AI Glossary

A hypothetical AI capable of understanding, learning, and performing any intellectual task at a human level or beyond — unlike today's models, which excel in specific domains but lack general-purpose intelligence. AGI has not yet been achieved, and its timeline is a subject of debate in the industry.

An LLM-based system that can not only answer questions but also independently plan steps, call external tools (search, code, APIs), and carry out multi-step tasks to reach a goal with minimal human involvement.

A situation where a model confidently states false or entirely made-up information as fact — for example, a nonexistent quote or fabricated statistic. It's the main reason AI-generated answers should always be fact-checked.

A technique where the model "thinks out loud" step by step before giving its final answer, instead of producing a result immediately. It noticeably improves accuracy on tasks that require logic or math.

The maximum amount of text (measured in tokens) a model can "see" at once — including the prompt itself, the conversation history, and any attached files. The larger the window, the longer the documents the model can process in one go.

A type of generative model that learns to create images (or other content) by gradually "cleaning up" random noise until a coherent picture emerges. It underlies most image generators, including Midjourney, Stable Diffusion, and DALL-E.

A numeric representation of text, an image, or another object as a vector, where semantic similarity between objects is reflected in the closeness of their vectors. It's used to find similar documents and underlies RAG and vector databases.

A model's ability to perform a new task after seeing just a few examples (few-shot) or with no examples at all, based purely on a text description of the task (zero-shot) — without any additional training on new data.

Additional training of an already-trained model on a narrow dataset to adapt it to a specific task or style — for example, legal documents or a brand's tone — without training the model from scratch.

A large model trained on a broad dataset that serves as the base for many narrower applications through fine-tuning or prompting — instead of training a separate model from scratch for every task.

The process of using an already-trained model to generate a response to a new input — as opposed to training, when the model adjusts its parameters. Inference is what happens every time you send a message to a chatbot.

A neural network trained on massive amounts of text that can understand and generate human language. Examples include GPT, Claude, Gemini, and Grok. It powers most modern chatbots and AI assistants.

An open standard introduced by Anthropic that lets AI models connect to external data sources and tools — databases, file systems, APIs — in a uniform way, without writing a separate integration for every model-and-service combination.

The numeric parameters inside a neural network that are adjusted during training and determine how the model processes input. "Open-weight" means these parameters can be downloaded so anyone can run the model themselves.

A model that can understand and/or generate multiple types of data at once — text, images, audio, and video — rather than just text. For example, it can analyze a photo and answer a question about it in text.

A mathematical model loosely inspired by the connections between neurons in the brain: many layers of simple computational units that together learn to find complex patterns in data. It's the foundation of nearly all modern AI systems.

The art and practice of crafting prompts to an AI so that it produces the most accurate and useful response. It includes phrasing the task clearly, adding examples and context, and specifying the desired output format.

A technique where the model first retrieves relevant documents from an external knowledge base, then uses them as context to generate its answer. It reduces hallucinations and lets the model answer questions about recent or highly specialized data.

A fine-tuning method where humans rate the quality of a model's responses, and the model is adjusted to more often produce answers that humans rate highly. It's a key step in making an AI model helpful and safe, not just technically functional.

The smallest unit of text a language model processes — usually part of a word, a whole word, or a punctuation mark. Usage costs for most AI models are billed by the number of input and output tokens.

The neural network architecture introduced in the 2017 paper "Attention Is All You Need" that underlies almost every modern LLM. Its key idea is the attention mechanism, which lets the model weigh the relationships between all words in a text simultaneously.

A database optimized for storing and quickly searching embeddings (vectors) by semantic similarity rather than exact text matches. It's a key component of RAG systems and semantic search.