Yann LeCun and Others Discuss: LLMs are the New Compilers, Performance is a Function of Compute

yisongyue · x · 2026-07-30

Prominent AI researcher Yann LeCun reposted and endorsed a profound analogy: "LLMs in general are the new compilers."

The argument suggests that a model's performance is simply a function of the compute spent on the "compilation" process (inference or training). As long as you can verify the program's output or behavior, you can continuously push its performance to the absolute limit.

This discussion originated from a demo by AsariAILabs, where their self-improving agents optimized the vLLM inference stack, achieving up to 16% more throughput and interactivity for DeepSeek and GLM models on B200s. This perfectly illustrates the logic of pushing performance limits through compute and iterative verification.

Original post →

More from AGI Musings

AGI Musings channel →