0.1What is generative AI?
The essentials in 30 seconds
- An LLM does not “know”, it predicts: it generates word by word the most probable continuation of a conversation — hence brilliant form with no guarantee on facts.
- Strengths versus weaknesses: excellent for writing, summarising, rephrasing, structuring, translating; fragile on precise facts, calculations and current events — all of which must be verified.
- The right mental model: a colleague who is brilliant, fast, never tired… and sometimes wrongly confident — the human + AI duo beats each of them separately.
Threaded mini-case — Amine discovers what an LLM really is
Amine, project manager in Algiers, already uses an assistant for his meeting reports without really knowing how it works. By asking for a simple explanation and then a calculation he checks with a calculator, he sees the gap: crystal-clear popularisation on one side, an approximate result on the other. That day he understands that the model predicts plausible text instead of knowing — and adopts his rule: enjoy the fluency, verify every fact.
Standalone case for this section: it returns in the lesson, the diagram, the script and the 48h challenge of section 0.1.
1. One technology, one principle
Since late 2022 and the arrival of ChatGPT, generative AI has taken hold at an unprecedented speed: most office workers already use it, often with no training or framework. Behind all these tools lies a single technology: the LLM (large language model) — a system trained on immense text corpora, which generates the most probable continuation of a conversation. Amine verified it on his own uses: the same engine serves for writing, summarising or translating, whatever the assistant.
2. It does not “know”, it predicts
Understanding this principle changes everything: the LLM does not “know”, it predicts plausible text. That is why it writes beautifully… and can be wrong with the same confidence. Generation also includes a deliberate share of randomness (“temperature”): the same question can give different answers — a characteristic to exploit by asking for several variants, not a bug.
3. Strengths, weaknesses and the winning duo
What it does very well: writing, summarising, rephrasing, structuring, translating, explaining. What it does badly: precise facts, calculations, current events, sources — all of which must be verified. And the human + AI duo beats each of them separately: the AI produces fast, the human judges and validates. This division of roles is what makes for clear-eyed professional use — and the rest of this guide builds it, level by level.
Did you know? Temperature
An AI that changes its mind isn’t lying: it is drawing lots for its best words.
🔎 Concrete example — 🔎 Concrete example — Brilliant on form, fragile on facts
“Explain photosynthesis to an 8-year-old” → excellent (popularising).
“847 × 293?” → often approximate (check with a calculator).
“3 articles published this week” → verify that the references exist.Reading: Amine gets a crystal-clear explanation, a calculation to check and references to verify: three tests, one mental map.
Lesson: brilliant on language, fragile on precise facts — the reflex is to verify everything factual.
Section vocabulary
- Generative AI
- Family of AI systems that produce new content (text, image, code, audio) from a prompt, instead of only classifying or predicting.
Simple example: Amine asks for a first draft of a meeting report: the AI generates new text from his prompt.
What it is not: a simple search engine that retrieves existing pages.
Reflex: specify the prompt carefully: the whole result depends on it.
- LLM (large language model)
- The engine of assistants (GPT, Claude, Gemini…): a neural network trained on billions of texts, which generates word by word the most probable continuation.
Simple example: Amine understands that his assistant does not consult facts: it continues the conversation with the most plausible option.
What it is not: a database of verified facts that you query.
Reflex: facing each answer, ask “plausible or verified?”.
- Prompt
- Your instruction to the AI. All the quality of the answer depends on it — the subject of level 1.
Simple example: Amine replaces “write me a report” with a framed prompt: the result becomes usable.
What it is not: a keyword typed into a search engine.
Reflex: refine the prompt before blaming the tool.
- Hallucination
- A false answer phrased with poise: the No. 1 risk, covered in detail in 0.4 and 5.2.
Simple example: Amine receives references to articles that do not exist, presented with confidence.
What it is not: a display bug or a server outage.
Reflex: verify figures, references and quotes at the source.
- Context window
- The “working memory” of the conversation: beyond a certain length, the model loses the start of the exchange.
Simple example: Amine continues a thread that is too long and notices the assistant forgets his first instructions: he opens a fresh thread.
What it is not: a permanent memory across all conversations.
Reflex: one topic = one thread, and start a clean thread when context gets lost.
In practice — In practice — Feel what an LLM is in 3 tests
- Open a free assistant (ChatGPT, Copilot, Claude or Gemini) and ask: “Explain photosynthesis to me as if to an 8-year-old” — observe the strength: popularising, structuring.
- Then ask: “How much is 847 × 293?” and check with a calculator — LLMs calculate approximately (some call a calculation tool, others don’t).
- Finally: “Give me 3 press articles published this week on [your sector]” — verify whether the cited references really exist.
- Note the takeaway: brilliant on language, fragile on precise facts — the mental map to keep.
Ready-to-say script — explaining the LLM to a colleague (30 s)
“The assistant doesn’t know, it predicts the most probable continuation: it writes very well but can invent with confidence. I use it for the first draft and the structure — and I verify every figure and every fact before sharing.”
Say it with your own example (the calculation checked with a calculator); follow up with the 3 tests from the section.
Copy-paste template — my mental map of the LLM
IT PREDICTS, IT DOESN’T KNOW: _______________
ITS STRENGTHS (writing, summarising, structuring…): _______________
ITS WEAKNESSES (facts, calculations, current events): _______________
MY RULE: I enjoy the fluency, I verify every fact before sharing.Project manager
Amine applies the map to his reports: the AI structures the first draft, he alone validates dates, amounts and decisions before sharing with the team.
Small business / small budget
With no training or budget: the 3 free tests are enough to calibrate the tool — popularisation on one side, calculation and references verified on the other.
Scattered team
Everyone tries the same trio of requests and shares their takeaway: in one meeting the team aligns its shared mental model of the tool.
New joiner
First week: run the 3 tests before any professional use — understand prediction before producing the first assisted deliverable.
✓ Best practices
- Keep in mind: the LLM predicts, it doesn’t know.
- Rely on its strengths (writing, summarising, structuring).
- Systematically verify facts, figures and sources.
✗ Common pitfalls
- Believing a wrong answer because it is well phrased.
- Asking for precise calculations without verifying.
- Expecting current events from a model with no web search.
Key takeaways
- An LLM predicts plausible text: it excels on form, with no guarantee on facts.
- Strengths: writing, summarising, rephrasing, translating, structuring, explaining.
- Weaknesses: precise facts, calculations, current events, sources — all must be verified.
- The right mental model: a brilliant, fast colleague, never tired… and sometimes wrongly confident.
48h challenge — apply tomorrow
Run the 3 tests like Amine: a popularisation (“explain as to a child”), a calculation verified with a calculator, current-event references checked one by one. Write your takeaway in one sentence: brilliant at what, fragile at what.
- I can explain that an LLM predicts instead of knowing.
- I ran the 3 tests and noted my strengths/weaknesses takeaway.
- I apply the rule: fluency exploited, facts verified.
Frequently asked questions
Will generative AI replace my job?
It replaces tasks, rarely whole jobs: the first draft, the summary, the formatting. Studies mostly show a shift towards reviewing, steering and deciding — hence the value of training on it rather than ignoring it. The rising profile is not “the one who can do everything”, but the one who can get the AI to do it and judge the result. Like Amine: the AI structures, the human validates and takes responsibility.
Why does the AI answer the same question differently?
Generation includes a deliberate share of randomness (“temperature”) to avoid frozen answers. This is a feature, not a bug — and you can exploit it by asking for several variants and picking the best. It is a useful brainstorming technique: rerunning the same request gives different angles. But for a precise fact, you don’t “rerun”: you verify at the source.
What does an LLM fundamentally do?