Sudden understanding

June 2, 2026

Anyone who has learned a language knows the plateau. For weeks you make no visible progress, and then one morning a radio programme that used to be noise is suddenly speech.

It is tempting to think nothing happened during the flat part. More likely, a lot happened that was not yet visible. Small pieces — a sound here, a grammatical pattern there — were being assembled, but none of them were useful on their own. Understanding a sentence needs most of the pieces at once. When the last few fall into place, the whole thing switches on.

Researchers training neural networks have described something similar. A model can sit at the same performance for a long time and then, with no change to the training procedure, rapidly generalize. The usual explanation is the same shape: internal structure was slowly forming underneath a flat surface measurement.

What this suggests in practice

If progress really is assembled from pieces that only pay off together, then judging an effort by its short-term curve is misleading. The flat part is not evidence that nothing is working. It is also not evidence that something is — plateaus can be genuine dead ends.

The honest position is uncomfortable: from the outside, a plateau before a jump and a plateau before nothing look identical. The only useful question is whether the underlying pieces are still accumulating.