00 — You Said AI What?

Status: Required on-ramp · 10 minutes · no artifact

Purpose

The part of the program where we admit AI is confusing, most people are winging it, and you are allowed to know nothing yet. No jargon, no quizzes, no judgment. Just the minimum mental model required so the rest of this curriculum does not waste your time.

What this is

One page. Read it. Then decide which problem you are actually solving next and walk into the right section.

What you will be able to decide after this

Whether a given AI claim is worth a second thought.
Which of the six sections you should enter first.
Whether you are ready for Core Judgment or still circling the airport.

Rule

You do not need this page if you already know what a prediction engine is, why fluency is not evidence, and why confident is not correct. If you don't — read it. It's ten minutes.

The short version

What "AI" actually is

A prediction engine trained on large amounts of data. It generates the most statistically likely continuation of the text (or other data) you give it. It does not know, believe, intend, or understand in the way a person does. When the output looks intelligent, that is pattern matching plus scale. When it looks wrong, that is the same mechanism operating without grounding.

Let's kill the expectations first

What it isn't

Not conscious or agentic by default.
Not a reliable source of truth.
Not a replacement for judgment, accountability, or domain knowledge.
Not magic. It is statistics with better marketing and a chat interface.

Confident ≠ correct. That distinction is the entire game.

Take these personally

Myths that need killing

"AI thinks like a human." No. It predicts. Thinking, deciding what matters, and owning the outcome remain yours.
"The AI is biased, so I can't use it." It is biased because the data is. That is a reason to verify and constrain, not a reason to avoid it entirely.
"If the AI says it, it must be true." It produces fluent text. Fluency is not evidence. Treat every important claim as provisional until checked.
"AI will do my whole job." It does fragments of jobs. The people who keep the judgment, the review criteria, and the accountability are the ones who stay valuable.
"I need to understand how the model works inside to use it." You need to know what it is good at, what it fails at, how to constrain it, and how to verify the result. Internal weights are irrelevant for almost every practical use.

The 7% that does the work

Vocab that actually matters — only the terms you will meet repeatedly in the rest of the curriculum.

LLM / model — The prediction engine itself.
Prompt — The input you control. Quality of output is heavily downstream of this.
Hallucination — Confident fabrication. Assume it can happen on any factual claim.
Context window — How much prior text the model can still "see." Exceed it and earlier information is lost or degraded.
RAG — Retrieve relevant documents first, then generate. One of the main ways to reduce pure hallucination.
Agent — A system that can call tools and take multi-step actions. Higher power, higher need for oversight and guardrails.
Token — The unit the model processes. Relevant mainly for cost, length limits, and latency.

Everything else can be introduced when it becomes necessary.

One track. That's it. For now.

How to choose your path

Do not start all six. Core Judgment first — it's required. Then pick one applied pack based on the decision you actually need to make next.

You can move between packs later. Starting in two at once is usually a sign you have not yet decided what problem you are solving.

The rules of the house

Operating principles

Evidence over attendance. You demonstrate capability by producing artifacts, reasoning traces, verification records, and revision. Finishing modules is irrelevant. Submitting work that shows judgment is the only currency.
Capability over product. We teach durable judgment (what to trust, when to reject, how to constrain, how to measure). Specific tools and interfaces are temporary overlays. When the product changes, the judgment should still transfer.
Honest uncertainty. "I do not have enough evidence" is a complete and acceptable answer — in learning and in production decisions. Pretending otherwise is the real failure mode.