You're about to understand AI.
Actually understand it.

The type of AI behind Claude, ChatGPT, and most chat tools. A neural network trained on hundreds of billions of tokens of text. The knowledge isn't stored as facts — it's encoded as patterns in billions of numerical weights. When you send a prompt, it flows through those layers and the model predicts, token by token, what should come next.

The most capable AI models available right now. GPT-5, Claude 4, Gemini Ultra — these are frontier models. The term just means "state of the art today." It shifts as new models are released, usually every few months. Frontier models are the ones pushing the boundary of what AI can do.

What you type to the AI. Before the model reads it, your words are converted into numbers by the tokenizer. The quality of your prompt directly shapes the output — not because the model is judging you, but because more context produces better probability distributions over what to say next.

Every token in your conversation — your prompts and the model's responses — accumulates in the context window. The model can only attend to what's inside it. As it fills, the model's ability to focus on the right parts degrades. Earlier instructions get diluted. Responses start to drift. Very long conversations break down for this reason — the model isn't forgetting, it's losing focus.

The unit the tokenizer uses to break text into numbers. Roughly 4 characters or ¾ of a word. "Hello" is 1 token. "tokenization" might be 3. Your entire prompt becomes a sequence of numbers — each token mapped to a unique integer — before the model ever sees it. API costs and context limits are both measured in tokens.

The specific version of an AI you're using. Claude Sonnet, GPT-4o, Gemini Pro — each is a different neural network with different training, different parameter counts, different strengths and costs. Choosing the right model for a task matters: a smaller, faster model is often better for simple tasks; a frontier model for complex reasoning.

A way for software to talk to other software over the internet. When you use Claude Code or Cursor, they're sending your prompts to the AI's API and getting responses back. You can call it directly too — that's how you build AI-powered apps. You send a request with your prompt, you get a response with the generated text.

A text-based interface for your computer. Instead of clicking icons, you type commands. Looks intimidating, isn't. It's just a more direct way to control your machine. Claude Code and Codex CLI run here — you type a prompt, the agent reads your files, makes changes, runs commands, and reports back.

The workspace where developers write code. Think of it as a very powerful text editor with tools built in — file explorer, terminal, error highlighting, version control. Cursor and Kiro are AI-native IDEs: the AI isn't a plugin added on top, it's built into the core of how the tool works.

An AI that can take actions, not just answer questions. A chat model responds to your message. An agent reads your files, runs commands, browses the web, edits code, and works through multi-step tasks on its own — checking its own work along the way. Claude Code and Kiro run as agents.

A technique where the AI looks up relevant information before generating a response, rather than relying only on what it learned during training. The retrieval step finds relevant documents or data, injects them into the context window, and the model generates a response grounded in that fresh information. Perplexity uses this for web search.

Continuing to train a model on a specific dataset after its initial training, to make it better at a particular task or domain. A general model fine-tuned on medical records gets better at clinical language. Fine-tuned on customer support transcripts, it gets better at support. The base weights stay, but the specialization gets baked in.