How LLMs work · 1 of 141 / 3
Neural networks
The conceptHow a phone learns to spot a cat no one described to it
A neural network is layers of simple units called neurons. Each multiplies its inputs by weights (how much each input matters), adds a bias (an offset), and bends the total with an activation, e.g. "keep if positive, else 0".
Nobody writes the rules; the weights are learned. In a cat detector, pixels go in, each layer feeds the next, and the last outputs a "cat" score. Weights start random and are tuned on labelled photos until the scores match.
What it’s forComputes what one neuron outputs from its inputs.
y = f(w₁x₁ + w₂x₂ + … + wₙxₙ + b)
- xᵢ
- the inputs, e.g. pixel values or the previous layer's outputs
- wᵢ
- the weights, learned numbers saying how much each input counts
- b
- the bias, a learned offset added to the sum
- f
- the activation, e.g. ReLU: f(z) = max(0, z)
What it givesOne number that is passed on as an input to the next layer.
- Inputs 4 and 1, weights 0.5 and 2, bias −1
- Sum: 0.5×4 + 2×1 − 1 = 2 + 2 − 1 = 3
- ReLU: max(0, 3) = 3, so the neuron outputs 3