AI Glossary

Imagine teaching a dog tricks. Instead of showing the dog *all* the tricks, you start by showing the tricks the dog struggles with the most. Active le...

Think of a massive language model as a fully built Lego castle. An adapter is a small plug‑in block you snap onto the existing walls to give the castl...

Like a targeted trick designed to intentionally fool an AI model. These are crafted inputs that look normal to humans but cause the model to make a wr...

An AI system that can take actions autonomously — reading files, calling APIs, running code, making decisions — not just generating text. Agents do wo...

A process where AI operates with autonomy: planning steps, using tools, evaluating results, and iterating without human intervention at each step. The...

A structured way for software to talk to other software. When you send a prompt to Claude through code instead of a chat window, you're using an API.

The mechanism that lets AI focus on relevant parts of the input when generating output. It's why AI can connect "the cat sat on the mat" to a question...

Backpropagation is how neural networks learn. Imagine adjusting the knobs on a mixing board, listening to the music, and tweaking the knobs to make it...

A prompting technique where you ask AI to reason step by step instead of jumping to the answer. It improves accuracy on complex tasks by forcing the m...

Asking the model to show its work step by step, like solving a math problem on paper instead of guessing the final answer. It improves accuracy by bre...

Splitting documents into smaller pieces for storage in a vector database. Chunk size is a critical design choice: too big and you retrieve irrelevant ...

The text that an AI model generates in response to a prompt. You send a prompt; you get back a completion.

The practice of carefully curating what information you feed to AI — not just how you ask, but what data you include. More strategic than prompt engin...

The maximum amount of text an AI can "see" at once — both your input and its output combined. Think of it as the model's working memory. Bigger window...

A way to measure how different the AI's predictions are from the actual truth. It's like judging how well a weather forecast matches the real weather—...

Like teaching a child: start with simple concepts before moving to complex ones. Instead of feeding an AI model all the data at once, you gradually in...

Data augmentation is like photocopying a picture, then cutting it up and reassembling it to create new versions. This helps train AI models to be more...

Data leakage is like having a cheat sheet for a test - it's when your model gets access to information it shouldn't have, like labels or test data, du...

Think of data preprocessing like preparing ingredients for a recipe. You need to clean, chop, and mix them in the right way so the model can use them ...

A data silo is like a separate kitchen in a large restaurant - it's a isolated collection of data that's not easily accessible or usable by other part...

Data sparsity refers to the issue of having too few data points to effectively train or fine-tune a model. Think of it like trying to learn a new lang...

Dataset shift is like moving a restaurant to a new location - the menu stays the same, but the customers and their preferences change. In AI, it refer...

An AI that generates images by starting with random noise and gradually refining it into a picture. Used by DALL-E, Midjourney, and Stable Diffusion.

A way to convert text (or images) into a list of numbers that captures meaning. Similar concepts get similar numbers, which is how AI can understand t...

New capabilities that appear in a model as it scales, often unpredictably, and are not present in smaller models of the same family. These abilities a...

Ensembling is like asking multiple doctors for a diagnosis before making a decision. Instead of relying on a single model, you combine the predictions...

Explainability is about understanding *why* an AI system arrived at a particular output or decision. It's like asking a black-box model to show its wo...

Federated Learning allows AI models to be trained across multiple decentralized devices or servers holding local data, without the data ever leaving i...

A prompting technique where you provide 2-5 examples of the desired input/output format before asking AI to process new input. Dramatically improves c...

Training an existing AI model on your own specific data to make it better at a particular task. Like teaching a general-purpose chef to specialize in ...

A very large, pre-trained model designed to be adaptable to a wide range of tasks and applications. Think of it as a powerful general-purpose engine; ...

Imagine you're pushing a car up a hill. Gradient accumulation is like pushing the car a little bit, noting the effort, then repeating this process sev...

Gradient clipping is a technique to prevent exploding gradients during neural network training. Think of it as putting a speed limit on how much the m...

Gradient descent is like hiking down a mountain in thick fog. You can't see the bottom (the lowest error), so you feel around to find the steepest dow...

Anchoring AI responses in real data (retrieved documents, search results, database records) rather than relying on training data alone. RAG is the mos...

Rules and constraints you set to keep AI output safe, on-topic, and within boundaries. "Do NOT modify existing tests" is a guardrail. They prevent AI ...

When AI generates information that sounds confident but is factually wrong. It doesn't "know" it's making things up — it's pattern-matching, not fact-...

A pattern where sensitive AI actions (send email, delete files, deploy code) require human approval before execution. The agent proposes, the human ap...

Hyperparameters control the training process itself. Think of them as the settings on your oven. Unlike the ingredients (the data) or the recipe (the ...

The process of running a trained AI model to generate output. When you send a prompt and get a response, that's inference. Training teaches the model;...

Instruction tuning is like giving a student a cheat sheet of how to answer different types of questions before the exam. You train a language model on...

Knowledge distillation is like having a experienced teacher simplify complex lessons for a new student. It's a technique where a smaller, simpler mode...

A knowledge graph is like a map of related concepts. Imagine a big network of interconnected ideas, where each idea is a node, and the connections bet...

The time between sending a prompt and receiving the first token of the response. Lower latency means faster responses. Affected by model size, server ...

A massive AI model trained on enormous amounts of text data. Claude, GPT, and Gemini are all LLMs. "Large" refers to the billions of parameters (adjus...

An open protocol that lets AI models connect to external tools and data sources in a standardized way. Like USB for AI — one protocol, many tools. Cre...

A lightweight program that exposes tools to AI via the Model Context Protocol. One MCP server for Gmail means every AI app can send/read email. Build ...

Like DevOps, but specifically for machine learning workflows. MLOps is a set of practices that automates and standardizes the entire lifecycle of ML m...

When an AI model disproportionately favors certain outcomes or groups due to skewed or unrepresentative training data. Like a biased referee, it makes...

Think of a Model Card as a nutrition label for AI models. It's a document that reports key information like how the model was trained, its intended us...

Like making sure a new app works with your phone’s operating system — not every AI model can run on every platform or tool. Check compatibility before...

Model drift happens when a model's performance slides downhill because the real‑world data it encounters gradually differs from the data it was traine...

A model registry is like a library catalog for your AI models. It records each model version, its training data, performance metrics, and who approved...

An AI that can process multiple types of input — text, images, audio, video — not just words. A multimodal model can look at a photo and describe what...

An AI agent that coordinates other agents, breaking complex tasks into subtasks and assigning them. Like a project manager for AI teams. In multi-agen...

Think of it like a GPS navigation system that's only been trained on city roads, but then you take it off-roading. It's when a model encounters data t...

Overfitting is like a student who memorizes the answers to a practice test instead of learning the material. They might ace the practice test, but the...

The adjustable numbers inside a neural network that the model learns during training. More parameters generally means more capability, but also more c...

A RAG system built on your own knowledge base — blog posts, talks, documents, investor memos — so an AI can answer questions grounded in your actual e...

The input you give to an AI model — the question, instruction, or context that tells it what to do. Better prompts get better outputs. This is the ski...

The art and science of crafting prompts that get the best results from AI. Includes techniques like role prompting, chain-of-thought, few-shot example...

A security attack where malicious instructions are hidden in data that AI processes. Like SQL injection, but for AI prompts. Important to understand i...

Quantization is like shrinking a high‑resolution photo into a smaller file by reducing the number of colors, but for neural nets. It drops the precisi...

A technique that gives AI access to external data (documents, databases) at query time. Instead of relying only on training data, the model retrieves ...

A cap on how many API requests you can make in a given time period. Prevents abuse and ensures fair access. If you hit one, you need to slow down or u...

It's the proactive process of intentionally trying to break, trick, or find vulnerabilities in an AI system before it's deployed. Think of it as hirin...

Regularization is like adding a stabilizer to a bike - it helps the model not to overfit the training data by reducing its complexity. Think of it as ...

Reinforcement Learning (RL) is training an AI by rewarding desired behaviors and penalizing undesired ones. Think of it like teaching a dog tricks wit...

A training technique where humans rate AI outputs, and the model learns to produce responses humans prefer. It's why modern AI feels more helpful and ...

A collection of tools, libraries, and documentation that makes it easier to build with an API. The Anthropic SDK lets you use Claude in your Python or...

Self-attention lets a model weigh the importance of different parts of the input sequence when processing it. Think of it like reading a sentence and ...

Instead of keyword matching, semantic search understands the *intent* and *context* behind your query. It's like asking a librarian 'where can I find ...

A multi-step workflow an agent knows how to execute by chaining tools together. "Deploy a website" is a skill (read files → build → deploy → verify). ...

A secondary AI agent spawned by a primary agent to handle a specific subtask. Like delegating work to a team member. Multiple subagents can run in par...

Artificially generated data that statistically mimics real-world data without containing actual sensitive information. It's often used to augment trai...

Hidden instructions given to an AI that set its behavior, personality, and rules before the user starts chatting. It's the difference between "generic...

A setting that controls how random or creative AI output is. Low temperature (0.0-0.3) = deterministic and focused. High temperature (0.7-1.0) = more ...

The basic unit AI uses to process text. A token is roughly 3/4 of a word. "Hamburger" is 3 tokens. Tokens determine cost (you pay per token) and fit w...

The process of splitting text into tokens — the fundamental units an LLM processes. Different models use different tokenizers, so the same text can pr...

The ability of AI to call external functions — search the web, read files, run code, access APIs. This is what separates agents from chatbots. Without...

Using knowledge gained while solving one problem and applying it to a different but related problem. It's like a carpenter who's great at building hou...

The neural network architecture behind all modern LLMs. Introduced in the 2017 paper "Attention Is All You Need." Uses attention mechanisms to process...

A database designed to store and search embeddings (number representations of text/images). Used in RAG systems to find relevant documents quickly. Ex...

Instead of exact keyword matches, vector search finds results based on semantic similarity. It's like searching for "best Italian restaurants near me"...

Synonym for parameters. The numerical values inside a neural network that determine how it processes input. Training adjusts the weights; inference us...

Asking AI to do something without giving it any examples. "Translate this to French" is zero-shot. Contrast with few-shot (providing examples first). ...