AI Jargon Buster: 10 Terms You'll Hear This Week

Before an AI system can do anything useful it has to learn, and that process involves four terms that get used constantly and often imprecisely. Training data is the collection of examples an AI system learns from: for a language-based AI this might be billions of sentences from books and websites, while for a medical imaging AI it might be thousands of labelled X-rays. The quality, diversity, and accuracy of training data directly shape what the resulting system can and cannot do. A persistent misconception is that AI tools learn from your conversations in real time, updating their knowledge as you chat; in standard deployment this is not how it works, because the training data was fixed before you ever opened the app. Think of it like a student who revised extensively before an exam: the revision happened beforehand, and answering your questions does not add new pages to the textbook. A model is the mathematical structure that encodes the statistical patterns found in the training data. When people say the AI said something, they are usually referring to the model underneath, and the important distinction is that the model is not the same as the product you are using: ChatGPT, Copilot, and Gemini are products built on top of models, where the model is the engine and the product is the car. Inside every model are parameters, sometimes called weights, the internal numerical values that were adjusted during training to capture patterns in the data, and a model might have billions of them. Parameters are fixed once training is complete and determine how the model responds to any given input, though most deployed systems add a degree of controlled randomness when generating, so identical prompts can still produce different wordings. The common misconception is that more parameters always means a smarter model; in practice a well-trained smaller model can outperform a poorly trained larger one, so parameter count is a measure of scale, not a guarantee of quality. A useful analogy is to think of parameters as the dials on a mixing desk in a recording studio, each controlling a specific aspect of the output. A skilled engineer sets those dials during the recording session, which is training, and once the record is pressed, which is deployment, the dials are locked, so playing the record does not change the settings. Fine-tuning is the process of taking an existing trained model and giving it focused additional training on a narrower task or domain, such as fine-tuning a general-purpose model on medical literature for clinical contexts. What fine-tuning does not do is completely rewrite the model: the base model's knowledge and capabilities remain, and fine-tuning adjusts emphasis and adds specialised competence. It is less starting over and more continuing education.