AI has a vocabulary problem. Everyone uses the terms, almost no one explains them. This is the glossary that should have come with every AI tool you have ever downloaded.
How AI thinks
Machine Learning. AI that learns from examples instead of being explicitly programmed. You show it thousands of emails labeled spam or not spam, and it figures out the pattern on its own.
Deep Learning. A type of machine learning that uses layered networks loosely inspired by the brain. The reason AI got so good so fast — it can find patterns in enormous amounts of data.
Neural Network. The architecture behind deep learning. Data passes through layers of nodes, each one transforming it slightly, until the model produces an output. Think of it as a very complex chain of if-then decisions.
Training Data. The material a model learns from. Text, images, code, audio — whatever you feed it shapes what it knows and how it behaves. Bad data in, bad model out.
Parameters. The internal settings a model adjusts during training to get better at its task. A model with more parameters can capture more nuance — but also costs more to run.
Inference. What happens when a model actually runs — takes your input and produces an output. Training is the learning phase. Inference is the using phase.
Types of AI
Generative AI. AI that creates new content — text, images, audio, video, code. It doesn't retrieve stored answers; it generates something new each time based on patterns it learned.
LLM (Large Language Model). The engine behind most AI text tools. Trained on massive amounts of text to predict and generate language. GPT-4, Claude, Gemini — all LLMs.
Multimodal AI. A model that can work across more than one type of input or output — text and images, or text and audio. Claude can read images. GPT-4o can hear you speak.
Agentic AI. AI that does not just respond to a single prompt — it takes sequences of actions, uses tools, and works toward a goal with some autonomy. Less chatbot, more assistant that actually does things.
AI Agent. A specific instance of agentic AI built to complete a task or set of tasks. Agents can browse the web, write and run code, send emails, and chain steps together without you doing each one manually.
Foundation Model. A large model trained on broad data that can be adapted to many tasks. Most AI products you use are built on top of a foundation model — they don't train from scratch.
Reasoning Model. A model designed to think through problems step by step before answering, rather than just pattern-matching to a response. Better at math, logic, and multi-step problems.
ANI / AGI / ASI. Three levels of AI capability. ANI (narrow) is what we have now — good at specific tasks. AGI (general) would match human-level reasoning across domains. ASI (superintelligence) would exceed it. We are firmly in ANI territory.
How models work
Token. The unit of text a model reads and generates. Roughly a word or part of a word. Models have a limit on how many tokens they can process at once — that limit is the context window.
Context Window. How much text a model can hold in its working memory at one time. Older models had small windows — you had to summarize conversations. Modern models can hold entire documents.
Temperature. A setting that controls how predictable or creative a model's outputs are. Low temperature = consistent, conservative. High temperature = more varied, sometimes surprising.
Embeddings. A way of representing words and concepts as numbers so a model can understand relationships between them. The reason a model knows that king minus man plus woman equals queen.
RAG (Retrieval-Augmented Generation). A technique where a model pulls in relevant information from an external source before responding, instead of relying only on what it learned during training. How AI can answer questions about your specific documents.
Fine-Tuning. Taking a pre-trained model and training it further on a specific dataset to make it better at a particular task or tone. Different from prompting — you are actually changing the model's weights.
Using AI
Prompt. The input you give an AI model. The quality of your prompt directly affects the quality of the output. Prompting well is a real skill.
System Prompt. Instructions given to a model before the conversation starts, usually by the developer or product. It sets the tone, persona, and constraints. The reason Claude behaves differently inside different apps.
Hallucination. When a model generates something confident and wrong — a fake citation, a made-up statistic, a person who doesn't exist. It happens because models predict likely text, not because they know facts.
Open Source vs Closed Source. Open source models (like Llama) release their weights publicly so anyone can run or modify them. Closed source models (like Claude or GPT-4) are proprietary — you access them through an API or product.
API (Application Programming Interface). The way developers connect to an AI model programmatically. When an app uses Claude under the hood, it is calling the Claude API.
Automation vs AI. Automation follows fixed rules — if this, then that. AI makes judgments based on patterns. Automation is deterministic. AI is probabilistic. They are often used together but are not the same thing.
The bigger picture
Alignment. The challenge of making sure AI systems do what humans actually want — not just what they were literally told to do. A major area of research because a misaligned powerful AI is dangerous.
Guardrails. Constraints built into a model or product to prevent harmful outputs. Content filters, refusal behaviors, safety classifiers — all guardrails.
Synthetic Data. Training data that was generated by AI rather than collected from humans. Used to fill gaps, protect privacy, or scale up datasets cheaply.
Knowledge Cutoff. The date after which a model has no training data. Ask Claude about something that happened last week and it genuinely does not know — its knowledge stops at a fixed point.
Benchmark. A standardized test used to measure and compare model performance. Companies use benchmarks to show how their model stacks up. Worth knowing because they can be gamed.
MCP (Model Context Protocol). An open standard that lets AI models connect to external tools and data sources — your calendar, your files, your apps. The reason Claude can actually do things in the real world, not just talk about them.