I was on the first SAE paper in 2023 and have been pursuing solving ambitious mech interp ever since. Recently I've come across tensor-transformer variant as a more interpretable architecture that's ~as performant. I believe this is one of the most underrated research directions.