microLoss

Interactive tutorial that explains softmax, cross-entropy loss, and gradient descent from first principles.

Live demo

What you'll learn

Walk through the complete prediction → loss → learning pipeline that runs inside every modern language model:

1. Logits — raw compatibility scores from the network
2. Exponentiation — mapping to positive values with e^z
3. Softmax — normalizing into a probability distribution
4. Target — encoding the correct answer as one-hot
5. Cross-entropy loss — measuring prediction quality with −log(p)
6. Loss in logit form — the tug-of-war between two forces
7. Gradient — the elegant p − y
8. Gradient descent — updating logits to reduce loss
9. Temperature — controlling distribution sharpness
10. Attention — softmax's second job in transformers

Running locally

bun install
bun run dev

Stack

Vue 3, TypeScript, Vite, Tailwind CSS v4, ECharts