1 pointby scaledsystems4 hours ago1 comment
  • scaledsystems4 hours ago
    Large language models can keep producing confident, well-structured answers even when they’re wrong. Because mistakes don’t slow them down or force revision, errors tend to accumulate quietly instead of triggering correction. This piece explains why that happens and why fluency alone is a weak indicator of alignment.