AI conversations are full of confusing terms. LLMs, tokens, hallucinations, fine-tuning—it can feel like everyone’s speaking another language.

This glossary cuts through the jargon. Every term is explained in plain English with practical context for how it affects you as a user.

Bookmark this page. You’ll probably come back to it.

Core AI Concepts

Artificial Intelligence (AI)

The broad field of creating machines that can perform tasks typically requiring human intelligence—understanding language, recognizing images, making decisions, solving problems.

In everyday conversation, “AI” usually refers to the new generation of tools like ChatGPT, not robots or science fiction scenarios.

Machine Learning (ML)

A subset of AI where systems learn from data rather than being explicitly programmed. Instead of writing rules for every situation, developers feed the system examples and let it find patterns.

When someone says an AI was “trained,” they’re talking about machine learning.

Deep Learning

A type of machine learning using neural networks with many layers (hence “deep”). This approach powers most modern AI—image recognition, language understanding, voice assistants.

You don’t need to understand the technical details. Just know it’s the underlying technology making today’s AI breakthroughs possible.

Neural Network

A computing system loosely inspired by the human brain. It consists of interconnected nodes (like neurons) that process information in layers.

Think of it as the architecture that allows AI to find complex patterns in data.

Training

The process of teaching an AI system by exposing it to large amounts of data. During training, the AI adjusts its internal parameters to get better at its task.

When someone says a model was “trained on the internet,” they mean it learned from massive amounts of text data.

Inference

When an AI applies what it learned to new inputs. Training is the learning phase; inference is the using phase.

Every time you chat with ChatGPT, that’s inference—the model is using what it learned to respond to you.

Language Models

Large Language Model (LLM)

An AI trained on massive amounts of text that can understand and generate human language. ChatGPT, Claude, Llama, and Gemini are all LLMs.

“Large” refers to both the training data (billions of words) and the model size (billions of parameters).

GPT (Generative Pre-trained Transformer)

A specific type of language model architecture developed by OpenAI. GPT models are trained to predict the next word in a sequence.

ChatGPT uses GPT models. The name has become somewhat generic—people sometimes use “GPT” to mean any similar model.

Transformer

A neural network architecture that revolutionized natural language processing. It’s especially good at understanding context and relationships in text.

The “T” in GPT, BERT, and other model names stands for Transformer.

Parameters

The internal values that a neural network adjusts during training. A model with more parameters can potentially learn more complex patterns.

When you hear “GPT-4 has over a trillion parameters,” that’s describing the model’s size and complexity.

Context Window

How much text an AI can consider at once. If a model has an 8,000-token context window, it can only “see” about 6,000 words of conversation at a time.

Larger context windows mean the AI can handle longer documents and maintain context in extended conversations.

Token

A unit of text that AI processes—roughly ¾ of a word in English. “Artificial intelligence” might be 3-4 tokens.

Tokens matter because AI pricing and limits are often based on token count.

AI Behaviors and Outputs

Hallucination

When an AI generates information that sounds plausible but is factually incorrect or completely made up. It’s not lying—it’s generating text that fits patterns without verifying facts.

Always verify important information from AI, especially specific facts, numbers, and citations.

Prompt

The input you give to an AI—your question, instruction, or request. The quality of your prompt significantly affects the quality of the response.

Response / Completion

The AI’s output—what it generates in reply to your prompt.

Temperature

A setting that controls randomness in AI responses. Higher temperature means more creative and varied outputs; lower temperature means more consistent and predictable.

Most users don’t adjust this directly, but it’s why the same prompt can give slightly different answers.

Alignment

How well an AI’s behavior matches human intentions and values. An “aligned” AI does what humans actually want, not just what they literally ask for.

This is why AI safety researchers work on—making sure powerful AI behaves helpfully and safely.

Training and Customization

Pre-training

The initial training phase where an AI learns general knowledge from massive datasets. This gives it broad capabilities before any specialization.

Fine-tuning

Additional training on specific data to specialize a model for particular tasks or domains. A general model might be fine-tuned on medical texts to become better at healthcare questions.

RLHF (Reinforcement Learning from Human Feedback)

A training technique where humans rate AI outputs, and the model learns from those ratings. This helps AI give responses that humans find more helpful and appropriate.

ChatGPT and Claude both use RLHF to improve their responses.

LoRA (Low-Rank Adaptation)

A efficient fine-tuning method that adds small trainable pieces to a model without changing the whole thing. Popular in the Stable Diffusion community for adding styles or concepts.

Embedding

A way of representing text (or images) as numbers that capture meaning. Similar concepts have similar embeddings.

This is how AI can understand that “happy” and “joyful” are related even though the words look different.

Vector Database

A database optimized for storing and searching embeddings. Used for semantic search—finding information based on meaning rather than exact keywords.

Types of AI Systems

Generative AI

AI that creates new content—text, images, music, video. ChatGPT generating text and Midjourney generating images are both generative AI.

Multimodal AI

AI that works with multiple types of data—text, images, audio. GPT-4V can see images and discuss them; some models can generate both text and images.

Diffusion Model

A type of generative AI that creates images by starting with noise and gradually refining it. Stable Diffusion, DALL-E, and Midjourney all use diffusion models.

Foundation Model

A large AI model trained on broad data that can be adapted to many tasks. GPT-4, Claude, and Llama are foundation models—they’re not built for just one thing.

Open Source / Open Weight

AI models where the code and/or trained weights are publicly available. Anyone can download and use them. Llama and Stable Diffusion are open; GPT-4 and Claude are not.

Closed Source / Proprietary

AI models that are only available through the company’s own services. You can use them but can’t download or modify them.

Usage and Access

API (Application Programming Interface)

A way for software to communicate with AI services. Developers use APIs to build apps powered by AI—the app sends requests to the AI and receives responses.

If you’re just using ChatGPT in a browser, you’re not using the API. But apps you use might be.

Endpoint

The specific URL or address where an API can be reached. Different endpoints might offer different models or capabilities.

Rate Limit

Restrictions on how quickly or how much you can use an AI service. Prevents any one user from overloading the system.

Latency

The time delay between sending a request and receiving a response. Lower latency means faster responses.

Streaming

Receiving AI responses word by word as they’re generated, rather than waiting for the complete response. This is why ChatGPT responses appear gradually.

Safety and Ethics

Content Moderation

Systems that filter AI inputs and outputs to prevent harmful content. Why AI will refuse certain requests.

Jailbreaking

Attempting to bypass AI safety measures through clever prompting. Generally against terms of service and can result in account restrictions.

Bias

Systematic unfairness in AI outputs, often reflecting biases in training data. AI can perpetuate stereotypes or treat groups differently.

Guardrails

Built-in limits on what an AI will do. These prevent the AI from helping with harmful tasks or generating inappropriate content.

Infrastructure

GPU (Graphics Processing Unit)

The hardware that AI runs on. GPUs are much faster than regular CPUs for AI tasks because they can do many calculations simultaneously.

Cloud AI

AI services running on remote servers. You access them over the internet—no special hardware needed.

Local AI / On-device AI

AI running on your own computer or phone. Requires appropriate hardware but offers privacy and no internet dependency.

Quantization

Reducing the precision of model numbers to make them smaller and faster. A quantized model might be 4-bit instead of 16-bit—less accurate but more efficient.

This is how large models can run on consumer hardware.

Putting It Together

You don’t need to memorize all these terms. But having a reference helps when you encounter them in articles, tutorials, or product descriptions.

The AI field moves fast, and new terms emerge regularly. The core concepts—models, training, prompts, tokens—remain relevant even as specific implementations change.

Understanding the vocabulary helps you make better decisions about which tools to use, what their limitations are, and how to get the most from them.

When in doubt about a term, come back here—or just ask your favorite AI to explain it.