The third rung of Module 5 · finding the bottom by feel
↓ → 0
Picture this: you're blindfolded on a hill, and somewhere down there is the spot with the least error. You can't see a thing. All you've got is your feet. Feel which way is downhill — that's just the slope from last rung — take one small step, feel again, repeat. Do that enough times and you're standing on the valley floor. That, exactly, is how an AI trains.
Why this quietly runs every AI on Earth
Swap "hill height" for "how wrong the AI's guess is," and swap "where you're standing" for "the AI's current settings." Training an AI is just this: feel the slope of the error, nudge the settings a little downhill, feel again, repeat — millions of times — until the error goes flat. No mysticism. Just a blindfolded hiker who never gets tired — and that hiker is exactly what's behind the AI writing your emails, picking your next song, and driving the car. Every "it learned" you've ever heard is this loop, running.
Step 1: you can't see the hill, only feel it
No blindfold jokes — this is the actual setup. You don't get a map of the whole hill. You only ever know one thing: what the ground feels like right where your feet are standing this instant. Tilted left? Tilted right? Flat? That's it. That's the entire information budget.
Step 2: the tilt under your feet is last rung's slope
Here's the callback: that tilt you feel is the derivative you met last rung — how steep the ground is, and which way is uphill. If it dips down to your left, downhill is left. You already know how to read it. You just never called it "gradient descent" before.
Step 3: take ONE small step downhill
Feel the tilt, step a little bit that way — not a leap, a step. Then stop and feel again, because the ground under your new spot is different. Try it on the actual hill below: press Step and watch the ball creep toward the bottom, one small nudge at a time.
Blindfolded, feeling for downhill. Press Step.
Step 4: feel again, step again, feel again, step again…
That's the whole loop, forever: feel → step → feel → step. Notice something in the demo — the steps get smaller the closer you get to the bottom. That's not you getting tired. The ground is genuinely getting flatter, so there's less "downhill" left to chase.
Step 5: flat ground means you've found a bottom
Eventually the tilt under your feet gets so close to zero it might as well be a pool table. That's not "stuck" — that's arrival. Flat ground = no downhill left = you're standing on a valley floor. You stop stepping because there's nowhere better to step to from here. One honest footnote: on a big landscape, flat means you've found a bottom — not always the lowest bottom on the whole map, because there can be other valleys past ridges you can't feel. For the enormous landscapes AI trains on, a good-enough valley is the working answer, and that's what training actually settles into.
Step 6: onward to millions of knobs at once
You just did this with one hill and one pair of feet. An AI does the exact same feel-and-step loop, but for millions of knobs at the same time. (Next rung, backprop, is how it feels the tilt for all of them at once — that part's a story for the next hill.)
Prove it · get one wrong and I’ll walk you through it
Two quick ones and the rung is yours.
Blindfolded on the hill, your left foot lands lower than your right foot. Which way do you step?
You keep stepping downhill and each step gets tinier, until the ground feels almost perfectly flat. What does that mean?
Answer both to unlock the next rung ↓
Gradient descent — mastered.
You didn't need a single equation — you felt the tilt and stepped downhill, over and over, until the ground went flat. That's the entire training loop hiding inside every neural network on Earth. Anywhere "training," "learning," or "loss going down" shows up again, you skip it.
Next rung → Backprop