one of the reasons I think deepseek is so efficient is because they have ML researchers who are cracked systems programmers and bare metal silicon wizards while other labs might have a separation layer between ML researchers and systems programmers

I know many ML researchers who haven’t heard about vector extensions, SIMD and some haven’t seen the computer world below the compiler. 

This really limits the amount of juice you can squeeze out from a machine

Curious | building robots | https://anurajrp.io/about/

yea squeezing-outs are at or below the compile target