That’s just a straw man, because there’s no reason why you wouldn’t be looking through your code. What LLM does is help you find areas of the code that are worth looking at.

It’s not a strawman because classifying unperformant code is a different task than generating performant replacement code. LLM can only generate code via it’s internal weights + input it doesn’t guarantee that that code is compilable, performant, readable, understandable, self documenting or much of anything.

The performance gain here is coincidental simply because the generated code uses functions that call processor features directly rather than get optimized into processor features by a compiler. LLM classifiers are also statistically analyzing the AST for performance they aren’t actually performing real static analysis of the AST or it’s compiled version. It doesn’t calculate a BigO or really know how to reason through this problem, it’s just primed that when you write the for loop to sum, that’s “slower” than using _mm_add_ps. It doesn’t even know which cases of the for loop compile down to a _mm_add_ps instruction on which compilers and which optimization levels.

Lastly you injected this line of reasoning when you basically said “why would I do this boring stuff as a programmer when I can get the LLM to do it”. It’s nice that there’s a tool that you can politely ask to parse your garbage and replace with other garbage that happens to use a function that’s more performant. But not only is this not Software Engineering, but a performant dot product is a solved problem at EVERY level of abstraction. This programming equivalent of tech bros reinventing the train every 5 years.

The fact that this is needed is a problem in and of itself with how people are building this software. This is machine spirit communion with technojargon. Instead of learning how to vectorize algorithms you’re feeding your garbage code through a LLM to produce garbage code with SIMD instructions in it. That is quite literally stunting your growth as a Software Engineer. You are choosing to ignore learning how things actually work because it’s too hard to parse through the existing garbage. A SIMD dot product algo is literally a 2 week college junior homework assignment.

Understanding what good uses for it are and the limitations of the tech is far more productive than simply rejecting it entirely.

I quite literally pointed several limitations in the post you replied to and in this post from a Software Engineering perspective.

Hey so I read your comments and found them insightful. Me being a Software Engineer who just started his first job, what would be your advice for the right approach to grow and learn as a software engineer? Both in general and with respect to using LLMs while learning/coding.