RAM-to-arithmetic cost ratio went from 1:1 to 1000:1 — time to make LLMs actually optimize code

CatAstro_Piyush · x · 2026-11-20

@bzogrammer notes most code is extremely inefficient: the constant-factor cost ratio of RAM loads vs arithmetic was about 1:1 50 years ago but is now 1000:1, while much algorithm and data structure design never caught up.

He argues RLVR is the main way LLMs improve today, and making code faster is both trivially easy to grade and directly yields faster, more efficient RL training. With the "hand-optimizing code is too much work" excuse gone thanks to LLMs, users should ask their models to actually optimize code — and expect them to get better at it over time.

Original post →

More from AGI Musings

AGI Musings channel →