The Sheep Knows

A sheep sees a wolf and runs. Ibn Sina asked a question about that which nobody had quite asked before. What did the sheep actually see?

Not a wolf. Light hit its eye carrying shape, colour and movement, and that is all light can carry. The danger was never in the light.

The faculty he had to invent

His answer was that something else must be doing the work. Sight delivers the form. A separate faculty grasps what the form means. He called it wahm, usually translated as estimation, and Latin translators later rendered it estimatio.

This one is his own. The other inner faculties he inherited from Aristotle and the Greek commentators and rearranged. Estimation appears in no earlier Greek, Christian or Muslim philosopher. He needed something for a job nothing in the existing scheme could do, so he added it.

The job is this. Part of what you perceive arrives in the sensory signal and part of it does not. The wolf's grey coat is in the signal. The wolf's hostility is not. Ibn Sina called the second kind ma'ani, meanings, and held that one faculty perceives them and another stores them. The storing faculty is the one I wrote about in The Rep You Skipped.

The grey of the coat is in the light. The danger is not.

Meaning is not in the signal

That sounds like a technicality. It is the whole thing.

Think about what it takes to read a two line message from your manager that says, can we talk tomorrow. The words are the form and every letter is right there in front of you. What it means depends on what happened in yesterday's meeting, whether your project slipped, how this person usually phrases things, and whether they used your first name. None of that is in the message.

You do not reason your way to the meaning either. You arrive at it immediately, the way the sheep arrives at danger. Ibn Sina would say your estimation did that, working on inputs the words never carried.

Where AI sits

Now the useful part.

A language model is extraordinarily good at form. It may be the best form reader ever built. Pattern, structure, style, the shape of an argument, the statistical texture of how sentences go: that is what it was trained on and what it does superbly.

What it does not have is your afternoon. It was not in yesterday's meeting. It does not know that this manager writes short sentences when annoyed, or that the deadline moved on Tuesday. Paste the message in and ask what it means, and the model reads the form beautifully, then fills in the meaning from the average of every similar message ever written.

Often that average is fine. Sometimes it is precisely wrong, and it will be wrong fluently, because nothing in the text announces that the missing context existed.

Two things worth taking from this

The first concerns your own judgment.

Intuition has a poor reputation at work. It sounds like the opposite of evidence. Ibn Sina's account suggests something more careful: a gut feeling is a faculty operating on real inputs and producing a real reading faster than you can put it into words. That does not make it right. Estimation misfires constantly, which he knew too, and he used the same faculty to explain why animals fear harmless things. But it is information rather than noise, and the useful response is to ask what it picked up on instead of waving it away.

The second is a rule for using AI. Ask one question before trusting an answer: does this depend on something that is not in the text I gave it?

If no, the model is on home ground and will often do better than you. If yes, you have asked a form reader for a meaning, and what comes back is a well written hypothesis rather than an answer.

The gap that has not closed

Ibn Sina added a faculty to his model of the mind because the model he inherited could not explain how a sheep knows to run. A thousand years later we have built machines that do the form half of his scheme superbly and the meaning half by imitation.

That is not a permanent limit, and I would not bet on it holding for long. For now it is the most reliable line to reason along. Whatever depends only on what is written, hand over. Whatever depends on what was in the room, keep.

What to do this week

1. Before pasting something into an AI, ask what the answer depends on. If it needs context that is not in the text, either supply that context or discount the answer.
2. When you get a gut reaction in a meeting, write it down before you talk yourself out of it, then ask what specifically set it off. That is usually recoverable and usually real.
3. Watch for the moments a model sounds most confident. Fluency measures how well it read the form, never how well it read the situation.
4. For anything turning on history, relationships or intent, brief the model the way you would brief someone joining on Monday. Everything a new colleague would need, it needs too.