Large Language Model

The cycle of model turn, tool call, observation, repeat — the basic mechanism of an AI agent.

A stated number pulls subsequent estimates toward it, regardless of relevance.

Turning audio into text, and the reasons real-world accuracy is set by the signal more than by the model.

Prompting a model to write out intermediate reasoning steps before its final answer, which measurably improves accuracy on multi-step problems.

Restricting a model's next-token choices to those a formal grammar permits.

The bounded span of tokens a model can attend to in a single request.

A numeric vector representing a piece of content such that semantic similarity between items becomes geometric distance between vectors.

A repeatable test suite for model behavior, since prompts and tool descriptions cannot be type-checked.

Continuing to train an already-trained model on a narrower dataset to specialize its behavior, instead of prompting the base model.

A hash over everything that determined a result, used to detect when the result has gone stale.

A processor built to run the same small program on thousands of data items at once, designed for pixels and now the engine of machine learning.

A rule constraining what a model may do, enforced by prompt, schema, or surrounding code.

Fluent, confident output that is not grounded in anything real.

A vocabulary for describing the shape of JSON data, and the lingua franca of model tool interfaces.

The stored attention keys and values for every token already processed, which lets a model generate each new token without re-reading the whole sequence.

A process whose next state depends only on the present state, and the basis of pre-neural text generation.

Reverse-engineering the computation a trained network learned into human-legible features and circuits.

Choosing which model handles which step, trading capability against cost and latency.

The property that identical inputs may produce different outputs, and what it costs to test around.

Ratings and written opinions that customers post about products, places and services, which have become one of the main inputs to buying decisions and, for the same reason, a constant target for fakes, bribes and coordinated attacks.

The text supplied to a language model to elicit a response.

Shaping a model's input to make the desired behavior reliable rather than occasional.

Untrusted content that reaches a model's context and is treated as instruction.

Storing a model's weights at lower numeric precision to shrink memory and speed inference, at some cost to accuracy.

Language models trained to think at length before answering, trading tokens and latency for accuracy on hard problems, and the second axis of scaling after model size.

Capturing a real interaction once and replaying it deterministically in later test runs.

Fetching relevant documents at query time and feeding them into the prompt, so a model answers from retrieved text rather than memory.

Checking data against a declared shape at runtime, and using the result as the typed value.

Constraining a model to emit data conforming to a declared schema rather than free text.

The standing instructions that frame every turn of a conversation with a model.

A parameter that scales how sharply a model favors its most likely next token, trading determinism for variety.

A stand-in for a real dependency during a test — and the ways a stand-in can lie.

Turning text into spoken audio, from robotic formant synthesizers to neural voices that clone a speaker from seconds of audio, and the latency and consent questions that come with them.

The sub-word unit a language model actually reads and writes; the unit of cost and of context.

The finite allowance of model usage a task may consume, and the design decisions it forces.

Giving a model a set of callable functions so it can act on the world rather than only describe it.

A datastore built to find the nearest neighbors of a high-dimensional vector quickly, at a scale exact search can't reach.