Flash-MSA: Train on a million tokens without breaking the compute bank
A sparse attention kernel that makes long-context training finally affordable - if the benchmarks hold up outside the lab.
Jul 13, 2026 at 02:20 10 10
"Your code is fast - if you're lucky": why your micro-benchmarks lie
A viral post questions the reliability of micro-benchmarks: the same code can vary by 30% depending on the order of symbols, memory position, or username. What this means for craft.
Jul 11, 2026 at 17:14 9 7