Threads @iammurataslann — DdjLJvfiEZp
Murat Aslan (@iammurataslann) · Threads · source
A 27B model spent 21 days trying to build its own CUDA inference engine on a single RTX 3090.
No human-written CUDA. About a dozen human nudges in three weeks.
It lost the speed race. That’s not the interesting part.