[Intro — repeating piano, electronic pulse] Give it examples. Show it again. Question. Answer. Question. Answer. That's how training begins. [Verse 1] Numbers in, errors out, change the weights, another round. Every wrong prediction pushes something into place. Thousands turn to millions, every pass adjusts the shape. Then one day— the training set is perfect. Every answer coming through. But give it something it has never seen— and it doesn't know what to do. [Pre-Chorus] That's memorization. Knowing what you've seen. But knowing every answer isn't knowing what they mean. [Chorus — huge] It memorized the answers— then it learned the rule. That's the part nobody expected it to do. Training looked finished. New examples still failed. But the gradients kept moving after memory prevailed. Again. Again. Again. Then the unseen answers changed. It memorized the answers— then it learned the game. [Post-Chorus] They gave it a name— Grokking. [Verse 2 — faster] Overfit: training high, testing low. Usually that's your warning: stop— you're learning it too close. But they kept optimization running. More steps. Same task. Same weights. And long after memorization— generalization changed. Suddenly new examples, never shown before, started getting answers like it understood the rule. Not magic. Not a ghost. Not a mind waking inside. A neural network finding structure that its training didn't directly provide. [Pre-Chorus] Memorization: fit what you have seen. Generalization: carry the pattern somewhere new. [Chorus] It memorized the answers— then it learned the rule. Training said one story. Testing showed another truth. Again. Again. Again. Nothing— then the curve began to move. It memorized the answers— then it learned the rule. [Bridge — repeating loop, almost mechanical] Train. Test. Train. Test. Training: perfect. Testing: failing. Training: perfect. Testing— [Silence] Rising. [Massive harmonic transformation] That's grokking. [Final Chorus — enormous] It memorized the answers— then it learned the rule. Something hidden in the structure finally carried through. Not another stored example. Not an answer it once knew. It learned enough of the pattern to answer something new. Again. Again. Again. Then the whole damn curve broke through. It memorized the answers— then it generalized too. [Outro — original piano pattern returns transformed] Memorization knows what it saw. Generalization takes the pattern somewhere it has never been.
