## The descent is not the opportunity

There is a story everyone is telling right now, and it is basically correct, so here it is fast. The cost of intelligence is collapsing. It rides two curves stacked on top of each other, cheaper silicon underneath and more efficient models on top, and it ends where the smartphone ended, in nearly everyone's hands. A capability that cost a fortune eighteen months ago is cheap today and close to free eighteen months from now. Assume all of it. I do.

That is the boring half of the story, and it is the half everyone keeps writing. The descent happens to you and to every competitor at the same time, on the same schedule, so on its own it hands no one an advantage. The curve is not the interesting thing. The person standing at the bottom of it is, and what that person builds while it falls.

Every time a machine gets better at something people were paid for, the same fear arrives on cue. If a machine supplies expert judgment for almost nothing, what is left for the expert to do. The fear hides an assumption, that there is a fixed amount of work to go around. There never has been. When something valuable gets radically cheaper we don't use less of it, we use wildly more, and we invent uses for it that were unthinkable when it was scarce. Real people in real roles will be dislocated, and pretending otherwise is a lie. But the work does not vanish. It moves.

## Judgment changed its address

When expertise was scarce, the bottleneck was reaching it. You needed the credential, the firm, the salary, the appointment. The hard part was getting to the person who knew.

When expertise is abundant, the bottleneck flips. Now the hard part is knowing what to do with it. Which question is actually worth asking. Which answer to trust. Which problem deserves solving at all. And the one no machine will take off your hands, who owns the consequence when the call turns out to be wrong.

This is the move up the stack. You go from producing the analysis to deciding what the analysis is for. A machine that can reason does not remove the need for a human to own the outcome. It multiplies that need, because now far more decisions can be made by far more people than ever before, and every one of them still needs someone to stand behind it. If you have ever managed anyone, you already know the shape. The junior wants the answer. The senior owns the question. Cheap intelligence does not erase that difference. It makes it the only difference that matters.

## The commodity trap

Here is where most companies are about to hurt themselves, and they will not see it coming, because it will feel like progress.

If the only thing you do with cheap intelligence is consume it the way your competitors also consume it, generic, off the shelf, identical, then you have not built an edge. You've erased one. You took the thing that used to make you different and swapped it for the exact commodity everyone else is buying from the exact same vendor.

The best companies never won on generic capability. Everyone had the same software, the same consultants, the same playbooks. They won on something proprietary, a way of doing things that was theirs alone. The distributor who read their own market better than anyone. The insurer with a feel for their own risk that no one could copy. That edge is the real asset.

## The edge was trapped in a head

For all of history the Edge lived in a frustrating place. In the heads of experienced people, in habits no one ever wrote down. Picture the buyer at a €30M distributor who knows, without a model in front of her, when to fill the warehouse with steel and when to hold, because she has watched that price move for twenty years and has a feel for it she has never once put into words. When she retires, the feel leaves with her. It could never be fully captured, because capturing it took engineering, and engineering was scarce and expensive. So every business ran on a handful of irreplaceable people and prayed they did not leave.

Now the constraint is dissolving. When the cost of building intelligence into working systems collapses, you can finally take the thing that actually makes you special and build it into the systems that run you. Not a sliver of it, which is all anyone could afford before. The whole of it. The judgment, the sequence, the exceptions, the move your best person makes on instinct and has never been able to explain to a new hire. This is the difference between renting a brain and building one, which [another essay](/essays/rent-models-own-judgement) is about in full. The short version here. The business that pours its proprietary edge into systems it owns gets deeper every single day. The one that pipes generic intelligence into generic workflows becomes interchangeable with every other business doing the same, quietly, so it feels efficient right up until the moment there is nothing left that is yours.

## Own the means of production

There is a threat in this that the essays about cheap intelligence quietly skip. The entity that owns the model layer has every incentive to move up it. Watch where value is being created on top of your model, then walk in and take it directly. It's not a conspiracy. It's the most rational thing a model company can do.

You don't have to imagine it. In April 2026 Anthropic launched Claude Design, a tool that competes head on with Figma, weeks after asking Figma to be a partner in the launch announcement. Figma's founder said Anthropic had not been consistently honest with them, and Anthropic's chief product officer left Figma's board three days before Claude Design shipped. Figma's stock fell around seven percent. The Information reported the whole thing. And it fits a pattern. Claude Design, Claude Code, and a steadily widening set of Claude-branded products, each one a category that other companies had been building on top of the model. Dominate the model layer, then use that position to walk up into the lucrative verticals above it.

The labs will also tell you that open models powerful enough to compete with them are dangerous. Dangerous to whom. Not to the business that wants to keep control of its own data and workflows. Dangerous to a business model that depends on customers having few real alternatives at the model layer.

So here is what real safety looks like for a business, and it is not abstract alignment research or a certification from some government-run department of motor vehicles for AI. It is control. Over your compute, your models, your data, and your alpha. It is knowing that the means of production are yours, and that the proprietary knowledge you feed into your systems is not quietly becoming someone else's next product. The safe position is not trusting that a lab's roadmap will never include your business. It is keeping the ability to choose, at the model layer, who gets to see and use the thing that makes you money.

## Taste and the will to act

If capability is becoming cheap, then capability is no longer where a person's value sits. What is left are the two things the machine is not taking.

The first is taste. Knowing which of a thousand competent options is the right one for this person, in this context, at this moment. Taste stays scarce precisely because everything upstream of it got cheap. Great work has always been arresting because it is rare, and nothing that is too easy stays tasteful for long.

The second is agency. Not what you can do, and not even what you know you should do, but what you will actually go and do, decide, and stand behind. This is where we put our own weight, because we only ever get paid when the operator wins, so the will to own the outcome is not a virtue we admire from a distance, it is the thing we are built on. In a world where anyone can summon competent output, the person who decides, commits, and owns the result is worth more, not less.

There is a trap in all of this, and it catches companies hardest. A business spends years perfecting the thing it ships and slowly forgets why anyone buys it. The thing it ships, the product, the finished job, the report, is exactly the part cheap intelligence will soon reproduce for almost nothing. The reason people come, the judgment underneath the work and the read on a market no competitor shares, is not. Mistake the one for the other and you will defend the wrong thing while your product gets copied and undercut. Know what you actually sell, build that into something you own, and a cheap copy of your product stops mattering, because the product was never the point.

## What to build now

So stop asking whether AI will replace your people. Wrong question, and it keeps you passive.

Ask instead what your company knows that no one else does, and then ask why that knowledge still lives in three people's heads instead of in a system you own and control. That is the question with money in it.

Then be honest about which path you are on. There are only two. You can pipe the same generic intelligence as your competitors through the same rented pipes and slowly become indistinguishable from them, and from the lab that sells to all of you. Or you can take your hardest-won edge, the thing you could never afford to systematize before, build it into something you own, and keep the means of production, and your alpha, on your side of the line.

The descent of intelligence is not the opportunity. It happens to you and to everyone else at once, and it confers no advantage on its own. The opportunity is what you build out of your own edge, on infrastructure you control, before cheap intelligence makes edges easy to copy and easy to take. That window is open now, and it is the widest it will ever be.

I have spent twenty years watching people mistake the tool for the work. Our souls have always needed proof of work, something real that we made and can stand behind and call our own. That need doesn't go away when intelligence gets cheap. It gets sharper, because the thing you built and kept becomes the only part no one can take.

## Definitions

**The Edge**

The proprietary way a business does things that no competitor can copy. For all of history it lived in experienced people's heads. Cheap intelligence is what finally lets you build it into systems you own.

**Enterprise AI safety**

Not abstract alignment research or a government certification, but a business keeping control of its own data, model weights, and compute, so a frontier lab cannot absorb its proprietary knowledge into a competing product.

**Your alpha**

The proprietary knowledge and workflow that makes a business money, and that no competitor, including a model provider, can be allowed to see and reuse.

**Moving up the stack**

The shift from producing the analysis to deciding what the analysis is for, which question is worth asking, which answer to trust, and who owns the consequence.

**Taste**

Knowing which of a thousand competent options is the right one for this person, in this context, at this moment. Scarce precisely because everything upstream of it got cheap.

**Agency**

Not what you can do, and not even what you know you should do, but what you will actually go and do, decide, and stand behind.

**Tacit knowledge**

The judgment, sequence, and exceptions a business runs on that were never written down, and that walked out the door when the good people retired.

## FAQ

**If AI can do expert work for almost nothing, what is left for the expert?**

The work does not disappear, it moves up the stack. From producing the answer to deciding what to ask, which answer to trust, and who owns the consequence. Real roles get dislocated, that is true, but the demand for judgment multiplies, because far more decisions now get made and every one still needs someone to stand behind it.

**Why is cheap intelligence not an advantage on its own?**

Because it arrives for every company at the same time. If you consume it the same generic way your competitors do, you have not built an edge, you have erased the one you had. The advantage is what you build on top of it, out of knowledge that is yours alone, on infrastructure you control.

**What does real AI safety mean for a business?**

Not abstract alignment research or a government certification. It is control over your own data, model weights, and compute, so a frontier lab cannot absorb your proprietary knowledge and ship it as its next product. The safe position is keeping the ability to choose, at the model layer, who gets to see and use what makes you money.

**Will the AI lab really compete with its own customers?**

It has every incentive to. Whoever owns the model layer can watch where value is being created on top of it and move in directly. It has already happened. In April 2026 Anthropic launched Claude Design against Figma weeks after inviting Figma to partner on the announcement, and Figma's founder said the lab had not been consistently honest. Assume the roadmap can include your business, and keep control accordingly.

**What should a company actually do first?**

Ask what you know that no one else does, and why it still lives in three people's heads instead of a system you own and control. That is the question with money in it. Then build that edge into owned systems, on infrastructure you keep, before cheap intelligence makes edges easy to copy and easy to take.