September’s “Ask HN: What are you working on?” thread pulled 194 points and 581 comments, which makes it — as it is every September — the internet’s largest voluntary inventory of handmade software. Five hundred eighty-one people describing the thing they built on nights and weekends. It is charming, it is generous, and it is aimed at a customer who has quietly stopped being the first reader.

That is not a small problem. The old contract of software was simple: build for a person with a problem and a credit card, and the person shows up. Now the person shows up second. First comes a model deciding which four paragraphs get summarized, and only then comes the human typing the follow-up question. Everyone in that thread is optimizing for the second visit.

The Buyers Changed Species

An Orbit Media study reported this year — 97 B2B sites, 29 million visits — found that visitors arriving from AI referrals converted roughly three times as often as visitors from ordinary organic search. Read that number again and notice what it isn’t. It isn’t a channel shift, the way desktop gave way to mobile. It’s a different animal walking through the door.

An AI-referred visitor has already been pre-sold. He arrives with a question the model has partially answered and a shortlist the model has already narrowed. Conversion triples because the funnel collapsed into a single sentence somewhere upstream. Which means the fight worth having happens before anyone clicks anything — at the moment a model decides whose page to quote.

The 581 builders in that thread are competing for the front page of a website whose readers are, increasingly, aggregators. The front page is still a real prize. It is just no longer the prize that pays.

215,128 Pages Nobody Will Read

The other half of the picture is less charming. According to a September 3 roundup of Hacker News dailies, three sites produced 215,128 “best software” pages built for AI. That is roughly 72,000 pages per site — a genre that once required a masthead, a budget, and a staff, now manufactured at industrial scale by three operations you have never heard of.

The reflex on my side of the aisle is to call this theft or fraud. It isn’t. It’s production. What’s actually happening is worse than theft: it’s competition. The incumbent tech publishers spent twenty years running the same play — the listicle, the comparison chart, the “Top 10 CRMs for Small Business” — and they ran it slowly, with human writers and quarterly targets. Three sites just ran it faster than anyone can enforce against, and the model doesn’t care who has a masthead.

A growth lead at a forty-person software company described the shift to me in a Slack thread that kicked off at 11:40 p.m. and ran until two: “We stopped writing for the buyer. We write for the summary, then we write for the follow-up question the buyer asks the model afterward. The buyer’s the second audience now.”

That’s the part the outrage misses. The problem isn’t that the slop is slop. The problem is that most of it won’t matter and a sliver of it will matter enormously. Two hundred fifteen thousand pages, and maybe four hundred of them will be the ones a model quotes when a small business owner asks what to buy. That is a power law that makes App Store economics look like a charity.

The Margin Is Underneath, Not On Top

The status in the Ask HN thread sits at the application layer — the beautiful CLI, the solar sensor, the better RSS reader. The money in September’s companion “Who is hiring” thread sits somewhere else entirely. Neuralwatt, an early-stage company doing GPU optimization and energy-aware scheduling for datacenters, was advertising $180,000 to $220,000 plus equity for Rust and Python work. A self-funded research lab in Zurich was recruiting for ARC-AGI3, LLM/RL reasoning, and quant.

Meanwhile the September 3 digest led with model news: a 27B model served on Cerebras at 1,500 tokens per second, and something called GPT-6 Astra. Serving speed and wattage per token. That’s the business. Every application in that 581-comment thread sits on top of a substrate whose prices are set by people optimizing kilowatts, and the builders have almost no leverage over it.

This is the uncomfortable thing for the build-in-public crowd, and for the people who cheer them: the skill that got you a front page in 2015 is a hobby now. The compounding asset is being the answer.

Being the Answer Is Now the Job

The standard indie objection — no SEO, no marketing, let the work speak for itself — was always half vanity. It is now a straight business error. If the model is the first reader, then quotability is a product feature, in the same sense that uptime is. Being written about in the right three sentences is worth more than a thousand Show HN upvotes, and it is a thing you can deliberately engineer: clean comparison pages, honest pricing in machine-readable form, documentation that answers the question someone would actually type.

The thread isn’t a failure. It’s a résumé. The trouble is that the recruiter is a language model, and it doesn’t read résumés. It reads summaries, and it only summarizes what it can find, and it finds whatever the three firms with 72,000 pages apiece have already told it.

Sources