Topic 3 of 4
How Learning Works
How does an AI learn? It measures its error, traces improvements backwards through its computation and changes its weights step by step.
What this topic covers
This topic is about the process that turns an unpracticed computational structure into a useful model. You will trace how training tasks are created, how errors become measurable, and how many small parameter updates gradually improve behavior.
The sequence makes an otherwise hidden cycle tangible: predict, measure the error, calculate each weight’s influence, and adjust the weights deliberately.
Topic 3 of 4
Lessons in this topic
Every lesson stands on its own. In this order, the ideas connect especially clearly.
Text Becomes Many Practice Problems
Why nobody writes down answers to train a language model, how many texts become sequences of a fixed length, and which text gets in at all.
Read≈ 16 minForward Pass and Loss: How the Model Measures Its Own Error
How a language model’s percentages for the token that actually follows turn into penalty points, how a whole stack of text becomes one single number, and what that number does not measure.
Read≈ 14 minBackpropagation and Gradients: How the Model Knows What to Change
How training turns a single error number into a direction and a strength for every fader, why one pass backwards through the model is enough, and why no fader moves along the way.
Read≈ 16 minOptimization: What Actually Changes the Weights
How the optimizer turns the notes from backpropagation into real steps: with momentum, with a step size of its own for every fader, with a gentle pull toward zero and on a schedule for the learning rate.
Read≈ 16 minBatch, Epoch, Step, Token Budget: How to Measure Training Progress
How a training run is counted in steps and tokens, why a language model like GPT-3 sees some sources not even halfway and others several times, how a held-out text tells real learning apart, and how much text goes with how many faders.
Read≈ 15 min
Look it up
Terms from this topic
Short explanations of the technical terms used by the lessons already available.
- Backpropagation
- Batch
- Bias
- Context Window
- Epoch
- Forward Pass
- Gradient
- Hyperparameter
- Intermediate Value
- Label (Target)
- Layer
- Learning rate
- Learning rate schedule
- Loss (Cross-Entropy)
- Momentum
- Neuron
- Optimizer
- Overfitting
- Parameters
- Probability
- Self-supervised learning
- Softmax
- Special Token
- Step
- Token
- Token budget
- Tokenizer
- Training algorithm
- Training data
- Validation loss
- Weight decay