Rolling downhill gradient descent

Read the slope where you stand; the gradient arrow points uphill; step the other way; repeat. Step size matters, shallow valleys catch walkers — and this walk, done on a loss, is how essentially every model trains.

step 0 Tap the curve to place a start.

Learning rate 0.15

tiny → crawl · big → bounce · past 2 on the bowl → runaway

Two walkers (quartic)

Try: the default start (x = 3.4) settles in the shallow valley. Find a start that reaches the deep one. Then step three times and press Back — the arithmetic walks backwards with you.