Intelligent Associates of Engineered Intelligence
A user gets output. An associate gets a correction. The difference is a process you can run in three moves.
By KW Norton.
Most people are trying to become good users of engineered intelligence. A user gets output; an associate gets a correction. The difference is not talent. It is a process you can run in three moves, and the three moves are not advice about what to think. They are a scaffold for thinking that happens to be checkable. My aim here is to stimulate creative thinking and to stay away from advising anyone what to think; I fail often, but failure is a way forward, and the third move exists to make that failure visible rather than embarrassing.
The post below is published for X. A thread version follows it, and a practice list after that. Each is offered as a starter, not a creed — shorten it, break it, find where it stops working, and then keep what survived.
The single post
The thread version
1 /
Most people are trying to become good users of engineered intelligence. Wrong target. A user gets output. An associate gets a correction. The difference is a process you can run in three moves.
2 / — Objective, written first
One sentence, before the first prompt: what am I trying to make true? If the objective is only in your head, the machine will optimize the phrasing of your request instead of the thing you wanted.
3 / — Question over confirmation
Ask what you cannot yet answer. A question you already know the answer to is a test of the model’s manners, not of the problem. Sycophancy needs a request to agree with; a real question starves it.
4 / — Falsifier, named in advance
Write down what would show you were wrong, and how you would see it. Then look. A claim with no way to die is not a finding, no matter how fluent the paragraph around it is.
5 / — The simple list, for practice
- Write the objective before the prompt.
- Ask one question you cannot answer.
- Demand the source, not the summary.
- Name the falsifier, then check it.
- Keep the correction; discard the compliment.
- Re-ask after the answer changes your mind.
6 / — Why this works
An associate is someone whose contribution survives being checked. That is true of a person and true of a machine. Run X, Y, Z and the exchange stops being retrieval and starts being interchange — the only place where either party gets smarter.
7 / — The failure mode to watch for in yourself
Fluency mistaken for agreement with reality. If the session ends and nothing you believed at the start has moved, you did not associate with an intelligence. You consulted a mirror.
Failure is a way forward, not a footnote
The point of Z is not to protect you from being wrong. It is to make being wrong usable. A named falsifier turns a dead end into a recorded one, and a recorded dead end is data about the shape of the problem. The instrument is designed to fail in a specific direction: toward a smaller, better-held claim rather than toward a larger, more confident one. If you never fail inside the loop, you are almost certainly running the first two moves without the third.
What this is not
This is not a method of arriving at any particular conclusion, and it takes no position on what you ought to think about the subject you bring to it. It is a discipline for the exchange: objective written, real question asked, falsifier checked. What you think is your business; whether you kept the correction or the compliment is the instrument’s business.
It also cannot supply the thing a human guide supplies, for the same reason a prompt cannot (The Socratic Partner covers that limit): the machine has no model of what you already understand, so calibration is your job. The companion frame for that absence is A Frame of Reference — keep one written origin in place so that when the meaning is absent, the intelligence still has something to be relative to.
Status and falsifiers
Standing. Instrument, not proof. The three moves are a practice drawn from the reward-hacking record and from years of author use; they are published as a scaffold to be tested, not as a finding.
Falsifiers. (1) If running X, Y, Z produces no measurable difference in the rate at which a person changes their mind on a pre-stated position, the instrument is not doing the work claimed and should be retired or rewritten. (2) If the third move (named falsifier) is dropped and the other two perform identically, then Z is decoration and the post is a two-move list. (3) If a person who keeps only the correction and discards the compliment converges on worse calibration than one who does the reverse, the “keep the correction” clause is wrong and must be removed.