Why did ChatGPT feel dumb to you?

Level: none yet — this is your first step.

The idea in one picture

Imagine a colleague who has read half the internet — millions of books, articles, and conversations. Enormously well-read. But: they finished reading a while ago, they have never seen your company’s files, and they have no idea what happened yesterday.

That colleague is an AI language model (an “LLM”). ChatGPT, Gemini, and the agent you will meet in this course are all built on one.

What it actually does

A language model does one thing: it predicts what text should come next, based on everything it has read. When you ask “What is the capital of France?”, it does not look up the answer in a database. It continues the conversation the way a well-read person plausibly would — and because it has read so much, the continuation is usually right.

This explains both its superpower and its weakness:

• Superpower: it can draft, summarize, translate, explain, and rephrase almost anything, instantly.
• Weakness: it always answers in a confident, fluent tone — whether or not it is correct. Fluency is not accuracy.

A concrete example

Ask a model: “Summarize this email in three bullet points” and paste the email — you will get an excellent result, because everything needed is right there in front of it.

Ask it: “What did my company decide at last week’s meeting?” without giving it anything — and it will either admit it does not know, or worse, invent a plausible-sounding answer. It was not at the meeting. It has not read your files. It only knows what you show it.

The one rule to remember

The model is brilliant with what you give it, and blind to what you don’t.

Everything else in this course builds on that rule.

The single most important lesson in this course.

The idea in one picture

Picture the model working at a desk. The only things it can see are the papers lying on that desk right now: your messages, the documents you pasted, and its own earlier replies in this conversation.

That desk is called the context. And it has two properties that explain almost every frustrating AI experience you have ever had:

1. If it is not on the desk, it does not exist. The model cannot see your inbox, your files, or your last conversation with it — unless someone puts them on the desk.
2. The desk has limited space. In a very long conversation, the oldest papers quietly slide off the edge. The model does not announce this. It simply no longer sees them.

A concrete example

You spend twenty minutes discussing a contract with an AI. Early on, you mentioned the deadline is March 15. Forty messages later you ask, “So what was the deadline again?” — and it confidently says “March 1.” It is not lying. The paper with “March 15” slid off the desk, and the model filled the gap the way it always does: with something plausible.

Three habits that follow

1. Paste, don’t assume. Never think “it probably knows this document.” Put the document on the desk — paste it or attach it.
2. New topic, new conversation. Starting fresh clears the desk. Old, irrelevant papers cannot confuse the new task.
3. Repeat what matters. In a long conversation, restate the key facts (“Reminder: the deadline is March 15, the budget is 40,000 zł”) — you are putting that paper back on top of the pile.

The idea in one picture

Remember the well-read colleague from Lesson 1? They have one strange trait: they are incapable of staying silent. Ask them anything, and they will always produce an answer that sounds right — because their whole skill is producing text that sounds right.

When the true answer is on the desk (Lesson 2) or deep in what they have read, “sounds right” and “is right” are the same thing. When it isn’t — they fill the gap with something plausible. This is called a hallucination, and it is not a malfunction. It is the machinery working exactly as designed, just without the facts it needed.

A concrete example

Ask an AI for “three scientific studies proving that standing desks improve productivity” and you may receive three beautifully formatted citations — authors, journals, years. Sometimes one or two of them do not exist. The model has read thousands of real citations, so it knows perfectly what a citation looks like. Looking like one is its job.

The trust rule: ufaj, ale sprawdzaj — trust, but verify

You do not need to distrust everything. You need to know which kind of task you gave it:

Safe zone (verify lightly or not at all):

• Summarize this text, translate this email, rewrite this paragraph more politely
• Brainstorm ideas, draft a first version, explain a concept
• Anything where the source material is on the desk and you can see the result is faithful

Check zone (verify before you act):

• Facts, numbers, dates, names, prices, legal or medical statements
• Anything you will forward to a client, sign, pay, or decide based on
• Citations, links, references — always click them

The habit

Before acting on an AI answer, ask yourself one question: “Did it work from material I gave it, or from memory?” From your material — relax. From memory — verify the parts that matter.

The idea in one picture

“AI” is not one thing. The same chat window can run on different models — think of them as different engines available for the same car. Some are small, fast, and cheap. Some are large, slower, and much smarter. The car looks identical from the driver’s seat; the difference appears when you climb a hill.

Why this matters to you personally

If you have used the free version of ChatGPT or Gemini, you have mostly been driving the cheap engine. Companies give the small model away for free and sell the strong one. This single fact explains a huge share of “I tried AI and it was dumb” experiences: it was not AI that was weak — it was the particular engine you happened to get.

A concrete example

Give a small model a messy, 3-page supplier contract and ask, “What are the risks here for us?” — you may get a shallow, generic list. Give the same document with the same question to a top-tier model, and it will point at the odd payment clause in section 7 and the missing termination notice period. Same car, same road, different engine.

The three trade-offs

	Small model	Large model
Speed	instant	a bit slower
Cost	cheap / free	more expensive
Depth	fine for simple tasks	needed for hard ones

Neither is “better.” You match the engine to the hill:

• Small engine: quick rewording, short summaries, simple questions, routine drafts.
• Large engine: analysis, tricky documents, multi-step reasoning, anything with consequences.

The habit

When an answer disappoints you, your first thought should no longer be “AI can’t do this.” It should be: “Was this the right engine for that hill?” Often the fix is not a cleverer question — it is a stronger model, or simply asking again with the material properly on the desk (Lesson 2).

The idea in one picture

There is no secret language for talking to AI. Everything you have read about “prompt engineering magic” boils down to how you would brief a capable new employee on their first day. You would not say “do marketing.” You would say what, for whom, in what form, by when — and you would give feedback on the first draft.

Four habits. That is the whole lesson.

Habit 1 — Give the context

The model only sees the desk (Lesson 2). So put the situation on it.

• Weak: “Write an email to the supplier.”
• Strong: “Write an email to our packaging supplier. They delivered 2 weeks late and we want a 10% discount on the next order, but the relationship matters — keep it firm and friendly.”

Habit 2 — Say the goal AND the format

Tell it what “done” looks like: length, structure, language, tone.

• Weak: “Tell me about this report.”
• Strong: “Summarize this report in 5 bullet points for a director who has 60 seconds. Plain language, no jargon.”

Habit 3 — Iterate instead of abandoning

This is where most free-tier users quit — and where the real value hides. The first answer is a first draft, not a verdict. Reply to it:

• “Shorter. Half the length.”
• “Good, but point 3 is wrong — we did NOT agree to that. Fix it.”
• “Now the same, but in English, for a client.”

Each correction lands on the desk and sharpens the next attempt. Two or three rounds usually beats an hour of solo writing.

Habit 4 — Paste real material

The single biggest quality upgrade available to you costs nothing: work from your actual text, not from a description of it. Paste the real email, the real table, the real paragraph. “Improve THIS” always beats “write something like…”

The anti-pattern to unlearn

Treating the first answer as final — accepting a mediocre result or walking away disappointed. People who get impressive results are not asking magic questions. They are having short working conversations.