saved
There's no reason for software to be slow anymore
gist
Dan Luu argues that coding agents have made specialized performance work cheap enough that slow software is no longer an inevitable default. Formerly expert-only optimizations—JITs, multithreading, workload-specific engines—can now be launched in minutes of human time. Bounded problems and personal workloads see quick wins; open-ended experimental design still needs humans. The practical consequence is more custom, workload-fitted software rather than one-size-fits-all engines.
ideas
- Performance expertise is no longer the scarce input. Agents collapse the human time to try tricky optimizations that once needed rare skills, so many more 2% wins become worth attempting.
- Optimize for a class, then for a workload. FRE’s AOT path and ripgrep-derived holdouts show class-level gains; minutes of agent work can then specialize further to real query distributions.
- Bounded problems beat open judgment. Agents can outpace strong performance engineers on well-defined takehomes, but still need humans to set benchmarks and experimental design.
- Custom software becomes the default shape. As Brooker and Malis note, fitting engines to particular workloads—or customer traces—looks more economical than shipping one general optimizer.
- Cheap does not mean automatic generality. Overfitting and regime change remain risks; holdouts and careful measurement still gate whether a speedup is real.
quotes
“Dynamic custom software, fitted to a particular workload rather than a class of workloads, seems like a very likely outcome.”
“A lot of the tedium it used to take to get a tricky optimization like this working is gone.”
“LLMs have lowered the barrier to entry and made it much easier to write a JIT compiler.”
“you can afford to use a coding agent that can beat him on a bounded optimization problem.”