Comment by BenoitP

For context: Goodfire has expertise in interpretability through a Sparse Auto Encoder technique of their own

https://www.goodfire.com/blog/a-practical-guide-to-saes