“The lizard that jumped from the high iroko tree said that if nobody praised him, he would praise himself.”
Igbo proverb.
The lizard appears to have had admirable self-esteem. Its connection to artificial intelligence will become clearer later. Probably.
At 5:47 p.m., Mireille has no interest in understanding how a large language model works. She does not want to learn what a token is, distinguish a generative model from a predictive one, or donate her Saturday morning to a course called “From Zero to AI Expert”.
Mireille wants to go home.
Unfortunately, one email remains. It is one of those meticulously polite messages whose politeness somehow feels sharpened. It blames her for a delay she did not cause and, because she knows that answering while irritated would probably produce either three unnecessarily hostile paragraphs or a painfully diplomatic surrender, she opens an AI assistant instead.
She explains the situation, removes anything confidential, provides the relevant dates and asks for a firm but professional reply. The first draft is too soft. The second is almost right. She rewrites three phrases, shortens a sentence and presses Send.
It is 5:54 p.m.
Mireille still cannot explain what a transformer is. Her problem, however, is gone.
And perhaps that is exactly how we began learning artificial intelligence backwards.
We turned AI into another school subject
Every major technology eventually acquires its own vocabulary, experts and training industry. Artificial intelligence has followed the pattern enthusiastically. Courses, certificates, masterclasses, prompting frameworks, tutorials and enough videos titled “You’re Using ChatGPT Wrong” now exist to make perfectly competent people wonder whether they have somehow failed an exam they never signed up for.
Some of that education matters enormously. If you are building AI systems, automating sensitive workflows, handling confidential data or making decisions about security and governance, improvisation is not a serious operating model. Understanding limitations, costs, controls and failure modes matters.
Most people, however, are not trying to build the engine. They are trying to get somewhere.
They have a contract whose third paragraph appears to have been written specifically to discourage human comprehension. They have a spreadsheet whose columns are reproducing at an alarming rate. They need to answer a complaint, prepare for an interview, summarise a report, compare two offers, organise scattered notes or understand why their monthly budget becomes mysteriously philosophical around the third week.
Yet we often introduce AI as something people must first study before they are allowed to use it.
That is rather strange.
Nobody required millions of people to understand SMTP before sending their first email. We did not demand database theory before allowing people to use search engines. Even Excel, responsible for countless small workplace crises, is mostly learned in contact with actual problems: one formula, one spreadsheet and one deeply personal encounter with #REF! at a time.
AI may end up being learned the same way.
The real “wow” is usually much less glamorous
AI demonstrations have trained us to expect spectacle. Models generate video, write software, create voices, compose music and occasionally appear ready to reorganise entire companies before breakfast. Against that backdrop, ordinary uses can feel almost disappointingly small.
Yet the real shift often happens far away from the demos.
It happens when a shopkeeper turns photographed sales notes into a usable table. When a student asks for a difficult chapter to be explained differently. When a parent converts school material into revision questions. When a job seeker compares a résumé with a vacancy and notices skills that were buried rather than highlighted. When somebody untangles an administrative letter, understands an invoice or turns thirty lines of midnight thoughts into something coherent.
None of this needs dramatic music. It is simply useful.
And usefulness has always had the unfair advantage of changing more lives than spectacle.
We talk constantly about the power of models, yet the more interesting unit of measurement for ordinary users may be the half hour returned to them. Twenty minutes today, fifteen tomorrow and another ten on Friday eventually stop looking like isolated conveniences. They become postponed tasks finally completed, small frustrations removed and mental space reclaimed.
Documents, deadlines, details. Daily work has always had a remarkable ability to let tiny things consume very large days.
AI becomes interesting when it starts taking a few of them back.
Perhaps the problem should come first
The traditional way to learn software is orderly: study, understand, practise, master. That still works, but conversational AI introduces another route.
You can begin with use.
A person encounters a problem, explains it badly, receives an imperfect answer, corrects it, tries again and gradually learns how to work with the tool.
The order has changed. We do not always learn before doing; sometimes we learn because we did.
That difference matters because curiosity becomes more powerful when attached to an actual frustration. Understanding hallucinations becomes much more interesting after a model has confidently informed you of something that never happened. Data privacy becomes less theoretical when you are about to paste a professional document into a chatbot. The difference between searching for information and asking a model to synthesise it becomes clearer the first time you discover that AI can beautifully explain a fact that does not exist.
Training stops being a waiting room. It becomes an answer.
Marcus Aurelius wrote that what stands in the way can also become part of the way forward. The Stoics were dealing with matters somewhat more serious than badly organised spreadsheets, but the principle travels surprisingly well: the problem is not always what delays learning; sometimes the problem gives learning its direction.
The obsession with the perfect prompt may have made this harder than necessary
Then came prompt engineering.
It is an impressive phrase. It can make communication with AI sound like an occult discipline involving roles, context, seven constraints, three examples, an output format and possibly the current position of Mercury.
In professional systems, carefully designed instructions absolutely matter.
For planning your cousin’s birthday menu, probably less so.
For ordinary use, there is a simpler principle: explain the problem properly.
Here is what I have. Here is what I need. Here are the constraints. Here is what I have already tried. Here is what I do not understand.
And if the first answer is poor, say so. The machine will survive. It is one of the few colleagues you can tell “No, you misunderstood me” at 11:18 p.m. without receiving a meeting invitation from HR the following morning.
Paradoxically, learning to use AI may therefore force us to practise a very human skill: clarifying our own thinking.
A vague request tends to produce vague usefulness, while a clearly framed problem forces us to identify what we actually want. That is not magic; it is articulation.
And occasionally, the act of explaining a problem to a machine helps us understand the problem ourselves.
Imagine a brilliant library with the confidence of a cousin who “knows a guy”
There is, however, an inconvenient detail: AI can be wrong.
Not always modestly.
These systems occasionally possess the fascinating ability to know absolutely nothing with extraordinary confidence.
So perhaps the best metaphor is not an oracle or an all-knowing assistant. Imagine instead a vast library that has somehow acquired a mouth, imperfect memory and the confidence of that relative who begins every answer with, “Actually, I know a guy…”
It can be brilliant. It can also confuse, simplify, extrapolate and invent.
Judgement therefore becomes more important with AI, not less.
You can delegate the first draft of an email without delegating your intention. You can ask for help understanding a contract without mistaking the explanation for legal advice. You can simplify medical terminology without promoting the chatbot to physician. You can analyse a budget without expecting your bank to accept “the model said I could afford it” as an overdraft policy.
Delegating some of the work does not require outsourcing judgement.
That principle may be worth teaching before many advanced prompting techniques.
Try the fifteen-minute test
There is a simple way to discover what AI can actually do for you. It does not require a certificate or a sacrificed weekend.
It requires fifteen minutes and one genuine annoyance.
Take the task you have been postponing for three days: the email you still have not answered, Tuesday’s chaotic meeting notes, the document from which you need to extract the important points, the monthly expenses you approach with the emotional preparation normally reserved for horror films, or the comparison currently distributed across eleven browser tabs.
Describe the result you want, provide only the context that is necessary, then ask for a first attempt.
After that, work on it. Ask why. Ask for something shorter. Ask for clarity. Ask what may be missing. Verify what matters.
Perhaps the tool will be terrible at your particular task. Useful information. Perhaps it will take three exchanges before the answer becomes worthwhile. Also useful information.
Or perhaps somewhere between the second and third response, you will stare at the screen and think: “Wait. I could do this the whole time?”
That is the “wow”. Not in a laboratory, not in the distant future, but inside the annoying task that just became twenty minutes shorter.
So what happens to training?
It comes afterwards.
Not because training is unnecessary, but because it becomes far more valuable once it answers questions you actually have.
Once people genuinely use AI, they begin asking better questions. Why does it invent facts? How should I protect my data? How do I verify a source? When should I use traditional search instead? How can I automate a repetitive process? What information should never be uploaded? How do I provide enough context without exposing too much? How do I tell the difference between a plausible answer and a well-supported one?
At that point, learning AI stops being an abstract project.
Every concept has somewhere to land.
Security corresponds to an actual risk. Privacy is no longer a footnote. Hallucination is no longer a funny technical term. Automation stops meaning “technology of the future” and starts meaning “perhaps I will never manually copy these thirty lines every Friday again”.
That may be the healthiest path into AI: begin simply enough to use it, then learn seriously enough not to become gullible about it.
Perhaps we simply started at the wrong end
Mireille may continue to have no idea what an embedding is.
Nothing catastrophic will happen.
Tomorrow she may use AI to compare two documents. Next week, she may organise meeting notes with it. A month later, she may have learned to anonymise information, demand sources and recognise situations where the model is simply the wrong tool.
Learning will have happened. It will simply have begun with practice.
We have spent a great deal of time asking: “What should I learn before I can use AI?”
Perhaps the more useful question is the reverse: “What is annoying me enough today that I should try solving it differently?”
That spreadsheet. That email. That research task. That repetitive piece of work.
Start there.
The rest will have somewhere to follow.
As for the lizard that fell from the iroko tree, perhaps it understood something before the rest of us: beginning does not always require permission, applause or complete mastery of the theory.
It praised itself.
Mireille closed her laptop.