Code is clay: AI as a power tool for experts
A post called Code is clay argues that code is now malleable rather than precious. Commenters add the old line that code is a liability and the data it produces is the asset. A map tool that triangulates 3.7M polygons in 489 ms faster than earcut was built with LLM help, and the earcut author joined in to compare notes. Commenters keep asking which optimizations the model found and which needed a human.
The Switch 2 jailbreak thread has the same shape. Commenters set aside how much the AI mattered compared with the user, an experienced security expert. A Rust style guide thread gets a comment that style guides now mostly exist to feed to a model. WorldTUI trained a 360M model to pretend to be a terminal, and commenters ask what the point is when real terminals exist, while others see a wide field in fine-tuning small models.
The pattern: experts are getting leverage, and the debate is how much credit the tool deserves versus the person steering it.
So what?
The edge goes to people with deep domain knowledge who can direct a model and verify its output. Write down your conventions and tests, because they are now input to the model as much as guidance for humans. Hire for judgment and taste, not typing speed.
Read these
Code is clay now
Faster than earcut: triangulating 3.7M map polygons in 489 ms
Jailbreaking the Switch 2
Rust Style Guide
WorldTUI: I trained a 360M model to pretend to be a terminal