How to walk down a mountain you cannot see. · Source

Night, fog, no map — and a mountain to get down

The fog rolled in before the light went. Now Mira stands somewhere high on the mountainside — no map, no lamp, no stars, fog so thick she cannot see her own boots. Somewhere below is the valley: warmth, water, safety. She cannot see one step of the way down. But there is one thing the mountain cannot hide from her…

The ground tells you one thing: which way is down

The ground under her boots. She turns slowly in place, and her ankles report the tilt: this way climbs, that way falls, and one direction drops away steepest of all. One boot-sized patch of mountain — that is everything she can know. And it is enough to choose a single step. But how big should that step be?

Lunge and you overshoot; creep and dawn finds you

She tries bold: a long lunge downhill — and her boot lands climbing, past the bottom of the little dip and up its far side. Bolder leaps would carry her higher each time, not lower. Timid, then: shuffles the length of a hand — but at that pace, dawn will find her still up here. The step must fit the slope. A rhythm begins…

Feel, step, stop. Feel again.

Feel the tilt. Take one measured step down the steepest line. Stop. The ground under her boots is new now, with new advice — so she feels again, steps again. No move needs the whole mountain: each one uses only the patch she stands on. Hundreds of small honest steps, and the slope slowly pays out beneath her — until the ground goes strange…

Flat under every boot — but is it the bottom?

Suddenly every direction is level; her ankles report nothing. Down has run out — she has arrived! But the air is still thin and cold, no water sounds, no trees. It is a hollow: a small dip high on the mountain, flat enough to fool her feet. She traces the rim — such bowls are rarely sealed. One edge gives way, ever so slightly…

Her night has a name: gradient descent

\theta \leftarrow \theta - \eta \nabla L

Over the rim, down at last to running water. What carried her down is how machines learn. The mountain is a model's error; her position, its settings; the tilt underfoot, the gradient; her stride, the learning rate. Feel, step, repeat: gradient descent. The line below is her whole night — new place = old place minus stride times slope.

🌱 The fog never lifts for a machine

At dawn the fog thins and Mira finally sees the mountain she came down. A learning machine never gets that dawn: the fog is permanent, the landscape has millions of directions, and it simply stops where every way is up. She found a valley — was it the deepest one? And if the water here is sweet, how much should the question matter?