Did you know vast majority of AI development happening right now is on python? The thing that literally consumes billions of dollars of even-beefier-than-4090 GPUs like A100. Don't you think if they could do this more efficiently and better on C or assembly, they would do it? They would save billions.

Reality is that it makes no benefit to move away from python to lower level languages. There is no poor optimization you seek. In fact if they were to try this in lower level languages, they'll take even longer to optimize and yield worse results.

[–] pipe01@programming.dev 5 points 3 days ago (1 children)

Every single high performance python library is written in C, you just don't see it

[–] Grumpy@sh.itjust.works 2 points 3 days ago

I've already addressed this same argument to the other commenter in length. But if what you say is true, which it is not, then python is not inefficient since it's already equivalent to C. Thus OP point of something being inefficient because it was made in python is wrong.

[–] Alexstarfire@lemmy.world 6 points 3 days ago (1 children)

TBF, using AI as an example isn't the best choice when it consumes an ungodly amount of power.

[–] Grumpy@sh.itjust.works 4 points 3 days ago

It makes the best example because there's that much more money to be saved.

[–] RarePossum@programming.dev 2 points 3 days ago (1 children)

They do do it in C. The packages are written in C, python is just used as the wrapper to allow less coding skilled data scientists to easily use it.

That's like the entire data science joke. It's C in a python trench coat.

[–] Grumpy@sh.itjust.works 3 points 3 days ago

Nearly every languages' every core packages are written in C. And almost every higher packages have some amount of C. That doesn't mean we get to say every program is done in C. And if you keep drilling down, everything is just machine lang. And certainly still disproves the OPs point of inefficient python.

Saying it's all done in C hardly even true. Just look at xformers library on GitHub. Only 2.7% of the code is C. And the entire library is about optimizing.

Additionally, vast majority of the great leaps in ML efficiency changes hasn't come from better programmed packages, though they too certainly made big strides. How we calculate itself has changed. That's what makes the greatest optimizations in anything. It doesn't matter what language it is, doing a loop 1000000 times to add 1 is going to be worse performance than just doing 1 multiplied by 1000000. How we calculate, what we choose to give up (such as determinism in some implementations if SDP attention changes) and such makes big differences.

Optimizations also has to be done by someone. Whether that be data scientists or otherwise. The ability for higher level languages to enable them to do so like you say also makes a big difference. If all the programmers had to optimize in C only, we'd still be way behind where we are now in performance.

Just swapping languages doesn't yield better results like OP is implying.