The film is by Daniel Roher, who made Navalny, with Charlie Tyrell, and it is produced by the team behind Everything Everywhere All at Once. The setup is simple and disarming. Roher finds out he is about to become a father, and he gets scared about the world his kid is going to inherit. So he does what a documentary maker does. He goes and asks everyone.

Everyone, and I mean everyone

He interviews more than forty people, and the casting is the whole point. Sam Altman and Dario Amodei and Demis Hassabis, the people building the thing. Eliezer Yudkowsky and Connor Leahy and Yoshua Bengio, the people who think it might kill us. Timnit Gebru and Emily Bender and Karen Hao, the people who think the real damage is quieter and already here. Guillaume Verdon and Peter Diamandis, the people who think we are about to walk into paradise if we would just get out of the way.

Nobody in that lineup is a crank. They are all serious, and a lot of them flatly contradict each other, and the film does not rescue you from that. It just sits you down in the middle of it. You leave without a tidy answer, which is exactly the experience of actually paying attention to this field.

The word

Here is what Roher lands on, and what I have not stopped thinking about since. An apocaloptimist is someone who takes the catastrophic risk completely seriously and refuses to be paralyzed by it. You hold both. The odds of this going badly are real and not small. And the only useful response is to roll up your sleeves and try to steer it toward the good outcome anyway.

That is not a compromise between the doomers and the optimists. It is a posture you adopt on purpose because the other two available postures are useless.

The two cop-outs

I see leaders default to one of two positions, and both are a way of not having to do anything.

The first is denial. It is all hype, the models are just fancy autocomplete, this will blow over like crypto did. This is comfortable because it means you can keep doing exactly what you were doing. It is also wrong, and the people saying it in 2026 are going to feel the way the people who dismissed the internet in 1998 feel now.

The second is doom. We are cooked, the machines win, nothing I do matters. This one feels sophisticated because it sounds like you have thought hard about it. But it lands in the same place as denial. If nothing matters, you are off the hook. You get to feel smart and do nothing.

Apocaloptimism is the only one of the three that comes with a to-do list. That is why I think it is the right one, and it has nothing to do with being a positive person. It is the only stance that treats the future as something you have a hand in.

Why this is a leadership posture, not a vibe

I run fractional executive engagements, and this is the conversation underneath most of them. A board wants me to tell them whether AI is salvation or catastrophe so they can file it and move on. The honest answer is that both are genuinely on the table, and which one your company walks into is partly up to the choices you make this year.

That is uncomfortable, because it means you do not get to outsource the outcome to fate or to the labs. You have to make decisions while holding real uncertainty, which is the actual job of leadership and always has been. The apocaloptimist does not know how it ends. They just refuse to sit it out.

What to do this week

1. Notice which cop-out is yours. Denial or doom. Everyone leans one way. Name yours, because it is quietly shaping every AI decision you make or avoid.

2. Write down the good outcome you are actually working toward. Not a vague better future. A specific one for your company and your people. You cannot steer toward a place you have not named.

3. Do one concrete thing that assumes you have a say. Ship a small AI project, set one real guardrail, have one honest conversation with your team about what changes. Action is the whole difference between the postures.

4. Watch the film. Under two hours, and it will do more for your thinking than a stack of think pieces. Bring someone who disagrees with you.

The part that got me

The reason the new-baby framing works is that you cannot be a nihilist with a kid in the room. A child is a bet on the future by definition. The moment Roher has to look at his, the doom posture stops being available, and the only move left is to try.

I think that is true whether or not there is a baby involved. Somewhere downstream of the choices you are making right now, there are people who will live in the world those choices build. That is reason enough to be an apocaloptimist. Take the risk seriously. Then go build the good version anyway.

The AI Doc: Or How I Became an Apocaloptimist is directed by Daniel Roher and Charlie Tyrell, from Focus Features. It is streaming on Peacock and available to rent or buy on Apple TV, Prime Video, and elsewhere (check where to watch). Go watch it.

Some links above are affiliate links. As an Amazon Associate I earn from qualifying purchases.

Everyone wants to argue about which model to use. Almost nobody wants to talk about how the documents get in. That is backwards. The model is a commodity you can swap in an afternoon. What you feed it is the whole game.

I have a plain way of saying it. Garbage in, confident nonsense out. A weak model on clean, well-labeled information will beat a frontier model drowning in duplicates, stale drafts, and mislabeled files every single time. The intelligence people think they are buying lives in the ingestion pipeline, which is the least glamorous part of the stack and the part that decides whether you can trust the answer.

The boring plumbing that does all the work

Ingestion is four unglamorous jobs. Get them right and the AI feels like magic. Get them wrong and it feels like a confident liar.

Provenance. Every answer has to be able to show its work. When the assistant tells you something, you need to click through to the exact document it came from. No source, no trust. This is the difference between "the AI said so" and "the AI found this in the signed contract, page four." One of those you can take to a board. The other you cannot take anywhere.

Freshness. The gap between when a document changes and when your AI notices is a lie waiting to be told. Someone updates the pricing page Monday. If the system does not pick that up until Friday, your assistant spends the week quoting last week's prices with total confidence. Somebody has to own the question of how fast changes flow through.

Deduplication. The same policy exists in nine slightly different versions across the company. Which one is real? If you feed all nine to the AI, it will average them into something that was never true and quote it back to you. Duplicates are not clutter. They are a way to manufacture wrong answers.

Metadata. Every document needs to carry its own labels. Who owns it. Who can see it. When it was last true. What it supersedes. This is the layer that lets the machine tell a signed contract from a scratch draft, and it is almost always the layer nobody has bothered to create.

What the nine refund policies do to you

Here is deduplication as it actually plays out, because "manufacture wrong answers" sounds abstract until it costs you a customer.

A company has a refund policy. Over five years it got pasted into an onboarding deck, a help-center article, two Notion pages, a sales enablement doc, and a contract template, and each copy got edited a little by whoever owned that surface. The window is fourteen days in one, thirty in another, "case by case" in a third. All nine are in the pile you fed the assistant. A customer asks about a refund. The AI does not pick the right one, because it has no way to know which is right. It blends them into a confident, reasonable-sounding answer that matches none of the nine and commits you to a promise no human approved.

Nobody wrote that policy. The machine did, by averaging your mess. That is what duplicates buy you. Not clutter. A liability with good grammar.

Where the file approach quietly wins

I keep coming back to plain files in this series, and ingestion is where the argument gets concrete.

In a folder of markdown files tracked with normal version control, three of these four jobs come mostly for free. Provenance is the file history. You can see who changed what, when, and why, because that is what version control does. Freshness is a habit, not a rebuild. You do not re-process the whole knowledge base to reflect a change. You edit the file. Metadata is a few lines at the top of the document, sitting right next to the content it describes, in plain text a human can read and correct.

Compare that to the standard approach, where every change means re-processing the document into a specialty database, and provenance is something you have to engineer back in on purpose. The file approach does not make ingestion trivial. Deduplication is still real work, and someone still has to care. But it starts you much closer to trustworthy, because the audit trail is built into how the files already work.

What I am seeing in the room

The companies with AI they actually trust did not buy better models. They did the janitorial work. They killed the duplicate policies. They put one owner on freshness. They made every answer cite its source, and they made "no source" a failure, not a shrug.

The companies that do not trust their AI skipped all of that and are now surprised that a system fed on mess produces mess. It is not the model. It never was.

What to do this week

1. Pick one important question and trace the answer. Ask your system something that matters, then find the exact document it used. If you cannot, provenance is your first project.

2. Hunt the duplicates. Take one core policy and count how many versions exist across your tools. The number will bother you. That is the point.

3. Put a clock on freshness. For your top five documents, how long between a change and the AI knowing? If nobody can answer, nobody owns it.

4. Add labels to the top ten. Owner, audience, last-verified date. Ten documents. One afternoon. Start there.

The uncomfortable truth is that most of the work in a trustworthy AI system happens before the model ever runs. It is intake, labeling, and cleanup. It is not exciting. It is the difference between an assistant your team relies on and one they have quietly stopped believing.

Next in the series I get concrete about the storage fight everyone has too early: why I keep replacing the vector database with a plain folder of markdown files, and the honest cases where a folder is not enough.

The Corporate Second Brain, a six-part series

1. Your Company Already Has a Second Brain

2. RBAC Is the Hard Part of AI

3. Why Ingestion Is the Whole Ballgame (this post)

4. I Replaced the Vector Database with a Folder

5. From What We Know to What You Should Do

6. Build a Corporate Second Brain in a Weekend

That is not a knock on him. It is a knock on the fact that the job exists at all. A hotel that nice pays someone to walk that same twenty feet of concrete, over and over, all day, looking for small pieces of trash and removing them. It is honest work and it is completely solvable, and once I saw it I could not stop seeing versions of it everywhere I looked.

The actual machine, spelled out

Here is the whole system, and there is nothing science fiction about any piece of it.

Mount a camera high on the building, out of the way, where no guest ever notices it. Point it at the entrance. It watches that stretch of ground all day and knows what clean looks like. The moment a wrapper blows in or someone drops a receipt, it sees the difference, marks the exact spot, and sends a small floor robot straight to it. The robot rolls over, picks it up, and goes back to its dock. Then in the quiet hours after midnight, the same fleet comes out and sweeps and polishes the whole entrance while the city sleeps.

Nobody sweeps the front of that hotel again. The lobby cleans itself.

That is one building, one small idea. The reason it matters is that the same shape is hiding under a thousand other jobs, and once you learn to see the shape you find it on every block.

None of this is waiting on invention

The reason we file this under someday is that we have watched self-cleaning everything in movies for forty years, so the whole idea feels like set dressing for the future. It is not. Every part of it already ships.

Computer vision that can pick a bottle cap out of a photo is a solved problem you can rent by the hour. Robots that navigate a space on their own and scrub floors overnight are already humming through big-box stores and airport terminals after closing. Cheap compute to run the whole loop is sitting in a box the size of a paperback. We are not waiting on a breakthrough. We are waiting on someone to bolt the existing pieces together and for the price to drop below what that sidewalk currently costs in wages.

Both of those are happening right now. That is why I think the self-cleaning entrance is a five-to-ten-year thing, not a someday thing. The hard part was never the science. It was the integration and the math, and the math gets better every quarter.

Learn to see the shape

The sweeping man is a stand-in for an entire category of work. Strip his job down and it is this: watch a defined space, notice a small problem, go handle it, repeat. That exact pattern is everywhere.

The parking lot that gets walked for stray carts. The grocery aisle checked for a spill. The pool deck cleared of towels. The warehouse row scanned for a box in the walkway. The office that gets tidied every night by someone going surface to surface. Every one of those is the same loop, and every one of them is the kind of narrow, physical, repeatable task that cheap cameras and cheap robots are about to be very good at.

The mistake leaders make is looking for the giant, glamorous AI project. The real opportunities look like a broom. They are small, boring, and repeated a hundred times a day, and that repetition is exactly what makes them worth automating.

The part I will not skip over

That sidewalk is somebody's paycheck. I am not going to wave that away with a line about everyone moving on to better things, because that is not always how it goes, and pretending otherwise is how technologists lose the room.

Here is what I will say. The work that is easiest to hand to a machine is the work we most wish we did not have to pay a person to stand and do all day. The honest move is to see it coming, say so plainly, and put real thought into where those people go next, instead of acting surprised when the robot shows up. The change is coming either way. The only choice is whether you meet it with a plan or with a shrug.

What to do this week

1. Walk your own operation and count the brooms. Every task that is watch-a-space, notice-a-thing, go-fix-it. Write them down. That list is your automation roadmap, in priority order.

2. Start with the most repeated one, not the most impressive one. The task someone does forty times a day beats the flashy moonshot every time.

3. Price the status quo honestly. What does that repeated task cost you a year in wages, turnover, and attention? That number is the budget you have to work with, and it is usually bigger than you think.

4. Assume the pieces already exist. Before you decide something is too futuristic, go check. Most of the time the camera, the robot, and the model are all sitting on a shelf waiting to be wired together.

The future of AI is not going to arrive as one humanoid butler walking through the lobby. It is going to arrive as a camera you never notice and a Roomba you never think about, and one morning you will realize nobody has swept that sidewalk in months.

Once you see the shape, you cannot unsee it. Look around on your next walk. They are everywhere.

The demo always works. It works because in a demo, the AI can see everything. You ask it anything and it answers, because there are no walls in a sandbox.

Then you roll it out to the whole company, and you remember that your company has walls for good reasons. HR data. Board materials. The comp spreadsheet. The deal that has not closed. The moment the assistant can read all of it, you have not built a productivity tool. You have built the fastest way to leak your own company to itself.

This is the part nobody wants to hear. The hard problem in enterprise AI is not the model. It is who is allowed to see what.

Retrieval is a permission decision, not a search

Here is the reframe I give every technical leader. When your AI pulls a document to answer a question, that is not a search. It is an authorization decision. The system is deciding this person is allowed to see this. Most teams build it as a search and bolt permissions on afterward, and that is exactly how the comp spreadsheet ends up quoted in a chat with a summer intern.

The rule is simple to say and hard to do. A document the user is not allowed to open must never enter the AI's context in the first place. Not filtered out of the answer after the fact. Never retrieved at all. If the machine read it to write the answer, it does not matter that it "chose not to mention it." It already leaked in the phrasing.

The good teams enforce this in two places. At ingestion, when a document comes in, it is tagged with who can see it, and anything sensitive is walled off before it is ever indexed. At retrieval, every lookup is scoped to the person asking. Two gates. OpenAI reportedly runs a permission system across tens of billions of documents to do exactly this. The scale is enormous. The principle is the same one your file server has used for thirty years.

The permission that outlived the person

Let me make the failure concrete, because it does not look like a hack. It looks like an ordinary Tuesday.

An analyst moves from finance to marketing. Nobody revokes the finance access, because nobody ever does. A quarter later a document from the finance drive gets copied into the AI's index, and the permissions ride along frozen as they were the day of the copy. Now marketing can ask the assistant about margins they were never supposed to see, and the assistant answers, helpfully, with a citation. No alarm goes off. Nobody chose this. It is just three small bits of drift lining up, and the machine treating stale permissions as current truth.

That is the whole risk in one scene. Access control is not hard because the concept is hard. It is hard because permissions go stale, and an AI applies stale rules faster and more confidently than any human ever would.

Why governance is where projects rot

The trap is copying permissions into a second place. The day you duplicate "who can see this" into a separate search index, you have signed up to keep two systems in sync forever. They will drift. They always drift.

A document gets copied into the AI index with the permissions frozen as they were the day it was copied. Six months later the source says "confidential" and the index still says "everyone." Now your AI is enforcing a rule that was true in March. This is the quiet failure mode that turns a working system into a liability nobody notices until it quotes the wrong thing to the wrong person.

The underrated advantage of plain files

This is where I break from the standard playbook, and it is the thread running through this whole series.

If your knowledge lives in a folder of files instead of a specialty database, the permission system is the one your operating system already ships. The folder either lets this person in or it does not. There is no second copy of the rules to keep in sync, because the files are the rules. Put the deal room in a directory the deal team can open and nobody else can. The AI acting on your behalf inherits your access. It cannot read what you cannot read, because it is walking the same filesystem you are.

I want to be honest about the edge. Scoping an agent so it truly cannot wander into a folder it should not touch is real engineering, and the tooling for doing it cleanly is still young. Anyone who tells you this is solved is selling something. But starting from "the files carry their own permissions" is a much better place to stand than "we copied everything into a search engine and now we manage access twice."

What to do this week

1. Ask for one number. How many places does "who can see this document" currently live? If the answer is more than one, that is your drift risk, named.

2. Run the intern test. If you gave the AI assistant to your newest hire today, what is the first thing they could ask it that they should not be able to see? If you do not know, you are not ready to turn it on.

3. Decide where retrieval gets scoped. Before generation, not after. Write that requirement down before you evaluate a single vendor.

4. Keep permissions in one place. Wherever the document actually lives. Do not let a shiny tool talk you into a second copy of the rules.

The board question I would ask this week is short. When we turn this on, whose access does it use? If the answer is "everyone's," you do not have an AI project yet. You have an incident waiting for a date.

Next up, the layer that decides whether the assistant is trustworthy or just fluent: ingestion, and why garbage in means confident nonsense out. Access control keeps the AI from reading what it should not. Ingestion decides whether what it does read is even true.

The Corporate Second Brain, a six-part series

1. Your Company Already Has a Second Brain

2. RBAC Is the Hard Part of AI (this post)

3. Why Ingestion Is the Whole Ballgame

4. I Replaced the Vector Database with a Folder

5. From What We Know to What You Should Do

6. Build a Corporate Second Brain in a Weekend

Now you keep going. You warm up your tan with a slider. You nudge your smile a hair wider. You tap the tourist standing behind you and he vanishes, replaced by a stretch of beach that was never actually empty. You send it to your family. This is my trip.

Somewhere in that sequence, was there a moment it stopped being a real photo? I have been chewing on this, and the honest answer bothered me. There was no such moment, because it was never a real photo to start with.

Your phone never took one picture

Here is the part almost nobody realizes. When you press the shutter, your phone does not capture a single frame. It captures a burst of them, at different exposures, and merges them into one computed image before you ever see it. That is what HDR is. Apple, Google, and Samsung all do it, by default, on every shot.

So the picture you think of as the raw, untouched original is already a composite that no single instant of reality ever looked like. "Straight out of camera" stopped meaning anything years ago. The sunset was fixed before you got a vote.

That is the floor. Everything else is a question of degree.

From fixing to fabricating

There is a real line in here, and it is worth naming precisely. Some processing reconstructs what was actually in front of you: pulling detail out of shadow, steadying your shaky hands, balancing a sky your eyes also adjusted to in the moment. Fair enough. Your own vision does versions of that.

Then there are features that invent things that never happened, and they are already in your pocket.

Google's Best Take will take a group photo and swap in each person's face from a different second, so everyone is smiling in a frame that never existed. Magic Eraser and Apple's Clean Up will delete the tourist and generate fresh sand where he stood, painting in pixels no lens ever saw. In 2023 someone pointed a Samsung phone at a deliberately blurred photo of the moon on a monitor, and the phone handed back crisp craters that simply were not in the source. Samsung's explanation was that its software recognizes the moon and enhances the detail. Which is a careful way of admitting the detail was added.

None of that is editing a photograph. It is manufacturing one and calling it a photograph.

Two camps, and they are moving apart

Ask who gets to decide where the line sits and you find two worldviews heading in opposite directions.

Google's Pixel camera lead, Isaac Reynolds, has said the quiet part out loud. His framing is memory, not record:

Your memories are your reality. What's more real than your memory of it?

The goal, in his words, is to produce the moment you wanted and remembered, not the one the sensor happened to catch. Under that view, erasing the tourist is not a lie. It is restoring the picture to what the day felt like.

Now the other camp. World Press Photo, which governs the most serious photojournalism on earth, spent 2026 tightening its rules in the exact opposite direction. To them a photograph is light recorded on a sensor, full stop. No generative fill. No adding, moving, or removing people or objects. They will let you crop and clean a dust speck, and that is nearly the whole list.

Same technology, two irreconcilable answers. One says a photo is a feeling you are allowed to perfect. The other says a photo is a fact you are not allowed to touch.

So how do you even know?

This is the part that should worry you more than the sliders. When someone hands you a photo now, you usually cannot tell from the image itself what was done to it. The pixels look clean either way.

The industry's answer is provenance: proving where a picture came from instead of judging the pixels. C2PA Content Credentials attach a signed, tamper-evident history to a file, a kind of nutrition label showing what device shot it and what edits touched it. Google is starting to stamp video from recent Pixels this way. Its SynthID system hides an invisible watermark inside AI-generated images and has tagged a hundred billion of them, and it is starting to show up in Search and Chrome as an "is this AI?" signal.

That is real progress, and it is worth supporting. But do not mistake it for a lie detector. Content Credentials record history, they do not rule on truth, and the label can be stripped off a file entirely, so its absence tells you nothing. And researchers have already shown you can forge one of these invisible watermarks onto a picture that was never watermarked, so its presence does not prove much either. For now, the honest position is that you often cannot know, and anyone who tells you a photo is provably real is overselling.

Where I land on "too far"

The line was never really about the pixels. It is about the promise.

Every photo comes with an implied claim. A shot you send your family that says this is roughly how the trip felt is a memory, and warming the light or smoothing your tan is inside that promise. A shot that says this is what happened, believe it, is a record, and the moment you invent a smile, delete a person, or add craters that were not there, you have broken the deal, whether a court, a customer, or your kids are the audience.

So my rule is short. Decide what the picture is for before you edit it, and be honest with whoever sees it about which one it is. Fixing how it felt is fine. Fabricating what happened and passing it off as evidence is the line.

What to do this week

1. Learn your phone's defaults. Open your camera settings and see what is on. You are already shipping edited photos and calling them originals.

2. When it has to be real, keep the original. For anything that might matter later, a receipt, a claim, a record, save an unedited copy and note that it is untouched.

3. Stop trusting the pixels. Assume any photo you are shown could be fabricated, especially the emotionally convenient ones. Look for content credentials, but treat them as a clue, not a verdict.

4. Say what you did. If you erased the tourist, it costs you nothing to mention it. Honesty about the edit is the whole ballgame.

The camera used to be the thing we reached for when we needed proof. Point it, click, done, argument over. That era is ending. The camera is becoming a suggestion, a draft of reality that each of us finishes to taste.

The useful question is no longer whether a photo is real. It is what the photo is for, and whether the person showing it to you is being honest about that. Everything now runs on that honesty. Guard it.

Your company already knows how to solve the problem in front of you. Someone wrote it down eighteen months ago. It is sitting in a Slack thread, a Notion page, a closed Zendesk ticket, and a Google Doc that three people have edit access to and nobody has opened since.

That is the real state of most companies I walk into. Not a knowledge gap. A retrieval gap. The knowledge exists. It is just scattered across forty tools, and no human can hold the map.

So leadership hears "AI" and buys a chatbot. Six weeks later it is answering questions confidently and wrong, and the project quietly dies. Depending on whose numbers you trust, most corporate AI projects never make it to production. I believe it, because I have watched the same failure up close. The tool was fine. There was no architecture underneath it.

The answer was already in the building

Here is the shape of it, over and over. A support team spends two days fighting a customer's integration problem. They escalate it, pull in an engineer, burn a chunk of everyone's week. Somewhere in that same company, eighteen months ago, a different engineer solved the identical problem and wrote up exactly how. It is sitting in a closed ticket nobody thought to search, under a title that does not match the words this team used.

That is not a knowledge problem. The knowledge was there. It is a retrieval problem, and retrieval is the thing no human can do across forty tools and a decade of history. You cannot hold the map. Nobody can. The company knows more than any single person in it can find.

This is the actual case for a second brain, and it is why the chatbot on top is the least interesting part. The value is not a machine that sounds smart. It is a machine that can find the answer your company already paid to produce, before you pay to produce it a second time.

A second brain is an architecture, not a product

You cannot buy a second brain. You build one, and it has four layers. I draw this on a whiteboard for every board that asks me where to start.

The canonical layer. The source of truth. The actual documents, decisions, contracts, and records, stored somewhere durable and never quietly rewritten by a machine. This is the floor. If it is wrong, everything above it is wrong faster.

The semantic layer. The meaning on top of the documents. What does "active customer" mean here? Which "Q3 plan" is the real one? Glossaries, tags, ownership, the relationships between things. This is where your company's actual vocabulary lives, and it is usually the layer nobody has ever written down.

The context layer. The rules that tell the AI how your company works. What to trust, what is stale, how to behave, what it is never allowed to touch. In practice this is a plain configuration file that a machine reads before it does anything. It is boring and it is load-bearing.

The engagement layer. The part people actually see. The chat box, the agent, the answer in the flow of work. Everyone wants to start here because it demos well. It is the top of the stack for a reason. It is worthless without the three layers holding it up.

Where the projects die

They die because someone bought the engagement layer and skipped the other three.

A chatbot with no canonical layer makes things up. A chatbot with no semantic layer does not know that your "churn" and finance's "churn" are two different numbers. A chatbot with no context layer will read the intern's abandoned draft with the same confidence it reads the signed contract. The demo hid all of this, because in a demo you hand it three clean documents and ask it three easy questions.

The company is not three clean documents. It is forty tools and a decade of mess.

What I am seeing in the room

The teams getting real value did the unglamorous work first. They picked the twenty or thirty documents that actually matter. They wrote down what the words mean. They decided who owns freshness. Only then did they put a chat box on top.

The teams that failed did it in the exact opposite order. They are the ones who tell me AI "does not really work for us." It works. They built the roof before the foundation.

What to do this week

1. List your canonical sources. Not all of them. The ten to thirty documents a new executive would need to actually run your business. If that list takes more than an hour, that is your finding.

2. Write one page of definitions. The ten terms your teams argue about. "Active user." "Closed deal." "Done." One agreed sentence each.

3. Name an owner for freshness. One person whose job is deciding when a document is out of date. Not a committee.

4. Do not buy anything yet. You are two layers away from needing a tool.

For now, the point is this. You are not missing knowledge. You are missing the architecture that lets a machine find it without embarrassing you. That part you have to build, and it starts with a list you could write today.

Next in the series I take the layers apart one at a time, starting with the one everyone underestimates: access control, and why retrieval is really a permission decision. After that, the boring ingestion plumbing that decides whether any of it is trustworthy, and why I keep telling companies to start with a folder of plain text files instead of a vector database.

The Corporate Second Brain, a six-part series

1. Your Company Already Has a Second Brain (this post)

2. RBAC Is the Hard Part of AI

3. Why Ingestion Is the Whole Ballgame

4. I Replaced the Vector Database with a Folder

5. From What We Know to What You Should Do

6. Build a Corporate Second Brain in a Weekend

Dictation has been built into your phone and your laptop for a decade, and almost nobody uses it, because it was never good enough to trust. You would speak a paragraph, then spend more time fixing the mess than you saved. So we all went back to typing and forgot the whole idea.

What changed is not the listening. It is the cleanup.

The built-in dictation was never the problem, and never the fix

Apple, Windows, and Google all ship voice typing. They are decent at turning sound into words and bad at everything after that. No real punctuation. No sense of what app you are in. Filler words and false starts land on the page exactly as you said them, the "um" and the "so, like" and the sentence you started over halfway through. You still go back and fix all of it by hand. That editing tax is why dictation never stuck.

The built-ins also behave differently in every app, so you never build a habit. It works one way in Notes, another in your browser, and not at all in half the tools you actually live in.

What Wispr Flow does differently

Wispr Flow sits on top of everything and cleans your speech into finished text as you talk. You hold a key, you talk like a normal person, and what lands on the page is punctuated, formatted, and stripped of the "ums" and the false starts. It works the same way in every app, your email, your code editor, a Slack message, a Google Doc, because it types into whatever field your cursor is already in.

A few things that make it stick for me:

• It edits while you ramble. You do not have to speak in clean sentences. You think out loud, and it hands you clean prose. That is the whole unlock. The editing tax is gone.

• It knows where you are. A quick Slack reply comes out casual. A long document comes out structured. It formats to the context instead of dumping a wall of text.

• It learns your words. Names, jargon, product names, the acronyms your industry beats to death. You teach it once and it stops mangling them.

• It works in over a hundred languages and syncs your setup across your Mac, your Windows machine, and your phone.

• On the paid tier you can edit by voice. Select this, rewrite that, without touching the keyboard.

There is also a real privacy story, including a HIPAA agreement you can sign inside the app, which matters if you or your clients care where the words go. I do.

The honest numbers

The company says four times faster than typing. Be careful with that number. It assumes a 40-word-a-minute typist, and if you write for a living you are quicker than that. I type somewhere around 80 to 110, and I dictate around 150 to 180 words a minute of clean output, so for me it is closer to twice as fast, not four. Still a real gain, and it depends on how quiet your room is and how much you proofread, which you still should.

But the speed is not even the main point. The main point is that talking is a lower-effort way to get a first draft out of your head than typing is. I am more willing to start the email, the doc, the reply, because starting costs less. Over a week that adds up to more than the raw words-per-minute math suggests.

When to talk and when to still type

I am not going to tell you to throw your keyboard away. Voice wins for first drafts, long messages, emails, notes, and thinking out loud. Typing still wins for code, for anything in a loud room, and for the last careful pass on something that has to be exact. Use the right one for the job. Most people are just leaving the voice half completely on the table.

If you read my last post on centaurs, this is a centaur tool. It makes you faster and you stay completely in charge. It is not doing your thinking. It is getting your thoughts onto the page at the speed you actually have them.

What to do this week

1. Try it on the free tier. It costs nothing to start, no card, and the free plan gives you enough words a week to feel whether it fits.

2. Use it for one thing first. Pick email, or Slack, or your notes. Do that one thing by voice for a few days before you judge it.

3. Teach it your words. Add the names and jargon it trips on. Ten minutes up front and it stops fighting you.

4. Watch the editing tax, or the lack of it. That is the real test. If you are shipping the text with barely a fix, it is working.

If you want to try it, here is my referral link, and it gets you a free month of the paid tier: wisprflow.ai. Full disclosure, that is a referral link, so I get credit if you sign up, and you get the free month either way. I am recommending it because I use it every single day, not because of the link.

Talk to your computer. The technology finally earned it.

A steak.

The two numbers, side by side

One six-ounce steak carries a water footprint of about 1,800 gallons. That is roughly 6,800 liters, for one piece of beef.

One AI-generated image uses about 0.0002 gallons. That is around 0.85 milliliters. Less than a teaspoon.

Do the division and a single steak uses on the order of nine million times more water than generating one AI image. Not double. Not a hundred times. Nine million.

Where 1,800 gallons goes

The steak number is not a trick. It is a lifecycle footprint, and here is what it is made of:

• Rainfall to grow feed and pasture: about 1,465 gallons. This is the water that falls on the fields that grow what the animal eats over its whole life.

• Irrigation from rivers and groundwater: about 246 gallons. Water pulled from the actual freshwater supply to grow crops.

• Drinking and farm operations: about 39 gallons. What the animal drinks and what the farm uses around it.

• Processing and transport: about 50 gallons. Slaughter, packaging, moving it to your plate, and diluting the pollution that comes with all of that.

Add it up and you get 1,800 gallons before it hits the grill.

The honest caveat, because I will not cheat the comparison

I care about being right more than I care about winning the argument, so here is the fine print.

Most of that steak number is green water, meaning rainfall that fell on grazing land anyway. That is not the same category as the treated blue water a data center pulls from a municipal supply to cool servers. A hydrologist will tell you that comparing green rainwater to blue cooling water is not perfectly apples to apples, and the hydrologist is right.

So account for it. Throw out the rainfall entirely. Count only the irrigation, drinking, and processing water for the steak, the parts that hit the real freshwater supply. That is still about 335 gallons of blue and gray water per steak, against 0.85 milliliters for the image.

That is still well over a million to one. The caveat shrinks the gap. It does not close it. It does not come close to closing it.

Why I am the one saying this

I do not eat steak. I read these exact water and land numbers years ago and they were part of why I changed how I eat. So I am not here defending beef. If anything I am the last person who would.

That is the point. A vegan is telling you that the water panic aimed at AI images is pointed at the wrong plate. If water is the thing that moves you, and it moves me, then the single highest-leverage water decision most people make in a day is the one with a fork, not the one with a keyboard.

You can generate an AI image every ten seconds for a year and not touch the water footprint of one steak dinner.

What to actually do about it

If you want to spend your water conscience well this week, here is the ranked list, biggest lever first.

1. Look at your plate before your prompt window. One or two plant-based meals a week moves more water than never touching an AI tool again.

2. Fix the leaks you can see. A running toilet or a dripping line at home outruns your entire year of AI usage in a couple of days.

3. Push the aggregate where it belongs. Data center water is a real engineering problem, and it is on the companies and the grid to keep improving cooling and energy mix. That pressure is worth applying. Per-image guilt is not.

The takeaway

Worrying about the water in an AI image while a steak sits on the plate is like worrying about a teaspoon while the bathtub overflows behind you.

I am not asking you to feel bad about dinner. I am asking you to keep the scale straight. When the concern is real, and water is real, put it where the water actually goes.

So the value moved. It's not in making the thing anymore. It's in knowing which thing to kill.

That's taste. And taste is the moat.

When output is cheap, judgment gets expensive

Every company I work with can produce more now. Ten times more. The bottleneck used to be production. Now the bottleneck is deciding what deserves to exist.

Most teams are drowning in on-brand-looking output. The colors are right. The font is right. The tone sounds close enough. And it's all forgettable.

Here's the trap. AI is very good at producing things that pass the surface check. It matches your style guide. It does not match your judgment, because your judgment lives in your head, not in a PDF.

The output I told a client to kill

A client had a campaign built almost entirely by a model. Landing page, ad set, email sequence. It looked correct. Brand colors, right voice, clean layout. The CMO loved it. They wanted to ship that week.

I told them to kill the whole thing.

Not because it was ugly. Because it said nothing. It was the average of every competitor in their space. A model trained on the internet will regress to the mean of the internet, and the mean is mush.

Shipping it would have cost them the one thing they actually had: a reason to be picked over the guy next to them. Saying no cost about two weeks and an awkward meeting. Saying yes would have cost the position.

Your brand is a guardrail, not a coat of paint

Most people treat brand as decoration. Colors, logo, a nice deck. That's the paint.

The real brand is a set of rules about what you refuse to do. What you won't say. What you won't sell. The tone you'll never take. That's the guardrail, and a guardrail only works when it stops you from doing something you could have done.

AI makes the paint free. Anyone can match your palette in a prompt. So the paint is worthless as a moat now. The guardrail is the whole game.

When you can generate a thousand versions, the person who knows which one is right, and can defend why, wins. That person is rare. That's the moat.

Taste doesn't scale, and that's the point

People keep asking me how to systematize taste. Put it in a prompt. Make the model do it.

You can't, fully. You can feed a model your best work, your no's, your voice. It helps. Claude and ChatGPT get closer when you show them what you've killed and why. But the final call is a human who has seen enough to feel when something is off.

That feeling is built from twenty years of being wrong and learning why. It doesn't fit in a context window.

What to do this week

Pick one piece of AI-generated work sitting in your pipeline. Something that looks fine.

Ask one question: if a competitor shipped this exact thing, would anyone notice the difference? If the answer is no, kill it or rebuild it.

Then write your kill list. Ten things your brand will never do. Give it to whoever is running your prompts. Now your generation has a guardrail instead of a spray can.

Close

The companies that win the next few years won't be the ones who generate the most. Everyone can do that. They'll be the ones who throw the most away, on purpose, for a reason they can name. Get good at no.