I've been working on a project that started with what I thought was a transformer problem. People usually talk about attention sinks, representation collapse, low-rank activations, weird key norm distributions, etc. as separate attention pathologies that need separate fixes. I don't think they're ac