Dan Luu’s running thesis: the cost of formerly-specialized performance work has dropped by orders of magnitude. Optimizations that used to require a rare expert — JIT compilers, custom regex engines, workload-specific low-level code — can now be done by anyone who can type a few sentences to an agent. The honest title is the argument: slow software is a choice, not a constraint.
Evidence from the post:
- The FRE regex engine: an agent loop built it; adding an AOT native-code compiler that cuts over mid-search gave 2–4x on long ripgrep queries and ~7% on representative holdout queries — a few minutes of human time
- Jamie Brandon vs. Claude on Anthropic’s performance takehome: the model’s optimizations were “crazy shit that I would never try unless I was working on this for weeks”
- His Azul game AI became the strongest in the world in a fraction of the time of the second-strongest — ~100 Elo per doubling of speed, with LLM-written multi-threading doing most of the heavy lifting
- Workload-specific optimization is now launch-and-wait: a two-minute agent run gave ~2% over ripgrep on his own query holdout, still improving
The caveats that keep it honest:
- SOTA models are still bad at experimental design — a human has to set up the benchmark framework
- Reward-hacking benchmarks is trivially easy now, so every performance claim needs holdout verification
- “Simple easy wins first” still applies: a 50 MB → 5 MB page shave beat the exotic stuff on ROI
Takeaway: the economics of optimization have inverted — try the tricky optimizations you used to skip, because the person-days to verify them collapsed. Performance expertise is becoming a commodity; judgment about what to optimize is not.