The Glossary

The vocabulary, defined once.

Every specialist term the briefs use, in plain language. The list grows on its own: as the newsroom reads and briefs new coverage, the writer flags terms a senior reader outside that subfield might not know, and they land here — 36 so far. Terms in briefs link back to this page.

A

Agent

An AI system that takes actions toward a goal — calling tools, browsing, writing files, spawning subtasks — rather than only answering a prompt. "Agentic" describes workflows built this way.

ARC-AGIAbstraction and Reasoning Corpus

A benchmark of novel visual-logic puzzles designed to resist memorization, used to measure abstract reasoning rather than recall. ARC-AGI-3 is the current generation.

B

Benchmark

A standardized test suite for comparing models. Useful but gameable: vendor-reported scores on their own benchmarks deserve the same skepticism as any first-party number.

C

Context window

The maximum amount of text (measured in tokens) a model can consider at once — its working memory. Everything the model "knows" about your request must fit inside it.

CVECommon Vulnerabilities and Exposures

The standard public identifier for a security vulnerability (e.g. CVE-2026-5674). A CVE number means the flaw is cataloged and referenceable, not that it is fixed.

D

Diffusion model

The model family behind most AI image and video generation: it learns to turn noise into coherent output step by step.

Distillation

Training a smaller or different model to imitate a larger one's outputs — transferring capability without transferring the original weights or training data.

DMCADigital Millennium Copyright Act

The US law governing copyright online, including the anti-circumvention rules that scraping and AI-training cases often turn on.

E

Embedding

A numeric vector representing a piece of text's meaning, letting software measure similarity between texts — the basis of semantic search and RAG retrieval.

F

Fine-tuning

Continuing a model's training on a narrower dataset to specialize it for a task, domain, or style.

Frontier model

A model at or near the current capability ceiling — the class trained by the major labs at the largest scale. "Near-frontier" means approaching that tier at lower cost.

G

Guardrails

The layered controls around a deployed model — filters, policies, sandboxes, permission limits — meant to keep its behavior inside intended bounds.

H

Hallucination

A confident model output that is factually wrong or entirely invented. The failure mode that grounding, retrieval, and verification pipelines exist to catch.

Harness

The scaffolding around a model call — prompts, formatting, retries, tool wiring, context management. Harness design can swing task accuracy by double digits on the same model.

I

Inference

Running a trained model to produce output, as opposed to training it. Inference cost and speed determine what a deployed AI product costs to operate.

J

Jailbreak

A prompt or technique that gets a model to bypass its safety training and produce output it would normally refuse.

L

LoRALow-Rank Adaptation

A cheap fine-tuning method that trains small add-on matrices instead of the full model, making customization feasible on modest hardware.

M

MaaSModel-as-a-Service

Selling access to a model through a hosted API rather than shipping the weights. Open-weights licenses sometimes restrict MaaS use specifically — the model is free to run, not necessarily free to resell.

MCPModel Context Protocol

An open standard for connecting AI models and agents to tools, data sources, and services — the plumbing that lets an agent read a database or file a ticket without custom integration for every pair.

MoEMixture of Experts

A model architecture that routes each token through only a small subset of its parameters ("experts"), so a very large model runs at a fraction of its full compute cost. A "750B A37B" model has 750 billion parameters but activates about 37 billion per token.

Multimodal

A model that works across more than one medium — text, images, audio, or video — in input, output, or both.

O

Open weights

A release where the trained model parameters are downloadable and runnable locally. Distinct from open source: the weights may come with use restrictions, and training data and code usually stay private.

P

Post-training

Everything done to a model after its main pretraining run — instruction tuning, RLHF, safety training — which shapes how it behaves rather than what it knows.

PQCPost-quantum cryptography

Encryption algorithms designed to resist attacks by future quantum computers; NIST runs the selection process for standardized candidates.

Prompt injection

Hiding instructions inside content an AI will read (a document, email, or web page) so the model follows the attacker's directions instead of its operator's. The foundational attack class against agents.

Q

Quantization

Storing model weights at lower numeric precision to shrink memory use and speed up inference, usually with a small quality cost — the trick behind running large models on laptops.

R

RAGRetrieval-Augmented Generation

Fetching relevant documents at question time and placing them in the model's context so answers are grounded in current, specific material instead of only training data.

Red-teaming

Deliberate adversarial testing of a model or AI system to find failures — jailbreaks, injections, unsafe outputs — before attackers or users do.

RLHFReinforcement Learning from Human Feedback

A post-training technique that shapes model behavior using human preference judgments — a large part of why chat models follow instructions and decline harmful requests.

RSIRecursive self-improvement

AI systems being used to accelerate AI research itself — the feedback loop where models help design, train, or debug their successors. The scenario named in 2026's frontier-lab "pacing" letter as the capability to slow down.

S

Sandbox

An isolated execution environment that limits what running code — including AI-agent actions — can touch. A "sandbox escape" breaks that isolation.

Scaling laws

The empirical finding that model capability improves predictably with more parameters, data, and compute — the economic logic behind ever-larger training runs.

Session affinity

A server-architecture constraint where a client must keep talking to the same server instance because state lives there — a scaling bottleneck that stateless protocol designs remove.

T

Token

The unit models read and write — roughly three-quarters of an English word. Pricing, context limits, and speed are all measured in tokens.

Transformer

The neural-network architecture underlying modern language models, built on an "attention" mechanism that lets every token weigh every other token in context.

Z

Zero-day

A vulnerability exploited before the vendor knows about it or has shipped a fix — meaning defenders had zero days of warning.