this article is really interesting and got me thinking about the limitations of current language models. i'm curious to see how this approach could unlock new capabilities while still preserving safety and control.