The future might not have source files

For decades, we optimized code for humans. We said that we write code for other humans to read, not for the machines.

Meaningful names, clean abstractions, consistent formatting, comments when necessary. The assumption was simple: someone else, or future you, would eventually read it.

Even when LLMs first appeared, that assumption mostly held. I'm no longer convinced that's where we're heading.

A good example is Google's Antigravity IDE. It started as an AI-powered IDE. Today, it has largely removed the editor and focuses on reviewing what the AI produced. The interface itself reflects a shift: less writing, more verification.

Another instance is that people are already experimenting with asking LLMs to generate assembly directly, skipping human-readable source code altogether. If humans never need to touch the code, why optimize it for us?

We're not there yet.

Human-readable code still has real advantages. It costs fewer tokens for both reading and writing, it's easier to inspect when something goes wrong, and manual debugging is still invaluable. Today's AI also benefits from well-structured source code.

But the trend is interesting. As models become better at generating, understanding, and fixing software, human-readable code may become more of a convenience than a requirement.

Interesting times; let's see where this trend leads us.