AI's new complexity: Why all code is now write-only and how teams should adapt

Phillip Mortimer, speaking at InfoQ, draws an unexpected parallel between the ultra-dense C code that Arthur Whitney crafted in 1989—a full working interpreter for the J programming language written in a single afternoon—and the volume of code that modern AI systems generate autonomously. Both create what programmers call "write-only code": software so complex that it defeats comprehension, though for entirely different reasons.

The evolution of write-only code#

Traditionally, write-only code emerges from languages engineered for power over readability. APL, used heavily in finance during the '70s and '80s, could express Conway's Game of Life in a single line where most languages require 40 or 50. Regular expressions—the poster child of opaque code—pack powerful pattern matching into compressed syntax that's nearly impossible to parse after writing it down.

The key insight: these dense notations force developers to rely on two principles. First, extensive tests become the true documentation, defining behavior through input-output pairs since reading the code itself is futile. Second, code becomes disposable; when debugging is nearly impossible, it's faster to rewrite than to modify.

AI flips the complexity problem#

Today, an entirely new kind of write-only code dominates. AI systems like Claude write readable, well-commented, comprehensive code—hundreds of thousands of lines across thousands of commits—far too voluminous for humans to review line-by-line. Anthropic's recent example of Claude autonomously developing a C compiler in Rust involved 500,000 lines added and 277,000 deleted. No individual line is arcane; the sheer volume is incomprehensible.

The result: the same two principles that governed old-school write-only code apply again. Yet traditional code review—the pull request scrutinized by human eyes—becomes a bottleneck rather than a safeguard.

Practical strategies for the AI era#

Detailed test suites guide autonomous development. Nicholas Carlini, the Anthropic researcher who supervised Claude's compiler project, didn't try to review individual functions. Instead, he assembled high-quality test cases and built a harness that guided the model's work. The tests defined success; the implementation was merely a means to pass them.

Automate code review with AI. Counter-intuitively, using the same AI system to review its own output works remarkably well. A different prompt, context, and framing let models catch errors they made during generation—a practical reality already in use at Carta. In one instance, AI-powered code review detected catastrophic backtracking in a generated regex that could cause denial-of-service vulnerabilities.