There’s no reason for software to be slow anymore. LLMs have lowered the barrier to performance optimization, so specialized tuning once reserved for experts can now be done by anyone typing a few sentences. As [PERSON_NAME] noted, dynamic custom software fitted to particular workloads is emerging, reminiscent of FFTW and demoscene techniques built for speed. Historically, writing JIT compilers was too difficult, but AI changed that.

We tested this by creating FRE, a regex engine optimized via an agent loop. Initially overfit, it generalized after holdout warnings. To improve efficiency, we ran a native code compiler in a background thread while ripgrep executed normally. This trades short-query speed for major gains on longer searches.

A simple prompt let an agent implement the threading experiment and benchmark real data. It delivered a 2x-4x performance improvement. With AI handling complexity, ambitious optimization is now universally feasible.