How Models Learn
Last topic your filter could call an email spam, but nobody told it how. It starts out guessing, gets one number back for how far off it was, and quietly tunes itself, over and over, until the guesses stop missing.
- ▸Nobody hand-writes the fix; the size of the miss is what tells the model which way to nudge
- ▸Bigger nudges are not faster: too big and the guess overshoots and bounces, too small and it crawls
- ▸Past a certain point, more passes buy almost nothing; the miss has bottomed out
Your spam filter makes a call, spam or not-spam, by leaning on a single dial inside it: how hard a pile of links should push an email toward spam. Right now that dial is set wrong. Nobody set it by hand, it started at some arbitrary value, so its very first calls are close to random.
Here is the one move that starts everything. The filter makes a guess with the dial where it is. Then it checks that guess against the graded inbox you already built, the one where every email carries the answer. The gap between what it guessed and what you graded becomes a single number: how wrong it was, all its mistakes rolled into one figure.
That number is the whole point. It is not a verdict on the email, it is a verdict on the dial. A big number means the dial is badly set; a small number means it is close. The filter cannot see the right setting directly. All it can see is this one measure of how far off today's setting is. Everything the model does next is an attempt to drive that number down.