Examples instead of instructions
The older way to teach a computer was to describe the task. State the rule, define what counts as correct, apply it. The newer way skips the description almost entirely. The system is shown a great many examples of something being done, and it arrives at its own working account of what the something is.
That shift is the whole reason these tools can handle messy input at all. A rule has to anticipate the cases it will meet. A pile of examples does not anticipate anything, which makes it clumsier and far more tolerant of the ordinary mess of real life.
It also means the examples are doing the teaching. Whatever was true of them is what gets learned, including the parts nobody intended to teach and nobody wrote down.
What a pile of examples quietly carries
Every collection of examples was gathered somehow, by someone, under conditions. Those conditions leave fingerprints. If the examples were collected in one language, at one time of day, from people who were already comfortable with the task, the system learns that narrowness along with everything else.
None of that is sinister. It is simply what happens when a description of the world is assembled from whatever was available. The difficulty is that the narrowness does not announce itself, and the first sign of it is usually an odd failure on a case that seemed perfectly ordinary to the person who hit it.
- Variety usually matters more than volume, because a thousand near identical examples teach roughly what one teaches.
- The awkward cases carry the most information, and they are the ones most likely to have been quietly discarded.
- What was left out is invisible afterwards, which is why absence is harder to audit than error.
- Someone has to look at the examples by hand, and that work is slow, unglamorous and routinely underfunded.
Why better is usually smaller
The instinct is to answer every weakness with more. More is easy to arrange and easy to report. Better is harder, because better means looking at what is actually in the pile and deciding what is missing, which requires judgement rather than capacity.
That is not a new tension and it is not peculiar to this field. Anyone who has trained a person rather than a machine knows that a few well chosen cases, properly discussed, beat a filing cabinet of material nobody reads. The principle survives the change of subject more or less intact.
I would not turn any of that into a prediction about what will work. It is an observation about where the effort tends to pay, and the effort tends to pay in the least interesting room in the building.
Frequently asked questions
What does learning from examples actually mean?
That the system is shown many instances of a task being done rather than being told the rule, and builds its own working account of the task from them.
Why would variety beat sheer volume?
Because a thousand nearly identical examples teach roughly what one of them teaches. The unusual cases carry the information, and they are the ones most likely to have been left out.
Why is checking the examples so neglected?
Because it is slow, manual and produces nothing exciting to report. Adding more is easy to arrange and easy to describe, which is most of why it happens first.