AI Starter Pack

AI has become one of those fields where everyone seems to speak in acronyms. So before we dive any deeper, let’s learn the vocabulary. Think of this as your AI phrasebook.

AI isn’t one thing

Artificial intelligence has become the catch-all term for everything from ChatGPT to self-driving cars, but it’s really just the umbrella.

Underneath that umbrella is machine learning, where computers learn patterns from data instead of following hand-written rules.

A subset of machine learning is deep learning, which uses neural networks—mathematical models inspired, very loosely, by the human brain. Deep learning is what made the recent explosion in AI possible.

Then came foundation models: large models trained on enormous amounts of text, code, images, and other data. Instead of building a different model for every task, researchers built models that could adapt to thousands of tasks.

When a foundation model specializes in language, you get an LLM, or large language model. ChatGPT, Claude, Gemini, and Llama are all powered by LLMs.

You’re not searching. You’re collaborating.

One of the biggest mindset shifts is understanding that talking to an AI isn’t like using Google.

SearchFind existing information

You enter keywords and choose among results.

CollaborateCreate and reason together

You give context, direction, feedback, and examples.

Your instruction is called a prompt. Think of it less like a search query and more like giving directions to a very capable intern. The clearer your instructions, the better the result.

For a while, everyone talked about prompt engineering: the art of writing better prompts. More recently, people have started talking about context engineering, which is arguably even more important.

Instead of finding the perfect wording, the goal is to give the model the right context: your documents, previous conversations, company knowledge, tools, examples, and instructions. In real-world AI products, getting the context right often matters more than wording the prompt perfectly.

AI has a memory… sort of

One misconception is that ChatGPT remembers everything. It doesn’t.

Think of the context window as a whiteboard.

Everything currently on the board is visible to the model. Once the board fills up, older information may be removed or summarized.

Behind the scenes, the model isn’t reading words the way humans do. It breaks information into small chunks called tokens. Every conversation, PDF, and line of code ultimately becomes tokens—the working language of the model.

You may also hear about temperature, a setting that controls randomness. Lower temperatures produce more predictable answers; higher temperatures can make outputs more varied and creative.

And yes, AI still makes mistakes. When it confidently invents facts or cites papers that don’t exist, that’s called a hallucination. Despite the dramatic name, it’s usually the model generating something that sounds plausible but isn’t true.

Why AI suddenly feels smarter

One reason today’s models feel more capable is that they don’t always respond immediately. Many use additional reasoning before answering, spending more computation working through a problem before producing a response.

You’ll also hear three terms that describe how models are created and adapted:

Training: The expensive process of teaching a model from very large datasets.
Fine-tuning: Taking an already-trained model and teaching it a specialized behavior or skill.
Distillation: Teaching a smaller model to imitate a larger one, making AI cheaper and faster to run.

Modern AI doesn’t work alone

A language model by itself is limited to what it learned during training and what you put in its context. That’s why modern AI systems combine models with tools.

If the AI needs today’s weather, it can call a weather service. If it needs your calendar, it can connect to your calendar. If it needs to calculate something, it can use a calculator instead of trying to do the math from memory. This is called tool calling, and it’s a major reason AI has become more useful.

Prompts, skills, and agents

A skill is a reusable capability that teaches an AI how to perform a specific kind of task. Skills can bundle instructions, tools, workflows, and domain expertise into a repeatable way of working.

An agent has a goal. It can make a plan, use tools or skills, gather information, remember previous steps, and decide what to do next. Ask an agent to plan a vacation and it might compare flights, build an itinerary, create a packing list, and add events to your calendar—not simply answer a question.

Some systems use multiple agents, each with a role: one researches, another writes, another reviews, and another checks for mistakes. An orchestrator coordinates their work.

Giving AI your knowledge

Sometimes the model needs information it was never trained on, like your company documentation. Instead of memorizing everything, it can search for relevant information first and answer using those documents. That’s called retrieval-augmented generation, or RAG.

Those documents are often represented as embeddings and stored in a vector database, allowing the system to search by meaning rather than only exact keywords. That’s known as semantic search.

Connecting everything together

As AI becomes integrated into more software, every application speaking a different language creates another challenge.

MCP is like USB-C for AI.

The Model Context Protocol gives AI systems a common way to connect with calendars, databases, file systems, design tools, development environments, and other services.

Instead of building a custom integration for every model and every tool, MCP provides a shared connection pattern. That makes it easier for AI to work with the information and software people already use.

The short version

You don’t need to memorize the dictionary.

You only need enough shared language to ask better questions, spot the useful opportunities, and participate in the decisions shaping how your organization uses AI.

Bring this learning to your team →

Reference library

The AI glossary, A–Z

Every definition has its own link. Bookmark the page—or send someone directly to the term they need.

AGI (Artificial General Intelligence)#: A proposed kind of AI able to learn and reason across many domains at roughly human level or beyond.
Agentic AI#: AI that can pursue a goal by planning steps, using tools, responding to feedback, and taking actions with some independence.
AI alignment#: The work of making an AI system behave in line with human intentions, values, and rules.
AI benchmarks#: Shared tests used to compare how models perform on tasks such as coding, image recognition, or instruction following.
AI fluency#: The practical judgment and skills needed to work productively, critically, and responsibly with AI.
AI safety#: The field concerned with making AI dependable and preventing harm, including in unusual or high-impact situations.
AI workflow#: The connected steps used to create and operate an AI system, from preparing data through testing, deployment, and upkeep.
Algorithm#: A defined sequence of instructions for completing a task or solving a problem.
Artificial intelligence (AI)#: The broad field of building machines that perform tasks associated with intelligence, such as learning, reasoning, or perception.
Attention mechanism#: A technique that lets a neural network weigh which parts of its input matter most for the output it is producing.
Automation#: Technology carrying out a task according to instructions and boundaries established by people.
Backpropagation#: A training process that traces prediction errors backward through a neural network so its internal values can be corrected.
Bayesian network#: A map of variables and their probabilistic relationships, used to reason about how factors may influence one another.
Bias (in AI)#: A systematic skew in an AI system that can produce unfair or discriminatory outcomes.
Big data#: Datasets so large or complex that specialized storage and analysis techniques are needed to find useful patterns.
Chatbot#: Software designed to hold a conversation with a person through text, voice, or both.
Closed source#: Software whose underlying source code is controlled by its owner and is not publicly available to inspect or modify.
Computer vision#: The area of AI focused on extracting meaning from images and video.
Context window#: The amount of information a model can consider together during one interaction or task.
Contrastive learning#: A training method that teaches a model by showing which examples belong together and which are different.
Data augmentation#: Expanding a training set by creating useful variations of existing examples rather than collecting entirely new ones.
Data mining#: Using statistical and computational methods to uncover patterns, relationships, or trends in large datasets.
Decision tree#: A model that reaches a prediction through a branching series of questions about the input.
Deep learning#: Machine learning that uses neural networks with many layers to learn complex patterns.
Diffusion model#: A generative model that learns to create content by reversing a gradual process of adding noise.
Dimensionality reduction#: Compressing complex data into fewer variables while trying to preserve its most useful structure.
Embeddings#: Number-based representations that place similar pieces of text, images, or other data near one another in a mathematical space.
Ensemble methods#: Techniques that combine several models so their collective prediction is often stronger than any one model alone.
Ethical AI#: Designing and using AI with attention to fairness, transparency, human values, and social consequences.
Expert system#: A program that applies a stored body of domain knowledge and rules to imitate specialist decision-making.
Explainable AI (XAI)#: Methods that help people understand why an AI system produced a particular output or decision.
Federated learning#: Training a shared model across separate devices or servers while keeping the original data in its local location.
Few-shot learning#: Learning or performing a new task from only a small number of examples.
Fine-tuning#: Further training an existing model on specialized examples to shape it for a particular domain or behavior.
Foundation model#: A large model trained broadly enough to serve as the starting point for many different applications.
Generative AI#: AI that creates new material such as text, images, audio, video, or software code.
Generative adversarial network (GAN)#: A system in which a generator creates synthetic examples while a second network judges them, pushing the generator toward realism.
Generative pre-trained transformer (GPT)#: A family of language models based on the transformer architecture and trained to generate content from prompts.
GPU (Graphics Processing Unit)#: A chip built to perform many calculations in parallel, making it useful for training and running AI models.
Hallucination#: An incorrect or invented AI output presented with the appearance of confidence or factuality.
Human-centered AI#: An approach that makes human needs, agency, values, and well-being central to AI design and use.
Human-computer interaction (HCI)#: The study and design of how people use and experience computer systems.
Human in the loop#: A system in which people review, guide, correct, or approve part of an AI process.
Hyperparameter#: A model setting chosen before training, such as a learning rate or number of training passes.
Inference#: The stage when a trained model uses new input to generate a prediction, decision, or response.
Interpretability#: How readily a person can understand the factors and process behind a model’s output.
k-Nearest Neighbors (k-NN)#: A method that predicts an item’s category or value by looking at the most similar examples nearby.
Knowledge graph#: Information organized as connected entities and relationships, forming a network that software can query and reason over.
Large language model (LLM)#: A model trained on very large collections of language data to interpret and generate text and related content.
Latent space#: A compressed mathematical space in which a model represents important features and relationships in complex data.
LLMOps#: The practices used to deploy, monitor, evaluate, secure, and maintain applications built around large language models.
Loss function#: A calculation that tells a model how far its predictions are from the desired answers during training.
Machine learning (ML)#: A part of AI in which systems learn patterns from data rather than relying only on explicitly written rules.
Markov chain#: A model of sequential events where the next state depends on the current state rather than the full path taken to reach it.
MLOps#: Engineering practices for reliably deploying, monitoring, updating, and governing machine-learning models in production.
Model#: A learned mathematical or computational representation used to identify patterns, make predictions, or generate outputs.
Model drift#: A decline in model performance as real-world conditions and data move away from what the model learned during training.
Multimodal AI#: AI that can work across more than one kind of information, such as text, images, audio, and video.
Natural language processing (NLP)#: The area of AI concerned with understanding, analyzing, and generating human language.
Neural network#: A layered computational system of connected units that learns patterns by adjusting the strength of its connections.
Optimization algorithm#: A method for adjusting a model during training so that its errors become smaller.
Overfitting#: When a model learns its training examples too specifically and performs poorly on unfamiliar data.
Open-weight model#: A model whose learned numerical weights are released for others to download and run, though its full training process may remain closed.
Open source#: Software whose source code is made available under a license that permits people to inspect, use, and often modify it.
Parameter#: An internal numerical value learned during training that helps determine how a model transforms inputs into outputs.
Predictive analytics#: Using historical data and statistical or machine-learning techniques to estimate likely future outcomes.
Prompt engineering#: Designing and refining instructions, examples, and context to guide a model toward a useful response.
Prompt injection#: A security attack that places deceptive instructions in model input to override intended behavior or expose protected information.
Retrieval-augmented generation (RAG)#: A pattern in which a model first retrieves relevant external information and then uses it to produce a grounded response.
Reasoning model#: A model optimized to spend additional computation working through multi-step or difficult problems before answering.
Reinforcement learning#: A training approach in which an agent learns through actions and feedback that rewards better outcomes.
Responsible AI#: The policies and practices used to develop and operate AI safely, fairly, transparently, and accountably.
Robotics#: The engineering and computing field focused on machines that sense, decide, and act in the physical world.
Spatial intelligence#: The capacity to understand three-dimensional environments, including where objects are and how they move or interact.
Self-supervised learning#: Training on raw data by creating learning signals from the data itself, such as predicting a hidden piece.
Semantic analysis#: Determining meaning in language by examining words, phrases, context, and their relationships.
Scaling laws#: Observed relationships that estimate how model performance changes as data, computing power, or model size increases.
Supervised learning#: Training a model with examples paired with the correct labels or target answers.
Synthetic data#: Artificially generated examples created by software or simulations rather than recorded directly from the real world.
Tensor#: A multi-dimensional arrangement of numbers used to represent and process data inside machine-learning systems.
Tokenization#: Splitting text into processable units that may be whole words, word fragments, characters, or punctuation.
Traditional AI#: Earlier AI approaches built largely from explicit logic, rules, and human-authored knowledge.
Training data#: The examples a learning system studies to develop its behavior, such as text, images, audio, or structured records.
Transfer learning#: Reusing knowledge learned for one task as the starting point for a different but related task.
Transformer#: A neural-network architecture that uses attention to model relationships across an input efficiently.
Turing test#: Alan Turing’s proposed test of whether a machine’s conversational behavior can be distinguished from a human’s.
Unsupervised learning#: Training with data that has no supplied answers so the model must discover structure or groupings on its own.
Vector database#: A database designed to store and quickly search embeddings, often by similarity of meaning.
Vision transformer (ViT)#: A version of the transformer architecture adapted to analyze images.
Weights#: The learned numbers that control the strength of connections in a neural network and shape its behavior.
Zero-shot learning#: Performing a task or recognizing a category without having been given direct training examples for that exact case.

Terminology coverage informed by the Stanford HAI AI glossary; definitions rewritten in the AI Starter Pack voice.