Reference
AI Explainers
Your guide to every term in AI — plain language, no hype. 30 concepts and counting.
Browse by Category
Agents
- AI Agent An AI Agent is a specialized system that uses a large language model as its core 'brain' to autonomously plan and execute multi-step tasks. Unlike a standard chatbot that only responds to text, an AI agent can interact with external tools—like web browsers, code editors, and software APIs—to achieve a specific goal on behalf of a user.
- Multi-Agent System A Multi-Agent System (MAS) is a framework where multiple independent AI agents work together to solve a complex task. By assigning specialized roles—such as a 'Project Manager,' 'Coder,' and 'Reviewer'—to different agents or models, the system can achieve higher quality and more' robust' results than any single AI could on its own.
- Tool Use Tool Use (also known as Function Calling) is the specialized capability that allows an AI model to interact with external software and APIs. Instead of just answering a question with text, the model can generate a specific command to trigger a calculator, search the web, run a piece of code, or update a database in real-time.
Applications
- Grounding In the world of artificial intelligence, 'Grounding' is the process of anchoring an AI model's response to a specific, verifiable source of truth—such as an external document, reaching the live web, or a trusted database. By 'grounding' its answers, the AI can drastically reduce hallucinations and provide user-verifiable citations for its information.
- Prompt Engineering Prompt Engineering is the strategic process of crafting precise, structured, and context-rich instructions to guide a large language model toward a specific, high-quality output. It involves selecting the right words, formatting, and iterative refinement to 'unlock' the most effective and accurate responses from an AI system.
- Retrieval-Augmented Generation (RAG) Retrieval-Augmented Generation (RAG) is a technique that connects an AI model to an external, trusted source of information. Instead of relying only on its internal memory, the system dynamically looks up facts, documents, or data in real-time to provide more accurate and up-to-date responses.
Architecture
- Attention Mechanism Attention is a technique that lets a neural network dynamically weight the relevance of different parts of the input when producing each part of the output.
- Mixture of Experts (MoE) Mixture of Experts (MoE) is a neural network architecture where a model is divided into specialized sub-networks, or 'experts.' For any given query, only a small fraction of these experts are activated, allowing the model to have a massive amount of 'knowledge' without the enormous computational cost of running every parameter for every word.
- Transformer Architecture The Transformer Architecture is a sophisticated neural network design characterized by its 'self-attention' mechanism, which allows AI models to process entire sequences of data simultaneously. This breakthrough replaced sequential processing with parallel computation, forming the foundational blueprint for nearly all modern generative AI systems, including ChatGPT and Claude.
Data
- Knowledge Graph A Knowledge Graph is a structured representation of facts that maps out the relationships between 'entities'—such as people, places, concepts, and things. By organizing data as a network of interconnected points, AI systems can reason across complex links and provide more accurate, context-aware, and factually verifiable responses.
- Vector Database A Vector Database is a specialized storage system designed to manage and search through high-dimensional vectors, or 'embeddings'. Unlike traditional databases that search for exact keywords, vector databases search for mathematical similarities in meaning, allowing AI systems to find relevant information by context and intent.
Foundation Models
- Diffusion Model A Diffusion Model is a type of generative AI that creates high-quality images, video, or audio by gradually 'reverse-engineering' noise into a structured representation. By learning to remove random pixels from a blurry image step-by-step, the model can synthesize realistic visuals from a simple text prompt.
- Embeddings Embeddings are a numerical way to represent the 'meaning' of a piece of data—like a word, sentence, or image—as a long list of numbers called a vector. By turning language into math, AI models can calculate how similar two concepts are, allowing them to 'reason' that 'king' is related to 'queen' in the same way 'man' is related to 'woman'.
- Large Language Model (LLM) A Large Language Model (LLM) is a sophisticated neural network trained on massive datasets to understand, generate, and manipulate human language with human-like proficiency. These models form the core of modern conversational AI, enabling systems to write code, summarize documents, and engage in complex reasoning across diverse topics.
- Multimodal Model A Multimodal Model is an AI system that can process and generate information across multiple formats, such as text, images, audio, and video, within a single architecture. Unlike standard 'text-only' models, multimodal systems can 'see' a photo, 'hear' a voice, and respond with a written explanation, allowing for more natural and versatile interactions.
- Reasoning Model A Reasoning Model is a type of AI designed to slow down and think through complex problems step-by-step before providing an answer. Unlike standard large language models that generate text intuitively, reasoning models use an internal 'chain-of-thought' to check their own logic, identify potential errors, and solve multi-stage tasks in math, coding, and science.
- Tokenization Tokenization is the first step of AI processing, where raw text is broken down into smaller, machine-readable units called 'tokens'. These tokens can be words, characters, or even sub-word fragments like 'ing' or 'ment', allowing the model to represent language as a sequence of numbers.
Inference
- Chain-of-Thought (CoT) Chain-of-Thought (CoT) is a prompting technique and reasoning strategy where an AI model is encouraged to break a complex problem into a sequence of logical intermediate steps before providing a final answer. By 'thinking out loud' or showing its 'work,' the model can significantly improve its accuracy in math, logic, and multi-stage reasoning tasks.
- Context Window The Context Window is the maximum amount of information—measured in tokens—that an AI model can 'remember' and process at any one time. It acts as the model's short-term working memory, determining how much of a conversation, document, or dataset the AI can analyze simultaneously before it begins to 'forget' earlier parts of the input.
- Inference Inference is the phase where a trained AI model is put to work to generate a response, solve a problem, or analyze data. While 'training' is the model's education, 'inference' is the actual live execution of that knowledge in response to a user's prompt or a new piece of information.
- Model Quantization Model Quantization is a technique used to compress an AI model by reducing the numerical precision of its internal 'weights'. By converting high-precision numbers into smaller, less precise formats, quantization significantly cuts down the model's memory footprint and speeds up inference without a proportional loss in accuracy.
Safety & Alignment
- Alignment AI Alignment is the technical and philosophical challenge of ensuring that an artificial intelligence system’s goals, behaviors, and outputs are in sync with human values and intentions. The 'alignment problem' asks: how do we build an increasingly powerful AI that is helpful, harmless, and follows our commands without causing unintended consequences?
- Guardrails AI Guardrails are the safety layers and constraints built into or around an AI model to prevent it from generating harmful, biased, or 'off-policy' content. They act as a digital 'fence' that steers the AI away from restricted topics—like hate speech, legal advice, or chemical weapon instructions—and ensures the model remains helpful and aligned with its intended purpose.
- Hallucination An AI hallucination occurs when a large language model generates a response that is grammatically correct and fluent, but factually incorrect or nonsensical. These errors aren't intentional 'lies'—they are the result of the model's probabilistic nature prioritizing the most likely next word over verifiable reality.
- Jailbreak An AI Jailbreak is a specialized prompting technique designed to bypass an AI model's built-in safety filters and guardrails. By using creative roleplay, hypothetical scenarios, or 'adversarial' language, a user can trick the AI into generating restricted content—like toxic speech, phishing instructions, or malware code—that it would otherwise refuse.
Training
- Benchmark An AI Benchmark is a standardized test or dataset used to evaluate and compare the performance of different AI models. By testing a system on its ability to solve math problems, write code, or reason through common-sense scenarios, benchmarks provide a way for the industry to rank models like GPT-4, Claude, and Llama.
- Fine-Tuning Fine-Tuning is the process of taking a pre-trained AI model and training it further on a smaller, specialized dataset. This 'bonus' training adapts the model's general knowledge to excel at specific tasks, follow a particular style, or understand the nuances of a specialized field like law or medicine.
- Model Distillation Model Distillation (also known as Knowledge Distillation) is a training technique where a small, efficient AI model (the 'Student') is taught to mimic the behavior and outputs of a much larger, more powerful model (the 'Teacher'). This process 'compresses' a massive model's knowledge into a smaller format that is faster and cheaper to run.
- Pre-Training Pre-Training is the first and most intensive phase of an AI model's education, where a neural network is exposed to a massive dataset like the entire public internet. During this stage, the model learns the statistical patterns of language, general facts about the world, and basic reasoning skills by predicting the next word in a sequence trillions of times.
- Reinforcement Learning from Human Feedback (RLHF) Reinforcement Learning from Human Feedback (RLHF) is a training method that uses human evaluations to align an AI model's behavior with human values. By ranking different AI responses, 'human trainers' help the model learn to be helpful, polite, and safe, transforming a raw text predictor into a conversational assistant.