Code is, but running application aren't "math". They have side effects - like uploading your private keys to hostile servers - which "math" conveniently ignores.
This is an underappreciated possibility that opened up. It now becomes at least viable to thoroughly inspect large code bases. It's not easy and never will be, but now we're not limited to the amound of literal eyeballs available.
Be careful of prompt injection. Some people think nothing of adding harmful instructions to their repositories. Sol and Opus seem to be smart enough not to fall for it, but I don't know about other models.
Just a friendly reminder that input device drivers are among the most security sensitive piece of code you can imagine. Literally everything you do goes through them.
That's the very reason I vibed my own daemon. I don't want the Logitech binary blob on my system. It needs too many permissions, it's huge, hogs memory and telemetry built in.
My vibe coded daemon is just a few hundred LOC of go code I can inspect, modify and trust.
A "nice site" was a proxy metric for "thoughtful competence". It was always flawed, but LLMs put the final nail in that particular coffin.
bellard.org style text-only, get-to-the-point-please content will the way forwards. "browsers" can then style and rewrite it till the cows come home - with fancy slideshows, LLMisms, carrousels, the works - or not and nobody will be worse off.
The way we structure content these days is insane. Separate content from presentation, please.
