AI Makes Building Easy. Failing First Is the Hard Part

Or: how to be wrong quickly, cheaply, and on purpose.

Everyone has a guy. A friend, a cousin, an uncle. Let's call him Dave. Dave is made up, but you know him anyway.

Last spring Dave discovered AI, and by Thursday he was an “app guy.” He built an app that matches dog owners who live exactly 0.4 miles apart, in an afternoon, on his phone, during the game. He has a logo. He has a waitlist: three cousins and a man from his gym who was being polite.

Ask Dave who the customer is and he says “dog owners.” Ask how he'll know it's working and he says “we'll see how it goes.” Ask what would make him quit and he looks at you like you asked him to disown a child. Dave isn't stupid. Dave is fast. AI taught him how to build. It didn't teach him how to find out whether anyone wants the thing.

That's the gap this post is about. Building is easy now, but innovation was never about the building. It's about knowing what to build, for whom, and how to learn you're wrong before you fall in love with it. The fundamentals didn't get easier. They got easier to skip.

It doesn't mean “screw up constantly and write a thread about it.” The idea is simple: if something is going to go wrong, find out immediately and cheaply. It began in engineering. In 1985, Jim Gray described “fail-fast” modules at Tandem Computers: components that stop the instant they detect a fault instead of quietly corrupting everything downstream. Jim Shore brought the phrase to everyday programmers in 2004. Then Silicon Valley turned it into a bumper sticker, with help from Steve Blank, Eric Ries's The Lean Startup (2011), and Facebook's “move fast and break things.”

The bumper sticker skips the part that matters: failure only helps if you can see it and learn from it. A failure nobody notices isn't a lesson. It's just a failure with good PR.

Zappos is almost insultingly cheap. In 1999, Nick Swinmurn believed people would buy shoes online without trying them on. The grown-up test was a warehouse and an inventory. He didn't have the money to be that grown-up, so he photographed shoes in local stores and put the photos on a website. When someone ordered, he bought the pair at full price and shipped it himself. He lost money on every sale, and it didn't matter. He was testing one question, and he had his answer before he had anything to sell. The site became Zappos, and Amazon bought it in 2009 for a reported $1.2 billion.

Humane's AI Pin bet that people would ditch their phone screen for a voice-driven device on their lapel. A rough prototype and a few dozen volunteers could have started answering that in a week. Instead, Humane raised hundreds of millions for custom hardware and launched in 2024 at $699 plus a subscription. The first moment a customer could say “no thanks” came after the money was spent. In February 2025, Humane sold its team and technology to HP for $116 million, and the Pins stopped working within days. Nothing says “we hear you” like a very expensive lapel ornament.

The idea may well have been wrong, and that's allowed. The tragedy was the price of finding out. Zappos paid almost nothing to learn its hypothesis was true. Humane paid everything to learn it was false.

AI adds a new way to fail, because building got cheap. In a postmortem published in August 2026, solo founder Gus Chiriboga described using AI coding agents to build a tool for supervising AI coding agents. He shipped 128 releases with CI, tests, bilingual docs and governance policies. External users: zero. Very thorough governance for a country with no citizens. His second attempt had 442 passing tests, and broke the moment he used it for real, because the tests checked his own model of the system.

His diagnosis, after about two years: technical work was how he avoided real users. “Agents make building so cheap that building becomes the procrastination.” And: “Tests validate your model of the system. Only strangers validate the system.” Put both on a mug.

The new loop is: build it, ship it, “learn from the data.” It sounds responsible. It's a trap wearing a lanyard. You can't learn from data when you never wrote down what you expected, so every result looks like a signal and nothing ever clearly fails. A flat chart is “early.” A silent launch is “building awareness.” Soon you're curating a graveyard of half-used features, each one “kind of promising.”

Failing fast is cheap now. Failing legibly is still your job. Cheap building made thinking the only scarce resource left, and the machine will happily build the wrong thing at astonishing speed without ever asking, “Are you sure?”

So write the hypothesis. Set the kill date. You have the most powerful building tools in history, so aim them at the right thing, and learn something every time. Every honest failure is a course correction, and course corrections are how you reach the moon.

Now go put your work in front of a stranger.