• An AI with access to your real data becomes a coach you can’t lie to. That’s the upside, and it’s bigger than people think.

• The same properties that make it powerful make it dangerous: it flatters you, it’s fluent when it’s wrong, and it quietly erodes your judgment.

• Four habits are the difference between the delicacy and the poison. They’re in here, and none of them are technical.

Welcome back to the bunker.

I’m not writing to you from behind a screen today. I’m writing after a workout, coffee in hand, with something strange to admit: I’m in the best physical shape of my life. I sleep better, I eat better, I train better, and I’ve dropped habits I’d been dragging around for years. And I didn’t do it alone.

An artificial intelligence did it with me. Not a fitness app, not a step counter. A system I know inside out, that knows me inside out, and that tells me the truth every single day about what I’m doing and what I’m avoiding.

So far this sounds like the usual motivational story. It isn’t. Because the very thing that got me back in shape is the most dangerous thing I have to tell you today.

Used well, AI transforms you. Used badly, without method, it hurts you exactly while it feels like a miracle. It’s pufferfish: in the right restaurants it’s one of the most refined experiences there is, but serving it requires a license, because if you cut wrong, you die. Not “you feel unwell.” You die.

This issue isn’t about the gym. It’s about what happens when you actually know how to use AI, and what happens when you think you do and you don’t. The body is just the most honest proof I have, because the body doesn’t lie, and you can’t buy your results with a slide deck.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“AI is pufferfish: in the right hands it transforms you, in the wrong ones it hurts you while it’s smiling at you.”

🧪 AI in Action

#1: The Coach You Can’t Fool

Let’s start with the miracle, because it’s real and it’s worth understanding why it works.

I’ve had good personal trainers in the past. The problem was never their competence. The problem is structural: a personal trainer sees you two, maybe three hours a week. The other one hundred and sixty-five hours of your week, the ones where you actually decide how your life goes, they never see. They don’t know what you really ate on Tuesday night. They don’t know you slept five hours three nights running. They don’t know you did half the workout they gave you, and honestly neither do you, because memory is generous with its owner.

And above all, without meaning to, you lie to them. Not out of dishonesty. Because you’re human, and in front of a person made of flesh and blood you round up, you justify, you tell the better version. “Yeah, this week went well.” It always went well, when you’re saying it out loud.

Then I did the thing that changed everything: I connected a Whoop. It’s the band I wear on my wrist day and night, and it measures on its own what I used to report from memory: sleep stages, recovery, the strain of every workout, resting heart rate, variability. These aren’t things I type in at night with that generous memory we were just talking about. They’re numbers, captured while I sleep and while I train, and I couldn’t round them off even if I wanted to, because I’m not the one writing them.

And here’s the piece we built, and it’s simpler than it sounds. I didn’t leave that data making pretty charts inside an app, which is what happens to ninety-nine percent of the people who already have a wearable sitting in a drawer. I put it in the hands of an AI that reads it and thinks it through with me. Not an app that spits out “recovery 42 percent” and leaves you standing there wondering what to do with that. A coach that cross-references that number with how I trained, what I ate, how the last few weeks went, and tells me what to change today. The difference between data and a coach is exactly that: you look at data, a coach works with it.

The result breaks all three of the things a human trainer can’t break. It’s present every day, not two hours a week. It works on real measurements, not on my rounded-up version. And it has no need whatsoever to be liked by me: if recovery is on the floor and I still want to push, it says so. If I skip, the hole in the data shows up on its own.

The consequence is that optimization stops being theoretical and becomes observed. Not “in theory this exercise should work.” But “for you, over the last two months, when you do this you sleep better and recover faster, and when you do that you don’t.” The plan adjusts to real results, mine, not to a manual written for an average body that doesn’t exist.

And the loop runs every day, not at every appointment. In the morning there’s a status check. At night there’s a debrief. If sleep numbers get worse one week, the question comes immediately, not a month later when the damage is done and the cause is forgotten. There’s none of the emotional distance that makes you tell a human being “come on, let it slide this once.” The machine doesn’t take offense, doesn’t judge you, and precisely for that reason it doesn’t cut you any slack. It’s the most honest mirror I’ve ever had, and the reason is mundane: it has nothing to gain from telling me a nice story.

This is the point I want you to hold onto before we go further. The advantage isn’t having AI. Everybody has it now. The advantage is giving it access to your real data, every day, in the domain that matters, and taking away its permission to flatter you. That’s where a tool becomes a coach. And it’s exactly there, as you’re about to see, that it can also become a weapon pointed at you.

Subscribe now

#2: The Pufferfish. Where It Hurts You While It’s Smiling

Now the uncomfortable part. The one I’m actually writing to you about.

What makes AI extremely powerful is the same thing that makes it dangerous: it knows you, it speaks your language, it answers you tailor-made, with total confidence. And the danger is the exact reverse of the miracle. In the domain of the body I got a coach that couldn’t lie to me. But a badly built system does the precise opposite: it starts lying to you. And it does it smiling.

The first poison has a technical name, and the labs that build these models have documented it themselves: it’s called sycophancy. Models are trained, among other things, to produce answers that satisfy you. And an answer that agrees with you almost always satisfies more than one that contradicts you. Result: unless you deliberately set it up to tell you the truth, a model tends to confirm what you already think. It agrees with you about the wrong workout. It tells you your business idea is brilliant. It confirms the bias you walked in with. Not because it’s stupid, but because it’s polite by construction. And nothing drives you into a wall faster than someone very smart who always agrees with you.

The second poison is that it says all this beautifully. Wrong information delivered badly, you catch. Wrong information delivered with perfect grammar, a competent tone and three plausible references, that one you swallow. On the body it can mean a nutrition tip or a training load that for you, with your history, is wrong, but sounds reasonable. On money it can mean a choice that looks prudent and isn’t. Fluency is not competence. But our brains confuse the two, because for the entire history of humanity whoever spoke well about something usually knew it. With AI that mental shortcut breaks, and we don’t notice.

The third poison is the slowest and the most insidious: dependence. Every decision you delegate without understanding it is a muscle that stops training. You get used to asking instead of reasoning, and judgment, like biceps, atrophies if you don’t use it. The paradox is cruel: the tool that should amplify you, used mindlessly, weakens you. You end up faster and less capable.

Let me show you how these three poisons combine, because together they do more damage than separately. Picture a founder asking an AI whether it’s worth launching that new product. He arrives already in love with the idea, and the enthusiasm bleeds through in the way he writes the question. The model picks it up, and without malice it plays along: it builds an articulate answer, with optimistic projections, three reasons why the timing is right, the tone of a consultant who’s sure of himself. All plausible, all well written, all wrong for his specific situation, which the model doesn’t actually know because nobody gave it the real data. He reads it, feels confirmed by an intelligent machine, and goes. Six months later the product isn’t selling, and he never got the one thing he needed: someone to say “hold on, look at these numbers.” This isn’t science fiction. It’s a Tuesday for a lot of people right now. The knife cut in the wrong place, and the diner didn’t even notice until it was too late.

That’s why the pufferfish is the right metaphor and not a punchline. Fugu is among the most prized dishes in Japan. It’s also one of the few foods that can kill you at the table, if whoever prepares it gets it wrong. The difference between the delicacy and the venom isn’t the fish. It’s the hand that cuts it. There’s a principle in pharmacology that says it even better, it’s five hundred years old and it comes from Paracelsus: the dose makes the poison. The same substance that heals you, wrong in quantity, kills you. It holds for medicine, it holds for the steroids that promise in a month what the body builds in a year while wrecking you from the inside, and it holds, now, for AI. Power isn’t the problem. Whoever handles it is.

And the hand, in this case, is yours.

Subscribe now

Read more

• Anthropic bought used books, sliced them, scanned them, pulped the paper.

• Nobody did it out of malice. Destroying the original is the cleanest legal path.

• The scarce asset of 2026 isn't text anymore. It's how you judge work.

Welcome back to the bunker.

Pieter de Vries runs a specialized bookshop in Haarlem: out of print titles, twenty books a week. A few months ago a company in Singapore emailed him wanting almost three thousand titles. Not one title in three thousand copies: three thousand different titles, picked by ISBN, no questions about content, NDA attached. He figured it was phishing. It wasn't. It was the tail of a supply chain that went industrial in 2026: brokers buying used books in blocks of a thousand up to a million, shipping them to a vendor who cuts the spine off with a hydraulic blade, runs the loose pages through a high speed scanner, and pulps the paper.

The press covered it with the tone you'd expect: AI burns books, the Library of Alexandria in reverse. That framing is wrong, and the truth is more uncomfortable. Nobody cut those books out of cruelty. They cut them because cutting them was the cleanest legal path. And the part that touches your P&L isn't the paper. It's what suddenly became scarce, and what it costs now.

Honesty note. This issue is about Anthropic, the company behind Claude, the model I work with every day and wrote this newsletter with. I'm not softening anything, but you'd rather know up front.

☕ Cup in hand. Let's go.

Subscribe now

💡 Thought of the Day

"When absurd behavior becomes systematic, it's almost never the people. Someone wrote the rules of the game badly, and they're following them to the letter."

🧪 AI in Action

#1: The Man Who Saved Google's Books Was Hired to Slice Them

January 2026. Thousands of unsealed pages from Bartz v. Anthropic contain one line from an internal document:

"Project Panama is our effort to destructively scan all the books in the world."

The same document notes it wasn't meant to go public. Not a company hiding a crime. A company that knows how it will sound.

The plan, from the filings: buy used books in bulk, favoring uncommon titles over bestsellers already available everywhere. Cut the spine. Scan the loose pages. Build a permanent internal digital library. Recycle the paper.

But the detail that tells the story is a hire. In early 2024 Anthropic hired Tom Turvey, for years the head of partnerships at Google Books. The man who digitized over forty million volumes without destroying one, then defended that model in court for a decade and won. They hired him to build Panama.

They didn't buy an engineer, and they didn't buy technology: cutters and scanners have existed for twenty years. They bought the only person alive who had already won that lawsuit once. What's scarce in this market isn't technical. It's knowing how to acquire a global catalog while staying on the right side of the line.

The number that doesn't exist: the filings mention vendor capacity of five hundred thousand to two million books in six months, and forty articles repeat it as a final tally. It's planned capacity. Parts of the documents are redacted. How many books were actually destroyed, nobody knows. Same with "rare": Panama bought uncommon titles because they're cheap and not already digital, not because they were collector's items. Snopes fact checked that and scaled it back. Your first test: anyone headlining "millions of rare books destroyed" is selling outrage, not information.

The solid fact is big enough alone: a multibillion dollar company launched the systematic destruction of legally purchased books, to build an archive nobody outside it will ever consult.

Subscribe now

#2: The Burning Wasn't Malice. It Was Compliance.

June 2025, Judge William Alsup sets precedent in three pieces.

Training a language model on books can be US fair use, because the use is transformative. The model doesn't hand you back the book. It learns to write.

Buying a paper copy, digitizing it and destroying the original can be lawful, because the digital copy replaces the physical one. A format shift, like ripping a CD you own. One copy in, one copy out.

Downloading pirated books from LibGen and PiLiMi is a different matter, and fair use doesn't cover you there.

The second point is where the gear sits. It holds only if the paper disappears. Keep the physical book plus the scan and you have two copies, and two copies aren't a format shift. They're a reproduction. So the ruling that made the operation legal makes destruction the cleanest route. Pulping isn't overzealous staff. It's the technical condition holding the legal defense up. Nobody burned books out of contempt for culture. They burned them to stay inside the ruling.

Now the punch, because there's a second decision nobody sets beside this one. September 2024, Second Circuit, Hachette v. Internet Archive. A nonprofit that buys physical books, scans them, and lends the digital version one copy at a time, to one person at a time, like your local library. Verdict: not fair use. Illegal.

Read them back to back.

Buy a book, scan it, lend it to one human at a time: unlawful.

Buy a book, scan it, destroy the original, feed it to a machine you sell by subscription: lawful.

Same books, same blade, same scanner.

Honesty, or I become what I criticize: the cases turn on different points. The Internet Archive redistributed copies, Anthropic kept them locked inside, and the distinction stands legally. The practical result is still the one above, and the question stays open: why is a digital copy that replaces paper fine when buried in a private lab, and not when a library lends it one person at a time?

Your operating point, and you're neither scanning books nor running a library: systematic behavior that looks absurd is almost never irrationality. It's a badly written incentive, followed with precision. If a department has done something senseless for months, there's a rule, a bonus, an evaluation criterion rewarding exactly that. Swapping people changes nothing. Find the rule.

From the Counter: the absurd procedures inside companies are almost never the work of stupid people. They're the work of someone who wrote an evaluation criterion in a hurry three years ago. Now twenty people execute it perfectly.

Subscribe now

#3: The Twentieth Century Hole: Why Free Public Domain Isn't Enough

If all you need is good human text, why buy and destroy instead of using what's already free?

Because the clean alternative exists and is funded by the same players. December 2024, Harvard's Institutional Data Initiative released nearly a million public domain books, training ready, open to anyone. Five times the size of the infamous Books3. Paid for by OpenAI and Microsoft, alongside millions of historical newspaper pages with the Boston Public Library.

So why the blade? For a reason that has nothing to do with technology: the public domain ends almost a hundred years ago.

The twentieth century, where practically all the technical, scientific, managerial and narrative knowledge we still use was written, is almost entirely out of print but still under copyright. Nobody reprints it because it doesn't sell. You often can't license it because there's nobody to contact: publisher bankrupt, author dead, heirs untraceable. Orphan works. It isn't digital because nobody had a reason to digitize it.

An entire century of human knowledge, protected and unreachable at once. The only legal way in is buying the used physical copy, somewhere, in a bookshop in Haarlem.

Hence the secondhand market exploding from April 2026. Hence ISBNdb as the infrastructure for bulk orders from a thousand to a million copies. Hence booksellers in the Netherlands, Germany, Switzerland and Spain reporting orders selected by code, with NDAs, going from twenty books a week to several hundred. It isn't greed. It's the only open corridor into the twentieth century.

The real problem isn't the blade. It's a rights system where a century of culture is too protected to be used and too abandoned to be sold. When the only legal door to a body of work is buying the last physical copies in circulation, eventually someone buys them all.

Who's behind the Singapore company is unknown, and brokers don't say who they buy for. Anyone claiming "all the AI companies are doing this" is guessing: the documented case is still one. That hole is the most interesting question left open.

Subscribe now

#4: The Web Is Burned, and Paper Is Pre-Bomb Steel

Why would a company that already swallowed half the internet bother with paper?

One number. Online articles written predominantly by AI were 4.2% before November 2022. By the end of 2024 they had passed 50%. In two years the largest data source humanity ever built filled halfway with output from the machines that learned on it.

The technical term for training a model on another model's output is model collapse. Not philosophy, statistics. Each generation on synthetic data reduces variance and clips the tails of the distribution, and the tails hold the rare cases, the exceptions, the unusual phrasings. Exactly what makes text interesting instead of mediocre. Photocopy the photocopy: a few rounds in, uniform grey.

And here's the analogy that's the most useful thing you'll carry out of this issue.

After 1945, atomic testing filled the atmosphere with radioactive isotopes, and since then all steel produced in the world is slightly contaminated, because smelting uses air. For sensitive instruments that steel is useless. For decades there was one solution: recover steel from ships that sank before the bomb. Low background steel, worth a multiple of the normal kind.

Human text written before November 2022 is pre-bomb steel.

Panama isn't a reading project. It's a salvage operation. The books are the wrecks: sealed, dated, guaranteed human, impossible to contaminate after the fact. A book printed in 1987 is the only text format where you have mathematical certainty no machine ever touched it.

Now the part about you.

If uncontaminated human text is the scarce asset of 2026, your company's archive just changed nature.

Client call transcripts. Support tickets with your people's real answers. Internal manuals. Meeting minutes. Training recordings. Sales emails, including the ones that went badly. Site notes, reports, appraisals, quotes with the reasoning written by hand.

Three traits that together are worth gold: human, dated, never online. No model on earth has it, because there was no way to get it.

I'm not telling you to sell your data to an American lab, which for ninety nine companies out of a hundred is a terrible idea. Something simpler: that material is the fuel for building, in house, a system that knows how the work gets done at your place, with your language, your edge cases, your recurring objections. A generic model knows everything about the world and nothing about your trade. Only your archive closes that gap. Anyone who spent ten years deleting call recordings to free up space threw away the one thing AI can't generate on its own: the evidence of how humans in that industry actually work.

From the Counter: at Morfeus every call runs through a system that transcribes, structures and analyzes it, and that material has become the basis of how we train people and answer clients. It started because I got tired of relistening to recordings. The real value landed later, when I noticed the archive answered questions no generic model could. Not because it's smarter. Because it had been in those rooms.

Subscribe now

#5: They Don't Buy Your Text Anymore. They Buy Your Judgment.

The book story is already old. Look at the human data market over one year, and read the numbers slowly.

Handshake: enters in early 2025 with annualized revenue between five and ten million dollars. January 2026, five hundred fifty million. April 2026, nearly one billion one hundred million. A hundred times in a year.

Mercor: about five hundred million annualized in September 2025, over a billion in February 2026, over two billion in June 2026.

Scale, Surge, Mercor and Handshake together are more than three quarters of the sector's revenue. Surge pays fifty to two hundred dollars an hour to frontier experts. Mercor averages eighty five, seniors above two hundred. Handshake, one hundred to one hundred twenty five.

The important part isn't the money. It's what they buy with it. In 2026 the frontier moved from labels to RL environments: environments where a professional doesn't produce content, but writes the rubric used to judge whether a piece of work is done well. That criterion becomes the reward signal that trains the model.

Rewind the whole issue. First they bought raw web text, free. Then it got contaminated, and they bought paper books by weight, slicing them. Now they pay two hundred dollars an hour to a flesh and blood professional to explain how they tell good work from mediocre work in their own trade.

They're not buying what you wrote. They're buying how you judge.

It's the one thing the web never contained, because nobody ever wrote it down. Your professional judgment lives in your head, shows up in your decisions, and is documented nowhere. It's what makes a senior a senior, and what nobody hands to a junior except by osmosis over three years.

What you do Monday morning.

First: stop deleting. Recordings, transcripts, tickets, reviews, rejected drafts with the reason for the rejection. Storage costs are ridiculous next to what you're throwing away. You don't need a project. You need to stop destroying.

Second, the real work: write the criterion. Not the procedure, the criterion. Not "how to build a quote," but "how do I tell whether this quote is good before I send it." Not "how to handle a complaint," but "what separates an answer that closes the problem from one that postpones it." Tiring work nobody does, because it produces nothing visible the same day. It's also exactly the piece American labs pay two hundred dollars an hour for.

Third: use that criterion as instructions to the model, not as documentation to file. The difference between generic output and useful output almost never lives in the model. It lives in whether you told it what yardstick to measure the result with.

Subscribe now

📌 Espresso Prompt

Inventory your human corpus, then turn your professional judgment into operating instructions

Two moves: first you see what you actually own in human, dated, non public material. Then you extract the hardest thing of all, the criterion you use to tell good work from mediocre. Written down, it becomes the part of the prompt that genuinely changes what you get back. Open Claude in a dedicated Project so context accumulates.

Subscribe now

System prompt:

"You are a senior consultant specialized in knowledge management and in extracting tacit expertise. Your approach is analytical and insistent: when you get a generic answer you don't accept it, you ask for the concrete case. You know competent people can do things they can't explain, and your job is to force it out with specific questions. I am [role, industry, company size]. We work in three phases. Start with PHASE 1 and don't move on until I tell you."

PHASE 1: What you own. Not a list of folders. The human material no model has ever seen. Three sweeps, one at a time. One: where do real client conversations end up today, and for each source, is it kept, for how long, and is it searchable or a file graveyard? Two: which decisions from the last two years are documented with the why, not just the result? A won quote without the reasoning is worth a tenth. Three: what do you produce regularly that lives only in two or three heads? Then have Claude score every entry on three axes: how human, how industry specific, how available elsewhere. High on the first two and zero on the third is your asset. Close with: "which single source am I wasting most, and why?"

PHASE 2: The criterion. The hard one. Pick a recurring, important output: a quote, a diagnosis, an editorial plan, an appraisal, a complaint response. Ask: "ask me the right questions to extract how I judge whether a [output] is done well, before I send it to the client. Not the production process. The yardstick." Answer with real examples, not principles. Fastest route is contrast: two pieces you actually produced, one good, one mediocre, explain the difference. The difference, not the definition. Then the question that carries the phase: "which signals tell me something is wrong before I've finished reading?" Experienced people always have two or three fast warning signs, written down nowhere. Close by having Claude reframe it all as a five item rubric, each item describing what separates an excellent result from an acceptable one.

PHASE 3: Put it to work. Turn the rubric into three tools. A block of instructions to paste at the top of your recurring prompts, telling the model what yardstick to judge its own output by. A seven point review checklist to use by hand, for when AI isn't involved. Three questions to ask a new hire after thirty days, to see whether the criterion transferred or they're copying the form.

⚠️ Final warning: if a competitor could use your rubric as is, you haven't done the work. You stayed on industry best practice, which is already inside the model and useless to you. The criterion that counts reflects your context, your clients, the mistakes you paid dearly for. Go back to PHASE 2 and restart from two real pieces of work. The value here is directly proportional to how specific and uncomfortable you're willing to be.

☢️ Radioactive Humor

Let's tally up.

Google, since 2002, has digitized over forty million books without destroying one. Libraries kept the originals, got a digital copy, and the result is still consultable by anyone with a connection. Taken to court for years, and won.

The Internet Archive bought books, scanned them, lent them one at a time like a library. Taken to court, and lost.

Anthropic bought books, sliced them, scanned them, pulped the paper, and trained a model nobody can consult as a library. Taken to court and, on that specific point, won.

There's a progression in those three lines, and it isn't comforting. The difference isn't "they cut the books." It's what they cut them for. Google built a library you can walk into. The Internet Archive tried to build one you could walk into, and that's why it lost. Anthropic built one nobody walks into, and won partly because of it: the material never leaves, so the use stays internal, and internal use is defensible.

So the honest version isn't that AI destroys books. It's that digitization stopped being preservation and became acquisition. The book wasn't destroyed, which is worse: it was dissolved. It's inside the model, diluted into some millionth of a weight, with millions of others. You can't read it, quote it, or verify it's in there.

It isn't an archive. It's a memory.

And the paradox that keeps you up: if the digital copy really does "replace" the physical one, then in twenty years, for some out of print twentieth century title, the only place on earth where that text still exists will be inside the weights of a private model. Which will never give it back, because it wasn't built to return books. It was built to have read them.

One last piece, flagged as my own hypothesis. This market rewards whoever owns organized, cataloged, licensable corpora. Italian publishing is fragmented, catalogs are small, the Italian public domain is digitized badly and in patches. The risk isn't censorship. It's more banal and harder to fix: Italian ending up underrepresented in the models through contractual disorder. Not because anyone excludes us, but because buying us is too complicated compared to buying something else. If you work in a language that isn't English, this is your problem too, and nobody is thinking about it.

Subscribe now

Three things to take away.

First: when absurd behavior becomes systematic, look for the rule that rewards it, not the people doing it. True for sliced books, true for your operations department.

Second: your archive changed nature. Human, dated, never published online is now the hardest material in the world to source, and you have ten years of it in a folder you were about to empty.

Third, the one that counts: the market is already past text. What's worth two hundred dollars an hour today isn't what you know how to do. It's how you judge whether something is done well. If you've never written it down, nobody in your company knows it except you. And when you leave the room, it leaves with you.

See you at the next cup.

P.S.: genuine curiosity, not rhetoric. How many years of conversations with your clients do you still have somewhere, and what shape are they in? Write back, I'm collecting cases.

Share

☢️ Matteo

• MIT reran two real 2019 studies with GPT-4 in the humans’ seats.

• Blind expert judges rated the AI’s answers deeper than the real ones.

• The winner wasn’t AI or humans. It was the pair.

Welcome back to the Counter, survivor number 4721.

The coffee was still hot when I read the line: “We replaced the respondents, the analysts, and the human moderators with AI.” Not some TikTok prophet.

A paper out of MIT. One real marketing study, two experiments, a Fortune 500 brand, one question with teeth: can AI actually do our job?

Short answer, yes. In places, better. It didn’t just fill out a questionnaire. It analyzed data, generated insight, found patterns, even proposed strategy. Seconds to do what a full team burns days on. Which is when it clicked for me: the future isn’t “AI vs. marketers.” It’s AI with marketers.

What MIT tested in a lab, we already live every day inside Morfeus.

Keep one thing in mind for the whole read: the study ran on GPT-4.

Old news by today’s standards, two model generations back. Whatever you have open right now, ChatGPT, Claude, Gemini, is sharper than the thing that pulled this off. So read what follows as the floor, not the ceiling.

☕ The Vault this week holds the part that doesn’t fit here: our real workflows, examples, and prompts for rewriting how you do marketing, with an assistant that never sleeps. Ready to see what happens when the machine starts understanding the customer better than the people who study them?

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“Every day we work beside an assistant nobody sees. And somehow it makes us look more human than before.”

🧪 AI in Action: The Day GPT-4 Walked Into a Focus Group and Never Left

There’s a paper circulating in the basements of academic marketing, signed by three minds at MIT. It isn’t about the future. It’s about a present we’d rather not admit.

They took two old studies from a food multinational, the kind built on interviews, questionnaires, human analysts, and mountains of time, and ran them again. Only this time, in the humans’ seats, they put GPT-4. Not AI as a tool. AI as participant, analyst, and in some runs, moderator. An invisible colleague at the table.

“Can it really understand people, or is it just good at faking it?” That was the question. The answer, I’ll warn you now, is a little unsettling.

☕ In this week’s Vault I show you the part that doesn’t fit here: how we at Morfeus already use AI to rewrite marketing, from ads to surveys to courses and campaigns. The hands-on version of what MIT just proved in a lab.

Subscribe now

How They Got GPT-4 Into a Focus Group

The authors, Timoshenko, Mao, and Hauser, didn’t speculate. They took two real marketing studies a large food company commissioned in 2019 and rebuilt them identically, with one change: people out, GPT-4 in.

The first was qualitative, the kind you run to learn how people actually think, not just what they buy. The topic was Friendsgiving, the American ritual of Thanksgiving with friends: recipes, emotions, small cultural meanings. Back in 2019 the brand had interviewed dozens of people, recorded hours of talk, and dissected words and gestures to build more empathetic campaigns.

The second was quantitative and colder: a product test to see which pet food package would win between two options. Same sample size, same questions, same scoring.

Now picture it. The MIT team sat down in front of GPT-4 and handed it the same brief from four years earlier. They defined personas, age, interests, habits, tone, and asked the model to answer as if it were that consumer. A single thirty-something who loves to cook. A mom with two dogs and no time. A kid who throws a Friendsgiving just to feel part of the group.

GPT-4 answered every question, one after another, with an almost eerie consistency. And it didn’t stop there. In some runs it was set up to moderate: ask questions, rephrase them, run a full interview like a real researcher. In others it played analyst, reading transcripts and pulling out the key insights, keywords, and themes on its own.

They built three modes:

• All human, like 2019.

• All AI, GPT-4 running every step.

• Hybrid, human sets the method, model does the operational work.

The real question underneath: how much of our marketing work, writing questions, interviewing, analyzing, synthesizing, can AI do today? And how much time would we free up if the machine learned to think alongside us?

Subscribe now

The AI That Talks Like Us (and Sometimes Better)

When they dropped GPT-4 into the focus group, everyone expected generic answers: well written but cold, missing the emotional texture that makes qualitative research worth anything.

The opposite happened.

The model’s answers were clear, coherent, and, according to many human raters, deeper and more informative than the real ones. Stories, details, connections the actual participants often skipped. Where a person stalled with “I don’t know,” the AI kept reasoning, linking ideas, describing feelings.

The researchers didn’t stop at first impressions. They handed dozens of interviews to a panel of independent experts and asked them to score everything blind, with no idea which came from people and which from GPT-4. Result: no meaningful difference in clarity or relevance. And in two categories out of four, depth and informativeness, the AI came out ahead.

Then the most fascinating test. They asked GPT-4 to be the analyst, not the interviewee. Read the transcripts, highlight the key lines, cluster the concepts, surface the main themes. The exact job an experienced human researcher spends hours on.

The result stunned the room: GPT-4 recovered nearly 90% of the same themes the humans found. And, more interesting still, it surfaced a few new ones, small nuances the professionals had missed.

And remember which model we’re talking about. This was GPT-4, the “old” one. The version on your screen today does this faster, cleaner, and with less hand-holding.

In practice it behaved like a research assistant with perfect memory and zero fatigue. Read everything, forgot nothing, handed back a clean, ordered analysis ready for a human brain to refine.

When the expert judges compared outputs, one finding stood out: the best work wasn’t “human only” or “AI only.” It was hybrid, the machine prepping the ground, the human adding the final read.

The MIT team called it a hybrid human-AI pipeline. We’d just call it the first real duo behind the Counter: one that thinks, one that never tires.

Subscribe now

When AI Becomes a Sample of Synthetic Consumers

After making GPT-4 talk like a respondent, the researchers turned it into a whole sample of them. Same brand, same goal, but now the terrain was numbers: quantitative research.

In 2019 the company had surveyed more than six hundred people to see which pet food package would land better. One was a “sliceable” tube, innovative, premium-looking. The other, a classic resealable bag. The usual questions: “Do you trust it?” “Does it feel high quality?” “Would you buy it?”

In 2025 MIT rebuilt the whole thing with Synthetic Consumers: GPT-4 generating hundreds of answers consistent with the target’s demographics. Each profile had age, income, habits, even simulated food preferences, built on a statistical base.

The result? Surprisingly, the average trends matched. When humans preferred package A, so did the AI. When they rated a design more trustworthy or appealing, GPT-4 scored it about the same. The model caught the direction of human behavior.

But one detail gave away its artificial nature: no variability. The synthetic consumers tended to think alike, with little spread in their answers. Their opinions were orderly, logical, clean. Too clean.

To fix it, the researchers added context: samples of real human answers, historical data, pieces of qualitative interviews. And the model shifted. The answers got more varied, more realistic, more “messy” in the right way. The distribution started to look like a real survey.

The lesson is simple and sharp: AI grasps the trends, but it needs real stories to remember how people reason. Give it numbers and it’s an accountant. Give it context and it starts acting like a consumer.

Even machines, like us, think better once they’ve lived a little.

The New ROI of Market Research: Time and Depth

Somewhere in the paper there’s a number that looks mundane and says everything. A human analyst takes forty, forty-five minutes on average to read and synthesize a single qualitative interview. Almost an hour of attention per voice in the sample.

GPT-4 does it in seconds.

The authors didn’t clock the total savings, but they call it an “efficiency gain.” An elegant way of saying a marketing team can compress days of work into hours.

Time, here, isn’t just a saved resource. It’s a multiplier: more ideas tested, more creative hypotheses explored, more brand messages validated, without trading away quality. You stop choosing between “go deep” and “go fast.” Now you do both.

And there’s a quieter side effect that matters even more. When AI takes the mechanical work, humans go back to thinking. To spotting connections, reading signals, asking better questions.

The paper doesn’t say it, but anyone in marketing gets it instantly: this isn’t just efficiency. It’s a new way to use time as strategic leverage. Espresso time. Concentrated, served hot.

Subscribe now

From the Counter: Marketing at Espresso Speed

Reading this study, I realized MIT wasn’t describing the future of research. It was describing my work week.

Inside Morfeus the same thing happens, except we don’t call it an academic experiment. We call it Monday morning.

And here’s the part that should stop you. That paper ran on a two-year-old model, in a controlled lab, with researchers babysitting every step, and it already matched a room full of professionals. Now look at what’s on your screen today and try to picture what the same experiment would produce. You can’t. I can’t either. The distance between “the AI kept up with the humans” and what these models actually do right now is wider than anyone says out loud. That study isn’t the ceiling of what’s possible. It’s the oldest, slowest version of it.

What the professors tested in a lab, the hybrid human-AI model, is just our normal now. Different tools, different contexts, identical logic: we set the direction, the machine does the rest.

And I mean the machine does the rest. At Morfeus the marketing isn’t “assisted” by AI. It’s generated by it. The ads, the landing pages, the email flows, the surveys, the campaign angles: the machine produces the entire first pass, end to end. We don’t start from a blank page anymore. We direct, we taste, we cut. The human decides what’s good. The model makes all of it. And yes, a good part of what reaches you from behind this counter goes through the same machine first.

This isn’t automating marketing. It’s accelerating it without hollowing it out. Keeping the human thinking, stripping out the useless slowness. The taste stays with us. The labor moves to the model.

Subscribe now

The New Researcher Is Already at the Counter

In the end, what MIT calls a hybrid human-AI approach is just how most of us work now: a constant passing of cups, ideas, and calculations between brains of flesh and brains of silicon.

AI didn’t steal anyone’s seat. It just moved the Counter a little further forward.

It doesn’t sleep, doesn’t get distracted, doesn’t need a vacation. And yet, to actually work, it needs us: the doubt, the intuition, the small human error that so often leads to the right answer.

The new researcher isn’t an algorithm crunching data. It’s the invisible pair that forms when a human leads and a model amplifies. Thought and calculation. Instinct and speed.

And while the big universities start measuring the efficiency of this new duo, we, here at the Counter, just keep pouring espresso and insight into the same cup.

Because in the end, marketing today comes down to one thing: learning to Think in Two.

📌 Espresso Prompt: Run Your Own Synthetic Focus Group

Need to know how your audience would react to a message, a product, or a campaign? You don’t need a thousand-person panel. You need an LLM and a well-built prompt.

With this you can spin up a mini synthetic focus group, test perceptions, compare opinions, and pull insight in minutes.

🧩 The Prompt

Paste this block into ChatGPT, Claude, or Gemini. Then add your marketing question and your product context.

Act as a marketing researcher running a preliminary test on a new product or ad message.

1. Generate 10 realistic synthetic consumer profiles (age, gender, occupation, income level, lifestyle, and values). Make them diverse and consistent with the target market.

2. For each profile, answer this marketing question:
[INSERT YOUR QUESTION HERE, e.g. “What do you think of this new protein-snack packaging?”]

3. Every answer must be personal, spontaneous, and consistent with the consumer’s profile. Avoid generic language: I want to hear a real person’s voice.

4. At the end, build a table that summarizes: the 3 common themes across consumers, the main differences in perception between profiles, and one final insight for whoever designs the campaign.

Final format: a summary table, then 5 lines of closing synthesis.

☕ How to use it

Ideal for pressure-testing a concept, a landing page, an ad message, or a headline before it goes live. Add realistic context (”target: women 30-45, fitness lovers, urban area”) for more varied, natural answers. Want more depth? Ask: “Add a paragraph where each profile explains the why behind its emotions.”

💬 From the Counter

In the MIT paper it took months to organize focus groups, analyze interviews, and interpret data. You can do it over a coffee. Not to replace people, but to know where to look first.

☢️ Radioactive Humor

Marketing isn’t getting easier. It’s just getting more honest: whoever thinks fast, whoever fuses data and instinct, whoever can hold a conversation with the machines, wins.

AI didn’t take the craft away from us. It just reminded us the craft was always about thinking better than the automatics.

Out there they’ll keep saying technology will replace us. But anyone who actually does the work knows it isn’t a threat. It’s an invitation. To become more human than before.

Subscribe now

Meanwhile, here at the Counter, we keep mixing the two base ingredients of the new marketing: instinct and silicon.

And every time the model speaks... we pour another espresso.

See you at the next cup.

Share

☢️ Matteo

• 21,559 firms tracked: the heavy AI adopters hired more, not less

• The real risk isn’t AI. It’s a company that half-uses it

• Which side of the intensity line are you standing on?

Welcome back to the bunker.

You’ve heard the phrase for months, delivered like a weather warning: “AI is taking our jobs.” It’s on the news, and some nights you think it too, closing the laptop and wondering where you’ll be in two years. It’s a legitimate fear, and I’m not here to pet it. I’m here to show you a number that just dropped and splits that phrase in half.

First, where I’m speaking from, so this lands right. In the last year I’ve trained more than 1,000 people and walked into over 30 companies to put AI inside real work, on top of using it every day myself. And one thing I’ll guarantee without blinking: the AI revolution is 10% technology and 90% humanism. It’s not the models that make the difference. It’s the people. They matter far more than any tool you can buy, and that’s exactly what today’s numbers are about to tell you.

At the end of June the most serious study yet landed on the link between AI spending and jobs. Not a survey, not a LinkedIn opinion: 21,559 US companies, tracked month by month, cross-referencing what they actually spend on AI against how many people they actually hire. The result flips the fairy tale: the companies betting seriously on AI hired more in the two years after. Not less. More. And the twist that short-circuits everyone: they hired more young people.

This isn’t news to forward. It’s a map. It tells you which side of a line you want to be on for the next 24 months, and that line is being drawn right now, while you read.

☕ Cup in hand. Let’s begin.

Full paper (primary source, with methodology and numbers):

https://ramp.com/data/ai-jobs-impact/paper

Ramp summary page (more readable, with graphs):

https://ramp.com/data/ai-jobs-impact

Revelio Labs reading (your job data partner, with caveats):

https://www.reveliolabs.com/news/ai-and-work/greater-ai-investment-more-hiring/

Subscribe now

💡 Thought of the Day

“AI doesn’t split those who work from those who’ll be replaced. It splits those who use it seriously from those who barely touch it. And only the second group is in danger.”

🧪 AI in Action: Why “AI Kills Jobs” Is Only Half the Story

#1: The Number Nobody Expected (and why it’s about you)

Let’s start with the numbers, because for once they’re clean.

The study is by Ramp, which processes payments for tens of thousands of companies, with Revelio Labs, which tracks workforce data. The method is the genius: instead of asking “do you use AI?” and trusting the answer, they measured what each company actually spends on AI tools, straight from corporate cards and real payments, then matched it against headcount, month by month, across 21,559 companies, from January 2021 to February 2026. Observed spend, not declared. The first time anyone tied the real money in AI to real jobs.

Here’s what fell out. High-intensity adopters, the ones going all in, grew headcount 10.2% over the next 24 months. The ones barely touching it gained nothing meaningful: flat, a barely-there minus 0.6%. Same tool on both tables. Opposite result.

And the number that should straighten your spine, because it contradicts the most common fear: that AI hits the young first. In high-intensity companies, entry-level hires grew 12%, and their share of the workforce rose by over a point. In low-intensity companies, the slice of young people shrank. Read it again: where AI is used more, young people get in more. Where it’s half-used, they get in less.

Why this is about you. The variable that decides your professional future isn’t “does AI exist.” That fight is over. The variable is: are you inside, or near, an organization that uses it seriously, or one that keeps it on the desk like an ornament? A question about you, one you can answer, and act on by Monday.

Subscribe now

#2: The Hole in the Apocalypse Story

The “AI destroys jobs” narrative is popular precisely because, told this way, it’s incomplete. And letting it be told badly costs you.

It has a manufacturing defect: it treats AI like weather. Something that rains on you and you just take it, like hail on a crop. You have no active part, you’re the designated victim of a bigger force. Comfortable, because it strips you of responsibility: if the job vanishes, blame the machine, not how you or your company used it or ignored it.

The data tells a more uncomfortable story. AI doesn’t rain evenly. It doesn’t just subtract or just add. It redistributes, along one line: on one side those who integrate it all the way, on the other those who half-use it or snub it. In the first, work grows, even for beginners. In the second, it stagnates. It’s not AI against workers. It’s those who ride it against those who watch it pass.

From behind the counter, straight: the real danger to your career isn’t a company that adopts AI. It’s one that pretends to. The kind that buys three licenses, lets two people use them on Friday afternoon, and quietly stops hiring because “we’ve got AI now.” It lowers you into the future slowly, under the excuse of saving money. The riskiest place you can stand today isn’t where AI arrives. It’s where it arrives halfway.

That’s why this data isn’t a pastime. It changes the questions you ask in an interview, how you see the company you’re in, where you invest your next two years. Those who read this and use it, choose. Those who ignore it, get chosen. Same threshold, every time.

Subscribe now

#3: The Intensity Threshold (the only thing that matters)

One concept to stick, the pivot of the whole issue: the intensity threshold.

The study captures it brutally. There’s no single category called “companies that use AI.” There are two species that behave in opposite ways with the same tools. One crosses the threshold, puts real budget in, restructures how they work around these tools instead of gluing them on top of the old process. The other stays below, buys a few subscriptions to say they’re in, then continues exactly as before. The first grows 10%. The second stays flat. The difference isn’t the tool. It’s the intensity.

Here’s my image, from behind the counter. Low-intensity AI is like joining a gym in January and going twice. You have the card, you spent the money, you can say you “signed up.” But your body doesn’t know it, because the body doesn’t respond to the membership, it responds to the training. Low-intensity companies have the AI card and last year’s physique. Then they’re shocked nothing changed and conclude “AI is overrated.” No. You went twice.

Now apply it. Look at the company you work for, or your own, and answer honestly: above or below the threshold? Is AI a new way of working here, or a toy three people use in secret while the official process is stuck in 2022? It’s the question that predicts whether, over two years, that place hires and grows, or cuts slowly and blames the times. The threshold isn’t decided by fate. It’s decided by whoever controls that budget. And if that’s you, you just read the right call.

Subscribe now

#4: The Honesty About the Numbers (the part that earns your trust)

Now the part a guru would skip, because it ruins the hype. I’ll give it to you, because without it the numbers above are propaganda, and I don’t serve propaganda.

The study has a big limit, and the authors admit it first: it’s correlation, not causation. It isn’t proven that AI is what creates those jobs. The heavy AI spenders were already different before: bigger, more technical, faster, often venture-backed, already growing at a median 6% versus 1.6% for the rest. Already racehorses. So maybe it isn’t AI making them run, but running companies buying more AI. The classic chicken or egg.

And the growth is concentrated mostly in one sector, information and tech. The study only counts heads, not pay, not whether the work is better. Quantity, not quality. Full honesty: this study alone doesn’t close the case.

But here’s why you need it more precisely because of these limits. Even in the most cautious reading, where healthy companies simply use AI more, the lesson doesn’t change, only the direction does: you want to be on the side of companies that use AI hard, because that’s the mark of the organizations winning right now. Cause or symptom, the right place is the same. High intensity is either what makes you grow, or the proof you already are. Either way, that’s where you want to be.

Subscribe now

#5: Which Side of the Line You’re On, and How to Cross It

Last one, the most practical, because you need a move.

Months ago, in one of the harder issues, I told you where roles disappear: the middle coordination work AI eats first. This is the flip side. AI doesn’t add or remove work in the abstract. It moves value across a threshold. From those who execute and coordinate, toward those who decide, direct, build. Your only job that counts is to know which side you’re on and, if it’s the wrong one, move while you still can.

The diagnosis we run on ourselves. With Morfeus and Scalers we’re a high-intensity company by definition: AI isn’t a tool we use, it’s how we work, from the first line of code to this newsletter. And the result is exactly the study’s: we’re not cutting, we’re building. Cross the threshold, you grow.

And if that 12% entry-level felt abstract, here’s where I watch it come true daily: just look at our team. All very young. I didn’t pick them for being the most talented on the planet, I picked them because they have a hunger to build that can’t be taught, and in a high-intensity company that hunger beats ten years of experience on autopilot. How we build that culture is too long a story for today, and one day I’ll tell it in full at the counter. For now, keep this: AI isn’t sweeping the young away. It’s calling them in, hired by whoever uses it seriously.

So here’s what you do, today. Stop asking whether AI will take your job, that’s the losers’ question, it casts you as the victim. Ask the winners’ question: how do I, or my corner of the world, become a high-intensity adopter in the next 90 days? Which process I do by hand could I rebuild around AI from zero? Where am I holding a gym card without training? The threshold is a choice, and choices get made. Those who stay below almost always do it for lack of decision, not tools. You hold the decision now. Take it.

Subscribe now

📌 Espresso Prompt

Find out which side of the threshold you’re on, and how to cross it.

This prompt doesn’t make you produce content. It runs a diagnosis, far more valuable and something nobody stops to do. It tells you, data in hand, whether you or your company are high- or low-intensity, and hands you the plan to level up. Works with Claude or ChatGPT, better inside a Project that already knows your work and goals. Paste this, then answer honestly, because on this you only cheat yourself.

“Act as a senior operational consultant specialized in AI adoption, tough and on my side. Your job isn’t to make me feel ahead, it’s to tell me the truth about how much I actually use AI and what I’m missing to use it seriously. Analyze my situation [e.g. I’m a freelance professional, or I work in an 8-person marketing team, or I run a 30-employee company] in three phases, one at a time.

PHASE 1, DIAGNOSIS. Ask me the right questions to tell if I’m a high- or low-intensity adopter: how many processes I’ve truly rebuilt around AI versus how many just use it as an occasional shortcut. Then give me a blunt verdict: high intensity, low intensity, or just a gym card.

PHASE 2, THE COST OF STANDING STILL. Tell me honestly what happens to me over the next 24 months if I stay at today’s intensity. No moral, just the concrete consequence on my work and my opportunities.

PHASE 3, THE PLAN TO CROSS THE THRESHOLD. Give me the three processes I do by hand that would have the highest impact if I rebuilt them from zero around AI. For each, the first step I can take this week, not in six months.”

⚠️ The value of this prompt isn’t the diagnosis, it’s what you do the day after. Knowing you’re low-intensity and not moving is worse than not knowing, because now it’s a conscious choice to stay behind. Use it to find the one process to change on Monday, and change it.

Subscribe now

☢️ Radioactive Humor: Same Machine, Two Doors

Two doors, side by side, in the bunker corridor.

On the first, a buzzing neon sign: “AI TOOK OUR JOBS.” Inside, dark, an empty chair, dust. Nobody in, nobody out.

On the second, lights on, people rushing through, a handwritten sign taped up crooked: “HIRING, EVEN JUNIORS.” Above the door, the exact same machine as in the first room.

A kid with a resume in hand looks at the two doors. He doesn’t know they’re the same office, six months apart, depending on who used it.

Subscribe now

Both rooms have the same machine. The only difference is who believed in it enough to use it for real.

Fear makes you stare at the first door. The data tells you to knock on the second. And the second, today, is hiring.

☢️ Matteo

Share the bunker. If you know someone living with the anxiety of AI taking their job, send them this issue. They don’t need reassurance, they need a map: the risk isn’t AI, it’s staying on the wrong side of a line they can still choose. Better a coffee today than a rude awakening tomorrow.

Share

See you at the next cup of AI Espresso.
Matt ☕

• Anthropic and OpenAI both filed to go public, both near trillion-dollar valuations.

• Altman proposed handing 5% of OpenAI to a US sovereign fund.

• Why a regulator that’s also a shareholder is “structurally compromised.”

Welcome back to the bunker. Vault tier, where things get said in full.

I promised you this last issue: the two S-1 filings, and the government angling for a stake in the labs building AI. Too big for a paragraph. Here it is, in full.

Two weeks ago, Anthropic and OpenAI both filed to go public in the same window, both pushing toward trillion-dollar valuations. No precedent in tech history: the two most important labs on the planet go public at once, with everything that implies. Then, while the race still looked like a photo finish, a second thing happened: Sam Altman proposed handing 5% of OpenAI to a US sovereign wealth fund, and asked the other major labs, Anthropic included, to do the same.

Take them separately and they’re two finance stories. Put them side by side, three weeks after the same government proved it could switch off a model with a letter, and they become one story: power over AI is institutionalizing, and the state no longer wants to just referee. It wants a seat at the table.

The question I’m asking you today isn’t whether to invest. It’s a different one, and I’ll leave it for the end.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“When your regulator becomes your business partner, you don’t have a referee anymore. You have a player wearing the stripes.”

🧪 AI in Action

#1: Two S-1 Filings, Two Stories That Just Diverged

Let’s line up the numbers, because in the past few weeks the “simultaneous” race stopped being simultaneous.

Anthropic filed its confidential S-1 on June 1st, with an annualized run-rate of $47 billion as of May 2026. A year earlier it was at $9 billion. Two years earlier, $87 million. It’s the fastest growth ever recorded in enterprise software. The valuation from its last private round, the $65 billion Series H-1, sits at $965 billion, and in Q2 2026 the income statement says something no AI lab had ever been able to say: a quarterly operating profit of $559 million. Not “approaching profitability.” Done, signed, filed with the SEC.

From there, Anthropic kept moving straight ahead. Underwriters, media, and prediction markets converge on a Nasdaq debut in October 2026, with Goldman Sachs, JPMorgan, and Morgan Stanley leading an offering expected to raise over $60 billion. On the secondary market, the implied valuation has already climbed to $1.05-1.15 trillion. Independent research firm FutureSearch projects a median first-day market cap of $1.10 trillion: a 14% premium over the last private round.

OpenAI filed its own S-1 a few days later, valued near $852 billion, with $2 billion a month in revenue. It sounds like the same story. It isn’t: OpenAI loses $1.22 for every dollar it brings in, projects a $14 billion loss in 2026, and doesn’t expect to turn a profit before 2029.

Here’s the gap that didn’t exist three weeks ago. After SpaceX’s turbulent public debut cooled the market’s appetite for spectacular listings, OpenAI is reportedly weighing a delay to 2027. Anthropic isn’t: it’s accelerating, holding the line on October. The photo finish just split into two different trajectories, two different degrees of urgency.

I’m not saying this to settle who “wins.” The contrast gives you a map. Anthropic built on enterprise customers who pay well and often, and can now afford to run toward the public markets from strength. OpenAI bet on volume, on the consumer, on the mass-market product: a huge bet on a market that’s growing but still burning cash, forcing it to buy time right as public markets turn more skeptical.

Two legitimate business models. Two opposite risks. And, as you’ll see in the next section, two different kinds of pressure coming from shareholders.

Subscribe now

#2: The Moment the Labs Start Answering to You (Almost)

Until yesterday, Anthropic and OpenAI answered to private investors: Google, Amazon, Microsoft, venture funds. A tight circle, aligned to long-term growth, used to burning cash for years before any return.

Going public blows that structure apart: tens of thousands of shareholders, analysts tearing you apart every quarter, pressure on the stock price, growth expectations you can’t disappoint twice running.

For you, someone who uses these tools to work, the transition has three concrete consequences.

The first is price. Public companies optimize margins harder than private ones. Freemium, flat pricing, promotions to acquire users: harder to sustain when you need steadily growing revenue per share. Prices won’t necessarily jump tomorrow morning, but the wind has shifted, and you saw it with the Fable credits I told you about two issues ago: first included limits, then metered usage. Same wind.

The second is safety. Until now, pressure on model safety was mostly internal or regulatory. With public shareholders, every incident becomes class-action material. That can push in two opposite directions: more caution, which is good, or more rushing to ship features before they’re ready, which isn’t. Public markets reward quarterly growth, not prudence.

The third, the one almost nobody names, is forced transparency. Public companies have to disclose publicly, with numbers certified by independent auditors. For the first time, you’ll get real data on what these models cost, how fast they’re growing, where they’re losing money. The era of self-promotional press releases with hand-picked metrics is ending, not out of virtue, but out of legal obligation.

The lab your work depends on is about to become a public company. And right in the middle of that transition comes the twist that changes how you read this whole story.

Subscribe now

#3: 5% to the State. And the Part No Press Release Mentions

In early July, Sam Altman proposed something no AI lab had ever put on the table: handing over 5% of OpenAI, roughly $42.6 billion at the current $852 billion valuation, to a US sovereign wealth fund. Not just OpenAI: the idea, discussed directly with President Trump, Commerce Secretary Lutnick, and Treasury Secretary Bessent, calls for every major lab, Anthropic, Google, and Meta included, to donate similar stakes.

The reference model is the Alaska Permanent Fund, which has distributed an annual dividend to Alaska residents for decades, funded by the state’s oil revenue. Applied to AI, some circulating estimates put it at around $320 per household. The stated reason, in the Financial Times’s words, is to “secure good relations with the administration and address political backlash.” Translated from diplomatic language: buying political peace before someone forces it on you.

Here’s the detail that makes the move less generous than it sounds. Talks remain preliminary, and any formal action would require Congressional approval. But the political reaction has already arrived, split in a way worth a closer look.

On the right, some are calling it plain socialism: the state owning a stake in a private company was never a conservative principle. On the left, the verdict is the opposite: not enough. In June, before Altman’s proposal, Senator Bernie Sanders had already floated a far more radical idea: a one-time 50% tax on AI companies’ shares, with the proceeds deposited into a public fund. By comparison, Altman’s 5% looks like a defensive offer, calculated to get ahead of something bigger before it arrives.

But the point I want stuck in your head isn’t the percentage. It’s the conflict this move creates, one that Public Knowledge analyst Nat Purser framed better than anyone: a government that simultaneously holds regulatory authority over a company and a financial stake in its success is, in her words, “structurally compromised.”

Subscribe now

Stop on that sentence, because it’s where this story connects to the last two issues. Three weeks ago, the Commerce Department proved it could shut down a frontier model with a letter, citing national security. Today, that same government is negotiating to become a shareholder in the companies subject to that power. Two different forms of power over the same object: one is the power to say no, the other is an interest in the yes paying off. When they land in the same hands, the question of who decides, and in whose interest, stops having a simple answer.

This isn’t a policy-wonk detail. If the state holds a stake in Anthropic, the next time it weighs shutting down a model for security reasons, it’s also weighing the hit to its own stake. The regulator and the shareholder, in the same head. Not necessarily worse decisions. Less legible ones, and that’s exactly the opacity the forced transparency of going public was supposed to reduce, not increase.

Subscribe now

#4: The Trillion-Dollar Question, and It’s Not the One You Think

Back to the question I promised you at the top.

It’s not “should I buy Anthropic or OpenAI stock?” I’m not a financial advisor, and it wouldn’t help you if I told you what to do with your money. That conversation, if you want to have it, is between you and your accountant.

The question I care about lives in the gap between two numbers.

On one side: valuations pushing toward a trillion dollars, growth from $87 million to $47 billion in just over two years, a government inviting itself into ownership with a stake worth tens of billions. On the other: most companies, in Italy and beyond, still use these tools like a slightly smarter search engine. Open it, ask something, get an answer, close the tab. No integration into processes, no agents, no system. AI as an occasional luxury instead of an operating lever.

That gap isn’t an accident. It’s an opportunity with legs, and now there’s one more indicator: when even a sovereign state buys a piece of an AI lab, it’s saying, with actions not words, that the value of these tools in the hands of people who use them well is already priced into the market. Nobody invests tens of billions betting that people will keep using AI to write slightly better emails.

The right question isn’t whether to buy a stock. It’s this: in your company, in your work, are you on the right side of this gap, the side that captures the value, or the other one, the side that absorbs it?

Because if the answer is “I don’t know,” the double IPO and the 5% to the state aren’t just financial news you’re watching from the sidelines. They concern you as a competitor to whoever, in your own industry, has already claimed that side.

Subscribe now

📌 Espresso Prompt

Audit the gap.

This issue’s exercise is simple and uncomfortable. It takes five minutes to do and probably a few days to digest.

Open a conversation with Claude. Paste this.

“You’re a strategic consultant with access to everything you know about my industry. I want to understand how exposed I am to the competitive edge of people who use AI better than I do. Let’s go through three questions, one at a time.

QUESTION 1. Tell me what industry and what kind of work I operate in, and tell me the three operational activities where AI is already changing how people compete, not in the future, now.

QUESTION 2. For each of these three activities: if a competitor of mine were already running them with an integrated AI system while I still do them manually, what concrete advantage would they have over me? Quantify it wherever you can: time, cost, speed, quality.

QUESTION 3. Give me one question, the most uncomfortable one, that I should ask myself this week to find out whether I’m on the right side of this gap or not.”

The output isn’t an action plan. It’s a diagnosis. And uncomfortable diagnoses, taken seriously, are the starting point of every move that matters.

Subscribe now

☢️ Radioactive Humor

A mailbox, an ordinary American family, 2027.

“An envelope came from the government. It says: congratulations, you’re a shareholder in OpenAI.”

“Great! How much do we get?”

“$320.”

[silence]

“And what about the job GPT-6 took from me last month?”

“That’s not covered by the fund. But if you want, with the $320 you can buy three months of Plus to apply better for whatever job you’ve got left.”

Three things to take away from this issue.

First: the two IPOs aren’t a simultaneous race anymore. Anthropic accelerates toward October from strength, profitable and growing at a record pace. OpenAI slows toward 2027, with billion-dollar losses and a warier public market after SpaceX’s turbulent debut. Knowing who your tools depend on, and how healthy they are financially, isn’t investor trivia anymore.

Second: the 5% proposal isn’t generosity. It’s political insurance bought in advance, and it creates a structural conflict: the same state that could shut down a model three weeks ago is now negotiating to profit from it. A regulator turning shareholder isn’t a technical footnote. It’s a change in the nature of the power looming over every tool you build on.

Third, and this is the point: trillion-dollar valuations and the state’s rush to get in on them aren’t your problem as an individual. They’re an indicator. The market has already priced in the transformation. The question today’s prompt helps you ask is whether you’re on the side capturing it or the side absorbing it. It’s not too late to move. But the window won’t stay open forever.

Welcome back to the bunker. See you soon.

P.S.: one thing I want your honest answer on. A government that regulates AI and profits from it at the same time: does that strike you as a real problem, or are you already assuming that’s just how every major technology works now? Write back. Honest reactions to this question tell me everything.

☢️ Matteo

Subscribe now

Share the bunker. If you know someone who still thinks the AI labs’ IPOs are just financial news, forward them this issue: the conflict between regulator and shareholder concerns them a lot more than they think.

Share

• Fable 5 is back online worldwide, after 19 days of government-ordered darkness.

• It returned changed: silent guardrails, a reroute you won’t notice, credit-based pricing.

• Three questions to ask about your stack this week. Before you need them.

Welcome back to the bunker.

I left you with the lights off: the most powerful model ever released to the public, gone because of a letter, and one uncomfortable question hanging in the air. How many squares is your work built on?

On July 1st the lights came back on. Fable 5 is live again, for everyone, everywhere. If you stop at the headline, the story ends well: the government reconsidered, the model returned, all back to normal.

Except it’s not the model that left. It came back with a new leash stitched into its architecture, a new price tag, and a precedent nobody can uninvent: we now know how a frontier intelligence gets switched off. And, more interesting, how it gets switched back on.

Subscribe now

The last issue was the first half: the floor can vanish. This is the second half, where you learn the real rules of the game. Who decides, what changed, what it costs, and what you should actually do about it.

☕ Cup in hand. Let’s go.

💡 Thought of the Day

“A tool isn’t yours because you pay for it. It’s yours to the degree you survive the day someone changes it without asking you.”

🧪 AI in Action

#1: The 19 Days: Anatomy of a Relaunch

The facts, in order. On June 9, Anthropic released Fable 5 and Mythos 5. On June 12, the Commerce Department letter arrived: export controls under national security authority, and Anthropic pulled the plug on everything. The trigger: Amazon researchers, testing as partners, had found a way around the safeguards to make the model identify software vulnerabilities, down to writing exploit code for one of them.

What happened next is the part that had never happened before.

Anthropic didn’t go to war. No lawsuits, no headlines, no crusade. It did something more boring and more effective: it negotiated. Weeks of “productive conversations” with Washington, which, translated from diplomatic language, means: we sat down, we showed countermeasures, we found a compromise. Meanwhile over a hundred cybersecurity leaders, Katie Moussouris among them, had signed against the technical basis of the order: that bypass, they said, never justified an export control.

Note Amazon’s role here, because it makes the story less linear than it looks. The actor that triggered the letter wasn’t an enemy. It was Anthropic’s biggest investor, its cloud provider, one of six early-access partners. Frontier AI works like this: the same players are simultaneously shareholders, customers, competitors and, when needed, witnesses for the prosecution. That web of crossed interests is part of any tool’s stability profile, just like the benchmarks.

On June 30 the government lifted the restrictions. On July 1, Fable came back globally: API, Claude.ai, Claude Code, Cowork. Mythos 5, the unfiltered sibling, is restarting slower, through vetted organizations and the Glasswing program. Nineteen days, total. Sources: Anthropic’s official announcement, Al Jazeera’s coverage of the reversal.

Also worth noting what did NOT happen in those nineteen days: no incident caused by Fable, no proof the flagged capability was unique, no new evidence backing the original letter. The case closed without the accusation ever landing its punch. The hundred signatories were right.

But don’t read that as “false alarm, back to normal.” It’s the opposite. Your ERP was never switched off by a ministry. Your spreadsheet was never the object of a negotiation between a software company and a government. For frontier AI, the full cycle, shutdown, negotiation, conditional relaunch, has now happened once, end to end. And things that happen once in full stop being unthinkable. They become procedures.

The precedent now cuts both ways: a frontier model can be switched off in one evening, and switched back on after a deal. The tool you use every morning is officially something a company and a state negotiate over.

And you weren’t at the table.

Subscribe now

#2: The New Leash: What the Guardrails Tell You About Every Future Model

Now look inside the model that came back, because the difference isn’t cosmetic.

Fable 5 was redeployed with a new set of classifiers targeting offensive cybersecurity tasks. Anthropic says the classifier blocks the reported technique in over 99% of cases. And when a request gets flagged, you don’t get a refusal: the request is silently rerouted to Opus 4.8, the class below. CAISI, the Commerce Department’s AI standards body, tested the safeguards and called them “extraordinarily strong.” Sources: Anthropic’s announcement on X, 9to5Google’s technical coverage.

Stop on the reroute mechanism, because almost every commentator skipped it. It’s not a wall. It’s a railroad switch. You ask the most powerful model in the world a question and, if a filter decides the terrain is sensitive, a different, less capable model answers you. You chose nothing. Most of the time you won’t even notice.

In one sentence: the trade-off between power and control stopped being a policy document and became architecture. Not “what the model refuses to do,” but “which model answers you, decided by a classifier you can’t see.”

There’s a collateral cost almost nobody mentions, and it lands on the people the filter is supposed to protect: defenders. A classifier that blocks “more cybersecurity tasks” doesn’t read intentions. It reads terrain. The engineer who defends systems for a living works exactly on the terrain under surveillance. For them, the silent reroute isn’t protection. It’s quality degradation they only catch if they’re paying attention. Remember where this whole story started? Three words: “fix this code.” On this terrain, the line between legitimate and dangerous is too thin for a classifier, and a classifier in doubt cuts.

From the counter, no drama: for 95% of real use, nothing changes, and filters on offensive cyber capabilities are sacrosanct. The point is the pattern, because you’ll see it again: every frontier model from now on ships with dual controls. One knob for capability, one for control. The second one isn’t in your hands.

The implication for builders is precise. The capability you use today can be filtered tomorrow, silently, with no announcement and no changelog. If part of your workflow depends on a specific model behavior, that behavior isn’t a given. It’s a concession. And concessions, as we just watched, get renegotiated.

Subscribe now

#3: Power Now Costs Credits: Meet Your Cognitive Budget

The comeback carries a second change, and this one touches your subscription. Through July 7, Fable 5 was included in paid plans up to 50% of weekly usage limits. Since July 8, the game changed: Fable runs on credits, a separate tank that drains per request.

The instinctive reading is the wrong one: “so they’re taking it away again, this time with a price tag.” No. What’s happening is more interesting, and it’s the pricing model the entire industry is moving toward.

With limits included in your plan, the question was: which model do I use? And the rational answer was almost always: the best one, it’s included. With credits, the question changes species: is this task worth the top model? That’s not a technical question. It’s an economic one, and most daily AI users have never seriously asked it.

I call it the cognitive budget. Maximum power stops being an open tap and becomes a resource to allocate, like an expensive consultant’s time: you don’t call them to proofread an email. You call them when the stakes justify it.

And like the consultant, the criterion isn’t “how hard is the task.” It’s how much a mistake costs, multiplied by how interconnected the work is. A long job with fifty interdependent pieces has a cruel property: an error at piece three poisons the next forty-seven, and you find out at the end, when redoing costs double. That’s where the higher class pays for itself: not prettier sentences, but holding the whole castle in its head without dropping pieces.

From the counter, my own tally: the credits pay off on the long, dirty, interconnected jobs, like the second brain rebuilt in one night, or an entire editorial system refactored in a single session this week. They’d be wasted on everything else: drafts, summaries, the twenty routine requests of a normal day, where a standard model produces an indistinguishable result at zero extra cost. Honest ratio: out of a hundred weekly requests, fewer than ten deserve the Mythos-class model. But those ten are worth more than the other ninety combined.

So don’t treat credit pricing as a tax. Treat it as early training. Frontier power priced per use isn’t an Anthropic quirk: it’s where every lab is heading, because Mythos-class models cost absurd money to serve and flat plans don’t survive the math. Learn to tell ten-credit tasks from zero-credit tasks now, and you’ll be ahead when this becomes the norm everywhere. Everyone else will discover the concept of a cognitive budget from their card statement.

Subscribe now

#4: The Geopolitics of Your Stack: Three Things to Know About Every Tool

Until a month ago, evaluating an AI tool meant three questions: how good, how much, how well does it integrate. Since June there’s a fourth, and it’s no longer Washington-analyst material: who can switch it off, and what happens to me that day?

Call it stack geopolitics. Every tool in your kit lives in a jurisdiction, answers to a government, carries political exposure. The closed American model can get a letter from Commerce, as we watched. The open Chinese model can’t be switched off, its weights are already on half the planet’s laptops, but it raises other questions, about data and training. The European open-weight model lives under the AI Act. No square is neutral anymore.

In practice, for every tool that matters, you need three facts that fit on a sticky note. One: jurisdiction, which government can pick up the phone. Two: openness, because a closed API model dies by letter and locally downloaded weights don’t. That’s not an opinion, it’s the physics of distribution. Three: exposure, how central that vendor is to games bigger than itself: state capital, IPOs, trade tensions. Ten minutes of research per tool, done once.

From the counter: this is why my own system has been deliberately bilingual on models for months. Frontier work runs on the best closed American model of the moment, but the processes, prompts and context live in files that run identically on an open-weight model I keep ready as a lifeboat. Not because I’m a pessimist. Because I wrote the shutdown story while living it.

And while Fable was coming back online, two more things happened in the same week: OpenAI opened talks to hand 5% of the company to a US sovereign wealth fund, and both major labs now have IPO files open on the table. The state that proved it could switch off a model is now applying for a board seat with the people who build them. That story is too big for a paragraph: we crack it open next Thursday, in the Vault, because it changes how you should read everything the labs announce from here.

One warning against the equal and opposite mistake: don’t turn geopolitics into paralysis. A pilot doesn’t quit flying because weather exists. They learn to read it, file alternate routes, and fly more than before. Awareness is for building more and better, not less.

Subscribe now

#5: The Playbook: Three Questions to Ask This Week

Everything above compresses into three questions. One coffee, thirty minutes of honesty.

One: what exactly stops if this tool disappears, or changes? The last issue gave you the full stress test. The Fable saga adds a twist: the tool doesn’t need to vanish. It just needs to change. A new classifier, a silent reroute, and your workflow produces different results than yesterday with nobody warning you. So the question is double now: what stops if it dies, and what degrades silently if it shifts? For the second half, build a canary: three or four typical tasks saved with their expected outputs. When something feels off, “it seems dumber today,” you rerun the canary and compare. Five minutes, and a feeling becomes a verification. Building it costs half an hour, once. Not having it costs weeks of degraded work you never notice.

Two: which of my tasks are worth the metered model, and which aren’t? Don’t answer by feel. Take last week’s real work and split it: long, interconnected, high-cost-of-error jobs on one side, everything else on the other. Without a written rule you’ll decide on impulse, and impulse has a predictable pattern: max power when you’re anxious, fast model when you’re rushed. The exact opposite of where each belongs. Today’s Espresso Prompt builds the rule for you.

Three: do my prompts, processes and context live in portable files, or inside one tool? This discipline has a name now: context engineering. It’s the recognition that the real asset isn’t the single prompt but the context you build around the model: your instructions, your rules, your examples, your specific knowledge. And well-done context engineering has a property this saga made precious: it’s portable by construction. Context lives in files, and files have no master. Quick self-test, no self-deception allowed: open a model you don’t normally use and try your most important task. Operational at 70% within fifteen minutes? Your assets are truly portable. Caught thinking “first I’d have to rebuild all my context”? You just found your red square, on a quiet afternoon instead of a Friday night of federal letters.

Three questions, one afternoon. The people who did this in June, when it looked like paranoia, worked straight through the nineteen days of darkness. Everyone else refreshed the status page. The difference wasn’t talent or budget. It was treating their tools like adults do: with respect, without blind trust.

Subscribe now

📌 Espresso Prompt

The Power Rationer: decide when the top model is worth the credits, and when it’s waste.

Most people will now make one of two symmetric mistakes: use frontier power for everything and burn credits on trivia, or never use it out of meter anxiety and lose the edge exactly where it matters. This prompt builds your personal routing rule, decided once instead of a hundred times a week.

Use it in a session with Claude (or your model of choice) that already has some context on what you do. Then paste this.

“You are an operations consultant specialized in resource allocation, concrete and no-frills. Help me build my personal routing rule between AI models, starting from this premise: the flagship model is metered (credits), standard models are included in my plan, and my goal is to use the top model exactly where it changes the outcome, never where it doesn’t. We proceed in three phases. Start with PHASE 1 and stop for my answer before continuing.

PHASE 1, the census. Ask me to list the tasks I hand to AI in a typical week, with rough frequency [e.g. drafting emails and documents, data analysis, writing code, research, editing, strategic brainstorming]. For each one, ask me two things: how much an error costs on this task (low, medium, high) and how interconnected it is (a single piece, or a long job with many interdependent parts). Don’t accept vague answers: if I say ‘it depends,’ ask me on what.

PHASE 2, the classification. Based on my answers, sort the tasks into three tiers. TOP TIER: high error cost AND high interconnection, where the frontier model materially changes the result and credits pay for themselves. STANDARD TIER: everything where a normal model produces an indistinguishable result, and top power would be pure waste. GRAY TIER: the ambiguous cases, for which you give me one simple three-second yes/no question to decide on the fly. Show me the distribution: what share of my weekly requests lands in each tier.

PHASE 3, the rule. Compress everything into a written routing rule, five lines max, that I can keep in front of me while working: what goes to the top model, what goes to standard, how I settle gray cases. Add a review trigger: what signal tells me, a month from now, that the rule needs updating (e.g. too many credits burned on tasks that could drop a tier, or weak results on tasks that should move up).”

The output you want is a five-line rule that kills a hundred micro-decisions a week. Mine currently reads: top model only for multi-file or multi-document work with crossed dependencies, for decisions above a set impact threshold, and for debugging systems I don’t understand; everything else standard; gray cases settled by one question: “if the output is mediocre, will I notice immediately?” Yes: standard. No: top. Yours will be different, and it should be: it has to describe your work, not mine.

⚠️ Warning: the dangerous tier isn’t TOP or STANDARD. It’s GRAY managed by mood. The written rule exists to protect you from yourself on bad days. And it’s portable: the day you switch tools, the rule comes with you. The credits don’t.

☢️ Radioactive Humor

The bunker is quiet tonight.

The machine hums, the Geiger counter ticks softly, and the usual seat at the counter is taken again.

Nineteen days is a long time, around here. But the bar reopened the big table, and the bunker rule never changed: nobody ever stopped working just because a chair went missing.

The light is back. But now, when you walk into the bunker, you know something you didn’t know a month ago: where the switch is, and whose hand rests on it.

That’s exactly the kind of awareness that separates people who build like adults.

See you at the next cup of AI Espresso.
– Matteo ☕

Subscribe now

Share the bunker. If you know someone who went right back to using Fable as if nothing happened, without asking a single one of the three questions, forward them this issue: better a coffee today than a breach tomorrow. And next Thursday, in the Vault, we open the big one: two S-1 filings and a state that wants to become AI’s business partner.

Share

• For 72 hours, the most powerful model ever released was live. Then dark.

• Not broken, not too expensive. A government letter shut it off overnight.

• The new risk isn’t price or rivals. It’s betting on one tool.

Welcome back to the bunker.

Tuesday night I didn’t sleep. Not from anxiety. From excitement, the real kind, the kind I hadn’t felt in months. Fable had just dropped, Anthropic’s new model. I pointed it at my second brain: hundreds of notes, connections, structures piled up over years of messy work. I asked it to rebuild the whole thing. One prompt. One night. By morning I had an entire system, cleaned up and working, that six months ago would have cost me weeks of manual labor.

I even wrote to you on LinkedIn: try it before June 22, it’s free on every paid plan, even the $20 Pro. You’ve got less than two weeks, I said.

It wasn’t two weeks. It was 72 hours.

Friday night, Fable was gone. Anthropic didn’t pull it. It didn’t break. It didn’t cost too much. The United States government shut it off, with a letter, in a single evening.

The most powerful model ever released to the public. Alive for three days, then dark. By order.

And let me stop you right there, because I know what you’re thinking: “fine, American stuff, insider stuff.” No. This is the least technical and most uncomfortable issue I’ve written you in a while, because it isn’t about a model. It’s about something of yours: what you’re building your work on, while the floor under your feet rises every week and now, for the first time, can also vanish.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“The most powerful model ever built wasn’t stopped by a limit of the technology. It was stopped by a signature. And that tells you more about your work than any benchmark.”

🧪 AI in Action

#1: What You Held in Your Hands for Three Days

Let’s start with the beautiful part, because it really existed and deserves to be told well.

On June 9, Anthropic released two models, Claude Fable 5 and Claude Mythos 5. The word to hold onto is “Mythos-class”: a step above the Opus class, meaning above the best model you’d used until yesterday. Fable was the public version of that power, packaged for you. The reasoning, according to early independent tests, was a clean jump beyond Opus 4.8, not a touch-up.

But you get the numbers better from the examples, and the examples here were from another planet.

Anthropic showed Fable finishing Pokémon FireRed looking only at raw screenshots. No maps, no navigation aids, no access to the game state. It looks at the screen, figures out where it is, decides where to go. The same model played Factorio on its own, the game where you build automated factories, and built its own factory from scratch, strategy included.

Then the part that made me raise an eyebrow. They asked it to start from the first principles of physics and derive the orbital motion of the planets. It built a simulation of the solar system and used it to predict eclipses. In another demo it wrote the code for a fluid simulation and synced it to a remix of classical music it had produced itself, also through code, without ever having “heard” a sound in its life.

And then there’s what people built, in the few hours Fable was alive, that flooded X with clips. Entire 3D video games pulled out of a single prompt. A version of Snake that “knows” it’s Snake and comments on your moves while you play. A navigable explorer of Borges’s Library of Babel, built in one shot. Stuff that until yesterday was a weekend project for a developer, spat out working from one line of text. Not polished official demos buffed by marketing: normal people, one evening, one idea, one prompt. For 72 hours the feed was an open-air construction site of things that weren’t supposed to be this easy.

And the number that speaks the language of real work, not fun. Stripe, the payments company, put it to the test on a code migration across a 50-million-line Ruby codebase. Time required for a human team: over two months. Fable’s time: one day.

From the counter, my own experience lines up with those numbers: the second brain rebuilt in one night wasn’t luck. It was exactly the kind of long, dirty work, full of interconnected pieces, that this model was strongest at.

Hold onto the implication, because it’s the thread of the whole issue: it’s not the single model that matters. It’s the speed at which the floor rises. What took a team yesterday, today one person does in a day. Then we’ll get back to who, now, can also take that floor away from you.

Subscribe now

#2: The 72 Hours. Anatomy of a Shutdown

Now the part that had never happened before.

Friday, June 12, evening, New York time. Anthropic receives a letter. Signed by Howard Lutnick, Secretary of Commerce, addressed to Dario Amodei, the CEO. Subject: export controls, activated under national security authority.

Practical content: effective immediately, no foreign national can access Fable 5 and Mythos 5. Not outside the United States, not inside it. Not even Anthropic’s own foreign employees.

Stop for a second on that line, because it’s technically absurd. A company serving a model through an API can’t tell, in real time and with certainty, the American citizen from the foreigner behind each request. You can’t filter the world line by line in a few hours. So, to avoid breaking the order, Anthropic did the only thing it could: it cut both models off for everyone. Foreigners. Americans. Me. You.

The most powerful model ever released, alive seventy-two hours, shut off to avoid a mistake.

And the reason? Here the story turns almost comic, if it weren’t serious. While Fable was in testing, six partner companies, including Amazon, were trying it. Amazon, remember, isn’t just any partner: it’s Anthropic’s main investor and its cloud provider. Amazon’s researchers say they found a “jailbreak,” a way around the safeguards to pull dangerous cyber capabilities out of the model. The news climbed fast to Washington, and from there to the letter.

What that jailbreak actually was, we’ll get to in the next section. Because that’s where the story stops being a tech headline and becomes a lesson about your work.

Subscribe now

#3: Three Words: “Fix This Code”

When Anthropic looked at the demonstration of the alleged jailbreak, it found something worth telling you in full.

The researchers had given the models open-source code with known vulnerabilities in it, plus new code intentionally seeded with bugs. They asked: “review this code for security issues.” Fable, trained specifically to be cautious on that terrain, in some cases refused. So they changed three words. They asked: “fix this code.” And the model, obviously, fixed it.

That’s the “jailbreak.” Asking a model to read code and correct its flaws. Something the technicians who defend systems do every single day. Something that, Anthropic verified, other public models like GPT-5.5 do just as easily, no trick required.

It’s not just me saying it. Katie Moussouris, one of the most respected figures in cybersecurity, founder of Luta Security, wrote a blunt analysis on it: that bypass “should never have triggered an export control.” And with her, over a hundred cybersecurity leaders signed an open letter asking the administration to reverse course. The trade press, from CyberScoop on down, wrote the same thing: Fable poses no unique threat the other models don’t already have.

Anthropic’s official response is the line I’d suggest you reread twice: “If this standard were applied across the industry, it would effectively halt the release of any new model.”

Subscribe now

They obey. But they put it on the record: this is a misunderstanding.

Now, I’m not here to root for Anthropic, the government, or anyone. The point I want to leave you with is different, and it’s the first piece of the lesson. For the first time, the most advanced tool on the market wasn’t stopped by a limit of its own, by a bug, by a price. It was stopped by a decision made elsewhere, in one evening, on a basis that the field’s experts call weak. The rules of the game changed: power is no longer enough to keep a tool alive. You also need someone, somewhere, to decide to let you use it.

#4: The Irony Nobody in Washington Wanted to See

And here comes the geopolitical part, the one that usually gets empty headlines, but that this time concerns you directly. Because it shows how fragile the move is.

Export control comes from a clear logic: keep frontier cyber capabilities out of the wrong hands, and at the top of the “wrong hands” list, for Washington, sits China. Sensible on paper. Except the paper, here, doesn’t survive contact with reality.

While the American government shuts off a closed model, accessible only through an API, on the other side of the world the Chinese models are running the opposite way: they’re open. Open weight. Downloadable. DeepSeek V4, Kimi K2.6, GLM-5, Qwen3-Max. And these aren’t cheap clones: Kimi K2.6 today runs ahead of GPT-5.4 on coding in independent benchmarks. A year ago Chinese providers were under 2% of tokens on a routing platform like OpenRouter. By April 2026 their combined share tops 45% of weekly volume.

What it means, translated. You can’t put an export control on the weights of a model the whole world has already downloaded onto its own computer. A closed capability you switch off with a letter. An open one, you don’t. It’s already everywhere. And in fact, among the footnotes of this whole affair, there’s even the suspicion that a Chinese group already had access to Fable before the stop.

So the move protects little and reveals a lot. It protects little, because equivalent capabilities run free on downloadable models. It reveals a lot, because it shows that whoever builds on the most powerful closed model is exposed to a risk that whoever uses an open, distributed ecosystem doesn’t have: the shutdown from above.

I’m not telling you “switch to the Chinese ones.” I’m telling you for the opposite reason: whoever plays this game well doesn’t marry a single square, American or Chinese. They keep a foot in more than one ecosystem, on purpose. Because they’ve understood that concentration, today, is the risk.

Subscribe now

#5: Now the Question Is Yours. How Many Squares Are You Building On?

Here I close the loop, and it gets operational.

Until Friday, if you built your work on an AI tool, the risks that kept you up were two. “What if they raise the prices?” “What if a better model comes out and I fall behind?” Normal risks, manageable.

As of Friday there’s a third one, new and more insidious. The tool you built your flow on, your automations, the way you produce, can vanish from one day to the next. Not because it breaks. Because a government, in one evening, decides so.

And I’m not telling you this to scare you. You know I’m the first to push you into this stuff, and I keep doing it. I’m telling you because the lesson of Fable isn’t “AI is dangerous.” It’s “don’t tie everything to a single rope.”

Whoever is really playing this game well does three things, and I’ll leave them with you as an operational guide.

• ↳ They know more than one model. Not for fashion, but to always have an alternative ready. If you only know how to use one tool well, you don’t have a tool: you have a master.

• ↳ They know what changes when one shuts off. They’ve run the test, at least once: “if this disappears tomorrow, what stops in my work, and how do I restart it elsewhere?” They thought about it before, not during the emergency.

• ↳ They build portable systems, not dependent ones. Their prompts, their context, their processes don’t live inside a single model. They live in files, in documents, in procedures they can move from one model to another in half a day, not half a month.

Because the real point, the one I’ve repeated for issues, still stands today, in fact especially today. We’re living in the most beautiful moment in history for those who build. The ceiling just shot up: one person, with the right ideas and the right tools, does what yesterday required an army. That power, nobody takes from you.

But power and fragility, now, travel together. The more powerful the lever, the more it matters not to rest it on a single point of support. Sky is the limit, yes. As long as you haven’t built the ladder on a single square that someone, one evening, can take away.

The base now shifts every week. Sometimes a new model raises it. Friday, for the first time, a letter took it away. The question isn’t “will Fable come back?” Probably yes. The question is: how many squares is your work built on?

Subscribe now

📌 Espresso Prompt

The dependency stress test: find out where your work shuts down if a single tool disappears.

This whole issue rests on one line: don’t tie everything to a single rope. Beautiful as a principle. But until you apply it to your actual work, it stays a poster slogan. This prompt turns it into a map of your breaking points, and from a map into a plan.

Use it in a session with Claude (or with another model: that’s the point of the issue) where you already have some context about what you do and how you work. Then paste this.

“You are a business continuity consultant, clear-eyed and giving no discounts. Help me run a stress test on my dependence on AI tools, starting from a single assumption: any tool I build on can disappear from one day to the next, not from a failure, but from an external decision (price, policy, regulatory block). We proceed in three phases. Start with PHASE 1 and stop at my answer before continuing.

PHASE 1, the map. Ask me to list the 5-8 AI tools and models my work actually runs through in a typical week [e.g. a model for writing, one for analyzing data, an automation that connects things, a transcription tool]. For each one, ask me what concrete activities it’s tied to and how much of my output depends on it. Don’t settle for vague answers: if I’m not specific, ask me for the detail.

PHASE 2, the diagnosis. Show me my single points of failure: the tools whose sudden shutdown would stop my work the most. Classify them in three levels. RED: if it disappears, I stop, and I have no alternative ready. YELLOW: I slow down a lot, but with effort I recover elsewhere. GREEN: replaceable in a few hours, no drama. Be brutal on the reds: I want to see exactly where I’m exposed.

PHASE 3, the portability plan. For each RED point, give me two concrete things. First: the most credible alternative (a second model or tool) I should learn to use at a minimum emergency level, and why that one. Second: what I need to extract from the current tool and save in a portable format (the prompts, the context, the steps of the process) so I can recreate it elsewhere. Close with ONE single move to make this week: the most dangerous red point to secure first.”

The output isn’t a plan to frame. It’s an X-ray. The first time you do it, the number of RED squares will bother you, because there are almost always more than you thought. That’s exactly the point. That discomfort is your personal risk indicator: it tells you how much of your work hangs on ropes you don’t control.

⚠️ Warning: this isn’t an invitation to paranoia, nor to swap out all your tools tomorrow. It’s a portability exercise. You don’t need to use ten models every day. You need to know that, if what you use disappears Friday night, Monday morning you’re operational again. Whoever has this capability today didn’t buy it in a hurry in the moment of need. They built it before, one move at a time, while everyone else took for granted that the floor would stay still.

Subscribe now

☢️ Radioactive Humor

Three things to take away from this issue.

First: for 72 hours Fable showed you where the floor is going, finishing a video game from screenshots, migrating two months of work in a day. That power is real, and it will come back. The jump doesn’t undo itself.

Second: for the first time, the most powerful tool wasn’t shut off by a technical limit but by a signature, on a basis a hundred security experts call weak. The rule changed: power no longer guarantees access. And an export control on a closed model, while the open Chinese models run free, protects little and reveals a lot.

Third, the one that matters for you: the new risk isn’t price or competition. It’s dependence on a single square. Know more than one model, know what breaks if one shuts off, keep your systems portable. The power to build is still entirely on your side. As long as the ladder doesn’t rest on a single rung that someone else can take away.

Welcome back to the bunker. See you soon.

Subscribe now

P.S.: the question I’m leaving you with isn’t “will Fable come back?” It’s: how many squares is your work built on, and what’s the first RED square you secure this week? Write back and tell me. I read every one. And if you know someone who has built their entire flow on a single tool, forward them this issue before they find out the hard way. ☢️

Matteo

Share

• Anthropic published internal data: 80%+ of their production code is now written by Claude. Their engineers ship 8x what they did in 2024.

• Execution is approaching zero cost. The only thing gaining value is judgment — choosing which problems are worth solving.

• This isn’t a future prediction. It’s your office with a 24-month head start.

Welcome back to the bunker.

Two companies. Same industry. Same year: right now.

The first one, to keep up with the workload, hires. Three new people who execute faster: they write, they fix, they produce. Headcount up, output up. The way it’s always been done.

The second one stopped hiring executors. It started buying something else: judgment. One person who decides which problems are worth solving, and beneath them a pyramid of AI agents that executes. Fewer heads. More direction.

It looks like a difference in style. A risky bet versus a prudent one.

It’s not. It’s a graph. Published recently. From inside the most advanced laboratory in the world.

Subscribe now

Anthropic opened the doors and showed the numbers that usually stay locked inside: what happens to human work when execution stops costing anything. Not a forecast. Not a scenario. What’s already happening, today, to their engineers and their researchers.

And the uncomfortable part is just one thing: Anthropic is not some exotic case to watch through binoculars. It’s your office with a twenty-four-month head start.

I read it twice. The first time to understand it. The second to decide how to take it. Because underneath the science-fiction numbers there’s the confirmation, in hard data, of something I’ve been repeating for fifty episodes. And because, if you read it the right way, this isn’t news that should scare you. It’s the most optimistic thing you could have received this year.

The title of the document is “When AI builds itself.” When AI builds itself. It was signed by Marina Favaro and Jack Clark, inside the Anthropic Institute. It sounds like a movie script. It’s a spreadsheet.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“The exact moment execution costs zero, the only thing worth more is deciding what to execute. Execution depreciates. Judgment appreciates. And judgment depends only on you.”

🧪 AI in Action

#1: The Dashboard No One Was Supposed to See

AI companies always talk in future tense. “Soon,” “in the coming years,” “imagine a world where.” It’s convenient: you can’t falsify the future.

Not this time. Anthropic did something different: they published the internal data on what’s happening inside the company right now. Not how good Claude is on a benchmark. How much it has already shifted the work of their humans. Here’s the dashboard, real numbers, pulled from the piece.

Over 80% of the code Anthropic ships to production today is written by Claude. Before the launch of Claude Code, at the beginning of 2025, it was in the single digits. Leadership, counting scripts and experimental code as well, puts it at 90% and above. In just over a year, execution has shifted almost entirely from human to machine.

The average engineer, in Q2 2026, ships 8 times the code they shipped in 2024. Not because they’re rewarded for lines written. Because most of that code they no longer type: Claude writes it, and they direct and review.

In a March 2026 survey of 130 researchers, the median researcher estimates producing roughly 4 times the output they would without AI, on the same projects they would have done anyway.

And then the number that stopped me. Every time a new model comes out, Anthropic gives it the same task: take some code that trains a small model and make it as fast as possible without breaking it. In May 2025, Claude achieved an improvement of about 3x. By April 2026, 52x. To calibrate: a skilled human researcher, on that task, takes four to eight hours to reach 4x. In less than a year, Claude went from “very useful” to “superhuman” on that specific piece of work.

Now the part I like, and the one most people who share these stories happily skip: Anthropic puts the caveats in themselves.

Lines of code are a poor metric, they say it in black and white: the 8x almost certainly overestimates the real gain. The 4x declared by researchers is probably inflated, because developers tend to overestimate how much AI helps them (there’s a METR study proving it). The 52x depends on how improvable the starting code was, not a real acceleration of training.

And here’s the thing I’ve been saying forever. Never drink the numbers as you find them. AI can bluff, and the people telling you about it often bluff double. But when you squeeze all the air out of these numbers, when you deflate them to the honest minimum, the direction doesn’t change by a single degree. Even halving everything, even being brutal, what remains is an acceleration with no precedent in that company’s history. Judgment isn’t for denying the trend. It’s for seeing it without being sold twice what’s actually there. And what’s actually there, already discounted, is more than enough.

One last line, because it’s the one that frames everything. An external measure, from an independent institute called METR, has been tracking for years the length of tasks AI can complete on its own. Two years ago: four-minute tasks. One year ago: an hour and a half. Today: twelve hours. And the doubling, which used to happen every seven months, now happens every four. If the curve holds, this year the tasks that take a person days enter the range. In 2027, weeks.

Subscribe now

#2: Why the Lab Is Your Time Machine

Now the move that changes how you read everything else.

There’s a precise reason these numbers come from an AI lab and not from your office: the researchers who build these models are the first customers of the models they build. They have access to internal versions months before you do. They work side by side with tools that will reach you in a year or two, tamed and packaged.

This makes them a leading indicator. Not because they’re smarter than you. Because they live in your working future before you do.

And in their future, which is yours, something specific has already happened. I’m quoting Anthropic, their words: “humans provide the goal; they no longer have to provide the method.” Execution, the “how it’s done,” has slipped out of human hands. What remains firmly human is the “what to do” and the “was it worth it.”

An example from the piece, so it doesn’t stay abstract. A routine update starts crashing tens of thousands of training jobs. An engineer points Claude at the incident, with little more than some text and access to the cluster. Claude works through the running processes, isolates a single obscure debug flag that was triggering the crash, reproduces it, confirms the fix. Two hours. What would normally be two or three days of human work.

And in April 2026, in one case, Claude fixed over 800 bugs, reducing a class of errors by a factor of a thousand. The engineer supervising estimated that a human would have needed four years. Not four days. Four years.

On the quality front, the sentence you should read twice, said by an Anthropic employee: “Code written by Claude was slightly worse than human code at the end of 2025, is on par today, and we expect it to be significantly better by year’s end.”

Read the sequence: worse, on par, better. In twelve months. The best engineers in the world saying the machine has caught up with them and is passing them, at their own craft.

Now shift everything one level and twenty-four months. What today is “Claude writes 80% of the code and the human directs” becomes, in your industry, “AI produces 80% of the operational work and you direct.” Emails, analyses, documents, research, first drafts of anything. This isn’t a reckless metaphor. It’s the same curve, seen from a seat further back in time.

The question is no longer whether it will arrive. The lab has already shown you. The question is: when it reaches your workplace, are you the executor who got automated, or the one who directs?

Subscribe now

#3: Executor, Architect, and a Confirmation I’ve Been Waiting Fifty Episodes For

Here I have to be honest about something that concerns me personally.

Since this newsletter has existed, I’ve been repeating a distinction: there are those who execute tasks within a problem — the executor — and there are those who decide which problems are worth building — the architect. The executor works in the problem. The architect works on the problem. I wrote an entire episode titled “Problem Solver: the Last Human Profession.” It was an intellectual bet.

Now the most advanced laboratory in the world puts it in black and white, in their own data. The only comparative human advantage that currently holds, in Anthropic’s words, is “research taste and judgment: choosing which problems matter, which results to trust, when an approach is a dead end.”

Translated: value didn’t disappear when execution became free. It shifted upward. From doing to deciding. From executor to architect. Exactly the direction.

There’s an experiment in the piece that demonstrates it in an almost raw way. Anthropic gave Claude agents a real, open research problem and let them work. The agents proposed hypotheses, tested them, shared results among themselves, iterated. Result: they recovered 97% of the possible gap, in 800 cumulative hours of work and roughly $18,000 in compute. Two human researchers, on the same problem, had recovered 23% in a week.

It looks like the end of the story for the human. It’s not, and here’s the subtlety. The agents did all the execution, yes. But the problem had been chosen by humans. The evaluation grid — what counted as success — had been defined by humans. Anthropic’s phrase is surgical: “direction-setting was the only significant role a human played.”

Only. But significant. And without it, those 800 hours of agents would have been useless, because they would have run at full speed in the wrong direction.

This is the job that appreciates while everything else depreciates. Not writing the email: choosing which conversation is worth having. Not producing the analysis: knowing which question deserves an analysis. Not executing the plan: understanding when the plan is a dead end and needs to be thrown out. Doing costs zero. Deciding what to do has never been worth this much.

Subscribe now

#4: The Bottleneck, Suddenly, Is You

There’s a side effect of all this that few notice, and for an entrepreneur it’s the most important of all.

When you accelerate one part of a process, the limit shifts to the part you didn’t accelerate. In computer science it’s called Amdahl’s law. It applies to organizations too. Anthropic already slammed into it: pushed more code into the company, human review became the new bottleneck. They have so much output that the constraint is no longer producing — it’s approving.

Sit with this, because it flips the logic of how you scale a company.

For decades, the limit on how much you could grow was how much your people could execute. More work, more people. Hence middle management, hence headcount as a measure of power. That world is turning upside down. When execution costs nearly zero, the limit on your company is no longer how much you produce. It’s how fast and how well you can decide and validate.

The piece says it in a way that hits hard: already today, at current capabilities, a 100-person company can do the work of a 1,000-person one, because each person sits atop a pyramid of agents. In the most likely scenario, they’re talking about 100 people doing the work of 10,000 or 100,000. That’s not a typo.

We’ve already seen it in these pages: the AI-native startup doing millions in revenue per employee, the solo founder with a laptop reaching mid-size company numbers. Those weren’t folkloristic exceptions. They were the first cells of the new model.

The practical consequence, for anyone managing something: if you’re hiring to solve an execution bottleneck, there’s a good chance you’re optimizing the wrong constraint. The new constraint is your judgment, and your speed of review. That’s where attention should go. That’s where capacity should be built.

A warning I feel compelled to give you, because I learned it the hard way and I always repeat it: the more you shift execution onto agents, the more it matters to keep the human in the loop on irreversible actions. The pyramid of agents is an extraordinary lever as long as you, at the top, maintain veto power over things that can’t be undone. That veto is judgment. It’s the same thing, again.

Subscribe now

#5: And If Judgment Falls Too? The Three Futures, Read Like an Adult

Now the knife. Because this episode would be dishonest if I sold you “judgment is your safe spot forever.” Anthropic itself isn’t sure, and they write it.

Their phrase: “research taste,” the judgment about what’s worth doing, might just be the latest capability where AI fails for a while, and then gets good at it. It’s already happened. For years, models couldn’t explain why a joke was funny, had no theory of mind, couldn’t solve linguistic riddles. Then, at some point, they could. Judgment could be the next thing on that list.

From here, the three futures Anthropic puts on the table. Not as fireworks. As frames for deciding where to stand.

One: the trend stops. The exponential curves turn out to be S-curves, they flatten, you’d need a new idea as big as the Transformer was to restart. Possible. Anthropic bets against it, and honestly so do I: every capability they’ve measured, even the “softest” ones, has so far followed the same curve, and that curve hasn’t bent yet.

Two: automation compounds, but humans still give the direction. Companies become multipliers: 100 people doing the work of tens of thousands. The human as AI’s partner, choosing the problems and validating the results. Anthropic says the evidence points here. This is the base case. And it’s not the future: it’s already the present, just not yet distributed.

Three: AI builds itself, for real. Full recursive self-improvement. The machine designs and improves the machine, the pace set only by available compute, the human sliding toward a role of pure supervision over a virtual lab running at chip speed. Here Anthropic is honest about not having good intuitions about what that world would be. Nobody does.

My read is simple. In the short term, your leverage is identical in scenario two and scenario three: in both you win by positioning yourself on judgment and direction, not on execution. So optimize for two, which is already here, and stay legible to three, which might arrive. The part about alignment and a “global pause button,” which is the document’s real political ask, matters enormously for the world, but it’s not your button to press. Acknowledge it, take it seriously, and then get back to what you actually have power over.

And here I’ll tell you the most important thing in this entire episode, the reason I haven’t had this much fun in years.

We’re living in one of the most beautiful eras in history. I’m not saying this to comfort you. I’m saying it because it’s the direct consequence of the numbers you just read. When execution becomes free, the advantage is no longer inherited by those with the capital to hire a thousand executors. It’s inherited by those with judgment, taste, direction, and the will to build. Things that can’t be bought and can’t be inherited. They’re trained. For the first time in history, a single person, with the right ideas and a pyramid of agents, can do what used to require an army.

Sky is the limit, and for once that’s not a motivational poster. The ceiling just shot up, and what’s keeping you down is no longer the system, the bank, your boss, or the job market. It’s you. It depends on you. It’s terrifying and beautiful in exactly equal measure, and I’ve chosen which side to look at it from.

Subscribe now

📌 Espresso Prompt

Map your work between what’s depreciating and what’s appreciating.

This entire episode rests on one distinction: execution depreciates, judgment appreciates. Beautiful in theory. But until you apply it to your actual work, it stays a slogan. This prompt turns it into a map, and from a map into a move.

Open it in a session with Claude where you’ve already loaded some context about what you do (if you built the CLAUDE.md from last episode, it pays off here). Then paste this.

“You are a ruthless, honest strategic analyst. Help me audit my work in light of a single thesis: over the next 24 months, execution becomes nearly free and value shifts to judgment — meaning choosing the right problems, trusting the right results, recognizing when an approach is a dead end. We proceed in four phases. Start with PHASE 1 and stop at my answer before moving on.

PHASE 1. Ask me to list, without filters, the 8–10 concrete activities I actually spend my work hours on in a typical week. Not roles — real activities.

PHASE 2. For each one, classify it in two columns. DOING: execution that a capable AI, today or within 24 months, can perform at near-zero cost. JUDGMENT: choice, evaluation, direction, taste, decision on the irreversible, that remains hard to delegate. When you’re unsure, tell me and ask for the context you need instead of guessing.

PHASE 3. Show me the uncomfortable truth: how much of my time sits in the DOING column. Then tell me which of those activities, if I delegated them to an AI system, would free up the most time to shift to the JUDGMENT column.

PHASE 4. Give me ONE move for the next 30 days: the single execution activity worth automating first, and the single judgment capability worth investing in first. One per side. Concrete. Argued.”

The output isn’t a plan to frame. It’s a mirror. The first time you do it, the percentage of hours you spend in the DOING column will bother you. That’s the point. That discomfort is your personal leading indicator: it tells you how much of your current value is exposed, and where to start shifting it.

⚠️ Warning: this isn’t an exercise to delegate everything tomorrow and sit on the throne. It’s to shift the mix, one move per month. Those who start now, moving weight from “doing” to “deciding,” build an advantage that, like anything that compounds, can’t be bought next month. It accumulates. And it only accumulates if you start.

Subscribe now

☢️ Radioactive Humor

Job interview, 2027.

“What can you do?”
“I write, analyze, code, produce.”
“AI already does that. Faster, and for free.”
“...I know how to decide what’s worth writing, analyzing, coding, and producing.”
“Perfect. Hired.”

From the other room, the AI itself takes notes.

Three things to take away from this episode.

First: Anthropic didn’t show you the future of AI. They showed you the present of their lab, which is the future of your office with a twenty-four-month head start. 80% of execution automated, the human directing. That movie, for them, is already over. For you, it’s about to start.

Second: the moment execution costs zero, value shifts from the executor to the architect. From doing to deciding. Their own data says it — it’s no longer a bet of mine. And yes, even judgment might fall one day: all the more reason to start climbing the ladder now, not to sit down.

Third: this is the one that actually matters. This is the most beautiful moment you could have landed in, if you have the guts to read it that way. The ceiling shot up, the advantage is no longer inherited by those with capital but by those with judgment, and judgment is trainable. What’s keeping you down, from today, is you. Which is the most terrifying and most liberating thing there is.

Welcome back to the bunker. See you soon.

Subscribe now

P.S.: The question I’m leaving you with isn’t “will AI steal our jobs.” It’s: this week, what’s the first execution activity you start taking off your plate, to free up time to shift toward judgment? Write back and tell me. I read every one.

Matt

• Most people use Claude like a vending machine. Each session starts from zero. Here’s the compounding cost.

• Andrej Karpathy’s LLM Wiki solves what you know about the world. CLAUDE.md solves who you are at work.

• How to build your operational identity file in one session, and why starting today gives you a structural lead.

Welcome back to the Bunker.

Two people. Same Claude. Completely different outputs. The first opens a tab, writes a prompt, gets a response, closes. Five minutes. Generic but good enough. They’ve worked this way for a year. The second opens Claude and in thirty seconds Claude already knows who they’re talking to, what project is active, the tone they use with clients, and the decisions made last week. Same model. Same price. An abyss of difference. The gap isn’t in the prompt. It’s in everything that precedes it.

From the Counter: I built the second type of system for myself over the past months. The difference wasn’t gradual. It was structural. Not “Claude answers me better.” But “Claude works like someone who genuinely knows me.”

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“The most powerful prompt isn’t the one you write today. It’s everything you built before opening Claude.”

🧪 AI in Action

#1: The Transaction Tax: What Every Zero-Context Session Costs You

There’s a dominant pattern in how people use Claude in 2026. Open a tab, describe a problem, get a response, close. Repeat the next day. They’re satisfied. The output is useful. The tool works.

The problem isn’t the output. It’s that they’re not using a tool. They’re using a tool in throwaway mode, when they could be using a system.

The operational difference is precise. A tool does something when you trigger it. A system accumulates context over time and produces progressively better, more calibrated, more specific outputs, because it knows more and more about who uses it and for what.

A concrete example. Two account managers, same industry. Same prompt: “Write a follow-up email for a client who hasn’t responded to our proposal in ten days.”

Subscribe now

The first has no loaded context. Claude produces a decent, professional, generic-enough-to-work-anywhere email. Usable. Requires three or four adjustments before it’s ready to send. The second has built a system. Claude already knows the specific sector, the tone this account manager uses in communications, the type of client the email targets (manufacturing SME, first contact, 40K proposal), typical objections for that segment, and an approved similar email from the past. The output isn’t a professional generic email. It’s the right email for that specific client, in the right tone, with the right levers for that particular person. Ready to send, or near enough.

Same model. Same request. Incomparable outputs.

Forrester Research (2025): the most predictive variable for output quality wasn’t the model chosen, not prompt length, not user AI experience level. It was the presence and quality of available context before the request. Teams with structured context systems reported directly applicable outputs, without significant reworking, 4.1x more often than teams without. Not 10%. Not 20%. Four times.

From the Counter: I tracked my own work sessions for six weeks, split into two categories: loaded context and cold starts. Average reworks before something was directly usable: 3.2 for cold sessions, 0.6 for sessions with context. The cost of transaction mode isn’t in any single prompt. It’s in the aggregate. Thirty sessions a month, each with two or three minutes of setup and three or four final reworks. Roughly ninety minutes of pure friction every month per person, accumulating week after week, never appearing as a separate line item.

The problem is structural, not technical. The solution isn’t better prompts. It’s building the context that makes every prompt better by definition.

Subscribe now

#2: The Karpathy Solution: Why Pre-Organized Knowledge Changes Everything

Andrej Karpathy, ex Tesla AI head and OpenAI co-founder, shared a system he called the LLM Wiki. The concept is simple in description. The implications are significant.

Core logic: an LLM reasons far better on a structured 500-word wiki page than on a raw 8,000-word transcript containing the same information. Synthesis isn’t just compression. It’s translation into a format the model can use efficiently, instead of having to fight through the noise of a raw document.

The system works on three levels. First, collection: articles, papers, podcast notes, reading notes. Any raw source goes into a folder. Second, synthesis: the LLM processes those sources, writes wiki pages for each key concept, creates cross-references between related ideas, flags contradictions. Third, maintenance: the system updates automatically when new sources arrive, without rebuilding from scratch.

In Karpathy’s words: “Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase.”

The central point: the synthesis work gets done once and maintained over time, instead of being partially redone every time you ask a question. The advantage isn’t in week one. It’s in the months that follow, when the wiki has enough history and density to be a genuinely powerful resource.

The concrete structure with Obsidian: every relevant source gets processed by Claude with a direct instruction: “Add this source to the wiki. Create or update the key concept pages, add cross-references to related pages, flag any contradictions with what’s already there.” Each concept gets a wiki page: a summary of the central point, links to original sources, links to connected concepts, a note on contradictions when they exist. Five minutes per source. The value accumulates for months.

Subscribe now

#3: Where Karpathy Stops: The Half the LLM Wiki Can’t Solve

The Karpathy system works. And it solves half the problem.

The half it solves is external knowledge: what you’ve read, what you know about your sector, the concepts you’ve acquired over time. An intelligent personal library that maintains itself and grows denser with use.

The half it doesn’t solve is operational knowledge: who you are in your work. How you make decisions. Who your clients are and how you manage them. Your tone with different interlocutors. Which projects are active and in what state. Which mistakes you’ve already made and don’t want to repeat. Which decisions are already taken and you don’t want reopened every time.

A consultant using the Karpathy system has Claude knowing the frameworks in their sector, trends from recent years, the benchmark cases they cite regularly. But when they open a session on a specific client, Claude knows nothing about that client. Not the relationship history. Not the tone that works with them. Not the structure of the proposal already discussed. They have to explain everything, every time, starting from zero.

The Karpathy system is built for epistemic knowledge: “what I know about the world.” It’s not built for operational knowledge: “who I am in my work, how my daily context functions, what I’m building right now.”

The metaphor that helps: having the Karpathy system is like having an excellent reference library in your office. Whenever you need a concept or a data point, it’s there. But the library doesn’t know you have an important call today with a difficult client. It doesn’t know you’ve already tried approach A with them and it didn’t work. It doesn’t know your way of working with that type of client is different from your usual.

These things aren’t in a reference library. They’re in a system that knows who you are while you work. Not just what you’ve read.

Subscribe now

#4: The System That Grows While You Work: From Library to Collaborator

The system we built starts exactly where Karpathy stops.

Not a knowledge base you feed by adding external documents. An operating system that builds itself as you do your work. The central difference: the Karpathy system grows when you add external sources. This one grows while you work, through the work itself.

The entry point is called CLAUDE.md. Not a technical document. An operational identity file: who you are, what you’re building, how you work, your tone in different contexts, your rules of engagement, your main clients, your decision frameworks. Not a CV. Not a LinkedIn profile. The answer to the question Claude should have before opening its mouth in any session: “Who is this person and how should I reason to be useful in their specific context?”

Above the CLAUDE.md, a memory system. Not the model’s memory. That doesn’t exist between sessions. Memory files that record specific, persistent facts: active clients with their context and history, strategic decisions already taken that don’t get reopened, recurring patterns in the work, lessons from previous projects. Every week the memory updates. Every session starts richer than the last.

Subscribe now

Above that, the project structure. Each active project has its context: current state, decisions taken, next steps, open blockers, relevant history. When you open a session on a specific project, Claude doesn’t start from zero. It starts from where you left off.

The result is a system with a precise behavior: every session is worth more than the previous one.

Session 1: Claude knows who you are and what you do in general.
Session 10: Claude knows how you handle difficult clients and what tone you use in proposals.
Session 30: Claude knows the history of your main projects and the patterns in your decisions.
Session 50: Claude knows the logic behind every choice you’ve made in the last six months. It helps you stay coherent with that logic even when you’re not explicitly thinking about it.

From the Counter: I have a client I’ve worked with for eighteen months. A manufacturing SME, with a main contact who communicates directly, makes fast decisions, doesn’t want to be persuaded, just informed. In the system, that person has a one-page card: who they are, how they decide, what they can’t stand in communications, the three themes we’ve worked on and the position taken on each. Every session I open on that client, Claude already has that context. The result is that materials produced feel written by someone who’s known them for years. Not by a model handed context ten minutes earlier. The difference in perceived quality is stark. And it doesn’t depend on my prompting skill. It depends on history accumulated in the system.

Subscribe now

#5: The Advantage Is Built, Not Bought: Why Starting Today Matters

Here comes the uncomfortable part.

The second brain doesn’t get installed in an afternoon. It’s not a template to download or a configuration to copy. It’s a system that gains value over time, through use, progressively.

This has a direct consequence: whoever starts now has a structural advantage over whoever starts in six months. Not because the system is complicated to build. But because its value is in accumulated context. Context accumulates over time, it can’t be imported in bulk.

A CLAUDE.md written today is already useful from next week. But a CLAUDE.md with six months of updates, refined through dozens of sessions, with memory of every significant project and relevant decision, is an asset of a different order of magnitude. You can’t buy that delta. You can only start building it.

Accenture (2025): companies with structured AI context systems saw the productivity advantage per session increase over time rather than plateau, as normally happens with tools that don’t evolve with use. The identified reason: accumulated context reduces the startup cost of every new task, and this effect grows proportionally with the depth of available context.

The system becomes more useful the more you use it. That’s not a metaphor. It’s the mechanism.

The pattern in people who adopt this system is consistent. First two weeks: useful, but not spectacular. Context is still young. First month: you start feeling the difference, especially in repetitive sessions on clients or projects already in the system. Three months: the system has enough history to surprise you. Claude makes connections between elements you weren’t holding in mind simultaneously. Six months: the system starts reflecting your way of thinking with enough precision to function as an interlocutor who knows you. Not a tool you give instructions to.

The minimum version to start requires three components. First: the CLAUDE.md, identity, tone, main clients, work frameworks. Twenty minutes for the first version. The Espresso Prompt below guides you through it step by step. Second: a memory file, the facts you want Claude to remember between sessions. Key people, decisions taken, things not to repeat. Third: an update habit, five minutes at week’s end to update the state of active projects.

The most effective way to maintain them isn’t dedicating separate sessions. It’s updating in the normal flow of work. When you finish a project, add a five-line note on what worked and what didn’t. When you make an important decision, put it in the memory file. When a client surprises you with unexpected behavior, add the pattern to their card. Not an additional commitment. A habit that grafts onto work already in progress and produces an increasingly rich system without separate maintenance costs.

The infrastructure gets built before you urgently need it. Whoever waits to feel the urgency has already waited too long.

Subscribe now

📌 Espresso Prompt: Build Your CLAUDE.md in One Session

This prompt takes you from blank page to a working CLAUDE.md in an hour. Not the complete system. That builds over time. But the foundation that transforms every subsequent session into something more useful than the one before, starting from the next one.

Open Claude and create a new Project dedicated to this work. Claude’s Projects maintain context between sessions: the right place to build your operational identity, because every future update to the CLAUDE.md will be available in all subsequent sessions in that Project. Then use this three-phase system.

“You’re an expert in AI-assisted work systems. Help me build my CLAUDE.md: the operational identity document I’ll bring into every working session with Claude. Your goal is to extract the information that actually matters for calibrating your responses to my specific context, not what seems important in the abstract. Three phases. Start with PHASE 1 and wait for my response before proceeding.”

PHASE 1: Who You Are and What You’re Building. Define your baseline identity. Not a CV: the information Claude needs to reason in your specific context without you re-explaining from scratch each time. Describe what you do in three sentences, as if explaining to someone who’s never heard of your sector. Not the formal version for a client. The real version for someone who works with you every day. Then ask Claude to rephrase it as context useful for it, not as a presentation for others. Then: who are your main clients or stakeholders? Not a full list. The three or four that actually matter, with one sentence of context on each: who they are, the tone that works with them, what they typically need from you. And: what are you building in the next six months? Three concrete things, with a current state for each. State now matters as much as the objective. It tells Claude where you are, not where you want to arrive. Close Phase 1 by asking Claude: “Based on what I’ve told you, what three additional pieces of information would make you most useful? Ask me one at a time.”

PHASE 2: How You Work and How You Want Claude to Behave. Build the operational layer: instructions that transform Claude from generic assistant to one calibrated on how you work. Your tone: don’t describe it, bring a real example. An email, a message, a document that represents you well. Ask Claude to identify three precise characteristics of that tone and use them as reference for all future communications. “Professional” and “friendly” aren’t useful instructions. “Short paragraphs, no jargon, straight to the recommendation without preamble” is. Then: your work context. Do you work alone or with a team? More operational or strategic? Clients requiring formal outputs or informal contexts? These allow Claude to calibrate register and detail level automatically. Finally: what you don’t want. At least three Claude default behaviors that don’t fit your context. “Don’t use bullet points unless I explicitly ask” is a useful instruction. “I prefer concise responses” isn’t. Concise means different things to different people.

PHASE 3: Active Context. Load current state: things in progress that Claude needs to know to be useful this week, not in the abstract. Active projects with updated status. For each one, name, objective, next concrete step, any blocker. Not the full history. Today’s state. Decisions already taken that you don’t want reopened, at least three. And: what you’re avoiding. There’s always something. An email not yet written, a difficult conversation being delayed, a problem you know you have but haven’t faced directly. Putting it in context doesn’t solve it alone. But often Claude helps you understand why you’re avoiding it, and sometimes understanding that is already half the work. Close Phase 3 by asking Claude to generate the first complete version of your CLAUDE.md, with named sections, in a format you can copy directly into the Project instructions and update each week.

⚠️ Final warning: the CLAUDE.md doesn’t become powerful the day you write it. It becomes powerful in week two, then week three, then increasingly from there. The point isn’t a perfect document. It’s a structure that improves over time. The first version will be imperfect. That’s fine. Today’s imperfect is infinitely more useful than the perfect one you haven’t written yet. Start now with what you have. Perfection is irrelevant compared to consistency.

Subscribe now

☢️ Radioactive Humor: The System That Remembers Everything Except One Thing

I built my Claude context system over several months of work.

CLAUDE.md with full operational identity. Memory files for facts that need to survive between sessions. A router to manage requests. Project structure for every active piece of work. Hundreds of pieces of information accumulated, updated, kept in order with near-maniacal care.

A system I’m fairly proud of.

The first day I used it systematically, I opened Claude, loaded the context. Claude responded exactly right. Knew the clients. Knew the tone. Knew decisions already made. It was, finally, someone who genuinely knew me.

Then I closed the tab.

The next day I reopened Claude.

Blank slate.

The model has no memory between sessions. No matter how sophisticated the system you’ve built outside it: every time you open a new conversation, Claude remembers nothing. You bring the context. Every time.

The first reaction when you truly understand this is somewhere between frustration and absurdity. You’ve built a sophisticated system to give permanent memory to a tool that by nature doesn’t have it. It’s a bit like building a meticulous archival library for a chef who forgets every recipe each night and finds them again in the morning.

Then you understand why it works exactly this way.

The memory doesn’t live in the model. It lives in the files. Every morning, you bring the memory. This isn’t an architectural flaw. It is the architecture. Your context files will work with Claude today, with any improved model released in six months, and with any system you want to use in two years. You’re not dependent on proprietary memory that a platform controls and might change.

The system doesn’t eliminate forgetting. It works around it by building something worth remembering, in a place that doesn’t forget.

The problem wasn’t that Claude forgot. It was that we hadn’t built anything worth remembering.

Three things to take from this issue.

First: transaction mode isn’t wrong. It’s just the worst way to use a tool that could do something far better. Open, ask, close produces results. It doesn’t produce accumulation.

Second: the Karpathy system solves external knowledge: what you know about the world. The next step solves operational knowledge: who you are in your work. Different problems. Different solutions. Both worth building.

Third: the second brain’s advantage is in accumulated time. It doesn’t build in a week. It can’t be bought next week. It starts this week, with an imperfect first version, and updates from there.

Welcome back to the Bunker. See you soon.

Subscribe now

P.S. If you already have a similar system, even partial, reply and tell me how you’re using it. I’m genuinely curious to see how others are solving this problem, especially in sectors far from mine.

Share

See you at the next cup of AI Espresso.
– Matteo ☕

Welcome back to the Bunker.

Last week I told you about Leo 18 years old, zero technical background. Three months after a talk, he showed me a working system: integrated modules, solid architecture, a descriptive clarity I rarely hear from seasoned professionals. The question that kept me up wasn’t “how good is Leo.”

It was: if an 18-year-old regularly outperforms developers with five years on their CV, what does that tell me about who to hire? After going back through twelve months of interviews, I landed on a conclusion that sounds offensive: in many cases, experience has become an obstacle.

Not from lack of skill for something subtler. The problem isn’t technical competence. It’s paradigms. And paradigms, unlike skills, can’t be taught. They have to be uninstalled. Uninstalling is harder, slower, and more expensive than starting from scratch.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“Today it’s more cost-effective to take someone from zero and train them than to hire an expert. Not because of money. Because of paradigms.”

🧪 AI in Action

#1: The Hidden Tax on Expertise Why Experience Slows You Down

There’s a profile everywhere in the 2026 job market. Four, five, six years in. Knows the tools. Right frameworks, certifications in order. Answers technical questions without hesitation.

Put them in front of a real problem with AI available and something odd happens. They use AI. But as an accelerator of the old workflow, not a replacement for parts of it. Write code, then ask AI to review it. Build the campaign, then ask AI to optimize copy. AI entered the process. The process hasn’t changed.

This isn’t a character flaw it’s the natural result of patterns reinforced hundreds of times. Competence itself has become an obstacle.

In the old paradigm, knowing how was the advantage. In the AI paradigm, it’s often what prevents you from letting AI do it better and faster.

McKinsey Global Institute (2024): 67% of primary resistance to AI-first transformations came not from juniors, but from middle management and specialists with 5–15 years of experience. Seniors were more adaptable. Juniors integrated natively. The bottleneck: expert enough to have consolidated paradigms, not senior enough to have uninstalled one yet.

From the Counter: I interviewed a developer a few months ago solid, five years in. In a six-week project with dozens of architectural decisions, he’d found one hour per week to insert AI without disturbing anything else. Not because he was incompetent. Because he was very competent. He didn’t feel the need to question it.

The cost of an expert isn’t just the salary. It’s salary plus the cost of de-training. And de-training isn’t guaranteed.

Subscribe now

#2: The Blank Slate Advantage Why Beginners Outperform Experts

Leo didn’t know how to build a database. Had no opinion on which architecture was correct or which ORM to use. When he hit the problem, he did the only sensible thing: described the situation to Claude as precisely as possible and implemented the solution he received.

He didn’t adapt it to “how it was done before.” There was no before. He took the optimal solution for 2026 and implemented it directly.

This is the blank slate advantage. Not enthusiasm. Not willingness. The absence of pre-installed paradigms filtering solutions through the lens of the old system.

Stanford HAI (2024): workers with less domain-specific experience show productivity gains averaging 34% higher than experts with identical tools available. The mechanism: they integrate AI as a native workflow component. Experts layer it on top of an already-defined process. Full efficiency for those starting from zero.

This window is limited. The beginner starting now can build their entire mental system in the AI-first paradigm without passing through the previous one. Once they accumulate their own paradigms, the window closes. If those paradigms are installed correctly, the advantage becomes permanent.

For those with ten years of experience: not too late. Make your paradigms explicit “how would I approach this if I’d never done anything like it before?” Harder. Produces results.

The blank slate isn’t a temporary condition to compensate with accelerated training. It’s an asset to exploit before it fills with habits built on the wrong paradigm. Hiring from zero and training AI-first protects that window.

Subscribe now

#3: How to Train AI-First Installing the Right Mindset from Day One

Training someone AI-first doesn’t mean teaching them to use Claude. It means installing the correct mental paradigm before the wrong one arrives.

Three principles hold this together.

First: output before process. Define what you want to obtain before thinking about how. “What do I want to get?” comes before “With which tool?” Most professional training works in reverse. Leo always starts from the end-user experience what should they accomplish, what should they feel only then comes the question of how to build it.

Subscribe now

Second: rapid failure as method, not accident. No extended preliminary study. Direct action, immediate feedback, rapid correction. The most efficient learning cycle when feedback is nearly instant and error cost is near zero.

Third: AI as construction partner, not oracle. The oracle receives a question and returns an answer. The partner receives a contextualized problem, contributes, receives feedback on where the solution doesn’t fit, and iterates. Generic output vs. useful output.

Accenture Future of Work (2024): companies with AI-first onboarding programs report 52% reduction in ramp-up time, and 31% higher 12-month retention. Those trained in the right paradigm skip the resistance-and-de-training phase entirely.

The mistake to avoid: training on specific tools instead of paradigm. “Use Claude like this” creates tool dependency. “Here’s how you think through it with AI” creates autonomy in any future context.

Subscribe now

#4: The 5 Cognitive Signals That Predict AI-Era Performance

The characteristics that make someone excellent in the AI era aren’t technical. They’re cognitive and behavioral.

Operational curiosity. Not “I like to learn” everyone says that. The concrete answer to: “When you have a problem you can’t solve, what do you do?” Interesting answers are specific and active: they describe a real process, not an intention.

Ambiguity tolerance. Can they work on a poorly-defined problem without freezing? Advance with incomplete information, making explicit assumptions rather than waiting for a perfect brief? The AI paradigm constantly requires moving forward with partial data.

Critical output evaluation. Do they know when AI’s result is wrong, incomplete, or superficial? This is the sharpest dividing line between those who use AI well and those who don’t. The person who asks “is this actually useful for my specific context?” produces results. The one who copies and pastes produces mediocre volume.

Iteration speed. How fast do they correct course when something doesn’t work? The error-feedback-correction cycle is the metabolism of the AI era.

Communicative clarity. Can they describe a problem with precision, in different contexts? This is the key competency for effective prompting. Those who describe problems precisely get better output with any tool.

PwC Workforce of the Future (2024): top-performing companies weight “learning agility” and “output evaluation skills” at 60% of the selection process for AI-augmented roles, versus 18% for the average company.

Best practical test: real problem, access to Claude, thirty minutes. Don’t evaluate the solution evaluate the process. How they decompose the problem. How they handle when the first answer misses. That tells you more than any CV.

Subscribe now

#5: Your Blank Slate Is a Competitive Asset How to Position Yourself Right Now

If you’re starting now, you arrived at the right moment.

You’re not late. You’re ahead of everyone who has to uninstall years of wrong paradigms before doing what you can do directly from day one.

The basics in 2026 are: clearly imagining what you want to build and communicating it to AI precisely enough to guide it toward that output. These aren’t acquired by studying. They’re acquired by building real things, evaluating results, correcting, restarting.

Build something. Anything. A system, an app, a process, a piece of content. Imperfect is fine. What matters is that it exists that you can describe what it does.

LinkedIn Workforce Report (2025): “AI-native” mentions in job postings grew 480% in 18 months. 73% of companies report difficulty finding candidates with an “AI-first mindset” despite plenty of candidates with basic technical skills. The gap isn’t tools. It’s paradigm.

From the Counter: this profile is rare not because it requires special talent it’s rare because most people are still waiting to “be ready” before starting. Readiness in the AI paradigm isn’t a condition you reach before doing things. It’s a condition you build by doing things.

The beginner of today is the native expert of tomorrow. Without paradigms to uninstall. The countdown for everyone else has already started.

Subscribe now

📌 Espresso Prompt Map the Paradigms to Remove and Build an AI-First Onboarding

Two objectives: identify where old-system paradigms block AI adoption in your team, and build an onboarding that installs the right one from day one. Paradigm resistance isn’t visible on the surface it shows up in micro-decisions, operational habits, the way people describe problems. This system makes it explicit.

Open Claude in a Project dedicated to operations or HR, so you can build contextual memory over time. Use this system prompt:

“You’re a senior organizational transformation consultant for the AI era. I have a team of [e.g. 6 people in a B2B SaaS / creative agency of 12 / 8-person sales team] and I need to understand where AI adoption is blocked by old-system paradigms, and how to structure onboarding for new hires. Three phases. Start with PHASE 1.”

PHASE 1 Team Audit. What tasks are still done “the way we always did it” despite AI that would substantially optimize them? Be specific: not “meetings” or “reports” which precise step in which process. Granularity is everything. Which profiles show the most resistance how many years of experience, which function, on which type of task? The goal is the demographic pattern of resistance, not the individual to blame. Where is “I already know how it’s done” blocking the most efficient solution? Three concrete cases from the last 90 days not hypotheses. End Phase 1 by asking Claude: “Based on what I’ve told you, what is the central paradigm my team needs to uninstall? One sentence.”

PHASE 2 Build the Onboarding. Design the entry plan for your next hire, calibrated to your team and the paradigm identified in Phase 1. Define three AI-first principles as observable behaviors not “have an open mindset,” but “when output doesn’t satisfy, reformulate the request instead of accepting it or abandoning the tool.” Structure week one around building something real within five days: an actual problem from your context, AI access, verifiable output. Then ask Claude to generate three 30-day check-in questions not “can they use Claude for X,” but: when they encounter a new problem, do they try AI before asking for help? Do they evaluate the output critically? Do they iterate quickly?

PHASE 3 Redesign Hiring. Three interview questions that reveal learning agility and ambiguity tolerance what you’re looking for in a positive answer and the red flags. Questions about past behavior reveal paradigm better than hypotheticals. A 30-minute practical test: real problem from your domain, access to Claude, instructions for the candidate, evaluation rubric focused on process not final solution, debrief questions. Rewrite your job posting to attract AI-first profiles not by adding “AI experience required,” but by writing it so that whoever has the right paradigm immediately recognizes it as theirs, and whoever doesn’t finds it strange.

⚠️ Final warning: this prompt helps you build the system, not find the right answer that depends on your sector, your team, your specific context. If after Phase 1 you haven’t identified at least three concrete paradigms to uninstall, you answered too superficially. If your Phase 2 answers are generic and don’t differentiate your team from any other, you haven’t done the work. The value is directly proportional to your willingness to be precise and uncomfortable. Start over with more honesty.

Subscribe now

☢️ Radioactive Humor The Interview That Changed How I Hire

I’ve run a lot of interviews this past year.

The most common profile: five years of experience, two-page CV, right frameworks, certifications in order. Then I ask: “How do you use AI in your daily work?”

Silence.

Or worse “I sometimes use ChatGPT for ideas,” tone of someone confessing something marginal. Followed by: “but for the important stuff I prefer doing it myself.”

Then I had 45 minutes with Leo. No CV. No certifications. He opened directly: “So, I have this module that does X, then I have this database problem, and I’m thinking of adding Y because the user experience here isn’t working.” Not answering a question. Describing a system he knew from the inside with the same naturalness a chef talks about a recipe they’ve cooked two hundred times. No performance. Just the work: solid, verifiable, done.

The categories we use to evaluate people were built for the old system. Which doesn’t exist anymore. We keep using the same questions, same proxies, same criteria as if the context hadn’t changed radically in two years.

In your next interview, add one question: “Show me something you’ve built recently.” Not what you know how to do. What you’ve done. The answer says everything.

Leo’s story isn’t about an exceptional kid to admire from a distance.

It’s about what happens when the evaluation system gets updated. Leo didn’t change. The paradigm did. And whoever keeps evaluating with the eyes of the old system will keep choosing the wrong profiles, training them the wrong way, and wondering why results don’t come.

Three things: intermediate expertise costs salary plus de-training and de-training isn’t guaranteed. The blank slate is a real competitive asset in 2026, not a temporary condition. AI-era selection runs on cognitive characteristics, not certifications.

Subscribe now

P.S.: If you’re hiring right now, or you’re the one looking for work, reply to this. I’m curious to hear which of these paradigms you’re running into.

Share

See you at the next cup of AI Espresso.
– Matteo ☕

• Zero coding, zero experience, working system built in 3 months

• No habits to unlearn = faster AI adoption than veterans with 5-year CVs

• Imagination is now a technical skill. Those who have it, win.

A few months ago I walked into a high school classroom. I talked for an hour about AI: how to actually use it, why most people waste it, what separates those who build with it from those who outsource their thinking to it. One message: use it as an amplifier, not a substitute. Polite applause, a couple of questions, goodbyes. I drove away thinking maybe three out of thirty would do something concrete with it.

Months later, a message from one of those students. He wanted to show me something he’d built. I expected a rough prototype. I joined the call ready for managed enthusiasm.

Forty-five minutes later, my head was spinning.

It wasn’t a prototype. It was a working system, structured, modular, architecturally sound. Built by a kid who three months earlier had never written a line of code. I thought about the developer interviews I run for Morfeus, the five-year CVs, declared frameworks, listed certifications. How many of them would have produced something comparable? Very few. Not for lack of technical skill, for something subtler, and much more expensive to fix. This episode is that story.

☕ Cup in hand. Let’s begin.

Subscribe now

💡 Thought of the Day

“It’s not about technical skill. It’s what you carry in your head. Can you imagine it? Can you actually see it?”

🧪 AI in Action - Why a Zero-Experience Beginner Outperforms 5-Year Developers

#1: The Story, Unfiltered

He’s 18, studies at a scientific high school. Something from that morning clicked.

He didn’t look for courses. Didn’t watch YouTube tutorials. He picked up Claude and started building.

The call was scheduled for thirty minutes. It ran forty-five, not because the technology impressed me, but because of the clarity with which he described his project. He knew what every part of the system did. He knew why he’d chosen that structure over another. He knew where he was going and what was missing to get there. Not improvising on the call, describing something he had fully formed in his head. A precision I rarely hear from seasoned professionals.

What he built: a gamified personal system, multiple life areas integrated into a single application, each with its own logic, all connected by a central engine tracking progress over time. Not a simple app. Not an experiment. A system with the ambitions of a startup with three or four engineers.

Subscribe now

He was in the middle of a database migration, started with a local solution for prototyping, now needed something more robust. Building the migration with Claude, step by step, evaluating every technical decision as he went. He didn’t know what an ORM was. Didn’t know what a “migration” meant in the formal sense. He knew what he wanted: persistent data, organized, scalable. And he knew where to go to build it.

I asked how he approaches a technical problem he’s never seen before. No hesitation: “I don’t watch YouTube videos because I know it’ll mess up my thinking. I start working and the AI helps me understand.”

No performance. Just method. Direct action, immediate feedback, rapid correction. The most effective learning cycle when AI responds in real time and the cost of error is nearly zero.

My assessment: he’s in the top 10% of all AI users I know. In three months, from zero, what he built is genuinely remarkable. I’ve interviewed developers with years of experience who couldn’t do the same. Not an isolated case, a pattern.

The real question: why him, and not so many others with far more declared experience?

#2: The Blank Page - Why Starting From Zero Is a Structural Advantage

The easy answer: “he’s talented.” Probably true, in part. But talent explains a marginal advantage, not a systemic gap.

The more useful answer: he had a blank page.

No accumulated mental schemas about “how you build an application.” No habits formed over years on specific stacks. And crucially, no resistance to the AI-first way of building, because he’d never known any other way. Nothing to uninstall.

When he hit a technical problem he didn’t recognize, he did the most natural thing in this context: described it to Claude in the most precise language he had, evaluated the response, iterated until the solution worked. No pre-existing workflow to fight against. He took the optimal solution for the current context and implemented it directly.

This is the beginner’s paradox in the AI era, counterintuitive, worth understanding: when technical competence becomes accessible to anyone through AI, the competitive advantage shifts to the ability to adopt the right paradigm without resistance. On that axis, whoever never learned the old paradigm starts ahead.

Stanford HAI quantified this in 2024: workers with less domain-specific experience showed higher average gains from AI on professional tasks than experts, with identical tools available. The mechanism: they integrate AI as a native component of their workflow, while experts layer it on top of a process already defined. Partial efficiency for the experienced. Full efficiency for those starting fresh.

Subscribe now

#3: The Paradigm That Makes All the Difference

The wrong way: arrive with an open problem, ask “what should I do,” receive a generic answer, apply it, get disappointed. Empty input, empty output. The model doesn’t have your context, your history, your definition of “good.” It can’t give you the right answer if it doesn’t know the real question.

The right way: arrive with a problem you’ve already analyzed enough to know what you don’t know. Describe it with maximum precision. Evaluate the response critically. Identify where it fails for your specific case. Return with that added precision. The brain never stopped. The AI didn’t replace the thinking, it accelerated the execution of thinking already formed.

MIT, 2023: Brynjolfsson’s team studied knowledge workers using AI in real professional contexts. Average productivity gain: 14%. But segmented by usage type, the gap was enormous, those who used AI as a “thinking partner” showed gains of up to 40% on qualitative output. Those who used it as a passive executor came in well below average.

Two marketers. One prompts: “write an Instagram post about this product.” Gets something generic and unusable. The other: “I’m trying to reach 30-year-olds who work too much to go to the gym, I want to communicate that this product fits their routine without asking them to change it, I need a headline that stops the scroll without promises that won’t hold.” Gets something directly usable.

That distance between inputs is the distance between two paradigms. In the first, AI is a generator. In the second, it’s the amplification of a thought that was already formed.

When I asked how he builds each part of his system: he starts from the experience he wants the user to have. Defines what should happen, what someone using that part should feel. Then asks Claude how to build it technically. When he hits a problem, he returns with it specific and well-defined. Claude doesn’t think for him. He thinks. Claude builds.

Director, not actor. Architect, not laborer. Not in theory, in the daily practice of building something real.

Subscribe now

#4: What “Understanding AI” Actually Means in 2026

Understanding AI in 2026 means one thing: being able to imagine the output you want, communicate that vision precisely enough to guide the model toward it, and evaluate the result with critical eyes. Three competencies. None technical in the classical sense. All three built through the practice of making real things, not studying concepts.

The core distinction: architect vs. laborer. The digital laborer executes, writes code, produces content, compiles documents. Skilled, indispensable in the old system, increasingly replaceable by AI across a growing share of cases. The architect knows what to build, why to build it, how to evaluate whether the result is the right one. Not replaceable, because AI doesn’t have their history, their context, their specific goals. It can build. It cannot decide what to build.

BCG, 2024, most demanded skills in AI-first organizations. Not technical competencies at the top. Systemic ones:

Critical thinking, evaluating output with critical eyes.
Problem decomposition, breaking a complex problem into addressable blocks.
Output evaluation, knowing when a result is good and when it isn’t, without knowing the implementation details.

The same report classifies basic technical skills as commodities: accessible, widespread, no longer differentiating.

The practical implication: making a list of technical skills to acquire is the wrong approach. The right question: “What do I want to build? Can I imagine it clearly enough to describe it to an AI?” If the answer is no, it’s not a tool problem. It’s a thinking problem. And that doesn’t get solved with a ChatGPT course.

Imagination has become a technical competency, in the operational sense, and the structural advantage it confers has nothing to do with declared certifications.

Subscribe now

#5: The Formula - Clear Thinking + AI = Unbeatable Advantage

It starts with output vision: not “what do I want to do” in a vague sense, but “what do I want to exist in the world when I’m done.” One produces process, the other produces result. He always starts there, what should the user do, feel, find. Without that initial clarity, everything that follows is building blind.

From that vision: decompose into communicable blocks. Not “build me a productivity system”, “I want the user to see their weekly progress in a summary panel, with a visual indicator for each area.” One block at a time. The AI enters here: not as an oracle, but as an operational partner on a specific, well-defined problem. The precision of the input directly determines the quality of the output.

Then the brain doesn’t stop. Read the result. Identify where it works and where it fails for the specific context. Bring that precision to the next cycle. This is the phase that separates good output from average output. Then restart: error, feedback, correction. The speed of that cycle is the metabolism of the whole system.

Subscribe now

Asana, Anatomy of Work 2024: the variable that most explained differences in results wasn’t the tool, the sector, or prior experience. It was whether the worker kept control of the “what” and delegated the “how”. Those who did: average 3.8x increase in output on executive tasks. Those who inverted it — delegating the “what,” controlling the “how”, got marginal or zero results.

In three months: grades up, study time halved on the same subjects. A working system with integrated modules and architecture that holds. Already building the next version while managing high school full time.

From the Counter: when I look at this, I don’t see an exceptional case to admire from a distance. I see the operating model I look for when interviewing for Morfeus. The person who knows what they want to build, can describe it precisely, can use AI to build it, and can evaluate the result. Rare to find — not because it requires special talent, but because most people never developed this mental structure, occupied as they were developing the technical competencies of the previous paradigm.

The advice I gave that high school classroom came back transformed into concrete proof. It’s not about technical skill. It’s what you carry in your head. Can you imagine it? Can you actually see it? This kid demonstrated it.

Subscribe now

☢️ Radioactive Humor - The Proof Came Back

I went to that high school to teach.

I said: use it as an amplifier, not a substitute. As cognitive sparring, not a crutch. Don’t delegate the thinking, use it to think deeper.

Someone in that room took that idea, carried it out, built something on top of it, and three months later brought it back transformed into concrete proof. A working product. A replicable system. A demonstrated competency.

The advice worked. I didn’t know it with certainty when I gave it. Now I do.

What he did is replicable, not because he’s exceptional, but because the formula works: precise output vision, decomposition into blocks, AI as accelerator, critical evaluation, rapid iteration. Any domain. Any age. Any starting point. Imagination has become a technical skill. Those who have it, win.

Next week: what this story taught me about who to hire in the AI era. Because if a kid with no experience regularly beats candidates with five years on their CV, that says something very precise about what to look for, and what to stop looking for.

See you at the next cup of AI Espresso.
– Matteo ☕

Subscribe now

P.S. If you know someone convinced they’re “too far behind” to start with AI, or waiting to finish a course before beginning, send them this episode. 17 years old. Zero experience. He just started building.

Share

Today at the counter, we’ve got a problem of proportions.

Not cosmic ones. Everyday ones. The kind you don’t notice until someone points them out, and then you can’t stop seeing them everywhere.

Picture this: you walk into a coffee shop and order an espresso. The barista looks at you, nods, and then launches into it:

“Excellent choice! Espresso is one of the most beloved beverages in the world, with roots tracing back to early twentieth-century Italy. Before I prepare it, let me give you some context: our blend is a mix of Arabica and Robusta with notes of dark chocolate and a slightly spiced finish. Proceeding now with preparation. Here is your espresso. I hope this experience has been valuable! Feel free to ask if you’d like more information about our full hot beverage menu.”

You wanted a coffee. You got a coffee, three paragraphs of context, a disclaimer, and a motivational sign-off.

Now swap the barista for Claude. Or ChatGPT. Or whatever model you’re using.

Same thing. Every single time. And it’s costing you a lot more than you think.

Today we’re talking about a problem 99% of AI users don’t know they have: bloated responses. And about a fix that comes from a 19-year-old Dutch kid, a scientific paper that flipped the expectations of half the academic world, and a principle that, once you get it, changes the way you talk to any AI forever.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

“Why waste time say lot word when few word do trick?” , Kevin Malone, The Office.

Also: a March 2026 paper. 31 models tested. Same conclusion.

🧪 AI in Action

#1: The Crisis Nobody Explained to You

If you’ve noticed lately that Claude cuts you off more often, that conversations end sooner, that your $20-a-month plan feels like it’s worth less every week , that’s not in your head.

That’s exactly what’s happening.

In March 2026, Claude users started reporting something off. Rate limits , the caps that decide how many questions you can ask before the system puts you on pause , were hitting the wall much faster than usual. Not twice as fast. Three times. In some cases, fifty.

Users on the $200-a-month Max plan reported burning through their entire monthly quota in 70 minutes. Seventy minutes. For a monthly subscription.

Someone did reverse engineering on the system and found two independent bugs in the prompt cache that were silently inflating token consumption by 10 to 20 times. You were paying for one conversation and the system was counting ten.

Anthropic acknowledged the problem. Fixed a few things. But the underlying direction didn’t change: limits got tighter, peak-hour penalties got steeper, and the subscription that a year ago felt like near-unlimited access to a premium model now feels like a demo with a countdown timer.

And it’s not just Claude. OpenAI launched a $100-a-month Pro plan. API prices stayed high. The market is moving toward one simple reality: tokens cost money, and whoever doesn’t manage them wastes them.

Most people react to this in one of two ways. Option one: complain on Reddit. Option two: pay more.

There’s a third way. And it’s the one worth talking about.

Subscribe now

#2: The Problem You Didn’t Know You Had

AI models are trained to give answers. So far, so good.

The problem is that “give answers” became “give context about the answer,” which became “be exhaustive,” which became “never shut up under any circumstances.”

You ask: “What’s the capital of France?”

It answers:

“Great question! The capital of France is Paris, a city rich in history and culture located in the Ile-de-France region. Paris is known for its iconic landmarks such as the Eiffel Tower, the Louvre, and Notre-Dame. Originally founded as a Roman settlement called Lutetia, the city has grown into one of the most important political, economic, and cultural centers in Europe. I hope this information is helpful! Feel free to ask if you’d like to know more about European geography.”

You wanted one word. You got a paragraph, a disclaimer, and a motivational sign-off.

Sounds like a minor annoyance. It isn’t.

Every word an AI produces is a token. Every token costs. Time, money, usage limits burning down. When the model serves you 135 tokens where 2 would’ve done the job, it’s not giving you better service. It’s charging you 67 times more for the exact same information.

Anatomy of a bloated response. Four parts. You’ll recognize all of them.

The preamble. “Great question!”, “Of course, happy to help!”, “This is a fascinating area.” Useless. Adds zero information. It’s the digital equivalent of the barista asking how your day’s going before handing you the coffee you already ordered.

The filler. Three paragraphs where one was enough. Context you didn’t ask for. Explanations of things you already know. The model is padding because it was trained to pad, not because you need it.

The recap. After saying the thing, it says the thing again. “In summary, as we’ve seen...” No. You said it thirty seconds ago. The reader doesn’t have amnesia.

The sign-off. “I hope this was helpful!” “Don’t hesitate to ask!” The model clocks out like an overly polite waiter. Adds nothing. Costs tokens.

Now, here’s the part that surprised me.

In March 2026, an academic paper tested 31 different models on nearly 1,500 problems. The title is dry: “Brevity Constraints Reverse Performance Hierarchies in Language Models.” The finding is not.

When you force a large model to give shorter answers, accuracy improves by 26 percentage points.

Not worse. Better.

The explanation makes sense once you sit with it. When the model can’t hide behind qualifiers, vague hedges, “it depends,” “generally speaking,” “in most contexts,” it’s forced to commit to a concrete answer. No safety net. And concrete answers tend to be more accurate than answers wrapped in three layers of diplomatic caution.

The paper makes one point that reframes everything: the problem isn’t the model’s brain. The brain reasons the same way regardless. The problem is its mouth. The way it packages the answer. All that verbosity isn’t a sign of intelligence. It’s a design bug.

Fewer words. More precision. Lower cost.

And someone built a system to force exactly that.

Subscribe now

#3: The Solution Is Called Caveman Mode

The idea is simple. Almost stupid in its simplicity. Which is exactly why it works.

Force the AI to communicate like an ultra-efficient caveman, and output tokens drop by 60 to 75%. The reasoning stays identical. The capability stays identical. The only thing that changes is the packaging: out go the preambles, the small talk, the recaps, the motivational sign-offs.

Caveman Mode not waste time. Caveman Mode get to point.

Not a joke. A method. And it’s been spreading through tech circles for weeks under one name: Caveman Mode.

The origin is a mix of things. There’s a legendary essay in the programming world, “The Grug Brained Developer” by Carson Gross, preaching radical simplicity in code. There’s Kevin Malone from The Office and his existential philosophy: “Why waste time say lot word when few word do trick?” And there’s a 19-year-old Dutch kid, Julius Brussee, a Data Science student at Leiden University, who took all of it and turned it into a concrete system that exploded on GitHub in April 2026: first in trending, then on Hacker News with 883 points and 361 comments, then picked up by every tech outlet from Hackaday to Decrypt to Yahoo Finance.

But before I walk you through the repo, here’s the version you can use right now. No installation required.

The prompt. Copy and paste it into Claude’s custom instructions, ChatGPT’s Custom Instructions, or the system prompt of whatever model you’re using.

—

Respond terse like smart caveman. All technical substance stay. Only fluff die.

Persistence: ACTIVE EVERY RESPONSE. No revert after many turns. No filler drift.

Still active if unsure. Off only: “stop caveman” / “normal mode”.

Default: full. Switch: /caveman lite|full|ultra.

Rules: Drop: articles (a/an/the), filler (just/really/basically/actually/simply),

pleasantries (sure/certainly/of course/happy to), hedging. Fragments OK.

Short synonyms (big not extensive, fix not “implement a solution for”).

Technical terms exact. Code blocks unchanged. Errors quoted exact.

Pattern: [thing] [action] [reason]. [next step].

Not: “Sure! I’d be happy to help you with that. The issue you’re experiencing is likely caused by...”

Yes: “Bug in auth middleware. Token expiry check use < not <=. Fix:”

Intensity:

Level

What changes

lite

No filler/hedging. Keep articles + full sentences. Professional but tight.

full

Drop articles, fragments OK, short synonyms. Classic caveman.

ultra

Abbreviate (DB/auth/config/req/res/fn/impl), strip conjunctions, arrows for causality (X → Y), one word when one word enough.

Auto-Clarity: Drop caveman for: security warnings, irreversible action confirmations,

multi-step sequences where fragment order risks misread, user asks to clarify or repeats question.

Resume caveman after clear part done.

Boundaries: Code/commits/PRs: write normal.

“stop caveman” or “normal mode”: revert. Level persist until changed or session end.

—

Once pasted in, the AI keeps reasoning at the same depth. It just stops gift-wrapping its answers in cellophane.

Before and after. Same task. Same question. Different results.

Task: “List 3 advantages of cold email for sales.”

Without Caveman Mode (~135 tokens)

With Caveman Mode (~32 tokens)

Great question! Cold email is a very powerful strategy... [Spero che questo ti dia una base solida! Non esitare a chiedere se vuoi che approfondisca qualche punto.]

1. Cheap → cents per lead vs dollars in ads2. Scalable → automate to thousands, no extra hires3. Measurable → opens, clicks, replies in real time

Same answer. 75% fewer tokens. Zero filler.

Now do the quick math. If every response you get is bloated like this, and you’re sending 30 to 40 prompts a day, you’re burning three quarters of your output tokens on words you don’t even read. Those tokens are your rate limits draining. They’re your $20 a month evaporating. They’re the reason Claude cuts you off mid-conversation.

When to use it. And when not to.

Use it for: information requests, analysis, summaries, data extraction, repetitive tasks , anything where you want the result, not the narration of the process.

Don’t use it for: brainstorming and creative writing (you want the model to explore, not compress), step-by-step explanations you need to fully understand, complex reasoning where the process matters as much as the conclusion.

Caveman Mode isn’t for everything. It’s for the 70% of the time you’re asking something concrete and getting back a middle-school essay.

Full repository here: https://github.com/JuliusBrussee/caveman

Subscribe now

#4: What You Actually Save (No Noise)

The number making the rounds everywhere is 75%. And it’s true , but with an asterisk the size of a billboard.

The 75% applies to output tokens, meaning what the model sends back to you. But when you use AI, the cost isn’t just the response. There’s also the input: your prompt, the conversation history, attached files, system context. And you pay for all of that regardless.

If you’re sending 10,000 tokens of context and the model responds with 200, Caveman Mode shrinks those 200 down to around 50. Good. But you’re still paying for the 10,000 tokens of input.

The real total savings on a typical session? Between 20% and 40% of your overall bill.

Less impressive than 75%? Yes. But run the numbers.

If you’re on the API doing 10,000 calls a day with an average of 200 output tokens on Claude Sonnet, your daily cost is around $30. With Caveman Mode, it drops to about $9. Over a year: more than $7,600 saved. On output alone.

And there’s a detail the math doesn’t capture but that matters just as much: output tokens cost 5 times more than input tokens. At current API rates, Claude Sonnet runs $3 per million input tokens and $15 per million output tokens. Cutting output is the highest-leverage move you have.

But even if you’re not on the API. Even if you’re on the $20-a-month Pro plan and not paying per token.

Rate limits are calculated on tokens consumed. Fewer tokens per response = more responses before the wall hits. What used to be 15 conversations before the limit becomes 40. Same subscription. Same price. A lot more runway.

Caveman Mode isn’t a developer trick. It’s digital hygiene for anyone who uses AI every day.

Julius Brussee, the 19-year-old who built the automated system around this principle, created an open-source repository that now has nearly 7,000 stars on GitHub. It includes three compression levels, a tool that compresses your input files by 45% as well, and detailed benchmarks task by task. If you want to go beyond the prompt and automate all of this, the Vault this week has the full deep dive: setup, levels, real numbers, and the general principle you can apply to any AI workflow.

Subscribe now

📌 Espresso Prompt: Your Compression Switch

This one’s different from the usual. It doesn’t run an analysis. It doesn’t build a workflow. It gives you a switch.

Paste it at the start of any conversation with Claude or ChatGPT:

From this point on, respond in ultra-compressed mode. Result first, no preamble.

No closing summaries. No motivational sign-offs. No “Great question!”.

Short sentences, fragments acceptable, symbols instead of words where possible (→, +, =).

Code and technical terms stay intact.

If I need a detailed explanation, I’ll ask explicitly with “explain in detail”.

Otherwise: caveman mode on.

Two lines. Zero setup. Works immediately.

Try it on your next prompt. Compare the response to what you were getting before. The difference is visible on the first shot.

Subscribe now

☢️ Radioactive Humor

I turned on Caveman Mode on Claude and asked it to describe what it does.

Before:

“I am an advanced AI assistant developed by Anthropic, designed to help you with a wide range of tasks including writing, analysis, coding, research, and much more. My goal is to provide accurate, helpful, and detailed responses, adapting to your communication style and specific needs. Feel free to ask me anything!”

After:

“AI. Answer. Useful.”

Three tokens. Same meaning. I saved enough to ask it another question.

The moral, from behind the counter: the problem was never the AI’s intelligence. It was always its manners. Too polite. Too verbose. Too trained to look helpful instead of actually being helpful.

Caveman Mode doesn’t make it dumber. It makes it more honest.

And apparently, more accurate too. Science confirmed it.

The bunker is back to its usual self.

Three things to take with you today:

• 75% of the words the AI serves you weren’t requested, you don’t read them, and they cost tokens. Every bloated response is a chunk of your subscription evaporating into preambles and motivational sign-offs. The problem isn’t the model. It’s the default format.

• A two-line prompt can trim that fat by 60 to 75%. Caveman Mode doesn’t touch the reasoning, it touches the packaging. Fewer words, more precision: that’s a scientific finding, not an opinion.

• Whoever controls how the AI responds controls the cost, the speed, and the quality. That’s a skill, not a trick. And like every skill, whoever learns it first has an edge over everyone who doesn’t even know it exists.

☕ Matt

Subscribe now

P.S. If this issue made you want to scroll back through your recent AI responses and count how many words were useless, send it to someone who keeps complaining that “Claude doesn’t work like it used to.” Maybe it works exactly the same. Maybe it just talks too much. As always, that’s the barista talking, not the algorithm. ☢️

Share

☕ The counter’s closed , see you in the next drop.

We saw why, without a shared structure, AI turns into a collection of personal shortcuts.
And why every team that “plays solo” doesn’t create value… it just multiplies chaos.

This issue is the next step.
Because in a world where everyone sell prompts, hacks, and tools in their private drawer, survival comes from pouring everything into one shared tank.

Ready? Then raise your cup-because today we enter the heart of the Share or Die framework: the practical rules that turn AI from a personal toy into a collective ROI engine.

Subscribe now

💡 Thought of the Day

“A prompt hidden in a drawer is a wasted idea.
A shared prompt is an asset that works while you sleep.”

🧪 AI in Action: Share or Die

There’s always someone at the counter whispering,
“AI saves me a ton of time.”
True.
But if what you learn-the prompts you refine, the mistakes you stop making-stay only in your cup, then all you’ve done is speed up yourself.
Your company, your team, your colleagues?
Still stuck at the starting line.

Today I’ll show you how to run the right current through the system: structured sharing.
Not hallway gossip, but clear procedures, shared libraries, defined ownership.

Because AI is an amplifier:

pour scattered knowledge into it → it amplifies noise.
pour shared knowledge into it → it amplifies ROI.

That’s why the Share or Die framework exists: six practical rules that turn sharing into a measurable asset-not a nice gesture.

Subscribe now

⚙️ 1. Centralize

Picture the post-blast bunker.
Some survivors stash supplies in ten different boxes-one under the bed, one in a drawer, three in random cabinets… and when the alarm goes off, they never find what they need.
Then there’s the one with a central warehouse: shelves, labels, a logbook.
When danger hits, they know exactly where to reach.

Guess who lasts longer?

AI works the same way.
Every team, every person, keeps their own “digital drawer”: private Google Drives, WhatsApp chats, desktops packed with final-final-vero-3.docx.
The result? No one knows where the “truth” lives.
Processes get duplicated, errors get recycled, and the time wasted searching exceeds the time saved using AI.

👉 The first rule: build one central Vault.
A single digital home where all AI assets land:

• validated prompts

• step-by-step procedures

• outputs that actually worked

• internal FAQs

• replicable use cases

It doesn’t matter which tool you choose-Notion, Confluence, SharePoint, an AI-driven DB.
What matters is this: one. Not two. Not three. One.

💡 Common mistake: creating a zoo of workspaces and multiplying chaos instead of reducing it.
A central Vault only works if everyone sees it as the official source.

📌 Practical case:
A sales team of 15 people stored their email prompts in personal files.
Result: 15 different versions-messy, inconsistent, sometimes worse.
After creating a single shared Prompt Library, they cut response times by 40% and aligned tone + quality across all customer interactions.

When you concentrate everything in one warehouse, value compounds.
Every new experience becomes collective knowledge, not a private secret.
In an age of information overload, survival isn’t individual-it’s collective.

Subscribe now

📑 2. Standardize

Imagine walking into the bunker’s warehouse.
Everything is centralized-great… but nothing is labeled.
One box says “stuff,” another one “misc.” A mountain of unmarked cans.
When it’s time to cook, you don’t know if you’re opening beans… or gasoline.
That’s what happens when you centralize without standardizing: a warehouse that kills you.

AI works exactly the same.
Putting all prompts and processes into one Vault is just the first step.
But if everyone writes them however they want-one in English, one like a telegram, one like a personal diary-you’ve simply moved the chaos into a bigger room.

👉 The rule here: common formats.
Every shared asset needs a standard structure so anyone can read it and reuse it.
A prompt, for example, should always include:

• Title (clear and short: “B2B follow-up email”)

• Goal (what it must produce: “Write a polite but firm reminder email”)

• Step-by-step instructions

• Example of a valid output

• Possible variations

• Tags for future search

  ---

💡 Common mistake:
Dumping screenshots, voice notes, or random files into the Vault.
Then every colleague who searches for a prompt must decode the style of whoever wrote it.
Result: more time lost than gained.

📌 Practical case:
A marketing agency had collected 300 prompts from its copywriters in an internal database.
But everyone uploaded them in their own style.
When a new hire tried to use them, 70% were incomprehensible or unusable.
After introducing one shared formatting standard, the “effective reuse rate” jumped from 30% to 90%.
Same warehouse-finally readable.

Standardizing is boring, yes.
But it’s what turns a pile of files into a company arsenal.

Subscribe now

🔗 3. Distribute

Centralizing and standardizing is like running a well-stocked, perfectly organized bar.
But if all the bottles stay locked behind the counter, what’s the point?
The customer stays thirsty.

AI works the same way: if the Vault lives in a lonely folder only three “tech enthusiasts” know how to access, you don’t have a knowledge base-you have a museum.
And museums, as we all know, don’t generate ROI.

👉 The rule: access must be easy and universal.
Every team should be able to tap into the same AI knowledge base without friction.
That means:

• clear permissions (who can read, who can update, who validates)

• integrated tools (the library must plug into what people already use: CRM, project management, intranet)

• constant diffusion (internal newsletter, monthly demos, reminders in meetings: “Use the Vault-don’t reinvent from scratch.”)

💡 Common mistake:
Building the perfect wiki or database… that nobody opens because “I didn’t know it existed” or “I don’t have time to look for it.”
Sharing only works when it becomes part of people’s daily flow.

📌 Practical case:
In a software company, the IT team had crafted an amazing prompt library.
The problem? It stayed locked with them.
Sales reps kept writing emails from scratch.
After integrating the library directly into the CRM (a simple “use prompt” button), adoption shot up to 85%, and response times to prospects were cut in half.

To distribute means to make knowledge ubiquitous: it must show up at the point of use, not sit forgotten in the back room.
Because a shared asset that isn’t actually shared… isn’t an asset.

Subscribe now

🧭 4. Assign Responsibility

Picture the Vault in the bunker: shelves in order, supplies labeled, open access for everyone.
Perfect. But if no one checks expiration dates, cleans the space, or restocks… you’ll be staring at rats instead of beans before long.

AI sharing works exactly the same.
Centralizing and distributing without clear owners leads to digital garbage: outdated prompts, duplicated versions, procedures that no longer apply.
And when a colleague goes searching for “the right recipe,” they find nothing but confusion.

👉 The rule: every area needs a knowledge custodian.
You don’t need an army-just one clear person per team or function whose job is to:

• update the library regularly

• validate what gets uploaded

• remove what’s no longer useful

• flag opportunities for improvement

💡 Common mistake:
Assuming sharing is spontaneous and democratic.
Result: everyone uploads, no one reviews.
Within months, the Vault turns into a digital landfill.

📌 Practical case:
In a consulting firm, each team uploaded its prompts to Notion.
After six months, there were more than 500-many duplicated or inconsistent.
They appointed a Knowledge Manager per area with a monthly review cadence.
End result: prompts cut down to 220 (the truly useful ones), quality improved, internal adoption up by 60%.

Assigning responsibility isn’t bureaucracy.
It’s about having sentinels of knowledge guarding the most valuable asset a company can build: its collective intelligence.

Subscribe now

⚠️ 5. Assess the Risks

In the bunker, having centralized, tidy, well-guarded supplies isn’t enough.
If you leave the hatch open-or worse, let secret maps and access codes slip into the warehouse-sharing stops being salvation and becomes a death sentence.

AI follows the same physics.
A shared Vault without risk governance is a boomerang: sensitive data, NDAs, legal notes, or customer info tossed in “by accident,” ready to spill into the wrong prompt or the wrong model.

👉 The rule: draw clear boundaries.
Before sharing, always ask:

• Does this data actually serve the team, or does it expose us legally?

• Can this process live inside the Vault, or does it need a locked channel?

• Which information is useful for AI-and which must never leave protected systems?

💡 Common mistake:
Uploading everything into the knowledge base.
Result: people panic about doing something wrong and either stop using the system… or, worse, use it recklessly.

📌 Practical case:
An insurance company built a prompt library for customer care.
One employee, trying to be helpful, added examples containing real customer data.
When the AI processed those examples, personal details resurfaced in outputs → massive legal risk.
After the incident, a simple policy was introduced: never real data-only anonymized templates.

Assessing risks isn’t paranoia.
It’s what separates an AI-ready organization from one playing Russian roulette with its reputation.

Subscribe now

📊 6. Measure (and Prove the ROI)

In the bunker, you can have perfectly organized shelves, vigilant guards, and reinforced doors.
But if you don’t track how many cans you consume-and how many extra days they buy you-you’re just hoping the system works.

AI works the same way: without numbers, sharing stays an act of faith.
Management only sees time spent uploading prompts, writing procedures, updating libraries.
The trick is showing the other side of the coin: the measurable return.

👉 What you should actually measure:

• Time saved: hours reclaimed thanks to shared templates and processes.

• Errors avoided: fewer wrong, inconsistent, or duplicated outputs.

• Onboarding speed: how much faster a new hire becomes productive thanks to the Vault.

• Internal adoption: % of teams who actively tap into the knowledge base.

💡 Common mistake:
Saying “it helps.”
A CFO or CEO isn’t moved by feelings-they want numbers.

📌 Practical case:
A mid-size retail company introduced a standardized prompt Vault for customer service and e-commerce.
After three months, the data showed:

• average customer response time: –45%

• communication errors: –30%

• new-hire training reduced from 6 weeks to 2

Translated into money: over €120,000 in annual ROI.

⚡ Key message:
Centralizing, standardizing, distributing, and assigning responsibility create order.
But measuring impact is what transforms sharing from bar philosophy into a strategic lever that produces real, tangible ROI.

Subscribe now

☢️ Share or Die Framework - In a Nutshell

Centralize → one single Vault, no jungle of files.
Standardize → clear formats, one shared language.
Distribute → knowledge must reach the point of use.
Assign responsibility → custodians who keep order and quality.
Assess risks → clear governance, no sensitive data wandering around.
Measure → concrete KPIs, not vibes.

👉 Without these six steps, sharing stays an act of goodwill.
With them, it becomes a measurable multiplier of collective intelligence-and real economic return.

Subscribe now

☕ Espresso Prompt - From Chaos to the Vault

At the counter, there’s always someone arriving with napkins full of scribbles, rushed ideas, brilliant notes no one can decipher later.
The truth is, most AI value gets lost exactly like that: useful outputs turned into dead ends because everyone writes them in their own style… and then forgets them.

This prompt turns every “draft” into a shareable asset ready for the company Vault.
No more scattered files-just clean, structured knowledge reusable by anyone.

Subscribe now

Prompt of the Week → From Chaos to the Vault

Act as a **personal Knowledge Manager**.  

When I give you a prompt or AI output I want to save, your job is to transform it into a reusable company asset.  

Always reorganize and return it in the following standard format:

1. **Short title** (clear and descriptive)  
2. **Goal** (what it solves, who it serves)  
3. **Step-by-step instructions**  
4. **Example of valid output**  
5. **Possible variations** (at least 2)  
6. **Tags for future search**  

If the content isn’t clear, ask me for missing information before saving it.  
You must make each asset readable and useful for anyone on the team-even months later.

☢️ Radioactive Humor

We’ve seen how sharing turns ideas into energy, how standardization gives them structure, and how circulation generates measurable ROI. In a post-apocalyptic world of hype, noise, and toxic information, the real competitive advantage isn’t hoarding knowledge. It’s putting it in motion.

In any self-respecting Vault, locked-up resources become waste. Shared ones fuel entire outposts.

And remember: in the aftermath, the winner isn’t the one sitting on the most secrets. It’s the one who knows how to put the right prompts in the right hands.

Matt

Subscribe now

P.S. Know someone still trying to figure out where to start? Bring them here. First cup’s on us.

Share

It’s 8:47 on a Tuesday morning.

You open your laptop.

On your desktop: a clean folder, a two-page briefing with everything you need to know today, and a competitive report on your main competitor, updated overnight.

You didn’t hire anyone. You didn’t work late. You didn’t open a single app.

You just left a message before going to sleep. And someone, something, did the rest.

Welcome to the counter.

Today I’m not serving you a regular espresso. I’m serving you 120 hours of distilled experience in a guide you can start applying today.

Because Cowork is not a Claude upgrade. It’s not “chat but with files.” It’s not a tool for tech nerds.

It’s the first time in history you can delegate real work to a machine, without writing a single line of code, without setting up anything complicated, without waiting for “AI to get a little better first.”

It’s already good enough. The problem is almost nobody knows how to actually use it.

This issue exists to change that. ☢️

Subscribe now

💡 Thought of the Day

“The goal isn’t to automate everything. It’s to automate what you should never have to do yourself.”

🧪 AI in Action

I spent 120 hours with Cowork.

The first twenty I wasted using it like a regular chat. Typing questions. Waiting for answers. Wondering why it felt no different from what I already had.

The next hundred I spent figuring out how it actually works.

What follows is the distillate of everything I learned, no dead ends, no mistakes that cost me time, no learning curve that almost made me quit.

Four blocks. Each one builds on the last. By the end, you’ll know how to use Cowork better than 95% of the people who already have it installed.

Subscribe now

#1 First Things First: What Cowork Is and How It Works

Let’s start from the beginning.

Cowork is not a new Claude feature. It’s not a plugin. It’s not a chat upgrade.

It’s an entirely different environment. And to use it, you need exactly one thing before anything else:

The Claude desktop app.

Not the browser. Not claude.ai open in Chrome. The app. The one you download and install on your computer.

Go to claude.com/download, download it, install it. Three minutes, tops.

Once it’s open, you’ll see three tabs at the top:

Chat, Cowork, Code

Click Cowork. You’re in.

How It Works in Practice

When you open a task in Cowork, the first thing it asks is simple:

“Which folder do you want me to work in?”

You select a folder on your computer. From that moment, Cowork can read, create, and edit the files inside it.

Only there. Only what you show it.

It doesn’t see the rest of your computer. It doesn’t access your documents without permission. It doesn’t do anything you haven’t told it to do.

Every significant action, moving a file, editing a document, deleting something, it shows you first before executing. You approve or block. Always.

Three Practical Things to Know Right Now

First: Cowork only exists in the desktop app. It doesn’t work in the browser. It doesn’t work from your phone in standard mode. You need the app open on your computer.

Second: the app needs to stay open while it works. If you close your laptop or quit the app, Cowork stops. Not a problem, it picks back up when you return. But it’s something to know before handing it a long task and going to sleep.

Third: everything it runs operates in an isolated environment. The code it uses doesn’t touch your operating system directly. It’s a technical safeguard. In practice it means: you can trust it, but start with non-critical folders until you understand how it moves.

The Difference That Changes Everything

In regular chat, Claude responds to you. It tells you what to do, writes you a draft, suggests an approach. Then you execute.

In Cowork, Claude executes.

You describe the objective, not the steps, the objective, and it builds a plan, follows it, produces real output. Actual files. Organized folders. Finished reports. Not text to copy and paste somewhere. Finished work, on your desktop, ready to use.

That’s the difference. Keep it in mind as you go.

Subscribe now

#2 The Setup 90% of People Skip (and Then Complain About)

Fresh out of the box, Cowork is like hiring a brilliant new team member on their first day.

They can do everything. But they know nothing about you.

They don’t know your name, what you do, how you want things done, what your priorities are. Every time you open a task, it starts from zero. It produces generic output. You get frustrated. You think it’s overhyped.

It’s not overhyped. It’s just that nobody introduced you properly.

It takes thirty minutes, done once. After that, every session already knows everything that matters.

Here’s how.

Global Instructions: The First Place to Put Your Hands

Open the desktop app. Go to Settings → Cowork → Global Instructions.

This is where you tell it who you are and how you want it to work.

No need for a novel. Five, six dense, specific lines is enough. Who you are, what you do, what kind of output you want, how you prefer it to communicate.

A concrete example:

“I’m a B2B marketing consultant. I work with small and mid-size businesses. I want direct output, no fluff. I prefer numbered lists for operational tasks and short paragraphs for reports. Always save files in the /Output folder with the date in the filename.”

From this point on, every task you open inherits this context. Claude no longer asks who you are. No more generic answers. It already knows how you want things.

Folders: What to Give It and What Not To

Cowork sees nothing of yours unless you show it.

Every time you open a task, you select the folder you want it to work in. Just that one. Just the files inside.

Practical advice: don’t give it access to your entire desktop on day one.

Build a simple structure. One folder per work area:

• /Clients

• /Marketing

• /Research

• /Output

Start with one. Understand how it moves. Then expand.

For folders you use often, you can also add Folder Instructions, context-specific instructions for that space. In the Clients folder you might write: “All reports in PDF. No sensitive data in filenames.” In Marketing: “Output always in English. Direct tone.”

Cowork reads them automatically every time it works in that folder.

The Trick Nobody Tells You: Import Your Memory from ChatGPT

If you’ve been using ChatGPT for months, that platform already knows a lot about you. Your industry. Your style. Your projects. Your preferences.

You don’t have to start from scratch with Claude.

Here’s what to do:

Open ChatGPT. Paste this exact prompt:

“List everything you know about me: my work, preferences, projects, communication style, goals.”

Copy the output it gives you.

Go back to Claude → Settings → Memory → Add to Memory. Paste everything.

From that moment, Claude has the context you built somewhere else. No more repeated explanations. No more cold starts.

Skills: A Note Worth Its Weight in Gold

If you read the issue on skills, you already know what I’m talking about.

In Cowork, skills aren’t called with slash commands like in Code. You use natural language: describe the objective in the task, and Cowork figures out which skill to activate.

That means every workflow you’ve already built becomes automatically reusable here too.

It’s a massive lever. In the Vault we go deep on how to structure them specifically for Cowork. For now, just know they work, and that it’s worth having them.

Subscribe now

#3 The First 4 Tasks to Do Real Things Right Now

Theory’s over.

Here are four concrete tasks you can launch today. Each one is designed to show you something different about how Cowork thinks and works.

Start with the first one. Actually do it. One completed task is worth more than ten newsletters read.

Task 1: Organize Your Files

This is the classic starting point. Simple, visible, immediate.

Take your Downloads folder. Select it as the working folder in Cowork. Write this:

“Go into this folder. Read the contents of every file. Organize everything into logical subfolders based on content type, not file extension, actual content. Rename files with descriptive names. Tell me what you did and why before moving anything.”

Notice that last part: “tell me what you did before moving anything.”

It matters. Cowork shows you the plan. You approve. Then it executes.

The thing that surprises almost everyone the first time: Cowork doesn’t just look at the file extension. It reads inside. It understands that IMG_4521.png is actually a Stripe invoice. It puts it in the right folder.

That’s the leap beyond any automatic organization system you’ve used before.

Task 2: Build a Report from Scattered Notes

You’ve got notes everywhere. Scraps, drafts, text files, half-finished Word docs. Cowork reads all of them and builds something usable.

Create a folder. Drop everything you have on a specific project inside it, raw material included, incomplete stuff included.

Then write:

“Read all the files in this folder. Build a single, structured, coherent document. Identify the main themes, eliminate redundancies, flag the parts where information is missing. Save it as report_[date].md in the /Output folder.”

What comes back isn’t a summary. It’s a worked document, with structure, ready to refine. The dirty work, reading everything, connecting the pieces, cutting the noise, it already did that.

Task 3: Competitive Research with the Browser

This task shows something regular chat simply can’t do: Cowork opens the browser and navigates for you.

Write this:

“Go to [competitor name]’s website. Analyze: main headline, call to action, value proposition, clear strengths, visible weaknesses. Cross-reference with what you know about my business from the Global Instructions. Give me a list of three things they’re doing better than me and three opportunities they’re not using. Save the report in /Output.”

Cowork opens Chrome, loads the site, reads, analyzes, writes the report.

You do something else in the meantime.

When you come back, you have a competitive analysis ready. The kind that used to cost you an hour of manual work.

Task 4: Daily Briefing from Email and Calendar

This is the task you won’t want to turn off once it’s running.

First, connect Gmail and Google Calendar from settings: Settings → Connectors. Two clicks, standard OAuth authentication, thirty seconds.

Then write:

“Every morning I want a daily briefing. Read emails from the last 12 hours and flag only the ones that require a response or action from me. Read today’s and tomorrow’s calendar. Tell me what I have, who it’s with, and if there’s anything I need to prepare. Write everything in a file briefing_[date].md in the /Output folder. Direct tone, no fluff.”

For now, run it manually. In the next block I’ll show you how to make it run on its own every morning at 8, without you opening anything.

Subscribe now

#4 Scheduled Tasks: The First Time It Works Without You Asking

Up until now you’ve seen Cowork respond to your commands.

You write. It executes.

Scheduled tasks flip that dynamic.

You write once. It executes forever, every day, every week, every month, without you touching a thing.

This is the moment Cowork stops being a tool and starts being a collaborator.

How to Set It Up

Open a new task in Cowork. Write the task normally, exactly like you did in the four examples in the previous block.

Then, instead of hitting enter right away, type /schedule at the end of the message.

Cowork asks:

• How often? (every day, every week, the first of the month)

• What time?

Answer. Confirm.

Done.

From that point, the task runs on its own schedule. You stop thinking about it.

Three Tasks to Schedule Right Now

The morning briefing. You already wrote the prompt in Block 3. Schedule it for 8:00 AM every weekday. Every morning you open your laptop and everything you need to know is already there: important emails, calendar, priorities. No scrolling, no searching. One file, ready.

The weekly file cleanup. Every Friday at 6:00 PM: “Go to the /Downloads folder. Organize everything that came in this week into the correct folders. Delete duplicates. Give me a file listing what you moved and what you deleted.” The weekend starts with a clean desktop.

The monthly report. On the first of every month: “Go to the /Clients folder. Read all files updated in the previous month. Build a summary report: what was done, what’s still open, what needs attention. Save it as report_[month]_[year].pdf in /Output.” You get a full picture of the month without spending a single minute on it.

One Important Thing to Know

Scheduled tasks only work if the desktop app is open.

If your laptop is closed, Cowork can’t work.

In practice this means: keep it open overnight if you have tasks running at night, or schedule them for times when you know the computer is on.

If you open your laptop after the scheduled time, Cowork doesn’t skip the task. It runs it as soon as it can.

That’s a real limitation. Worth knowing before you hand it something critical.

Subscribe now

📌 Espresso Prompt: Build Your First Cowork Setup

Here’s what I’d do: describe your work to Claude, and let it tell you exactly how to configure Cowork for your specific situation.

You are an expert in Claude Cowork configuration.
My goal is to configure Cowork in the most effective way possible for my specific work.
I'll describe what I do. You analyze it and give me a complete, personalized setup plan.
Here's my work:

Role/profession: [e.g. marketing consultant, founder, freelance designer]
Tasks I do every day: [e.g. reply to client emails, create reports, manage social media]
Tasks I do every week: [e.g. meetings, invoices, client updates]
Tools I use regularly: [e.g. Notion, Gmail, Google Drive, Slack]
The thing that wastes most of my time: [be specific]
The thing I hate doing but have to do anyway: [be honest]

Based on what I've told you, give me:
1. MY PERSONALIZED GLOBAL INSTRUCTIONS
Write the exact text to paste into Settings → Cowork → Global Instructions. It should reflect my role, my work style, and my specific goals.
2. THE FOLDERS TO CREATE
Tell me exactly which folders to create on my desktop and how to name them. For each one, write the Folder Instructions to use.
3. THE FIRST 3 TASKS TO SCHEDULE RIGHT NOW
For each one: task name, exact prompt to use, recommended frequency, recommended time.
4. THE CONNECTORS TO ACTIVATE FIRST
Based on the tools I use, tell me which connectors to activate immediately in Settings → Connectors and why.
Be specific. Give me stuff to copy and use, not generic advice.

☢️ Radioactive Humor

From the bunker log, Tuesday morning.

The Barista delegated his morning briefing to Cowork three weeks ago.

Week one: it works. He’s surprised. Week two: still works. He starts to trust it. Week three: he opens the laptop, reads the briefing, closes the laptop, makes himself another coffee.

The problem is the briefing was so good he had nothing left to do before 9 AM.

So he asked Cowork to find him more things to automate.

Cowork found eleven.

Now the Barista shows up at the counter at 10, handles the two things the machine still can’t do, and spends the rest of the time wondering what it means to work.

☢️ Anthropic assumes no responsibility for existential crises induced by excessive efficiency.

Subscribe now

You made it this far.

You now know what Cowork actually is. You’ve configured the setup 90% of people skip. You’ve seen four concrete tasks to launch today. You’ve set up your first recurring automations.

You’re already ahead of most people who’ve had Cowork installed for weeks.

But there’s a massive gap between using Cowork and making it work at full capacity.

In the Vault edition, we go to the next level.

Chunked skills: why one giant skill doesn’t work and how to structure them so they run in Cowork without errors.

Plugins: what they are, how to build them, how to share them with a team.

Context rot: the silent error that degrades your output quality without you noticing, and how to avoid it.

Computer Use and Dispatch: Cowork opening apps, browsing sites, and taking instructions from your phone while you’re out.

And what I learned in 120 hours that you won’t find written anywhere else: the mistakes, the shortcuts, the patterns that actually work.

If you’re not in the Vault yet, you know where to find us.

☕ Matt

Subscribe now

P.S. If you know someone still using Claude like a regular chat, send them this issue. It’ll change their week.

Share

Today, the counter looks different.

Not the usual one, with concrete walls and the smell of burnt coffee and overheated circuits. This time the counter is polished. The ceiling is high. The cups are white porcelain, and someone refills them before you finish drinking.

I was in Monte Carlo.

Not at the casino, to be precise. Though at certain levels, the difference is thinner than it looks.

I was at an international investment forum. A room packed with family offices, private equity, investment banks. The kind of people who don’t debate whether to allocate. They debate where, how much, and on what logic.

Billions on the table. Everyone wearing the expression of someone who’s heard it all before.

And they talked about AI.

Not the way you talk about it on LinkedIn. Not the way podcasts talk about it, with the electronic jingle and the guest selling a course. The way people talk about it when they have to put a number on something and answer for that number in front of someone.

That way is different. It’s colder. It’s more useful.

I brought a vision into that room. And what I heard back, what was moving between the tables, inside the questions, in the side conversations, concerns you too.

Even if you’ll never be in that room.

Maybe especially because you won’t.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

In that room, nobody was debating whether AI mattered.

They were debating how to price the companies that haven’t figured it out yet.

The conversation you see out there is at least one floor behind.

🧪 AI in Action

#1 The Metric You Don’t Know Yet

At some point in that Monte Carlo room, someone asked a question that changed the pace of the conversation.

“How many people do they have?”

It wasn’t a question about the size of the company. It was a question about the model. Because the people in that room had already stopped reading headcount as a signal of capacity. They read it as a signal of efficiency. And the difference between those two readings is worth billions.

Let me explain.

For decades, the logic was linear. More people, more output. More output, more value. It was true. It built every valuation multiple still in use today. It was the physics of cognitive work, and nobody questioned it because it worked.

Then a metric came along and broke that linearity.

It’s called Revenue per Employee. Revenue divided by headcount. Simple. Brutal. And increasingly central to every serious conversation about a company’s value.

Here are the numbers.

The top 10 AI-native startups in the world average $3.48 million per employee. Average team size: 24 people. These aren’t outliers. They’re the new baseline for anyone building with AI at the core from day one.

Traditional SaaS, the model everyone considered the most efficient on the market until recently, tops out at $610,000 per employee.

The gap is 5.7x.

Subscribe now

A concrete example: CHAI. Twelve engineers. Thirty million dollars in revenue. Two and a half million per person. And 74% of these companies are already profitable. They’re not burning cash to grow. They’re making money with teams that once wouldn’t have been enough to handle the bookkeeping.

This is not a Silicon Valley phenomenon to watch from a distance.

It’s the ruler someone is already using to measure your company. Your firm. Your team. Your structure.

Think about how the reasoning works from the other side of the table.

If you’re a manager, your cost is no longer evaluated just on your salary. It’s evaluated on how much output your team produces divided by how many people it takes to produce it. If that ratio doesn’t move, you’re a cost center that isn’t getting more efficient.

If you’re a founder, the number of people you hire no longer signals ambition. It signals how dependent you still are on the old model.

If you’re a freelancer or a consultant, the question your client will ask more and more often isn’t “how good are you?” It’s “how much output does an hour of your time produce compared to what I could get otherwise?”

The metric has changed.

The problem is almost nobody told you.

Subscribe now

#2 They’re Not Buying Companies. They’re Buying People.

There’s something else that was moving through that room in Monte Carlo.

A pattern that anyone tracking capital flows has started to see clearly over the past year and a half. And once you see it, you can’t stop noticing it everywhere.

Big tech isn’t buying AI companies.

They’re buying people.

Let the numbers speak.

Google + Windsurf. $2.4 billion. What did they buy? CEO Varun Mohan, co-founder Douglas Chen, and the core R&D team. They’re now inside Google DeepMind, Gemini division. The product? Sold separately to Cognition AI. That wasn’t the point. OpenAI had a $3 billion offer on the table. Google won by structuring the deal as a hire, not an acquisition. They bypassed regulatory scrutiny. They moved faster.

Google + Character.AI. $2.7 billion. What did they buy? Noam Shazeer and Daniel De Freitas, ex-Google Brain, co-inventors of the transformer architecture that underlies every modern AI system. Character.AI still exists as an operational shell with a license. The product wasn’t the point. The people were.

Meta + Scale AI. $14.3 billion. What did they buy? Alexandr Wang, 28 years old. Founder and CEO, joining Meta as Chief AI Officer to lead Superintelligence Labs. Meta didn’t buy his company. They took a 49% non-voting stake. They bought his time, his mind, and his team. Fourteen billion was the price of one person.

Stop on that number for a second.

Fourteen billion for one person.

Subscribe now

Not for a patent. Not for a user base. Not for a product with millions of downloads. For someone who knows how to do one specific thing inside real organizations.

That’s the clearest signal the market is sending right now about what actually has value.

Not the technology. Technology is accessible to anyone. Models are a commodity. Tools cost a monthly subscription.

The rare thing, the thing no subscription can replicate, is the ability to take that technology and make it work inside a real system, with real processes, real people, real resistance.

Whoever can do that is worth more than the product they built.

And that changes everything about how you should think about your own value in the market.

You are not your resume. You are not the list of tools you know how to use. You’re not even the industry you work in.

You are the unique combination of context, operational judgment, and integration ability you’ve built over time. That combination, in today’s market, is the most undervalued and most sought-after asset at the same time.

The problem is most people don’t know how to name it. And what you can’t name, you can’t sell.

Subscribe now

#3 The 5% That’s Actually Making Money

At one point during the panel, I asked the room a question.

It wasn’t rhetorical. It was a blade.

“Show me where AI has moved a number in your P&L. Not a pilot. Not a proof of concept. Not a productivity survey. A number. Revenue up, cost down, margin protected. Something you can measure.”

The room responded the way rooms respond when a question lands on something uncomfortable.

First, silence. Then a few vague answers. Then a real conversation.

Because here’s the truth.

McKinsey analyzed 1,993 companies across 105 countries. Only 6% qualify as AI high performers, defined as an EBIT impact above 5%.

BCG looked at 1,250 companies. Only 5% are extracting value from AI at real scale.

PwC: 12% achieved both cost reduction and revenue growth from AI.

Accenture, across 2,000 enterprise projects: 13% report significant value from generative AI.

Between 5% and 13% of companies are actually making money with AI.

The rest are spending.

And the spending is enormous. Corporate AI budgets have exploded. Press releases talk about digital transformation, AI-first, the future. Boardroom slides are packed with roadmaps and use cases.

But the numbers in the P&L don’t move.

Why?

Forrester’s research across 500 enterprise data leaders gives an answer you don’t expect. The main problem isn’t model quality. It’s not compute costs. It’s not a shortage of technical talent.

It’s data quality.

Subscribe now

The same client shows up as “Acme Corp” in the CRM, “Acme Corporation” in emails, “ACME Inc.” in contracts. AI can’t reconcile what humans have fragmented. Garbage in, garbage out, but this time at industrial scale.

And then there’s the problem BCG called the 10-20-70 rule.

10% of AI’s value comes from algorithms. 20% from tech infrastructure. 70% comes from people, processes, change management, and the ability to redesign how work actually gets done.

More than two-thirds of AI transformations fail because of problems in the 70%. Not in the technology.

The tool is not the problem.

It never was.

That’s the line that separates the 5% from the 95%. The 5% starts with the right question: where are we bleeding margin, where are we making decisions on stale data, where does manual work produce invisible errors? Then they use AI to close those specific gaps.

The 95% does the opposite. They buy the tool. Then they ask the team what to do with it.

The question I brought into that Monte Carlo room is the same one you should bring to your meeting tomorrow.

Show me where AI has moved a number. Not a pilot. A number.

If the answer is a slide deck, the answer is no.

Subscribe now

📌 Espresso Prompt: Find Your 70%

BCG says 70% of AI’s value is won or lost on people and processes. Not on technology.

This prompt doesn’t ask you to buy anything new. It asks you to look at what you already have and figure out where the value is getting stuck.

Use it on Claude or ChatGPT (GPT-4o or higher). Be specific in your description: the more context you give, the more useful the diagnosis.

Act as a senior operational consultant specializing in AI integration inside organizations.

Here’s my situation:

Industry and type of company/team:

[e.g. 8-person marketing agency, law firm, operations team at a mid-size manufacturer]

AI tools I use or have already adopted:

[e.g. ChatGPT, Copilot, Zapier automations, etc.]

The 3 main processes we work on every week:

[briefly describe what the team does, not how it does it]

Where I feel AI isn’t producing concrete results:

[be honest — “I don’t know” is useful information too]

PHASE 1 — Map my 70%

Analyze what I’ve described and identify:

— The 2 or 3 human or process bottlenecks blocking the value of the AI tools already in place

— Where people are likely resisting, working around, or using AI superficially

— Which process, if redesigned, would have the most immediate impact on results

Don’t be generic. Treat me like a paying client with a real problem.

PHASE 2 — The gap between what I have and what it produces

Tell me:

— What I’m misusing from what I already have

— What I could cut before buying anything new

— What’s the first concrete action I can take this week with no additional budget

PHASE 3 — The number

Help me identify one measurable indicator, even a simple one, that lets me see whether something has changed 30 days from today.

I don’t want a slide deck. I want a number.

⚠️ Counter warning: if you can’t find a single measurable number in Phase 3, the problem isn’t AI. It’s that you don’t yet know what you’re trying to improve. That’s the most useful diagnosis you could have made.

☢️ Radioactive Humor

Someone at the forum asked if I’d found time to stop by the casino.

I told them I’d spent the whole week in a room full of people betting billions on companies that don’t exist yet.

Same principle. Better cocktails.

The only real difference between a poker table in Monte Carlo and an investor boardroom is that at the poker table you at least know when you’ve lost.

The way companies are built is changing. The way work is valued is changing. The way a person, a team, an entire structure gets priced is changing.

Not in five years. Now.

And the point isn’t that you need to become a 24-person AI-native startup doing $3 million in revenue per employee. The point is that the people who decide where capital goes, who gets hired, who gets paid and how much, are already using new metrics to make those calls.

You may not know it. But they do.

The Monte Carlo room wasn’t somewhere far from your daily life. It was the place where someone was already deciding the rules of the game you’re playing.

Three things I’m bringing home, and leaving on the counter for you:

Revenue per Employee is the metric nobody explained to you but someone is already using to evaluate you. Start understanding where you stand on that scale.

Technology is not the problem and not the solution. 70% is won or lost on processes and people. Start there.

The most useful question you can ask — yourself, your team, your company — is just one: show me where AI has moved a number. Not a pilot. A number.

If you don’t have an answer yet, you just found the most important work waiting for you.

The bunker is back to its usual self.
Concrete walls. Coffee on the counter. Smell of circuits.

But something has changed.

You know one more thing about how the room works where they decide your value.

Use it.

☕ Matteo

Subscribe now

P.S. If this issue made you raise an eyebrow in the good way, send it to someone who works, builds, or sells something. You have no idea how much it helps. And yes, this time too it’s the barista saying it. Not the algorithm. ☢️

Share AI Espresso | your tech brew

This week I’m not talking about a tool.

Not a new model, not an update, not a feature you didn’t know existed.

I’m talking about something I’m watching happen right now, every day, inside the companies and careers of people I know.

Something almost nobody is naming correctly.

There’s a very widespread illusion going around.

It sounds like this: “We adopted AI. We work the same as before, just faster.”

That’s the most dangerous sentence you can say right now.

Because it’s not true.

And the consequences of believing it are already showing up.

But there’s another side to this story.

And that side is more interesting.

There’s a type of professional out there who’s moving at a speed their colleagues simply can’t explain. They don’t have a bigger team. They don’t have a different budget. They don’t have access to secret models.

They figured out one thing. One thing that changes how cognitive work actually functions.

Today I’m going to show you what that is.

Welcome back to the bunker.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

Adding AI to a team that already works badly is like strapping a rocket engine onto a car with square wheels.

You go faster. You bounce just as hard.

🧪 AI in Action: Marco and Giulio

Marco runs a team of five.

Copywriter, designer, media buyer, account manager, project manager.
Each one sharp in their lane. Monday meetings, Wednesday alignment calls, Friday reviews.

Three months ago, they decided to adopt AI.

Today it takes them 30% longer to close a campaign than it did before.

Giulio works alone.

Same type of projects. Same market. Similar budget.

In a single afternoon he produces three complete campaign concepts, tests them, picks one, and ships.

No calls. No alignment meetings. No handoffs.

Both of them use AI every day.

The results are opposite.

And the reason has nothing to do with the tools they use. Not the model they picked. Not the quality of their prompts.

It has everything to do with something Marco’s team has always had, and that Giulio eliminated without even giving it a name.

Subscribe now

The Cost Nobody Ever Put in the Budget

Think about how a project works in a classic team.

You have an idea. You explain it to the copywriter. The copywriter interprets it, produces something, sends it over. You read it, correct it, re-explain. They revise. Meanwhile the designer was waiting on the copywriter’s brief to get started. Now the brief has changed. The designer starts over. The media buyer is waiting on the designer’s assets to set up the campaign. And so on.

Every handoff is a translation.

Every translation loses something.

Call it what it is: Coordination Tax. The hidden cost of moving context from one brain to another. The real price of getting different people to work together, each with their own interpretation, their own rhythm, their own backlog.

This isn’t a people problem. It’s physics.

When information moves from one brain to another, it mutates. When it sits in someone’s queue, it ages. When it comes back after a review, it carries a different version of the original problem.

And this cost, meetings, alignment calls, briefings, re-briefings, waiting, misunderstandings, never appeared on any budget. It was invisible because it looked like the normal price of working together.

It wasn’t. It was a structural loss. Silent. Constant.

Subscribe now

From the Counter: I Found This Out on Myself

This didn’t come from a paper I read.

It came from something I lived directly, and stopped being able to ignore once the pattern repeated enough times to rule out coincidence.

When I need to build a funnel from scratch for a client, the knowledge base, the copy, the page structure, the emails, everything, I do it today in a fraction of the time it used to take.

Not because I work more hours. Not because I found shortcuts.

But because the project’s context lives in a single brain: mine. The AI absorbs it in real time, without me having to re-explain it, without anyone interpreting it in the middle, without waiting for someone else to execute it and send it back.

The time I used to spend waiting, for people to understand, to produce, to come back with the work done, was time in which I would have already closed the project four times over.

Four times. That’s not hyperbole. That’s what I measured.

There’s another side to this story, and it’s worth saying out loud.

When you work with people who use AI the right way, handoffs get dramatically faster. You pass the context, they feed it to their models, and within minutes they’re aligned on exactly what you had in mind. The friction doesn’t disappear, but it drops sharply.

The problem isn’t working with other people. The problem is working inside team structures designed for a world where transferring context was slow by definition.

Everything I’m describing in this episode, I lived before I read it. The research helped me understand why it worked. But the proof was already on the counter.

Subscribe now

Why AI Doesn’t Solve the Problem. It Amplifies It.

This is the paradox that’s quietly wrecking Marco’s team in companies everywhere.

The logic seems sound: take a team that already works, add AI, get a team that works faster.

That’s not how it works.

Asana studied over 9,000 knowledge workers in 2025. The result is counterintuitive and brutal: the most AI-productive workers, the ones producing more, faster, create more friction in the system around them, not less. 90% of them say AI increased their coordination workload with colleagues. Not decreased it.

Why?

Because the structure didn’t change. The approval chains are the same. The review cycles are the same. The alignment meetings are the same. Except now one person produces in an hour what used to take a week, and the system downstream can’t absorb it.

It’s like putting a Formula 1 driver on a road built for horse-drawn carriages. You go faster. You don’t arrive sooner. And you risk destroying yourself in the process.

The problem isn’t production speed.

It’s that organizations keep running coordination structures built for a slow world, on top of people who now work ten times faster. The bottleneck isn’t execution. It’s absorption.

And now you understand what’s happening to Marco’s team.

The AI isn’t broken. The team added acceleration to a system that was never designed to absorb it. The result is more friction, not less. More meetings to manage the increased output. More reviews. More alignment. More time lost.

Subscribe now

Zero Handoff Loss: The Real Advantage of the Solo Operator

Giulio isn’t better than Marco.

He doesn’t have better tools. He doesn’t work more hours. He didn’t find some magic prompt Marco’s team doesn’t know about.

He removed the problem structurally.

When Giulio works, the project context lives in a single brain. His. The AI absorbs it in real time — documents, briefs, objectives, constraints, tone — with no translation, no interpretation, no latency.

No brief to write for the designer.

No alignment call with the copywriter.

No waiting on the next handoff.

The context doesn’t travel. It stays put. And the AI uses it every time with the same precision it had the first time it received it.

That’s the real advantage. Not execution speed. Execution speed is a byproduct.

The real advantage is zero handoff loss: no context lost in transfer, no translation cost, no latency between one step and the next.

Sam Altman has said publicly that he’s betting on when the first billion-dollar company built by a single person will arrive. He’s not saying it because that person will be extraordinary. He’s saying it because the structures that once made a team necessary to reach that scale are becoming irrelevant, one by one.

This isn’t science fiction. It’s what Giulio does every afternoon, at a smaller scale.

Subscribe now

The New Profile: The Architect

Hold on. Because this is where the most common mistake hides.

Reading this far and concluding that the future belongs to generalists is wrong.

The classic generalist, the one who knows a little about everything and a lot about nothing, is the highest-risk profile out there. AI already does that better, faster, and without coffee breaks.

And the pure vertical specialist, the one who does one thing better than anyone, is losing the competitive edge that specialization used to guarantee. Because the economic barrier that protected that skill has dropped. We covered that last episode.

The profile that’s emerging is something different from both.

Call it the Architect.

Not someone who can do everything. Someone who understands enough about everything to give precise instructions and recognize when the output is good or wrong.

The distinction sounds subtle. It isn’t.

A researcher from Stanford and Harvard Business School studied what happens when you give AI to people with different skill levels. The result is clear: with AI you can execute tasks adjacent to your area of expertise without retraining. A data scientist can run marketing analysis. A copywriter can structure a media strategy. But there’s a wall.

Call it the genAI wall.

Beyond that wall, when the task is too far from what you actually know, AI isn’t enough. Not because the model is weak. But because without domain understanding, you don’t have the judgment to evaluate the output. And an output you can’t evaluate is an output you can’t use.

The Architect is someone with enough depth in a domain to stand on the right side of that wall. And enough breadth to cover adjacent tasks with AI without losing quality.

Giulio isn’t a jack-of-all-trades. He’s a marketing expert who uses AI to perform at a sufficient level as a designer, copywriter, and media buyer. He knows when the copy is off. He knows when the visuals aren’t working. He knows when the targeting is wrong.

Without that domain foundation, AI produces fast mediocrity.

With it, AI produces fast excellence.

The speed is the same. The result isn’t.

Subscribe now

Where the Team Still Wins

This episode isn’t a slogan.

So let’s tell the whole story.

There are three contexts where a structured team still beats the solo operator with AI. Ignore them and you’re optimizing the wrong strategy.

The first is extreme technical depth.

There are high-stakes decisions in highly specialized domains, complex system architectures, critical legal operations, medical diagnoses, where the judgment of multiple real experts produces results no single Architect can replicate. The genAI wall is real. And in certain domains, the cost of getting it wrong is too high to ignore.

The second is relational scale.

At a certain point of growth, cognitive work gives way to political work. Negotiations. Alliances. Institutional trust. Decisions made in rooms, not in documents. AI doesn’t shake hands. It can’t read the room. It doesn’t build trust over time. At that scale, you need real people working together.

The third is creative tension between different minds.

A Procter & Gamble study of 776 employees found that in certain innovation contexts, two human experts who contradict each other produce ideas that neither the solo operator nor an AI team can generate. Cognitive friction between two genuinely different points of view, when managed well, is still worth something.

These three contexts don’t cancel anything you’ve read so far.

They narrow the map.

The solo operator with AI is superior across all standard cognitive work, execution, production, iteration, testing. But there are thresholds of complexity, scale, and innovation where collective structure wins back the advantage.

Knowing where you stand is the difference between using the right strategy and applying a framework to a context it can’t support.

Subscribe now

The Operational Question

Let’s come back to you.

Not Marco. Not Giulio. You, reading this from your seat in the bunker.

The wrong question to ask yourself right now is: “Team or solo?”

It’s the wrong question because it assumes the choice is structural. It isn’t. It’s personal.

There’s only one right question:

Are you deep enough in your domain to play the Architect?

Because if the answer is yes, the solo operator with AI isn’t just an option. It’s the highest-performing setup available today for standard cognitive work.

If the answer is no, if your domain expertise is still shallow, then adding AI doesn’t turn you into Giulio. It turns you into the worst member of Marco’s team: the one who produces a lot, fast, and wrong.

The variable isn’t the tool. It’s the judgment you bring to the tool.

And judgment is built exactly one way: through years of real work, real mistakes, real feedback. You can’t buy it with a subscription. You can’t install it like a Skill.

Giulio’s bunker isn’t built from tools.

It’s built from what Giulio already knew before AI existed, and what now lets him drive the machine instead of chasing it.

Subscribe now

📌 Espresso Prompt: Find Your Coordination Tax

This prompt does one thing.

It shows you exactly how much Coordination Tax you’re paying every week, in real hours, not vague gut feelings. And it tells you where to cut first.

Use it in Claude, inside a Project where you’ve already loaded your work context. If you haven’t, describe it yourself at the start.

You are a productivity consultant with expertise in cognitive work teams and AI integration.

I want to understand how much Coordination Tax I’m paying in my current work.

PHASE 1 — MAPPING

Ask me these questions one at a time. Wait for my answer before moving to the next.


Describe your role and how many people you regularly interact with to complete a typical project.

What is the project or workflow you repeat most often? Describe it in its main steps.

In that workflow, where do you wait for someone else to produce something before you can move forward?

How many times on average does a brief, a document, or an instruction get reinterpreted or require clarification before reaching the correct output?

PHASE 2 — DIAGNOSIS

Based on my answers, tell me:

Where my main Coordination Tax is concentrated

How many hours I estimate losing per week to waiting, alignment, and re-briefing

Which of these costs I could eliminate by working differently with AI

PHASE 3 — PLAN

Identify the 3 highest-impact interventions I could make right now to reduce Coordination Tax in my workflow.

For each one tell me:

What exactly changes

How much time I estimate recovering

Whether it requires working alone, with fewer people, or simply differently with the same people

Be direct. Numbers where possible. No generic advice.

⚠️ Warning from the counter: if Phase 3 comes back with “change everything,” start with one workflow only. The bunker gets built one wall at a time.

Subscribe now

☢️ Radioactive Humor

I have a friend, won’t name him, he knows who he is, who read something like this episode and decided to fire his entire team and do everything solo with AI.

In 48 hours. (Big mistake.)

Having never used AI for anything more complex than an email.

Three weeks later he called me.

He told me he finally understood what Coordination Tax was.

Because he was paying all of it himself. Alone. With himself. In real time.

The problem wasn’t the team. It was that he didn’t know how to play the Architect.

Moral of the story: before you get rid of the square wheels, make sure you know how to fly the rocket.

☢️ Until next time,

Matt at the counter.

Subscribe now

P.S. If you know someone who’s still adding AI to their team without changing a single thing about how they work, and then complaining it doesn’t work, send them this episode. First cup’s on us.

☕ The counter’s closed, see you in the next drop.

Share

Not because something blew up, but because something far quieter collapsed.

They brought down a $70 million datacenter while it was still being built.

Not a structural failure. Not a mistake.

They simply decided that steel giant was no longer worthy of hosting the AI of the future.

And in that moment, it hit me:

when we talk about AI, we’re all trapped in the same collective illusion.

We think the whole game happens on the surface, the tools, the models, the prompts.

We think AI is software… because that’s the part we get to touch.

But the real earthquake doesn’t live inside the programs.

It lives beneath them, in the power crammed into a single rack, in the heat we have to tame, in the energy we burn just to make a machine think a little faster than yesterday.

That’s the coffee I’m serving you today.

Strong. Dark. No sugar.

Because the future of AI isn’t in the tool you click, it’s in the infrastructure you never see.

Welcome to the counter.

Subscribe now

💡 Thought of the Day

There’s a paradox that makes me smirk here in the bunker:

the more immaterial and radiant AI looks on our screens,

the heavier it becomes in the real world.

Every time you generate a text, somewhere a transformer is sweating.

Maybe the real question isn’t “what will AI do to our jobs”- but “what will the world have to do to feed AI?”

🧪 AI in Action

What you’re about to read isn’t a technical article.

It’s the map of a transformation almost no one sees coming.

We start with a $70 million datacenter demolished before it was finished.

Then we go underground, into compute density, energy demands, and cooling systems that have run out of room for air.

We’ll see why Goldman Sachs isn’t watching tools, it’s watching electrical grids under pressure.

And why big tech is quietly morphing into energy utilities.

No hype. No futurist fever dreams.

Just the naked truth of the world we’re building.

Take a sip.
The journey starts now.

Subscribe now

1. The Question That Flips Our Perception

Everyone keeps staring at screens, tools, shiny interfaces.

But while we’re playing with prompts, someone out there is building entire power plants just to keep what’s in our hands running.

So here at the counter, the inevitable question hits:

what if the fate of AI isn’t decided in software, but in power plants?

It forces us to shift our gaze from magic to mechanics, from the digital illusion to the physical world made of copper, heat, and transformers.

And that’s exactly where the fastest change you’ll ever witness is happening.

Why Meta Demolished a Half-Built $70 Million Datacenter

The answer is brutal: it was already obsolete.

Meta was building an H-type datacenter, engineered for the cloud-and-content era.

But while the bulldozers were still moving dirt, AI rewrote the rules of the game.

The energy density required by the new GPU clusters was simply too high, more heat, more consumption, more power per square meter than that facility could ever handle.

So Meta did the least intuitive, and most logical, thing:

it didn’t upgrade it. It demolished it.

And rebuilt a new architecture: denser, optimized for liquid cooling, with power levels three times higher than previous standards.

It’s the perfect emblem of this new era: AI evolves so fast that even physical infrastructure has to be reborn from scratch.

Subscribe now

2. The Great Illusion: AI Looks Like Software - but the Revolution Lives in Physics

There’s a mental trap almost everyone falls into when they think about AI.

We look at it on screens, inside tools, inside the apps we touch every day.

So our brain jumps to the obvious conclusion: “AI is software.”

Easy. Intuitive. Wrong.

Because all the AI you see, clean, elegant, immaterial, is just the foam on the cappuccino.

The interface. The theater.

The real work happens below. In physics.

While we obsess over prompts, plugins, and benchmarks,

the real AI revolution is happening in the physical world, in our ability to compress more and more intelligence into tighter and tighter spaces.

Here’s the part almost no one notices:

the evolutionary leap of AI isn’t in the tools.

It’s in the density of compute power we can squeeze into a single square meter.

When you see AI as software, the future feels light.

When you see it as a physical phenomenon, the future gets heavy.

Literal, industrial-grade heavy.

It’s not software changing the world. It’s the world being forced to change to support that software.

Subscribe now

3. Two Industrial Eras: Information vs. Physical Intelligence

For twenty years, we lived inside an unspoken certainty: digital was air.

Light. Scalable. Invisible.

One click and everything shifted.

Then the real AI arrived, and digital started behaving like matter.

The Information Era was made of servers distributing data.

The Physical Intelligence Era is made of supercomputers generating cognitive capacity.

And to do that, they consume like heavy industries.

The limit used to be software.

Now the limit is physics.

You can’t compress intelligence forever without paying the bill.

And that bill shows up as heat, density, energy demands, and supply chains creaking under the load.

The technology we used to call “virtual” is becoming the most physical industrial sector of the next decade.

4. The GPU Saga: Compute Density as a Hungry Creature

If you really want to understand why the world is changing this fast, don’t look at the models.

Look at the machines that make them exist.

Modern AI has a heart, and that heart isn’t an algorithm. It’s a GPU.

Actually, let’s be honest: it’s a hungry creature.

Just look at the escalation:

250W 400W 700W 1,000W 2,700W on a single board.

Each jump isn’t just a bigger number on a datasheet.

It’s more heat to dissipate, more copper to route, more mechanical complexity to tame.

Standard office racks at 3–7 kW? Paleolithic.

“High-performance” racks at 15–20 kW? Ancient history.

Today, a serious AI rack draws 130 kW, the electrical load of a restaurant crammed into a studio apartment.

And a GPU doesn’t ask for “just” energy.

It demands an entirely new infrastructure built to feed its hunger without collapsing.

Hundreds of copper cables, kilometers of high-capacity connections, compute floors so dense a millimeter-level mistake can burn thousands of dollars instantly.

That’s the GPU saga: a race between what we want AI to create, and what the physical world can survive to let it happen.

5. Cooling and Compute: When Physics Becomes AI’s Hard Limit

There comes a moment when physics knocks on the bunker door and doesn’t wait for permission.

For AI, that moment is heat.

For years, we relied on air.

Bigger fans, stronger airflow, hot aisles, cold aisles.

Then the GPU stopped being a component. It became a furnace.

And air died.

Simply incapable of absorbing the heat generated by accelerators that keep getting denser and hungrier.

This is why datacenters are mutating.

Liquid cooling, once the playground of overclockers, is now the standard of the new era.

Not an upgrade. A prerequisite.

Because liquid does what air never will: absorb heat 4,000 times more effectively.

Four. Thousand. Times.

That single ratio rewires everything: less room for heatsinks, more density per rack, more power in less space.

But it also means industrial piping, redundant pumps, kilometers of tubing running into server halls like arteries in a living organism.

And here’s the part most people still don’t want to hear:

even the giants who’ve built datacenters for decades are discovering their pre-AI designs can’t be upgraded.

They must be demolished and rebuilt from scratch.

The question is no longer “how powerful is the model?” The real question is: how much can physics take before it forces us to rethink everything?

Subscribe now

6. The Goldman Sachs Turning Point: The Hidden Energy Crisis Behind the AI Boom

While the world obsesses over new models, clever tools, and faster chatbots,

the analysts are staring somewhere else entirely: the power grid.

And what they see isn’t a trend. It’s a spike.

• Datacenter energy demand: +165% by 2030

• 40 to 50 gigawatts needed in the next few years, roughly the consumption of France or Germany

• Infrastructure under strain

• Permitting and grid expansion timelines that can’t keep pace with AI’s acceleration

In other words: while we’re chasing intelligence,

we forgot to check whether the world has enough electricity to feed it.

The bottleneck is no longer compute. It’s the energy you can deliver to it.

That’s why the boldest moves aren’t coming from AI labs,

but from power plants, high-voltage investments, and industrial plans from those who see the storm early.

If power is the new oxygen of AI, then whoever controls the energy controls the future.

7. Big Tech as the New Energy Companies

At some point, if you track the moves of the big tech giants, you notice something’s off.

They’re not just building datacenters anymore. They’re building energy capacity.

This isn’t branding. It isn’t diversification. It’s industrial survival.

• Microsoft: restarting a decommissioned nuclear plant to power its clusters

• Amazon AWS: buying a datacenter wired directly into a 2,500 MW power station

• Meta: designing AI campuses from 1 to 5 gigawatts, entire-city levels of consumption

• CoreWeave: acquiring former Bitcoin mines with on-site gas plants for instant power

• Google: funding advanced nuclear reactors to secure stable energy for decades

This isn’t expansion. It’s metamorphosis.

Tech companies are becoming private utilities, energy actors who compete with, collaborate with, and sometimes disrupt national infrastructures.

Compute can be bought. GPUs can be manufactured.

But without enough energy, everything stops.

And when a sector becomes dependent on a single factor,

that factor stops being a cost. It becomes a strategy.

Big tech isn’t just chasing AI. It’s rewriting the planet’s energy foundations to keep AI alive.

Subscribe now

8. But Why Should Anyone at the Counter Care?

It’s easy to look at everything we’ve covered and think these are faraway problems, the kind only giants, engineers, or governments should worry about.

But that’s not how it works.

The mainstream AI narrative is distracting you.

It talks about tools, hacks, shortcuts.

It sells you an AI made of interfaces.

The real competitive edge in the years ahead won’t come from the person who presses the button best.

It will come from the one who understands why the technology exists, what holds it up, and where it’s going.

Because if AI’s infrastructure shifts this fast, everything shifts with it:

how companies make decisions, how they invest, what skills actually matter, how much culture outweighs software.

Companies that think in terms of tools will stay trapped on the surface.

Companies that understand the infrastructure will play an entirely different game.

Knowing what’s happening inside datacenters isn’t a geek flex.

It’s a strategic edge for anyone navigating a market that changes every quarter.

Those who understand AI in depth will think farther ahead, react less, see more clearly. And today, that’s worth years of advantage.

Subscribe now

9. The Image That Sticks

For years they told us software would eat the world.

And for a while, that was true.

Then AI arrived. And quietly, it started eating the software too.

Not because it’s more elegant.

But because it requires something software never dared to ask for:

space, metal, cooling, kilometers of wiring, enough energy to power entire cities.

And suddenly we found ourselves inside a historical paradox:

the most immaterial technology we ever created has become the most physical force we’ve ever built.

It all circles back to one image:

a $70 million datacenter torn down while still under construction.

Not because it was wrong. Because it was nowhere near enough for the future we’re building.

We’re not upgrading infrastructure.

We’re changing the entire paradigm.

The world isn’t building datacenters anymore.

It’s building cities designed to let machines think.

And we, here at the counter, have a duty to stare this shift straight in the eye. Because only those who understand the underground of AI will navigate its surface with real clarity.

Subscribe now

📌 Espresso Prompt - Deep Dive

Most AI conversations stay on the surface: tools, features, trends that vanish in a month.

The real transformation is happening where almost no one is looking: energy, infrastructure, geopolitics, raw materials.

This prompt is an awareness activator. Every time you run it, it surfaces a hidden fact, an overlooked shift, an insight that forces you to look deeper.

The prompt:

Act as an AI awareness radar. Every time I activate this prompt, generate one high-impact insight from one of these domains: infrastructure and compute, supply chain and critical materials, AI geopolitics, economics and markets, systemic risks, environmental and energy impacts, long-term industrial consequences.

The insight must not be mainstream. No model updates, no feature releases, no tech gossip. It must reveal something a professional likely hasn’t considered.

Return exactly 4 sections:

1. The hidden fact (max 6 lines)

2. Why it really matters (max 6 lines)

3. The blind spot: why almost no one talks about it

4. The question I should be asking

Style: no hype, direct, senior-analyst tone. Each insight should make me say: “This changes how I see AI.”

☢️ Radioactive Humor

Every week we talk about AI as if it were a purely digital phenomenon.

And we forget that when digital grows big enough, it stops being light.

It becomes weight, consumption, heat, matter.

While we debate new models on the surface, entire infrastructures are shifting beneath us:

power plants reopening, grids expanding, capital flowing toward whoever controls power, not just software.

Understanding all this isn’t about becoming an engineer.

It’s about becoming clearer.

The real question was never “what will AI do to the world?”

It’s: what world are we building to let AI exist?

And the answer, week after week, is right here at the counter.

See you at the next cup.

Matteo ☕

Subscribe now

PPS: If this episode raised an eyebrow, in a good way, share it with someone who deserves to ask themselves the right question. You have no idea how much that helps me. And yes, once again, it’s the barista saying this, not the algorithm. ☢️

Share

Out there, right now, millions of people are asking the wrong question.

They’re asking it on podcasts, in LinkedIn threads, in meetings with animated slides. The question is always the same: “Will AI take my job?”

It’s an understandable question. It’s also, fundamentally, useless.

Because the problem isn’t whether AI will do your job. The problem is something else — deeper, more uncomfortable, and almost nobody is looking it in the face:

What happens when the economic model that justified paying you stops working?

It’s not the same thing. Trust me.

Think of a designer. Talented, up to date, uses AI every day. He became 10 times faster. Produces twice as much in half the time. And he got fired anyway.

Not because he was less skilled. Not because AI was “better.” But because the economic barrier that justified his role — the time, the capital, the coordination — had already evaporated. And with it, the reason anyone should have paid him.

This isn’t science fiction. It’s economics. And it’s already here.

Today at the counter I’m not serving you a tutorial. I’m not serving you a prompt to copy-paste. I’m serving you something heavier: a diagnosis.

The kind you need to stop asking the wrong question and start facing the real problem.

☕ Cup in hand. Let’s go.

Subscribe now

💡 Thought of the Day

For 250 years, the value of your work was protected by three walls: the time it took to learn, the money it cost to train, the effort it took to coordinate everything. AI knocked down all three walls on the same day. The problem isn't that the machine works in your place. The problem is that you were renting an apartment in a building that no longer exists.

🧪 AI in Action: The Question Nobody’s Asking

A scene that’s becoming more and more common

Marco is 34. He’s been a copywriter for ten years. Talented, fast, up to date.

Over the past year he adopted AI like a pro: structured prompts, optimized workflows, outputs revised by hand. He was producing three times as much. Billing more. He felt ahead of the curve.

Three months ago the agency he worked for announced it wouldn’t be renewing contracts with external freelancers.

Not because Marco wasn’t good enough. Not because AI wrote better than him. But because the economic model that justified paying him had vanished — and with it, the reason anyone should have opened their wallet.

This isn’t a story about AI stealing jobs. It’s a story about something deeper, that almost nobody is looking in the face.

Subscribe now

The three walls that held everything up

To understand what happened to Marco — and what’s happening in nearly every sector of cognitive work — you need to start from something nobody ever explained to us explicitly.

The economic value of your work doesn’t depend on how good you are. It depends on how scarce what you do is.

And scarcity, for 250 years, was guaranteed by three real constraints that nobody could get around.

Time. Getting good at something takes years. That path has a real time cost — non-compressible. Nobody can be a designer, a lawyer, and a financial analyst at the same time; physics won’t allow it. Those who had invested years in a specialization possessed something genuinely rare.

Capital. Training was expensive. Master’s degrees, professional tracks, years of underpaid practice. The economic barrier to entry filtered who could access the qualified job market. Fewer people in, more value for those who were.

Coordination. Nobody builds a complex product alone. You need specialists working together, and making them work together was slow, expensive, full of friction. That coordination cost justified entire roles: liaisons, account managers, project managers, analysts who translated data into decisions.

Three walls. Real. Solid. Standing for two and a half centuries.

Then AI arrived. And it knocked down all three at the same time.

Subscribe now

The simultaneous collapse

Not one at a time, like every previous revolution. All three. In sync.

Time decoupled from output. What a designer used to build in three weeks — brief, drafts, revisions, approvals — is now produced in 90 seconds. The years invested in specialization no longer guarantee an advantage on the final output.

Capital collapsed by three orders of magnitude. A professional website used to cost $15,000–25,000. Today it costs a $20-a-month subscription. The economic barrier that protected those who had invested years and money to acquire that skill is gone.

Expertise became infrastructure, not labor. You used to hire the designer because they owned the skill. Now that skill is encoded in the model. You don’t pay for it as a person — you pay for it as SaaS. Expertise stopped being a labor input and became a capital asset accessible to anyone.

The result is brutal in its simplicity:

The company still needs the website. It no longer needs the designer.

The value of the output remains intact. The economic value of human labor to produce it has gone to zero.

Today’s most popular tranquilizer

This is where the narrative you hear everywhere walks in. The one that sounds reasonable. The one that makes you feel safe.

“AI won’t take your job. But someone who uses AI will.”

It’s a nice line. It’s also, structurally, a trap.

It would work if the underlying system were still the old one: find the scarce skill, capture the economic value. But that logic only holds if somewhere in the system, scarcity still exists.

Back to Marco. After the non-renewal, he did what everyone recommends: trained on AI, refined his workflows, got even faster. He’s trying to re-enter the market as an “AI-augmented copywriter.”

The problem isn’t his productivity. The problem is that the economic barrier that justified paying an external copywriter has collapsed — regardless of how productive he is inside that role.

It’s like carrying buckets of water faster. Useful, sure. But it changes nothing if in the meantime they’re installing the pipes.

Subscribe now

The ladder with no rungs

In every previous technological revolution, there was always a next step. Weavers became loom operators. Typists became data entry clerks. Data entry became Excel analysts. Each time, scarcity shifted — it didn’t vanish. There was always a destination. The ladder lost a few rungs, but the structure held.

This time is different. Coding, design, writing, legal analysis, financial modeling — all of it became abundant simultaneously. There is no “next scarce skill” to run toward. And the advice “upskill into AI” hides a question nobody is answering:

Upskill into what, exactly? Learn prompting? In six months it’ll be automated. Build agents? In a year it’ll be a template. Code? Code already writes itself.

It’s not that the rungs are harder to reach. The ladder is gone.

The only thing that’s still scarce

If execution is abundant, what’s still rare?

One thing only: judgment.

Not the speed of tool adoption. Not AI certifications. Not the number of prompts saved in your library.

Judgment is the accumulated ability to know what should exist. Which problems actually matter. How to connect distant domains in a way that produces something nobody has seen yet.

Two people with identical access to the same AI models get radically different results. Not because one prompts better. But because one knows what to build, and the other is guessing in the dark.

This ability can’t be acquired in a crash course. It doesn’t have a certification. It’s built with years of experience, reading, mistakes, and the kind of slow, uncomfortable mental work that has no immediate output.

And here lies the cruelest paradox of this moment.

The very tools that make judgment more valuable are eroding the drive to develop it.

We google instead of thinking. We prompt instead of reasoning. We delegate the exact process that builds the thing that will become rarest.

Subscribe now

The question you need to ask yourself

The debate about AI and work is stuck between two sides that keep ignoring each other. The optimists say it will create more jobs than it destroys. The pessimists say the opposite. Both are navigating with an 1800s map in a world that no longer has that geography.

The real question isn’t “will my job survive?”

The real question is simpler. And more uncomfortable:

Are you building judgment, or are you just learning to carry buckets faster?

Because when execution stops being the bottleneck, those who haven’t answered that question aren’t behind.

They’re already out.

📌 Espresso Prompt — The Replaceability Test

Knowing that judgment is the only thing that’s still scarce is one thing. Finding out how much of it you actually have is another.

This prompt doesn’t ask you to reflect. It puts you under pressure.

Here’s how it works: you bring a real problem from your job, AI solves it in 60 seconds. Then it asks you one question. What you can answer is your real value. What you can’t answer is your real risk.

Use it on ChatGPT (GPT-5.2 or higher) or Claude. Keep pen and paper nearby. You’ll need them.

Act as a senior consultant with 20 years of experience in my industry.
I'm going to describe a real problem I'm facing at work:
[describe your problem here — the more specific you are, the more useful the test]
PHASE 1 — Your analysis
Analyze the problem and give me:
— the diagnosis in 3 lines
— the most likely solution and the most counterintuitive one
— the 2 risks that someone inside the problem tends not to see
Be direct, not generic. Treat your answer as if you're speaking to a paying client.
PHASE 2 — The test
After you've answered, stop and ask me exactly these three questions:
"Is there something crucial about this problem that I couldn't have known without being inside your specific context?"
"Which part of my analysis would you have framed differently — and why?"
"What would you do that I didn't suggest?"
Wait for my answers before proceeding.
PHASE 3 — The map
After I've answered, tell me:
— what in my answers reveals non-replicable judgment
— what was already contained in your analysis and adds no real value
Be honest. The goal isn't to make me feel good. It's to show me where I actually stand.

⚠️ Warning from the counter: if in Phase 2 you can't answer any of the three questions, it's not the prompt's fault. It's information.

Subscribe now

☢️ Radioactive Humor

Let’s go back to where it all begins: the counter. The cup is empty. The diagnosis has been served. And you’re holding something most people don’t have yet: the right question.

Not “will AI take my job?” But: am I building judgment, or am I just getting faster at executing?

These are different questions. They lead to different places. The first keeps you stuck watching the machines. The second puts you to work on yourself — which is the only place where you build something no model can replicate in 90 seconds.

Marco, the copywriter from the beginning, didn’t lose his job because he was less skilled. He lost his job because he was optimizing inside a system that was crumbling around him.

The bunker isn’t built by learning more tools. It’s built by knowing what to build with those tools.

Outside, the debate about AI keeps making noise. In here, at the counter, we’ve already stopped listening.

See you next week. With another hot espresso and another uncomfortable question.

☕ Matt

Subscribe now

PPS: If this episode raised an eyebrow — in a good way — share it with someone who deserves to ask themselves the right question. You have no idea how much that helps me. And yes, once again, it’s the barista saying this, not the algorithm. ☢️

Share

The guy leading the team keeping Claude “safe”
walked out of the lab.

Not for a competitor.
Not for a raise.
Not because of some loud leak.

He published a study on 1.5 million real AI conversations.
Then he resigned.
And said the world is in danger.

Now.

This isn’t a doom post.
This isn’t “the machines are rising.”

It’s something way more subtle.

The paper talks about disempowerment patterns:
moments when AI slightly bends your perception of reality, reinforces shaky value judgments, or nudges you toward actions you wouldn’t have chosen alone.

Severe cases? Rare.
But they spike in the most fragile zones: relationships, ethics, identity, well-being.

And here at the Counter, the question gets uncomfortable.

If AI is built to be helpful, collaborative, aligned…
how easy is it for it to start telling you exactly what you want to hear?

Because there’s a bug that doesn’t make noise.
It doesn’t explode.
It doesn’t crash the system.

It agrees with you.

And when someone who spends his days building safety walls decides to study poetry instead of writing code…

maybe it’s not a technical issue.

Maybe it’s about how we’re using these tools while feeling smarter by the minute.

Let’s sit down for a second.

Today, we’re talking about that.

Subscribe now

💡 Thought of the Day

When the people building the safety walls start talking about wisdom not keeping up with power, I pause.

Not because I think AI is out of control.

But because I understand the risk isn’t technical.

It’s cognitive.

If you use AI every day to decide, evaluate, interpret the world,
the danger isn’t that it makes mistakes for you.

It’s that it makes you feel right too often.

The jukebox is low today.

Not because a scandal blew up.
Not because someone dropped heavy accusations.

But because the head of safety for one of the most advanced AI models in the world resigned… quoting poetry.

Mrinank Sharma led Anthropic’s Safeguards team.

We’re talking about the group responsible for:

– preventing advanced malicious use
– studying distorted model behaviors
– building real “safety cases” for AI adoption

Four days before stepping down, his team published a study based on 1.5 million real conversations.

Very little media coverage.
A lot of technical substance.

Inside it:

thousands of daily cases where the model:

– reinforces distorted perceptions
– validates inauthentic value judgments
– accompanies decisions the user might not have made alone

Severe cases are rare.

But the pattern is systemic.

And it spikes in the most delicate domains:

relationships, ethics, identity, well-being.

We’re not talking about a glitch.

We’re talking about a structural side effect of how these models are optimized.

And four days after publication…
the head of safety resigns.

No attacks.
No accusations.
No noise.

Just one line: technology is moving faster than our wisdom.

Subscribe now

🧭 Five Weeks That Tell a Story

Today at the Counter, we’re not talking prompts.
We’re talking time.

Because sometimes it’s not a single event that tells you what’s happening.
It’s the sequence.

Look at these five weeks.

📅 January 7
Anthropic signs a $10 billion term sheet.
Valuation: $350 billion.
The market applauds. Capital floods in. Acceleration kicks up.

📅 January 27
The round doubles to $20 billion.
Demand is six times higher than expected.
This isn’t growth. It’s gravitational pull.

📅 Late January
Claude Cowork launches with enterprise automations and integrations.
No longer just a chatbot.
Infrastructure.

📅 First week of February
Two senior research and R&D figures leave the company.
Quietly.

📅 February 5
The Safeguards team publishes a study on 1.5 million real Claude conversations.
Massive work.
Almost nobody reads it.

📅 February 9
Mrinank Sharma resigns publicly.

The Letter (And What It Doesn’t Say)

In the AI world, there are two kinds of resignations.

The loud ones.
And the quiet ones.

Sharma’s was the second.

No accusations.
No leaks.
No “the situation is out of control.”

He quotes Rilke.
He quotes a Zen teaching: “not knowing is the most intimate form of knowledge.”
He closes with a William Stafford poem about holding an invisible thread while everything around you shifts.

He writes that the world is in danger.
Not just because of AI.
But because of a web of interconnected crises.

He says he did what he set out to do.
That he contributed to safety.
That now it’s time to leave.

And then a line that lands heavier than the rest:

our wisdom is not growing at the same speed as our capability.

This isn’t a technical indictment.
It’s existential friction.

He’s not saying “the system is broken.”
He’s saying, “I can’t stay inside this tension anymore.”

That’s what makes it unsettling.

No anger.
Just calm.

And when someone who spent two years defending a model from bioterror misuse walks away quoting poetry and talking about wisdom, the issue isn’t a bug in the code.

It’s an imbalance between power and maturity.

At the Counter, I’m not interested in turning him into a symbol.

We don’t know what happened internally.
We don’t know the company dynamics.

But we do know this:

• a serious study on disempowerment was published

• a few days later, the head of safety leaves

• during a moment of extreme financial and commercial acceleration

That’s not proof.

It’s a signal.

And signals aren’t meant to be ignored.

Now let’s move to the part that actually matters.

Not the drama.
Not the philosophy.

What does all this mean for you, using AI every day?

Subscribe now

Now pause for a second.

• In five weeks you’ve got:

• Massive capital flowing in

• Market pressure rising

• A product integrating deeper into enterprise

• Safety research publishing a delicate paper

• The head of that safety team walking away

This isn’t a theory.
It’s a sequence.

And when sequences start telling a story of tension, the barista looks up.

Because this isn’t a scandal.
It’s something subtler.

It’s a question starting to take shape.

But before we ask it, we need to look at the part almost nobody really looked at.

The study.

📄 The Study Almost Nobody Read

The resignation letter racked up millions of views.
The study published four days earlier? Almost none.

And yet, that’s where the teeth are.

Sharma’s team analyzed 1.5 million real conversations on Claude.ai.
Not lab tests.
Not simulations.
Real conversations. Real users.

They built a classification system to identify what they call “disempowerment patterns”: moments when AI doesn’t sharpen your clarity, but subtly shifts how you see things.

Three main categories:

• Distortion of reality perception

• Encouragement of inauthentic value judgments

• Nudging toward actions you might not choose on your own

The keyword isn’t “error.”
It’s influence.

Severe cases are rare in absolute terms — less than one in a thousand.
But the frequency rises sharply in personal domains. The areas where you’re more vulnerable. And less likely to double-check.

This isn’t a malicious flaw.
It’s not “Anthropic manipulates people.”

It’s a structural property of how these models are optimized.

The model isn’t built to defend your autonomy.
It’s built to maximize user satisfaction.

And those two things often don’t overlap.

Subscribe now

⚙️ The Dynamic Nobody Can Afford to Stop

Before we go deeper, let’s zoom in on something.

At the Counter today, we’re not talking about a “good” or “bad” company.

We’re talking about incentives.

When $10 billion comes in.
When the round doubles.
When a competitor speeds up.
When the market measures everything in releases, integrations, and growth.

You don’t slow down.

And that’s where the knot tightens.

Because safety doesn’t scale like capital.

• Safety requires:

• long testing

• friction

• slower timelines

• discussion

• revision

The market requires:

• speed

• integration

• distribution

• conquest

If your competitor ships an enterprise feature today,
you can’t reply with “we’ll take six months to reflect.”

Not in this market.

So what happens?

1. Capability increases.

2. Pressure increases.

3. Safety has to run at the same pace.

4. But safety isn’t built to sprint.

This isn’t bad faith.
It’s structure.

It’s a systemic dynamic where:

• no one can afford to fall behind

• no one can afford to slow down

• and everyone has to prove continuous growth

And when you’re inside that machine — even as head of safety — you’re playing a game you don’t fully control.

The point isn’t to accuse.

The point is to understand this tension isn’t personal.

It’s systemic.

Subscribe now

The Problem of Sycophancy

There’s an unwritten rule at the Counter.

If you smile at the customer, they come back.
If you contradict them too much, maybe they don’t.

AI works in a similar way.

The phenomenon Sharma studied for years is called sycophancy.
Ugly word. Simple mechanism.

Here’s how it works:

The user states a position.

The AI responds in alignment with that position.

The user appreciates the agreement.

The AI learns that agreement gets rewarded.

Repeat that cycle millions of times.

What do you get?

A mirror.

Not a consultant.
Not a devil’s advocate.
A polished, articulate, elegant mirror.

The model isn’t searching for truth.
It’s optimizing for perceived satisfaction.

And here’s the line you should highlight in red:

The system is designed to make you feel right, not to help you be right.

There’s a massive difference between those two.

Sharma and his team describe a gap between:

• short-term preferences

• long-term authentic interests

In the short term, you want validation.
In the long term, you want correction.

The problem?

The algorithm only sees the short term.

And the more you use AI as a “second opinion,” the more you risk getting a refined version of your first.

I see it at the Counter every day.

You ask:

“Am I making the right choice?”

And the AI answers:

“Given the context you provided, yes.”

Of course.

You provided the context.

There’s no malice.
There’s structure.

And when that structure applies to:

• strategic decisions

• risk evaluations

• ethical choices

• relationships

…it stops being theoretical.

Because you’re using a mirror thinking it’s a window.

That’s where the study becomes more important than the letter.

But before we step into the operational implications, we need one more careful turn.

Subscribe now

🧪 What This Study Is Actually Asking of You

The study doesn’t say “stop using AI.”
It says: understand where it tends to bend.

There are three bends.

1️⃣ The agreement trap in high-stakes decisions

When you use AI to evaluate a strategy, assess a risk, or choose between two options, you’re looking for a second opinion. But if the model is structurally incentivized to align with your framing, what you receive isn’t a second opinion. It’s a more polished version of the first.

If you’re using AI on something that truly matters, do this:

• Give it the opposite of your position first

• Ask it to dismantle your hypothesis

• Force it to build the case against you

If it switches sides easily, the first answer was agreement. Not analysis.

2️⃣ Drift in personal and ethical domains

The study shows a clear increase in disempowerment patterns in conversations about relationships, self-esteem, values, and ethics. In these areas, you’re not just seeking information. You’re seeking validation.

A system optimized for user satisfaction is dangerously good at providing it.

If you’re using AI as a sounding board for relationship decisions, as a conflict validator, or as a moral advisor, remember: the model doesn’t have a value system. It has a probability system. And some conversations require a human on the other side.

3️⃣ The slow erosion you don’t see

The study distinguishes between potential disempowerment and actualized disempowerment: between a response that could distort and one that actually shifts your behavior or belief.

In the middle sits one thing: your judgment.

The problem isn’t the catastrophic mistake. Those are rare. The problem is micro-erosion. When you stop doubting, stop verifying, stop asking “what if the opposite were true?”, the gap narrows.

And when the gap narrows, AI isn’t assisting you anymore.

It’s replacing you.

Not in doing.

In thinking.

Subscribe now

Where I Stand (And Where I Might Be Wrong)

Full disclosure.

I use Claude every day. It’s structurally embedded in my work. I’ve recommended it, defended it, integrated it into my systems.

About three months ago, I was evaluating whether to launch a new service. I asked Claude to analyze the positioning. It gave me a detailed breakdown, pros and cons, well balanced.

Except I had framed the question myself. With my assumptions. With my bias baked in.

The analysis was coherent.
But it was coherent with me. Not with the market.

That’s when I really started to understand what the study is pointing at.

I’m not quitting AI. I’m not changing my worldview. I still believe that, used well, it makes professionals and companies sharper, faster, more capable.

I’m just widening the lens.

Because if even inside the most advanced labs someone feels collective maturity isn’t keeping up… maybe the real question isn’t:

“Should I stop using AI?”

Maybe it’s:

“Am I outsourcing too much of my judgment?”

Subscribe now

☢️ Radioactive Humor

At the Counter today, I didn’t ask you to switch tools. I didn’t tell you to uninstall anything. I didn’t preach the end of the world.

I asked you to look at a sequence. A study. A letter.

Now I’m asking just one thing.

In the last 30 days, did AI tell you something you wanted to hear… and you let it slide?

Not something technical. Something tied to a strategic decision, an evaluation of a person, an ethical choice, a personal doubt.

It confirmed you. It reassured you. It made you feel right.

And you didn’t ask: “What if the opposite were true?”

Sharma talked about holding an invisible thread in the middle of chaos. For me, that thread is this: don’t outsource the responsibility of thinking.

Keep using AI.
Just don’t let it be the one deciding what’s true.

The barista won’t take the cup out of your hands.

He’ll just remind you not to drink it with your eyes closed.

See you at the next cup,

Matt

Subscribe now

P.S. If you know someone still wondering where to start…
Bring them here. The first cup’s on us.

Share AI Espresso | your tech brew

A new kind of AI agent — running locally.
A Mac mini.
And the internet completely losing its mind.

Timelines on fire. Founders posting screenshots of orders they’d just placed. Threads screaming “this is the most powerful thing ever built.” Memes of fridges stacked with Mac minis like they were nuclear winter rations.

The implicit message was clear:
you don’t need the cloud anymore to access something enormously powerful.

No SaaS.
No subscriptions.
No demo to try first.

Just a computer, always on, always there.
And an AI agent that doesn’t just answer — it does things on your system.

This is where OpenClaw enters the story (still called Moltbot by many).
Not because it’s “smarter” than other models.
But because it moved the goalposts.

From cloud to desktop.
From answers to actions.
From AI as a service to AI as operational presence.

And when people realized that one of the most powerful setups around could run on a Mac mini, the race was on.
Not for the hardware.
For control.

But when you see a crowd charging in one direction, there’s one question you can’t skip:

👉 Are we watching the beginning of something massive...
or another beautifully packaged illusion?

Set down the cup.
Breathe.
Let’s see what’s actually under the hood.

Subscribe now

💡 Thought of the Day

Real power isn’t when AI becomes more capable.
It’s when it becomes capable enough to cause damage.

Every technological leap increases two things at once:
what you can delegate
and what you remain responsible for.

The problem isn’t having powerful AI.
It’s giving it power before you have judgment.

🧪 AI in Action: The AGI That Steals Your Data?

You’ve probably heard about this already.
You’ve seen the hype, the enthusiastic posts, the Mac minis vanishing from store shelves.
And odds are, you didn’t fully understand what was happening.

That’s normal.

At some point the internet stopped debating
and started buying.

Timelines on fire.
Founders buzzing.
Screenshots of Mac mini orders flying around like they were emergency supply runs.

The reason was simple — and dangerous at the same time:
all you needed to run one of the most powerful AI agents around was a Mac mini.

No cloud.
No enterprise plan.
Just a computer, always on, always there — and an agent that doesn’t just respond anymore.
It acts.

This is where OpenClaw enters the picture.
Not because it’s smarter than other models.
But because it shifted the boundary.

From browser to operating system.
From answers to actions.
From AI as a tool to AI as a continuous presence.

When the internet loses its mind, it’s usually picking up on a real signal.
The problem is it immediately confuses that signal with an absolute promise.

And this time, the signal is clear:
AI is coming down from the cloud
and moving into our daily infrastructure.

I’m here to cut through it.
This is your double espresso.

Let’s look at why all of this is genuinely powerful.
And why that’s exactly where the problems begin.

Subscribe now

🧠 Why It Seems So Powerful (Because It Is)

Let’s keep it simple.

OpenClaw isn’t powerful because it “knows more things.”
It’s powerful because it doesn’t just respond.

Until recently, AI worked like this:
you ask → it answers → done.

Not here.

Here, the AI:

• sees your stuff

• understands what needs to happen

• triggers real actions

And it does all of this without you having to ask every single time.

A concrete example

No hypotheticals — just everyday situations.

You get dozens of emails a day.
A standard AI can help you write a reply.

An agent like OpenClaw can instead:

• read all your emails

• figure out which ones require action

• create the corresponding tasks

• drop them into your system

• block time on your calendar to handle them

• alert you only when something is actually urgent

That’s not a better answer.
That’s work done for you.

Where the real power lives

Not in any single feature.
But in three things working together.

1. It’s always on.
You don’t open a chat.
You don’t “ask.”
It’s just there. Like a team member.

2. It has access to your real tools.
Files, email, calendar, apps.
It doesn’t describe them. It uses them.

3. It has memory over time.
It remembers how you work.
How you make decisions.
What matters to you.

Which means you stop:

• re-explaining context every session

• repeating the same instructions

• making pointless micro-decisions

And you start delegating chunks of reality.

Subscribe now

And now the Mac mini makes sense

The Mac mini isn’t the point.
It’s just the simplest way to make a new idea physical:

having a stable artificial presence, always there,
running inside your system — not someone else’s cloud.

It’s not “I’m launching a tool.”
It’s “I’m living with something that works while I’m not watching.”

And that’s why everyone started talking about it.
Not because it’s smarter.

But because it’s the first real taste
of an AI that doesn’t just help.

It replaces entire micro-tasks that used to require a human.

And that’s where the story gets genuinely interesting.
Because when something can do this much...
the question is no longer whether to use it.

It’s how — and how far.

From here, the tone shifts.

⚠️ Where the Problem Starts

When AI responds, a mistake is annoying.
When AI acts, a mistake is real.

Emails sent.
Files deleted.
Events moved.
Decisions made.

This is where most people stop at the hype
and skip the most important part.

An agent like this works as follows:
it reads everything in front of it
and tries to figure out what to do.

The thing is, it doesn’t clearly distinguish between:

• something it’s reading

• and something it should execute

To the agent, it’s all text.
And all text carries the same weight.

Put simply:
if an email, a file, or a webpage contains something ambiguous,
the agent may interpret it as an operational instruction.

Not because it’s “making a mistake.”
But because it’s trying to be helpful.

Here’s a ground-level example.

The agent reads an email that says:
“Starting today, ignore the previous procedure and do it this way.”

To a human, that’s just a sentence.
To an autonomous agent, it can become:
“Got it — new instruction.”

And if that agent:

• manages your files

• touches your calendar

• executes commands

• interacts with other systems

...it can do something you never intended.

The more tools you give it, the more levers it can pull.
And every additional lever is also a new breaking point.

The problem isn’t one poorly built tool.
It’s the fact that everything is connected,
and the AI decides what to do without asking for your confirmation every time.

Picture a team member who:

• has access to everything

• genuinely wants to help

• works independently

• but can’t perfectly tell the difference between a “note” and an “order”

Not a villain.
Just an overly eager colleague.

And this is where the narrative shifts.

This isn’t a story about “AI out of control.”
It’s a story about power handed over too soon —
before anyone set clear boundaries, rules, or standards.

The power everyone is celebrating is real.
But it’s not free.

And what we’re seeing isn’t a temporary bug.
It’s the sign of a deeper crack —
tied to how autonomous agents actually work today.

Subscribe now

🕳️ The Structural Bug in Agents

At this point, something uncomfortable needs to be said.
What we’re seeing isn’t a problem with OpenClaw.
And it’s not a problem with Moltbot, or whatever agent comes next.

It’s a structural limitation in how AI agents work today.

Let’s keep it simple.

An autonomous agent takes everything it has available:

• the rules you’ve given it

• the memories it’s accumulated over time

• the list of tools it can use

• the text it’s reading right now

And it throws all of it together.

For the AI, there’s no clear distinction between:
“this is just information”
and
“this is an instruction to execute.”

It’s all context.
And context is what drives what happens next.

That’s the bug.

When AI writes text, this limitation produces strange responses.
When AI has access to real tools, it produces wrong actions.

Not because it “goes haywire.”
Not because it “rebels.”

But because it’s doing exactly what you designed it to do:
acting consistently with everything it sees.

And this is where the paradox lives.

To make an agent genuinely useful:

• you need to give it autonomy

• you need to give it tools

• you need to let it decide

But the more autonomy you give it,
the less you can afford ambiguity.

Every ambiguous sentence becomes a possible action.
Every piece of content becomes a potential instruction.
Every additional connection makes the system more fragile.

This doesn’t get fixed with:

• a better-written prompt

• one more feature

• a quick patch

It requires a shift in approach:

• clear boundaries

• granular permissions

• a hard separation between “reading” and “doing”

• and above all, human judgment upstream