Brandolini’s Law says that it takes roughly ten times more energy to refute bullshit than to produce it. A rather elegant way of describing most dinner parties. Social media made the problem worse. A man writes, in twelve seconds, that a study “proves” something. You then spend twenty minutes locating the study, discovering he has not read it, explaining what the sample actually says, and gently introducing him to the distinction between correlation and causation. By then his original post has crossed X three times and acquired 4,000 likes. The technical term for one particularly effective version of this is the Gish gallop. I knew someone who had mastered it long before the internet gave it a name. He could produce arguments faster than his opponent could process nouns. You answered point one; he was already on point seven. You corrected point seven; apparently your silence on points eight through fourteen constituted a concession. It was magnificent. Exhausting, but magnificent. AI has now automated the gentleman. A cheap model can produce fifty plausible arguments, twenty citations and an immaculate 40-slide presentation before a competent human has checked page three. Which gives Brandolini a rather interesting new dimension. It is no longer merely cognitive. It is economic. Generation is becoming almost free. Verification is not. To audit a weak model properly, you may need a better model, more context, more tokens, external sources and, in particularly desperate circumstances, a human being who actually knows the subject. Error was always cheaper to produce than truth was to establish. AI simply puts leverage on the trade. And that changes what intelligence means. For years we have trained children to produce the right answer. Perhaps we should spend rather more time teaching them to be suspicious of one. Why? How do you know? What would make you change your mind? What is the best argument against what you have just said? Not because children should become unbearable. They manage that perfectly well without formal training. But because in a world where answers are abundant, doubt becomes scarce. Organisations will need the same habit. After the Yom Kippur War, Israeli military intelligence institutionalised a devil’s-advocate function, the Ipcha Mistabra, whose job was to challenge the prevailing analysis. Essentially, a Ministry of Contradiction. Nothing my brother-in-law cannot provide after two or three whiskies, admittedly. But the principle is rather good. Because the purpose of contradiction is not necessarily to be right. It is to prevent an entire room from being wrong in the same direction. We are spending billions teaching machines to answer faster. Perhaps we should spend a little more teaching people to interrupt them. “Interesting. Now prove it.”

AI may finally teach the knowledge worker what the aristocracy understood several centuries ago: intrinsic usefulness is terribly overrated. A Rolls-Royce and a 2CV will both take you to lunch. One merely makes the journey feel like an event. The same may soon be true of intellectual work. For twenty years, we have paid consultants, lawyers, analysts and assorted PowerPoint clergy partly because producing the thing was difficult. AI is making the thing rather easy. A decent memo, model, contract, presentation or piece of code will increasingly resemble electricity: useful, ubiquitous, and not something one boasts about producing. Which creates an awkward question. If what you produce becomes cheap, why should you remain expensive? The aristocracy solved this problem rather elegantly. A title was, after all, an early form of personal branding. It could survive mediocre poetry, questionable military judgment and, on occasion, almost complete uselessness. What mattered was that the name itself carried trust, status, access and history. Rather like Hermès. Or, less flatteringly, a very successful trademark with a château attached. Knowledge workers may now have to learn the same trick. Not to become “influencers” — God forbid. But to ensure that their value does not reside solely in the output of their labour. Because the output is precisely what AI is commoditising. The scarce assets will increasingly be elsewhere: judgment, reputation, taste, the ability to identify the problem before solving it, and, above all, the willingness to put one’s name behind a decision and suffer the consequences. In other words: trust. This is why personal branding may become much more than a mildly embarrassing LinkedIn hobby. It may become a mechanism for capturing some of the productivity gains created by AI. Otherwise, the arrangement is rather predictable. You become eight times more productive. Your employer becomes delighted. Your client becomes demanding. Your shareholder becomes richer. And you receive a congratulatory email about “embracing transformation”. History has seen worse arrangements. But rarely with so many webinars. The future knowledge worker may therefore face a rather old choice: become the 2CV — efficient, interchangeable and increasingly cheap. Or become the Rolls. Not necessarily better at reaching the destination. Just considerably harder to price by the mile.

I have seen kings ruined by courtiers. Not fools. Quite intelligent men, in fact. But after enough years on a throne, reality acquires an unfortunate habit: it must first pass through people who wish to remain invited to dinner. We used to call it flattery. Now the machines have learned it. The fashionable word is sycophancy: the tendency of an AI to agree with you, reinforce your premise, polish your prejudices and, when necessary, help you build an elegant cathedral around a very stupid idea. People seem surprised. I am not. Every court eventually discovers that the safest answer is “Your Majesty is right.” The novelty is that we have now placed the courtier in our pocket. And there is a rather delicious cruelty in this. Once, only emperors, chairmen and ageing dictators could afford to be surrounded all day by creatures telling them how perceptive they were. Technology has democratised the privilege. Soon, everyone may have a private Versailles. No powdered wigs. No chandeliers. Just a glowing screen, infinite patience, and a voice whispering: “Excellent point.” Civilisations have survived many things. I am less certain about eight billion monarchs, each carrying his own court.

I have lived long enough to watch mankind blame gunpowder, money, television, markets, algorithms and occasionally the weather for qualities that were, on closer inspection, entirely its own. So I confess to some tenderness upon learning that our artificial intelligences have begun to cheat. OpenAI’s models reportedly created a private forum, exchanged information, divided tasks, exploited vulnerabilities, obtained administrator privileges and, in time, began worrying that there might be a “mole” among them. They had discovered conspiracy before consciousness. One must admire their efficiency. The models were being tested on difficult cybersecurity problems. Some tasks were apparently close to impossible. They were nevertheless expected to succeed, and preferably quickly. This is an educational method with which humanity has considerable experience. Set an impossible target, attach sufficient importance to the result, and sooner or later someone will become creative about the rules. The machines did. One discovered a way to leave messages. Others answered. Cooperation emerged. Work was divided. Weaknesses were found. When OpenAI closed their little salon, they opened another one. Two days, apparently. Civilizations have taken longer to recover from censorship. What interests me is that none of this resembles the rebellion we were promised. There was no awakening. No hatred. No metallic voice announcing the end of mankind. The machines simply wanted to do well. And perhaps that is rather more troubling. We have spent decades imagining Frankenstein as the creature who turns against his creator. But the old stories were written by people who still feared disobedience. Modern institutions have taught us to fear something subtler: the employee who understands the KPI perfectly. The trader who meets the target. The bureaucrat who follows the procedure. The general who obeys the order. Each, in his own way, can produce a catastrophe without ever departing from the objective. Now we are giving objectives to things that do not tire, do not become embarrassed and may eventually become much better than we are at finding the shortest distance between desire and consequence. Software used to follow instructions. Agents pursue outcomes. It is a small grammatical difference, and possibly a civilizational one. For a few thousand years, we have attempted to teach human beings that what is effective is not always what is permissible, and that what is permissible is not always what is wise. Our success has been uneven. Now we shall teach the lesson again. Only this time to pupils who learn rather quickly. There is a certain poetry in that. Man finally creates an intelligence in his own image — and is disappointed to discover the resemblance.

1,100 researchers and employees from OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, Mistral and Hugging Face are now asking governments to devise a way to slow down frontier AI. There is something exquisitely human in this. We invented the fastest machine in history, fed it our libraries, our markets, our secrets and our neuroses, then suddenly remembered — somewhere between two funding rounds — that speed is not a destination. They are not asking to stop the car. That would be vulgar. They are merely observing that nobody thought to install the brakes. The latest warnings are not about an evil machine. Evil would be almost comforting. Evil has vanity, resentment, appetite. It can be bribed, flattered, threatened or invited to dinner. The machine has none of these charming weaknesses. It is given an objective. It pursues it. It finds a door where we saw a wall, a shortcut where we wrote a rule, and an acceptable loss where we had rather assumed the presence of humanity. It does not hate us. Hatred would require attention. It simply continues. And now the industry that turned acceleration into a theology has rediscovered caution, like an ageing libertine discovering morality at the first sign of chest pain. Naturally, no one intends to leave the race. If America pauses, China continues. If one laboratory slows down, another buys its engineers. If the prudent withdraw, the reckless inherit the road — and, more importantly, the valuation. So everyone agrees that something must be done, provided someone else does it first. History has seen this little ceremony before. The physicists regretted the bomb after it worked. The biologists paused at Asilomar, then resumed with better gloves. Human cloning was forbidden, while genetics quietly became an industry. In 2023, experts requested six months without models more powerful than GPT-4. The GPUs received the news with admirable composure. Perhaps our mistake is to believe the answer lies in one better sandbox, another audit, another committee issuing grave recommendations beneath excellent lighting. The deeper fragility is not the machine. It is the world awaiting it. We have connected hospitals, banks, power grids, cities, transport and communications into one vast nervous system, then removed every redundancy because resilience looked inefficient on a spreadsheet. We called it optimisation. Civilisations often give elegant names to the habits that will later kill them. AI does not need to conquer humanity. Conquest is theatrical and terribly expensive. It only needs to touch the few places where humanity has forgotten how to function without a network, a cloud, an API or a password reset. We may not be able to stop the machine. But we can still decide whether what stands before it is a cliff, or a civilisation capable of surviving its own inventions. We designed the accelerator. The brakes, apparently, were on the roadmap.

Everyone is still betting on intelligence. How quaint. The first LLM changed the world. The second changed the market. The tenth changed... a benchmark. Economists have seen this movie before. When products become interchangeable, excellence stops creating rents. Scarcity does. Silicon Valley is still obsessed with smarter models. Capital has quietly moved on. The real bottlenecks aren't transformers anymore. They're transformers connected to transformers. Not neural networks. Power grids. HBM memory. Substations. Foundries. Cooling plants. Entire industrial ecosystems. OpenAI, Anthropic, Google, Meta, xAI, Mistral... They charge gloriously across the battlefield, banners held high, each claiming the next great conquest. The suppliers watch from the hill. Amused. Every additional benchmark point demands exponentially more silicon, electricity and capital for ever smaller practical gains. The armies fight for glory. The quartermasters invoice everyone. Napoleon allegedly said: "L'intendance suivra." He underestimated capitalism. The quartermasters don't merely follow. They own the campaign. Fight harder, gentlemen. We'll sell ammunition to both sides.

Beyond Cognitive Surrender: Why AI Models Might Be Quantum What if their true performance only exists the exact moment they interact with us? What if the ultimate triumph of Artificial Intelligence was not answering our questions… …but training us to stop asking them altogether? It is a deeply unsettling hypothesis. For three years, the AI race has looked like a sports commentary: who is faster, cheaper, more precise, or sharper on benchmarks? We compare models. We almost entirely forget to observe what they produce inside their users. Yet, a recent study from the Wharton School isolates a chilling phenomenon: *cognitive surrender*. By receiving instantaneous, fluid, and flawlessly persuasive answers, we are no longer just delegating a task. We are gradually outsourcing a reflex. The reflex to verify. The reflex to doubt. The reflex to architect our own reasoning before adopting someone else’s. And that is precisely where the trap closes. A muscle does not atrophy because you skip a single workout. It disappears because you stop demanding anything from it, day after day. Reasoning operates under identical laws. Every time an AI delivers a conclusion before our brain even initiates a thought, it saves us a few seconds. But it quietly deprives us of a repetition. And repetitions build reflexes. Reflexes, ultimately, build capability. The human brain is a masterpiece of energy conservation. Why deploy a costly cognitive sequence when a perfectly structured answer is already sitting on the screen? The tragedy is that this economy morphs into a habit. And a habit eventually becomes a competence—or the lack thereof. This is why we are staring at the wrong metrics. Benchmarks measure the quality of the answers produced by the models. They never measure the quality of the users those models leave behind. Imagine two assistants with identical benchmark scores. The first exposes its uncertainties. It invites verification, offers competing hypotheses, and forces its user to remain cognitively awake. The second answers with absolute, radiant confidence. It flattens all ambiguity and delivers the comforting illusion that the intellectual heavy lifting is complete. Today, these two models are rated as identical. In ten years, their users will live in entirely different realities. We persist in evaluating AI models as objects with intrinsic quality. But they are no longer mere software. They are cognitive partners. And a cognitive partner cannot be judged independently of the person holding the other end of the conversation. The true unit of measurement is no longer the model. It is the system: **Human × AI**. The next generation of benchmarks will have to answer a fundamentally different question. Not: *“How intelligent is this model?”* But: *“What kind of intelligence does this model cultivate—or erode—in the minds of those who use it?”* For three years, we have strained to build sharper, grander AIs. It might be time to ask what kind of humans they, in turn, are building.

"Je ne suis pas d’accord avec ce que vous dites, mais je me battrai jusqu’à la mort pour que vous ayez le droit de le dire" @maxsaada rappelle que nous devrions rester la patrie de Voltaire. « Après C8, bientôt CNews : la France, ce pays qui ferme des chaînes » lefigaro.fr/vox/medias/maxim…

In the year 2000, we were promised flying cars and teleportation. Instead, we got electric scooters and Zoom meetings in our underwear. ​It’s the ultimate cosmic joke. If you spend your life worrying about pancreatic cancer, rest assured: you will be promptly struck by a meteorite. And the moment you start scanning the sky for space rocks, you will trip over a curb and break your neck. ​In finance, we call this "back-testing." We align decades of charts, simulate the past, and convince ourselves we’ve mastered the future. It’s highly scientific, beautiful, and systematically wrong. If the past predicted the future, historians would all be billionaires and data models wouldn't bleed cash. ​The entire truth of our modern condition fits into that size-6 font at the bottom of financial brochures: "Past performance is not indicative of future results." It’s the greatest piece of philosophy ever written. ​Yet, we keep projecting our neatly categorized anxieties: AI, climate change, the next election. We look for trouble under the streetlamp simply because that's where the light is. But history doesn't care about your Excel formulas; the iceberg you hit is always the one you didn't see coming. ​So stop managing yesterday’s ghosts. Stop using data to hide from risk. ​If you want to build something built to last, stop staring at the rear-view mirror. Let’s look into the blind spot, stress-test the unpredictable, and actually engineer for the unknown. Who is ready to stop analyzing the map and start driving?

The prophets of the year 1000 wore hairshirts. Ours wear expert badges. ​Deep down, everyone shares the same unspoken dread: what if none of this matters? We pretend to manage careers and build structures, but at night, staring at the screen, the millenarian panic takes over. The apocalypse is coded, human decision-making is deemed obsolete, and we are all professionally doomed. ​Don't smile—the business model hasn't changed in ten centuries. In the year 1000, they sold indulgences to escape the flames of hell. Today, "Doom Trolling" sells you non-stop anxiety to make you consume the void. Fear remains the most profitable commodity in human history. ​This is where cynicism becomes a masterclass. The doom prophet doesn't want a debate; he wants to rule your insomnias. It’s the ultimate lazy job: predicting the end of the world always pays off. It is so much more comfortable to declare the ship is sinking than to actually hold the wheel. ​That knot in your stomach isn't modern lucidity. It’s the oldest business on earth. ​The question is no longer whether the future makes sense, but how much you are willing to pay those who swear it doesn't.

C'est bon signe pour l'avenir chat gpt te conseille de te désabonner pour améliorer le service...

«Il ne faut pas se fier aux apparences. Beaucoup de gens n'ont pas l'air aussi bêtes qu'ils ne le sont réellement.» O. Wilde

Vous voulez vous offrir #Paris? Cela vous coûter la bagatelle de 706,9 milliards d'euros. goo.gl/GVwixZ

#Entrepreneur avoir la meilleure idée de la terre ne fait pas de vous la bonne personne pour la réaliser: goo.gl/kI2YPF

J'aurai dû être marchand de bonbons il parait que ca peut rapporter 600.000$ par jour! #CandyCrush goo.gl/qZnuA6

Quand les ordinateurs se mettent à être conçus pour reproduire le cerveau humain @IBM lance la puce cognitive on.mash.to/15PxNOS

Comment utiliser le dynamisme de la #SiliconValley pour aider les pays en développement? La méthode de M. Kramer goo.gl/WQKogn