Is AI interpretability close to a breakthrough? A new essay has researchers divided

An interpretability researcher argues that neural networks are already much cleaner than the field assumed, with neurons standing for recognizable features. Nick Cammarata reinterprets the famous cat-car neuron as a detector for pointed shapes, and blames the apparent mess on old models and weak tools. Other researchers are divided on whether the claim holds.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.