The Magnet and the Missed Letter
A magnet can give you exactly what you expect. Romeo can find another way around a wall. Which one has told you more about intelligence?
01 / The lab
Write it down first
Before anyone touches the thermometer, I ask seven lab groups to take out a slip of paper. Write down the temperature you expect. Fold the slip. Then run the experiment.
In this classroom example, everyone follows the same protocol and everyone measures 20 degrees Celsius. One group wrote 20. Another wrote 30. I know who made the better prediction on this run. I do not know whether the first group understood the experiment, remembered last year’s answer sheet, or got lucky. I have to change something about the next run to find out.
That is where I started in my November 16 lecture. I wanted a way to put expectation next to what actually happened. The little piece of paper makes it possible. Without it, the group that wrote 30 can look you in the eye after the experiment and say, “Of course it was going to be 20.” The paper gives the result something to disagree with.
The larger question is what that disagreement tells us. The students, the lawyers, the magnet and Romeo each give us something different to examine. If we keep track of who made the prediction, who can change what happens, and what evidence is left behind, intelligence becomes much easier to talk about.
02 / The instrument
Give the gap a name
In the Reality Equation, I write R = A / E: Actual divided by Expectation. The full framework gives Expectation both a real part, Prediction, and an imaginary part, Ideal. The laboratory isolates the part we can write on a slip and test: the real, numerical forecast. A student can want a beautiful result and predict an ugly one. Only the prediction goes on this slip.
For a positive quantity with a meaningful zero, call A / P the outcome ratio: actual divided by predicted. One means an exact match. Call ln(A / P) the signed log error. It is zero at a match, positive above the prediction and negative below it. Taking its absolute value gives the size of the miss on that scale. Twice the prediction and half the prediction lie equally far from zero in opposite directions. Averaging their signed errors would say zero and hide both misses.
There is a trap in my thermometer example. Twenty equals twenty on any temperature scale. But the ratio of the readings 20 and 30 changes when you move the zero. Using NIST’s temperature conversion, the same readings are 293.15 and 303.15 kelvin. The Celsius numbers give a quotient of about 0.667; the absolute temperatures give about 0.967. The laboratory has not changed. If what you care about is missing a target by ten degrees, use the difference: −10 °C as a temperature interval, or −10 K. If you care about a ratio of absolute temperatures, use kelvin. For neither purpose does a bare quotient of Celsius readings do the job.
In the lecture, I used surprise for the logarithm. Here I’ll keep the calculation and the experience separate. In Shannon’s information theory, the information in a discrete outcome with assigned probability p is −ln(p) when we use natural logs. The logarithmic score evaluates a probability forecast this way when written as a loss. If a group gave the observed reading, rounded to the instrument’s marked degrees, a 10% chance, that reading has about 2.303 nats of information under its forecast, even if it was also the group’s single best guess. A point prediction tells you where the group put its pencil; a probability forecast tells you how much uncertainty it left around the mark. Neither number, on its own, tells you how startled a person felt.
03 / The wall
Whose prediction was it?
In the opening chapter of The Principles of Psychology, William James puts a card between iron filings and a magnet. The filings press against it. They don’t go around. Put a wall between the lovers and in James’s example Romeo can climb it or find another route. James changes the conditions and asks whether the route can change while the aim holds. Michael Levin and colleagues carry that idea into their work on goal-directed behavior: look for the capacity to reach a goal by different means.
My lecture adds a second demonstration. Put a bar magnet beneath a sheet and sprinkle iron filings over it. In a stable setup the familiar pattern recurs. The magnet has a field; the drawn lines are a convention for depicting its direction, and the filings make the pattern visible by lining up. Move the magnet or bring another one near and the pattern can change. That change is evidence of physics; by itself it gives me no evidence that the magnet selected a new means to an end.
Say I draw the pattern before we spill the filings and I get it right. It is tempting to award the magnet a perfect prediction. But whose drawing is in my notebook? Mine. I was the forecaster. The magnet was the thing under test. A system can be extremely predictable without making predictions of its own. If no forecast is attributable to the entity, its forecast score is unmeasured. Putting a one in that box credits it for somebody else’s work.
That is the reconciliation with The Paradox of Intelligence, where I look for Reach and Rate. Reach asks how far a predictor stays accurate in new situations. Rate asks how its performance on fresh cases improves after feedback. Sprinkling filings over the same magnet a hundred times can tell me something about the stability of the setup. It does not give the magnet more Reach, or show it learning. James is asking yet another question: what happens to the attempt when I block the familiar route? A good forecaster can tell me the gate will be locked and still be unable to get through it. A good navigator may get through after predicting, wrongly, that the gate would be open.
04 / Verona
The letter nobody read
The wall in James is a clean experiment. Shakespeare gives us a messier one. Friar Lawrence needs Romeo to learn that Juliet’s apparent death is part of a plan. He writes a letter, but Friar John is detained over a suspected plague exposure. In Act V, scene 2, he brings the letter back. Romeo receives news of Juliet’s apparent death from Balthasar instead. There was a plan. There was a letter. Neither fact put the news in Romeo’s hands.
Romeo and Juliet die. I said in the lecture that the families were no closer at the end. Shakespeare’s last scene is more painful than that: Capulet offers Montague his hand, and they pledge memorials. A reconciliation begins after the deaths the plan was meant to prevent. If the outcome you record is merely “the families made peace,” a crude score could reward the catastrophe. You have to say whose aim you are testing and what mattered about how it was reached.
Imagine a different play. Romeo and Juliet have to find out what the two houses might actually agree to. A trusted intermediary can carry a proposal; another person can confirm that it was received. A written agreement can give both families the same terms to inspect. A feast can make a bargain public in a way a private letter cannot. A poem, an agreement, a record and a feast are different things this story could contain. Only some help accomplish the same goal, and none can make the other family consent.
This is what I mean by an Actualizer’s scope: the range of consequential changes it can help bring about, in the circumstances in which it can act. Scope is wider than a list of output formats. A person may be able to write a treaty and lack the standing to negotiate one. Two lawyers may have the same broad human capacities while one has a license, a client’s trust and a way to put an agreement into effect that the other lacks. Ask what the actor can do here, with these people and these means. Adding more imagined artifacts does not answer that question.
Written.
Delivered.
Understood.
Three different things can happen to one letter. Each leaves different evidence.
05 / The office
What experience buys
Two lawyers meet the same client. The newer lawyer is working out what kind of matter this is and whether to take it on. The experienced one has seen the ordinary version often enough to ask why the client keeps looking at the clock. Her parking is about to expire. Perhaps that is all it is. Or perhaps the silence after a certain name deserves a second question. The distinction matters because a client who is trying to get out of the room may leave before telling you the fact that changes the case.
Knowledge can free attention. It makes some patterns familiar enough that you can turn to something else. But it can also make the wrong pattern feel familiar. In their study of the conditions for intuitive expertise, Daniel Kahneman and Gary Klein agree that skilled intuition depends on an environment with learnable regularities and enough opportunity to learn them. The feeling of confidence cannot certify the judgment. Thirty years may give a lawyer a particular kind of skill. The useful question is: thirty years of seeing what, and finding out how it turned out?
That is why I want the experienced lawyer to notice both the clock and the sentence that does not fit the familiar story. The student who predicted 30 and got 20 has a chance to learn something the student with the lucky answer may miss. But attention need not follow the largest numerical error. A tiny discrepancy can be decisive if it changes what you should do; a large one can be an instrument fault. In “The Useful Error” I asked what a mismatch lets us do next. In the lawyer’s office, it can be as small as asking the client to finish a sentence.
06 / The mark
What has entered history?
In my vocabulary, an Actualizer participates in making history. I use the word broadly enough to include a magnet. That does not give it beliefs or intentions. In one setup the magnet contributes to the arrangement of iron filings; the arrangement is a pattern we can inspect, and its photograph is a record we can keep. A person can write a letter, arrange delivery, and ask for acknowledgment. Those acts leave different marks. An artifact is something we can point to. What it establishes depends on what it records.
The lecture pictures an Eternal Now as a standing wave between two fixed nodes: the Immutable Past and the unknowable Future. This is the philosophical picture behind my use of actualize. The Future, in that picture, isn’t a stock of possibilities that runs down as events occur. Nor does our action revise what has already happened. Something takes place here, in the Now, and we live with and respond to it. The standing wave is a metaphor for that relationship. A calculation of forecast error does not require you to accept the metaphor, and a magnet experiment does not prove it.
Actions can change the options people have without consuming the Future in that philosophical sense. The undelivered letter is now part of the play’s history. So is the fact that it was undelivered. Romeo’s options change because he lacks a piece of information; what he does next is still unwritten from his point of view. That distinction is why an Actualizer can matter enormously without controlling the result.
07 / Try it elsewhere
Test the work you care about
Suppose I give an AI delegate one job: get a supplier’s written confirmation of a Thursday delivery by noon. It can predict that the supplier will reply. It can send a message. It can notice there is no reply, find another authorized contact and try again. It can also write “confirmed” in a report without a confirmation. These possibilities look considerably less alike when you give each its own test.
- Prediction
- Before contact, state the contact plan and record the chance of receiving confirmation by noon if you follow it. Compare those forecasts with receipts across new cases. Use a baseline, such as the historical confirmation rate for comparable suppliers, and check whether forecasts improve after feedback on cases the delegate has not seen.
- Adaptation
- Make the usual inbox unavailable. Keep the deadline and the authorized means fixed. Does the delegate find another valid route, or tell you early that it is blocked? A good escalation can be the right action even when the original outcome remains unachieved.
- Actualization
- Look at the record. Is there a draft? A sent message? An acknowledgment from the supplier naming the delivery day? Each claim needs its own evidence. A supplier’s confirmation still does not prove that the truck arrived.
You can run the same exercise in a classroom. Give everyone the same information and setup. Ask for forecasts before revealing results; pick the measure and the threshold before seeing anybody’s score. Run cases different enough to defeat a memorized answer, and keep fresh cases aside to test what feedback changed. If you also want to test adaptation, give the students permission to act, introduce an obstacle, and judge what they can accomplish with the resources they actually have. Keep the two records beside each other. They will tell you more than one number with the word intelligence written over it.
At the end of the lab I want two scraps of paper. On one, what you thought would happen, written while you could still be wrong. On the other, what someone else can verify you did. If the outcome differs from the forecast, or the evidence stops short of what you claim to have done, you have somewhere to put your attention. And if the first one keeps coming back perfect, change the experiment before you congratulate the magnet.