Module 1: What is generative AI? · 10 min
How a model learns
A neural network is a large collection of numbers, called parameters or weights. At the start they are random. Training slowly adjusts them so the model makes fewer mistakes.
- Show the model an example from the training data.
- Let it make a guess (for a language model: guess the next word).
- Measure how wrong the guess was. This measure is called the loss.
- Nudge all the weights a tiny bit to make that guess less wrong.
- Repeat billions of times.
Big models have billions of parameters and need thousands of special chips (GPUs) running for weeks. This is why training a top model costs millions of dollars, while using one is cheap by comparison.
Three things decide how good a model is: the amount and quality of data, the size of the model, and the computing power spent on training.
← AI, machine learning and generative AI | The main kinds of generative models →