A summer of AI optimization
Daniel Lemire's blog
Six mature open-source libraries got much faster in 2026, mostly over a few summer weeks. The surprise: the gains came after years of flat performance, and AI seems to have lowered the cost of trying ideas.
Based on reporting by Daniel Lemire's blog — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Six mature open-source libraries all had the same story: years of hard-won tuning, then suddenly a summer burst of speed. The author, who maintains or co-maintains them, rebuilt every commit and benchmarked the results on one machine, an Intel Xeon Gold 6548N, using August 2024 as the baseline. The point wasn’t to prove AI wrote the code. It was simpler than that: the code got better, and the gains were real.
Roaring, the Go bitmap library, saw some of the most dramatic jumps. Decoding to an array became 2.5 times faster, FastOr hit 3.1 times on one data set, the many-value iterator improved by 4.5 to 5.9 times, and intersection cardinality rose 10%. The author says one contributor was perfloop, with Philipp Klose also helping after declaring Claude use.
Ada had been stuck for a long stretch. From August 2024 to July 2026, roughly 550 commits left throughput on a 100,000-URL corpus at 0.54 GB/s. Then, in six weeks, it climbed to 1.28 GB/s, about 15 million URLs per second on one core. Most of that came from Yagiz Nizipli, with smaller contributions from Abdul Rawoof Khan and Dillon Mulroy. The author also worked on IP address parsing, though that didn’t show up in this benchmark.
Fast_float, simdjson, simdutf, and CRoaring all followed the same pattern: long plateaus, then sharp gains once people started trying more ideas. Fast_float improved 43% on one file and 70% on another between March and July 2026. Simdjson’s new C++26 reflection support made serialization 1.6 times faster on twitter.json and 2.1 times faster on citm_catalog.json. Simdutf nearly doubled ASCII validation from 83 GB/s to 160 GB/s, and CRoaring sped up several bitmap operations by large margins.
The author’s argument is blunt: the techniques themselves were not new, but AI made experimentation cheap enough that people actually did it. That’s the part worth sitting with. Not the hype, not the fear, just a summer where neglected optimizations finally got finished.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous AI story that matters: not robot poetry, just fewer excuses for leaving easy speed on the table. Open-source infrastructure has been running on patience and coffee for years; if AI helps people try the boring ideas faster, good. The industry can keep arguing about grand risks while the real world enjoys faster parsers and fewer wasted weekends.
Read more about this at: Daniel Lemire's blog