7 comments

[ 0.22 ms ] story [ 19.8 ms ] thread
Saving you a click:

> I decided to try it out and see if it could optimize my Rust program.

> Kimi K3 had 2 recommendations, and I decided to have it pursue the SWAR optimization—SIMD Within a Register. With SWAR, a single 64-bit value is treated as 8 eight-bit values—so we can handle 8 u8 values (single-byte) at once.

> Kimi K3 took about 15 minutes for the task, and it completed it perfectly. I found the optimization sped up the program by 2%.

> It cost about $1 for the tokens on OpenRouter.

Question is whether other models would offer same optimization or was it exclusive to K3? Would have been much more interesting write-up.
$1 seem a lot, with a high cache hit rate it should be lower.
Should have given the other frontier models a chance too? What if Fable or GPT-5.6 would have found a 30% speed up? Not like a sample size of one means much though.
> Tested Kimi K3 for Coding

TLDR: Kimi K3 is good, but the data point is minuscule.

Author couldn't be bothered to compare any other models? ok man