Attention.

Somewhere in the last few years the scarcest input to software flipped. It used to be compute — you shaped programs around what the machine could afford. Then intelligence got cheap. You can now rent a competent mind by the token, at three in the morning, in whatever quantity you want. What you cannot rent, buy, or mint is another hour of your own life.

I said this on a podcast once and it's still the realest thing I know about building with AI: I can buy more tokens. I cannot buy another day of my life. Comparing the cost of tokens with the cost of human time is not even a comparison.

Take that seriously and it stops being a sentiment and becomes an engineering constraint — the engineering constraint. Every design decision reduces to one question: whose attention does this spend? A loading spinner spends yours. A refresh button spends yours. An agent that asks a question it already asked last month is spending the only resource in the whole system that doesn't come back.

It cuts both ways. The machines have attention too — a context window is exactly that — and it obeys the same law. Tell a model to pay attention to everything and it will pay attention to nothing. The craft of working with these systems isn't giving them more information or more power; it's directing a finite beam. Context quality beats context quantity every time anyone bothers to measure.

So the rule I build everything by: spend machine attention lavishly so human attention is never spent twice on the same thing. Tokens by the million to save one correction. That ratio sounds absurd until you price the correction honestly.

Most software is still built the other way — frugal with compute, profligate with you. Every notification, every re-login, every "are you sure," every re-explanation to a fresh session: small withdrawals from the one account that takes no deposits.

The withdrawals are optional. That's the part almost nobody builds like they believe.