Researcher Queries Dimensionality of Model Misalignment

Herbie Bradley, a Cambridge PhD student, posted a reply on X addressing @_sholtodouglas and @TheStalwart. He asked how much a referenced point aligns with the conclusion from the most recent alignment blog post. That conclusion states misalignment is high-dimensional and a misaligned model may be an outlier in only a few dimensions. Bradley also asked whether such misalignment is expected to become more low-dimensional over time. The post includes a generated headline fragment on the topic of AI alignment discussion and the dimensionality of model misalignment.