Amdahl's Law for LLM generated code

[akiarie]

LLMs may theoretically be able to generate millions of correct lines of code.

But for any important code the only way to know that it's correct is to hand-audit every line, which is harder than writing every line because this allows you to build a conceptual hierarchy and model within which you can think about the code.

So there's a fundamental limit to the speedup from LLMs when writing anything you care about.

(The main objection will be what about when you have a human engineer working on your behalf. But this is a distraction: sometimes you can fully trust an engineer, which is not at all true for LLMs.)

[akiarie]

Never mind, someone else already made this point: https://electric.ax/blog/2026/02/19/amdahls-law-for-ai-agent...

[srbsa]

Thank you! This changed how I was thinking about my product: a knowledge harness for coding agents. While I was discussing incidents & context-switching tax, the structural compounding is likely the real win.

[dnnddidiej]

Yep. AI is a freeway. But there is still bumper traffic in the city at the end of it.

Or like your 1 hour flight has 1 hour to check in etc.

[phatnhse]

so cool