How AI Models are Trained
Training an AI model is the process of showing it enormous amounts of data and adjusting it until it gets good at a task. Here is a simplified look at how that works.
The basic loop
Imagine you are teaching a friend to sort apples: "That is a good one. That one is bruised. No, that is not bruised — look more carefully." Over many repetitions, they get better. AI training works similarly:
- The model makes a prediction
- That prediction is compared to the correct answer
- The model is adjusted slightly in the direction of better predictions
- Repeat millions or billions of times
What changes during training
Modern AI models are made of billions of numerical parameters — think of them as dials that can be turned up or down. During training, these dials are adjusted automatically to reduce errors. By the end of training, the configuration of those dials encodes everything the model has learned.
Why training is expensive
Training large models requires enormous computing power running for weeks or months. GPT-4 is estimated to have cost over $100 million USD to train. This means only a small number of very well-funded organisations can train cutting-edge models from scratch — which has significant implications for who controls AI.
Fine-tuning and adaptation
Most businesses do not train AI from scratch. Instead, they take an existing model and fine-tune it on their specific data and tasks. A New Zealand legal firm might take a general language model and fine-tune it on New Zealand case law. This is far cheaper than training from scratch and often produces excellent results for specific domains.
Once trained, a model is frozen. It does not continue learning from new conversations unless it is retrained. This is why AI can seem unaware of recent events.