LLM 101 — 03: When Examples Are Necessary
Specifying a task and calibrating its semantics are different jobs — when a category word does not pick out a unique position on its abstraction axis, a few labeled examples locate it through relative structure.
The previous note ended with a qualification: if the target can be stated directly, state it directly.
But there is another kind of ambiguity.
Sometimes the instruction is perfectly clear about what the model should do, and the output is still underdetermined.
Consider:
Identify the skills required for this job.
As a task, nothing is missing. The verb is clear, the object is named; any competent reader knows what to do.
Now look at some perfectly reasonable outputs:
pandas Python programming software development
A library, a language, a capability, a discipline. None of these is wrong — and they barely disagree about what capability is involved. What they disagree about is where to draw the boundary around it. Each answer represents the same underlying capability at a different resolution.
The model’s problem is not that it doesn’t understand skill. It understands skill at every resolution at once. The instruction told it what to do. Nothing told it at which resolution to answer.
So what is missing is not part of the task. It is a variable the instruction never mentions:
= the abstraction level at which “skill” should be read.
The instruction specified the task completely and calibrated not at all — because the word skill itself does not pick out a unique position on that axis.
A task can be fully explicit while its semantics remain underdetermined. Specifying the task and calibrating its semantics are two different jobs.
Can be fixed with words? Certainly:
Treat programming languages such as Python as skills — not individual libraries such as pandas, and not broad disciplines such as software development.
This works. But look at how it works. It does not define the level. It names a point below the target and a point above it, and says: between these. More specific and broader are words for directions along the axis; the position itself has no word of its own — so the sentence locates it by pointing at its neighbors.
Some semantic variables are much easier to specify relationally than absolutely. And once you see that, the honest form of that sentence is already familiar:
software development — too broad programming — this level Python — too specific
These are not demonstrations. Two of them are labeled wrong; nothing here shows how to perform the task.
What they do is calibrate. Each point sits at a different resolution, which makes the hidden axis visible; the labels say which direction along it goes too far. And the information is not in any single example — programming, alone, is just one more possible answer. The information is in the relative structure between the points. One marked right, one too broad, one too specific: together they locate — the variable the category word left open.
This is the role reversal. In the previous note, a concrete point was the hazard: it fixed many semantic dimensions at once, most of them accidents. Here, the thing that is missing is precisely a position on one dimension — and a concrete point is the one kind of signal that occupies a definite position.
A good example does not merely show an answer. It locates a concept.
And skill is not special. What counts as a risk? At what granularity is something an industry? Is answering one email a task, or is running the support queue a task? Every category word spans resolutions the word itself does not distinguish, and leaves the same variable unwritten.
So the series closes on a symmetry.
When the target can be described, a concrete example is a poor encoding of it: a point standing in for a region, fixing every dimension when your intent concerned only a few. When what is missing is a position that the category word does not encode, a few concrete points — arranged around the target, labeled with directions — are the most natural encoding there is.
Instructions describe a region. Examples give it coordinates.
Instructions describe. Examples locate.