We also think that honesty is useful as a first step – for example, if we could build honest systems we could use them to conduct research into other aspects of alignment without a risk of research sabotage.
I’ve made some steps towards this, with a technique for steering toward honesty via an adapter optimiser on internal representations. It has limitations (seed variance), but It’s also a method with some nice properties for alignment debugging (self-supervised, inner), and was designed for this exact purpose, so it may be of interest.
I’ve made some steps towards this, with a technique for steering toward honesty via an adapter optimiser on internal representations. It has limitations (seed variance), but It’s also a method with some nice properties for alignment debugging (self-supervised, inner), and was designed for this exact purpose, so it may be of interest.