AI glossary

Plain English

This glossary is a quick map—not a substitute for the full pages. Each bullet gives a short definition and points you to the encyclopedia entry for depth. Slugs appear as /learn/slug paths you can open in the Learn library.

Terms are grouped by theme matching the Wave 1 catalog: foundations, machine learning, deep learning, language models, prompting, agents, generative media, classical AI, and evaluation/society.

How to use this page

Skim the group that matches your question. Follow the /learn/... reference for diagrams, code, misconceptions, and sources. If two terms overlap, read both entries—AI vocabulary often stacks (tokens → embeddings → transformers → language models).

• L0 entries are conceptual; L1/L2 add mechanism and math.
• Prerequisites on each catalog page show a sensible reading order.
• Related pages at the bottom of each article cross-link further.

Foundations

Core vocabulary for what AI is, how the field evolved, and honest limits.

• Artificial intelligence (AI) — Software that performs tasks people associate with human intelligence; see /learn/what-is-ai.
• Machine learning (ML) — Systems that improve from data rather than only hand-coded rules; nested under AI in /learn/ai-vs-machine-learning-vs-deep-learning.
• Deep learning — ML using multi-layer neural networks; see /learn/ai-vs-machine-learning-vs-deep-learning and /learn/neural-networks.
• Narrow AI — Today's deployed systems tuned for specific tasks; contrast with AGI in /learn/narrow-vs-general-ai.
• AGI / ASI — Research goals for broadly capable or superhuman AI; definitions and hype checks in /learn/narrow-vs-general-ai.
• Training data — Examples a model learns from; quality and labels in /learn/data-and-labels.
• Labels — Correct answers paired with inputs in supervised setups; see /learn/data-and-labels and /learn/supervised-learning.
• Generalization — Performing well on new data, not only memorized training items; /learn/overfitting-and-generalization.
• Capabilities and limits — What current systems reliably do versus common failure modes; /learn/capabilities-and-limits.
• History of AI — Symbolic era, ML rise, deep learning, and modern LLMs; /learn/history-of-ai.

Machine learning

Learning paradigms and classical models that still matter on tabular and structured data.

• Supervised learning — Learn input→output mappings from labeled examples; /learn/supervised-learning.
• Unsupervised learning — Find structure without labels (clusters, components); /learn/unsupervised-and-self-supervised.
• Self-supervised learning — Create supervision from the data itself (predict masked tokens); /learn/unsupervised-and-self-supervised.
• Reinforcement learning (RL) — Learn via actions and rewards; /learn/reinforcement-learning.
• Features — Numeric representations of raw inputs; /learn/features-loss-and-optimization.
• Loss function — Score of how wrong predictions are; minimized during training; /learn/features-loss-and-optimization.
• Optimization — Adjusting parameters to reduce loss (e.g., gradient descent); /learn/features-loss-and-optimization.
• Overfitting — Memorizing training noise and failing on new data; /learn/overfitting-and-generalization.
• Regularization — Penalties or constraints that discourage overfitting; /learn/overfitting-and-generalization.
• Linear models, trees, ensembles — Classical ML algorithms; /learn/classical-ml-models.
• Train / validation / test split — Data partitions for learning, tuning, and honest scoring; /learn/benchmarks-and-evaluation.

Deep learning and compute

Neural networks, vision, sequences, and the hardware scaling story.

• Neural network — Layers of connected units learning representations; /learn/neural-networks.
• Weights and biases — Learnable parameters adjusted during training; /learn/neural-networks.
• Activation function — Nonlinearity between layers; /learn/neural-networks.
• Backpropagation — Algorithm to compute gradients through layers; /learn/backpropagation.
• Gradient descent — Update weights in the direction that lowers loss; /learn/backpropagation.
• CNN (convolutional neural network) — Architecture for images and spatial patterns; /learn/cnns-and-vision.
• Computer vision — Machine perception of images and video; /learn/cnns-and-vision.
• RNN (recurrent neural network) — Sequence model with hidden state over time; /learn/rnns-and-sequence-models.
• GPU / TPU — Accelerators for parallel math in training; /learn/gpus-tpus-and-training-compute.
• Compute — Compute budget (FLOPs, chip-hours) for training or inference; /learn/gpus-tpus-and-training-compute.
• Scaling laws — Empirical trends linking size, data, and compute to quality; /learn/scaling-laws-and-compute.
• Distributed training — Splitting work across many devices; /learn/distributed-training-basics.
• Batch size — Number of examples processed together per step; /learn/distributed-training-basics.

Language models

How modern LLMs represent text, attend across context, and are trained.

• Language model (LM) — Model trained to predict text (often next token); /learn/language-models.
• LLM — Large language model with billions of parameters; /learn/language-models.
• Token — Subword piece the model reads and writes; /learn/tokenization.
• Tokenization — Splitting raw text into tokens; /learn/tokenization.
• Vocabulary — Set of tokens a model knows; /learn/tokenization.
• Embedding — Vector representing a token or document; /learn/embeddings.
• Vector similarity — Comparing meaning via distance in embedding space; /learn/embeddings.
• Transformer — Architecture using attention for sequences; /learn/transformers.
• Attention — Weighted mixing of tokens based on relevance; /learn/self-attention.
• Self-attention — Each token attends to others in the same sequence; /learn/self-attention.
• Query, key, value — Attention components; /learn/self-attention.
• Context window — Maximum tokens visible in one forward pass; /learn/context-windows-and-memory.
• Pretraining — Broad training on large corpora; /learn/pretraining-and-finetuning.
• Fine-tuning — Narrower training for tasks, style, or safety; /learn/pretraining-and-finetuning.
• Mixture of experts (MoE) — Route tokens to specialist subnetworks; /learn/mixture-of-experts-and-modern-variants.
• Parameter — A single learnable weight in the network; /learn/scaling-laws-and-compute.

Prompting and use

Steering models at inference time and grounding outputs in facts.

• Prompt — User instruction plus context sent to the model; /learn/prompting.
• Zero-shot — Task without in-prompt examples; /learn/prompting-patterns.
• Few-shot — Task with a handful of examples in the prompt; /learn/prompting-patterns.
• Chain-of-thought — Asking the model to show intermediate reasoning steps; /learn/prompting-patterns.
• System message — Developer-set behavior instructions; /learn/prompting.
• RAG (retrieval-augmented generation) — Fetch documents into the prompt before answering; /learn/rag-retrieval-augmented-generation.
• Retrieval — Search over a corpus using embeddings or keywords; /learn/rag-retrieval-augmented-generation.
• Hallucination — Confident but incorrect fabricated output; /learn/hallucinations.
• Grounding — Tying answers to cited sources or tools; /learn/rag-retrieval-augmented-generation.
• Multimodal — Models handling text, images, audio, or more; /learn/multimodality.

Agents and systems

Loops that observe, plan, call tools, and remember across steps.

• AI agent — System that iterates toward a goal with tools and feedback; /learn/agents.
• Tool use / function calling — Model invokes APIs with structured arguments; /learn/tool-use-and-function-calling.
• Planning — Decomposing goals into ordered steps; /learn/planning-and-multi-step-reasoning.
• Multi-step reasoning — Working through subproblems before a final answer; /learn/planning-and-multi-step-reasoning.
• Agent memory — Short chat context vs long-term stores; /learn/memory-for-agents.
• Agent evaluation — Testing success, safety, and cost beyond one-shot QA; /learn/evaluation-of-agents.
• Observability — Logging traces of agent decisions for debugging; /learn/evaluation-of-agents.

Generative media

Images, video, audio, and diffusion-style generation.

• Generative model — Learns to produce new samples (text, pixels, audio); /learn/diffusion-models.
• Diffusion model — Iterative denoise-from-noise generation; /learn/diffusion-models.
• Denoising — Removing noise stepwise to reveal structure; /learn/diffusion-models.
• Text-to-image — Prompt-conditioned image synthesis; /learn/image-and-video-generation.
• Text-to-video — Prompt-conditioned clip generation; /learn/image-and-video-generation.
• Inpainting / outpainting — Edit regions or extend borders of an image; /learn/image-and-video-generation.
• ASR (automatic speech recognition) — Speech-to-text; /learn/audio-and-speech-models.
• TTS (text-to-speech) — Text-to-spoken audio; /learn/audio-and-speech-models.
• Voice cloning — Synthesizing a specific speaker's voice; /learn/audio-and-speech-models.
• Synthetic media — AI-generated visual or audio content; /learn/image-and-video-generation.

Classical AI

Search, logic, and probability—ideas predating deep learning that still apply.

• Search — Exploring state spaces to reach a goal; /learn/search-and-planning.
• BFS / DFS — Breadth-first and depth-first graph search; /learn/search-and-planning.
• A* search — Heuristic best-first search; /learn/search-and-planning.
• Heuristic — Estimate of cost-to-go for informed search; /learn/search-and-planning.
• CSP (constraint satisfaction problem) — Assign values under constraints; /learn/search-and-planning.
• Knowledge base — Stored facts and rules; /learn/knowledge-representation-and-logic.
• Inference — Deriving conclusions from a knowledge base; /learn/knowledge-representation-and-logic.
• First-order logic — Logic with quantifiers and relations; /learn/knowledge-representation-and-logic.
• Bayes' rule — Update beliefs with new evidence; /learn/uncertainty-and-bayesian-thinking.
• Prior / posterior — Belief before / after seeing data; /learn/uncertainty-and-bayesian-thinking.
• Bayesian network — Compact probabilistic graphical model; /learn/uncertainty-and-bayesian-thinking.
• MDP (Markov decision process) — Sequential decisions under uncertainty; /learn/search-and-planning.

Evaluation, safety, and society

Measuring systems, aligning behavior, fairness, and real-world impact.

• Benchmark — Standardized test suite for models; /learn/benchmarks-and-evaluation.
• Metric — Numeric score (accuracy, F1, pass rate); /learn/benchmarks-and-evaluation.
• Leaderboard — Public ranking on a benchmark; /learn/benchmarks-and-evaluation.
• Contamination — Test data appearing in training; /learn/benchmarks-and-evaluation.
• Red teaming — Adversarial testing for harm and jailbreaks; /learn/alignment-and-safety.
• Alignment — Matching system behavior to human intent; /learn/alignment-and-safety.
• RLHF — Reinforcement learning from human feedback; /learn/alignment-and-safety.
• Guardrails — Filters and policies limiting harmful outputs; /learn/alignment-and-safety.
• Bias — Systematic skew harming groups; /learn/bias-fairness-and-accountability.
• Fairness — Equitable outcomes across people and contexts; /learn/bias-fairness-and-accountability.
• Accountability — Who answers when systems cause harm; /learn/bias-fairness-and-accountability.
• Provenance — Tracking origin of content and models; /learn/ai-and-society.
• Labor and automation — Work impacts of deployed AI; /learn/ai-and-society.

Going deeper

Use reading tracks on the Learn home page—Curious, Builder, Pro—for ordered paths. Foundations start at /learn/what-is-ai; builders often jump from /learn/supervised-learning to /learn/transformers to /learn/agents; pro tracks add /learn/backpropagation, /learn/benchmarks-and-evaluation, and /learn/search-and-planning.

When terms collide (e.g., planning in classical search vs LLM chain-of-thought), read both pages and note whether the system guarantees exploration or generates plausible steps. That distinction matters for reliability in production.

Common misconceptions

This glossary replaces the full encyclopedia pages.

It orients you quickly; mechanisms, code, and sources live on each dedicated /learn/... page.

Every vendor uses these terms the same way.

Marketing language diverges; check each product's docs against the concepts here.

Key facts

• Wave 1 covers foundations through agents, generative media, classical AI, and society.
• Each term points to a slug such as /learn/transformers for full depth.
• Depth badges L0/L1/L2 signal how technical an entry is.
• Confused pairs (AI/ML/DL, pretrain/fine-tune, RAG/memory) have separate pages.
• The catalog in data/learn-catalog.ts is the authoritative topic list.