Meta and Microsoft cut 20,000 jobs. Both companies are spending a combined $250+ billion on AI this year.
The math isn't subtle. The companies building AI are using AI to shrink their own workforce. Over 92,000 tech workers have been laid off in 2026 alone, bringing the total to nearly 900,000 since 2020. CNBC
Everyone's asking the wrong question. It's not "will AI replace engineers." It's "what does an engineer even do in 2027?"
Here's a fact most people don't know.
Before the 1940s, "computer" wasn't a machine. It was a job. You'd see ads in newspapers "Computer Wanted." It meant a human being, usually a mathematician or clerk, who did calculations by hand. NASA's early trajectory math was done by rooms full of "computers." Most of them were women. They were highly skilled. Then the machines arrived, and the job title dissolved.
But here's what people forget: those humans didn't disappear. They became something else. They became programmers, statisticians, analysts, engineers. The skill of thinking mathematically didn't go away. The job of executing the math did.
I think that's exactly what's happening to us right now.
When people see the "will AI replace software engineers" my honest answer is yes. But not in the way they mean. The role is being replaced. The people aren't.
Here's why I'm confident.
LLMs don't understand code. They predict patterns in code. That's a real distinction, and it matters. A model can produce correct code. It can also produce code that looks correct, compiles, runs, and quietly returns the wrong answer at 3am on a random tuesday. I've shipped both. I've debugged both. And the second kind takes longer to find than anything I've ever written by hand.
So the bottleneck isn't typing speed anymore. It's judgment.
The hard parts of a job in 2026 are not what they were in 2021. They are:
Defining the problem precisely enough that "correct" is verifiable. Most production bugs aren't bugs. They're underspecified requirements that the model interpreted differently than the human did.
System design. How do these 14 services talk to each other? What happens when the LLM call times out? What's the fallback? Models don't design systems. They build pieces of systems.
Validating outputs. Knowing when the model is confidently wrong. This is the new senior-engineer skill. It's also the hardest one to teach.
Taste. Knowing why this approach and not that approach. The model will happily do either. It has no opinion. You have to.
This is the shift. We're moving from production work to supervision and design work. The same way "computer" went from a person doing arithmetic to a person directing the machine that does the arithmetic.
The engineers who get cut won't be the ones who can't code. They'll be the ones who can only code.
If your value is "I can write a React component," yeah you're in trouble. The model writes React components. If your value is "I can look at a half-built system, see why it'll break under load, and tell you what to build instead," you're not going anywhere. The model can't do that yet. It might never.
The 20,000 layoffs at Meta and Microsoft? Read between the lines. The cuts concentrated in content moderation, customer support, software testing, and certain engineering roles places where work is well-defined and outputs are easy to verify. That's the tell. Well-defined, easy-to-verify work is exactly what LLMs eat first.
The work that's hard to define and hard to verify? That's the work that's becoming more valuable, not less.
Free Resource of the Week
This is the prompt I run on every piece of AI-generated code before it goes near production. It catches the kind of mistakes that look right and aren't. I built it after shipping a bug that cost a client $2,400 in duplicate Stripe charges. Use it on any code an LLM hands you.
Grab it here: The "Confidently Wrong" Code Reviewer
Sign-off
That's it for this week.
See you next Thursday.
— Carl