AI Glossary —
Top 100 keywords.

The field of computer science focused on building systems that can perform tasks requiring human-like reasoning, perception, language, and decision-making.

Umbrella term covering ML, deep learning, and symbolic AI. Practical AI today is dominated by machine learning and neural networks.

A subfield of AI where models learn patterns from data rather than being explicitly programmed with rules.

Spans supervised, unsupervised, and reinforcement learning. Most modern AI products are ML systems.

Machine learning using neural networks with many layers, enabling models to learn hierarchical representations of raw data.

Powers computer vision, NLP, and generative AI. 'Deep' refers to the depth (number of layers) of the network.

A computational model loosely inspired by the brain, consisting of layers of interconnected nodes (neurons) that transform input data into output predictions.

The backbone of modern AI. Large language models are very deep neural networks with billions of parameters.

Training a model on labelled examples — pairs of input and the correct output — so it can predict outputs for new inputs.

Used for classification (spam/not-spam) and regression (price prediction). Requires expensive labelled data.

Training a model on unlabelled data to find structure, clusters, or representations without predefined correct answers.

Used for clustering, dimensionality reduction, and pre-training large models on raw text.

Training an agent to take actions in an environment to maximise a cumulative reward signal.

Used in game AI (AlphaGo), robotics, and RLHF for aligning language models to human preferences.

Running a trained model on new inputs to generate predictions or outputs — as opposed to training the model.

What happens when you send a message to ChatGPT. Inference is where AI cost is incurred at scale.

A learnable numerical value inside a model — weights and biases — that is adjusted during training to minimise error.

GPT-4 has ~1.8 trillion parameters. More parameters generally means more capacity but higher compute cost.

The basic unit of text that a language model processes — typically a word piece or subword (e.g. 'running' → ['run', '##ning']).

LLM costs and context limits are measured in tokens. ~1 token ≈ 0.75 English words on average.

A neural network trained on vast text corpora to understand and generate human language at scale.

GPT-4, Claude, and Gemini are LLMs. 'Large' refers to both parameter count and training data volume.

The neural network architecture underpinning most modern LLMs, built around self-attention mechanisms to process sequences in parallel.

Introduced in 'Attention Is All You Need' (2017). Replaced RNNs and enabled the current AI boom.

A technique that allows a model to weigh the relevance of different parts of the input when producing each output — 'paying attention' to what matters.

The core innovation of transformers. Enables models to handle long-range dependencies in text.

The maximum number of tokens a model can process in a single inference call — both input and output combined.

Claude 3.5 Sonnet has a 200K token context window (~150K words). Larger windows enable longer documents and richer agents.

A large model pre-trained on broad data at scale, intended to be adapted to many downstream tasks.

Term coined by Stanford in 2021. GPT-4, Claude, and Llama are foundation models. Fine-tuning adapts them to specific uses.

An AI system that can process and generate multiple data types — text, images, audio, video — within a single model.

GPT-4o and Claude 3 are multimodal — they can analyse images and generate text. Enables richer human-AI interaction.

The component of a model that converts raw input (text, images) into a dense internal representation (embedding).

BERT is an encoder-only model, good at understanding. Encoders produce embeddings used in semantic search.

The model component that generates output tokens sequentially, conditioned on an internal representation.

GPT models are decoder-only — they generate text left-to-right. Most chat LLMs are decoder-only transformers.

An architecture where only a subset of the model's components (experts) are activated for any given input, improving efficiency at scale.

GPT-4 is believed to be an MoE model. Enables very large models without proportionally higher inference cost.

Dense numerical vector representations of text, images, or other data that capture semantic meaning in a high-dimensional space.

Similar concepts have embeddings close together. The foundation of semantic search and RAG systems.

The initial large-scale training phase where a model learns general representations from a massive, broad dataset.

GPT models are pre-trained on trillions of tokens of text from the internet. Expensive; done once by labs.

Further training a pre-trained model on a smaller, task-specific dataset to specialise its behaviour.

Used to make a base LLM follow instructions, adopt a persona, or specialise in a domain like medicine or law.

A training method where human raters rank model outputs, and those preferences train a reward model used to fine-tune the LLM via RL.

Used by OpenAI for InstructGPT/ChatGPT and Anthropic for Claude. Key technique for making LLMs helpful and safe.

Anthropic's method of training Claude to follow a set of written principles (a 'constitution') — using AI feedback rather than solely human feedback.

Reduces reliance on human labellers for harmlessness training. Scales safety alignment more efficiently.

A parameter-efficient fine-tuning method that injects small trainable matrices into model layers, avoiding the cost of updating all parameters.

Makes fine-tuning LLMs practical on consumer hardware. Widely used for custom Llama and Stable Diffusion models.

A family of techniques (LoRA, prefix tuning, adapters) that fine-tune a small fraction of model parameters to reduce compute and storage costs.

Enables fine-tuning 7B-parameter models on a single GPU, democratising model customisation.

The optimisation algorithm that iteratively updates model parameters in the direction that reduces the loss (error) on training data.

The fundamental mechanism by which neural networks learn. SGD and Adam are practical variants used in LLM training.

The algorithm that computes how much each parameter contributed to the prediction error, enabling gradient descent to update them precisely.

The engine of neural network training. Computing gradients through billions of parameters requires enormous GPU clusters.

When a model learns the training data too well — including noise — and performs poorly on new, unseen data.

Regularisation, dropout, and more data mitigate overfitting. LLMs pre-trained on web-scale data rarely overfit.

Training a smaller 'student' model to mimic the outputs of a larger 'teacher' model, transferring knowledge more efficiently.

Used to produce efficient models for edge deployment. DeepSeek-R1-Distill and Phi models use this technique.

Reducing the precision of model weights (e.g. from 32-bit floats to 4-bit integers) to decrease memory and inference cost with minimal quality loss.

Enables running 70B-parameter models on consumer GPUs. 4-bit quantisation is standard for local LLM deployment.

AI systems that take sequences of actions autonomously — calling tools, browsing the web, writing code, and making decisions — to complete longer-horizon tasks.

The frontier of AI in 2025. Distinguished from chat by the ability to act, not just respond.

The ability of an LLM to invoke external functions or APIs — web search, code execution, database queries — as part of generating a response.

Enables LLMs to access real-time data and take real-world actions. Core capability of Claude, GPT-4, and Gemini.

An open protocol by Anthropic defining a standard interface between AI models and external tools/data sources.

Like USB-C for AI integrations. Lets any MCP-compatible tool plug into any MCP-compatible model. Growing ecosystem rapidly.

Architectures where multiple AI agents collaborate, each with specialised roles, to complete complex tasks beyond the capacity of a single agent.

Used for tasks requiring parallelism or specialist reasoning. Orchestrator + worker patterns are common.

Coordinating multiple LLM calls, tools, and agents in a workflow — routing tasks, managing state, and handling errors.

LangChain, LangGraph, and CrewAI are orchestration frameworks. Claude Code orchestrates agents natively.

The mechanisms by which an AI agent retains and retrieves information across steps or sessions: in-context (working), external (vector DB), and episodic memory.

Persistent memory enables agents to remember user preferences and prior work across conversations.

A hidden instruction block provided to an LLM before the user's message, shaping the model's persona, constraints, and capabilities.

Used by product teams to configure Claude's behaviour. Invisible to end users but sets the rules of the conversation.

The practice of crafting input text to guide an LLM toward the desired output — through instruction, examples, structure, and framing.

A critical skill for AI practitioners. Chain-of-thought, few-shot, and role prompting are common techniques.

A prompting technique that asks the model to reason step-by-step before giving a final answer, improving accuracy on complex tasks.

Adding 'Let's think step by step' measurably improves LLM reasoning. Basis of most modern reasoning models.

The ability of an LLM to learn a new task from examples provided in the prompt — without updating model weights.

Enables task adaptation at inference time. Few-shot and zero-shot prompting are forms of in-context learning.

A technique that retrieves relevant documents from a knowledge base at inference time and includes them in the LLM's context, grounding responses in current facts.

Solves hallucination and knowledge cutoffs. Standard pattern for enterprise AI systems using internal documents.

A database optimised for storing and searching high-dimensional embedding vectors by semantic similarity rather than exact match.

Pinecone, Weaviate, Chroma, and pgvector are popular choices. Core infrastructure for RAG systems.

Searching by meaning rather than keyword — finding results conceptually similar to a query even if they share no words.

Powered by embeddings and vector databases. Enables AI systems to find relevant context efficiently.

A metric measuring the angle between two vectors, used to quantify how semantically similar two pieces of text are.

Score of 1 = identical meaning, 0 = unrelated. The standard distance metric for embedding-based retrieval.

A structured representation of entities and their relationships, enabling AI systems to reason over connected facts.

Used in enterprise AI to represent domain knowledge (products, people, policies). Complements RAG with structured reasoning.

Connecting an LLM's outputs to verified, external sources of truth to reduce hallucination and improve factual accuracy.

RAG, tool use, and citations are grounding techniques. Critical for production systems where facts matter.

When an LLM generates factually incorrect, fabricated, or nonsensical information with apparent confidence.

The most cited risk of LLMs in production. Mitigated by RAG, structured outputs, and human review gates.

The process of splitting text into tokens (the atomic units a model processes) before embedding or inference.

Different tokenisers produce different token counts for the same text, affecting cost and context window usage.

A tokenisation algorithm that iteratively merges frequent character pairs to build a vocabulary of subword units.

Used by GPT models and many others. Balances vocabulary size, rare word handling, and sequence length.

A measure of how well a language model predicts a sample — lower perplexity means the model finds the text more probable.

Used to compare language models on held-out text. Not a direct measure of helpfulness or safety.

The challenge of ensuring AI systems reliably pursue goals that are beneficial, intended, and consistent with human values.

Central concern of AI safety research. RLHF, Constitutional AI, and interpretability are alignment techniques.

The field focused on preventing AI systems from causing harm — both near-term (misuse, bias) and long-term (loss of control, catastrophic risk).

Anthropic, DeepMind Safety, and OpenAI Safety teams work here. Growing regulatory focus post-EU AI Act.

Constraints applied to LLM inputs and outputs to prevent harmful, off-topic, or policy-violating content.

Implemented via system prompts, classifiers, or dedicated safety models. Balance safety with usefulness.

An attack where malicious text in user input or retrieved content hijacks an LLM's instructions, causing it to act against its system prompt.

Critical vulnerability in agentic systems that process external content. Defence requires input validation and output monitoring.

A prompt or technique designed to bypass an LLM's safety training and elicit responses it was trained to refuse.

Ongoing adversarial game between model developers and users. Red-teaming helps find and patch jailbreaks before release.

Systematically probing an AI system for vulnerabilities, harmful outputs, and failure modes — adversarially testing before deployment.

Standard practice at major AI labs. External red teams (including human experts) stress-test models before release.

System design that requires human review or approval before an AI takes consequential actions.

Critical safety mechanism for agentic systems. Trades speed for accountability and error correction.

Research into understanding what computations are happening inside neural networks — making AI decisions explainable.

Anthropic's interpretability team discovered 'features' inside Claude (e.g. the 'Golden Gate Bridge' feature). Active research frontier.

Systematic and unfair discrimination in AI outputs, typically inherited from biases in training data or label choices.

Particularly problematic in hiring, lending, and criminal justice AI. Requires diverse data and ongoing auditing.

The principle that AI systems should protect users' epistemic autonomy — presenting balanced views rather than nudging beliefs at scale.

One of Claude's core principles. Given LLMs interact with millions of users, homogenising views is a societal risk.

Graphics Processing Units (and Google's Tensor Processing Units) — massively parallel processors used to train and run AI models efficiently.

NVIDIA H100s are the primary AI training accelerator. A cluster of thousands powers LLM training runs.

The compute and financial cost of running a model on new inputs — the cost paid every time a user sends a message.

Measured per 1M tokens. Claude Haiku (~$0.25/Mtok) is 60× cheaper than Opus. Drives architecture and caching decisions.

The time between sending a request and receiving the first token of the response — a key UX metric for AI applications.

Streaming responses (showing tokens as they arrive) reduce perceived latency. Critical for real-time agent pipelines.

The number of tokens (or requests) a model serving system can process per second — a measure of serving efficiency.

Batching requests increases throughput but may increase latency. Key tradeoff in LLM serving architecture.

A parameter (0–2) controlling the randomness of an LLM's output: lower = more deterministic, higher = more creative and varied.

Temperature 0 is nearly deterministic (good for code). Temperature 1 is standard. Use lower temps in production for consistency.

A sampling strategy that restricts token selection to the smallest set whose cumulative probability exceeds p, balancing quality and diversity.

Often used alongside temperature. Top-p 0.95 is a common default. Works better than top-k for most tasks.

The infrastructure for deploying trained models to handle real-time inference requests at scale.

Includes load balancing, autoscaling, batching, and quantisation. vLLM and TGI are popular open-source serving frameworks.

Caching the KV (key-value) computation for a repeated prefix of a prompt so subsequent requests avoid reprocessing it.

Anthropic's prompt caching reduces cost by up to 90% and latency for long system prompts. Critical for production agents.

Delivering model outputs token-by-token in real time rather than waiting for the full response to complete.

Dramatically improves perceived response time. Server-Sent Events (SSE) is the standard transport for LLM streaming.

Processing multiple AI requests together in a single pass, improving throughput and reducing per-request cost.

Anthropic's Message Batches API offers 50% cost reduction for async, non-time-sensitive workloads.

Systematic tests measuring AI model quality, safety, and capability across a defined set of tasks or criteria.

The discipline of AI evals is maturing rapidly. Good evals are the foundation of responsible model development and deployment.

A standardised test suite used to compare AI models on specific capabilities — reasoning, coding, maths, factuality.

Models saturate benchmarks quickly. MMLU, HumanEval, and GPQA are widely cited. Real-world evals matter more.

Massive Multitask Language Understanding — a benchmark testing knowledge across 57 subjects from elementary to professional level.

One of the most cited LLM benchmarks. Top models now exceed 90% accuracy, approaching saturation.

A coding benchmark by OpenAI measuring how accurately LLMs generate Python functions that pass unit tests.

Standard measure of code generation quality. Claude 3.5 Sonnet scores >90%. Real coding evals are more nuanced.

Bilingual Evaluation Understudy — a metric comparing machine-generated text to reference translations by measuring n-gram overlap.

Originally for machine translation; now widely used for text generation quality. Imperfect: doesn't capture meaning well.

Providing a small number of examples in the prompt to demonstrate the desired task, allowing the model to learn in-context.

3–5 examples often dramatically improve accuracy. Used in evals to test models' ability to generalise from minimal guidance.

Asking a model to perform a task with no examples — relying entirely on its pre-trained knowledge and the instruction.

Modern LLMs are surprisingly capable zero-shot. Evals often compare zero-shot vs few-shot to measure generalisation.

An AI model's ability to perform multi-step logical, mathematical, or causal inference to reach correct conclusions.

Chain-of-thought prompting, o1, and Claude's extended thinking mode are designed to improve reasoning.

A mode where Claude generates an internal reasoning trace (thinking) before its final response, improving accuracy on hard problems.

Available in Claude 3.7 Sonnet and later. Traded off against latency and cost. Particularly strong on maths and coding.

A technique that injects information about the position of each token in the sequence into the model, since transformers lack inherent order awareness.

RoPE (Rotary Position Embedding) is the modern standard. Enables models to handle long sequences and generalise beyond training lengths.

An attention variant where each token only attends to a subset of other tokens, reducing the quadratic cost of full attention.

Enables very long context windows. Longformer and BigBird are examples. Critical for 100K+ token models.

A hardware-aware attention algorithm that reduces memory bandwidth usage, enabling faster and more memory-efficient attention computation.

Widely adopted in production LLMs. Enables larger batch sizes and longer sequences on the same GPU hardware.

A function that converts a vector of raw scores into a probability distribution — outputs sum to 1, with larger scores becoming higher probabilities.

Used in attention (to weight relevance) and output layers (to select the next token). Core building block of transformers.

Rectified Linear Unit — an activation function that outputs the input if positive, zero otherwise. Enables deep networks to learn non-linear functions.

Standard activation for hidden layers. SwiGLU and GELU are modern variants used in LLMs.

A regularisation technique that randomly deactivates neurons during training, forcing the network to learn redundant representations and reducing overfitting.

Less used in large-scale LLM training but common in smaller models and fine-tuning.

A generative model that learns to create data (images, audio) by learning to reverse a gradual noising process.

Stable Diffusion, DALL-E 3, and Midjourney use diffusion. Produces higher quality images than GANs at scale.

The subfield of AI focused on enabling computers to understand, process, and generate human language.

Pre-LLM NLP used rule-based and statistical methods. Modern NLP is dominated by transformer-based LLMs.

The AI field focused on enabling machines to interpret and understand visual data — images and video.

Powers autonomous vehicles, medical imaging, and facial recognition. Now merging with LLMs in multimodal systems.

AI models that produce new content — text, images, code, audio, video — rather than only classifying or predicting.

GPT, Claude, Stable Diffusion, and Sora are generative AI systems. Disrupting creative and knowledge-worker industries.

Application Programming Interface — in AI, typically the HTTP endpoint through which developers access LLM capabilities programmatically.

Anthropic, OpenAI, and Google expose their models via APIs. Enables building AI-powered products without operating model infrastructure.