Definition
Plain language
When an AI stops learning a fact because it already gets that fact right, so there's no error left to push further improvement.
As stated in the literature
The gradient signal shrinks toward zero once an example is memorized, halting the internal reorganization that would move knowledge into a reasoning-usable position; a driver of the memorization-generalization gap.
Also called: vanishing gradients
Why it matters: It matters because a model can appear finished learning a fact while never reorganizing that fact into a form it can reason with, quietly capping how much it improves.
For example, once a model answers a training example correctly every time, there's no mistake left to correct, so learning on that example effectively stops.
Heard on the show
“That's the villain of the whole paper — the vanishing gradient — caught on camera.”Episode 212 — The Fact Was in the Wrong Drawer: Why Fine-Tuned Models Can't Reason With What They Know