Stop reading every line of AI-written code. You can't keep up, and reading isn't understanding anyway. The Meta engineer who said it all feels like a black box skipped the step that matters, the design. I decide the interfaces and the plan, and the model fills in the bodies. Then I know what each part is for without reading it twice. Then I make it write the tests I'd have written, and nothing merges without a check from something that didn't write the code. In my own Claude Code setup that's a review gate: a second pass reads the diff cold and has to say yes, and a commit hook blocks secrets and client names before they leave my machine. The black box isn't the tool. It's what you get when the model writes the design as well as the code. If you let it do both, how did the code look three months in?