Balanced and well-written article on 'AI consciousness and model welfare' in @IEEESpectrum by Eliza Strickland @newsbeagle, with input from me, @camhberg, @mustafasuleyman, & @mmitchell_ai. One small amend - I *do* think that research on functional emotional vectors (&c) can tell us many useful things helpful for interpretability, prediction of model behaviour, and control. I just think that this research doesn't have much to say about 'AI consciousness' because the presence of emotional vectors (&c) seems very much to be expected given the architecture/training data/learning paradigm. In other words there are readily available deflationary explanations for these interesting (and useful) observations which don't involve consciousness at all, and are more parsimonious and plausible because of this. Anyway, quibble aside, here's the link: