The machines got good at writing code, so the story goes that the people who write code are finished. I think there will be more of them. Nobody was ever paying for the typing.
Most software anyone has paid for had to clear one bar: worth an engineer’s time at an engineer’s price. That bar has held for decades, so nobody knows how much software the world actually wants. We know how much was worth paying for. Nobody has seen the other number.
Will Bryk drew this line on Christmas Day 2024, days after OpenAI announced o3. He gave mathematicians seven hundred days. He asked whether AI would automate software engineers away soon, answered “No,” and gave the reason: engineers spend their days with customers and teammates, working inside “a ton of organizational context.” This month he wrote that all of mathematics could fall to AI within months, close to the seven hundred days he gave it. In the same post he held the other side: “we won’t fully trust them as software engineers.” I think he’s right, and I think I can say why.
Jevons saw half of it in 1865. Engines were getting more efficient, so Britain was expected to burn less coal. It burned more, because efficiency makes things worth doing that weren’t worth doing before. That explains demand. It says nothing about who gets hired. Coal is something engines burn. Software is work people were doing.
The other half is in Bryk’s own post, a few paragraphs above the part people quote:
The o3 class models are reeeaally good at optimizing for anything you can define a reward function for. Math and coding are pretty easy to design a reward function for.
What survives is what you can’t score. “Did it compile” is a score. “Was this the right thing to build for this business” isn’t. There’s nothing to check it against, and the answer turns up months later through people who will never put it as a grade.
So what got automated was the typing. Typing was never the job. The job was deciding what should exist, catching the confident answer that’s quietly wrong, and being the one answerable when it breaks. Those were always the parts that were hard to hire for.
Venkat Subramaniam, forty years into this, puts it from the other end:
When I ask AI to do something that I’m a novice at, I’m at awe at what AI is doing. But if I ask AI to do something that I’m an expert at, I find it awful.
What he’s describing is where the value sits: with whoever can tell good output from bad.
That has a cost, and Bryk names it in his post about maths, adding that it applies to software too: people who love the work for the craft, or for being better at it than everyone around them, “will feel diminished.” The ones who love it for what gets built won’t.
The harder question is whether fewer people will be needed to do that judging. If a project takes a quarter of the hours it used to, one senior engineer could approve four times the work while the rest of the team goes home. On some teams that will happen. Settled specs, good tests, known failure modes: automate all of it.
But what gets overseen is a business, and businesses don’t compress the way code does. Models can read every ticket and summarise every call, so recall isn’t the bottleneck. The customer is. People find out what they want by seeing the wrong version first, and that happens at their speed, not the model’s. Ten more modules in a system someone already knows is real leverage. Ten systems in ten companies is ten customers, each deciding at their own pace.
So the arithmetic comes down to one question. If a project needs a quarter of the human time, employment falls unless the cheaper price more than quadruples the number of projects. I think it brings in far more, because today’s price shuts out most of the businesses that could use software built around them.
And what people buy when they buy software is somebody whose problem it is when it breaks. Sometimes that’s a vendor, and somebody at the vendor still has to understand the thing well enough to fix it. Software fails quietly, the cost lands long after the code was written, and it has nothing to do with what the code cost. Cheap code raises that cost, because cheap code ends up running in more places nobody’s watching.
Here’s what I want. Software built around one particular business has mostly belonged to whoever could afford a team to build it and keep it running. Everyone else made do with spreadsheets, or bought the product built for a bigger customer and paid the difference in workarounds. That was arithmetic, and the arithmetic just changed. I want millions of companies of one and three people, each mattering enormously to a few hundred customers, instead of ten that matter slightly to everyone.
It only holds while intelligence stays cheap, and only while nobody owns enough of it to name the price. Hank Green has the clearest version of why that price matters. Some things saturate: we worked out what hot water was for and stopped wanting more. Other things never do, because there’s almost nothing you can’t do with them. If intelligence is the second kind, its price will shape nearly everything built on top of it, and it’ll decide more about who ends up owning all this than anything the models themselves do.
Bryk graded his own predictions a year later, thirteen of fifteen by his count, and said the next ones would be concrete enough to grade properly. So I’ll put a count on this one.
I opened by saying nobody was paying for the typing. That isn’t quite true. A lot of software work has been sold by the hour, much of it through large services firms. If typing is what’s being automated, that’s where it shows first. So I’m counting two things: software developer employment in the US Bureau of Labor Statistics’ annual occupational survey, and the tech industry headcount NASSCOM publishes for India each year. In 2025 they stood at 1,687,890 and about 5.8 million. My bet is that both are higher in 2029. I’ll post the numbers here every year until then, and if both come in lower, I was wrong.
I’ve already bet on the answer, which makes me the least neutral person to be telling you it. What I can say is that I looked at it long enough to stop having an opinion about it and start living on it. I’m full-time on this now. If the direction is wrong, I’ll find out in the most direct way there is, and I’d rather be the experiment than the commentary.