What is (and isn’t) gained by avoiding architectures with high opaque serial depth?

The post discusses the implications of avoiding architectures with high opaque serial depth, a metric for measuring opaque serial cognition in models. While low opaque serial depth helps with model oversight, it doesn't guarantee transparency, as some harmful actions don't require deep opaque reasoning, and models might still use incomprehensible reasoning. Current training methods risk creating 'garbled' reasoning processes that are hard for humans to interpret.