Tag
This paper introduces 'Rosetta Neurons'—universal neurons across diverse neural networks—and shows they scale as a sublinear power law, becoming more selective and monosemantic with scale, enabling data filtering that nearly matches oracle performance.