The Barbershop Is Safer Than Your Job
I lost one of my company's competitive advantages the week ChatGPT launched. I got it fast. Now I often asked about how to compete as a startup or as a person in the AI era. Here is the answer.
For most of the last decade the thing we were genuinely best at was writing to you about your own body.
Not the app. Not the charts. The sentences. Taking a pile of heart rate variability and sleep data and whatever you did last Tuesday, and producing a paragraph that was scientifically defensible, specific to you, and actually worth reading. That was the moat. It cost years and an uncomfortable amount of money.
Then the models arrived and I watched it evaporate.
Not really. Our paragraphs are still better, and I can show you why. The problem is that nobody can tell. A user cannot distinguish a rigorous interpretation of their data from one that merely reads like one. Mostly, neither can an investor. And a moat nobody can perceive has stopped performing the function of a moat, whatever it’s doing in the source code.
So we knew fairly early that we’d have to change, and change substantially. What I want to write about is where that thinking ended up, because it started as a question about my company and turned into a question about everyone’s career, including yours.
What we didn’t do
Everyone I knew started building wrappers.
We didn’t, and I’ve written a whole essay about why, so I’ll keep this short: there’s no defensibility in it. You build the thing, you raise easy money from people who aren’t looking closely, and then you die. Done cleverly it’s a decent way to make quick money, and I want to be fair about that — some people are running that play deliberately and they know exactly what they’re doing. It just isn’t a company.
Then the mood shifted and everyone decided SaaS was finished. I half agree. What’s finished is SaaS whose advantage was the interface: the buttons, the arrangement of the screens, some simple automation between two APIs. That advantage has genuinely evaporated, and I don’t think it’s coming back.
But that isn’t all software, and the distinction matters more than the doom does.
Investors used to have a standard test: what happens if someone just hires a hundred engineers and builds the same thing? If the honest answer was “they’d have it in six months,” you didn’t have a business. What changed is that the same test now runs faster and costs less. That’s all. It’s the same question with a new price.
And the interesting part was always the cases where the answer was no. You can hire any number of engineers and you will not reproduce Telegram’s speed and reliability. That’s not a feature list. It’s a decade of specific hard decisions compounding, and it doesn’t come out of a model, because it never came out of typing.
Which gives you the rule, and it’s almost stupidly simple. In a world with AI, don’t do what AI does. Do what it doesn’t.
Fine. So what doesn’t it do?
The safest person I know is a barber
I’ve been turning this over for a year and I keep landing in the same place, which is not where I expected.
Barbershops. Hairdressers. They’re in less danger than almost anyone reading this.
I’m not being cute. Look at what a haircut actually is. Someone puts their hands on your head for forty minutes. There’s conversation, most of it not important, some of it more honest than the conversations you have with people who matter. There’s a person who remembers what you asked for last time and what you complained about after. There’s physical contact, which does measurable things to your nervous system whether you consented to that or not. You leave slightly different than you arrived, and not only because of the hair.
None of that downloads.
The same is true of a good coffee place — and I mean the small kind, where the barista knows you and knows what you like. Not the American high-throughput version with a queue out the door and your name misspelled on a cup. That one is a logistics operation and it will be automated. I mean the ten-table place where the relationship is the product and the coffee is the excuse.
Now hold that next to a mid-level programmer who receives well-specified tickets and closes them.
I know which of those two jobs I’d rather be holding in five years, and it isn’t the one that pays better today.
Three things nobody can take
In technology, once you set hardware aside, I think everything defensible sits in one of three places. I’ve made this argument before at company level, so here it is compressed:
Relationships between actual humans. Whatever you think about what people feel toward AI, a living person still does something to another living person that nothing else does.
Datasets that are hard to acquire and attached to something worth knowing. Hard to acquire is the operative half. Anything scrapable is already gone.
Knowledge that isn’t public and can’t be reverse-engineered out of your product by looking at it.
What you process, how you process it, and what you build with the people using it. That’s the whole list.
But here’s the thing I only noticed recently, and it’s the reason I’m writing this instead of another company post.
The list transfers to people almost unchanged.
Your career, as a company
Swap the words and it comes out as four things:
The breadth of your context and how much of your domain you actually hold in your head.
Your own view of that domain — not what the field thinks, what you think, and whether you’ve earned the right to think it.
Relationships worth investing in, which means the ones where someone would recommend you without being asked.
And whether you can get real leverage out of AI rather than pretending you don’t need to.
The mental model I’ve found most useful, and I’d suggest trying it even if you’re happily employed: think of yourself as a one-person consultancy with your name on the door.
What does it specialise in? What does it know that the others don’t? What has it actually shipped, and what went wrong, and what did it learn that isn’t in any book? What’s the house view — the thing it will argue for that not everyone agrees with? And the question that decides the whole thing: would a client hand you to their friend without being prompted?
If those questions have thin answers, that’s the work. Not a certification. Not another framework. Those.
And while we’re here: classical education has almost entirely lost the part of its purpose that consisted of loading information into you. Information is not an advantage when it exists in books and on the internet. What survives is trained reasoning — which is why mathematics is worth more now than the ability to discuss which decade produced which theory. One of those is a capability. The other was always just retrieval, and retrieval got cheap.
Context is the only ground you can hold
Here’s the part almost nobody wants to hear, and I say it as someone who has watched very good people refuse it.
A programmer who has spent years executing tasks handed down from above is going obsolete. Not because they’re bad at the job. Because the job was the narrow bit, and the narrow bit is what got automated first.
To stay relevant they’ll have to go wide. Architecture, or product, or analysis, or something — but they have to grow a band of territory that lets them understand a wider context than the ticket they were given. Context is the only ground where we can confidently compete, because it’s the one thing the model doesn’t have. It has everything ever written. It doesn’t have your customer’s political situation, or the reason the last attempt failed, or which of your colleagues will quietly kill the project.
But going wide means leaving the thing you’re good at and being mediocre at something new, in public, at forty. Which is genuinely hard, and I understand why people don’t, and I think the ones who don’t are making a serious mistake anyway.
There’s one more thing on the human side of the line, and it took me a while to see it.
Accountability. AI cannot be responsible for an outcome. It will produce something confidently wrong, and when it goes badly it says the equivalent of oops — and there is nobody there. Nobody whose reputation moves. Nobody who has to sit in the room afterwards. Nobody with anything at stake.
Which means responsibility for the actual result stays with humans, and if you want to know where the value concentrated, it concentrated there. Not in producing the work. In standing behind it.
The arithmetic moved
There’s no longer a choice about whether to adapt to this. There’s adapting, and there’s being beside the point. I’d love to offer a third option and I don’t have one.
But I’ll finish with the thing that genuinely surprised me while I was working this out.
If you have ever quietly wanted to tend bar, or cut hair, or open a bakery — I’m not about to tell you to follow your dream. I don’t do that. I’m telling you the arithmetic changed underneath you while you weren’t looking, and that work is now more defensible than a clerk’s, and more defensible than a great deal of programming.
That isn’t a consolation prize. It’s a repricing, and I don’t think most people have noticed it yet.
Interesting times, and I mean that in every available sense.