AI Glossary

AI is the broadest term — any machine that exhibits intelligent behavior. Machine Learning is a subset of AI where systems learn from data. Deep Learning is a subset of ML that uses neural networks with many layers. Think of it as nested circles: AI contains ML, which contains Deep Learning.

Training (or pre-training) is the initial process where a model learns from a large dataset — this is expensive and time-consuming (weeks or months on hundreds of GPUs). Fine-tuning takes an already-trained model and adapts it to a specific task using a smaller dataset — this is much faster and cheaper.

Tokens are the basic units that AI models process — typically parts of words or whole words. "Tokenization" is roughly 4 characters per token in English. Tokens matter because AI pricing, context limits, and performance are all measured in tokens. Understanding tokenization helps you estimate costs and fit content within context windows.

RAG (Retrieval-Augmented Generation) combines LLMs with external knowledge retrieval. Instead of relying only on what the model learned during training, RAG retrieves relevant documents and includes them in the prompt. This reduces hallucinations, keeps information current, and allows the model to work with private data it wasn't trained on.

Parameters are the learnable weights in a neural network — essentially, the numbers that the model adjusts during training. More parameters generally mean more capacity to learn complex patterns, but also require more memory and computation. GPT-4 is rumored to have over 1 trillion parameters; Llama 3 has 8B to 405B parameter versions.

Prompts are text inputs you give to a model at inference time — the model's weights don't change. Fine-tuning actually modifies the model's internal parameters using your data, creating a customized version. Prompt engineering is faster and cheaper but limited; fine-tuning is more powerful but requires more effort and data.

An AI agent is a system that can take actions autonomously to achieve goals. Unlike a simple chat model that only responds to prompts, agents can use tools (search the web, run code, access databases), make decisions, and work on multi-step tasks without constant human guidance. Examples include AutoGPT, Claude's tool use, and OpenAI's function calling.

RLHF (Reinforcement Learning from Human Feedback) is a training technique that makes AI models more helpful, harmless, and honest. After initial training, humans rank model outputs by preference, and these rankings are used to train a reward model. The AI is then trained to maximize this reward. RLHF is what makes ChatGPT behave like a helpful assistant rather than just predicting text.