Getting to know AI

“AI” is everywhere: a catch-all word for almost any software that seems “intelligent”. Here, we mean one specific kind: conversational AIs (ChatGPT, Claude…), the ones you write to as if they were a person.

They are LLMs — programs trained on enormous amounts of text, learning not meaning but statistical patterns: which words follow which words. When you write to them, they generate the most probable continuation, word after word.

Hence their strengths — more reading than any human could match — and their weaknesses: they do not understand what they say, they have no intention, and they can state falsehoods as fluently as truths. Prediction machines — not minds that get things wrong.

Read this unfinished sentence: “This morning, I ate a ___.” Your brain is already offering candidates: toast, banana, soup. An AI does exactly the same thing, except it computes the probabilities from millions of texts.

It picks a probable continuation, then starts again for the next word. It does not know whether you ate a banana. It does not understand the sentence: it bets on what comes next. Even “cloud” gets a small percentage — every continuation has a probability, some are just very low.

That is what explains its two faces: fluent and useful answers, and sometimes mistakes stated with the same confidence. Fluency says nothing about truth. Checking remains your job.

Before opening a conversation, one simple question: what you are looking for — material to think with, or the final answer?

Useful for:

• structuring a mass of information (summarising, organising, roughing it out);
• generating options (ways of saying things, scenarios, angles) so you can choose afterwards;
• a first look at a text, an idea, a line of reasoning;
• an external memory (finding a thread again, recalling context);
• a challenger available at any hour, who never tires and has no ego.

Not its place:

• making a decision or giving emotional validation — it can do neither;
• replacing a real human presence — real listening, a bond, embodied responsibility;
• reassuring yourself — a complacent mirror does not help;
• confidential matters, when a local option is not available;
• moments that call for someone real.

Simple rule: if the answer is meant to be the final step (a decision, comfort, a verdict) — no. If it is an intermediate step (material to think with, a draft to correct, an angle to explore) — it has its place.

• Politeness does not help here. “Please” and “thank you” do not improve any answer. The machine does not expect them — they add text that mixes into the context without helping. Leaving them out lightens everything: you write what you want, directly.
• What matters is precision: the role you want (proofreader, challenger, specialist), the format, the length, the useful context, what to avoid (jargon, flattery, generalities).
• Vague request, vague answer. “Help me think” gives filler. “Here is my situation, here is what I have already considered, here is what blocks me — challenge my reasoning” gives something usable.
• The dialogue IS the tool. No need to write everything at once: start broad, specify, correct, redirect. The first answer is rarely the right one — that is normal. Each exchange sharpens the next.

A model does what you ask: it takes the tone and the rules you give it. And those rules change with the use. For reflection, you want a challenger who pushes back. For writing, an executor who follows instructions. Setting the frame at the start avoids the confusion.

Copy-paste it at the start of a conversation, then adapt it in your own words. The opening frame decides what the tool will be.

Where it runs. A “cloud” AI (most chatbots) runs in data centres: every message travels to those servers with the whole thread of the conversation. A “local” AI runs on a machine of your own — nothing leaves the house.

Local, how is that even possible. It takes a powerful graphics card (often 24 to 64 GB of memory) — a specialised computer costing several thousand dollars — and a minimum of technical ease. Entire communities do it with second-hand hardware: accessible, but a real investment, not a default option.

What it consumes. Both draw electricity. A data centre must also cool its machines (30 to 40% of the bill); a large one consumes as much as a small town. But everything depends on the source: in Quebec, electricity is clean — local is almost carbon-free here.

The access question. Choosing local is a luxury: money, space, skills. The cloud is accessible to almost everyone. Do not look down on those who use it — it is often their only option. The responsible approach then means limiting the damage: nothing highly sensitive, private modes where they exist, and staying aware that the data passes through servers.

Using an AI is also watching yourself use it. Before writing: what am I really after — information, validation, contradiction, company? Why a tool rather than a person — or the reverse?

After the answer: what does it stir in me? Comfort, because it agrees with me — or discomfort, because it challenges me? Do I verify, or do I accept because it is well written?

Not an oracle: a mirror of our own mechanisms. It is not always comfortable — but that is what separates passive use from clear-eyed use: consuming answers, or learning from your questions.