There's no reason for software to be slow anymore

522 points · 375 comments on HN · read original →

Points and comments are a snapshot, not live.

With LLM agents, performance optimizations that took days now take minutes.

Dan Luu argues that LLM coding agents have collapsed the human time and expertise needed for performance optimization, making formerly expensive work cheap enough for personal projects. He describes workload-specific regex optimization: a version of his FRE engine, optimized via agent for his own ripgrep queries, showed a 2% speedup on a holdout set after minutes of human time. He also built a world-leading Azul AI via agent, where multi-threading and other optimizations would have taken days by hand. He notes that dynamic, workload-fitted software is becoming feasible for larger companies as well.

What commenters are saying

Commenters split into two camps. One camp argues that despite the technical potential, software feels slower because business incentives still favor new features over optimization, and enshittification via VC funding and AI-generated mediocrity continues. The other camp notes that the RAM crunch from AI may force optimization back into focus, citing iOS 27's improved performance on old hardware. Several commenters critique the article's bare-bones website style as needlessly ugly.