Microsoft sent a "Voluntary Exit Offer" to roughly 60,000 employees on Monday, a little over a quarter of the company, with decisions due in early December and departures at year end. The reporting ties it to funding something like $140 billion in capital expenditure, most of it AI. Strip the memo language and it is a trade: payroll out, compute in.

Microsoft is the latest example of an industry-wide trade, and the people inside the number deserve better than being a hook. The trade of people for AI raises a question nobody at Redmond has to answer, so I will.

But if the data center can build whatever you ask for, where does the effort go, and what do you tell the people who were told, years ago, to learn to code? What about the people who learned to code before being told that this was the future they should long for?

The wish#

This is what people seem to imagine: Janelle goes to an AI, and says, "I want a roguelike. Here are my rules and constraints. Make me a roguelike." The software appears. No effort, no craft, no job, just tokens, maybe even some money changes hands where Janelle has paid for an account so she has some headroom.

Look at that most excellent prompt again. The rules and constraints of a roguelike are the roguelike. Permadeath, procedural levels, the hunger clock (if any)1, what the player can see and when, what the dungeon is allowed to do to you: that's the design, and the design is the game. The "wish" is a design document in a trench coat, and the game that comes back is bounded by how good the document was. The effort didn't vanish. It moved up a layer, to Janelle.

I've argued before that specification cost, not capability, is the limit on this tooling. This is the same claim, just made from the other side. When the cost of building drops toward zero, the cost of knowing what to build is all that's left, and it was always the expensive part.

It's also the part that requires the people, often some of the same people being jettisoned by the bean counters2.

Which layer we told people to learn#

"Learn to code" was never friendly advice. It was offered from the top, to people in mines and on factory floors whose work had just been taken away, by someone whose work had not: just move up, how hard could it be? It sold the one layer that was cheap to sell, syntax and a framework and twelve weeks and a take-home, which was also the most commoditized layer of the work, and mein Gott, that was visible at the time. And the advice pointed people at exactly the rung the compute has now reached.

We got told "learn to code"; with AI, we've managed to translate that to "learn to want," but what we should have heard is "learn to imagine." There's no syllabus for it and no take-home test, no certification available. But that's the key: it isn't the same as "wanting a roguelike."

I can want a better spreadsheet without having the faintest picture of what "better" would mean; the want is a direction, and the tooling can't do anything with a direction. I can imagine a roguelike with a rule I've never seen in one, and that's a picture. The picture is what the tool needs, and the picture is what can create something new and worthwhile, given a human and the AI to help build it.

a directiona picturea checkable documentyes or noWantImagineSpecifyVerify
Diagram source
flowchart LR
    W[Want] -->|a direction| I[Imagine]
    I -->|a picture| S[Specify]
    S -->|a checkable document| V[Verify]
    V -->|yes or no| I

"Specify" is the teachable aspect, which makes it the practical replacement for "learn to code." "Verify" is the one I've written about in "The Floor and the Multiplier," as the person who pushes back where the model affirms: somebody still has to be the one who says no. "Imagine" is the thing the other two depend on, and it's the one with no course and no guide.

Newness is where the tool helps#

Ask for "a spreadsheet, again" and the weight of every spreadsheet ever written drags the result to the mean; you get the median spreadsheet, competently, based on everyone who's ever written about a spreadsheet and how one works.

Ask for a game with a rule the model has never seen, and it has choices now, because your rule didn't settle everything. Every gap you left is a door, and it's a "Lady, or the Tiger?" game: either the specifier closes the door herself, or the model picks one from the largest body of other people's decisions ever assembled, which is the mean you just spent your imagination getting away from. The model will reach for something. Good luck, Janelle3!

The common fear is that these tools produce average work. They produce average work when you ask for average things. Imagination is the lever that moves them off the mean, and the newer the thing, the longer the lever.

This is why the current enthusiasm for using the tools to reverse-engineer commercial products reads to me as a failure of imagination, even when the results are cool, maybe even worthwhile.

Note

I use some of those reverse-engineered things - I have been enjoying playing cloned versions of games from the 1990s. They're good, sometimes really good, and they're good because somebody imagined the original and the clone inherits that imagination at a discount. There's nothing wrong with buying at a discount. The discount exists because someone paid full price once, and if everyone shops the clearance rack the rack eventually empties.

A clone is a fine exercise and a poor destination. If the only thing we do with a tool that can build the unprecedented is rebuild the precedented, the failure of imagination is ours, not the tool's.

The argument against all this#

The obvious rebuttal is the Jevons paradox. In 1865 William Stanley Jevons noticed that more efficient steam engines made Britain burn more coal, not less, because cheap steam found work that expensive steam never could; efficiency grew the number of things worth doing rather than shrinking the resource it took to do them4.

Thus: compilers were going to end programming. So were spreadsheets, and Rails, and every tool that made software cheap. Each time, the volume of software went up and the median practitioner moved up a rung.

On that reading, "learn to code" was fine advice that simply needs a sequel, and the sequel is the same thing it always was: learn the next layer.

I think that's probably right about volume and probably wrong about people: designing is the intriguing part, code-writing was the grunt work, and most humans turn out to be weak at design, because design is hard.

The ladder exists. Not everyone can climb it, and the floor below it is being removed. Whether that's a transition or a cliff I don't know, and anyone who tells you they do is selling something.

Probably more AI services, these days.

What walked out#

Microsoft's capex buys the capacity to reproduce. It buys the median spreadsheet at any scale you like5. It can't buy the picture of the thing that doesn't exist yet, because that isn't for sale; it lives in people, and some of it is going to accept an offer and leave by December 31. Nobody priced that in, and I don't think anyone knows how.

A voluntary exit offer selects for tenure: the people who take it are the ones with the years to make the package worth taking, and those are the people who know how the place actually works. The culture, the knowledge, the things that make the institution what it is, leave with them, and the models don't have those things to give back.

My advice, as cold as it is: learn to imagine. Then learn to specify, because an imagination nobody else can check is a daydream. Write it down. Record it. If necessary, hand it off. Then learn to say no to what comes back. That's the ladder, and I don't have a twelve-week course for the first rung either.

  1. I built a roguelike recently, and it has no hunger mechanics, because I don't like hunger mechanics. It's also for the Commodore 64, and thus I got to tell myself that hunger took up too much RAM, so out it went.

    ↩

  2. Yes, I said "bean counters." I am all for profit, but profit for this quarter that makes the next six quarters far worse is dumb, and I've been on the wrong side of this myself... and watched companies fail afterward because the accountants cut people the company needed, on the strength of salary lines and projections that didn't ask what those people were still doing. Counting beans is dumb. There can always be more beans.

    ↩

  3. For what it's worth, I made up Janelle. If there's an actual Janelle I'm addressing, it's the guy sitting in my office chair right now. I entertained the models' suggestions, but made the choices myself.

    ↩

  4. The rebuttal has a prominent advocate: Satya Nadella himself invoked the Jevons paradox in early 2025, when cheaper models looked like a threat to AI spending, to argue that efficiency would increase demand for AI rather than reduce it. He was talking about compute, which is the resource. The paragraph above is about the people.

    ↩

  5. Weird, too, because Microsoft already has the better spreadsheet - people are using AI to recreate Excel in its median form.

    ↩