New research challenges a common idea of how ML algorithms "think" about certain tasks
There’s a long-standing idea that as raw inputs get transformed to these intermediate representations, the system is trading prediction for compression, and building higher-level concepts through this information bottleneck.