The words behind the hype, defined in Plain English

The problem of getting a system to pursue what we actually intend, rather than a proxy that merely correlates with it.

A mechanism that lets a model weigh how much each part of the input should influence each output, learned rather than fixed.

The algorithm that works out how much each weight in a network contributed to the error, by applying the chain rule backwards through the layers.

Prompting a model to work through its reasoning step by step before answering, which measurably improves multi-step problems.

The maximum number of tokens a model can consider at once — everything it is allowed to "see" for a single response.

A network that slides small learned filters across an image, building up from edges to shapes to objects.

A generative model that learns to reverse a gradual noising process, turning random noise into an image step by step.

A list of numbers representing a piece of data, arranged so that similar things end up close together.

Giving a model a handful of worked examples in the prompt so it infers the pattern you want.

Continuing to train an existing model on a smaller, specific dataset so it adapts to a particular task, domain or style.

Two networks trained against each other — one generating fakes, one detecting them — until the fakes are convincing.

The optimisation method that trains most models: repeatedly step every parameter a little way downhill on the error surface.

When a model states something false with the same fluency and confidence it uses for things that are true.

Running a trained model to get an output — as opposed to training, which is producing the model in the first place.

A neural network trained on very large amounts of text to predict what comes next, which turns out to be enough to answer questions, write code and ho...

Training a small model to imitate a large one, keeping most of the capability at a fraction of the cost.

A model built from layers of simple weighted units, trained by adjusting those weights until the output matches the target.

When a model learns the training data so closely — including its noise — that it performs worse on anything new.

Writing the input to a model deliberately — with context, examples and constraints — to get a more reliable output.

Storing model weights at lower numerical precision to cut memory and speed up inference, trading a little accuracy for a lot of resource.

A network that processes sequences one element at a time, carrying a hidden state forward — the standard approach before transformers.

Learning by acting in an environment and adjusting behaviour based on rewards, rather than from labelled examples.

Fetching relevant documents at query time and putting them in the prompt, so the model answers from your data rather than its memory.

Tuning a model using human preference comparisons, so it produces the kind of answer people actually rate highly.

Training on examples that come with the right answer attached, so the model learns to map inputs to known labels.

A setting that controls how much randomness goes into choosing each token — low is predictable, high is varied.

The unit a language model actually reads and writes — usually a word fragment rather than a word or a character.

Reusing a model trained on one task as the starting point for another, instead of training from scratch.

The neural network architecture behind almost all modern language models, built around attention instead of recurrence.

Finding structure in data that has no labels — clusters, groupings, or a compressed representation.