AI is writing GPU code even human engineers can’t fully understand

For years, CUDA engineers who optimize Nvidia chips have been among tech’s most sought-after specialists. Now, they’re increasingly managing AI that does it for them.

These engineers specialize in using Compute Unified Device Architecture (CUDA) — Nvidia’s software for programming its AI chips, known as GPUs — to write code that can run AI as efficiently as possible.

The jobs are coveted because compute costs are massive; engineers who can squeeze the most value from a chip can save companies millions.

CUDA engineers traditionally spent their days writing kernels — the small pieces of code that tell a GPU how to do a single job as quickly as possible — and then testing them to find the fastest version. Now, AI is taking over much of that painstaking work by generating hundreds of kernels, testing them, and picking the best one.

As a result, CUDA engineers increasingly spend their time supervising coding agents. It’s part of a broader shift across the software industry, where developers are moving from writing code to managing AI.

This involves setting goals, checking results, and stepping in when the AI gets stuck, said Jeremy Nixon, founder of the startup Infinity, which builds AI that optimizes chip software.

Because AI can introduce “bizarre” bugs that humans wouldn’t have written, the job of reviewing AI code has become “more intense,” said Anne Ouyang, cofounder of the AI infrastructure startup Standard Kernel.

Despite the…