Glossary ●

Short, jargon-free definitions for the terms that show up in AI news. Tap a letter to jump.

A

Agent

Software that acts on your behalf: it takes a goal, makes a plan, uses tools (a browser, a terminal, an API), checks its own work, and keeps going until the task is done. Think of it as a model with hands.

Agentic

A task where an AI agent has to carry out many steps in a row to reach a goal — writing and testing code, booking a trip, doing research. "Agentic" benchmarks measure whether the model can actually finish multi-step work, not just answer one question.

Alignment

The effort to make AI systems behave in line with human values and intentions — helpful, honest, and safe — rather than pursuing goals we didn't ask for.

API

An interface that lets your code talk to a model or service. Instead of using the ChatGPT website, a developer sends prompts to api.openai.com and gets answers back programmatically.

ARC-AGI

A hard test for AI reasoning made of brand-new puzzles the model has never seen. Because the puzzles are novel, memorization doesn't help — the model has to actually figure them out.

Attention

The mechanism inside a transformer that lets the model weigh which words in a prompt matter most to each other. It's the core trick that made modern LLMs work.

B

Benchmark

A standardized test used to score AI models — MMLU for knowledge, SWE-bench for coding, OSWorld for computer use. Vendors love quoting the ones they win; always check who ran the test and on what settings.

C

Chain of thought

Spelling out intermediate reasoning steps — "let's think step by step" — which reliably improves a model's answers on hard problems. Reasoning models do this internally by default.

Computer use

A test of whether an AI agent can operate a real computer: clicking, typing, opening apps, and finishing tasks inside a virtual desktop, the way a person would.

Context window

How many tokens of conversation a model can keep in mind at once — its working memory. A 1M-token window can hold roughly a whole codebase or several books.

D

Distillation

Training a smaller, cheaper model to imitate a bigger one. The student model learns from the teacher's answers instead of from raw data, trading a little quality for a lot of speed.

E

Embedding

A list of numbers that captures the meaning of a piece of text. Similar meanings get similar lists, which is how search-by-meaning and RAG find relevant documents.

Eval

A measurement of a model's skill, taken by running it through a fixed test. "Evals" is the industry's word for the whole discipline of testing models properly.

F

Fine-tuning

Taking a general model and continuing its training on your own examples — support tickets, your codebase, your writing style — so it gets better at your specific job.

Foundation model

The biggest, most capable general-purpose models, trained on enormous compute. GPT-6, Claude Opus, Gemini — the ones that make headlines.

Frontier model

The current top tier of models, pushing the limits of what's possible. "Frontier" moves every few months as new models ship.

Function calling

Giving a model the ability to call functions or use tools mid-answer: look something up, run code, check a calendar. The model decides when to act and folds the result back into its reply.

G

Guardrails

Rules wrapped around a model in production: block disallowed content, require human approval for risky actions, limit what tools an agent can touch. The seatbelts of AI deployment.

H

Hallucination

When a model confidently states something false. It isn't lying — it has no notion of truth, just patterns. Retrieval (RAG) and tool use are the main fixes.

Harness

The test setup around a benchmark: the prompts, tools, and scoring rules the model runs under. Scores measured with different harnesses aren't comparable, even on the same benchmark.

HLE

Humanity's Last Exam: thousands of extremely hard questions written by experts across science, math, and the humanities. One of the toughest general-knowledge tests for models.

J

Jailbreak

Breaking a model's rules with a cleverly worded prompt to get it to do something it was told not to. A constant cat-and-mouse game between attackers and model builders.

L

Latency

The delay between asking a model something and getting a full answer. Lower latency matters for voice and agents; bigger models and longer reasoning raise it.

LLM

Large Language Model: a neural network trained on vast amounts of text to predict what comes next, which turns out to enable writing, coding, reasoning, and conversation.

LMArena

LMArena (formerly Chatbot Arena): real people chat with two anonymous models side by side and vote for the better answer. The votes produce an Elo rating — the closest thing to a popular-vote leaderboard for models.

Long-horizon task

A task that takes many steps over a long period — an agent working for an hour on a coding task, for example. Long-horizon performance is where today's models most visibly differ.

M

MCP

An open standard (started by Anthropic) that lets AI apps plug into outside tools and data — your files, a database, a ticketing system — through one common protocol instead of custom integrations.

Mixture of experts

A design where a model is made of many smaller "expert" sub-models, and only a few activate for each token. Big capability with a fraction of the compute per answer.

O

Open weights

A model whose trained weights are published for anyone to download and run. "Open weights" is not the same as open source — the training data and code often stay private.

Orchestration

Working with a model across several linked apps at once — email, calendar, docs, code — chaining tools together to finish real workflows. Zapier's AutomationBench measures exactly this.

P

Parameters

The adjustable internal numbers of a neural network, learned during training. More parameters usually means a more capable (and more expensive) model. GPT-class models have hundreds of billions.

pass@k

In coding benchmarks, the fraction of tasks solved within k tries. Pass@1 (first try) is the strictest and most commonly quoted number.

Prompt

The text you send a model: your question plus any instructions. Good results come mostly from good prompts, which is why the field takes them seriously.

Prompt engineering

The craft of getting better answers by structuring what you ask: giving examples, specifying format, asking the model to show its work. Less magic trick, more clear specification.

Q

Quantization

Shrinking a model so it runs cheaper and faster — using fewer bits per number, for example. Slightly lower quality in exchange for much lower cost.

R

RAG

Retrieval-Augmented Generation: instead of relying only on what the model memorized in training, you fetch relevant documents first and hand them to the model with the question. The standard fix for hallucinations and stale knowledge.

Reasoning model

A model tuned to think step by step before answering — working through logic, checking its own work. Slower and more expensive per answer, much better at hard problems.

Red-teaming

Deliberately attacking your own AI system to find weaknesses before someone else does — the AI equivalent of penetration testing.

RLHF

Reinforcement Learning from Human Feedback: training a model by having people rank its answers and rewarding the winners. The main technique behind models that feel helpful rather than just autocomplete-y.

S

Sandbox

An isolated, restricted environment where an agent's code runs safely — no access to your real files or network unless you grant it. Where coding agents do their work.

Scaffold

The supporting code around a model that turns it into an agent: the loop that lets it plan, act, observe results, and retry. The model is the engine; the scaffold is the car.

SFT

Supervised Fine-Tuning: training a model on example conversations showing exactly how it should respond. Usually the step before RLHF.

SWE-bench

A test of an AI agent's pure coding skill: given real GitHub issues from popular open-source projects, can it write a fix that passes the tests? Reported as a percentage of issues resolved.

System prompt

The hidden instructions prepended to every conversation that set a model's role, tone, and rules — e.g. "You are a helpful coding assistant." Users never see it, but it shapes every answer.

T

Temperature

A setting (0 to 1+) controlling how random a model's answers are. Low temperature: focused and predictable. High temperature: creative and unpredictable.

Terminal-Bench

A benchmark where an AI agent works inside a real Linux terminal: installing packages, running tests, fixing bugs, submitting patches. The practical exam for coding agents.

Throughput

How fast a model produces output, measured in tokens per second. Matters a lot for interactive apps and agents that generate long responses.

Token

The basic unit a model reads and writes — roughly a word fragment. "Hello world" is two tokens. Pricing, context windows, and speed are all measured in tokens.

Top-p

An alternative to temperature for controlling randomness: the model only considers the most likely next tokens adding up to probability p. Lower top-p means safer, more focused output.

Training

The months-long, massively expensive phase where a model learns from trillions of words of text. Everything after — fine-tuning, RLHF — is comparatively cheap refinement.

Transformer

The neural-network architecture behind every modern LLM, based on the attention mechanism. The 2017 paper "Attention Is All You Need" started the whole boom.

V

Vector database

A database designed to store embeddings and find the closest matches by meaning. The memory layer that makes RAG search work.