@rick.theengineer: Everyone says neural networks "learn." Almost nobody says what that actually means. Here's the unglamorous version: the model makes a prediction, measures how wrong it was, works out which direction would make it less wrong, and nudges its numbers a little that way. Then it does it again. And again. And again. That's gradient descent. The reason it isn't obvious is that the model can't see the landscape it's walking on — it only ever knows the slope directly under its feet. So this whole thing is shot in the dark. You never see the shape of the valley until the last four seconds, when the lights come up on the path it took. Everything on screen is real, by the way. 12,001 houses, an actual fitted model, an actual 400-step descent. The house that's worth $300,000 and gets valued at $500,000 is a real row in the data — it's off by exactly $200,000 because the starting model really is that wrong. And the three learning rates weren't chosen to look good: they were found by searching the real loss curve until one converged in 20 steps, one needed 251, and one never settles at all. Stick around for the ending. It's the shot the rest of the video is hiding. #machinelearning #gradientdescent #neuralnetworks #datascience #aiexplained

RickTheEngineer

Open In TikTok:

Region: US

Tuesday 01 September 2026 19:20:22 GMT