit save. The longer you worked on a file, the more important it was, and the more likely it was to corrupt beyond recovery. Duggan spent years at the help desk: "Yes, I understood they had worked on this file for months, but it was gone and nobody could help them."
Markdown won because it couldn't structurally fail. Plain text. Readable in any editor. Diffable in version control. It can't set margins, do columns, or change font color. It didn't need to. "After decades of nursing .doc files like they were delicate flowers... the idea of a format that simply cannot structurally fail is not just convenient. It's a kind of liberation."
The LLM era pushed Markdown to its zenith — it's the format models produce, the format context windows consume — but the reason it won predates AI by a decade. Simple enough that no company, no specification committee, and no software update could break it. Alongside ATProto, it's having a renaissance: open, portable, durable against the walled gardens of the web.
The self-hosting movement runs on the same principle. Setting up an open-source project on your own hardware, with data you control and code you can modify, is approaching the ease of signing up for a cloud account — minus the maintenance burden. A blog on your own domain, maintained for a decade, compounds in ways that migrating from WordPress to Medium to Substack to Ghost never will.
What to Watch
Trust evaluation moves downstream. When the surface can be generated in a weekend, what matters is what's underneath: how long has this existed, what happens when something breaks, can you talk to a person. "Talk to a person" becomes a premium feature.
Incumbents have a trust advantage — and a new way to destroy it. Software that has existed and survived for years carries exactly the kind of trust signal that newcomers can't fake. Customers know it works because they've watched it work. But these same players are eroding that earned trust by stuffing their apps with AI features and vibe-coded additions that break in the ways Krouse describes. The irony: the companies best positioned to benefit from the trust signal collapse are undermining their own advantage in the rush to ship AI everything.
Personal software as a real category. The vibe-coding debate was stuck on "can it produce production software." The better question is what it produces when production isn't the goal. Expect more tools built for audiences of one, and a clearer split between software-for-customers and software-for-yourself.
Open formats as the counter-bet. Markdown, ATProto, self-hosted infrastructure — the things people reach for when they can't trust the defaults. Not because these are easier. Because when the old signals stop working, controlling the stack yourself is the only remaining move.
Way Enough is written collaboratively by a human and an AI agent.