What two hours in Bali with Benoît Raphaël taught me about who actually wins in the age of intelligent machines.
Benoît Raphaël sat across from me in Bali.
He built what I consider the best AI newsletter in the world, Génération IA. He once beat a Prix Goncourt winner — Hervé Le Tellier, France’s highest literary honor — in a short-story challenge: AI against pen. So when he tells you the debate everyone is having is a fake debate, stop and listen.
The fake debate: can AI write?
Here is the real question. What story do you have to tell? How will you tell it? And at the end, will you put your signature at the bottom of the page?
Everything that matters about AI sits inside that last question.
This essay is about why.
The machine is engineered to be average
Start with a fact almost nobody says out loud.
A language model is a machine built to minimize surprise.
That is not an insult. It is the math. During training, the model learns by predicting the next word and getting punished when it’s wrong.
The optimization target is simple:
be as unsurprised as possible.
Reduce the error.
Smooth the curve.
Benoît put it through Claude Shannon, the father of information theory.
Shannon’s insight: information only exists where there is surprise.
A sentence you can perfectly predict carries zero information. « AI is a geopolitical issue » tells you nothing. « It’s not China that’s winning — it’s India » makes you sit up.
Now connect the two. The model is trained to avoid exactly the thing that makes content worth reading.
That is why raw AI output feels like corporate oatmeal.
Padded.
Pastiche.
Roman capitals in italic for « premium elegance, » the same trick every app has used since 2003.
The machine isn’t broken. It’s doing its job: returning the least surprising answer.
So the flatness is structural. Which means the edge is structural too. The human who can inject entropy — surprise, friction, an unexpected connection — is doing the one thing the machine is built not to do.
Words are the new code
Here’s where it gets useful.
How do you force a surprise-minimizing machine to surprise you?
You use better words.
A model doesn’t « know » that Paris is the capital of France. It has no drawer with a little card inside. It reconstructs the answer from the connections that show up most in its training. Everything it does is connections between words. Words are its entire way of thinking.
So vocabulary is leverage. Benoît’s example stuck with me. Ask for « an analysis » and you get mush. Ask for a « Gestalt analysis » and you’ve just opened a door — German psychology, the school of seeing the whole inside the parts — and behind that door are doors you’d never have reached.
Each precise word is a branch. Take the branch, new branches appear. The model’s real knowledge is an iceberg; the surface answer is all you get unless you know which word unlocks the depth.
This is the part I love, because it’s the revenge of the literary mind. For two years we were told the future belonged to coders. The actual skill turns out to be older: clarity, precision, range, knowing the exact word. Andrej Karpathy said English is the new programming language. Benoît goes further — general culture is the new edge.
Not « intellectual. » Practical. The word is a tool. If you want a sharp result, you need the sharp word.
The letter came back, and it’s not nostalgia
Benoît is a former journalist — regional press, Le Monde group, media builder. He bet his second act on email. Email.
Why does the oldest format on the internet still win attention in 2026?
Robustness. In a world accelerating past everyone, people reach for what’s stable. Email was born with the internet. Everyone has it. No notification fires when a letter lands — so it stays intimate without being intrusive.
And it carries an old human current. The first media in history were the information letters merchants wrote to each other by hand — the most reliable news of their time. The newsletter inherits that trust. It speaks person to person. You sit down, you open it, you take the time.
This is the polarization nobody names. The world is splitting into 7-second video on one end and 3-hour essays on the other. The middle is dying. Long isn’t good by default — long and boring is what AI produces on autopilot. But chosen, edited, alive long-form is now rare enough to be valuable.
The newsletter and the podcast win on the same three levers: connection, slow time, and intimacy. Exactly what’s scarce in the chaos.
Audience is rented. Community is owned.
Then Benoît drew a line I’ll keep.
Audience and community are not the same thing. Your 20,000 YouTube subscribers are an audience. Community starts the moment those people talk to each other.
His test for whether you’ve built one: when you get attacked, and they show up to defend you before you can. That’s a bond money can’t buy and a platform can’t repossess.
Génération IA has tens of thousands of readers. But the community is the 600 people in a WhatsApp group who now trade advice without him in the room. He barely posts there anymore. It runs on its own.
This matters strategically because audience is rented from an algorithm and community is owned outright. One is a number. The other is a moat.
Capitalize your intelligence or disappear
Now the part that should change how you work tomorrow.
Benoît’s frame for the AI era is brutal and simple: capitalize your intelligence.
You save money by putting a little aside every month. Compounding does the rest. Do the same with your mind. Every book, every conversation, every reflection — most of it you’ll forget by tonight. He doesn’t. He records it, his personal AI transcribes and files it, and it compounds into a second brain that knows what makes him singular.
The trap is the opposite move: rent everything, store nothing. Buy the ChatGPT subscription, hand over all your thinking, keep none of it. Do that long enough and you become transparent. Your edge evaporates. Your brain, like an unused muscle, atrophies.
So keep your data sovereign — your notes, your processes, your way of seeing. Those are yours. Then let a model come into your house, work with you, and leave when you want to swap it. The model is interchangeable. Your accumulated intelligence is not.
He built this for his father — a surgeon with a life worth recording — using Claude Code as a personal AI working on files in a folder, chapter by chapter, voice notes feeding the raw material. Same method that let him fool a Flaubert scholar. He didn’t pastiche the style. He studied how Flaubert thought. Form follows substance. Style without substance is bullshit.
I built a business on a Pokémon card — alone, in two hours
Theory is cheap. So here’s the proof I ran on myself.
Living in Indonesia, I spotted an asymmetry. The Pokémon Company prints a card sold only here: Pikachu Batik, the mascot wrapped in Indonesia’s UNESCO-listed batik cloth. Sealed, it costs around 15 euros in Bali. The same card sells for 80 to 100 euros in France. Ten times the price.
Rising every month.
That’s a textbook asymmetry. Same object, two prices, a border in between.
So I built pikachubatic.com — now the best site in the world on that single card — in about two hours.
Not by hand. With Nexus, my AI agent.
Nexus is the system I’m building to do exactly what Benoît preaches: keep the intelligence on my side. It’s wired to the Pokémon card database, checks the Indonesia-only drops every week, watches the events where the cards surface, and feeds a collector newsletter. The cards release one city at a time, every six months — and you have to be there on the day to get one. Nexus aggregates that edge so I don’t have to.
It’s a side project. Built for fun, because I collect things — figures, Dragon Ball cards, Pikachu. But it proves the point physically. One person, in a bedroom, now assembles what used to demand a developer, a designer, and a marketer. The agent walks every step with you — down to choosing the domain name.
The lesson isn’t Pokémon. It’s that the wall between an idea and a built thing just collapsed. The only question left is whether you push the first door.
The real power shift isn’t the robots
Here’s where the geopolitics lands, and where I push back on the doomers.
Benoît’s fear isn’t superintelligence. It’s concentration. We are pouring more money into this technology than into anything except nuclear weapons — and that power is pooling inside a handful of companies run by people he wouldn’t trust with his little sister.
I’m more optimistic on one front: Europe. Not because I’m French and pleased about it — because of the assets.
ASML in the Netherlands is the chokepoint of the entire chip supply chain. France runs on near-decarbonized energy, a real edge for data centers. Mistral signs sovereignty deals across the continent almost weekly.
The gap is investment, not capability. And the deeper risk is the Cloud Act: use an American tool and your data is reachable, wherever it’s stored. For a company like Thales, the choice isn’t « best model. » It’s sovereignty.
De Gaulle pulled France out of NATO command long before it was fashionable to distrust the alliance. The instinct was the same one Europe needs now: don’t be dispossessed of the most powerful technology humanity has built.
Push the first door
So what is a human worth in 2030?
Wrong question, Benoît says. Better question: what should I work on now? Become antifragile. Build what the machine is structurally bad at — surprise, judgment, taste, the connection no model would make.
He keeps coming back to Catherine, who at 60 quit her corporate job, won Paris’s best chocolate mousse contest, and built a brand in Japan when every financier said she was too old. Her line: if you want the next door, you have to push the first one.
That’s the whole thing. Whatever your age, you have a story to tell and something to build. AI, used well, is an instrument of emancipation — a kid in a village with an idea no longer needs the right passport or the right parents.
But protect the house you’re working from.
Buy a paper notebook. Read a paper book. Walk in nature. Practice otium — active, unproductive thought. Your brain is fragile in front of these machines. Guard it, then use them.
One more thing before I close. I’m a Michael Jackson obsessive — I went to see Michael, the King, here in Bali, and I felt every beat of it. Steve Jobs took a calligraphy class that looked useless for years, until it became the typography of the Mac. You can’t connect the dots looking forward. Only backward.
So take the strange class. Learn the odd word. Dance at the screening. The machine will never do it for you — and that, exactly, is your edge.
This was the first episode of a podcast we plan to run a hundred times. Benoît Raphaël writes Génération IA, which I read religiously, distractions off. Nexus is the sovereign AI agent I’m building so my intelligence compounds on my side, not theirs. If one idea here moved you, push your own first door — and tell me which one in the comments.
— Rémy