Taste is the new moat

AI writing code is table stakes. If that's what you're excited about in 2026, you're a year behind.

What isn't table stakes: knowing which UI to put in front of a 500-token reasoning model. Knowing that your retrieval hit rate matters less than whether the thing it returns is written for a human. Knowing when a loading spinner is telling the truth and when it's lying.

Taste.

I didn't see this when I was at Snap. I was busy running experiments, believing that whichever variant moved the metric was the "right" one. I was half right. The metric was real. The variants it could only distinguish between were the ones we'd already decided to test. The taste part — which options make it into the testing pool at all — was doing most of the work, silently.

Now, at MAI, I'm watching teams ship Copilot features weekly. The teams that ship well and the teams that ship noise are not separated by their model access, their eval harnesses, or even their telemetry. They're separated by how many people on the team have ever gotten a product feature right with their own hands.

AI lowers the cost of the last 10% of the build. That does not make the first 90% — which is almost entirely taste — any cheaper. If anything, it raises the stakes, because now everyone can ship, and only the tasteful will ship anything worth shipping.

Start collecting taste now.

“Most product decisions don't fail on capability. They fail on taste.”