The AI Glossary for Real People

An AI system that can take actions autonomously — browsing the web, writing code, calling APIs, managing files — to accomplish a goal with minimal human guidance. Unlike a chatbot that only responds to prompts, an agent plans multi-step workflows and executes them on its own.

The challenge of making AI systems behave in ways that match human values and intentions. Alignment research focuses on ensuring AI does what we actually want, not just what we technically asked for. It spans safety, ethics, and the gap between instruction and intent.

A set of rules that lets one software application talk to another. In AI, APIs are how you send prompts to models like GPT or Claude and get responses back — without needing to run the model yourself. Think of it as a waiter taking your order to the kitchen.

A technique that lets a neural network focus on the most relevant parts of its input when generating output. Instead of treating every word equally, attention assigns weights so the model can "pay attention" to what matters most for each prediction. It is the core innovation behind transformers.

Tools and techniques that automate the process of building machine learning models — from data preprocessing to feature selection to model tuning. AutoML lets non-experts create production-quality models without writing algorithms from scratch.

The algorithm that lets neural networks learn from their mistakes. After making a prediction, the network calculates how wrong it was (the loss), then works backward through each layer to adjust its internal weights so the next prediction is better. It is the fundamental training algorithm for all deep learning.

A language model from Google that reads text in both directions (left-to-right and right-to-left) simultaneously, giving it a deeper understanding of context. BERT revolutionized search, sentiment analysis, and question-answering when it launched in 2018.

Systematic errors in AI output caused by skewed training data, flawed assumptions, or unrepresentative samples. If a hiring model is trained mostly on resumes from one demographic, it may unfairly penalize others. Bias can be statistical (a technical modeling issue) or societal (reflecting real-world inequities baked into data).

The number of training examples a model processes before updating its internal weights. A batch size of 32 means the model looks at 32 examples, calculates the average error, and then adjusts. Larger batches are more stable but use more memory; smaller batches learn noisier but can generalize better.

A prompting technique where you ask the AI to show its reasoning step-by-step before giving a final answer. Instead of jumping to a conclusion, the model "thinks out loud," which dramatically improves accuracy on math, logic, and complex reasoning tasks.

OpenAI's conversational AI product, built on the GPT family of large language models. Launched in November 2022, ChatGPT brought AI into the mainstream by giving anyone a simple chat interface to interact with a powerful language model. It can write, code, analyze, summarize, and brainstorm.

A type of machine learning task where the model assigns input data to predefined categories. Is this email spam or not spam? Is this X-ray showing a tumor or healthy tissue? Classification models output a label (or probability of each label) rather than a continuous number.

Anthropic's family of AI assistants, designed with a focus on safety, helpfulness, and honesty. Claude models (Haiku, Sonnet, Opus) offer strong reasoning, long context windows, and careful handling of sensitive topics. Claude is a leading alternative to ChatGPT in both consumer and enterprise use.

The field of AI that teaches machines to interpret and understand visual information — images, videos, and live camera feeds. Computer vision models can detect objects, recognize faces, read text in photos, and understand scenes with near-human accuracy.

The maximum amount of text (measured in tokens) that an AI model can consider at once — both your input and its output combined. A model with a 200K context window can process roughly 150,000 words in a single conversation, while a 4K window limits you to about 3,000 words.

A mathematical measure of how similar two vectors are, based on the angle between them rather than their magnitude. In AI, it is the standard way to compare embeddings — two documents about "machine learning" will have vectors pointing in nearly the same direction, yielding a cosine similarity close to 1.

A subset of machine learning that uses neural networks with many layers (hence "deep") to learn complex patterns from data. Deep learning powers image recognition, speech-to-text, language translation, and generative AI. It thrives on large datasets and GPU computing power.

A type of generative model that creates images (or other data) by learning to reverse a process of gradually adding noise. During training, the model learns how to "denoise" — and at generation time, it starts from pure random noise and iteratively refines it into a coherent image.

One half of a GAN (Generative Adversarial Network). The discriminator's job is to distinguish between real data and fake data produced by the generator. As it gets better at spotting fakes, the generator is forced to produce more realistic output — creating a competitive training loop.

A technique for compressing a large, powerful AI model (the "teacher") into a smaller, faster model (the "student") that retains most of the teacher's performance. The student learns to mimic the teacher's outputs rather than training from scratch on raw data.

A way of representing text, images, or other data as a list of numbers (a vector) that captures its meaning. Words with similar meanings end up close together in this numerical space — "king" and "queen" are neighbors, while "king" and "toaster" are far apart.

A neural network architecture with two parts: an encoder that reads and compresses input into a representation, and a decoder that generates output from that representation. Originally designed for translation (English in, French out), this architecture influenced the design of transformers.

One complete pass through the entire training dataset. If you have 10,000 examples and train for 3 epochs, the model sees every example three times. More epochs can improve learning but risk overfitting — memorizing the training data instead of learning general patterns.

Standardized measurements used to assess how well an AI model performs. Common metrics include accuracy (% correct), precision (how many positives were truly positive), recall (how many actual positives were found), and F1 score (the balance between precision and recall).

The ability to understand and explain why an AI model made a specific decision. Explainability techniques let you peek inside the "black box" to see which features or inputs drove the output. This is critical in regulated industries where decisions must be justified.

A technique where you give an AI model a handful of examples in the prompt to teach it what you want — without any retraining. Show the model 3 examples of how to format a response, and it generalizes the pattern to new inputs. Zero-shot means no examples; few-shot means a few.

Taking a pre-trained AI model and training it further on a smaller, specialized dataset to improve performance on a specific task or domain. Instead of training from scratch (which costs millions), you start with what the model already knows and teach it your specific needs.

A large AI model trained on broad, diverse data that can be adapted to many downstream tasks. GPT-4, Claude, Llama, and Gemini are all foundation models. They serve as the "foundation" that gets fine-tuned, prompted, or extended for specific applications.

A pair of neural networks — a generator and a discriminator — that compete against each other. The generator creates fake data, the discriminator tries to spot the fakes, and both improve through competition. GANs were the dominant image generation technique before diffusion models.

A family of large language models from OpenAI. GPT models are "generative" (they produce text), "pre-trained" (trained on massive data before specialization), and built on the "transformer" architecture. GPT-4 and GPT-4o are the latest versions powering ChatGPT.

The optimization algorithm that neural networks use to learn. It works by calculating the slope (gradient) of the error function and taking small steps "downhill" to reduce error. Imagine being blindfolded on a hill and feeling the slope under your feet to find the valley — that is gradient descent.

Connecting an AI model's outputs to verifiable, real-world sources of truth — documents, databases, APIs, or live data. Grounding reduces hallucinations by forcing the model to base its answers on actual evidence rather than just its training data.

Rules, filters, and safety mechanisms that constrain what an AI model can say or do. Guardrails prevent harmful outputs, keep responses on-topic, enforce formatting rules, and block sensitive information from leaking. They are the "bumper lanes" that keep AI behavior within acceptable bounds.

When an AI confidently generates information that is factually wrong or completely fabricated. The model is not "lying" — it is predicting the most plausible next tokens, and sometimes plausible text is completely false. Hallucinations can include fake citations, invented statistics, or fictional events.

A graph-based algorithm for extremely fast approximate nearest-neighbor search in high-dimensional vector spaces. When you search a vector database with millions of embeddings, HNSW is the algorithm that finds the closest matches in milliseconds instead of scanning every single vector.

A setting you choose before training begins — like learning rate, batch size, number of layers, or dropout rate. Unlike regular parameters (which the model learns during training), hyperparameters are the knobs the engineer turns. Getting them right can mean the difference between a mediocre model and a great one.

The process of using a trained AI model to make predictions on new data. Training is when the model learns; inference is when it applies what it learned. Every time you send a prompt to ChatGPT or Claude and get a response, that is inference happening on a server somewhere.

The ability of large language models to learn new tasks from examples provided directly in the prompt — without any weight updates or retraining. You show the model a few examples, and it picks up the pattern instantly. This is what makes LLMs so flexible compared to traditional ML models.

A feature in AI APIs that forces the model to output valid JSON (JavaScript Object Notation) — structured, machine-readable data instead of free-form text. This is essential when the AI's output needs to be parsed by code, stored in a database, or passed to another system.

A model compression technique where a smaller "student" model is trained to replicate the behavior of a larger "teacher" model. The student learns not just the correct answers but the teacher's confidence levels across all possible answers — capturing nuanced knowledge in a fraction of the size.

One of the simplest machine learning algorithms: to classify a new data point, look at the k closest data points in your training set and go with the majority vote. If 4 out of 5 nearest neighbors are "spam," classify the new email as spam. No training required — just comparison.

A popular open-source framework for building applications with large language models. LangChain provides pre-built components for common AI patterns — chaining prompts together, connecting to databases, adding memory to conversations, and building agents. It is the "React of AI apps."

A neural network trained on massive text datasets that can generate, summarize, translate, and reason about human language. LLMs like GPT-4, Claude, Gemini, and Llama contain billions of parameters and can perform tasks they were never explicitly trained for, simply by following instructions in natural language.

A technique for fine-tuning large models efficiently by only updating a small number of additional parameters instead of the entire model. LoRA "freezes" the original model weights and trains tiny adapter layers, making fine-tuning 10-100x cheaper in compute and memory.

A mathematical function that measures how wrong a model's predictions are. The model's entire goal during training is to minimize its loss function. Different tasks use different loss functions — cross-entropy for classification, mean squared error for regression, etc.

A subset of AI where systems learn patterns from data instead of being explicitly programmed with rules. Instead of writing "if temperature > 90, then hot," you show the model thousands of examples and it learns the pattern on its own. ML includes supervised learning, unsupervised learning, and reinforcement learning.

An open protocol developed by Anthropic that standardizes how AI models connect to external tools, data sources, and services. MCP provides a universal interface — like USB for AI — so models can read files, query databases, call APIs, and use tools through a single, consistent protocol.

A model architecture that routes each input to a subset of specialized "expert" networks rather than running the entire model. A MoE model might have 8 experts but only activate 2 for any given token, giving it the knowledge of a huge model with the compute cost of a small one.

An AI model that can process and generate multiple types of data — text, images, audio, video, code — within a single system. GPT-4o, Claude, and Gemini are multimodal: you can show them a photo, ask a question about it, and get a text answer.

The field of AI focused on enabling computers to understand, interpret, and generate human language. NLP covers everything from spell-check and autocomplete to machine translation, sentiment analysis, and the conversational abilities of ChatGPT and Claude.

A computing system inspired by the human brain, made up of layers of interconnected nodes (neurons) that process information. Each connection has a weight that gets adjusted during training. Stacking many layers creates a "deep" neural network — the foundation of deep learning and modern AI.

A technique for scaling input data or internal network activations to a standard range, which helps models train faster and more stably. Batch normalization and layer normalization are common variants. Without normalization, neural networks can struggle with "exploding" or "vanishing" values during training.

The AI research company behind ChatGPT, GPT-4, DALL-E, and Whisper. Founded in 2015 as a nonprofit, OpenAI transitioned to a "capped profit" structure and has become the most commercially prominent AI company. Their products brought generative AI to the mainstream.

When a model memorizes the training data so well that it performs poorly on new, unseen data. An overfitted model is like a student who memorizes the answer key but cannot solve a new problem. It scores perfectly on practice tests but fails the real exam.

An open format for representing machine learning models that allows you to move models between different frameworks (PyTorch, TensorFlow, etc.) and deploy them anywhere. Think of ONNX as a universal file format — like PDF for AI models.

A single learnable value inside a neural network — essentially one number that gets adjusted during training. When people say a model has "70 billion parameters," they mean it has 70 billion tunable numbers. More parameters generally means more capacity to learn patterns, but also more compute to train and run.

A metric that measures how "surprised" a language model is by new text. Lower perplexity means the model predicts the text more confidently and accurately. A model with perplexity of 10 means it is, on average, choosing between 10 equally likely next words.

The initial, large-scale training phase where a model learns general knowledge from massive datasets — often trillions of tokens from the internet, books, and code. Pre-training is the expensive part (millions of dollars in compute) that creates a foundation model, which can then be fine-tuned cheaply for specific tasks.

The practice of crafting inputs (prompts) to an AI model to get the best possible output. This includes writing clear instructions, providing examples, setting context, specifying output format, and using techniques like chain-of-thought. It is part science, part craft, and part communication skill.

The most popular open-source framework for building and training deep learning models, created by Meta. PyTorch uses a "define-by-run" approach that feels natural to Python developers, making it the default choice for AI research and increasingly for production deployments.

A technique for reducing model size and speeding up inference by using lower-precision numbers. Instead of storing each parameter as a 32-bit floating point number, you can use 8-bit or even 4-bit integers. This can shrink a model by 4-8x with minimal quality loss.

In the attention mechanism, each token generates three vectors: a Query, a Key, and a Value. The Query represents "what am I looking for?" It gets matched against Keys from other tokens to determine which tokens to pay attention to, and the corresponding Values are combined into the output.

A technique that gives an AI model access to external documents so it can answer questions with up-to-date, source-backed information. RAG works in two steps: first, it searches a knowledge base for relevant documents; then, it feeds those documents to the LLM along with the user's question.

An AI model's ability to think through problems logically, break down complex questions, weigh evidence, and arrive at conclusions through multi-step inference. Reasoning capabilities separate simple pattern matching from genuine problem-solving — and have improved dramatically in recent models like o1, Claude, and Gemini.

A type of machine learning where an agent learns by interacting with an environment and receiving rewards or penalties for its actions. Instead of learning from labeled examples, it learns from trial and error — like a dog learning tricks through treats. RLHF (RL from Human Feedback) is how ChatGPT was trained to be helpful.

The most commonly used activation function in neural networks. It simply outputs the input if it is positive, and zero if it is negative: f(x) = max(0, x). Despite its simplicity, ReLU solved a critical training problem (vanishing gradients) and remains the default choice in most architectures.

The process of finding and fetching relevant information from a large collection of documents or data. In AI, retrieval has evolved from keyword matching to semantic search using embeddings — finding documents that mean similar things, not just contain the same words.

Search that understands meaning rather than just matching keywords. If you search for "how to fix a slow computer," semantic search also returns results about "optimizing PC performance" and "speeding up your laptop" — because it understands these mean the same thing, even though the words are different.

A specific type of attention mechanism where each element in a sequence attends to all other elements in the same sequence. In a sentence, each word looks at every other word to understand context. "The bank was steep" uses self-attention to figure out that "bank" means riverbank, not a financial institution.

An open-source AI image generation model that creates images from text descriptions. Unlike DALL-E (closed-source), Stable Diffusion can be downloaded, modified, and run locally. It sparked an explosion of AI art tools, workflows, and fine-tuned variants for specific styles and use cases.

A type of machine learning where the model learns from labeled examples — input-output pairs where the correct answer is provided. Show the model 10,000 photos labeled "cat" or "dog," and it learns to classify new photos. Most production ML systems use supervised learning.

A hidden instruction given to an AI model before the user's conversation begins, defining its behavior, personality, constraints, and role. The system prompt is like a job description for the AI — it sets the rules the model follows throughout the conversation, even if the user tries to override them.

A setting that controls how random or creative an AI model's output is. Temperature 0 makes the model always pick the most likely next word (deterministic, factual), while temperature 1+ makes it more willing to pick less likely words (creative, varied, but potentially less accurate).

The basic unit of text that AI models process. A token is roughly 3/4 of a word in English — "hamburger" becomes ["ham", "burger"]. Models read, think in, and generate tokens, not words. When a model has a "128K context window," that means 128,000 tokens, or roughly 96,000 words.

The component that converts raw text into tokens (numbers) that a model can process, and back again. Different models use different tokenizers with different vocabularies — this is why the same sentence might be 15 tokens in GPT-4 but 18 tokens in Claude. The tokenizer determines a model's "alphabet."

The neural network architecture behind GPT, Claude, Gemini, and virtually every modern AI model. Introduced in the 2017 paper "Attention Is All You Need," the transformer uses self-attention to process all words in parallel rather than one at a time, enabling massive scaling and superior language understanding.

Using knowledge learned from one task to improve performance on a different but related task. A model trained to recognize cats can transfer that knowledge to recognizing dogs much faster than starting from scratch. In LLMs, the entire pre-train/fine-tune paradigm is transfer learning at massive scale.

The dataset used to teach a machine learning model. Training data quality directly determines model quality — garbage in, garbage out. For LLMs, training data includes trillions of tokens from the internet, books, code, and curated sources. For custom models, it is your labeled, domain-specific examples.

When a model is too simple to capture the patterns in the data — it performs poorly on both training data and new data. If overfitting is memorizing the answer key, underfitting is not studying at all. Common causes: too few parameters, too little training, or overly aggressive regularization.

A type of machine learning where the model finds patterns in data without labeled examples. Instead of being told "this is a cat," the model discovers clusters and structures on its own. Unsupervised learning powers customer segmentation, anomaly detection, and dimensionality reduction.

A specialized database that stores data as numerical vectors (embeddings) and retrieves results by similarity rather than exact matching. When you search "documents about company growth," a vector database finds semantically similar content even if those exact words never appear. Popular options include Pinecone, Weaviate, Chroma, and Qdrant.

An AI model specifically designed to understand and process visual information — images, diagrams, screenshots, documents, and video frames. Vision models can describe images, extract text, identify objects, and answer questions about visual content. Most modern LLMs now include vision capabilities.

A type of generative model that learns to compress data into a compact "latent space" and then reconstruct it. VAEs can generate new data by sampling from this latent space. They are used as components in diffusion models (like Stable Diffusion) and for data compression, generation, and anomaly detection.

A numerical value in a neural network that determines the strength of the connection between two neurons. During training, weights are adjusted via backpropagation to minimize error. The "knowledge" of a trained model is entirely encoded in its weights — they are what make a model smart.

A pioneering algorithm from Google (2013) that represents words as numerical vectors, capturing semantic relationships. Word2Vec famously demonstrated that "king - man + woman = queen" in vector space. While superseded by transformer-based embeddings, Word2Vec proved that meaning could be represented as math.

The ability of an AI model to perform a task it has never been explicitly shown examples of. You simply describe the task in natural language — "classify this review as positive or negative" — and the model does it, without a single training example. This is a hallmark capability of large language models.