The only calculus you need

In Module 2 you set a knob. In Module 3 you shrank the mistake by hand — and an algorithm did it by rolling downhill. That left one open question: at any given spot, which way is downhill? Answering it is the whole of the calculus you need here.

Which way should I move?

Picture the loss as a bowl and your knob as a ball resting on it. Drag the ball.

• On the left wall → downhill is right.
• On the right wall → downhill is left.
• At the bottom → nowhere to go; you're done.

You knew which way to move without a single equation. That instinct is the idea.

The local tilt tells you everything

Look at the short coloured line that grazes the curve under the ball — its tilt. Drag the ball and watch it rotate. That tilt carries two facts:

• which way it leans → which way is downhill (the arrow),
• how steep it is → how big a step to take.

And notice: the grazing line only cares about right here. Zoom in on any point and the curve looks almost straight — the tilt is local feedback, nothing more.

Sign → direction. Leans up to the right (a positive slope, red) → go left. Leans down to the right (negative, blue) → go right. Flat (≈ 0, green) → you're at the bottom.

Micro-challenges — drag the ball until:

1. the slope is positive (red),
2. the slope is negative (blue),
3. the slope is ≈ 0 (green, flat).

You already do this

You read slopes all day without calling it math:

• leaning into a steep road on a bike,
• easing off as the hill flattens,
• nudging a shower knob hotter — gently, as it gets close.

You constantly sense direction and steepness. That sense is the entire concept.

The same tilt, as one number

The widget shows the tilt as a picture; here it is as a single number. Edit w and run.

Predict first: at w = 0, is the slope positive or negative — and should we step left or right?

That one number — which way + how steep — is everything you need. It has a name: the slope of the curve at that point. In calculus it's called the derivative, written dL/dw ("how the loss L changes as the knob w changes"). The symbol is just shorthand for the tilt you've been dragging — a label for an idea you already own.

Good news

Almost all of modern machine learning is built from this one move:

1. measure the local slope,
2. step downhill,
3. repeat.

Deep learning does exactly this — just with millions of knobs at once instead of one. Same instinct, more dimensions. And the instinct is already yours.