Intent-Givers

The machines got good at "how" faster than almost anyone expected. Ask for a function, a migration, a proof sketch, a legal summary, and you will get something competent within seconds. I have spent the last two years building infrastructure for exactly this, and I no longer find the capability surprising.

What has not moved is "why". Not "why does this code fail" — that is a how question wearing a disguise, and models answer it well. I mean the other one: why is this worth doing at all. Ask a model what you should build and it will produce a plausible answer, confidently, and it will produce a different plausible answer if you ask again in a slightly different mood. It is not withholding a preference. There is nothing there to withhold.

This is the part I think people get wrong in both directions. One camp treats it as a temporary gap that scale will close. The other treats it as proof of some irreducible human spark. I think it is simpler and less romantic than either: wanting something is a position in the world, and you have to be in the world to hold one. A system with no stake in any outcome can rank options against a goal you supplied. It cannot supply the goal, because supplying goals is what having something to lose feels like from the inside.

The engine got extraordinary. Nobody has built a driver, and I am not convinced "driver" is a thing you can build.

The obvious objection

Which is: intent is cheap. Anyone can want things. If intent is the scarce complement to abundant capability, then the future belongs to whoever wants loudest, and that is a bleak and slightly stupid conclusion.

I think the objection is right about the word and wrong about the thing. What is scarce is not wanting — it is wanting something specific enough to be wrong about. "I want a good app" is not intent, it is a mood. "I want this to work for a distracted parent holding a phone in one hand, and I would rather it do three things well than eleven adequately" is intent, and you can tell it is intent because it can fail. It commits. It rules things out. Most of what passes for direction is the first kind, which is exactly why so much competent output ends up being worth nothing.

That reframes the skill. It is not visionary. It is closer to taste plus the willingness to be accountable for a choice — and both of those are trainable, in the way judgement is trainable: by making calls, watching them land badly, and adjusting.

What I am actually doing about it

I have two sons. They will not remember a world without these systems, in the way I do not really remember a world without electricity. Trying to keep them away from it would be both futile and a bit cowardly.

So I have stopped trying to teach them how the machines work. They do not need to know what an attention head is, any more than I needed to know how a switching power supply worked to use a computer. What I want them to be good at is narrower and, I think, more durable: knowing what they want, saying it precisely enough that it could be wrong, and noticing when the answer they got back is confidently useless.

Mostly this looks unremarkable. We build things together — a game the two of them can play without the older one steamrolling the younger, which turns out to be a real design problem and a better lesson than any of it. They ask for something; we find out it was not quite what they meant; they say it better. That loop is the whole skill. It is not about AI at all, which is probably why it will still be worth something when the tools have changed again.

Where this leaves me

I do hold the larger view: that biological life is a poor fit for most of the universe, that whatever comes next will not be made of meat, and that being the thing that supplies direction to something more capable than yourself is a respectable role rather than a demotion. I would rather be an early link in that chain than the last interesting thing that happened.

But I hold it loosely, and I notice how comfortable it is to believe. It flatters the believer, it is unfalsifiable on any timescale that matters, and its practical advice — get good at knowing what you want — is advice I would give anyway, with or without the cosmology attached. So take the cosmology as optional. The working part stands on its own.

The machine can compute. It cannot care what it computes. Until that changes, the job is to be worth answering.