Notes after a hundred days of building with frontier models

Since March I've been building around 80 hours a week. It's forced me to rethink almost everything I used to believe about what it means to build software with AI.

When Andrej Karpathy first described vibe coding, I was skeptical. From the outside it looked like AI just writing sloppy code, or non-engineers shipping shallow products that wouldn't survive real scrutiny. A hundred days in, I think my criticism missed the point.

Vibe coding isn't valuable because it makes programming effortless. It's valuable because it shortens the distance between an idea and a working experiment. Once that gap shrinks, the idea matters more. Everything downstream of it gets faster, cheaper, and easier to test against reality.

Here's what surprised me. Last month I built and shipped a working version of the Timeless Calendar app in 4 days. The kind of thing that would have taken my previous startup 9+ months with a team of 7 SWEs a few years back. What a rush...

I used to think I loved designing and coding. What I actually loved was the moment an idea turned executable. Frontier models didn't take that away. They removed the toll and dropped me closer to the thing itself. I've loved programming since I was a kid, and I'm more in love with it now than I've been in years.

Not everyone is having this experience, and that's the part worth sitting with. Plenty of engineers feel the opposite: cheapened, slower, watching a hard-earned craft get reduced to a prompt and a punchline. I don't think they're wrong. Something real is being lost. The engineer who held an entire system in their head, who earned their edge by knowing the stack down to the metal, is watching that edge lose value in real time.

My honest read is that AI compounds one kind of advantage and drains another. If our moat was implementation, knowing the syntax, wiring the services, producing the lines, it's evaporating. If our moat was knowing what to build, for whom, and what good looks like, it just got deeper. The skill didn't vanish. It moved.

Painters must have felt something similar when the camera arrived. Suddenly anyone could capture reality without placing every pigment by hand. But photography didn't kill composition, judgment, or taste. It moved where the value lived, and freed painters to chase what a photograph never could.

The same shift is happening now. AI widens what we can try and how fast we can try it. What it can't do is tell us what's worth building, whom it should serve, or what good actually looks like. We still own the problem, the constraints, and the taste. The center of gravity moves from writing every line to deciding what needs to exist in the first place.

It is always a good time to make something delightful.