“We were looking at it wrong”: The explosive claim that’s upended the AI world

Explain Like I'm Normal

Goodfire researcher Nick Cammarata published a massive manifesto on mechanistic interpretability, arguing that neural networks aren't inscrutable black boxes but structured systems we can finally map. By using Sparse Autoencoders, researchers can now isolate specific 'concepts' within models, potentially solving the alignment and safety problems by actually seeing how a model 'thinks'.