Training

What changes when you flip the playground into training mode.

In this section

• Forward vs backward — The two sweeps per step, and what the backprop pulse shows.
• Cross-entropy loss — The objective: predicting the next token, measured in nats.
• Backpropagation — Gradients flowing output → input, one sweep per training step.
• Gradients — Reading gradient heat, and the 'largest ∇' callout.
• The optimization step — Applying the update, and the intuition behind learning rate.
• The loss curve — Reading the live cross-entropy ↓ curve as the model learns.
• Train on your own text — Training on a corpus you type, and what 'Build model' does first.