lol fuck Meta, but also, LLMs aren't better at coding than you

#software-dev #on-the-internet

Here's a thread of Bluesky posts from Craft Conference 2026 that opens with a story about Meta: https://bsky.app/profile/hillelwayne.com/post/3mnh5d5sovc2m

While the story itself is obviously bonkers, I'm somewhat haunted by this final paragraph:

Meta is an extreme case but things have changed a lot in the past six months. Many senior engineers feel like models are now better at coding than them.

I think devs too frequently forget that LLMs are simply faster at coding than we are. But they don't have intelligence, and they can't understand things, so they'll never be able to quickly generate code that understands all of the real-world shit surrounding software that humans understand intuitively, such as, "Maybe my support chatbot should not be allowed to change accounts' email addresses willy-nilly."

Too many AI-pilled devs encounter superficially impressive code output and react with, "Wow, I've never heard of that API/algorithm/syntactic sugar/etc. before; this AI is so smart." To me, that's the same as opening an encyclopedia and thinking, "Wow, this book knows everything! It's so smart!"

Software is, tautologically, the collection of code that composes it. But software is also -- and I'd argue this is more important -- the code that doesn't compose it. Like the code that lets support chatbots change random accounts' email addresses upon request, for example. Maybe generating as much code as possible as fast as possible isn't great for answering the "what should this software not do?" question.