alt Hacker NewsIt’s an interesting strategy, but I question how much it pays off, if at all. Very few parts of a program benefit from manually tuned assembly compared to the naive C implementation. Writing everything in assembly adds an extra layer of thought, which even for an LLM is additional effort that could have been used for targeted optimizations instead. It makes it harder to notice patterns that have been trained into the data set, from security problems to performance opportunities.
On a long enough time frame with enough tokens invested there’s probably not a difference, but being written in assembly by an LLM doesn’t imply optimal to me. I’d almost prefer having an LLM rely on higher level abstractions offered by a programming language rather than rolling everything itself. After reviewing a lot of LLM code, even at Fable and Sol levels, I just don’t trust that LLMs are writing optimal code. Assembly makes it harder to even review.
I do it find it very fun and entertaining. This is a component and I’m grateful that it was shared.
It will make the code slower.
Writing maintainable assembly is at odds with writing fast assembly in most circumstances.
A key optimization that's hard to pull off is inlining.
An optimizing compiler can see that a method is small enough that it can be pulled into the caller, it can then further eliminate from that smaller method branches that can't be executed due to the nature of the caller (Imagine calling a function with a `bool` parameter and you send in `true` at the call site).
To make the code faster in hand written assembly, you have to do the inlining, but that makes writing more structured code a lot harder. You are duplicating logic paths in the name of performance.
Not to mention the fact that the compiler gets updated and knows about more instructions and architectures then you do or then you could have. Hard to write the FMA instruction if it didn't exist when you were writing the assembly in the first place.