Dan Luu argues agent-driven optimization makes workload-specific performance work cheap, and measures 2 percent on holdout ripgrep queries

Sources: danluu.com post, HN discussion
Summary: The reported wins are 2x to 4x on a few long queries, and 2 percent on representative holdout queries for the workload-optimized build against standard ripgrep. A separate result, the cutover to the native code compiler, is given as about 7 percent on representative holdout queries for queries where AOT is enabled. The argument rests on the cost of attempting an optimization falling rather than on the size of any single result. The author states he set a half-hour budget for the writeup and expects data problems he did not find.
Why it matters: The claim is about the price of attempting an optimization, not the size of the win, so the holdout figures are the ones to carry, 2 percent for the workload-optimized build and about 7 percent from the AOT cutover where AOT is enabled, rather than the 2x to 4x on individual long queries.