I am not going to tell you what to think about the markets. I do not know, and neither, it turns out, did the people whose entire job it was.

Here is the part that stayed with me.

Aschenbrenner was not a fool who got lucky. He wrote the essay that gave the moment its name, in public, before most people were paying attention. He argued it clearly and then put real money behind a conviction that has, in the large, held up. The technology is real. He said so early, and loudly. He called the fund Situational Awareness, after the essay that made his name.

It did not save him.

That is not a story about AI. It is the oldest story I know, and I watch a smaller version of it play out on a floor most nights.

Knowing and doing are different jobs

There is a difference between knowing a thing is great and knowing what to do about it. The gap between those two is where most of the money, and most of the dignity, in my trade gets lost.

I can hand you a wine that is by every measure extraordinary. Correct provenance, singular site, a maker who does everything right. Identifying that bottle is table stakes. Any competent somm can do it, and a machine can now do it faster than I can.

Knowing that this extraordinary bottle is exactly wrong for the table in front of me, for the food they ordered and the night they are having and the number on the right side of the list they are quietly hoping I respect, that is the job. It is the only part that was ever hard. It is the only part anyone should pay for.

I have gotten it wrong. I have poured something objectively brilliant for a table that needed something honest and cheap, and watched a good night go slightly cold because I was busy admiring my own conviction instead of reading the room. The wine was right. I was wrong. Those are not the same sentence, and learning the difference cost me more than any exam ever did.

Everyone is right about the technology now

That is the cheap part. The AI is real, it is powerful, it is not the empty bubble the loudest skeptics keep promising. Fine. Agreed. You and I and the fund that just came apart all agree.

None of that tells you what to do on Tuesday.

It does not tell you how much to bet, or when, or with how much borrowed against it. It does not tell you which of your restaurant’s problems is worth automating and which one will quietly break the moment you touch it. It does not tell you whether the workflow that saved another operator a fortune will save you a dollar, because your business is different in ways that never make it into the pitch.

One widely cited study found that roughly ninety five percent of enterprise generative AI pilots showed no measurable impact on profit and loss. Almost none failed because someone was wrong about AI. They failed because someone was right about AI and mistook that for being right about their own situation.

I watch operators do the smaller version of this constantly. A sharp one reads that AI is real, which it is, and concludes he should therefore put it everywhere in his restaurant, which does not follow at all. He is right about the technology and wrong about his own floor, and the tool ends up automating the one warm thing his regulars actually came back for. Right about AI. Wrong about the room. The bill arrives either way.

Conviction is not judgment

The smartest bull in the room was right about the destination and undone by the path. Maybe the timing, maybe the borrowing, maybe both. It does not matter which, because any one of them is fatal on its own, and being correct about the big picture does not buy back a single one.

This is the thing I keep saying and the world keeps insisting on proving. Conviction is loud and cheap and everywhere right now. Judgment is quiet. Judgment is knowing which grape to leave alone, which table wants which bottle, which of your problems you must never automate no matter how real the technology gets.

The technology was never the hard part.

I did not need a language model to learn that. I needed a floor, and a few wrong pours in front of people I badly wanted to impress.

Read the canonical edition at woodywusommelier.com.

That much I am confident about, because I have watched it for twenty years and lived it for most of them. What I am no longer confident about is any specific number attached to it.

Here is a thing that happens if you read about restaurant technology. Someone tells you the average general manager loses eighteen to twenty one hours a week to administrative work. Someone else says fifteen to twenty of those go to inventory alone. A vendor tells you their tool returns a third of a manager’s week. The numbers are specific, they are confident, and they have decimal points.

I used to quote figures like these. I stopped. This essay is about why, and about what I think you should ask for instead.

Where the numbers come from

Almost every labour statistic in restaurant technology is one of three things.

It is a survey, in which busy people estimated their own hours from memory, which is a genre of fiction. It is a vendor’s own study, which is marketing wearing a lab coat. Or it is a real number from a real business that has been generalised into an industry average, which quietly assumes your restaurant is like the one that produced it.

None of those are lies exactly. They are just not measurements of your building.

Here is the part that should bother you more. The manager whose hours are being counted is the same person being asked to estimate them, and nobody in hospitality has ever accurately reported how long anything took. Ask a chef how long prep ran. Ask a GM how much of Sunday went to the schedule. The honest answer is always “more than I want to say.”

The number that actually matters

Around ninety five percent of AI projects produce no measurable return. That figure I do trust, because it keeps being reproduced by people with no product to sell, and because it matches everything I have seen.

But sit with what it implies. If almost all of these projects show no measurable return, then almost all of the confident hours-saved numbers in the market are describing projects that, measured properly, did nothing.

The failure is rarely the model. It is that someone automated the wrong thing, and then measured the wrong thing to prove it worked.

What I built instead

This year I built an AI system and put it into a working service context, and the part that took longest was not the intelligence. It was deciding what I would allow myself to claim about it.

So the measurement layer I built reports three things.

First, whether it held the standard. Not usage. A thousand conversations that resolved nothing is a cost, not a saving, and usage dashboards are the most common way a tool proves nothing while looking busy.

Second, what it plausibly saved, expressed in one specific form: inquiries resolved without staff time, multiplied by a handling time the operator measured themselves. Not an industry benchmark. Their number, from their building, which means they can argue with it. If they think three and a half minutes is wrong, we change it in front of them and the model still stands. A number a client can argue with is a number a client can believe.

Third, and this is the section I care most about, what it refuses to claim.

It does not claim it lifted revenue. Without a control period you cannot prove a tool caused that, and guest mix, season, menu and weather all move the same line. It does not claim it saved covers. It counts what it can stand behind, says so in writing, and leaves the rest alone.

I want to be precise here, since the whole essay is about precision. That system is a demonstration running against a fictional venue. It has no guests, so it has no results yet. What I can tell you about is the method, not the outcome, and I would rather hand you a method you can use than a number I would have to invent.

Desk work and floor work

None of this means the opportunity is not real. It is. But the honest split was never smart jobs against dumb jobs. It is desk against floor.

Everything that happens at a desk is genuinely in play. Invoicing. Inventory reconciliation. Pulling guest history together before service. First draft schedules. Review responses. The allergen sheet. The marketing copy that eats a Sunday. That work is real, it is expensive, and almost none of it is why anyone got into this business.

And then there is the floor, where none of it works.

The read of a table at eight o’clock. The recovery that turns a ruined night into a regular. The decision to comp, to push, to slow a table down, to leave someone alone. The two in the morning conversation with a cook whose life is coming apart. None of that automates, because none of it is information. It is presence, and presence is the job. A machine can draft the schedule. It cannot stand in a dining room and feel the weather change.

The re-concentration

So AI does not shrink the restaurant job. It re-concentrates it.

The desk work compresses toward zero and what remains is the part that was always the actual profession. The operators who should be nervous are the ones whose week was secretly all sediment. The ones who should be glad are the ones who came here for the room and have been drowning in paperwork ever since.

The reclaimed hours are not a cost saving to quietly pocket, either. That is the mistake I would most want to talk an operator out of. Take the hours back and spend them where a machine will never go: the pre shift, the new server who needs coaching, the recovery, hiring well instead of fast, standing in the room and reading it. That is the return. Not a smaller team. A team finally pointed at the work worth doing.

What to ask for

If someone is selling you AI for a restaurant this year, you now have a sharper question than “how good is your model.”

Ask what handling time they are multiplying by, and where that number came from. If it is an industry benchmark rather than something measured in your building, it is a guess with a decimal point.

Then ask what their reporting refuses to claim. Any honest measurement has a section on what it cannot prove. A vendor who shows you only numbers that go up is not measuring. They are marketing, and the difference will cost you a year.

The machines are not coming for your floor. They are coming for the paperwork that has been keeping you off it.

Give your managers their hours back. Then send them to the room, and count it honestly.

Woody

Operators: how many hours a week do you think you lose to admin, and have you ever actually measured it rather than estimated? Reply and tell me. I read everything.

Read the permanent edition on woodywusommelier.com.

I thought about that bottle last week, because the most expensive thing in artificial intelligence just became free.

On July 16 a Chinese lab called Moonshot released a model named Kimi K3. Two point eight trillion parameters, the largest open weight model anyone has put into the world, with the full weights promised by the end of this month. Open weight means you can download it. Not rent it, not call it through somebody’s meter. Download it, run it on your own hardware, and owe no one anything.

The instinct is to file this under geopolitics and move on. Resist that instinct for a minute, because there is something in it for anyone who runs a room, sells a bottle, or has ever been told a machine is coming for their job.

Third in the world, and free

Let me be precise, because precision is the whole point of what follows.

On GDPval-AA v2, a benchmark that scores real work across forty four occupations and nine industries, K3 scored 1,687. That is third in the world. Ahead of it sit Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8. Just behind it sits Claude Opus 4.8. On frontend code, in blind developer testing, K3 placed first, ahead of Fable 5. It leads the field on sustained coding work.

I am not going to tell you it is the best model in the world. It is not. The two above it are genuinely above it. What I will tell you is that it is close enough to the top that the gap has stopped mattering to anyone actually building something, and it is the only one of the three you can own outright.

Read that back. The frontier is now a place you can move into for the price of the hardware. A year ago the most capable intelligence on the planet was a subscription. Today something within arm’s reach of it is a download, and it came from a lab most of my guests have never heard of, in a country the industry spent three years insisting was a step behind.

The step-behind story is over. What replaces it is more interesting.

The cellar, and the person in front of it

Here is the frame I cannot stop seeing it through.

A great wine cellar used to be a moat. If you had the bottles, the allocations, the depth nobody else could assemble, you had something a competitor across the street simply could not match. The list was the edge. Spend enough, wait long enough, know the right people, and you owned an advantage that was made of scarcity.

That has been quietly ending in wine for a decade, and it just happened to AI in a single week. When the rare thing becomes something anyone can download, the rare thing stops being the edge. The cellar went open source.

So what is left.

What is left is the only thing that was ever actually scarce, which is the person standing in front of the cellar. The one who knows which of those thousands of bottles belongs on this table, tonight, in front of these people, with this food, at this moment in their evening. The bottles were never the hard part. Choosing was the hard part. Reading the table was the hard part. The cellar was just the raw material, and raw material has a price, and that price is now falling toward zero.

This is the thing I keep arguing, most recently in Palantir Copied the Dining Room, and the industry keeps proving it for me. Intelligence got cheap. You can hold two point eight trillion parameters of it on a machine in your back office. Judgment did not get cheap, and judgment is the entire job.

Why it matters that it was not built in English

There is a second thing in the K3 story that most of the coverage walked straight past, and it is the part a hospitality person should sit with longest.

The model was built by people thinking in Chinese. That is not a footnote. For three years the working assumption of this industry has been that the center of gravity is one valley in California and one language, and everyone else is downstream of it, translating, catching up, importing the frontier. A top-three open model out of Beijing breaks that assumption in public.

It should feel familiar, because it is the oldest story in wine. For a century the assumption was that the center of the world was a few hundred square kilometers of France, and everything else was an imitation of it or a discount from it. Then the imitations got very good, and then they stopped being imitations, and now the most interesting bottles on my list come from places the old map did not have names for. Growers, not houses. Countries the canon ignored. The center did not hold, because the center was never a place. It was a temporary agreement about where to look.

AI is having its version of that week right now, and it happened faster than wine’s version because software copies for free. The lesson is the same one a good sommelier learns early and painfully: the moment you believe the important thing can only come from one place, in one language, you have stopped being able to see it arrive from anywhere else. The people who miss the next great thing are almost never the people who could not afford it. They are the people who were certain they already knew where it would come from.

What an operator should actually do about this

I try never to write one of these without a sentence a working person can use on Monday, so here it is.

Stop paying for the model. Start paying for the deployment.

The intelligence is about to be the cheapest input in the entire stack. Within a year, paying a premium for access to a frontier model will look the way paying a premium for access to the internet looks now, which is to say it will look like something people used to do before it became plumbing. The bottle is going to nine dollars. It is not where the value is.

The value is where it has always been in my trade. It is in the person who can look at your specific business, find the one thing worth automating and the three things you must never touch, deploy it on the systems you already run, and prove in your own numbers whether it earned. That person is not selling you intelligence. Intelligence is free now, or it is about to be. That person is selling you the read, and the read is the only thing in this entire story that did not just get cheaper.

The most capable open model on earth is now a download, built by people who were supposed to be a step behind, proving that the frontier is not a place and never was. That is a story about China, and about open source, and about a benchmark leaderboard.

It is also, if you have spent twenty years in front of a cellar, the most familiar story you have ever heard. The rare thing became common. The judgment stayed rare. It always does, and knowing that in advance is the closest thing to an edge that any of us get.

Here is the part almost nobody mentions. When Palantir invented the role, they modeled it on a restaurant.

Not as a loose metaphor. As the founding image of the whole role. The story inside Palantir is that their CEO watched the way French waiters work, the way the floor is fused to the kitchen, the way a great waiter will tell you no when you order the wrong wine, and built the engineering practice on it. The engineer sits in the client’s office, learns the business the way a server learns a table, and earns the standing to say no to a bad request. They called the first ones Deltas, and until around 2016 they had more of them than they had ordinary software engineers.

I have spent twenty years in the room that job was copied from. So let me tell you what the AI industry just rediscovered, and why it decides who actually gets their money’s worth out of all this.

Intelligence got cheap. Judgment did not.

Start with what changed. For most of computing history, the intelligence was the scarce thing. You paid for the smart system.

That era is over. Every company on earth now has access to roughly the same frontier models. The intelligence is a commodity, priced by the token, available to your competitor at the same rate as to you. When everyone has the same brain for rent, owning the brain stops being an advantage.

So the value moved. It moved to deployment, which is a polite word for a much harder thing: knowing where the intelligence should and should not be pointed inside a real business, and shipping something that actually works there. The numbers are brutal about how rare that skill is. One widely cited MIT study found that 95 percent of enterprise generative AI pilots produced no measurable financial impact. And they almost never fail because the model was not smart enough. They fail because someone automated the wrong thing. A problem that did not matter. Or one that needed a human and never should have been touched.

Picking the right problem is the whole game. And picking the right problem is not an engineering skill.

What the dining room actually teaches

Here is why Palantir looked at a restaurant, and here is what a restaurant taught me before I ever wrote a line of code.

A great server is not taking orders. A great server is running a live diagnostic on a table they met four minutes ago. Is this a celebration or a negotiation. Who is paying, and do they need to look generous, or careful. Is the quiet one bored or just tired. Should I move faster, slow down, disappear, or become the whole evening. None of that is on the menu. All of it decides whether the night works.

And the best ones do the hardest thing of all: they tell you no. No, not that bottle with that dish. No, you do not want the tasting menu tonight, you are too hungry and too rushed, let me feed you properly instead. Saying no to a paying guest, in a way that makes them trust you more, is the highest skill on the floor. It is also, it turns out, the highest skill in deploying AI. The consultant who cannot tell a client “you should not automate this” is the one who sells them the 95 percent failure.

That is what the dining room teaches that a computer science degree does not. How to sit with someone, read what they actually need under what they are asking for, and have the standing to redirect them. Palantir could not find that skill in engineers, so they built a role around people who had it and taught them to ship. The AI labs are doing the same thing right now, at enormous cost, because the skill is rare and they know it is the bottleneck.

I did not have to learn that half. I have been doing it every night for twenty years.

The sommelier is the deployment

There is a sharper version of this, and the people building the AI stack keep reaching for it without quite knowing why. When they try to explain what a Forward Deployed Engineer actually does, they land, again and again, on the sommelier.

Think about why. A sommelier does not have one recommendation. A sommelier has a cellar, and a table, and the entire job is the match between them. You read the guest: what they know, what they can spend, what they are eating, whether tonight is a celebration or a Tuesday, whether they want to be led or left alone. Then you find the one bottle, out of hundreds, that fits this person, this plate, this night. The bottle that is perfect for table nine is wrong for table ten. There is no generic pairing. There is only the read, and the match, performed fresh every single time, by someone who does it on the floor, night after night, with real people and real stakes.

That is deployment. That is the whole thing.

Because there is no generic AI deployment either. The workflow worth automating at one restaurant is the one you must never touch at the next. The tool that saves a hotel group two hundred hours is useless to the room down the street with a different system and a different bottleneck. The frontier model is the cellar: vast, powerful, available to everyone at the same price. The value is not the cellar. The value is the person who can read this specific business and make the one match that fits it. A pairing, not a product. Done case by case, by someone who has spent their life learning to read a room.

The AI industry keeps using my job to describe its most valuable one. It is not an analogy. It is the same skill, pointed at a different table.

The two halves, and why almost nobody has both

The Forward Deployed Engineer works because it fuses two things that rarely live in one person.

One half is the build: you can ship a production system, against real data, on the systems a business already runs, with the guardrails and the human approval that make people trust it. I wrote a whole essay arguing that managing AI agents is just managing staff, because I had been doing both. That half I earned at a terminal, after service, building the tools I wished the industry had.

The other half is the read: you can walk into a working business, feel where the money is quietly leaking, know which of its problems is worth solving and which one to leave alone, and say so with enough authority that they believe you. That half I earned on the floor.

Most people chasing this role have the first half and are learning the second the hard way, one failed pilot at a time. A smaller number have the second and cannot ship. The people who have both are the ones getting paid a million dollars, and there are not many of them, and almost none of them come from hospitality, which is strange, because hospitality is where the model came from.

What this means for anyone trying to buy AI

If you run a restaurant, a hotel, a wine business, a group, and you are being sold AI right now, here is the test. The person selling it to you: have they ever worked your floor? Because if they have not, they cannot tell which of your problems is worth solving. They will sell you all of them, because selling you all of them is their business model, and most of it will fail, and it will not be the model’s fault.

The right person does the opposite. They sit in your business first. They find the one workflow that is quietly costing you hours or covers or comps, they build the thing that fixes it on the systems you already have, and then they prove it saved you money in your own numbers. And they tell you, plainly, which of your problems to leave alone. That last part is how you know they are worth hiring. Anyone who never says no is selling.

Where I stand

I am a sommelier and a hospitality operator who learned to build. I spent two decades in the room the most valuable job in AI was copied from, and I spent my nights learning to ship the tools. I am not an AI consultant. That phrase means nothing now; everyone is one. I am the person you commission to find the one thing worth automating, deploy it, and prove it earned.

The industry spent a fortune rediscovering that the dining room had the answer all along. Some of us never left it.

If your room, your list, or your group is trying to figure out where AI actually pays, that is the conversation I am built for. I will even tell you where it does not.

— Woody

They call it “agent design.” They write papers about it. They’ve coined a vocabulary (orchestration, delegation, human in the loop, evaluation, guardrails) and they present it as a new discipline, because for them it is. They’ve never had to run a Saturday service with a brand new server, a hungover line cook, and a sommelier who’s brilliant with Burgundy and a hazard with bills.

I run AI agents at night and a restaurant floor at night too, technically. And I’m here to report that they are the same job. If you have ever managed staff in a high pressure service environment, you already hold an intuition for working with AI that most engineers are still acquiring the hard way. Here’s the translation table.

1. The briefing is the product

A new server doesn’t fail because they’re incapable. They fail because the briefing was bad. “Take care of table twelve” produces chaos; “table twelve is a regular, anniversary, he’s allergic to shellfish, she chooses the wine, don’t rush the dessert” produces magic. Same human, different context.

Working with AI is briefing, full stop. The people getting mediocre results from these tools are writing “take care of table twelve” prompts and concluding the staff is bad. The people getting astonishing results are writing the second kind: role, context, constraints, examples of done right, what to do when uncertain. In the industry we’d call that a pre shift meeting and an SOP. The tech industry calls it prompt engineering and pays handsomely for it. Operators have been doing it twice a day, unpaid, forever.

2. Delegate the task, never the judgment

Every manager learns this through scar tissue: you can hand off the doing but not the deciding. Not at first. The new hire runs food before they take orders; takes orders before they handle complaints; handles complaints for months before they’re allowed near a comp. Trust is granted in layers, each layer earned by verified performance at the last one.

This is exactly, exactly, the correct posture with AI agents, and it’s where smart people fail in both directions. The doomers won’t let the new hire carry a plate (”it might drop it!”) and get no leverage. The reckless promote it to manager on day one, let it email clients unsupervised, and learn in public. The operator’s instinct is the right one: tight scope, verify the output, expand scope on evidence.

I learned this the hard way with the first real tool I built: a pipeline that reads photos of the bottles I’m bringing in and drafts the tasting captions for each one, in English and Mandarin, at volume. The first batch came back like a new server reciting a script he doesn’t believe. Fast, fluent, generic. Correct notes with no soul: red fruit, a touch of spice, a smooth finish. Copy that could describe four hundred wines and sell none of them.

The fix wasn’t a smarter model. It was a better pre shift. I narrowed the brief: lead with the grower and the place, name the family that farmed it, one true sensory detail, no marketing words, no em dashes. I cast the tool to the one thing it was good at and kept the rest for myself.

It got me to maybe seventy percent. Useful, not finished. I still read every caption, and every line of Mandarin gets checked by a native speaker before a single one goes out. The machine drafts the room; it does not open the doors. It earned a lane, not a promotion. That’s the whole posture: my agents earn their autonomy the way my staff do, gradually, with me checking the work until the error rate tells me I can stop.

3. Inspect what you expect

No chef trusts a station they haven’t tasted from. The plate gets looked at before it leaves the pass, every plate, every night, no matter how senior the cook. Not because the cook is suspect, but because the standard is the standard and entropy never sleeps.

AI work needs a pass. The single biggest production mistake people make with these tools is shipping unread output. It’s the equivalent of food going to the table straight off the stove with nobody’s eyes on it. My rule at the terminal is my rule at the kitchen window: everything gets tasted. The agent drafts; I plate. The moment you stop checking is the moment something goes out cold, and with AI as with food, the guest remembers the one bad plate, not the hundred good ones.

4. Staff have shapes, work with the grain

A great manager doesn’t fight a server’s nature; they cast them. The charming one gets the celebrating tables, the precise one gets the eight top with the complicated bill, the wine mad one gets the collectors. Skill isn’t generic. Deployment is half the craft.

Models have shapes too. Some are meticulous and slow; some fast and a little sloppy; some brilliant at structure and stiff at warmth. The operator’s move (cast by strengths, cover the weaknesses with process) translates directly. I don’t ask my fastest tool for nuance or my most careful one for speed, the same way I don’t put my most poetic somm on the table that just wants the check.

5. Escalation is a designed path, not a panic

The best service systems make one thing crystal clear to every staff member: when you’re out of your depth, here is exactly who you bring it to, and you will never be punished for bringing it. Ambiguity about escalation is how small problems become Yelp reviews.

Agents need the same explicit lane: here’s what you decide, here’s what you flag, here’s the line you never cross alone. The tech world calls this human in the loop and acts like it’s novel. It’s the oldest rule on the floor: comp the dessert yourself, but a furious guest gets the manager, every time, no heroics.

6. The manager’s hours change shape, not size

Here’s the part nobody on either side tells you. Hiring great staff doesn’t reduce a GM’s work; it transforms it, from doing the tasks to designing the system: hiring, briefing, checking, coaching, improving the SOP. The leverage is real and so is the new job.

AI is identical. The fantasy is “the agent does my work while I sleep.” The reality, the productive reality, is that your work moves up a level: from producing to specifying, from doing to reviewing, from task to system. You become the GM of a staff that costs pennies, never tires, and is only ever as good as your brief. Which means the limiting factor in the AI age isn’t the model.

It’s whether you ever learned to manage.

The punchline

And there it is. The inversion I keep circling in these essays: the floor and the terminal trading places when no one was watching.

For decades, the deal was clear: tech skills were the leverage, and hospitality was the fallback, the thing you did while you figured out your real career. Now look at the actual skill stack the AI age rewards: briefing under ambiguity, layered trust, relentless quality inspection, casting by strength, designed escalation, systems management under pressure. That’s not a computer science curriculum.

That’s a GM’s Tuesday.

The operators don’t need to learn how to think about AI. They need to notice they already know, and then walk over to the terminal and start hiring.

Woody

If you run a floor, a kitchen, a hotel, a crew, or a company: what management rule have you tried on a machine? Reply and tell me. The best answers become a follow up post, with your permission and credit. I read everything.