Blog

AI is the first abstraction you can't trust

The usual story about AI coding is that it's the next step up the ladder. Machine code gave way to assembly, assembly to C, C to memory-managed languages like Java and JavaScript. Now we write English and the model writes the code. Each step let you forget the one below, and this one is no different.

I think that story is wrong twice. The earlier steps never let good developers forget the layer below. And this step isn't like the earlier ones at all.

The layers below never stopped mattering

I remember George Hotz saying that if you hire a JavaScript developer, they should know what V8, libuv and Node are. Not as trivia. If you say "V8 is a JavaScript virtual machine", what is a virtual machine? If you say "JavaScript gets JIT compiled", what does JIT compilation mean?

You can be a JavaScript developer with no idea how a CPU works. Plenty of people are, and they ship. But I've never met a developer I'd call good who didn't understand the fundamentals. So I don't see why moving one more level up would suddenly make them optional.

This layer is different in kind

Every earlier layer is deterministic and well specified. An assembler, a C compiler and a JIT do the same thing every time, and they are almost always right. That is why most JavaScript developers could afford to ignore V8: it earned the trust.

An LLM hasn't. It gives different output for the same input, and it is confidently wrong a fair amount of the time. That makes it less like the step from C to Java and more like taking code from a fast junior who never tells you when they're guessing.

Joel Spolsky's law of leaky abstractions says every non-trivial abstraction leaks, and when it does you need the layer underneath. Compilers leak rarely and predictably. Models leak often, and you can't tell where in advance. Reviewing their output takes more understanding of the layers below, not less.

Where my own argument is weak

Three objections are worth taking seriously.

Fundamentals might be a sign, not a cause. The good developers I know may be good because they're curious and obsessive. Knowing what libuv is could just be what that curiosity leaves behind. Hotz's question works as a hiring filter because it's a proxy for the trait.

"Not very good" is not "not useful". Huge amounts of valuable software were shipped by people who couldn't explain JIT compilation. The market has always paid for that work, and it still will.

The real risk is the missing bottom rung. Most people didn't learn fundamentals from a textbook. They picked them up by accident, doing junior work: chasing a memory leak, reading a stack trace, finding out why the query was slow. That is exactly the work AI now does first.

Are we heading for a post-knowledge society?

The enrolment numbers get quoted as proof that we are. They're real, but smaller than the headlines.

  • Undergraduate computer science enrolment at US four-year colleges fell 8.1% in autumn 2025 and 8.4% in spring 2026, about 53,000 students (Chronicle of Higher Education).
  • Some places are hit much harder. Oregon State lost 23% of its computer science majors in a year, and graduate programmes have lost more than a quarter of their students since autumn 2024.
  • The University of California is down 6% in a year and 9% over two, its first sustained fall since the dot-com bust. The total is still close to double what it was a decade ago (GovTech).

And the students aren't giving up on hard subjects. Engineering grew 7.3% over the same period, nearly mirroring the fall in computer science (National Student Clearinghouse). That is people reading the job market, not walking away from knowledge.

The knowledge is concentrating

So no, I don't think it's post-knowledge. I think the knowledge is ending up in fewer heads.

The likely result is a split. A small group understands the whole stack and can tell when the model is guessing. A much larger group operates tools it can't debug. The first group gets rarer, and worth more, every year the second one grows.

If you're learning to code now, that tells you where to spend your time: on the layer the AI can't be trusted with.