Learning to code in the age of AI : what really matters

if i were learning to code now, i’d go as close to the hardware as possible. C, assembly, computer architecture. i want to know what happens after the nice function names disappear.

learn how instructions execute. how registers and caches work. how memory is laid out, what moving data costs, how a GPU actually runs your code. spend time with a debugger and the assembly your compiler produces.

i think we’d be much further along in ML if more researchers really understood how computers work. you’re making decisions about computation while treating the computer as somebody else’s problem. where are the tensors stored? how often are they copied? what is the GPU waiting for? i want those questions involved in the research itself.

write a small allocator. implement matrix multiplication. change the memory layout and measure what happens. if something gets faster, find out why. there is a lot to learn in code that looks completely unimpressive in a screenshot.

AI can write plenty of code for you. fine. use it. i’d spend the time that saves learning what the machine is actually doing.