Is this in any way similar to Goodfire's work? https://www.goodfire.ai/research/rlfr#