Goodfire researcher Curt Tigges demonstrates that individual neurons in large language models are directly interpretable and act as functional control gates. His team shows that models can shed ninety percent of active neurons while maintaining strong performance.
FireBrief
FireBrief