Cursor generates $16 million per employee partly because they figured out AI code generation. So the capability curve is steepening. It’s not leveling off. And if you’re reasoning from what AI could do in 2025, in 2024, you’re working from an expired map. But the cost of not knowing what to build, of specifying badly or vaguely or not at all, is compounding much faster than production cost is falling, which is a huge statement because production cost is falling really fast.

Yet, every framework people reach for to understand this moment tends to ask the incorrect question because it tends to ask whether AI replaces workers and jobs. But when the cost of production is collapsing like this, the more useful question is actually what is the new bottleneck where jobs are going to be useful? What is the new bottleneck where humans have to get really clear? And guess what? It’s around intent. It’s around those specifications that engineers struggle to write. All of knowledge work is becoming an exercise in specifying intent.

And this video is about what happens when those engineering mental models get out into the rest of the job space and we all have to think about where our value is moving when it is not doing the work. I think one place we need to start when we understand jobs and AI is the thinking of Francois Chollet the creator of Keras one of the sharpest thinkers in machine learning. He made an argument that’s become the default framework for understanding AI and jobs. He pointed to translation, a profession where AI can perform 100% of the core task and has been able to do so since 2023.

Translators did not disappear. Employment has held roughly stable since. The work has shifted in the last couple of the years from doing it yourself to supervising AI output. Now payment rates have dropped and freelancers have gotten cut first. So there’s new hiring freezes going on. So there’s impact on jobs. And yet, despite all of that, the Bureau of Labor Statistics still projects modest growth for the translation job category.

Chollet’s claim is that software is going to follow the same pattern. More programmers will be needed in 5 years, not fewer. The jobs will transform rather than vanish. I think the model is useful for thinking, but I think it’s also stuck on the wrong question yet again. Will software engineers keep their jobs is not the most interesting question when the cost of production is collapsing towards zero because so many of us as engineers frankly so many of us as knowledge workers all of our work has been in production and so if you’re going to take the cost of production to zero will we keep our jobs is really the wrong way to think about it.

It’s really what is our job going to turn into? And so the interesting question if we ask about job transformation not just for engineers but for everybody is what is becoming scarce and therefore what is becoming valuable when doing the work when building is no longer the hard part. Chollet doesn’t have a framework for that because translation’s capability plateau gave the market the time to find a stable answer in the translation job category.

AI coding and by extension AI knowledge work is on the steepest part of the curve right now. I’ve said before that I think benchmarks are fairly easy to game. I’m not the only person to say that. But the production evidence of coding capability gain is so unambiguous. You don’t need to pay attention to a benchmark to believe it. You get to look at Cursor’s ARR and how fast they’re growing. Look at Lovable. Look at the ability to now have agents review the code of agents.

Translation had a couple of years to adjust because the technology essentially solved translation and then you had to figure out what to do with it. Software may not get the same runway because the depth of what’s changing is much more profound and the pace is even faster.

We need a different model to understand how jobs in software and knowledge work are going to change. First, when cost goes to zero, demand goes to infinity. Every time in economic history that the marginal cost of production has collapsed in a given domain, demand has exploded. Desktop publishing did not eliminate graphic designers. It created a universe of design work that could not have existed at any price point prior. Cameras in all of our phones created a universe of photography that did not exist when cameras were very expensive and only a few people had them.

Mobile didn’t replace developers. It multiplied the number of applications the world needed by orders of magnitude. Software is about to go through the same expansion except bigger.

Right now, most of the world cannot afford custom software. Regional hospitals run on spreadsheets. Small manufacturers will track inventory by hand. School districts use tools designed for organizations 10 times their size or more, and some of them use nothing at all. The total addressable market for software is constrained not by demand because demand is functionally infinite. It’s constrained by the cost to produce.

We are underbuilt on software even after 30 years of software engineering 40 50 years. When the cost of production collapses, constraint that means that we are underbuilt lifts forever. Every business process currently running in email, spreadsheets, phone calls is up for grabs now. Every workflow that was never worth automating at a $200 an hour engineering rate becomes worth automating at two bucks in API calls.

The market for software is not going to contract. It is going to explode. And that is the best argument for why total software employment likely grows and not shrinks. Chollet is right about that. The demand for people who make software happen, however they make it happen, it may not be traditional coding, it won’t be. That has never been higher and the cost collapse is going to push it higher still.

But I do want to be honest, just because we can wave our hands and say Jevons paradox means employment grows does not mean your specific job is safe. And understanding the difference requires understanding what happens when the constraint shifts from production to specification.

So let’s talk a little bit more about the specification bottleneck. The majority of software projects that fail don’t fail because of bad engineering. They fail because nobody specified the correct thing to build. Make it user friendly is not a specification. It’s like Uber for dog walkers is not a specification either. It’s just a vibes pitch.

The entire discipline of software engineering, agile, sprint planning, etc. evolved as a way of forcing specification out of vague human language. We need mechanisms for converting vague human intent into instructions precise enough that code can be written against them.

That vagueness problem has always been there. What’s new is that the friction of implementation is changing. When building something took 6 months and at best a half a million dollars, organizations were forced to think really carefully about what they wanted. The cost of building acted like a filter on the quality of the spec.

If you take away the cost of building, as AI is doing, that filter is going to disappear. The incentive to specify just evaporated in all of your orgs and the cost of specifying really badly is going to keep compounding faster than ever because now you can build the wrong thing at unprecedented speed and scale.

A vibecoded app can take an afternoon and 20 bucks in API calls and if the spec is wrong, you did not save 6 months. You wasted an afternoon and perhaps launch something that will harm customers because the spec was never right.

This is the inversion we need to pay more attention to because it tells us a lot about where jobs are headed. The scarce resource in software is not the ability to write code. It’s the ability to define what the code should do. And funnily enough, that is part of why knowledge work is starting to collapse into a blurry job family. Because the ability to specify is something we all need to do, not just engineers.