AI Glossary

A ↑ top

Agentic AI: An umbrella term for AI systems that don't just answer a single question but pursue goals over multiple steps — planning, calling tools, checking their own work, and adjusting course. Where a classic chatbot waits for your next message, agentic AI decides what to do next on its own. The shift from single-turn chat to agentic workflows has been one of the defining AI trends since 2024.
AI Agent: A program built around a model — usually a large language model — that can perceive its environment, make decisions, and take actions toward a goal with limited human supervision: researching a topic, fixing a bug, booking a meeting. Agents typically combine a model with tool use, memory, and a loop that keeps them working until the task is done or a limit is reached.
AI Alignment: The research problem of making AI systems reliably pursue the goals their operators and society actually intend, rather than a literal-but-wrong reading of their instructions. Alignment work ranges from training models to refuse harmful requests today to long-term questions about steering systems smarter than their designers. Closely related to AI safety and RLHF.
AI Safety: The broader field concerned with preventing harm from AI systems. It spans technical work like alignment, evals, and red teaming, plus governance, policy, and deployment practices. "Safety" covers both near-term issues — bias, misuse, hallucinations — and longer-term concerns about highly capable systems.
Artificial General Intelligence (AGI): A hypothetical AI system able to match or exceed capable humans across most economically valuable cognitive work, rather than excelling at one narrow task. There is no agreed test for having reached AGI, which is why lab leaders can disagree in the same week about whether it is imminent or distant. Beyond it lies superintelligence.
Artificial Intelligence (AI): The umbrella discipline of building computer systems that perform tasks normally requiring human intelligence: understanding language, recognizing images, reasoning, planning, and learning from experience. Modern AI is dominated by machine learning, and in particular deep learning.
Attention Mechanism: The technique at the heart of the transformer architecture that lets a model weigh which parts of the input matter most to each other part — for example, which earlier words in a sentence a given word refers to. Self-attention is the reason modern models handle long-range context far better than earlier architectures did.

B ↑ top

Benchmark: A standardized test used to compare AI models: a fixed set of questions or tasks with a scoring method — math problem sets, coding challenges, graduate-level Q&A. Benchmark numbers headline every model release, but they can saturate (top models all near 100%) or leak into training data, so a single score deserves healthy skepticism. See also evals.
Bias: Systematic skew in a model's outputs, usually inherited from patterns in its training data — for instance, associating certain professions with certain demographics. Bias can also be introduced or amplified by fine-tuning choices. Measuring and mitigating it is a core part of model evaluation and of AI safety.

C ↑ top

Chain of Thought (CoT): A technique in which a model works through a problem step by step before giving its final answer — either because the prompt asks it to ("think step by step") or because the model was trained to reason that way. Reasoning models industrialize this idea, spending extra inference-time compute thinking before they answer.
Chatbot: A conversational interface to an AI system: you type (or speak), it answers, and the exchange continues across turns. Early chatbots followed scripts; modern ones are built on large language models and can hold open-ended conversations, write code, and analyze documents. The chatbot is the interface; the model behind it is what varies.
Compute: Shorthand for the processing power used to train and run AI models, measured in chip-hours or FLOPs (floating-point operations). Compute is one of the three ingredients of model capability alongside data and algorithms — see scaling laws — and access to it (chips, data centers, electricity) has become an industrial and geopolitical story of its own.
Context Engineering: The practice of assembling everything a model sees at inference time — instructions, retrieved documents, tool results, conversation history — so the model has exactly the information it needs inside its context window. Increasingly used as the successor term to prompt engineering now that agentic systems juggle far more than a single prompt.
Context Window: The maximum amount of text — measured in tokens — that a model can consider at once: its working memory. Everything you send plus everything the model writes back must fit inside it. Windows have grown from a few thousand tokens to millions, though long context reduces rather than eliminates the need for retrieval techniques like RAG.
Convolutional Neural Network (CNN): A neural network architecture designed for grid-like data such as images. Convolution filters scan for local patterns — edges, textures — and deeper layers compose them into higher-level features like faces or objects. CNNs powered the 2012–2020 computer-vision boom; transformers now dominate many vision tasks, but CNNs remain widely deployed.

D ↑ top

Data Labeling: The human (and increasingly automated) work of annotating training examples — tagging what's in an image, ranking chatbot answers, marking toxic content — so models can learn from them. It is a large, mostly invisible industry, and the quality of the labels caps the quality of the model. RLHF depends on it directly.
Deep Learning: The branch of machine learning built on neural networks with many layers ("deep"), which learn useful features directly from raw data instead of needing hand-engineered rules. Essentially every headline AI system of the last decade — image recognition, speech, LLMs — is deep learning.
Diffusion Model: A generative model that creates images — and increasingly video and audio — by learning to reverse a noising process: start from pure random noise and denoise it step by step toward an output that matches the prompt. Diffusion is the technique behind most AI image and video generators. Compare the transformer, which dominates text.
Distillation: Training a smaller "student" model to imitate a larger "teacher" model's outputs, preserving much of the capability at a fraction of the size and inference cost. It's how labs produce fast, cheap variants of their flagship models — and a controversy when done against a competitor's model without permission.

E ↑ top

Embedding: A list of numbers (a vector) that represents the meaning of a piece of text, an image, or other data, arranged so that similar things sit near each other in the vector space. Embeddings power semantic search, recommendations, and the retrieval step of RAG; they are stored and searched in a vector database.
Emergent Capabilities: Abilities that show up in larger models without being explicitly trained for — arithmetic, translation, code — seemingly appearing once scale crosses a threshold. Researchers debate whether emergence is real or an artifact of how capabilities are measured, but the term captures why each jump in scale manages to surprise even the people building the models.
Evals: Industry shorthand for systematic evaluations of a model's behavior: capability benchmarks, safety tests, and regression suites run before and after release. Frontier labs publish eval results with launches, and "we need better evals" is the standard refrain whenever models behave unexpectedly in the wild.
Explainability (XAI): Techniques for understanding why a model produced a given output — from visualizing attention to mechanistic interpretability, which tries to reverse-engineer the circuits inside a network. Explainability matters for debugging, regulation, and deciding whether to trust models in high-stakes settings.

F ↑ top

Few-Shot Learning: Getting a model to perform a task by showing it a handful of examples inside the prompt itself — no retraining involved. The ability to learn a task from examples on the fly is a signature capability of large language models. With no examples at all, it's zero-shot learning.
Fine-Tuning: Continuing a pre-trained model's training on a smaller, task-specific dataset to specialize it — a support bot trained on your tickets, a coding model trained on your codebase's conventions. Far cheaper than training from scratch, and cheaper still with parameter-efficient methods like LoRA.
Foundation Model: A large model pre-trained on broad data that serves as the base — the foundation — for many downstream uses via prompting or fine-tuning. GPT, Claude, Gemini, and Llama are all foundation models; most AI products are layers built on top of one.
Frontier Model: The most capable models at the current state of the art — the ones pushing the frontier of what AI can do. The term carries policy weight: "frontier AI" safety frameworks and commitments target models whose novel capabilities may pose novel risks, and several labs gate releases on frontier-capability evals.
Function Calling: A model API feature where the model can reply with a structured request to run a function the developer defined — get_weather(city) — instead of plain text. The application executes the function and hands the result back. Function calling (also called tool calling) is the plumbing that makes tool use and agents reliable.

G ↑ top

Generative AI: AI that produces new content — text, images, code, audio, video — rather than only classifying or ranking existing data. The post-2022 wave of chatbots and image generators that took AI mainstream is generative AI, built mostly on transformers and diffusion models.
GPU (Graphics Processing Unit): The chip class that made deep learning practical: thousands of cores doing parallel matrix math, originally for rendering games. Training a frontier model takes tens of thousands of them running for months, which is why GPU supply shapes the entire industry. NVIDIA dominates; Google's TPU is the best-known alternative.
Grounding: Connecting a model's answers to authoritative, verifiable sources — retrieved documents, live search results, databases — so its output reflects checkable facts rather than only what's compressed in its parameters. RAG is the standard grounding technique, and grounded answers can cite their sources. The opposite failure mode is hallucination.
Guardrails: The layer of checks wrapped around a deployed model that constrains what it will do: refusal policies, content filters, input and output validation, rate limits. Guardrails are product-level defenses; alignment works at the training level. Jailbreaks and prompt injection are the attacks they exist to stop.

H ↑ top

Hallucination: When a model states something false with complete confidence — an invented citation, a nonexistent API, a wrong date. It happens because LLMs generate statistically plausible text rather than looking facts up. Mitigations include grounding answers in retrieved sources, better post-training, and models that say "I don't know" — but no current model is immune.
Human-in-the-Loop: System design in which a human reviews, approves, or corrects an AI system's decisions at key points instead of letting it run fully autonomously. Standard practice for high-stakes uses like medical or legal work, and the pattern behind the approval steps in agent workflows. Compare data labeling, where humans improve the model rather than police its output.

I ↑ top

Inference: Running a trained model to get outputs — every chatbot reply, every generated image is inference, as distinct from training. Inference cost and speed now drive much of AI engineering (see quantization and distillation), and "inference-time compute" — extra thinking while answering — is the axis reasoning models scale on.
Instruction Tuning: Fine-tuning a raw pre-trained model on instruction-and-response examples so it learns to follow directions like an assistant instead of merely continuing text. It's the step that turns a base model into a chat model, typically followed by preference training such as RLHF.

J ↑ top

Jailbreak: A prompt crafted to bypass a model's safety training and make it produce output it would normally refuse — via role-play framing, encoded text, or gradual multi-turn manipulation. Jailbreaking is done by the user; prompt injection comes from third-party content. Defending against both is a core job of red teams and guardrails.

K ↑ top

Knowledge Cutoff: The date after which a model's training data ends. Ask about events past the cutoff and the model either admits ignorance or hallucinates. Products work around it by giving models live search or retrieval via tool use — which is how a chatbot can discuss yesterday's news despite a months-old cutoff.

L ↑ top

Large Language Model (LLM): A neural network — in practice always a transformer — trained on vast text corpora to predict the next token, and which as a byproduct learns to write, summarize, translate, reason, and code. LLMs are the core technology behind modern chatbots, agents, and most of the AI news cycle.
Latency: The time between sending a request and getting a response. For chat products, time-to-first-token matters most; for agents, whole-task latency does. Latency trades off against model size and reasoning effort — the reason providers ship both fast-and-cheap and slow-and-smart tiers of the same model family.
LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning method that freezes the base model and trains only small "adapter" matrices — often under 1% of the parameters — making customization drastically cheaper and the resulting adapters easy to swap in and out. Ubiquitous in the open-weights ecosystem for both language and image models.

M ↑ top

Machine Learning (ML): The approach of having computers learn patterns from data rather than following hand-written rules — show a system enough examples and it works out the rules itself. Its main paradigms are supervised, unsupervised, and reinforcement learning; deep learning is its dominant modern branch.
Mixture of Experts (MoE): An architecture in which a model contains many "expert" sub-networks but activates only a few of them for each token, so total parameter count can be huge while the compute paid per token stays modest. Many frontier and open-weights models use MoE to buy capability at lower inference cost.
Model Context Protocol (MCP): An open standard, introduced by Anthropic in late 2024, for connecting AI assistants to external tools and data sources through one common interface — so any MCP-compatible model can use any MCP server, instead of every integration being custom-built. Often described as "USB-C for AI tools," it was adopted across the industry through 2025 and underpins much of today's tool use.
Multi-Agent System: Multiple AI agents coordinating on one task — a planner delegating to workers, agents debating to check each other, or swarms of specialists handling research, coding, and review. Multi-agent setups can beat a single agent on complex work, at the price of orchestration complexity and errors that compound between agents.
Multimodal AI: Models that work across more than one kind of data — text, images, audio, video — as input, output, or both: describing a screenshot, holding a voice conversation, generating video from a script. Frontier models are now natively multimodal, trained on several modalities at once rather than having vision bolted onto a text model.

N ↑ top

Natural Language Processing (NLP): The field of getting computers to understand and generate human language. For decades it was a collection of separate specialized tasks — parsing, sentiment analysis, translation — each with its own models. The rise of LLMs largely unified it: one general model now handles most NLP tasks through plain prompting.
Neural Network: A mathematical function loosely inspired by the brain: layers of simple units pass signals forward, and the connection weights between them are adjusted during training until inputs map to correct outputs. Neural networks are the substrate of deep learning; an LLM is one with up to trillions of weights.

O ↑ top

Open-Weights Model: A model whose trained weights are published for download, so anyone can run it locally, inspect it, or fine-tune it — Llama, Mistral, and DeepSeek are prominent examples. Often loosely called "open source," though full open source would also include training data and code, and licenses vary from permissive to restrictive.
Overfitting: When a model memorizes its training data instead of learning patterns that generalize — acing training examples, failing in the real world. The classic machine-learning failure mode. Its modern echo in LLMs is benchmark contamination: a model that has seen the test in training looks smarter than it is.

P ↑ top

Parameters: The learned numeric weights inside a neural network — the model's knowledge, compressed into numbers. Counts run from millions to trillions, and more parameters generally means more capability and cost (see scaling laws), though data quality, MoE architectures, and distillation complicate the simple story.
Pre-training: The first and most expensive phase of building a foundation model: training on trillions of tokens of broad data to predict the next token, which instills general knowledge of language, code, and the world. Post-training — instruction tuning, RLHF, reasoning training — then shapes that raw capability into a usable assistant.
Prompt: Whatever you send the model: a question, instructions, documents to analyze, examples to imitate, images to inspect. The model's output quality tracks prompt quality closely — the observation behind prompt engineering and its successor, context engineering. See also system prompt.
Prompt Engineering: The craft of writing prompts that reliably get the output you want: clear instructions, good examples (few-shot), role framing, explicit output formats, and structured reasoning. As models improved, the elaborate tricks mattered less and the discipline evolved toward context engineering for agentic systems.
Prompt Injection: An attack in which malicious instructions hidden inside content the model processes — a web page, an email, a document — hijack it into ignoring its real instructions: the LLM-era cousin of SQL injection. It is a top security concern for agents that browse and act on the world, and it differs from a jailbreak, which the user performs directly.

Q ↑ top

Quantization: Compressing a model by storing its weights at lower numeric precision — 16-bit floats down to 8- or 4-bit integers — which shrinks memory use and speeds up inference with only modest quality loss. Quantization is what makes large open-weights models runnable on laptops and phones.

R ↑ top

Reasoning Model: A model trained — typically with reinforcement learning — to produce a long internal chain of thought before answering, trading extra inference-time compute for much stronger math, code, and logic. Brought mainstream by OpenAI's o1 in 2024 and now standard across frontier labs, usually surfaced as a visible or hidden "thinking" phase.
Red Teaming: Deliberate adversarial testing of a model: jailbreaks, injection attacks, misuse scenarios, and probes for dangerous capabilities, run before and after release. Frontier labs use internal and external red teams, and findings feed guardrails, safety training, and sometimes release decisions.
Reinforcement Learning (RL): Training by trial and reward: an agent acts in an environment, receives a reward signal, and updates itself to earn more reward. RL produced the game-playing milestones (AlphaGo) and is now central to post-training language models — both RLHF and the RL on verifiable problems that builds reasoning models.
RLHF (Reinforcement Learning from Human Feedback): The post-training recipe in which humans rank model outputs, a reward model learns those preferences, and reinforcement learning optimizes the model against that reward. RLHF is the technique that made chatbots helpful and polite rather than merely fluent; successors include RLAIF, where AI provides the feedback.
Retrieval-Augmented Generation (RAG): An architecture in which relevant documents are fetched — usually via embeddings and a vector database — and inserted into the prompt, so the model answers from your data and can cite sources. RAG is the default enterprise pattern for grounding, freshness beyond the knowledge cutoff, and reducing hallucination.

S ↑ top

Scaling Laws: Empirical findings that model performance improves smoothly and predictably as parameters, data, and compute scale up together — the thesis that justified ever-larger frontier models. How far pre-training scaling can go is actively debated; reasoning models opened a second scaling axis at inference time.
Small Language Model (SLM): A compact language model — typically a few hundred million to a few billion parameters — good enough for many tasks while running cheaply, fast, on-device, or in privacy-sensitive settings. Built through careful data curation and distillation from larger teachers; phones and laptops increasingly ship with one.
Superintelligence (ASI): Hypothetical AI decisively beyond the best humans at essentially every cognitive task — including AI research itself, which is what could make its arrival self-accelerating. Whether and when it is achievable is contested; it anchors the strongest safety arguments and, lately, the mission statements of several labs. One step beyond AGI.
Supervised Learning: Learning from labeled examples: each input comes with the correct answer — this email is spam, this photo is a cat — and the model learns the mapping. Most deployed machine learning historically worked this way, which is why data labeling became an industry. Compare unsupervised and reinforcement learning.
Synthetic Data: Training data generated by models rather than collected from the world: model-written math problems and solutions, simulated dialogues, auto-labeled images. Labs lean on it as high-quality human data grows scarce and to target specific skills — with careful quality control, because training models heavily on model output can degrade them (the "model collapse" concern).
System Prompt: The standing instructions a developer gives a model before any user input: persona, rules, available tools, output format, guardrail policies. Users don't normally see it, and much of jailbreaking and prompt injection amounts to trying to override it.

T ↑ top

Temperature: A sampling setting that controls how random a model's output is. Low temperature makes it pick the most likely next token — focused, repeatable, right for code and facts. High temperature spreads the probability — varied and creative, better for brainstorming. One of the few dials exposed by most model APIs.
Token: The unit a model actually reads and writes: a word piece averaging roughly four characters of English ("glossary" might be one token; "ZealVector" several). Tokenization is the process of splitting text into them. Pricing, context windows, and generation speed are all measured in tokens, which is why the word is everywhere in AI product news.
Tool Use: A model's ability to call external tools — web search, code execution, databases, browsers, other APIs — usually via function calling, to overcome its knowledge cutoff and act on the world rather than just describe it. Tool use is the defining capability of agents, and standards like MCP exist to make it interoperable.
TPU (Tensor Processing Unit): Google's custom AI accelerator chip family — an alternative to the GPU — used internally to train and serve Gemini and rentable through Google Cloud. TPUs are the most established example of the industry-wide push toward custom AI silicon as compute demand explodes.
Training Data: The corpus a model learns from: web text, code, books, images, licensed datasets, and increasingly synthetic data. Its composition determines what a model can do and what biases it carries; its provenance is the subject of copyright lawsuits and a wave of licensing deals between AI labs and publishers.
Transfer Learning: Reusing what a model learned on one task or domain as the starting point for another, instead of training from zero. It is the principle that makes pre-train-then-fine-tune work at all — general knowledge transfers, so a small specialized dataset goes a long way.
Transformer: The neural architecture introduced in Google's 2017 paper "Attention Is All You Need," built on self-attention and highly parallelizable — which is what made training on internet-scale data feasible. It is the T in GPT and the basis of virtually every frontier model, across text, vision, and audio.
Turing Test: Alan Turing's 1950 proposal: if a machine's side of a conversation is indistinguishable from a human's, treat it as intelligent. Modern chatbots pass casual versions, which mostly revealed the test's limits — fluent conversation turned out not to equal general intelligence — so the field measures progress with benchmarks and evals instead.

U ↑ top

Unsupervised Learning: Learning structure from unlabeled data — clustering similar customers, spotting anomalies — with no human-provided answers. The self-supervised pre-training of LLMs is its modern descendant: next-token prediction manufactures its own labels from raw text, which is what lets models learn from the whole internet without anyone annotating it.

V ↑ top

Vector Database: A database built to store embeddings and find nearest neighbors fast at scale — the retrieval layer behind RAG and semantic search. Examples include Pinecone, Weaviate, Milvus, and pgvector. (And no — ZealVector, despite the name, is an AI news digest, not a vector database.)
Vibe Coding: Building software by describing what you want to an AI coding agent and iterating on the results, without reading most of the generated code. Coined by Andrej Karpathy in early 2025, it went from half-joke to genuine workflow within months — and raised real questions about review, security, and maintainability of code nobody has read.

W ↑ top

World Model: A system's internal model of how an environment works, used to predict what happens next and plan actions — "if I drop this, it falls." Whether LLMs implicitly learn one from text alone is a live debate; building explicit world models is a major research bet for robotics, video generation, and game-like simulation.

Z ↑ top

Zero-Shot Learning: A model performing a task it was given no examples of — instructions only — relying purely on generalization from pre-training. "Translate this to German" with no sample translations is zero-shot. It is the everyday mode of using chat models; add examples to the prompt and it becomes few-shot learning.

A ↑ top

B ↑ top

C ↑ top

D ↑ top

E ↑ top

F ↑ top

G ↑ top

H ↑ top

I ↑ top

J ↑ top

K ↑ top

L ↑ top

M ↑ top

N ↑ top

O ↑ top

P ↑ top

Q ↑ top

R ↑ top

S ↑ top

T ↑ top

U ↑ top

V ↑ top

W ↑ top

Z ↑ top

Need AI in your product?