The idea that Large Language Models (LLMs) are leading to slower, bloated code is being challenged by the notion that these models will enable a rewrite in highly optimized assembly. While we're not at the stage of writing everything in assembly, the principle that performance can be chosen, similar to how bugs can be chosen, is becoming increasingly relevant.
This is partly due to the significant reduction in the cost of specialized performance work. Tasks that once required rare expertise can now be accomplished by anyone who can provide simple instructions to an AI. This makes optimizations that were previously too expensive for all but the largest projects feasible for a wider range of applications.
Marc Brooker noted that this trend could lead to dynamic custom software tailored to specific workloads, reminiscent of projects like FFTW and demoscene techniques that focused on extreme optimization for particular problems and hardware.
Michael Malis has also observed that while
