Golden abstract architectural details of a modern building facade in Jönköping, Sweden.

Glossary

Speak AI fluently

Plain-language definitions of the terms that matter, written to be used, not to show off. Each entry includes how we think about it in real client work.

Book a consultation Talk to us

Photo: Efrem Efre / Pexels

Models & training

Large Language Model (LLM)

An AI model trained on vast amounts of text that predicts language, letting it answer questions, write, summarize, translate, and reason.

Fine-Tuning

Continuing a trained model's training on your own data so it performs better on your specific tasks, terminology, or style than the base model.

Transformer Architecture

The neural network design behind modern AI that uses attention to weigh how parts of an input relate to each other.

Foundation Model

A large model trained on broad data at scale that serves as a base you can adapt to many different tasks instead of building from scratch.

Reinforcement Learning (RLHF)

A training method where models learn from feedback signals, including human preferences, to behave more helpfully and safely.

Training vs Inference

Training is teaching a model patterns from data; inference is using the finished model to make predictions. Each has very different costs and needs.

Model Distillation

A technique that trains a smaller, faster model to mimic a larger one, cutting cost while keeping much of the quality.

LoRA / QLoRA

Efficient fine-tuning methods that adapt a large model by training small add-on weights instead of the whole thing.

Federated Learning

Training a shared model across many devices or sites without the raw data ever leaving them; only model updates travel, not the data itself.

Synthetic Data

Artificially generated data that mimics real data, used to train or test AI when real data is scarce or sensitive.

Data Labeling

Tagging raw data with the correct answers so a model can learn from it or be measured against it. Label quality caps how good a model can get.

Annotation

Adding descriptive markup to raw data, such as boxes on images or tags on text, so it can be used to train, evaluate, or ground AI systems.

Transfer Learning

Reusing a model trained on one task as the starting point for a related task, instead of training from scratch.

Catastrophic Forgetting

The tendency of a neural network to lose previously learned abilities when it is trained on new data or a new task.

Contrastive Learning

A training approach that teaches models by pulling similar examples together and pushing dissimilar ones apart in embedding space.

Diffusion Model

A generative model that creates images, audio, or video by learning to reverse a gradual noising process, starting from pure noise.

Domain Adaptation

Adapting a model trained on general data so it performs well in a specific field, such as medicine, law, or a company's own vocabulary.

Early Stopping

Halting model training when performance on held-out data stops improving, preventing overfitting and wasted compute.

Graph Neural Network

A neural network designed for data shaped as networks of connected entities, learning from relationships as well as attributes.

Instruction Tuning

Fine-tuning a language model on instruction and response examples so it follows directions instead of merely continuing text.

Mixture of Experts (MoE)

A model architecture that routes each token to a small subset of specialized subnetworks, giving huge capacity without proportional compute cost.

Sparse Attention

Transformer attention variants where each token attends to a selected subset of others instead of all of them, making long contexts affordable.

Techniques

Retrieval-Augmented Generation (RAG)

A pattern that fetches relevant documents from your data and feeds them to a model so answers are grounded and citable.

Prompt Engineering

The practice of crafting instructions, examples, and context that reliably steer a model toward the output you want.

Tool Use (Function Calling)

Letting an AI model call external functions, APIs, or systems so it can take real actions and fetch live data instead of only generating text.

Chain of Thought

Prompting a model to reason step by step before giving its final answer, which measurably improves accuracy on math, logic, and hard problems.

Semantic Search

Search that matches on meaning rather than exact keywords, so it finds relevant results even when wording differs.

Temperature

A setting that controls how random or predictable a language model's output is, from focused and repeatable answers to varied and creative ones.

Active Learning

A training approach where the model picks the most useful examples for humans to label next, so you get more accuracy from far less labeling.

Few-Shot Learning

Teaching a model a new task with just a handful of examples shown in the prompt, no retraining needed. Often enough to reach useful accuracy.

Zero-Shot Learning

Getting a model to perform a task with no examples at all, working purely from an instruction, thanks to what it learned during pretraining.

Deterministic Output

Model behavior where the same input reliably produces the same output, essential for testing, auditing, and automated pipelines.

Ensemble Methods

Combining predictions from multiple models so their collective answer is more accurate and robust than any single model alone.

Grounding (AI)

Anchoring a model's answers in verified, current sources such as your documents or databases, rather than its training memory.

JSON Mode

A model API setting that constrains output to valid JSON, so software can parse responses reliably instead of scraping prose.

Prompt Caching

Reusing the computation for repeated prompt prefixes across requests, cutting cost and latency on the static parts of prompts.

Retrieval-Augmented Fine-Tuning

Fine-tuning a model specifically to work well with retrieved context, combining RAG's fresh knowledge with training's learned behavior.

Structured Output

Making a model return data in a guaranteed schema, typed fields and formats, so applications can consume AI results like any API response.

Core concepts

AI Agent

An AI system that plans multiple steps, uses tools and APIs, and works a task through to completion instead of just answering a single question.

Agentic AI

AI that pursues goals autonomously by planning, using tools, and adapting across multiple steps with minimal hand-holding.

Embeddings

Numeric representations of text or images that place similar meanings close together, enabling search and retrieval.

Hallucination

When an AI model states something false or made-up while sounding completely confident, presenting invented facts as if they were true.

AI Alignment

The work of making sure an AI system actually pursues the goals and values its operators intend, rather than optimizing for the wrong thing.

Multi-Modal AI

AI that can understand and combine more than one type of input, such as text, images, audio, and video, within a single model or system.

Knowledge Graph

A structured map of entities and the relationships between them, stored so that machines can query, connect, and reason over the connections.

AI Operating System

A unified layer that connects models, data, tools, and workflows so AI runs reliably across an entire organization.

Natural Language Processing (NLP)

The field of AI focused on getting computers to understand, interpret, and generate human language, from search and translation to chatbots.

Computer Vision

The field of AI that lets computers interpret images and video, identifying objects, text, and patterns in them.

Token

The basic chunk of text a language model reads and writes, often a word, part of a word, or punctuation mark. Models price and limit usage in tokens.

Context Window

The maximum amount of text, measured in tokens, a model can consider at once. Anything beyond the window is forgotten or must be summarized.

AI Readiness

How prepared your data, people, processes, and infrastructure are to adopt AI successfully, before you commit budget to building anything.

Chatbot vs AI Agent

A chatbot answers questions in conversation; an AI agent plans and takes multi-step action with tools. The difference shapes cost, risk, and value.

Autonomous Agent

An AI system that pursues a goal end to end, planning steps, calling tools, and correcting course without a human driving each action.

Data Flywheel

A compounding loop where product usage generates data that improves the AI, which attracts more usage and generates still more data.

Latent Space

The internal coordinate system a model learns, where data is represented as points and distance reflects similarity of meaning.

Multi-Agent System

An architecture where several specialized AI agents collaborate on a task, coordinated by an orchestrator or shared protocol.

Orchestration Layer

The software layer that coordinates models, data retrieval, tools, and business logic into one dependable AI application.

Operations & MLOps

MLOps

The engineering practice of deploying, monitoring, and maintaining AI systems so they stay reliable in production.

Vector Database

A database built to store embeddings and find the most semantically similar items fast, powering search and RAG.

Inference

The act of running a trained model to produce an output, such as generating an answer from a prompt, as opposed to the training that built it.

Quantization

Shrinking a model by storing its numbers at lower precision, making it faster and cheaper to run with minimal quality loss.

Data Drift

When the live data a model sees gradually diverges from what it was trained on, quietly eroding accuracy until someone measures and retrains.

Model Monitoring

Continuously tracking a deployed model's quality, behavior, and cost so problems are caught before users feel them.

AI-as-a-Service (AIaaS)

Renting AI capabilities through cloud APIs and subscriptions instead of building, training, and hosting the models yourself, paying as you go.

Edge AI

Running AI models directly on local devices such as sensors, cameras, and phones instead of sending data to the cloud for processing.

Benchmark

A standardized test or dataset used to measure and compare how well AI models perform on a task, from reasoning to coding to factual recall.

Evaluation

The process of measuring whether an AI system produces correct, safe, and useful outputs for its intended task.

Latency

The time it takes an AI system to return a response after receiving a request, usually measured in milliseconds or seconds.

Throughput

The volume of requests or tokens an AI system can process in a given period, such as requests or tokens per second.

Cost per Token

The price charged for each unit of text an AI model reads or generates, the core pricing unit for most language models.

Batch Inference

Running model predictions on many inputs at once as a scheduled or queued job, trading response speed for much lower cost.

Canary Deployment

A release strategy that sends a small slice of real traffic to a new model or prompt first, watching metrics before full rollout.

Data Mesh

A decentralized data architecture where domain teams own and publish their data as products, instead of routing everything through one central team.

Edge Inference

Running model predictions directly on local devices, phones, cameras, factory machines, instead of sending data to the cloud.

Feature Store

A central platform that stores, documents, and serves the computed data signals models use, consistently for both training and live predictions.

GGUF Format

The file format used by llama.cpp for packaging quantized language models into a single file that runs efficiently on ordinary hardware.

Model Registry

A version-controlled catalog of an organization's models, tracking lineage, approvals, and which version is running where.

ONNX Runtime

A cross-platform engine that runs models exported to the open ONNX format, decoupling how a model is trained from where it runs.

Governance & compliance

AI Governance

The policies, roles, and controls that ensure AI is used safely, lawfully, and accountably across an organization.

AI Red Teaming

Deliberately attacking your own AI system with adversarial prompts and misuse scenarios to find failures before real users or attackers do.

Guardrails

Rules, filters, and checks that constrain what an AI system is allowed to say or do, keeping its behavior inside safe and approved boundaries.

AI Safety

The practice of building AI systems that behave reliably and avoid harm, both today and as they grow more capable.

Responsible AI

A set of practices for building AI that is fair, transparent, accountable, and respectful of privacy and people.

Explainable AI (XAI)

Techniques that make an AI system's outputs understandable, so people can see why it decided what it did and challenge or trust the result.

Bias in AI

Systematic unfairness in AI outputs, usually inherited from skewed training data or flawed design choices, that harms specific groups or decisions.

SOC 2 for AI

Applying the SOC 2 trust criteria for security, availability, and confidentiality to AI systems and their data.

HIPAA for AI

Meeting US health-privacy rules when AI systems handle protected health information, covering safeguards, access controls, and vendor agreements.

EU AI Act

The European Union's law that regulates AI by risk level, from banned uses to strict rules for high-risk systems.

NIST AI RMF

A voluntary US framework from NIST for identifying, measuring, and managing AI risks across the full lifecycle of an AI system.

ISO 42001

The first international standard for managing AI responsibly. It defines an auditable AI management system, much like ISO 27001 does for security.

PCI-DSS for AI

Applying payment card security rules to AI systems that touch cardholder data, so models and pipelines meet the same controls as other systems.

PIPEDA for AI

Meeting Canada's federal privacy law when AI systems handle personal information, including consent, safeguards, and limits on how data is used.

AI Policy

Your organization's written rules for how AI may and may not be used, covering approved tools, data handling, review, and accountability.

Shadow AI

Unsanctioned AI tools employees adopt without IT or security oversight, creating data leakage and compliance risks the organization cannot see.

Constitutional AI

A training method where a model critiques and revises its own outputs against a written set of principles, reducing reliance on human labelers.

Model Card

A standardized document describing what a model does, how it was built and tested, where it works, and where it should not be used.

AI Watermarking

Embedding detectable signals in AI-generated content, or provenance metadata alongside it, so machine-made material can be identified later.

Let's build the intelligence that moves your business.

Tell us where you're headed. We'll show you what's possible, and exactly how we'd get there together.

Book a discovery call Talk to us