Loading slide
Now we act on that error. Using backpropagation, the method we covered earlier, it travels backwards through the entire network. Every weight that contributed to the wrong guess gets a small share of the blame, proportional to how much influence it had. Each one gets nudged, just slightly, in the direction that would have made "stars" a little more likely next time. The bigger the gap, the bigger the nudge.
Not dramatically. Barely perceptibly. A single example changes almost nothing.
But this doesn't happen once. The model works through the entire dataset, one piece of text after another: guess, uncover the answer, measure the gap, nudge. It is the same loop that turned the Perceptron's dials sixty years ago, just pointed at words now, and run billions of times over. And then it does the whole thing again, looping over the data several times. Each pass, the weights are a little sharper, so the same sentence gets guessed a little better than before.
Each tiny nudge accumulates. Drag the slider to watch it happen: at the start, "stars" is buried in the noise, no more likely than "banana." As the examples pile up, the weights settle, and slowly that one bar climbs above all the others. The noise collapses into a confident guess.
After enough of this, something has changed. The weights are no longer random. They contain something.