How LLMs Are Made / Glossary
How LLMs Are Made glossary
Plain-language definitions of How LLMs Are Made terms, each linked to a short visual lesson.
- AI alignment
- AI alignment is the work of making an AI system's behaviour match what people intend and value: helpful with everyday requests, honest about what it knows and careful to avoid harm.
- AI benchmark
- An AI benchmark is a fixed set of tasks with checkable answers, such as exam questions, maths problems or coding tasks, used to measure and compare how well models perform.
- Attention mechanism
- The attention mechanism is the part of a transformer that lets each token weigh how much every other token matters to it, then take in a mix of their information based on those weights.
- Base model
- A base model is a language model straight out of pretraining. It continues text instead of following instructions, so it needs further training before it works as an assistant.
- Chain of thought
- Chain of thought is the step-by-step working a language model writes out before its final answer. Reasoning models are trained to produce long chains of thought to solve harder problems.
- Context window
- A context window is the maximum amount of text, counted in tokens, that a language model can take in at once, including your messages, its replies, pasted files and hidden instructions.
- Data deduplication
- Data deduplication means finding and removing repeated copies of the same text from a dataset, so a language model does not see the same passage again and again during training.
- Embedding
- An embedding is a list of numbers that represents a word, token, sentence or image as a point in a space of meaning, so that things with similar meanings end up close together.
- Fine-tuning
- Fine-tuning means training an already pretrained model a little more on a smaller, focused set of examples, so it gets better at a task or behaves in a particular way.
- Gradient descent
- Gradient descent is the method that trains most AI models: it works out which direction lowers the loss for every weight, then nudges all the weights a small step that way, again and again.
- Knowledge distillation
- Knowledge distillation trains a small “student” model to copy the outputs of a large “teacher” model, so the student keeps much of the teacher's skill at a fraction of the size.
- Loss function
- A loss function is the formula that measures how wrong a model's prediction was. Training tries to make this number, the loss, as small as possible.
- Model card
- A model card is a short public document released with an AI model, like a nutrition label. It says what the model is for, how it was trained and tested, and its known limits and risks.
- Parameters (model weights)
- Parameters, also called weights, are the adjustable numbers inside an AI model. Training tunes them so the model's predictions get better, and together they store everything the model has learned.
- Pretraining
- Pretraining is the first and biggest stage of building a language model: it reads a huge amount of text and learns by predicting the next token, over and over, without any human labels.
- Quantization
- Quantization makes an AI model smaller and faster by storing each of its weights with fewer bits, for example 4 instead of 16, accepting a small amount of rounding error.
- Red teaming
- Red teaming is deliberately attacking your own AI system before release, trying to make it misbehave so its weak spots can be found and fixed before real users meet them.
- Reward model
- A reward model is a separate AI model that scores answers the way human raters would. It is trained on human comparisons and then used to guide a language model toward better answers.
- RLHF (reinforcement learning from human feedback)
- RLHF, reinforcement learning from human feedback, is a training method that teaches a model which answers people prefer, using many human comparisons between pairs of answers.
- robots.txt
- robots.txt is a small text file on a website that tells crawlers which pages they may and may not visit. Site owners can use it to ask AI crawlers not to collect their content.
- Scaling laws
- Scaling laws are patterns showing that a language model's loss falls smoothly and predictably as you add parameters, training data and compute. Labs use them to plan how big to build.
- Temperature (AI)
- Temperature is a setting that controls how predictable or varied a language model's output is. Low temperature makes it stick to the most likely tokens; high temperature makes less likely ones more common.
- Tokenizer
- A tokenizer is the tool that cuts text into tokens, the pieces a language model reads, and turns each piece into a number. It also turns the model's output numbers back into text.
- Tool calling (function calling)
- Tool calling is when a language model writes a structured request to use an outside tool, such as a search, calculator or calendar, and the software around it runs the tool and returns the result.
- Transformer
- A transformer is the neural network design behind almost all modern language models. It processes every token of a text in parallel and uses attention to let the tokens share information.
- Web crawler
- A web crawler is a program that visits web pages automatically, saves their content and follows their links to find more pages. Crawlers collect much of the text used to train language models.