What Does "Training" Mean?

What Does "Training" Mean?

You've probably heard that AI models are "trained" on data. But what does that actually mean? What's happening underneath?

Training is the process by which an AI model learns from examples. It's not magic — it's a mathematical process of making small adjustments, millions of times, until the model gets good at its job.

The Core Idea: Learning from Mistakes

Imagine you're learning to throw darts. You throw, miss, adjust your aim, throw again. Each throw teaches you something. Over time, your aim improves.

Training an AI model works the same way, but with numbers:

1. Show the model an example (e.g., a photo labelled "cat")
2. The model makes a prediction (e.g., "I think this is a dog")
3. Measure how wrong it was — this is called the loss
4. Adjust the model's internal settings to be slightly less wrong next time
5. Repeat — billions of times, across millions of examples

After enough repetitions, the model's predictions become accurate.

What Gets Adjusted: Weights and Biases

A neural network is made of billions of numerical values called weights (and biases). These numbers control how the network processes information.

At the start of training, weights are set randomly — the model knows nothing. Every training step slightly nudges these weights in the direction that reduces the loss. This process is called gradient descent.

Think of it like rolling a ball down a hill to find the lowest point. Each step moves slightly downhill. Eventually you reach a valley — a set of weights where the model makes good predictions.

How Long Does Training Take?

Training large AI models takes:

• Huge amounts of data — frontier models train on tens of trillions of tokens (Meta has disclosed that Llama 3 was trained on about 15 trillion tokens; most labs don't disclose the exact figure)
• Massive compute — thousands of specialised chips (GPUs/TPUs) running in parallel
• Weeks to months of wall-clock time
• Enormous energy — training a large model can use as much electricity as a small town

This is why frontier AI models cost hundreds of millions of dollars or more to train, and why only a handful of organisations can do it.

When you use a model via an API or app, you're using the trained result — not the training process itself.