Had a strange performance regression from a seemingly innocuous code change, and GPT-5.6 tracked it down to a microcode workaround for an Intel CPU bug. The workaround can make conditional jumps that straddle a 32-byte boundary much slower, so programs have to be laid out around it. Tried building with -mbranches-within-32B-boundaries and wouldn't you know it, we're fast again! The future is so fucking cool!
Muse Spark 1.3: Cheaper Than GPT-5.6, Better at Coding