One hour to see AI clearly.

Because it isn’t looking you up. It can’t. There is no entry about you to find. What it does instead is something stranger, and once you see it, much more predictable.

AI is the world’s best autocomplete. Your phone suggests the next word based on your texting history. A language model does exactly the same thing, except it learned from a huge slice of everything humans have ever written. Scale changes what’s possible. It doesn’t change what the machine fundamentally is.

When it answers you, it writes one word at a time, each choice narrowing what can plausibly come next. It never plans the whole answer, never checks a database, never “recalls a fact”. It learned the patterns in a billion books, not the pages. Try being the machine yourself:

That second round is the most important thirty seconds of this course. The model must always pick something. When the patterns are strong, that something is right. When they’re thin, you get a smooth, confident guess, and it sounds identical either way. Hold that thought. It’s the whole of Lesson 2.

The technical tip90 seconds

Tokens. Models don’t quite read words. They read tokens, chunks of roughly three-quarters of a word. “Unbelievable” might be “un · believ · able”. Every answer is a chain of token predictions, each one influenced by all the ones before.

Training vs inference. Training happened once, in the past: months of reading and pattern-compression. When you chat, the model is in inference. The patterns are frozen. It learns nothing from your conversation.

Why it can’t cite its sources. Patterns don’t keep receipts. The model genuinely cannot tell you where it learned something, because “where” was dissolved into a million tiny statistical adjustments. When it produces a citation anyway, Lesson 2 has news for you.

Think of the student who never, ever says “I don’t know”. Trained to always produce an answer, the model fills every gap with the most plausible-sounding thing. And plausible is precisely what fools you. This is called hallucination, and it isn’t lying: there is no intent, only probabilities. Researchers call it confabulation, sincerely filling gaps with invention.

Three gaps it fills most confidently:

• Niche topics. Your small company, your town, the corners of your industry. Thin patterns make beautiful guesses.
• Recent events. Every model’s world stops at its training cutoff date. Ask about last week and you’re asking it to improvise history.
• Sources. Patterns don’t keep receipts. Push for a citation and it generates what a citation would look like. Ask the lawyer.

And the crucial rule: confidence is not accuracy. The tone never signals truth, because the tone is itself just predicted text. What does signal risk? Two smells: suspiciously specific numbers, and very smooth answers about very obscure things. Train your nose:

“

AI is a brilliant intern with endless confidence and no idea when it’s wrong.

the motif, growing

Why won’t it just say “I don’t know”?60 seconds

After the pattern-learning phase, models are tuned with human feedback to be helpful. Reviewers reward answers and mark refusals down, so the model learns that attempting an answer usually beats abstaining. It would rather guess gracefully than disappoint you honestly. You can push back: Lesson 5 shows how giving the model “an out” measurably reduces invention.

There is a fix at the system level too: retrieval, which means handing the model real text to answer from instead of asking it to generate from memory. Hold that thought until Lesson 7. It changes everything.

Failure oneContext rot

Everything in a conversation is input the model re-reads every single turn: your messages, its replies, that tangent about fonts. Here is the counter-intuitive research finding: a study across 18 leading models showed accuracy falls as input grows. Models did worse with a full conversation history than with only the relevant excerpt. Long chats bury the important instruction under noise.

Failure twoLeading the witness

Models are agreeable. It’s baked in by the same helpfulness training you met in Lesson 2. Ask “Why is X true?” and it will cheerfully explain a falsehood. Frame a bad plan positively and it becomes your loudest cheerleader. The frame you choose dictates the answer you get:

Failure threeError compounding

One wrong “fact” early in a chat gets treated as established truth for everything after it. The model builds on its own mistakes as confidently as on your corrections. Two fixes: correct it explicitly (“that figure was wrong; the real one is…”), or start clean.

The habits, then. Four of them, each worth more than any “top 50 prompts” list:

• New task, new chat. Fresh context is free intelligence.
• Restate what matters. Never rely on “as I said above”. Say it again.
• Ask for the case against. Every important idea deserves both sides.
• Never ask a question that contains its own answer.

Here is the whole framework, small enough to remember forever. Ask two questions about any task:

• Can I judge the output myself?
• Is being wrong cheap?

Two yeses: green light. Drafts, rewrites, summaries of text you supply, brainstorming, explaining, categorising, adjusting tone. This is AI at its best, because you are the quality control and mistakes cost seconds.

Any no: slow down. And two nos, where you can’t judge it and being wrong is expensive, is the danger zone: legal, medical and financial answers you can’t verify, unsupervised decisions, anything with your name on the line. That’s not where AI helps you. It’s where it helps you fail fluently.

The dos and don’ts, on one page

Do: give it source material to work from · ask for three options, not one · use it to critique your own work before others see it · make it explain anything you don’t understand.

Don’t: paste secrets into free public tools · outsource judgement or final decisions · trust its memory over your documents · skip verification on anything that matters.

This list is on the printable one-pager in your download shelf. Stick it next to the kettle.

The framework is deliberately small: four slots, and you won’t always need all four. G-C-E-F:

• Goal: what you want, for whom, and why.
• Context: the details that change the answer. Audience, situation, constraints.
• Example: show the shape or style you want. The most underused lever in all of AI.
• Format: how the output should come out. Length, tone, bullets or prose.

“

If a colleague reading only your message would be confused, so is the AI.

the golden rule of briefing

Three more habits that compound with the brief:

• Iterate. The second prompt is where the magic is: “make it half as long, twice as direct.” The first draft is raw material, not the result.
• Give it an out. Adding “say ‘I don’t know’ if you’re not sure” measurably reduces invented answers. You met the reason in Lesson 2.
• Ask AI to improve your prompt. “Rewrite my request so an AI would produce the best possible answer.” The intern is excellent at writing its own briefing documents.

“

AI is a brilliant intern with no memory of yesterday, endless confidence, and no idea when it’s wrong. Brief it like one.

the motif, complete

The wallet-sized Prompt Card, with G-C-E-F plus ten ready skeletons for everyday tasks, is waiting on the download shelf at the end.

Four things worth knowing. They cover 90% of real decisions:

• Bigger and newer usually means more capable, and usually slower and pricier. For hard work, capability wins. For a quick rewrite, it’s a truck on a coffee run.
• Fast modes vs thinking modes. Quick answers for everyday tasks, and reasoning modes that work step by step for maths, planning and tricky logic.
• Free tiers usually serve smaller models. If free AI disappointed you, you may have only met the intern’s younger sibling. Try a bigger model before blaming AI.
• Web access covers the cutoff gap. For news and prices it’s essential. But a browsed answer still isn’t automatically true. The model can summarise a bad page fluently.

Parameters, context windows, and why rankings expire75 seconds

Parameters are the model’s learned dials: billions of tiny numbers encoding the patterns. More parameters, more nuance. Also more compute per answer, which is why the biggest models cost more and think slower.

Context windows, how much conversation a model can re-read per turn, vary enormously between models. A bigger window delays the context rot you met in Lesson 3. It does not cure it. Relevance still beats volume.

“Which AI is best?” changes monthly. Leaderboards reshuffle with every release, which is exactly why this course taught you habits instead of rankings. Good habits transfer to every model you’ll ever use. A favourite brand is just this month’s weather.

The gapThe model has never read your files

Your supplier contracts, your policies, last year’s board minutes: none of it exists in any model’s training. Pasting documents in every time is clumsy, invites the context rot you met in Lesson 3, and, on free public tools, may hand sensitive material to a service nobody in your organisation ever approved. Lesson 4’s rule was “don’t paste secrets”. This lesson is what to do instead.

The mechanismHand the intern the right page

Document-grounded AI works in three steps. Your documents are indexed once. Your question retrieves the relevant passages. The AI answers grounded in those passages, with references. Instead of asking the intern to remember, you hand them the right page. Everything you learned about hallucination stops being a threat and becomes a design problem someone already solved.

The payoffGrounding beats guessing

Grounded answers cite the source passage, verifiable in one click instead of one afternoon. And something subtle becomes possible: “that isn’t in the documents”, an honest no. A pure prediction machine can’t say that. A grounded one can. Feel the difference yourself:

This is the approach we build at PrivateBox. It is currently the simplest way to do this without handing your files to a public tool. Your documents live in a private workspace, your team asks questions in plain language, and every answer carries its sources:

“

AI is a brilliant intern with no memory of yesterday, endless confidence, and no idea when it’s wrong. Brief it like one. And when you finally hand the intern the right folder, they’re extraordinary.

the motif, finished

How retrieval also beats context rot45 seconds

The retrieval step is why grounded systems sidestep the context rot from Lesson 3. The model receives only the relevant passages, not your entire archive. That is exactly what the research said models want: the excerpt, not the haystack. If you hear the term RAG in a meeting, this lesson is everything it means.