- agent
- An AI system that takes several steps toward a goal, often planning, using tools, or calling other models along the way.
- backpropagation
- The method that works out how much each parameter contributed to the error, so training knows which way to adjust them.
- benchmark
- A standard test used to compare how well different AI models perform on the same task.
- bias
- Systematic unfairness in a model, often because the data it learned from was unbalanced, which can mean worse results for some groups of people.
- convolution
- A small filter slid across an image to pick up local features like edges and textures. The core operation in most vision models.
- diffusion model
- A way to generate images by starting from pure noise and removing it step by step until a picture appears. The method behind most modern image generators.
- embedding
- A way of turning words, images, or papers into lists of numbers so a computer can measure how similar they are.
- fine-tuning
- Taking an already-trained model and training it a bit more on a specific task or style.
- generative model
- A model that creates new content, such as images or text, rather than sorting or scoring things that already exist.
- gradient descent
- The basic training loop: nudge the model a tiny step in the direction that lowers its error, then repeat many times.
- hallucination
- When an AI states something false with confidence. Common in language models because they predict fluent text, not verified facts.
- inference
- Actually using a trained model to make a prediction or generate an answer (as opposed to training it).
- neural network
- A model loosely inspired by the brain, made of layers of simple units that pass numbers to each other to learn patterns.
- overfitting
- When a model memorises its training examples instead of the general pattern, so it looks great in practice and fails on anything new.
- parameter
- One of the adjustable numbers inside a model. Big models have billions of them.
- reinforcement learning
- Learning by trial and error: an agent takes actions, gets rewards, and works out a strategy that earns more reward.
- tokenization
- Splitting text into the small chunks, called tokens, that a language model actually reads. A token is often a word piece, not a whole word.
- training
- The process of showing a model many examples and nudging its internal numbers until it gets good at a task.
- transformer
- A neural-network design that learns which parts of the input to pay attention to. It powers most modern language models.