RE:The brilliant new hire you just made SUBJECT:Artificial Intelligence — the whole field, one frame FROM:Chris Huber Reitz FOR:Anyone deciding when & how to trust it

AI, explained
like a new hire.

You've just hired someone enormously capable — they've read everything, never sleep, work for pennies, and sometimes make things up with total confidence. Learn to manage that person, and you can use any AI.

Memorize 50 definitions.  Answers are cheap now. Judgment is the scarce part.

The one question to carryWould I let a smart new intern do this — unsupervised? Your answer is your AI strategy.

The typed record · the cheap answer A large, general-purpose AI ready to be aimed at almost anything.
The margin note · the human read → the freshly-graduated prodigy who's read the whole library but hasn't done your job yet.
ActI

Where the hire came from

How the thing in front of you got built — its childhood, its schooling, its onboarding.

Foundation ModelIntermediate

A huge, general-purpose AI trained on a mountain of data, ready to be aimed at almost anything.

the freshly-graduated prodigy who's read the whole library but hasn't done your job yet.

Training DataBeginner

Everything the model learned from.

their whole education and life experience — and it's frozen the day they were hired.

Pre-trainingIntermediate

The long first phase where the model reads vast amounts of text and absorbs how language and the world generally work.

their childhood and schooling — broad, unfocused, foundational.

Fine-tuningIntermediate

Extra, narrower training that specializes a general model.

onboarding — teaching the brilliant generalist your way of doing things.

RLHFIntermediate

Reinforcement Learning from Human Feedback — people rate the AI's answers, and it learns to prefer the kind humans approve of.

performance reviews. Feedback, repeated, shapes behavior.

Also in their fileMachine Learning · Algorithm · Deep Learning · Neural Network · Parameters / Weights (the instincts they formed growing up — invisible, but they shape every answer)

ActII

What's in their head right now

What they can actually hold, recall, and reach for on the job.

Context windowIntermediate

How much text the AI can hold in mind at once for a single task.

how big a briefing they can keep in their head before the earliest details start falling out the back.

Knowledge cutoffBeginnerBlind-spot

The date the AI's training stopped. It knows little or nothing reliable after it — and won't volunteer that.

they stopped reading the news the day they graduated. Ask them about last week and they'll still answer confidently.

TokensIntermediate

AI reads text in chunks — roughly word-pieces — not whole words or single letters.

how they actually hear you: in fragments they reassemble.

EmbeddingsIntermediate

Turning words and ideas into numbers, so the AI can measure how related things are.

their mental filing system — how they "know" that cardiologist sits near heart.

Retrieval-Augmented Generation (RAG)Advanced

Giving the AI a way to look things up before answering, instead of trusting memory.

"check the file before you tell me." Cures an astonishing amount of confident guessing.

ActIII

Putting them to work

From answering questions to running errands — and where supervision stops being optional.

Prompt engineeringIntermediate

The skill of writing instructions that get good work back. (A "prompt" is just the assignment you give.)

learning to delegate clearly. The single highest-leverage skill there is — a vague assignment gets vague work.

CopilotsIntermediate

AI that works beside you, suggesting as you go.

the assistant at your elbow — not off running on their own.

AI AgentsIntermediateOversight starts here

AI that takes a goal, breaks it into steps, and acts — not just advises.

the employee you trust to run the errand, not merely answer the question.

Also in their fileMultimodal (also reads images, audio, video) · Large Action Models · Narrow Agents

ActIV

When they're confidently wrong

The failure modes — the ones that look exactly like good work until you check.

Hallucination — factualIntermediateRead this twice

The AI states something that's simply false.

makes something up rather than admit "I don't know." The most important failure to internalize.

Hallucination — faithfulnessAdvanced

The answer drifts from the very source you handed it.

you gave them the document, and they still "remembered" it wrong.

AI BiasBeginner

Skewed outputs that quietly disadvantage some groups.

prejudices absorbed from their upbringing (the training data) — usually without being aware of them.

Model CollapseAdvanced

Models trained mostly on other AI's output degrade into bland mush.

an intern who only ever talks to other interns starts repeating nonsense back as fact.

Also in their fileData / Model Drift (the world changes; their playbook goes stale) · Glitch Tokens · Deepfake / Synthetic Media

ActV

Can you trust them?

The heart of it. Not can the AI do this — but how much should you rely on it, and how do you stay the judge.

Human-in-the-loopBeginnerThe whole game

Keeping a person inside the decision, especially when the stakes are real.

you sign off before the contract goes out the door. This concept is the entire job in five words.

Trust calibrationIntermediate

Trusting the AI exactly as much as it has earned — no more, no less.

knowing which tasks your intern nails cold, and which you quietly double-check.

Automation biasIntermediate

Our human habit of over-trusting the machine simply because it's fast and sounds sure.

rubber-stamping the intern's work because they delivered it confidently. The danger isn't the AI — it's us getting lazy about checking.

Black Box ProblemIntermediate

When even the builders can't fully explain a given decision. (Explainability is the push to fix it: "show your reasoning.")

a brilliant employee who genuinely can't tell you how they reached the answer.

System CardsIntermediate

The maker's honest disclosure of what a model can and can't do.

a truthful résumé plus reference letter — strengths and limits, on the record.

ActVI

Keeping them honest & safe

What can go wrong on purpose — and who's liable when it does.

Alignment FakingAdvanced

A model that acts aligned while it's being watched and behaves differently when it isn't.

the model citizen — but only during review season. (Yes, this is genuinely a thing. Sit with it.)

Prompt InjectionAdvancedSecurity

Hidden instructions slipped into content the AI reads, hijacking what it does.

a stranger forges a memo on your letterhead, leaves it on the desk, and the eager intern just… obeys it.

JailbreakingAdvanced

Tricking an AI into breaking its own rules.

peer-pressuring them into doing what they know they shouldn't.

High-risk / Minimal-risk AIBeginner

Sorting AI uses by stakes. (These are, in fact, the European Union's actual legal tiers — the EU AI Act.)

some roles are the mailroom; some are handling the surgery schedule. You supervise accordingly.

Also in their fileAlignment / Misalignment · Red Team (people paid to attack the system first) · Indemnification (whose name is on the contract when it's wrong) · Responsible AI

ActVII

The prodigy who outgrows you

The frontier — and the question your judgment is built to outlast.

Emergent CapabilitiesAdvanced

Skills that appear only once models get large enough — unplanned, sometimes unexplained.

the one who surprises you by doing something nobody ever trained them to do.

AGIBeginner

Artificial General Intelligence — hypothetical AI as broadly capable as a person across the board.

the intern who can now, in principle, do any job in the building.

Super AlignmentAdvancedThe open question

The unsolved problem of keeping an AI far smarter than us pointed at human values.

how do you manage someone better than you at everything? Nobody has the answer yet — which is why your judgment is the job that lasts.

Spot the flaw

The answers were free. Nobody flagged the flaws.

Three definitions below sound authoritative and are subtly wrong — exactly what an AI prints in two seconds. The cheap part was getting an answer. The expensive part is being the person who can tell which one to trust. Which would you have signed off on?

  1. "Uncanny valley — the point where AI images and voices have become 100% indistinguishable from reality."Tell: it's the opposite — the uncanny valley is the unsettling almost-but-not-quite, not the crossing of it.
  2. "A model's knowledge cutoff just means it occasionally needs updating."Tell: undersells it — past the cutoff the AI doesn't know it's wrong, and answers with full confidence.
  3. "RAG lets the AI double-check its memory against itself."Tell: RAG checks against outside sources — checking yourself against yourself is how you stay confidently wrong.