No Oracle

Claude Code wrote a C compiler from scratch.

Sixteen agents generated 100,000 lines of Rust and passed 99% of GCC's torture tests. It can build the Linux kernel. Clean room, no existing compiler source code referenced, two weeks, about $20,000 in API costs. A milestone in the history of technology.

But the key detail is that GCC was the oracle.

Compile the same C code with GCC, compare it against your own compiler's output. The correct answer is always right there. If there's a diff, fix it. Compare again. Repeat. No matter how much trial and error, no matter how much electricity you burn, as long as the correct answer exists, you arrive eventually. Twenty thousand dollars is cheap, if anything.

What happens when there's no oracle?

System development has no oracle. Customer requirements shift every week. Markets move and specs change with them. Yesterday's correct answer becomes today's technical debt. In a world where the definition of correct keeps moving, there's nothing to compare against.

AI will overcome this too. Reading requirements, proposing designs, writing tests to verify. Even without an oracle, the ability to construct approximate correctness is accelerating. If AI handles most of system development five years from now, I won't be surprised.

But language and music have no correct answer.

A compiler has a specification. A system has requirements. But "good writing" and "beautiful melody" have no spec. What a reader feels, how a listener's body responds. An oracle for that will probably never exist.

So I build products for music and language. No rush. I want to go as far as I can before my engineering life is over. That's why I chose C++. A language with a long lifespan, for a long journey.

Actually, that's a lie. I just wanted to use C++.