Under some sun umbrella, on some beach in Italy, there’s a man in a white vest, mirrored sunglasses and a newspaper hat. He knows everything about everything, and explains it to anyone who passes in front of his sun lounger.
I’ve been telling his story all summer and never gave him a name. It’s Rocco!
And Rocco isn’t new this year. He’s already been national football manager, spread expert, epidemiologist, prime minister and White House relations consultant. Every summer Italy graduates in something, and he’s always in the first session. This year it’s AI’s turn.
With one difference that counts. Getting the national team formation wrong all summer has no consequences for anyone. This year’s words, on the other hand, come back to the office in September — and there they start driving real decisions.
Rocco isn’t unpleasant and he isn’t stupid. He arrived at the beach carrying six months of noise. He’d read things on social media, watched reports on the news, had articles forwarded to him in group chats. One day AI replaces everyone, the next day it’s a bubble about to burst. One day it’s extremely dangerous, the next it’s just statistics with a PR department. Strong, conflicting signals — almost all delivered with exactly the same confidence.
With that material you inevitably build a position, because staying inside the noise without one is exhausting. And confidence is the most human shortcut there is: when the voices contradict each other, whoever picks one and says it loudly stops feeling at the mercy of things.
That’s who Rocco is. Not someone who misunderstood. Someone who put things in order as best he could, with what he had.
We all had a corner of Rocco this summer, on some word or other. Me first, on a couple of these thirteen, before going deep enough to write about them.
So here’s the toolkit. Thirteen words, three lines each: what gets said, what doesn’t get said, and the question to keep in your pocket.
Each of these thirteen had an article of its own this summer. If after three lines you want more, you’ll find the full version at the bottom.
The glossary
Prompt engineering
- What gets said: there’s a perfect prompt, you just need the right words.
- What doesn’t get said: there are no spells. There’s saying what you want — and where you don’t say it, the gap gets filled randomly.
- The right question: which two or three details, if I give them, actually change the answer?
Agents
- What gets said: an autonomous robot that does everything for you.
- What doesn’t get said: it’s a prompt with tools and a margin of autonomy. Going from prompt to agent, you don’t just gain capability — you gain consequences.
- The right question: what can it touch, and how much do I let it decide? And if I can’t answer, am I delegating or am I gambling?
Loop
- What gets said: AI that corrects itself and improves with every round, while you sunbathe.
- What doesn’t get said: it only works if something whispers “warmer.” A clear measure and the next round is better. A vague measure and the next round is just different.
- The right question: who whispers “warmer,” and how much does it cost me each time?
Training and cut-off
- What gets said: it learns as I use it, it updates itself.
- What doesn’t get said: it’s a photograph frozen at a date. Blind forward in time, and blind about you. Where it doesn’t know, it doesn’t stay silent — it fills the gap.
- The right question: what does it need to know about me and my work that it can’t know on its own?
Memory
- What gets said: it remembers me, so it knows me.
- What doesn’t get said: it doesn’t remember — it re-reads. Someone keeps a file on you and pastes it back in just before every answer.
- The right question: who keeps that file, what’s written on it, and have I ever written it myself?
System prompt
- What gets said: this AI is polite, that one is blunt — they have a character.
- What doesn’t get said: the character is a note of instructions written by the provider and slipped in at every call. Same model, different notes, opposite characters.
- The right question: who owns the layer between me and the machine, and what is it asking it to do?
GDPR
- What gets said: either you can’t use it because of privacy, or who’s going to read my chats anyway.
- What doesn’t get said: there’s no wall — there’s a chain of people and organisations behaving well. And a long chain never holds at a hundred per cent.
- The right question: how much does it hurt me if it gets out, and how appealing is it to someone else?
Context window
- What gets said: I upload everything about myself and it finally knows everything.
- What doesn’t get said: it isn’t a window — it’s a backpack with a bottom, and the provider has already half-filled it with their own stuff. More is less.
- The right question: what do I put in, in what order, and above all what do I leave out?
Token
- What gets said: a line item on the bill to keep an eye on.
- What doesn’t get said: it’s the way AI reads. Chunks of characters turned into numbers, not letters. That’s why it can’t count the R’s in “strawberry” — and why you pay for thought by weight.
- The right question: am I paying for context that’s useful, or weight I loaded out of laziness?
Hallucination and bias
- What gets said: an occasional bug, the next version will fix it.
- What doesn’t get said: it’s the engine, not the fault. It always produces the most plausible: when that matches the truth we call it a correct answer, when it doesn’t we call it a hallucination. And it has no way of telling you which case you’re in.
- The right question: instead of asking why it got it wrong, how do I verify?
AGI
- What gets said: we’ll get there in two years. No, five. It’s already here and we don’t know it.
- What doesn’t get said: it’s a threshold nobody knows how to draw, produced by a process not even its builders can fully explain. A prediction squared.
- The right question: what is the person giving me this date selling?
Universal basic income
- What gets said: AI will produce abundance, and then they’ll give us an income.
- What doesn’t get said: you don’t make everyone rich by making everything abundant. Scarcity doesn’t disappear — it changes address.
- The right question: what will remain scarce, and who already has it?
AI Act
- What gets said: European bureaucracy that bans, slows down and always arrives late.
- What doesn’t get said: it doesn’t regulate AI — it regulates the radius of action. The bigger and more irreversible the consequence, the more a human who answers for it must remain in the loop.
- The right question: what happens the day it fails with confidence, and who pays?
Three weeks later
Rocco is still under his sun umbrella. Same vest, same sun lounger, tan slightly worse. But something has shifted.
It’s not that he knows many more things than before. He knows thirteen, and thirteen words aren’t a fortune. What’s changed is the way he sits inside the noise. Before, when a headline arrived, he had to decide whether to believe it. Now, when a headline arrives, he has a question to put to it. It’s far less comfortable and it works far better: confidence gave him control in pretend, the questions give him a real piece of it.
Then the interesting thing happens. The words start pointing somewhere else.
As long as you say “AI will replace us” you’re talking about a debate — and debates are comfortable because they ask nothing of anyone. But Rocco reads “agent” and no longer thinks of the science fiction robot: he thinks of the quotes in his office that get redone from scratch every week. He reads that a loop is only as good as its measure and asks himself how he knows today whether a piece of work has gone well — and the honest answer is that he knows it by feel. He reads “who pays the day it fails with confidence” and no longer thinks of Brussels: he thinks of his biggest client.
And something occurs to him that he wouldn’t have noticed in June. In his company AI has already arrived — and no project brought it: people did, one at a time, each with their own personal subscription and their own way of using it. Some use it very well, some don’t use it at all, some use it and prefer not to say so.
He thinks about it for a while, under his umbrella, and arrives where he was supposed to arrive. N individuals in a company who use AI are not a company of N individuals that uses AI.
And he stops there, because he notices something uncomfortable: the thirteen words taught him how to use AI — and told him nothing about how a company uses it. They’re two different problems. The first is literacy, and you can cover it in three weeks under a sun umbrella. The second begins where summer ends.
One thing the glossary does tell him, though — and it’s the most useful of all. The questions at the bottom of the thirteen entries, looked at together, are always the same three. What I wanted to achieve. How I know whether I got there. Who answers if it’s wrong.
On a single person they weigh little: when you get it wrong, the damage is your morning. Inside a process they weigh everything — because a process repeats itself, involves people who weren’t in the room when you designed it, and reaches the client.
Then Rocco looks up from his phone and glances around. Under the other umbrellas, the same summer is playing out — on different professions.
There’s the man in the panama hat and linen shirt who picks his phone up from the waterproof case around his neck every twenty minutes, looks at it, puts it back down. There’s the woman with the snorkelling mask pushed up on her forehead for the past hour — perfect gear, never a dive. There’s the one playing beach tennis as if it’s a final — one more point and that’s it, one more point. And there’s the one who brought a cool box from home, because at the kiosk it costs three times as much.
The same thirteen words. The same three questions. And on each of them they land in a different place.
In June Rocco had an answer for everything. Now he has a good question for everything, and no certainty in a vest.
He hasn’t become an expert. He’s stopped being one.