99+ AI Terms, Actually Explained

Half the confusion in any AI conversation is vocabulary, not concepts -- once the terms are actually defined, most of what sounds complicated turns out to be a small number of ideas wearing different names.

The foundation layer starts with artificial intelligence itself: systems built to perform tasks that typically require human cognition -- reasoning, learning, perception, decision-making. Machine learning (ML) is the subset where a system learns patterns from data rather than following rules a person wrote by hand. Deep learning (DL) is the ML approach that uses multi-layered neural networks to learn hierarchical representations from large datasets, and a neural network itself is the underlying computational model: interconnected nodes, loosely called neurons, organized in layers that pass information forward.

Most of today's frontier systems share an architecture called the transformer, which uses self-attention to weigh the relevance of every part of an input against every other part -- it's the mechanism modern large language models (LLMs) are built on. An LLM is a neural network trained on massive amounts of text to understand and generate language; a foundation model is the broader category of any large model trained on broad data and adaptable to many downstream tasks. What a model actually learns lives in its parameters (the learnable weights set during training), tuned according to hyperparameters chosen before training even starts (learning rate, batch size, and similar settings). Training runs on training data over repeated epochs (one full pass through the dataset), processed in batches, guided by a loss function that measures how wrong the model's predictions are and gradient descent, the optimization method that nudges parameters to reduce that error, propagated backward through the network via backpropagation. A reasoning model is specifically trained to work step by step before answering, while a small language model (SLM) trades some capability for being compact enough to run on a phone or at the edge.

An AI agent is a system that perceives its environment, reasons, and takes autonomous action toward a goal, and agentic AI describes a model capable of multi-step planning and independent task execution rather than a single one-shot response. Tool use is an agent's ability to call external functions, APIs, or services to complete a task, often guided by chain-of-thought prompting -- techniques that elicit a model's step-by-step reasoning before it answers. Retrieval-augmented generation (RAG) grounds a model's output in real external knowledge fetched at query time instead of relying purely on training data, and the Model Context Protocol (MCP) is the open standard for connecting models to those external tools and data sources consistently. Agent memory is persistent context an agent keeps across sessions; an agentic workflow is a sequenced, multi-step process an agent runs autonomously; computer use is an agent's ability to control a computer interface directly rather than through an API; spec-driven development means an agent executes from a detailed specification instead of hand-written code; and human-in-the-loop (HITL) is the governance pattern where a person reviews or approves an agent's decisions at critical steps rather than letting it run fully unsupervised.

Generative AI is any AI that creates new content -- text, images, audio, video -- rather than just classifying or predicting from existing data. A prompt is the instruction that directs what a generative model produces; a token is the basic unit of text a model actually processes; the context window is the maximum amount of text a model can consider at once; and temperature is the parameter controlling how random or deterministic that output is. Inference is the act of using an already-trained model to generate a prediction on new input, as opposed to training it. A diffusion model learns to reverse a noise-adding process to create an image from randomness, while an autoregressive model generates output sequentially, one token at a time, each conditioned on everything generated so far -- the approach behind most text-generating LLMs.

How a model actually learns falls into a handful of paradigms. Supervised learning trains on labeled data where every input is paired with the correct answer; unsupervised learning finds hidden patterns in unlabeled data without being told what to look for; reinforcement learning (RL) learns through trial and error, optimizing actions based on reward signals; and self-supervised learning generates its own labels from the input data itself rather than needing a human to label anything. Transfer learning applies knowledge learned on one task to improve performance on another; fine-tuning adapts a pre-trained model to a specific task using smaller, targeted data; zero-shot learning performs a task with no task-specific training examples at all; and few-shot learning does it from a very small number of examples.

Natural language processing (NLP) is the broader field focused on understanding, interpreting, and generating human language. Embedding is a dense numerical representation of meaning, positioned so similar meanings sit close together in that space, which is what makes semantic search -- finding results by meaning rather than exact keyword match -- possible. Tokenization splits text into the tokens a model processes, and the attention mechanism is what allows a model to weigh the importance of different input parts when producing output. Reinforcement Learning from Human Feedback (RLHF) is the training technique that uses human preference ratings to align a model's behavior with what people actually want, rather than just what it learned to predict from raw text.

Computer vision is AI focused on interpreting visual information from images and video. Multimodal AI extends that to models that process and relate multiple data types together -- text, image, and audio at once. Optical character recognition (OCR) converts text visible in an image into machine-readable text; image segmentation partitions an image into distinct regions; and object detection identifies and locates specific objects within an image.

Several terms describe the specific ways models go wrong. Hallucination is a model generating plausible-sounding but factually incorrect content, confidently, with no signal anything is wrong. Overfitting is a model memorizing its training data so closely that it fails to generalize to new inputs; underfitting is the opposite, a model too simple to capture the underlying pattern at all. Bias refers to systematic errors baked in by flawed training data or design choices, and alignment is the ongoing work of ensuring a model's behavior actually matches human values rather than just optimizing for whatever its training objective technically rewards. Jailbreaking is an attempt to bypass a model's safety constraints through crafted prompts, and prompt injection is a related but distinct attack: malicious input designed to override a model's original instructions, often hidden inside content the model is asked to process. Drift is the gradual degradation of a deployed model's performance as real-world data diverges from what it was trained on, and prompt leaking is the unintentional exposure of a model's system prompt through what a user sees in its output.

Judging whether any of this works relies on a consistent set of measures. A benchmark is a standardized test comparing model performance across systems; accuracy is the proportion of predictions that are correct overall; precision is, of the things a model flagged as positive, how many actually were; recall is, of everything actually positive, how many the model caught; and the F1 score is the harmonic mean of precision and recall, used when one number needs to penalize a model for being lopsided toward either. Perplexity measures how well a language model predicts a sample of text -- lower is better. BLEU score evaluates generated text against a reference translation or summary, and ground truth is the verified, correct answer a model's output gets checked against.

Underneath the model sits infrastructure most users never see. An API is the interface software uses to talk to a model programmatically; latency is the delay between a request and a response; throughput is how many requests a system handles per unit of time; a GPU accelerates the parallel computation training and inference depend on, and a TPU is Google's own purpose-built version of the same idea. Quantization reduces a model's numeric precision to shrink its size and speed it up, usually trading away a small amount of accuracy; distillation trains a smaller model to mimic a larger one at a fraction of the cost; and edge AI means running models locally on a device rather than a cloud server. A vector database stores and retrieves the high-dimensional embeddings that power RAG, and inference cost is the ongoing compute expense of running a trained model in production, distinct from the one-time cost of training it. "Vibe coding" is the newer, informal term for building software by describing what you want in plain language and letting an AI agent write the actual code.

A separate cluster of terms covers accountability rather than raw capability. Explainable AI (XAI) covers methods for making a system's decisions genuinely interpretable to a person. Red teaming is adversarial testing meant to find a model's vulnerabilities before real attackers or users do. Constitutional AI, Anthropic's own training approach, uses a written set of principles to guide a model's behavior during training rather than relying purely on human-labeled preference data. Guardrails are constraints built in to prevent harmful outputs; data privacy covers protecting personal information used in training or inference; a model card is documentation describing a model's capabilities and intended use; sandboxing isolates a system so its actions can't affect production without explicit permission; and dual-use risk names the reality that many AI capabilities can serve both legitimate and harmful purposes.

A handful of terms describe where the field is headed. Latent space is the compressed internal representation where a model organizes concepts, similar ones clustered near each other. Emergent abilities are capabilities that show up in larger models but weren't present in smaller versions of the same architecture -- a real observed pattern, though researchers still debate how much of it is a genuine qualitative shift versus an artifact of how it's measured. Scaling laws describe the fairly predictable relationships between model size, data, compute, and performance that labs use to plan a larger model's likely capability before training it. In-context learning is a model adapting to a new task purely from examples given in the prompt, without updating its weights; synthetic data is artificially generated data used to train or augment training when real data is scarce. A mixture of experts (MoE) is an architecture where specialized subnetworks handle different kinds of input, activating only the relevant ones per request. AGI is the still-hypothetical idea of AI matching human-level reasoning across every domain, and ASI is the further, more speculative idea of AI surpassing human intelligence entirely -- neither exists today. A world model is an AI's internal representation of how its environment works, and mechanistic interpretability is the research effort to reverse-engineer a network's internal circuits well enough to explain concretely how it produces a specific output.

Standard AI/ML terminology as used across current industry and research usage; organized as a reference rather than tied to a single source.

Explore next: Knowledge Hub · career resilience · upskill · switch jobs · free AI assessment · AI risk by role · how we score