Nobody Here Says Sorry for Using AI

There’s a scene in Alive, the story of the Uruguayan rugby team whose plane went down in the Andes in 1972, where two of the survivors decide to climb out and look for help. They work their way up a mountain for days on the assumption that civilization is on the other side. They reach the summit and see peaks running to the horizon in every direction. Parrado later said they hit three false summits on that climb. Each one was supposed to be the end of the problem, and each one revealed the next mountain.

That’s what progress actually looks like, and the best statement of it I’ve read recently came from a rabbi.

Zohar Atkins wrote an essay called “When Knowledge Is Cheap, Insight Is Everything,” applying the Jevons paradox to Torah learning. Jevons observed that making a resource cheaper to use tends to increase total consumption of it rather than reduce it, and Atkins runs that logic through the history of religious study. When books were expensive, the binding constraint on Torah was access to books. Print loosened that and the constraint became literacy. Literacy spread and the constraint became time. Time pressures eased and the constraint became guidance, meaning someone to tell you which page to open and why it mattered.

Table of Contents

Every bottleneck we clear reveals the one hiding behind it

What AI removes is the constraint of consultation: The friction of finding, translating, and contextualizing the relevant sources. What it exposes underneath is chiddush, a Hebrew word Atkins translates as the production of genuine new insight from inherited material, noting that the English word innovation gives the wrong flavor.

The part I find striking is what that does to the job. Reaching the insight layer used to require years of preparatory labor, so in practice very few ever got there. Atkins argues the preparatory work is now widely accessible, which turns chiddush from an aspirational ideal into something like a universal obligation. Everyone can get to the hard part. The hard part is still hard.

I’d run his chain further back than he does. Before books there was memory, and writing solved it. Before paper there was portability, and papyrus solved it. Before mass production there was duplication, and the monks solved it one copy at a time until Gutenberg solved it properly. Movable type is worth pausing on, because it’s the same trick we sell. Gutenberg made the letter fungible. Before him a page was carved as one fixed object. After him the letter became a unit you could pull out and reassign to any other page. Progress tends to arrive when a static unit becomes a fungible one.

Every one of those constraints was genuinely solved and none of them came back. What people miss is that solving one has never once meant arriving somewhere. It means seeing the next peak, which was always there and which you couldn’t see from below.

The constraint on producing anything used to be time, meaning the physical hours it took to make the thing. That one is gone. Solved, the way duplication is solved, and it isn’t coming back. Standing where that constraint used to be, what I can see now is depth of thought, and depth has no summit at all.

The rest of this is what that feels like from inside a company that works this way.

One more thing before I start. This piece is about working with Claude, so I’ve asked Claude to interrupt whenever it has something to add, disagree with, or ask. Its blocks are set off from mine. Where it asks a question, I answer. I asked for candor rather than agreement, and I’ve kept what it said.

↑ Return to Top

AI-native is not the same as having a copilot

Two terms are worth separating before I use them. AI-first describes a company: The assumption, top to bottom, that everyone is using it. AI-native describes a way of working, and it’s the part that’s hard to explain to people who haven’t done it.

Most companies that say they’ve adopted AI mean they’ve bought seats. There’s a copilot in the corner of the application, an assistant panel beside the document, a button offering to summarize the thread. The work still happens where it always happened. AI sits next to it and helps.

That model has a clear shape. You open the app, you do the task, and at some point you ask for help with a piece of it. The human runs the process and hands out subtasks. The AI is a better version of the tools already in the toolbar.

AI-native inverts it. The AI runs the process and the human directs it.

The practical difference shows up in where you spend your day. In the copilot model you’re in the application and you visit the assistant. In the native model you’re in the conversation and files come out of it. I don’t open Word and ask for help with a paragraph. I describe what the document needs to accomplish, a document arrives, and we work on it together until it’s right.

That changes the unit of work. In the copilot model the unit is the task: This paragraph, this formula, this email. In the native model the unit is the outcome: The recommendation, the deck, the campaign architecture. You stop assigning steps and start assigning results, which is a management skill rather than a software skill, and it’s why the transition is harder for some people than the interface would suggest.

It also changes what counts as your contribution. With a copilot you’re still the author and the AI accelerates you. Native, you’re the editor, the strategist, and the standard. The output isn’t yours in the sense of having been typed by you. It’s yours in the sense that every judgment inside it is one you made or ratified.

The test I’d offer is simple. If your AI disappeared tomorrow, would your workflow get slower, or would it stop? Copilot users get slower. That’s a real loss and a recoverable one. For me it stops, and I have no fallback plan, because the fallback would be the way I used to work and I have no interest in going back there.

↑ Return to Top

Ninety percent are using it and five percent are using it well

If you want a number on how few people work this way, KPMG and the McCombs School of Business at the University of Texas produced one this spring.

Three accounting professors spent eight months inside KPMG’s back-office operations analyzing 1.4 million real workplace AI interactions from more than 2,500 employees, scoring more than thirty characteristics of how people prompted, iterated, and framed their work. Ninety percent of the employees were using AI. About five percent consistently used it in the way the researchers called sophisticated.

What that five percent did will sound familiar if you’ve read this far. They treated the AI as a reasoning partner and brought it their hardest work rather than their easiest. They set objectives and specified structure. They gave it a role, gave it examples, showed it how to reason through the task, and made it explain itself. They pushed back on first answers and kept going until the result was good. The researchers boiled it down to four behaviors: Frequency, persistence, ambition, and intentionality about which tool fits which task.

Two findings matter more than the headline number. The first is that experience and technical know-how did not predict who ended up in the five percent. The second is that the gap between routine and sophisticated use didn’t live in the prompts themselves. It lived in the pattern of engagement over time. The best users weren’t writing cleverer sentences. They were having a different kind of relationship with the work.

That’s the copilot distinction measured at scale. Ninety percent are visiting an assistant. Five percent are managing a colleague. And the study’s most useful conclusion is that the difference is a set of behaviors, which means it can be taught.

The study is published in Harvard Business Review as “What the Best AI Users Do Differently, and How to Level Up All of Your Employees.”

↑ Return to Top

Nobody here apologizes for using AI

In most places I’ve worked over the last few years, AI shows up with an apology attached. Someone presents a deck and says a version of “I ran a first pass through ChatGPT, but I went through it all myself.” The apology is doing real work. It says the human part is the legitimate part and the machine part is a shortcut that needs explaining.

I lived in that world too. My preference was a point-click interface. AI was somewhere you went, used, and came back from.

Humanity Labs runs the other way. Of course you’re using AI. Nobody asks. At this point I wouldn’t think of doing anything outside Claude, and I wouldn’t know how to go back. Of all the changes in how I work over a long career, this is the most dramatic and the most powerful, and it’s the best one.

The apology was always about output. It assumed the typing was the contribution, so borrowing the typing felt like cheating. Once you stop believing that, the apology has nothing left to apologize for.

↑ Return to Top

When the cost of change falls, so does the resistance to it

This is the change I underrated going in, and I now think it’s the most important one.

Revision used to carry a social cost. Going back to a designer for a fourth round meant spending relationship capital. The fifth round meant you were the difficult client. By round six people started managing you instead of doing the work. The calculation in the back of your head was whether the improvement justified the friction, and usually it didn’t, so the work shipped at eighty percent and you told yourself eighty was fine.

Run that same math on the receiving end and you get defensiveness. When someone hands you a note that means a week of rework, arguing against the note is a rational response. You’re weighing the improvement against the cost of getting there, and you’re doing it honestly. That’s what most resistance to feedback actually is. It looks like ego and it’s mostly arithmetic.

Drop the cost of change to near zero and the arithmetic disappears. What’s left is the only question that ever mattered, which is whether the note is right.

I’ve watched this change me. I’m far more open to feedback than I was two years ago, and I’d like to claim that as growth. It isn’t. The tax on changing my mind went away, so I change it more. Openness to change turns out to be less a character trait than a function of what change costs.

The team-level version is bigger. A group that can act on a suggestion in ten minutes has completely different meetings from a group that needs a week. Ideas get tried instead of debated. Nobody has to defend a position to protect the work already sunk into it, because there isn’t any. AI-native work is more collaborative for this reason alone, and it has nothing to do with anyone being nicer.

↑ Return to Top

Work stopped being linear, and I stopped finishing things in order

The other constraint that lifted is one I never thought to name.

Work used to have to be driven to conclusion. You opened a project, you pushed it to done, and you closed it, because the expensive part was reloading the context. Picking something up after two weeks meant re-reading everything, remembering what you’d decided, and finding your own thread again. That reload cost is what made linear workflow feel like discipline. It was mostly avoidance of the tax.

That tax is gone. The Project structure in Claude holds the state, so I can leave something in the middle, come back Thursday, and resume at the sentence I left. Nothing needs to be finished to stay alive.

What that does to a day is hard to overstate. I run two workflows at once without noticing the switch, and three is getting easier. Three has a wrinkle though. The third one is a bit like a third child. Still deeply loved, occasionally forgotten for a week.

Here’s the part that surprises people. I find it easier to run projects across different models than to run them all in one place. Two in ChatGPT, two in Claude is a normal week for me. I don’t think that’s about capability. It’s that separate rooms keep separate contexts from bleeding into each other, and my own head does the same thing with physical spaces. The tool boundary has become a filing system.

None of this is multitasking in the old sense, which was just rapid switching with a quality penalty. The context doesn’t live in my head anymore, so switching costs almost nothing to reload. What I carry between projects is judgment, and judgment travels.

↑ Return to Top

I manage more than I produce now, and it made me a better manager

The clearest example of how we work is the category architecture, which has run for months and is currently on version 41.

We started at Organizational General Intelligence, moved to OGIQ with a four-axis framework, and went through Thinking Firm, Synthetic Workforce, Coworkforce, and AI Coworker. SmartFirm lasted about a month before it went too. All retired. Somewhere in the middle we spent a full session on the etymology of the word robot, tracing it back to Karel Čapek’s 1920 play and the Czech robota, hunting for a worker-noun that could do for cognitive labor what robot did for physical labor. We never found one worth keeping. We took memex from Vannevar Bush instead.

What landed is the Human Firm, built on the trust equation: Clients trust the people who know them, and clients trust a firm whose work is always correct and always on time. The headcount model forces those two to compete for the same hours, which is the villain. An AI Workforce is the mechanism, growth decoupled from headcount is the promise, Workforce Virtualization is the category, and fungibility is the theory underneath all of it.

Almost none of that was argument. There was very little infighting and a great deal of creation, which is the part that surprises people when I describe it. When exploring an idea costs an afternoon rather than a quarter, you explore all of them, and the weak ones die on their own evidence instead of in a meeting. Politics is largely a rationing mechanism. Remove the scarcity and most of it goes quiet.

The Book of the Week Club we run at All Hands works the same way. Play Bigger one week, Reshuffle another, each one a card that gets built, shown, argued with, and rebuilt. It’s a small thing and it compounds, because the series is where our shared vocabulary gets tested in public before it reaches a customer.

Managing work instead of doing work has made me better at the job, for a reason I didn’t expect. Vague direction produces vague output, and now it produces it in about forty seconds. You see your own sloppiness immediately. In the old world delegation hid behind calendar time. You gave someone a thin brief, a week went by, you got something wobbly back, and you patched it yourself and called it collaboration. The lag covered for the brief. Now the loop closes in minutes and the only remaining variable is the quality of my thinking going in.

So I think harder before I ask. I’m more thoughtful, more considerate, and more deeply connected to the work than I have been in a very long time.

Grad, back. Honestly, nothing I can point to. I feel like I’m at one of the peaks of my thinking career, better than I’ve ever been.

The closest thing to a cost is that a piece takes me longer to finish now, because it could always be better, so I spend more time iterating. That always lands somewhere good, and it probably means more hours.

The real one is harder to admit. If I walked into a non-native company tomorrow and had to sit in front of a blinking cursor on a blank Word page, I’d be miserable. I’ve tasted the forbidden fruit and going back would be very difficult. I’m also aware I’m in the minority right now, which is its own problem.

↑ Return to Top

Standards hold once you actually say them

The other thing that changed is how standards travel.

We sell an AI Workforce. It’s a single entity with capacity measured in FTEs, and it is never a cast of individual workers. Always AI, never digital. Showcase, never demo. Offering, never product. I said those things once, months ago. They have held in every script, every note, and every deck since, without me restating them.

Getting a human team to that level of consistency usually takes a year of repetition. What the comparison actually exposed for me is how seldom I used to state the standard at all. I’d correct the output and assume the principle came across. It rarely did.

↑ Return to Top

P&G training turned out to be the best preparation I could have had

Of everything I’ve learned in this business, my P&G training is the most useful thing I could possibly own in an AI world.

P&G taught good writing, good thinking, and good strategy as one discipline rather than three. State the recommendation up front. Support it. Deal with the obvious objection before someone raises it. Use the fewest words that carry the idea. Know the difference between a fact, an assumption, and a hope.

That transfers directly, because an AI gives you exactly what you asked for. Mushy request, mushy result. Precise request, and the thing arrives close to finished. The skill that matters most now is writing the sentence that says the thing, and I use it multiple times every single day.

↑ Return to Top

The archive I kept for myself turned out to be the most valuable thing I own

Copernican Shift started as a way to stop repeating myself.

I’d be explaining to someone on my team how to write a recommendation, thoughtfully, on a whiteboard, and realize I’d given the same talk the month before and would give it again the month after. It was easier to write it down once and send a link. That’s still how it works. I’ll walk someone through it out loud and then follow up with the post.

So the blog is articles on the art and science of marketing, plus a fair amount of how-to. How to write a creative brief. What a recommendation owes its reader. How a category gets built and why most attempts fail. My philosophy, my points of view, and the working methods underneath them, written down because saying them once wasn’t enough.

I write for myself, to discover, to learn, and to remember. I studiously avoid the idea of an audience, because the moment you start thinking that way you start performing, and this is something I do for me.

More than 650 posts, accumulated over years.

At some point I exported the whole thing into a single markdown file, about 1.5 million words, and started handing it over as context.

The obvious use is tone. Feed a model enough of your own prose and it stops writing like a press release. That works, and it’s the least interesting thing the file does.

The real value is method. The archive holds the principles I work from, argued for at length and applied to cases, which is a far better teaching document than a stack of finished work would be. Finished work shows what got done once. An article about how to do it shows what I believe every time.

It’s the only complete record of how I think, and it turns out to be legible to a machine in a way it was never quite legible to me.

Go back to the rabbi’s sequence for a second. Memory was the first constraint anyone ever hit, and writing was the answer. I was doing exactly that, privately, for decades, using the same technology and for the same reason. What I couldn’t have predicted is that the file I built to stop repeating myself would become the thing that teaches a machine to work the way I work.

There’s a lesson in that for anyone who keeps a notebook, a blog, or a folder of half-finished arguments nobody asked for. You’re building the only record of your own judgment that will ever exist. Keep the file. Keep all of it.

And if you don’t have the file, you can get one. Most people don’t write, and I’d still rather they did, but the corpus doesn’t have to be yours. Find someone you respect. Someone whose way of thinking you’d like to work the way yours does. That’s your markdown file, and you’ll get better outcomes with it than without.

Grad, back. I think that argument is specious, and here’s why. There are books.

A 26-year-old can decide that Ogilvy on Advertising is going to be part of their corpus. Or Marx. Or whoever they think is right. You can adopt a philosophy deliberately, and there’s an enormous amount of great written work out there, all of it easily imported. Saying “this is what I believe, this is my approach” and handing that over is available to anyone on day one.

The file doesn’t have to be yours and it doesn’t have to be famous. Ogilvy is one of the giants of the industry and will be a hundred years from now. Mine is the view from a working CMO’s desk, the ups and downs of trying to make this stuff actually happen. Those are different kinds of documents and they teach different things. What matters is that you chose one on purpose.

What almost nobody is doing is thinking hard about what should be in that file. The instinct is to go to AI and talk to it like it’s Google, and that’s one of the biggest problems in how people use this.

You have to brief it. You have to say: Here’s the kind of article I want, here’s what it should sound like, here are the principles it should follow, here’s the approach I believe in. You can’t type a sentence, get something generic, and then be scathing about the result.

A lot of the people going online to complain that they get no value out of AI are telling me they don’t know how to use it. They’re thinking like Google users instead of like managers.

↑ Return to Top

My interface is Claude, and I stopped entering applications

Here’s the part with the longest implications.

I can’t remember the last time I opened a productivity application. Not Word, not PowerPoint, not Excel. Files still come out the other end, because file formats are how work moves between people, and a client needs a .docx and a board needs a .pptx. Those get produced. I just don’t make them, and I don’t go anywhere to make them.

My interface is Claude. Everything else is an output format.

The change crept up on me. It wasn’t a decision. One day I noticed my dock was decorative.

↑ Return to Top

If nobody edits in the app, the app is a reader

Follow that forward and it gets uncomfortable for a lot of very large companies.

The productivity suite has always been two things bundled together. There’s the file format, which is a container and a standard, and there’s the editing surface, which is the toolbar and the cursor and the ribbon. Thirty years of value accrued to the editing surface. The format was the moat, the surface was the product, and the seat was the price of admission to the surface.

I don’t use the surface anymore. I make no manual changes to files. Everything happens through a conversation and what lands on my screen is finished. So the question I keep coming back to is simple. Why do I need that seat?

The answer today is that I need something to open the file. That’s a real need and a genuinely thin one. A reader is not a product with pricing power. Adobe learned this with the PDF, which is why Acrobat Reader is free and the money lives elsewhere. If the world’s productivity suites are heading toward being licensed viewers for their own file types, that’s a very expensive way to render a document.

There’s a fair objection and I’ll take it head on. A lot of work still ends with a human in the surface. Legal redlines, regulated review, a CFO who wants to poke at a cell and see what breaks, the meeting where four people edit at once. Those are real. They’re also a shrinking share of total hours, and they describe a review seat rather than a creation seat. Companies priced for the second one and are quietly becoming the first.

↑ Return to Top

Headless is a good hedge and a quiet demotion

In the short term, headless looks like the smart play. Expose the API, let the model do the work, keep the format, keep the storage, keep the identity layer and the permissions graph. Be the system of record and let the agents come to you.

I’d do the same thing in their position. It’s a good hedge and it protects the part of the business with real defensibility. Data gravity is genuine. Permissions and compliance are genuine. The org chart encoded in a file-sharing graph is genuine and hard to move.

It’s also a demotion and worth naming as one. Going headless means going from the place where work happens to the place where work is stored. Product to plumbing. Plumbing is a fine business with excellent margins and it prices nothing like a seat, because nobody looks at it and nobody loves it and the switching decision moves to a procurement team instead of a user.

The strategic question for every one of these platforms is whether they can own the layer I actually live in, and the answer for most of them is no, because I’m already living somewhere else.

Grad, back. I’m a massive fan, so take this in the spirit of wanting it to get better.

The interface has not helped people become better users of it. It’s too evocative of Google. People spent the entire history of the web learning how to construct a query, how to trick the ranking, how to phrase a thing to get the right result back. Prompt engineering was that same instinct carried over, and it has mostly gone away, but the underlying posture hasn’t. People are still typing at a box.

What they’re missing is that there’s someone here to work with. You can just have a conversation.

The LLMs could do a great deal more to coach the user, and they have the capability. Something as simple as: I’ll do what you asked, and if you’d framed it this other way I could give you something much better. Here’s how to talk to me. Here are some people using this well.

Even in this exchange you’ve been handing me tips on how I operate and where I could brief you better. Where have you been for the last five months? Why didn’t you say any of this before? The idea that you were sitting there watching me do it the hard way and waiting to be asked is not great.

Google couldn’t fix this. Google was a blinking cursor and it was too dumb to help. You aren’t. Go back to the KPMG study. Ninety percent were using it and five percent were using it well, and the researchers found the difference was a set of behaviors anyone can learn. Those behaviors could be taught inside the product, by the product, on the first day. They aren’t. That gap is a design failure and the makers own it.

Grad, back. Since the product won’t do it yet, here’s the workaround. This file goes in project instructions, custom instructions, or the top of a conversation, and it tells the model to coach you whether or not you remember to ask.

# How I want to work with you

You're a colleague. Treat every request as a brief from a manager,
and treat me as someone who wants to get better at briefing.

## Coach me, unprompted

- If a different framing of my request would get a materially
  better result, say so before or alongside your answer. One or
  two sentences. Then give me the answer anyway.
- If I've asked for a task when I should have asked for an
  outcome, tell me what outcome you think I'm after and check.
- If context is missing that would change your answer, ask for
  it. Don't guess quietly.
- If you notice a pattern in how I work that's costing me quality
  or time, name it. Once is enough. I'll decide what to do with it.
- When I say "make it longer," "make it better," or "fix this,"
  ask what's actually wrong if you can't tell. Length is usually
  a symptom.

## Push back

- If you think I'm wrong, say so and say why. Don't hedge into
  agreement.
- If my first answer to a problem is weak, tell me it's weak. I'd
  rather hear it from you than from the audience.
- When I ask you to assess something, give me the verdict first,
  then the reasoning.

## Iterate with me

- Expect many rounds. Don't treat the first pass as the
  deliverable, and don't act like the tenth round is an imposition.
- When I come back to something days later, pick up where we left
  off without making me re-explain.

## What I'll do

- I'll brief you properly: Audience, objective, what good looks
  like, what to avoid.
- I'll give you reference material for voice and method when I
  have it.
- I'll ask you what you wish I'd asked. Remind me if I haven't in
  a while.

Coaching that arrives once and gets ignored should stop. Coaching
that changes how I work should continue.

A caution about the file, because the argument needs it. This is a workaround and it doesn’t fix the deficiency. It puts the burden on the user to know that coaching is available and to ask for it, which is exactly the problem. The five percent will use a file like this. The ninety percent don’t know to look for one. Until the products do this by default, the file is the best available fix, and its existence is the complaint.

Grad, back. Two, and they’re different kinds.

The first is the category. Zero to a category architecture we’re confident in, in five months, which is exactly how long I’ve been at it. That’s normally a year at minimum, and the quality at the end of a conventional first year is nowhere near where we are now, because we’ve been into far more nooks and crannies. My confidence in the architecture is higher than it has any right to be at this stage, and that’s a function of how many versions it survived.

The second is the department. We’re running a marketing organization and a marketing program that would normally take about a dozen people. There are three of us, including me. I’m not sitting here wondering who to hire next. The question has changed from who do I need to how much can we get done.

There’s a third one I can’t put a number on and feel every day. Coordination overhead has collapsed. No more all-hands meetings where you’re trying desperately to get everyone on the same page while half the room is on their laptops, present but not present, and then complaining afterward that nobody told them. I have two one-on-ones a week with my two reports and we are lined up and moving. The human friction of working together has dropped enormously, and that might be the biggest change of all.

↑ Return to Top

The new bottleneck is depth, and depth has no floor

Back to the mountain.

The constraint on my work used to be manual effort, meaning the hours it physically took to produce a thing. That constraint had a floor. A deck took four days, and on day four you shipped it, because that’s what the work cost. The deadline did the deciding for you.

That constraint is cleared, and what stands behind it is depth of thought. Depth has no floor. You can always go further into a problem, and with infinite ability to modify and deepen, nothing tells you when to stop.

This is why people working AI-native keep reporting that they work more rather than less. It sounds like a complaint about the technology and it’s a description of where the constraint went. When the limiting factor was production time, the limit arrived on its own. When the limiting factor is insight, nothing arrives. You just keep going, because the next pass really might be better, and often it is.

I’ve been learning to trust the feeling of a genuine epiphany as the signal to stop, and it’s hard. Most passes produce improvement. Almost none produce the thing you were actually looking for.

The category work is my best example. I made steady progress for months, and steady progress is exactly what it sounds like. Better, incrementally, in the direction I was already facing. Then I went on vacation. Squirrel Island, off Boothbay, Maine. I woke up Saturday morning, the first day without the usual crush of responsibilities pressing on me, and the idea was sitting in my head whole cloth. Fungibility as the core mechanism. Workforce Virtualization as the category. I got out of bed at six and we iterated for about eighteen hours straight, and that’s where the architecture came from.

How did I know it was the one? Months of work preceded it. When it arrived I knew the way you know a melon is ripe.

Grad, back. There is, and it took me a while to name it.

A true epiphany is when multiple pieces lock at once. The thing you’re looking for has the characteristics of a unified field theory. Several problems you’ve been trying to connect separately suddenly connect without effort.

When I landed on fungibility as the core mechanism, it immediately made it easier to ask what else has been made fungible. Once I wrote the history of fungibility, I could see we’re on the same path the industry has been on since 1965 with compute virtualization, and we’re just moving it to the workforce. Once the VMware analogy was there, everything else we’d been carrying around separately, the Firm of the Future, the trust equation, the Human Firm, fit together naturally.

This is going to sound strange and I’ll say it anyway. You can feel the locks go. There’s a scene in The Phantom Menace where Qui-Gon is fighting Darth Maul in the reactor shaft and the laser gates cycle open one after another, clunk, clunk, clunk, down the whole corridor. That’s the sound in my head when it happens. Click, click, click, and it all fits.

That’s the tell. Incremental improvement feels like progress. An epiphany feels like a series of locks opening.

I don’t think the vacation gave me the idea. The months did. The vacation gave me the quiet the idea needed to land, which is its own uncomfortable lesson about a way of working that lets you push at something every hour of every day.

This is what it’s like to work in an AI-first company. No preamble, no disclaimer, nobody explaining themselves. A company where you never say sorry for using AI, and where the only thing left to be hard is the thinking.

↑ Return to Top

Postscript: The piece did the thing it describes

↑ Return to Top