Dan Luu Argues LLMs Dropped the Cost of Performance Work by Orders of Magnitude, Making Per-Workload Custom Optimization Viable
Luu's post, which hit 631 points on Hacker News on August 20, extends Nolan Lawson's 'you can choose how many bugs you want' framing to performance: optimization that used to need a rare specialist or a whole team is now reachable by anyone who can write a few sentences. Marc Brooker's response in the piece sharpens the consequence, predicting dynamic custom software fitted to one particular workload rather than a class of workloads, comparing it to FFTW and demoscene techniques tuned to specific hardware. The counter-argument Luu is answering is the viral claim that people complaining about slow LLM-generated code will be proven wrong when everything gets rewritten in optimized assembly, which he says overshoots the actual mechanism.
↳ Follow the thread