The model is the saw, not the carpenter.

There's a way of talking about machine-learning models that treats them as proto-craftsmen — as if the right prompt is the only thing standing between us and a finished bookshelf. It's the wrong frame, and the wrong frame produces the wrong tools.

A saw is good at one thing. It cuts wood faster and more accurately than a hand can. It is useless without the carpenter who knows where to cut, why, and when to stop. Nobody pretends the saw is the carpenter. Nobody complains the saw doesn't know how to make a chair.

The interesting tools — the ones that actually become part of someone's working life — get the saw-and-carpenter relationship right. They make the cut faster. They don't pretend to know which cut to make. They surface the choice that the carpenter alone is qualified to make, and then they get out of the way.

The bad tools do the opposite. They obscure the choice. They flatter the user with the impression that the work is done. They reduce a craft to a button. And then, six months in, the user looks up and realizes the work the tool produced is mediocre, and the craft they used to have has rusted, because no one was practicing it.

We have a soft preference, when we ship anything, for the boring side of this trade-off. Make the cut faster. Surface the choice. Trust the carpenter. Refuse the temptation to be impressive at the cost of being useful.

It's a slower path. It produces less spectacular demos. It tends to disappoint the people who want a magic word on the homepage. But it's the only path we've seen that produces tools the people using them still respect a year later.