KI-Glossar

The basic unit of text that language models process and are billed by.

The maximum number of tokens a model can process in a single request.

Large Language Model — a neural network trained on vast text corpora to generate human-like text.

A large pre-trained model that serves as the base for many downstream applications.

The process of running a trained model to generate outputs from new inputs.

The text output generated by a language model in response to a prompt.

The practice of designing inputs to elicit optimal outputs from language models.

An instruction block sent before the conversation that configures model behavior.

A dense numerical vector that encodes the semantic meaning of text.

A sampling parameter that controls the randomness of model outputs.

A sampling strategy that limits token selection to the smallest set covering a cumulative probability threshold.

Delivering model output token-by-token as it is generated rather than waiting for the full response.

The time between sending a request and receiving the first token of a response.

The number of tokens or requests a model can process per second.

Enhancing model responses by fetching relevant documents from an external knowledge base at query time.

A database optimized for storing and searching high-dimensional embedding vectors.

Search that retrieves results based on meaning rather than exact keyword matching.

Prompting a model to perform a task without providing any examples.

Providing a small number of input-output examples in the prompt to guide the model.

A prompting technique that instructs the model to reason step-by-step before answering.

An LLM-powered system that autonomously takes actions in pursuit of a goal.

The ability for a model to call external functions or APIs to retrieve data or take actions.

The date beyond which a model has no training data and is unaware of events.

The algorithm that converts raw text into a sequence of tokens for a language model.

The per-million-token cost charged for tokens in your prompt.

The per-million-token cost charged for tokens the model generates.

A provider feature that stores a reusable prefix in memory to avoid re-processing repeated tokens.

Submitting requests asynchronously in bulk for a 50% price discount.

An architecture where only a subset of model parameters is activated per token.

The neural network architecture underlying virtually all modern LLMs.

The core operation in transformers that lets each token attend to all other tokens.

Compressing model weights to lower numerical precision to reduce memory and speed up inference.

A model that can process and generate multiple types of data, such as text and images.

A learnable weight in a neural network; model size is measured in billions of parameters.

The standard unit for expressing model scale, where larger usually means more capable.

A model variant that produces explicit step-by-step thinking before answering.

Continuing to train a pre-trained model on domain-specific data to adapt its behavior.

The initial large-scale training phase where a model learns language from massive text corpora.

Fine-tuning a model on examples of instructions paired with ideal responses.

Reinforcement Learning from Human Feedback — training models to align with human preferences.

When a model generates plausible-sounding but factually incorrect information.

Connecting model responses to verified, real-world information sources.

The field of ensuring AI systems behave according to human values and intentions.

Anthropic's method of training models to self-critique and revise outputs using a set of principles.

An attack where malicious text in the environment overrides a model's instructions.

The discipline of building AI systems that are reliably beneficial and avoid harmful outputs.

Massive Multitask Language Understanding — a benchmark testing knowledge across 57 academic subjects.

A coding benchmark measuring a model's ability to write correct Python functions from docstrings.

A metric of how well a language model predicts a sample of text — lower is better.