z → 0 or z…never negative

Why this quietly runs your life

Your phone unlocking because it recognized your face, the doctor's scanner flagging a suspicious mole, the spam filter catching that scam email, the voice in your car understanding "navigate home" — every one of those runs on nothing but this: multiply, add, one blunt yes/no switch, repeated by the billions.

That's the whole "brain" everyone's a little scared of. Not some inscrutable alien intelligence — just this same tiny decision you're about to build, wired into a network big enough to notice patterns no single neuron could ever catch alone.

Learn this one unit, and you own the atom of every AI system running today.

Step 1: a neuron is just a tiny decision, fed by clues

Say you're deciding whether to grab an umbrella. You've got clues: cloud cover, forecast percent, how the wind feels. Some clues you trust more than others. A neuron formalizes exactly this: a handful of input numbers, each backed by how much you trust it.

Step 2: multiply each clue by how much you trust it — the weight

Each input x gets its own weight — a number saying how much that input matters. Trust it a lot, the weight is big. Barely care, the weight sits near zero. Actively distrust it, the weight goes negative. You already did this move in Multiplication: x × w.

Step 3: add them all up — the weighted sum (hi, Softmax)

Multiply every input by its weight, then add every product together.

Recognize that? It's the exact weighted sum move from Softmax — before the exponent, before the probabilities, there was this same add-it-up step. Same arithmetic, new job. One extra passenger rides along at the end: the bias b, the neuron's built-in head start (or handicap) before a single clue arrives. Training tunes it exactly like the weights. (The playground below keeps b = 0 so the sliders tell the whole story.)

Step 4: the switch — meet ReLU

Now the one new idea. z hits a bouncer called ReLU (Rectified Linear Unit — ignore the name, it's a doorman). The rule is dumber than it sounds:

Positive vibe? You're in, unchanged. Zero or negative? Bounced — flat zero, no negotiating. Drag the sliders below and watch the neuron fire (or not).

Step 5: stack a billion of these and you basically get a brain

One neuron alone is trivial. Wire thousands of them into layers, each layer feeding the next, and you get a neural network. Here's why the switch matters so much: without that little kink at zero, stacking layers would collapse into one boring flat calculation — no smarter than a single layer. The kink is what lets depth actually buy you something.

That's the whole secret behind "deep" in deep learning. Not magic — multiply, add, kink — repeated at billion-neuron scale.