AI Terms from Zero to Pro: The Dictionary Most People Don't Have

Most people using AI today are driving a Formula 1 car with a learner’s permit. They know the steering wheel and the pedal, but they have no idea what the gearbox, differential, or downforce actually do. This video is the missing manual.

Ariel walks through the full dictionary of modern AI terms—not the buzzwords, but the machinery underneath: what a model really is, how tokens work, why attention matters, what makes an agent different from a chatbot, and why your AI keeps “forgetting” things that were in the conversation five minutes ago.

What is actually going on inside an AI

The word “AI” is a label, not a single technology. In the 1950s it meant a program that could check if a number was bigger than ten. In the 1990s it meant Deep Blue beating Kasparov. Today it mostly means large language models (LLMs) that generate text one token at a time.

A model is not magic. It is a giant table of parameters—mathematical weights learned during training. When you type a prompt, the model does one thing and one thing only: next token prediction. It guesses the next word, then the next, then the next. Everything that looks like reasoning, coding, or conversation is just this loop running fast enough to fool you.

Tokens, attention, and context

A token is the smallest unit the model counts. In English one word is roughly one to one-and-a-half tokens. In Hebrew it is often three to five tokens per word because of the grammar and morphology. That matters because you pay per token, and Hebrew prompts are more expensive than they look.

Attention is how the model decides which words matter to each other. When you ask “What is the age of ST?” the model gives more attention to “age” and “ST” than to the rest of the sentence. As the conversation grows, the attention budget gets spread thinner. Past a certain context size the model enters what Ariel calls the dumb zone: it stops tracking details accurately and starts guessing or hallucinating.

Stateless models vs. stateful systems

The model itself has no memory. Every request is stateless. What gives it memory is the harness—the software sitting between you and the model. The harness keeps the conversation history, reads files, manages tools, and resends context so the model feels like it remembers.

That is why the same model can act completely differently in Claude Code, Claude.ai, Cursor, or any other harness. The model provider supplies the brain. The harness supplies the hands, eyes, and notebook.

Agents, tools, and the environment

An agent is a model wrapped in a harness with memory and tools. Tools are functions the harness exposes to the model: read a file, search the web, run a test, query a database, send a message. The model does not call tools directly. It outputs a structured request, the harness parses it, executes the action, and feeds the result back as new input.

The environment is the world the agent lives in: your file system, a Docker sandbox, a cloud VPS, a test database. If you want the agent to be safe, give it a sandbox. If you want it to be useful, give it the right tools and the right context. If you want it to remember tomorrow what you did today, write the important stuff to a file or database. Do not expect the model to remember it for free.

Why AI makes mistakes, lies, and forgets

Two failure modes matter here:

• Factuality errors: the model invents functions, APIs, citations, or version numbers because its parametric knowledge is frozen at training time. The fix is fresh context: documentation, changelogs, code, or search results.
• Faithfulness errors: the model drifts away from your original goal because the attention budget gets diluted across a long conversation. The fix is to compact or restart the session, and to use auto-compact settings so the harness keeps summaries instead of carrying every token forever.

🔥 Roast Corner

Let us be honest: 90% of the people throwing prompts at ChatGPT do not know what a token is. They think the model “understands” them. It does not. It is a statistical parrot that has been trained to sound confident, helpful, and occasionally sycophantic.

The real crime is not that people are ignorant. The real crime is that vendors sell this ignorance as a feature. “Just ask it anything!” No. If you do not know what context window, system prompt, and tool calling are, you are not using AI. You are being used by it. And when it hallucinates a deprecated API or writes a whole file when it should have edited three lines, you blame the model instead of blaming your own workflow.

If you are building with AI and you cannot explain stateless vs. stateful, input vs. output tokens, and why your harness matters, you are not an AI engineer. You are a person who types into a chatbot and hopes.

🤖 AI for Humans

Here is the concrete takeaway: stop treating the model like a person and start treating it like a very fast, very expensive, very confident token-prediction engine.

Before your next project, answer three questions:

1. What is my context budget? Know the context window, estimate your input and output tokens, and keep the relevant files and docs loaded.
2. Where is my memory? Decide whether state lives in the session, in files like claude.md, or in a real database. If you need it tomorrow, write it somewhere durable today.
3. What tools does the agent need? Give the harness the exact tools it needs—file access, search, test runners, APIs—and put dangerous work inside a sandbox.

Get these three right and the same mediocre model will outperform a much larger model in a messy workflow. Get them wrong and even the best frontier model will disappoint you.